BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:A Hierarchical Statistical Model to Track the Performance of a Dis
tributed Industrial Fleet
DTSTART:20230913T082000Z
DTEND:20230913T084000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-582@conferences.enbis.org
DESCRIPTION:Speakers: Ignasi Puig-de-Dou (Statistics and Operations Reseac
h Dpt. Escola Tècnica Superior d'Enginyeria Industrial de Barcelona. Univ
ersitat Politècnica de Catalunya)\, Xavier Puig-Oriol (Statistics and Ope
rations Research Department. Escola Tècnica Superior d'Enginyeria Industr
ial de Barcelona. Universitat Politècnica de Catalunya)\n\nThe research p
resented showcases a collaboration with a leading printer manufacturer to
facilitate the remote monitoring of their industrial printers installed at
customer sites. The objective was to create a statistical model capable o
f automatically identifying printers experiencing more issues than expecte
d based on their current operating conditions. To minimize the need for ex
tensive data collection\, a unified model was developed for all printers\,
using a hierarchical approach. By incorporating a hierarchical set of ra
ndom effects\, information sharing among the installed printer base was en
abled\, while also accounting for each printer's unique characteristics. T
he model was implemented using a Bayesian framework\, enabling automatic i
dentification of out-of-control situations.\n\nhttps://conferences.enbis.o
rg/event/32/contributions/582/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/582/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Brownie Bee: An Appetizing Way to Implement Bayesian Optimization
in Companies
DTSTART:20230911T113000Z
DTEND:20230911T115000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-473@conferences.enbis.org
DESCRIPTION:Speakers: Morten Bormann Nielsen (Danish Technological Institu
te)\n\nDesign of Experiments (DOE) is a powerful tool for optimizing indus
trial processes with a long history and impressive track record. However\,
despite its success in many industries\, most businesses in Denmark still
do not use DOE in any form due to a lack of statistical training\, prefer
ence for intuitive experimentation\, and misconceptions about its effectiv
eness.\n\nTo address this issue\, the Danish Technological Institute has d
eveloped *Brownie Bee*\, an open-source software package that combines Bay
esian optimization with a simple and intuitive user interface. Bayesian op
timization uses a more iterative approach to solve DOE tasks than classic
designs but is much easier for non-expert users. The simple interface serv
es to sneak Bayesian optimization through the front door of companies that
need it the most\, particularly those with low digital maturity.\n\nIn th
is talk\, I will explain why Bayesian optimization is an excellent alterna
tive and supplement to traditional DOE\, particularly for companies with m
inimal statistical expertise. During the talk\, I will showcase the tool *
Brownie Bee* and share insights from case studies where it has been succes
sfully implemented in 15 Danish SMEs.\n\nJoin me to discover how you can i
ncorporate Bayesian optimization through *Brownie Bee* into your DOE toolb
ox for process optimization and achieve better results faster compared to
traditional DOE designs.\n\nhttps://www.browniebee.dk/\n\nhttps://conferen
ces.enbis.org/event/32/contributions/473/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/473/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Building Blocks for a Data Driven Organization
DTSTART:20230911T090000Z
DTEND:20230911T093000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-469@conferences.enbis.org
DESCRIPTION:Speakers: María Cristina Fernández (Grupo Santander)\n\nAppl
ying Machine Learning techniques in our business require several elements
beyond the Statistics and Math. The main building blocks which would enabl
e a real deployment and use of Machine Learning commonly imply data and s
tatistics but also expert teams\, technology\, frames\, tools\, governance
\, regulation and processes\, amongst other. Expert data scientists knowin
g the limits of the algorithms and data\, the nature of the problem and th
e optimal fit into the business are essential\, jointly to technology and
tools adequacy and a deep understanding of the process design. Ethical and
fair playing fields must be ensured by a prompt and timely regulation\, c
hanneled through the right internal governance in the companies. Therefore
\, numerous challenges are faced by the industry when transposing the late
st solutions on Machine Learning and Artificial Intelligence to become a D
ata Driven organization.\n\nhttps://conferences.enbis.org/event/32/contrib
utions/469/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/469/
END:VEVENT
BEGIN:VEVENT
SUMMARY:New Estimation Algorithm for More Reliable Prediction in Gaussian
Process Regression: Application to an Aquatic Ecosystem Model
DTSTART:20230913T071000Z
DTEND:20230913T073000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-491@conferences.enbis.org
DESCRIPTION:Speakers: Amandine MARREL (CEA)\, Bertrand IOOSS (EDF R&D)\n\n
In the framework of emulation of numerical simulators with Gaussian proces
s (GP) regression [1]\, we proposed in this work a new algorithm for the e
stimation of GP covariance parameters\, referred to as GP hyperparameters.
The objective is twofold: to ensure a GP as predictive as possible w.r.t.
to the output of interest\, but also with reliable prediction intervals\,
i.e. representative of its prediction error. \nTo achieve this\, we propo
se a new constrained multi-objective algorithm for the hyperparameter esti
mation. It jointly maximizes the likelihood of the observations as well as
the empirical coverage function of GP prediction intervals\, under the co
nstraint of not degrading the GP predictivity [2]. Cross validation techni
ques and advantageous update GP formulas are notably used. \nThe benefit b
rought by the algorithm compared to standard algorithms is illustrated on
a large benchmark of analytical functions (up to twenty input variables).
An application on a EDF R&D real data test case modeling an aquatic ecosys
tem is also proposed: a log-kriging approach embedding our algorithm is im
plemented to predict the biomass of the two species. In the framework of t
his particular modeling\, this application shows the crucial interest of w
ell-estimated and reliable prediction variances in GP regression.\n\n[1] M
arrel et al. (2022). The ICSCREAM Methodology: Identification of Penalizin
g Configurations in Computer Experiments Using Screening and Metamodel. Ap
plications in Thermal Hydraulics. Nucl. Sci. Eng.\, 196(3):301–321.\n\n[
2] Demay et al. (2022). Model selection for Gaussian process regression: a
n application with highlights on the model variance validation. Qual. Reli
ab. Eng. Int.\, 38:1482-1500.\n\nhttps://conferences.enbis.org/event/32/co
ntributions/491/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/491/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Self-Validated Ensemble Models (SVEM) – Machine Learning for Sma
ll Data Typical of Industrial Designed Experiments
DTSTART:20230912T150000Z
DTEND:20230912T152000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-522@conferences.enbis.org
DESCRIPTION:Speakers: Christopher Gotwalt (JMP Statistical Discovery LLC)\
n\nSelf-Validating Ensemble Modeling (S-VEM) is an exciting\, new approach
that combines machine learning model ensembling methods to Design of Expe
riments (DOE) and has many applications in manufacturing and chemical proc
esses. In most applications\, practitioners avoid machine learning method
s with designed experiments because often one cannot afford to hold out ru
ns for a validation set without fundamentally changing the aliasing struct
ure of the design. We present a technique that fractionally allocates rows
to training and validation sets that makes machine learning model selecti
on possible for the small datasets typical in DoE applications. The approa
ch with S-VEM is similar to Random Forests ™ except that instead of aver
aging a set of resampling-based bootstrap decision tree models\, one avera
ges fractional-random-weight bootstrap linear models whose effects have be
en chosen using forward selection or the Lasso. In this way\, we are able
to retain the interpretability of response surface models\, while being ab
le to obtain greater accuracy as well as fit models that have more paramet
ers than observations. Although our investigations have only applied the S
-VEM model averaging technique to linear least squares models\, the algori
thm is quite general and could be applied to generalized linear models\, a
s well as other machine learning methods like neural networks or decision
trees. We will present simulation results comparing independent test set a
ccuracy of S-VEM to more traditional approaches to modeling DoE data and i
llustrate the method with case studies.\n\nhttps://conferences.enbis.org/e
vent/32/contributions/522/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/522/
END:VEVENT
BEGIN:VEVENT
SUMMARY:dPCA: A Python Library for Dynamic Principal Component Analysis
DTSTART:20230913T094500Z
DTEND:20230913T100500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-428@conferences.enbis.org
DESCRIPTION:Speakers: murat kulahci (DTU)\, Sebastian Topalian (Technical
University of Denmark)\n\nAnalysis of dynamical systems often entails cons
idering lagged states in a system which can be identified by heuristics or
brute-force for small systems\, however for larger and complex plantwide
systems these approaches become infeasible. We present the Python package\
, dPCA\, for performing dynamic principal component analysis as described
by Vanhatalo et al.. Autocorrelation and partial autocorrelation matrices
can be constructed for which eigen decomposition can reveal important lags
in terms of large eigenvalues and subsequently which variables are highly
correlated across time in terms of eigenvector coefficients. \nTwo use ca
ses are presented – one employing synthetic timeseries data to demonstra
te a direct connection to ARMA systems\, and one employing two datasets fr
om the largest industrial wastewater treatment plant in Northern Europe. T
he second use case demonstrates a low-cost tool for analysing large system
dynamics which can be used for initial feature engineering for supervised
prediction tasks at the plant. The two datasets present different plant l
ayouts utilising different flow schemes\, and the approach and Python pack
age is then used to find delays between upstream production plants and dow
nstream operations. \n\nFinally\, a perspective is given on how the packag
e can be applied for identifying which lags to use for statistical process
monitoring as well as future work.\n\n\nE. Vanhatalo\, M. Kulahci and B.
Bergquist\, On the structure of dynamic principal component analysis used
in statistical process monitoring\, Chemometrics and Intelligent Laborator
y Systems. 167 (2017) 1-11. https://doi.org/10.1016/j.chemolab.2017.05.016
\n\nhttps://conferences.enbis.org/event/32/contributions/428/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/428/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Statistical Learning in Reproducing Kernel Hilbert Spaces
DTSTART:20230913T080000Z
DTEND:20230913T082000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-577@conferences.enbis.org
DESCRIPTION:Speakers: Ambrus Tamás (ELTE)\n\nKernel methods are widely us
ed in nonparametric statistics and machine learning. In this talk kernel m
ean embeddings of distributions will be used for the purpose of uncertaint
y quantification. The main idea of this framework is to embed distribution
s in a reproducing kernel Hilbert space\, where the Hilbertian structure a
llows us to compare and manipulate the represented probability measures. W
e review some of the existing theoretical results and present new applicat
ions of this powerful tool. Distribution‐free\, nonparametric results wi
ll be introduced for supervised learning problems (classification and regr
ession).\n\nhttps://conferences.enbis.org/event/32/contributions/577/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/577/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Transfer Learning across Biopharma Molecules\, Scales and Phases B
ased on Hybrid Semi-Parametric Modeling
DTSTART:20230913T065000Z
DTEND:20230913T071000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-655@conferences.enbis.org
DESCRIPTION:Speakers: Hugo Santos (DataHow)\, Moritz von Stosch (DataHow A
G)\n\nQuality by Design (QbD) guided process development is time and cost-
effective only if knowledge is transferred from candidate to the next\, fr
om one scale to the other.\n \nNowadays\, knowledge is shared across scale
s and candidates via technical risks evaluation. Though platform processes
are widely used\, this type of knowledge transfer is limited and every ne
w candidate requires some degree of process development from scratch\, lea
ving significant potential to accelerate process development.\n \nIn this
contribution\, we show how transfer machine-learning and hybrid modeling a
pproaches can be exploited to transfer knowledge between scales\, unit ope
rations and molecules for a number of examples from mAb processes\, Raman
and cell & gene therapies. Further\, we present a novel method for incorpo
rating information about similar processes into the model generation to su
pport selecting the ideal design for a new process in development. We will
also highlight the importance of creating a concrete\, standardized\, and
self-learning ecosystem so that all parties involved in process developme
nt and tech transfer may take benefit from such model-derived knowledge.\n
\nhttps://conferences.enbis.org/event/32/contributions/655/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/655/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Incremental Designs for Simultaneous Kriging Predictions Based on
the Generalized Variance as Criterion
DTSTART:20230913T071000Z
DTEND:20230913T073000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-494@conferences.enbis.org
DESCRIPTION:Speakers: Helmut Waldl (Johannes Kepler University Linz)\n\nIn
this talk\, the problem of selecting a set of design points for universal
kriging\,\nwhich is a widely used technique for spatial data analysis\, i
s further\ninvestigated. The goal is to select the design points in order
to make simultaneous\npredictions of the random variable of interest at a
finite number of\nunsampled locations with maximum precision. Specifically
\, a correlated random\nfield given by a linear model with an unknown para
meter vector and\na spatial error correlation structure is considered as r
esponse. A new design\ncriterion that aims at simultaneously minimizing th
e variation of the prediction\nerrors at various points is proposed. There
is also presented an efficient\ntechnique for incrementally building desi
gns for that criterion scaling well\nfor high dimensions. Thus the method
is particularly suitable for big data\napplications in areas of spatial da
ta analysis such as mining\, hydrogeology\,\nnatural resource monitoring\,
and environmental sciences or equivalently\nfor any computer simulation e
xperiments. The effectiveness of the proposed\ndesigns is demonstrated thr
ough a numerical example.\n\nhttps://conferences.enbis.org/event/32/contri
butions/494/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/494/
END:VEVENT
BEGIN:VEVENT
SUMMARY:ECAS-ENBIS Course: Conformal Prediction: How to Quantify Uncertain
ty of Machine Learning Models?
DTSTART:20230910T120000Z
DTEND:20230910T160000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-415@conferences.enbis.org
DESCRIPTION:Speakers: Margaux Zaffran (INRIA)\n\nhttps://conferences.enbis
.org/event/41/\n\nhttps://conferences.enbis.org/event/32/contributions/415
/
LOCATION:2.11
URL:https://conferences.enbis.org/event/32/contributions/415/
END:VEVENT
BEGIN:VEVENT
SUMMARY:ENBIS Live - Open Problem Session
DTSTART:20230912T074000Z
DTEND:20230912T084000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-622@conferences.enbis.org
DESCRIPTION:Speakers: Christian Ritter (Ritter and Danielson Consulting)\n
\nThis is a session in which we discuss two or three open problems propose
d by conference participants. As usual\, Chris Ritter will lead this sessi
on. He is looking for fresh cases for this session:\n\nCALL FOR VOLUNTEERS
\n\nWe need volunteers who have current open problems and would like to pr
esent them at this session.\n\nYou will present for about 5-7 minutes to d
escribe the context and the question/problem\n\nAfter that\, Chris will fa
cilitate several rounds of questions/suggestions. The curiosity and expert
ise of the other participants will then give you some new ideas.\n\nAre yo
u interested to propose a project? Then send him an email (ritter.christia
n@ridaco.be)\n\nhttps://conferences.enbis.org/event/32/contributions/622/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/622/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Predicting Indocyanine Green Retention at 15 Minutes (ICG15) in He
patocellular Carcinoma Patients Using Radiomics and Hematology
DTSTART:20230913T092500Z
DTEND:20230913T094500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-591@conferences.enbis.org
DESCRIPTION:Speakers: Jakey Blue (National Taiwan University)\, Ming-Chih
Ho (Department of Surgery\, National Taiwan University Hospital and Colleg
e Medicine\, National Taiwan University)\, Chih-Horng Wu (Department of Me
dical Imaging and Radiology\, National Taiwan University Hospital and Coll
ege Medicine\, National Taiwan University)\, Pei-Chun (Zoey) Chao (Institu
te of Industrial Engineering\, National Taiwan University)\n\nHepatocellul
ar carcinoma (HCC) poses significant challenges and risks globally. Liver
metabolism assessment\, reflected in Indocyanine Green Retention at 15 min
utes (ICG15)\, is crucial for HCC patients. This study aimed to predict IC
G15 levels using radiomics-based features and selected hematology test res
ults. A hybrid predictive model combining clustering and stacking models i
s developed to enhance ICG15 prediction precision.\n\nA total of 120 HCC p
atients were enrolled\, with 107 patients included after outlier handling.
Dimension reduction using the Least Absolute Shrinkage and Selection Oper
ator (LASSO) identified the 30 most influential predictors for subsequent
investigation. Gaussian Mixture Model (GMM) clustering was then employed t
o categorize patients into two groups based on radiomics and hematology fe
atures. Subsequently\, a stacking framework is built\, with XGBoost servin
g as the base model and XGBoost\, AdaBoost\, RandomForest\, and SVM regres
sor as the four meta-learners. Our research underscores the significance o
f integrating radiomics and machine learning models in treating liver canc
er. By improving the predictive accuracy of ICG15\, our model holds the po
tential to serve as a valuable tool for physicians in the preoperative eva
luation of liver function\, thus benefiting HCC patients.\n\nhttps://confe
rences.enbis.org/event/32/contributions/591/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/591/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Batch Manufacturing Datasets - Open Source Data for Academia and I
ndustry
DTSTART:20230913T063000Z
DTEND:20230913T065000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-426@conferences.enbis.org
DESCRIPTION:Speakers: Daniel Palací-López (IFF Benicarló )\, Francisco
Navarro (Imperial College)\, Mattia Vallerio (Solvay SA)\, Carlos Perez-Ga
lvan (Solvay SA)\, Philippe Neyraval (Solvay SA)\, Benjamin Katz (Solvay S
A)\n\nMachine Learning is now part of many university curriculums and indu
strial training programs. However\, the examples used are often not releva
nt or realistic for process engineers in manufacturing.\n\nIn this work\,
we will share a new industrial batch dataset and make it openly available
to other practitioners. We will show how batch processes can be challengin
g to analyze when having sources of information containing quality\, event
s\, and sensor data (tags). We will also introduce machine-learning techni
ques for troubleshooting and detecting anomalous batches at a manufacturin
g scale.\n\nhttps://conferences.enbis.org/event/32/contributions/426/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/426/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Big Behavioral Data - How Machine Learning Made Students Learn Mor
e DOE
DTSTART:20230912T163000Z
DTEND:20230912T165000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-567@conferences.enbis.org
DESCRIPTION:Speakers: John Tyssedal (NTNU)\n\nSome years ago the largest b
ank in our region came to the university and offered project and master th
esis on bank related problems and huge data sets. This was very well recei
ved by students and it became an arena for learning and job-related activi
ty. The students got practice in working with imbalanced data\, data pre-p
rocessing\, longitudinal data\, feature creation/selection and hyperparame
ter tuning. The presentation will focus on the lesson learned and in parti
cular the last item looking into methods like grid search\, random search\
, Bayesian optimization and DOE. Some advantages of using DOE will be high
lighted.\n\nhttps://conferences.enbis.org/event/32/contributions/567/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/567/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Recent Developments on Distribution-Free Phase-I Monitoring - An
Overview and Some New Results
DTSTART:20230913T090500Z
DTEND:20230913T092500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-543@conferences.enbis.org
DESCRIPTION:Speakers: Amitava Mukherjee (XLRI -Xavier School of Management
)\n\nPhase-I monitoring plays a vital role as it helps to analyse the proc
ess stability retrospectively using a set of available historical samples
and obtaining a benchmark reference sample to facilitate Phase-II monitori
ng. Since\, at the very beginning process state and its stability is unkno
wn\, trying to assume a parametric model to the available data (which coul
d be well-contaminated) is unwarranted\, and to this end\, nonparametric p
rocedures are highly recommended. Earlier research on nonparametric Phase-
I monitoring was primarily confined to monitoring location\, scale\, or jo
int location-scale parameters. Recent developments have suggested includin
g skewness or kurtosis aspects as well in monitoring. The current paper gi
ves a broad overview of various available charts and offers some new resul
ts on the adaptive choice between the charts when nothing is known. Some i
ndustrial applications are discussed.\n\nhttps://conferences.enbis.org/eve
nt/32/contributions/543/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/543/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Spectral Methods for SPC of 3-D Geometrical Data
DTSTART:20230911T115000Z
DTEND:20230911T121000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-544@conferences.enbis.org
DESCRIPTION:Speakers: Yulin An (Penn State U)\, Xueqi Zhao (Google)\, Enri
que del Castillo (Penn State University)\n\nWe present a summary of recent
ly developed methods for the Statistical Process Control of 3-dimensional
data acquired by a non-contact sensor in the form of a mesh. The methods h
ave the property of not requiring ambient coordinate information\, and use
only the intrinsic coordinates of the points on the meshes\, hence not ne
eding the preliminary registration or alignment of the parts. Intrinsic sp
ectral SPC methods have been developed for both Phase I (or startup phase)
and Phase II (or on-line control). In addition\, we review recently devel
oped spectral methods for the localization of defects on the surface of a
part deemed out of control that do not require registration of the part an
d nominal geometries.\n\nhttps://conferences.enbis.org/event/32/contributi
ons/544/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/544/
END:VEVENT
BEGIN:VEVENT
SUMMARY:AI's Adventures in Batchland: A Case Study in Massive Batch Proces
sing
DTSTART:20230911T093000Z
DTEND:20230911T100000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-584@conferences.enbis.org
DESCRIPTION:Speakers: Iñaki Ucar (Universidad Carlos III de Madrid)\n\nWe
often think of digitalization as the application of complex machine learn
ing algorithms to vast amounts of data. Unfortunately\, this raw material
is not always available\, and\, in particular\, many traditional businesse
s with well-established processes accumulate a large technical debt that i
mpedes progress towards more modern paradigms. In this talk\, we review a
complete case study\, from data collection to production deployment\, comb
ining old and new techniques for monitoring and optimizing massive batch p
rocessing.\n\nhttps://conferences.enbis.org/event/32/contributions/584/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/584/
END:VEVENT
BEGIN:VEVENT
SUMMARY:The Case against Generally Weighted Moving Average (GWMA) Control
Charts
DTSTART:20230911T100000Z
DTEND:20230911T103000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-600@conferences.enbis.org
DESCRIPTION:Speakers: William H. Woodall\, Sven Knoth (Helmut Schmidt Uni
versity Hamburg\, Germany)\, Víctor G. Tercero-Gómez\n\nWe argue against
the use of generally weighted moving average (GWMA) control charts. Our p
rimary reasons are the following: 1) There is no recursive formula for the
GWMA control chart statistic\, so all previous data must be stored and us
ed in the calculation of each chart statistic. 2) The Markovian property d
oes not apply to the GWMA statistics\, so computer simulation must be used
to determine control limits and the statistical performance. 3) An approp
riately designed\, and much simpler\, exponentially weighted moving averag
e (EWMA) chart provides as good or better statistical performance. 4) In s
ome cases the GWMA chart gives more weight to past data values than to cur
rent values.\n\nhttps://conferences.enbis.org/event/32/contributions/600/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/600/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Scalar-On-Function Regression Control Chart Based on a Functional
Neural Network
DTSTART:20230913T094500Z
DTEND:20230913T100500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-548@conferences.enbis.org
DESCRIPTION:Speakers: Gianluca Sposito (Department of Industrial Engineeri
ng\, University of Naples Federico II\, Naples\, Italy)\, Biagio Palumbo (
Department of Industrial Engineering\, University of Naples Federico II\,
Naples\, Italy)\, Antonio Lepore (Department of Industrial Engineering\, U
niversity of Naples Federico II\, Naples\, Italy)\, Giuseppe Giannini (Hea
d CBM Tool and Data Analysis\, Hitachi Rail Italy\, Naples\, Italy)\, Mura
t Kulahci (Department of Applied Mathematics and Computer Science\, Techn
ical University of Denmark\, Kongens Lyngby\, Denmark\; Department of Busi
ness Administration\, Technology and Social Sciences\, Luleå University o
f Technology\, Luleå\, Sweden)\n\nModern data acquisition systems allow f
or collecting signals that can be suitably modelled as functions over a co
ntinuum (e.g.\, time or space) and are usually referred to as *profiles* o
r *functional data*. Statistical process monitoring applied to these data
is accordingly known as *profile monitoring*. The aim of this research is
to introduce a new profile monitoring strategy based on a *functional* neu
ral network (FNN) that is able to adjust a scalar quality characteristic f
or any influence by one or more covariates in the form of functional data.
FNN is the name for a neural network able to learn a possibly nonlinear r
elationship which involves functional data. \nA Monte Carlo simulation stu
dy is performed to assess the monitoring performance of the proposed contr
ol chart in terms of the out-of-control average run length with respect to
competing methods that already appeared in the literature before. Further
more\, a case study in the railway industry\, courtesy of Hitachi Rail Ita
ly\, demonstrates the potentiality and practical applicability in industri
al scenarios.\nAcknowledgements: This study was carried out within the MOS
T – Sustainable Mobility National Research Center and received funding f
rom the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RES
ILIENZA (PNRR) – MISSIONE 4 COMPONENTE 2\, INVESTIMENTO 1.4 – D.D. 103
3 17/06/2022\, CN00000023). This manuscript reflects only the authors’ v
iews and opinions\, neither the European Union nor the European Commission
can be considered responsible for them. This research was also partially
supported by the Danish Data Science Academy.\n\nhttps://conferences.enbis
.org/event/32/contributions/548/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/548/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Utilizing Individual Clear Effects for Intelligent Factor Allocati
ons and Design Selections
DTSTART:20230911T093000Z
DTEND:20230911T100000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-613@conferences.enbis.org
DESCRIPTION:Speakers: Hongquan Xu\, William Li (Shanghai Advanced Institut
e of Finance)\, Qi Zhou\n\nExtensive studies have been conducted on how t
o select efficient designs with respect to a criterion. Most design criter
ia aim to capture the overall efficiency of the design across all columns.
When prior information indicated that a small number of factors and their
two-factor interactions (2fi's) are likely to be more significant than ot
her effects\, commonly used minimum aberration designs may no longer be th
e best choice. Motivated by a real-life experiment\, we propose a new clas
s of regular fractional factorial designs that focus on estimating a subse
t of columns and their corresponding 2fi's clear of other important effect
s. After introducing the concept of individual clear effects (iCE) to desc
ribe clear 2fi's involving a specific factor\, we define the clear effect
pattern criterion to characterize the distribution of iCE's over all colum
ns. We then obtain a new class of designs that sequentially maximize the c
lear effect pattern. These newly constructed designs are often different f
rom existing optimal designs. We develop a series of theoretical results t
hat can be particularly useful for constructing designs with large run siz
es\, for which algorithmic construction becomes computationally challengin
g. We also provide some practical guidelines on how to choose appropriate
designs with respect to different run size\, the number of factors\, and t
he number of 2fi's that need to be clear.\n\nhttps://conferences.enbis.org
/event/32/contributions/613/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/613/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Applied Research as a Tool to Influence Policy
DTSTART:20230912T113500Z
DTEND:20230912T120500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-559@conferences.enbis.org
DESCRIPTION:Speakers: Daphna Aviram Nitzan (Israel Democracy Institute)\n\
nMy main goals\, as the Director of the Center for Governance and the Econ
omy in the Israel Democracy Institute is to initiate\, lead and manage app
lied research and to professionally analyze the key developments within Is
raeli economy\, society and labor market. I work towards achieving these
goals on several tracks: \n\n1. Recruiting a team of talented professional
s\, experts in data analysis. .\n2. Constructing a unique database contain
ing anonymized administrative data from various government bodies\, mainly
from the ICBS (including demographic information\, educational and employ
ment history). The database is updated on an ongoing basis to ensure that
it includes the most recent data available.\n3. Using cutting edge resear
ch and advanced programming capabilities (including Machine Learning techn
iques\, Prediction Models and Big Data algorithms) while applying various
empirical methods. \n4. Adapting new research strategies\, and when necess
ary\, acquiring external data (in addition to the ICBS data) from data-c
ollecting bodies\n5. Providing intensive professional guidance to the youn
g data analysis researchers\, making sure they focus on an applied policy
recommendation. \n6. Maintaining a professional and unbiased approach to t
he research projects\, with an emphasis on professional integrity. \n7. Id
entifying relevant partners in the government\, business sector\, academia
and the civil service interested in the research outcomes and open to ada
pting our policy recommendation. In order to gain partners' collaboration
\, we usually establish a think tank. that accompanies the research and so
metimes serves as the research steering committee. \nOur research has led
to the completion of many project on the ground\, alongside the publishin
g of the following research papers: Intergenerational Mobility among Popul
ations in Israel\, The Evolving Tasks and Skills Necessary in the Israeli
Job Market. Return of the COVID-Unemployed to Work\n\nhttps://conferences.
enbis.org/event/32/contributions/559/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/559/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Modelling Curve Data: Functional Data Explorer Workshop
DTSTART:20230913T123000Z
DTEND:20230913T163000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-435@conferences.enbis.org
DESCRIPTION:Speakers: Chris Gotwalt (JMP Division of SAS Institute)\, Phil
Kay (SAS)\n\nhttps://conferences.enbis.org/event/42/\n\nhttps://conferenc
es.enbis.org/event/32/contributions/435/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/435/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Modelling and Forecasting Correlated Failure Counts
DTSTART:20230912T065000Z
DTEND:20230912T071000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-601@conferences.enbis.org
DESCRIPTION:Speakers: Antonio Pievatolo (CNR-IMATI)\n\nWe present a state-
space model in which failure counts of items produced from the same batch
are correlated\, so as to be able to characterize the pattern of occurrenc
e of failures of new batches at an early stage\, based on those of older b
atches. The baseline failure rates of consecutive batches are related by a
random-walk-type equation\, and failures follow a Poisson distribution. T
he failure process determined by this model rests on few assumptions\, so
that it can be adapted to different situations. Bayesian inference and com
putation are carried out by particle filtering.\n\nhttps://conferences.enb
is.org/event/32/contributions/601/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/601/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Explainable AI Time Series Forecasting Using a Local Surrogate Mod
el
DTSTART:20230912T082000Z
DTEND:20230912T084000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-610@conferences.enbis.org
DESCRIPTION:Speakers: Florian Sobieczky (SCCH Software Competence Center H
agenberg)\, ALFREDO LOPEZ (Software Competence Center Hagenberg SCCH)\, Th
omas Wetzelmaier (Software Competence Center Hagenberg SCCH )\n\nWe introd
uce a novel framework for explainable AI time series forecasting based on
a local surrogate base model. An explainable forecast\, at a given referen
ce point in time\, is delivered by comparing the change in the base model
fitting before and after the application of the AI-model correction. The n
otion of explainability used here is local both in the sense of the featur
e space and the temporal sense. The validity of the explanation (fidelity)
is conditioned to be persistent throughout a sliding influence-window. T
he size of this window is chosen by the minimization of a loss functional
comparing the local surrogate and the AI-correction\, where we make use of
smoothing approximations of the original problem to enjoy of differentiat
ion properties. We illustrate the approach on a publicly available atmosph
eric probe dataset. The proposed method extends our method of BAPC (Before
and After correction Parameter Comparison) previously defined in the cont
ext of explainable AI regression.\n\nhttps://conferences.enbis.org/event/3
2/contributions/610/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/610/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Variance-Based Importance Index for Systems with Dependent Compo
nents
DTSTART:20230912T134500Z
DTEND:20230912T141500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-606@conferences.enbis.org
DESCRIPTION:Speakers: Jorge Navarro (Universidad de Murcia)\, M. A. Sordo
(Universidad de Cádiz)\, Antonio Arriaza-Gómez (Universidad de Cádiz)\,
Alfonso Suarez-Llorens (Universidad de Cádiz)\n\nOur work proposes a var
iance-based measure of importance for coherent systems with dependent and
heterogeneous components. The particular cases of independent components a
nd homogeneous components are also considered. We model the dependence str
ucture among the components by the concept of copula. The proposed measure
allows us to provide the best estimation of the system lifetime\, in term
s of the mean squared error\, under the assumption that the lifetime of on
e of its components is known. We include theoretical results that are usef
ul to calculate a closed-form of our measure and to compare two components
of a system. Finally\, we illustrate the main results with several exampl
es.\n\nhttps://conferences.enbis.org/event/32/contributions/606/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/606/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Statistical Model for Wildfires and the Effect of the Climate Chan
ge
DTSTART:20230912T100500Z
DTEND:20230912T103500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-476@conferences.enbis.org
DESCRIPTION:Speakers: Kristóf Halász (Eötvös Loránd University)\, And
rás Zempléni (Eötvös Loránd University\, Budapest)\n\nWe have created
a wildfire-probability estimating system\, based on publicly available da
ta (historic wildfires\, satellite images\, weather data\, maps). The math
ematical model is rather simple: kriging\, logistic regression and the boo
tstrap are its main tools\, but the computational complexity is substantia
l\, and the data analysis is challenging.\n\nIt has a wide range of applic
ations. Here we show a very interesting one: based on our model and the av
ailable possible climate change scenarios\, we are able to estimate the po
ssible damage caused by wildfires for a given region in the future. This i
s based on skilful simulation from the possible weather patterns and the d
atabase of known historic wildfires in the region and on some simplificati
ons (e.g. there are no changes in cities\, roads\, costs).\n\nThe methods
are illustrated for South American regions\, using different climate model
s. We hope that the results may contribute to the climate change awareness
and we plan to use it for European regions as well.\n\nhttps://conference
s.enbis.org/event/32/contributions/476/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/476/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Set Estimation for Dimensional Control in Shipbuilding
DTSTART:20230911T134000Z
DTEND:20230911T141000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-549@conferences.enbis.org
DESCRIPTION:Speakers: Ricardo Cao (Universidade da Coruña)\, Nataly Romar
ís Lodeiro (Universidade da Coruña y Navantia)\, Salvador Naya (Universi
dade da Coruña)\n\nWithin the framework of the Mixed Research Center (CEM
I) between the company Navantia and the University of A Coruña\, one of t
he research lines consists of using statistical methods for dimensional co
ntrol of panel production. This paper will present some advances in the us
e of set estimation for detecting singular elements in panels and determin
ing their geometric characteristics (angles between elements\, lack of fla
tness\, welding defects\, etc.)\, which allow detecting deviations with re
spect to nominal parameters and minimizing industrial reprocessing in ship
building.\n\nThere exists currently a pilot system for obtaining point clo
uds using artificial vision for inspecting dimensional control and welding
quality. The datasets (point clouds) extracted from panel scanning have a
typical size of the order of hundreds of millions of points. As a consequ
ence\, traditional set estimation methods can be very time-consuming from
a computational viewpoint. Through the use of subsampling\, nonparametric
density estimation of projections of the point cloud\, as well as modern s
et estimation techniques (such as those existing in the R package *alphash
ape*)\, efficient algorithms have been implemented that allow carrying out
dimensional quality control for manufactured panels.\n\nhttps://conferenc
es.enbis.org/event/32/contributions/549/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/549/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Bayesian Approach to Network Classification
DTSTART:20230911T090000Z
DTEND:20230911T093000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-550@conferences.enbis.org
DESCRIPTION:Speakers: Abel Rodriguez\, Sharmistha Guha (Texas A&M Universi
ty)\, RAJARSHI GUHANIYOGI (TEXAS A & M UNIVERSITY)\n\nWe propose a novel B
ayesian binary classification framework for networks with labeled nodes. O
ur approach is motivated by applications in brain connectome studies\, whe
re the overarching goal is to identify both regions of interest (ROIs) in
the brain and connections between ROIs that influence how study subjects a
re classified. We develop a binary logistic regression framework with the
network as the predictor\, and model the associated network coefficient us
ing a novel class of global-local network shrinkage priors. We perform a t
heoretical analysis of a member of this class of priors (which we call the
Network Lasso Prior) and show asymptotically correct classification of ne
tworks even when the number of network edges grows faster than the sample
size. Two representative members from this class of priors\, the Network L
asso prior and the Network Horseshoe prior\, are implemented using an effi
cient Markov Chain Monte Carlo algorithm\, and empirically evaluated throu
gh simulation studies and the analysis of a real brain connectome dataset.
\n\nhttps://conferences.enbis.org/event/32/contributions/550/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/550/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Joint Modelling of Longitudinal and Event-Time Data for the Analys
is of Longitudinal Medical Studies
DTSTART:20230912T131500Z
DTEND:20230912T134500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-583@conferences.enbis.org
DESCRIPTION:Speakers: Ruwanthi Kolamunnage-Dona\n\nJoint modelling is a mo
dern statistical method that has the potential to reduce biases and uncert
ainties due to informative participant follow-up in longitudinal studies.
Although longitudinal study designs are widely used in medical research\,
they are often analysed by simple statistical methods\, which do not fully
exploit the information in the resulting data. In observational studies\,
biomarkers are measured at irregular follow-up visit times\, and in rando
mised controlled trials\, participant dropout is common during the intende
d follow-up\; which are often correlated with patient’s prognosis. Joint
modelling combines longitudinal biomarker and event-time data simultaneou
sly into a single model through latent associations. We describe the metho
dology of joint models with some applications in health research.\n\nhttps
://conferences.enbis.org/event/32/contributions/583/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/583/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Data Science Driven Framework for Leak Detection in LNG Plants usi
ng Process Sensor Data
DTSTART:20230912T080000Z
DTEND:20230912T082000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-618@conferences.enbis.org
DESCRIPTION:Speakers: Arvind Ravi\, Stephen Varghese (Shell India Markets
Private Limited)\, Rihab Abdul Razak\, Shirish Potu\, Jose M. Gonzalez-M
artinez\, Resmi Suresh\n\nLiquefied natural gas (LNG) is a promising fuel.
However\, a major component of LNG is Methane\, which is a greenhouse gas
. Shell aims to reduce methane emissions intensity below 0.2% by 2025.\nEx
isting leak detection techniques have limitations\, such as limited covera
ge area or high cost. We explore data science driven framework using exist
ing process sensor data to localize and estimate leak magnitude. However\,
sensor noise and process changes can make leak detection challenging. Alg
orithms developed are tested on synthetic flow and composition chemical me
asurements data generated using process simulations of an LNG plant (Fern
ández\, 2015).\nWe present a leak detection and localization framework co
mprising different techniques. First the use of wavelet analysis combined
with mass balance to localize leaks\, followed by a maximum likelihood est
imation of leaks (Bakshi\, 1998). Different optimization-based approaches\
, as well as Kalman filters with fine-tuned covariance matrices\, utilizin
g mass balance\, are also being adapted to determine the potential leak ma
gnitude in each unit\, enabling confirmation of leak detection and localiz
ation using hypothesis testing. Alternatively\, statistical metrics such a
s Kantorovich distance are being explored\, coupled with classical Multiva
riate Statistical Process Control methods (Kourti and MacGregor\, 1995)\,
for the analysis of mass balance residuals at each unit to detect and loca
lize leaks\, by studying deviations in the metric (Arifin et al.\, 2018).\
nReferences\nFernández\, E.\, MS. Thesis\, NTNU\, 2015\nBakshi\, B.R.\, A
IChE journal\, 44(7):1596-1610\, 1998\nKourti\, T.\, and MacGregor\, J.F.\
, Chemom. Intell. Lab. Syst.\, 28(1):3-21\, 1995\nArifin\, B.M.S.\, et al.
\, Comput. Chem. Eng.\, 108: 300-313\, 2018\n\nhttps://conferences.enbis.o
rg/event/32/contributions/618/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/618/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Automated Registration of Polarized Light Microscopy Images Using
Deep Learning Techniques
DTSTART:20230912T100500Z
DTEND:20230912T103500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-579@conferences.enbis.org
DESCRIPTION:Speakers: Erik Blasch (Air Force Office of Scientific Research
)\, Nathan Johnston (United States Air Force)\, John Wertz (Air Force Rese
arch Laboratory)\, Bruce Cox (Air Force Institute of Technology)\, Sean O'
Rourke (Army Research Laboratory)\, Nathan Gaw (Air Force Institute of Tec
hnology)\, Matthew Cherry (Air Force Research Laboratory)\, Laura Homa (Un
iversity of Dayton Research Institute)\n\nStudies have identified a connec
tion between the microtexture regions (MTRs) found in certain titanium all
oys and early onset creep fatigue failure of rotating turbomachinery. Micr
otexture regions are defined by their size and orientation\, which can be
characterized via scanning electron microscopy (SEM) Electron Backscatter
Diffraction (EBSD). However\, doing so is impractical at the component-sca
le. A novel method of characterizing MTRs is needed to qualify new engine
components. Researchers in the Air Force Research Lab Materials and Manufa
cturing Directorate have proposed fusion of two inspection methods (eddy c
urrent testing (ECT) and scanning acoustic microscopy (SAM)) to achieve th
e goal of MTR characterization\, which proves to be a significant challeng
e to minimal literature in the area.\n\nOur research focuses on developmen
t of a Convolutional Neural Network (CNN) to automatically register two po
larized light microscopy (PLM) images. Polarized light microscopy is a sur
rogate ground-truth method that provides data similar to EBSD for this ins
pection scenario. The baseline single-modality CNN will then be adapted to
jointly train and register the SAM and ECT images for MTR characterizatio
n. The method proposed CNN in this work involves receiving two PLM images
as input\, one an unaltered copy known as the moving image (i.e.\, the ima
ge to be transformed) and the other an artificially transformed copy known
as the fixed image (i.e.\, reference for image registration). The objecti
ve of the CNN is to evaluate the moving image with the fixed image and out
put parameters to produce an affine transformation matrix that registers b
oth.\n\nhttps://conferences.enbis.org/event/32/contributions/579/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/579/
END:VEVENT
BEGIN:VEVENT
SUMMARY:The Challenges in Building Meaningful Models with Publicly Availab
le Omics Data
DTSTART:20230912T152000Z
DTEND:20230912T154000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-593@conferences.enbis.org
DESCRIPTION:Speakers: Felix Feyertag (Oxford Biomedica)\, Eva Price (Unive
rsity College London)\, Dugyu Dikicioglu (UCL)\n\nOmics data\, derived fr
om high-throughput technologies\, is crucial in research\, driving biomark
er discovery\, drug development\, precision medicine\, and systems biology
. Its size and complexity require advanced computational techniques for an
alysis. Omics significantly contributes to our understanding of biological
systems.\n\nThis project aims to construct models for Human Embryonic Kid
ney cells used in industry for viral vector production by incorporating fi
ve types of omics data: genomics\, epigenomics\, transcriptomics\, metabol
omics\, and proteomics. With over 25 terabytes of publicly available data\
, the abundances of each data type vary significantly\, including more tha
n 15\,000 sequence runs covering the genome\, epigenome\, and transcriptom
e\, as well as approximately 300 proteomics experiments and only 6 metabol
omics experiments. Skewed data availability presents challenges for integr
ative multi-omic approaches for meaningful machine learning.\n\nData gener
ation technologies have advanced rapidly\, surpassing the computational ca
pabilities required for analysis and storage. Dealing with diverse data st
ructures and varying database information requirements poses significant c
hallenges. The absence of a comprehensive data warehouse incorporating mul
tiple omics data\, with standardised quality and metadata criteria\, compl
icates information extraction from diverse sources. The persistent issue o
f missing or inadequate metadata continues to impact data collection\, cas
ting doubts on adherence to the FAIR principles and raising significant co
ncerns about the reproducibility and credibility of included studies. Impl
ementing standardised criteria and improving documentation practices acros
s databases is crucial. Addressing these challenges and developing strateg
ies for integrating and analysing publicly available omics data from multi
ple sources have immense potential to advance our understanding of complex
biological systems\, furthering innovation in industry.\n\nhttps://confer
ences.enbis.org/event/32/contributions/593/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/593/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Cost-Sensitive Classifiers for Fraud Detection
DTSTART:20230912T161000Z
DTEND:20230912T163000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-478@conferences.enbis.org
DESCRIPTION:Speakers: Jorge C. Rella (Abanca Servicios Financieros and Uni
versidade da Coruña)\, Gerda Claeskens (KU Leuven)\, Juan M. Vilar (Unive
rsidade da Coruña)\, Ricardo Cao (Universidade da Coruña)\n\nFinancial f
raud detection is a classification problem where each operation have a dif
ferent misclassification cost depending on its amount. Thus\, it fall with
in the scope of instance-dependent cost-sensitive classification problems.
When modeling the problem with a parametric model\, as a logistic regress
ion\, using a loss function incorporating the costs has proven to result i
n a more effective parameter estimation compared to classical approaches\,
which only rely on the likelihood maximization. The drawback is that this
has only been empirically demonstrated in a limited number of datasets\,
thus resulting in a lack of support for their generalized application. Thi
s work has two aims. The first is to propose cost-sensitive parameter esti
mators and develop its consistency properties and asymptotic distribution
under general conditions. The second aim is to test the cost-sensitive str
ategy over a wide range of simulations and scenarios\, testing the improve
ment obtained with the proposed cost-sensitive estimators compared to a co
st-insensitive approach.\n\nhttps://conferences.enbis.org/event/32/contrib
utions/478/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/478/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Fair Solutions in Regression Models: A Bayesian Viewpoint
DTSTART:20230912T090500Z
DTEND:20230912T093500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-513@conferences.enbis.org
DESCRIPTION:Speakers: Emilio Carrizosa (Universidad de Sevilla)\, Rafael J
imenez Llamas (Universidad de Sevilla)\, Pepa Ramirez Cobo (Universidad de
Cádiz)\n\nIn today's society\, machine learning (ML) algorithms have bec
ome fundamental tools that have evolved along with society itself in terms
of their level of complexity. The application areas of ML cover all infor
mation technologies\, many of them being directly related to problems with
a high impact on human lives. As a result of these examples\, where the e
ffect of an algorithm has implications that can radically change human bei
ngs\, there is a growing need at both the societal and institutional level
to develop fair ML tools that correct the biases present in datasets. In
this work we present a new statistical methodology that results in fair so
lutions for the classic linear and logistic regression. Our approach takes
benefit from the Bayesian paradigm\, where the use of a prior distributio
n enables to control the degree of fairness in the solution. Both Empirica
l Bayes and Variation Inference techniques are explored. The new approach
shall be illustrated through real datasets.\n\nhttps://conferences.enbis.o
rg/event/32/contributions/513/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/513/
END:VEVENT
BEGIN:VEVENT
SUMMARY:The Class of Multivariate Bernoulli Distributions with Given Ident
ical Margins
DTSTART:20230911T100000Z
DTEND:20230911T103000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-499@conferences.enbis.org
DESCRIPTION:Speakers: Roberto Fontana\, Patrizia Semeraro (Politecnico di
Torino)\n\nThe main contributions of the work (joint with P. Semeraro\, Po
litecnico di Torino) are algorithms to sample from multivariate Bernoulli
distributions and to determine the distributions and bounds of a wide clas
s of indices and measures of probability mass functions. Probability mass
functions of exchangeable Bernoulli distributions are points in a convex p
olytope\, and we provide an analytical expression for the extremal points
of this polytope. The more general class of multivariate Bernoulli distrib
utions with identical marginal Bernoulli distributions with parameter p is
also a convex polytope. However\, finding its extremal points is a more c
hallenging task. Our novel theoretical contribution is to use an algebraic
approach to find a set of analytically available generators. We also solv
e the problem of finding the lower bound in the convex order of multivaria
te Bernoulli distributions with given margins\, but with unspecified depen
dence structure.\n\nhttps://conferences.enbis.org/event/32/contributions/4
99/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/499/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Design Risk Analysis and Importance of Involving a Statistical Min
d-Set
DTSTART:20230912T071000Z
DTEND:20230912T073000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-612@conferences.enbis.org
DESCRIPTION:Speakers: Sören Knuts (GKN Aerospace Sweden)\n\nDesign Risk A
nalysis is often resembled with doing a Design Failure Mode and Effects An
alysis (DFMEA). By doing a DFMEA a structure is defined where the customer
technical requirements are mapped to functions\, and the functions are ma
pped to failure modes that contains a cause and effect description. This i
s in a qualitative way ranked and managed.\nThe challenge in a Design Risk
Analysis work as well as when doing Reliability work is to get accurate q
uantitative numbers to express the probability of failure for a certain fa
ilure mode.\nIn the International Aerospace Quality Group and now in Suppl
y Chain Management Handbook a Guidance document has been written with the
aim to assist a standard AS9145 on Advanced Product Quality process\, wher
e the concept of Design Risk Analysis is used. This guidance material desc
ribes a process and framework for Design Risk Analysis\, where DFMEA is us
ed as recording tool\, but where a more elaborate uncertainty thinking is
used. This uncertainty thinking is referring to the concept of Knowledge S
pace and Design Space\, and the ability to predict outcome and robustness
of outcome. The toolbox therefore consists of Design of Experiments\, Mont
e-Carlo simulations and Geometry Assurance simulations as tools to be used
to map a Knowledge Space and to simulate effects of variation and the sea
rch for a Robust Design Solution.\nIn this presentation the existence of t
his guidance will be presented and discussed.\n\nhttps://conferences.enbis
.org/event/32/contributions/612/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/612/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Comprehensive Degradation Modelling: From Statistical to Artific
ial Intelligence Models
DTSTART:20230912T144000Z
DTEND:20230912T150000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-506@conferences.enbis.org
DESCRIPTION:Speakers: Arnaud Caracciolo (CETIM\, 52 Avenue Félix Louat\,
Senlis\, France)\, Amélie Ponchet Durupt (Roberval (Mechanics Energy and
Electricity)\, Centre Pierre Guillaumat\, Université de Technologie de C
ompiègne\, France)\, Nassim Boudaoud (Roberval (Mechanics Energy and Elec
tricity)\, Centre Pierre Guillaumat\, Université de Technologie de Compi
ègne\, France)\, Yun Xu (ALFI ADLER\, Route de la borde\, Crèvecœur Le
Grand\, France)\, Hai Canh Vu (Roberval (Mechanics Energy and Electricity)
\, Centre Pierre Guillaumat\, Université de Technologie de Compiègne\, F
rance)\, Patrick Leduc (ALFI ADLER\, Route de la borde\, Crèvecœur Le Gr
and\, France)\, Hasan Misaii (Roberval (Mechanics Energy and Electricity)\
, Centre Pierre Guillaumat\, Université de Technologie de Compiègne\, Fr
ance)\n\nIn the real world\, a product or a system usually loses its funct
ion gradually with a degradation process rather than fails abruptly. To me
et the demand of safety\, productivity\, and economy\, it is essential to
monitor the actual degradation process and predict imminent degradation tr
ends. A degradation process can be affected by many different factors.\nDe
gradation modelling typically involves the use of mathematical models to d
escribe the degradation processes that occur in materials or systems over
time. These models can be based on empirical data\, physical principles\,
or a combination of both\, and can be used to make predictions about the f
uture performance of the material or system.\nThis work is attempted to re
view previous degradation models\, and present some new deep learning base
d approaches for degradation modelling. First\, it deals with statistical
models\, like general path and stochastic models. Then\, because of some c
umbersomeness of statistical models\; like incompleteness modelling\, it m
oves to make comforts by some AI models.\nThe main advantage of AI models
is capturing possible nonlinearity in the observed degradation data\, but
they often suffer from limitations of available dataset.\nTo overcome limi
tations of statistical and machine learning models\, some mixed models con
sidering both simultaneously have been presented. \nThis work is aimed at
explaining briefly all models and then making a huge comparison between th
em for some irregular real degradation data. The mentioned data is related
to the wear of some chains producing glass wool.\n\nhttps://conferences.e
nbis.org/event/32/contributions/506/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/506/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Where are the Limits of AI? And How Can You Overcome these Limits
with Human Domain Knowledge?
DTSTART:20230912T144000Z
DTEND:20230912T150000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-418@conferences.enbis.org
DESCRIPTION:Speakers: Andrea Ahlemeyer-Stubbe (Ahlemeyer-Stubbe)\, Eva Sch
eideler (Technische Hochschule Ostwestfalen-Lippe)\n\nAI is the key to opt
imizing the customer experience. But without explicit industry knowledge\,
empathy\, knowledge of currents\, values and cultural characteristics of
the audience\, the cultivation\, and expansion of customer relationships f
alls short of expectations. AI and the segmentation and forecasting possib
ilities that come with it quickly become a blunt sword. Only in combinatio
n with human domain knowledge can campaigns be developed that ensure an op
timal\, hyperindividualised customer approach in a fully automated manner
and thus enable an inspiring customer experience. For decisive success\, i
t takes both man and machine.\n\nhttps://conferences.enbis.org/event/32/co
ntributions/418/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/418/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Is It a Bird? Is It a Plane? No\, It's a Paper Helicopter!
DTSTART:20230912T144000Z
DTEND:20230912T154000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-483@conferences.enbis.org
DESCRIPTION:Speakers: Jonathan Smyth-Renshaw (Jonathan Smyth-Renshaw & Ass
ociates Ltd)\n\nThe use of paper helicopters is very common when teaching
Six Sigma and in particular DoE (Design of Experiments). During the confer
ence in Turkey\, I used the paper helicopter demonstration to spur discuss
ion. Now is the time to revisit this topic and rejuvenate interest.\n\nDur
ing this session I will demonstrate how Statistical Process Control (SPC)\
, DoE (Plackett and Burman 8 runs) and single trials using the paper helic
opter\, to create a database of domain knowledge to be established which c
an be analysed using Regression.\n\nFollowing this\, there will be a discu
ssion on the application of the approach and how it could be embraced in o
ther applications.\n\nhttps://conferences.enbis.org/event/32/contributions
/483/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/483/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Computer Code Validation via Mixture Model Estimation
DTSTART:20230911T131000Z
DTEND:20230911T134000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-581@conferences.enbis.org
DESCRIPTION:Speakers: Pierre Barbillon (AgroParisTech\, université Paris-
Saclay)\, Kaniav Kamary (CentraleSupélec\, université Paris-Saclay )\, C
édric Goeury (EDF)\, Merlin Keller (EDF)\, Eric Parent (AgroParisTech\, u
niversité Paris-Saclay)\n\nWhen computer codes are used for modeling comp
lex physical systems\, their unknown parameters are tuned by calibration t
echniques. A discrepancy function is added to the computer code in order t
o capture its discrepancy with the real physical process. This discrepancy
is usually modeled by a Gaussian process. In this work\, we investigate a
Bayesian model selection technique to validate the computer code as a Bay
esian model selection procedure between models including or not a discrepa
ncy function. By embedding the competing models within an encompassing mix
ture model\, we consider each observation to belong to a different mixture
component. The model selection is then based on the posterior distributio
n of the mixture weight which identifies under which model the data are li
kely to have been generated. We check the sensitivity of posterior estimat
es to the choice of the parameter prior distributions. We illustrate that
the model discrepancy can be detected when the correlation length in the G
aussian process is not too small. The proposed method is applied to a hydr
aulic code in an industrial context. This code being non linear in its cal
ibration parameter\, we used linear surrogate illustrating that our method
can be used for more complex codes provided a reasonable linear approxima
tion.\n\nhttps://conferences.enbis.org/event/32/contributions/581/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/581/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Role of Data in Successful Transition into Bioprocess Industry 4.0
and Cognate Implications for Standardisation\, Storage and Repurposing of
Data
DTSTART:20230912T150000Z
DTEND:20230912T152000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-580@conferences.enbis.org
DESCRIPTION:Speakers: Duygu Dikicioglu (University College London)\n\nIndu
stry 4.0 opens up a new dimension of potential improvement in productivity
\, flexibility and control in bioprocessing\, with the end goal of creatin
g smart manufacturing plants with a wide web of interconnected devices. Bi
oprocessing involves living organisms or their components to manufacture a
variety of different products and deliver therapies and this organic natu
re amplifies the complexity of the process\, hence implementing novel solu
tions means higher risk and greater investment. In such a climate\, utilis
ing the existing information in the best possible way to drive novelty and
improvement in biomanufacturing becomes ever more important. A large segm
ent of the industry comprises the manufacturing of biopharmaceuticals and
advanced therapies\, some of the most expensive deliverables available to
date\, and these products undergo tightly regulated and controlled steps f
rom product conceptualisation to patient delivery. This implicates the gen
eration and storage of extensive amount of data. Despite this wealth of in
formation\, data-driven industry 4.0 initiatives have been unusually slow
in some sub-sectors hinting at an often overlooked underlying challenge im
plicating a bottleneck in the reusability of the collected data. In this t
alk\, some of the challenges around the nature of bioprocessing data\, and
its collection will be discussed and the potential solutions to overcome
such challenges will be highlighted with a specific focus in biomanufactur
ing new modalities of medicines.\n\nhttps://conferences.enbis.org/event/32
/contributions/580/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/580/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Unleashing the Potential of Data Modeling and Monitoring for a Sus
tainable and Digital Manufacturing Future: Challenges and Opportunities in
the Era of Green Targets and Industry 4.0
DTSTART:20230911T143500Z
DTEND:20230911T153500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-619@conferences.enbis.org
DESCRIPTION:Speakers: Bianca Maria Colosimo (Politecnico di Milano)\n\nThe
emergence of green targets is driving manufacturing to minimize environme
ntal impact\, optimize resource utilization\, reduce waste\, and achieve z
ero-net industries. On the other side\, the emergence of Industry 4.0 and
advancements in process technologies have led to the availability of compl
ex and massive data sets in various industrial settings. This has sparked
a new renaissance in digital manufacturing\, as industries leverage emergi
ng technologies such as additive manufacturing\, micro-manufacturing\, and
bioprinting\, coupled with advancements in sensing and computing capabili
ties. \nIn this evolving landscape\, traditional approaches to quality dat
a modeling\, monitoring\, and control\, need to be reevaluated to address
the unique challenges posed by this new paradigm shift. The talk discusses
open challenges and opportunities provided by functional data monitoring\
, manifold learning\, spatio-temporal modeling\, multi-fidelity data analy
sis\, and data reduction to unlock the potential of the green and digital
twin transition to pave the way for a more sustainable and efficient manuf
acturing future.\n\nhttps://conferences.enbis.org/event/32/contributions/6
19/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/619/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Developing a Composite Index of Environmental Consciousness: Evide
nce from Survey and Google Trends Data
DTSTART:20230912T063000Z
DTEND:20230912T065000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-578@conferences.enbis.org
DESCRIPTION:Speakers: Ida D'Attoma (Department of Statistical Sciences\, U
niversity of Bologna)\, Marco Ieva (Department of Economics and Management
\, University of Parma)\n\nEnvironmental consciousness is a complex constr
uct that involves multiple dimensions related to pro-environmental attitud
es\, beliefs and behaviours. Academic literature has attempted\, over the
last 20 years\, to conceptualize and operationalize environmental consciou
sness\, thus leading to a wide variety of measures. However\, the availabl
e measures are country-specific and with a predominant U.S. focus\, based
on convenience samples and fairly limited in terms of interpretability and
external validity. To overcome the above limitations the present study de
velops an index of environmental consciousness at both micro (consumer) an
d macro (country) level\, by considering the four main dimensions of envir
onmental consciousness: the affective\, cognitive\, active and disposition
al dimensions. By means of the analysis of more than 27 000 “Eurobaromet
er 92.4” responses from consumers belonging to the 28 EU member states i
n 2019\, the present paper develops a comprehensive measure of consumer en
vironmental consciousness that captures heterogeneity across European coun
tries. To assess the robustness of the index\, the link between environmen
tal consciousness and life satisfaction is also examined. The new survey-b
ased composite index is further compared to a big-data-based index based o
n Google Trends data on environmental-related search categories. Results s
hed light on differences in environmental consciousness across European co
untries. The link between environmental consciousness and life satisfactio
n is also supported\, confirming previous research in this area. Finally\,
the index appears to be strongly correlated with actual consumer search p
atterns on Google. Results provide implications for companies and policy m
akers on how environmental consciousness can be measured and assessed.\n\n
https://conferences.enbis.org/event/32/contributions/578/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/578/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Accelerated Stability Study with SestakBerggren R Package: Impact
of Statistics for Quicker Access to New Vaccines
DTSTART:20230911T100000Z
DTEND:20230911T103000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-532@conferences.enbis.org
DESCRIPTION:Speakers: Bernard Francq (GSK)\, Raymundo Sanchez (GSK)\, Oli
vier Schmit (GSK)\, Marilena Paludi (GSK)\n\nThe recent pandemic surged th
e emergency for quick access to new drugs and vaccines for the patients. S
tability assessment of the product may represent a bottleneck when it is b
ased on real-time data covering 2 or 3 years. To accelerate the decisions
and ultimately the time-to-market\, accelerated stability studies may be u
sed with data obtained for 6 months. We show that the kinetic Arrhenius mo
del is oversimplified to extrapolate the critical quality attribute over t
ime.\n\nOn the other hand\, the Ordinary Differential Equation (ODE) from
Sestak-Berggren model gives one overall model allowing the extrapolation o
f the degradation both in time and temperature. The statistical modeling o
f the ODE model (including bias and coverage probabilities\, from asymptot
ic theory and bootstrap) is here evaluated by simulations. Finally\, real
world data from vaccines development are analysed with the new R package S
estakBerggren. This will include decreasing and increasing trends like ant
igenicity\, residual moisture and pH.\n\nhttps://conferences.enbis.org/eve
nt/32/contributions/532/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/532/
END:VEVENT
BEGIN:VEVENT
SUMMARY:The Seven Deadly Sins of Data Science
DTSTART:20230913T101500Z
DTEND:20230913T111500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-534@conferences.enbis.org
DESCRIPTION:Speakers: Richard De Veaux (Williams College)\n\nAs we are all
too aware\, organizations accumulate vast amounts of data from a variety
of sources nearly continuously. Big data and data science advocates promis
e the moon and the stars as you harvest the potential of all these data. A
nd now\, AI threatens our jobs and perhaps our very existence. There is ce
rtainly a lot of hype. There’s no doubt that some savvy organizations a
re fueling their strategic decision making with insights from big data\, b
ut what are the challenges?\nMuch can wrong in the data science process\,
even for trained professionals. In this talk I'll discuss a wide variety o
f case studies from a range of industries to illustrate the potential dang
ers and mistakes that can frustrate problem solving and discovery -- and t
hat can unnecessarily waste resources. My goal is that by seeing some of t
he mistakes I have made\, you will learn how to take advantage of big data
and data science insights without committing the "Seven Deadly Sins."\n\n
https://conferences.enbis.org/event/32/contributions/534/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/534/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Electrical Load Curve Prediction for Non Residential Customers Usi
ng Bayesian Neural Networks
DTSTART:20230911T134000Z
DTEND:20230911T141000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-575@conferences.enbis.org
DESCRIPTION:Speakers: Anne Philippe (Nantes Université)\n\nWe explore sev
eral statistical learning methods to predict individual electrical load cu
rves using customers’ billing information. We predict the load curves by
searching in a catalog of available load curves. We develop three differe
nt strategies to achieve our purpose. The first methodology relies on esti
mating the regression function between the load curves and the predictors
(customers’ variables)\, using various feedforward neural networks. The
predicted load curve is then searched by minimizing the error between the
estimation and all the load curves available. The second and the third met
hodologies rely on dimensionality reduction on the curves using either an
autoencoder or wavelets. We then apply deep feedforward neural networks\,
Bayesian neural networks and deep Gaussian processes to estimate the regre
ssion function between the reduced load curves and the predictors. In the
second methodology\, we search for the load curve by minimizing the error
between the estimation and all the reduced load curves available in the ca
talog. In the third methodology\, however\, we reconstruct the load curves
using the estimated reduced curves\, and then we search for the predicted
curve as in the first methodology. We implement the methods mentioned abo
ve on a use-case from EDF concerning the scaled electricity consumption of
non-residential customers\, aimed at correctly predicting hours of sunlig
ht so as to size the customers’ potential photo-voltaic installations.\n
\nhttps://conferences.enbis.org/event/32/contributions/575/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/575/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Hybrid Modeling for Extrapolation and Transfer Learning in the Che
mical Processing Industries
DTSTART:20230912T131500Z
DTEND:20230912T134500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-502@conferences.enbis.org
DESCRIPTION:Speakers: Ricardo Rendall (Dow)\, Birgit Braun (Dow)\, Joel Sa
nsana (University of Coimbra)\, Caterina Rizzo (Eindhoven University of Te
chnology/Dow)\, Ivan Castillo (Dow)\, Marco P. Seabra dos Reis (Department
of Chemical Engineering\, University of Coimbra)\, Leo Chiang (Dow)\n\nHy
brid modeling is a class of methods that combines physics-based and data-d
riven models to achieve improved prediction performance\, robustness\, and
explainability. It has attracted a significant amount of research and int
erest due to the increasing data availability and more powerful analytics
and statistical methodologies (von Stosch et al.\, 2014\; Sansana et al.\,
2021). In the context of the Chemical Processing Industries (CPI)\, hybri
d modeling has the potential to improve the extrapolation capabilities of
existing models. This is a critical activity for CPI as new process condit
ions\, products\, and product grades are manufactured to handle shifting t
rends in market demand\, raw materials\, and utility costs. \nIn this work
\, we study the application of hybrid modeling for supporting extrapolatio
n and transfer learning\, both critical tasks for CPI. We study different
configurations of hybrid modeling (e.g.\, parallel\, series) and compare t
hem to benchmarks that include a physics-based model only and data-driven
models only. The physics-based model considers simplified reaction kinetic
s. The set of data-driven methods includes partial least squares (PLS)\, l
east absolute shrinkage and selection operator (LASSO)\, random forest and
boosting\, support vector regression (SVR)\, and neural networks (NN). A
simulated case study of biodiesel production (Fernandes et al.\, 2019) is
considered\, and hybrid modeling consistently shows improved results compa
red to using physics-based or data-driven models only. In particular\, ser
ial hybrid approaches are preferred for the extrapolation task. Regarding
the transfer learning task\, hybrid modeling also shows advantages\, requi
ring fewer samples than other benchmarks.\n\nhttps://conferences.enbis.org
/event/32/contributions/502/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/502/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Retrospective DoE Methodology for Guiding Process Optimization fro
m Historical Data
DTSTART:20230913T074000Z
DTEND:20230913T080000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-535@conferences.enbis.org
DESCRIPTION:Speakers: Peter Goos (KU Leuven)\, Joan Borràs-Ferrís (Unive
rsitat Politècnica de València)\, Sergio García Carrión (Universitat P
olitècnica de València (UPV))\, Alberto J. Ferrer-Riquelme (Universidad
Politecnica de Valencia)\n\nThe emergence of Industry 4.0 has led to a dat
a-rich environment\, where most companies accumulate a vast volume of hist
orical data from daily production usually involving some unplanned excitat
ions. The problem is that these data generally exhibit high collinearity a
nd rank deficiency\, whereas data-driven models used for process optimizat
ion especially perform well in the event of independent variations in the
input variables\, which have the capability to ensure causality (i.e.\, da
ta obtained from a DoE that guarantees this requirement).\nIn this work\,
we propose a retrospective DoE methodology aimed at harnessing the potenti
al of this type of data (i.e.\, data collected from daily production opera
tions) for optimization purposes. The approach consists (i) retrospectivel
y fitting two-level experimental designs\, from classical full factorial o
r fractional factorial designs to orthogonal arrays\, by filtering the dat
abase for the observations that were close to the corresponding design poi
nts\, and (ii) subsequently carrying out the analysis typically used in Do
E. We also investigate the possibility of imputing eventual missing treatm
ent conditions. Finally\, we conduct a meta-analysis with the results of a
ll the retrospective experimental designs to extract consistent conclusion
s. Here\, raster plots play a crucial role\, enabling the detection of ali
asing patterns as well as factors appearing consistently in the best model
s\, and thereby pointing to the potential active factors.\nThe proposed me
thodology is illustrated by an industrial case study. It is expected to be
useful for screening\, gaining some insights about potential active facto
rs\, and providing data-driven input towards efficient subsequent designed
experimentation.\n\nhttps://conferences.enbis.org/event/32/contributions/
535/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/535/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Tools Created with R and Python for Teaching Statistics in Blended
Learning
DTSTART:20230913T090500Z
DTEND:20230913T092500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-569@conferences.enbis.org
DESCRIPTION:Speakers: Sonja Kuhnt (Dortmund University of Applied Sciences
and Arts)\, Katharina Meiszl (Dortmund University of Applied Sciences and
Arts)\, Lara Kuhlmann de Canaviri (Fachhochschule Dortmund)\n\nBlended le
arning refers to the combination of online teaching with face-to-face teac
hing\, using the advantages of both forms of teaching. We will discuss tas
k generators developed with R and Python that support students in practisi
ng statistical tasks and can be easily extended in the future. The tools a
utomatically generate tasks with new data\, check the solutions and give s
tudents visual feedback.\n\nWe present an e-learning self-test programmed
with R on contingency tables and correlation measures. For the development
of the tool\, a so-called interactive tutorial from the learnr package is
used as the output format\, which generates a dynamic HTML-based web page
. Using the programming language Python\, a task generator for descriptive
statistics exercises was developed that covers location and scale measure
s\, histograms and boxplots. The graphical user interface is based on PyQt
5. The Qt GUI framework is written in the programming language C++ and is
offered platform-independently.\n\nFinally\, we give an outlook on researc
h within the project IPPOLIS\, which is part of the German federal-state f
unding initiative "Artificial Intelligence in Higher Education". The focus
of the overall project is on measures to improve higher education through
artificial intelligence-based support of teaching activities and learning
processes. To enable the use of case studies in statistics teaching\, a l
earning environment with R shiny is being developed.\n\nhttps://conference
s.enbis.org/event/32/contributions/569/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/569/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Time-Frequency Domain Vibration Signal Analysis to Determine the F
ailure Severity Level in a Spur Gearbox
DTSTART:20230913T090500Z
DTEND:20230913T092500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-585@conferences.enbis.org
DESCRIPTION:Speakers: Rene-Vinicio Sánchez (Universidad Politécnica Sale
siana)\, Antonio Pérez-Torres (Universidad Politécnica de Valencia)\, Su
sana Barceló Cerdá (Universidad Politécnica de Valencia)\n\nA gearbox i
s a critical component in a rotating machine\; therefore\, early detection
of a failure or malfunction is indispensable to planning maintenance acti
vities and reducing downtime costs.\nThe vibration signal is widely used t
o perform condition monitoring in a gearbox as it reflects the dynamic beh
avior in a non-invasive way. This work aimed to efficiently classify the s
everity level of a mechanical failure in a gearbox using the vibration sig
nal in the time-frequency domain.\nThe vibration signal was acquired with
six accelerometers located at different positions by modifying the load an
d rotational frequency conditions using a spur gearbox with different type
s and severity levels of simulated failure under laboratory conditions. Fi
rst\, the Wavelet transform with varying types of mother wavelet was used
to analyze the vibration signal condition in the time-frequency domain. Su
bsequently\, Random Forest (RF) and K nearest neighbor (KNN) classificatio
n models were used to determine the fault severity level.\nIn conclusion\,
RF was the most efficient classification model for classifying the severi
ty level of a fault when analyzing the vibration signal in the time-freque
ncy domain.\n\nhttps://conferences.enbis.org/event/32/contributions/585/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/585/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Learning User Preferences from Sensors on Wearable Devices
DTSTART:20230913T074000Z
DTEND:20230913T080000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-563@conferences.enbis.org
DESCRIPTION:Speakers: Aurélie Le Cain (EssilorLuxottica)\, Jairo Cugliari
(Laboratoire ERIC\, Université lumière Lyon 2)\, Simon Weinberger (Essi
lorLuxottica)\n\nThanks to wearable technology\, it is increasingly common
to obtain successive measurements of a variable that changes over time. A
key challenge in various fields is understanding the relationship between
a time-dependent variable and a scalar response. In this context\, we foc
us on active lenses equipped with electrochromic glass\, currently in deve
lopment. These lenses allow users to adjust the tint at will\, choosing fr
om four different levels of darkness. Our goal is to predict the preferred
tint level using ambient light data collected by an Ambient Light Sensor
(ALS). We approach this as an ordinal regression problem with a time-depen
dent predictor. To tackle the complexities of the task\, we use an adaptat
ion of the classical ordinal model to include time-dependent covariates. W
e explore the use of wavelets and B-splines functional basis\, as well as
regularization techniques such as Lasso or roughness penalty. In cases whe
re first-order information is insufficient\, we propose utilizing the ALS
signal's signature transform within the ordinal model to leverage second-o
rder information.\n\nhttps://conferences.enbis.org/event/32/contributions/
563/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/563/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Assessing Conditional Independence in Directed Acyclic Graphs (DAG
s)
DTSTART:20230913T063000Z
DTEND:20230913T065000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-485@conferences.enbis.org
DESCRIPTION:Speakers: Kristian Hovde Liland (NMBU)\, Ingrid Måge (Nofima
AS)\, Lars Erik Solberg (Nofima AS)\, Christian Holth Thorjussen (Nofima A
S)\n\nCausal inference based on Directed Acyclic Graphs (DAGs) is an incre
asingly popular framework for helping researchers design statistical model
s for estimating causal effects. A causal DAG is a graph consisting of nod
es and directed paths (arrows). The nodes represent variables one can meas
ure\, and the arrows indicate how the variables are causally connected. Th
e word acyclic means there can be no loops or feedback in the DAG\, meanin
g causality flows in one direction (w.r.t. time).\n\nAny DAG comes with a
set of implied (and testable) statistical conditions in the form of margin
al and conditional independencies. However\, testing of these statistical
conditions is rarely reported in applied work. One reason could be that th
ere are few straightforward\, easily accessible ways for researchers to te
st conditional independence. Most existing methods apply only to specific
cases\, are not well known\, or are difficult for the general practitioner
to implement. In addition\, there are some theoretical challenges to test
ing CI in DAGs with these methods.\n\nI will present a new method called B
ootstrapped Conditional Independence Analysis (BOOCOIN). This non-parametr
ic procedure aims to handle complex data-generating processes\, different
data types\, and small sample sizes. The method is compared to existing me
thods through simulations. The results show that BOOCOIN is an excellent t
ool for assessing implied conditional independencies in DAGs and it avoids
some of the theoretical challenges in CI testing.\n\nhttps://conferences.
enbis.org/event/32/contributions/485/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/485/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Monitoring Frameworks for ML Models
DTSTART:20230912T100500Z
DTEND:20230912T103500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-557@conferences.enbis.org
DESCRIPTION:Speakers: Alvaro Mendez (IBiDat)\n\nDespite the advantages of
ML models\, their adoption in banking institutions is often limited due to
regulatory restrictions. These regulations aim to ensure transparency and
accountability in decision-making processes and tend to prioritize tradit
ional models where interpretability and model stability are well establish
ed. This project studies the banking institution's existing workflow in te
rms of model deployment and monitoring and highlights the benefits of the
usage of ML models. The objective is to study the necessary changes when t
ransitioning from traditional models to ML models. Additionally\, we study
the existing approach for the analysis of the stability and predictive po
wer of the models and propose a series of improvements on the cases where
the current methodologies may have been outdated by newer advances or are
no longer valid in the ML context. By shedding light on the benefits and c
onsiderations associated with incorporating ML models into the finance ind
ustry\, this project contributes to the ongoing application of statistics\
, data analysis\, and ML in the industrial sector.\n\nhttps://conferences.
enbis.org/event/32/contributions/557/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/557/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Design and Inference in a RCT when Treatment Observations Follow a
Two-Component Mixture Model
DTSTART:20230911T093000Z
DTEND:20230911T100000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-496@conferences.enbis.org
DESCRIPTION:Speakers: Daniel Jeske (University of California\, Riverside)\
, Bradley Lubich (University of California)\, Weixin Yao (University of Ca
lifornia )\n\nA mixture of a distribution of responses from untreated pati
ents and a shift of that distribution is a useful model for the responses
from a group of treated patients. The mixture model accounts for the fact
that not all the patients in the treated group will respond to the treatme
nt and their responses follow the same distribution as the responses from
untreated patients. The treatment effect in this context consists of both
the fraction of the treated patients that are responders and the magnitude
of the shift in the distribution for the responders. In this talk\, we in
vestigate the design and analysis of a RCT that uses a two-component mixtu
re model for the observations in the treatment group.\n\nhttps://conferenc
es.enbis.org/event/32/contributions/496/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/496/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Estimation of the Infection Rate of Epidemics in Multilayer Random
Graphs: Comparing Classical Methods with XGBoost
DTSTART:20230913T063000Z
DTEND:20230913T065000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-466@conferences.enbis.org
DESCRIPTION:Speakers: Damján Tárkányi (Eötvös Loránd University\, Bu
dapest)\, Villő Csiszár (Loránd Eötvös University\, Budapest)\, Edit
Bognár (Eötvös Loránd University\, Budapest)\, András Zempléni (Eöt
vös Loránd University\, Budapest)\, Ágnes Backhausz (Eötvös Loránd U
niversity and Alfréd Rényi Institute of Mathematics\, Budapest)\n\nWe ad
dress the problem of estimating the infection rate of an epidemic from obs
erved counts of the number of susceptible\, infected and recovered individ
uals. In our setup\, a classical SIR (susceptible/infected/recovered) proc
ess spreads on a two-layer random network\, where the first layer consists
of small complete graphs representing the households\, while the second l
ayer models the contacts outside the households by a random graph. Our cho
ice for the latter is the polynomial model\, where three parameters contro
l how the new vertices are connected to the existing ones: uniformly\, pre
ferentially\, or by forming random triangles. \n\nOur aim is to estimate t
he infection rate $\\tau$. We apply two different approaches: the classica
l method uses a formula based on the maximum likelihood estimate\, where t
he main information comes from the estimated number of the SI edges. The s
econd\, machine learning-based approach uses a fine-tuned XGBoost algorith
m. We examine by simulations\, how the performance of our estimators depen
d on the value of $\\tau$ itself\, the phase of the epidemic\, and the gra
ph parameters\, as well as on the possible availability of further informa
tion.\n\nAcknowledgement: This research has been implemented with the supp
ort provided by the Ministry of Innovation and Technology of Hungary from
the National Research\, Development and Innovation Fund\, financed under t
he ELTE TKP 2021-NKTA-62 funding scheme.\n\nhttps://conferences.enbis.org
/event/32/contributions/466/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/466/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Complex Statistical Models for New Challenges in Life Insurance In
dustry
DTSTART:20230912T093500Z
DTEND:20230912T100500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-554@conferences.enbis.org
DESCRIPTION:Speakers: Maria Durban (Universidad Carlos III de Madrid)\n\nT
he modelling and projecting of disease incidence and mortality rates is a
problem of fundamental importance in epidemiology and population studies g
enerally\, and for the insurance and pensions industry in particular. Huma
n mortality has improved substantially over the last century\, but this ma
nifest benefit has brought with it additional stress in support systems fo
r the elderly\, such as healthcare and pension provision. For the insuranc
e and pensions industry\, the pricing and reserving of annuities depends o
n three things: stock market returns\, interest rates and future mortality
rates. Likewise\, the return from savings for the policyholder depends on
the same three factors. In the most obvious way\, increasing longevity ca
n only be regarded as a good thing for the policyholder\; a less welcome c
onsequence is that annual income from annuities will be reduced. In this t
alk\, we consider one of these three factors: the prediction of mortality.
The requirements of the insurance industry for forecasts of future mortal
ity are daunting\, because forecasts up to 50 years ahead are required for
pricing and reserving. Human mortality so far ahead depends on the impact
of such unknowables such as future medical advances. We will show how non
-parametric regression models can be used to forecast future mortality by
extrapolating past trends as well as create different scenarios to emulat
e the impact of future medical advances in mortality.\n\nhttps://conferenc
es.enbis.org/event/32/contributions/554/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/554/
END:VEVENT
BEGIN:VEVENT
SUMMARY:The Benefits of Classification: An Appointment Case Study
DTSTART:20230913T080000Z
DTEND:20230913T082000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-474@conferences.enbis.org
DESCRIPTION:Speakers: Yariv N. Marmor (BRAUDE - College of Engineering\, K
armiel)\, Boris Shnits (BRAUDE - College of Engineering\, Karmiel)\, Illan
a Bendavid (BRAUDE - College of Engineering\, Karmiel)\n\nIt is safe to as
sume that classifying patients and generating multi-type distributions of
service duration\, instead of using a general distribution for all patient
s\, would yield a better appointment schedule. One way to generate multi-t
ype distributions is by using data mining. CART\, for example\, will gener
ate the best tree\, from a statistical perspective\, nevertheless one coul
d argue that most times\, right from the base of the tree\, the marginal c
ontribution of each split decreases and at some point\, for practical uses
it is meaningless to continue further deep into the tree. Thus\, from an
operational perspective\, the question arises – what is the benefit of u
sing the whole tree compared to the much shorter (simpler) tree version? W
e explore and answer this question using an appointment case study. We sta
rt by comparing the operational measurements (i.e.\, end of day\, utilizat
ion\, idle time and over time) using the whole tree for the appointment sc
heduling vs. applying the shorter tree versions. The results show that for
all measurements there is a benefit in bigger trees until a certain point
. After that\, we can see some benefit\, but it is not statistically signi
ficant nor meaningful. We further investigate how well the findings are ro
bust under different daily patients mix. It seems that appointment schedul
ing based on bigger trees works better on average\, but it does not have a
relative advantage when patients' mix results in loaded days.\n\nhttps://
conferences.enbis.org/event/32/contributions/474/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/474/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Statistical Aspects of Kansei Engineering
DTSTART:20230912T074000Z
DTEND:20230912T084000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-546@conferences.enbis.org
DESCRIPTION:Speakers: Shirley Coleman (ISRU\, Newcastle University)\, Llui
s Marco-Almagro ( Universitat Politècnica de Catalunya)\, Simon Schütte
(Linkoping University)\n\nFollowing on from the Kansei Engineering (KE) sp
ecial session at ENBIS 2019\, we now present new work in this niche area d
ealing with design\, service and product development. \n\nThe role of affe
ctive engineering in product development. \nAffective aspects in products
are increasingly important for usability\, application and user purchase d
ecision-making. Therefore\, considering these aspects is crucial when desi
gning products and services particularly for small and medium-sized enterp
rises (SMEs). Given the current trends\, creating desire for innovative pr
oduct solutions driven by environmental adaptation\, technology advancemen
t or societal changes\, is in high demand. The talk gives a 6-step guideli
ne for KE methodology illustrated with examples.\nBy Simon Schütte \n\nAd
vances in statistical analysis and presentation of results in Kansei metho
dology.\nKE employs many interesting analyses of multi-dimensional data. E
stablished methods have been successful in extracting insight from extensi
ve data on product features and their relationship with users' emotional r
esponses. KE is continuously evolving and advances in machine learning and
artificial intelligence add new opportunities. Common users of KE are des
igners and artists\, people sometimes far away from data-driven decision m
aking\, and who value aesthetics. We present improvements in presentation
of results vital to connect with these users. \nby Lluís Marco-Almagro \n
\nPedagogic aspects of Kansei Engineering. \nKE cuts across many different
disciplines. It lends itself to small group and short project work. It is
also appropriate for more detailed\, longer-term dissertation projects. O
ur recently released textbook includes nice examples of posters produced b
y graduate students. The talk will showcase our results. \nby Shirley Cole
man\n\nhttps://conferences.enbis.org/event/32/contributions/546/
LOCATION:2.7/2.8.
URL:https://conferences.enbis.org/event/32/contributions/546/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Blending Statistics with Artificial Intelligence
DTSTART:20230912T124500Z
DTEND:20230912T131500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-595@conferences.enbis.org
DESCRIPTION:Speakers: Bart De Ketelaere (Catholic University of Leuven)\,
Yannis Kalfas (KU Leuven)\n\nIn the previous century\, statisticians playe
d the most central role in the field of data analysis\, which was primaril
y focused on analyzing structured data\, often stored in relational databa
ses. Statistical techniques were commonly employed to extract insights fro
m these data. The last few decennia have marked a substantial change in th
e way data are generated\, used and analyzed. The term data analysis is ma
inly replaced by data science now to encompass a broader scope that combin
es elements of statistics\, computer science\, and domain knowledge to ext
ract knowledge and insights\, including both structured and unstructured d
ata. This changing and expanding landscape requires a collaborative effort
involving computer scientists\, mathematicians\, engineers and statistici
ans\, inherently rendering the role of statisticians more limited as it us
ed to be. \nDuring the last few years this broader data science field was
revolutionized itself by the rapid expansion of Artificial Intelligence (A
I)\, where concepts such as Deep Learning\, Convolutional Neural Networks
and Large Language Models have proven to be nothing less than disruptive i
n many fields\, not the least in industrial applications and quality engin
eering – the home ground of industrial statisticians. \nIn this talk I
will share some of the opportunities I see for statisticians in the field
of Artificial Intelligence. I will touch upon aspects such as variable an
d sample selection (and relate it to Design of Experiments) and outlier de
tection (and relate it to robust statistics) and provide examples where we
blended statistics into an efficient AI learning strategy.\n\nhttps://con
ferences.enbis.org/event/32/contributions/595/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/595/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Near Real-Time Prediction of Hospital Performance Metrics Using Sc
alable Random Forest Algorithm
DTSTART:20230912T065000Z
DTEND:20230912T071000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-553@conferences.enbis.org
DESCRIPTION:Speakers: Richard Wood (National Health Service)\n\nWhile prev
ious studies have shown the potential value of predictive modelling for em
ergency care\, few models have been practically implemented for producing
near real-time predictions across various demand\, utilisation and perform
ance metrics. In this study\, 33 independent Random Forest (RF) algorithms
were developed to forecast 11 urgent care metrics over a 24-hour period a
cross three hospital sites in a major healthcare system in and around Bris
tol\, England. Metrics included: ambulance handover delay\; emergency depa
rtment occupancy\; and patients awaiting admission. Mean Absolute Error (M
AE)\, Root Mean Squared Error (RMSE) and Symmetric Mean Absolute Percentag
e Error (SMAPE) were used to assess the performance of RF and compare it t
o two alternative models: naïve baseline (NB) and Auto-Regressive Integra
ted Moving Average (ARIMA). Using these measures\, RF outperformed NB and
ARIMA in 76% (N = 25/33) of urgent care metrics according to SMAPE\, 88% (
N = 29/33) according to MAE and 91% (N = 30/33) according to RMSE. The RFs
developed in this study have been implemented within the local healthcare
system\, providing predictions on an hourly basis that can be accessed 24
/7 by local healthcare planners and managers. Further application of the m
odels by another healthcare system in South West England demonstrate the w
ider scalability of the approach.\n\nhttps://conferences.enbis.org/event/3
2/contributions/553/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/553/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Statistical Engineering: An Experience from Brazil
DTSTART:20230912T134500Z
DTEND:20230912T141500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-501@conferences.enbis.org
DESCRIPTION:Speakers: Andressa Siroky (UFRN)\, Carla Vivacqua (Universidad
e Federal do Rio Grande do Norte)\n\nIn this talk we share our experience
introducing Statistical Engineering as a new discipline in Brazil. We prov
ide an overview of the actions taken and the challenges we face. Our effo
rts have been mentored by Professor Geoff Vining\, an enthusiastic leader
in promoting the emerging subject of Statistical Engineering. The initiat
ive is led by the Federal University of Rio Grande do Norte (UFRN)\, locat
ed in northeast Brazil. Our approach targets two sectors: academia and bus
iness.\n\nA Statistical Engineering course was taught for the first time a
t UFRN to engineering and statistics graduate students. Initially\, the st
udents received training in the basic principles of Statistical Engineerin
g and discussed case studies. After a preparatory stage\, the group visite
d local companies to understand their needs. Our main challenge in the aca
demic setting is to engage more students since Statistical Engineering is
a non-required subject and demands extra time in a busy student schedule.\
n\nIn Brazil\, 99% of all businesses are performed by small and micro ente
rprises (SMEs). Our strategy to reach these companies considers their stru
cture and thinking. Like most small businesses in the country\, the compan
ies in the region lack a mindset of connecting with academia. Therefore\,
academia needs to take an active role disseminating the potential of Stati
stical Engineering. We are currently engaged in establishing partnerships
and starting to work on problems identified by the companies. In this sens
e\, the major challenge is to show the benefits of Statistical Engineering
to attract companies and then create lasting working collaborations.\n\nh
ttps://conferences.enbis.org/event/32/contributions/501/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/501/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Tricky Topics – a Focus on Niggling Challenges when Teaching
DTSTART:20230913T074000Z
DTEND:20230913T084000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-547@conferences.enbis.org
DESCRIPTION:Speakers: Sonja Kuhnt (Dortmund University of Applied Sciences
and Arts)\, Shirley Coleman (ISRU\, Newcastle University)\, Jacqueline As
scher (Kinneret College)\n\nThe active session will explore what topics we
find difficult to teach. Common examples include: what are degrees of fre
edom\; when should we divide by n and when by n-1? But moving on from thes
e classics\, we want to delve deeper into the things that trip us up when
performing in front of an audience of students.\n\nThe session will commen
ce with a short introduction and then settle into small groups for us to s
hare our niggling challenges. The challenges will be collated and together
we will review them and see what interesting solutions we come up with.\n
\nThe session will be co-ordinated by Jacqi Asscher\, Shirley Coleman and
Sonja Kuhnt who between them have many enjoyable (often exhausting) years
of explaining our wonderful subject to other people.\n\nhttps://conference
s.enbis.org/event/32/contributions/547/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/547/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Time Series Based Machine Learning Strategy for Wastewater-Based
Forecasting and Nowcasting of COVID-19 Dynamics
DTSTART:20230913T074000Z
DTEND:20230913T080000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-566@conferences.enbis.org
DESCRIPTION:Speakers: Mallory Lai (University of Wyoming)\, Shaun Wulff (U
niversity of Wyoming)\, Alexys McGuire (University of Wyoming)\, Bledar Bi
sha (University of Wyoming)\, Tim Robinson (University of Wyoming)\, Yongt
ao Cao (Indiana University of Pennsylvania )\n\nMonitoring COVID-19 infect
ion cases has been a singular focus of many policy makers and communities.
However\, direct monitoring through testing has become more onerous for a
number of reasons\, such as costs\, delays\, and personal choices. Wastew
ater-based epidemiology (WBE) has emerged as a viable tool for monitoring
disease prevalence and dynamics to supplement direct monitoring. In this t
alk\, I describe a time-series based machine learning strategy (TSML) whic
h incorporates WBE information for nowcasting and forecasting new weekly C
OVID-19 cases. In addition to WBE information\, other relevant temporal v
ariables such as minimum ambient temperature and water temperature are acc
ounted for via feature engineering in order to enhance the predictive capa
bility of the model. As one might expect\, the best features for short-ter
m nowcasting are often different than those for long-term forecasting of C
OVID-19 case numbers. The proposed TSML approach performs as well\, and so
metimes better\, than simple predictions that assume available and accurat
e COVID-19 case numbers from extensive monitoring and testing. As such\, m
achine learning based WBE offers a promising alternative to direct monitor
ing via testing for decision-makers and public health practitioners when p
reparing for the next wave of COVID-19 or a future pandemic.\n\nhttps://co
nferences.enbis.org/event/32/contributions/566/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/566/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Unravelling Sources of Variation in Large-Scale Food Production wi
th Power Spectral Density Analysis
DTSTART:20230912T163000Z
DTEND:20230912T165000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-552@conferences.enbis.org
DESCRIPTION:Speakers: Ingrid Måge (Nofima)\, Katinka Dankel (Nofima)\, La
rs Erik Solberg (Nofima)\, Jorun Øyaas (TINE SA)\, Jens Petter Wold (Nofi
ma)\n\nQuality testing in the food industry is usually performed by manual
sampling and at/offline laboratory analysis\, which is labor intensive\,
time consuming\, and may suffer from sampling bias. For many quality attri
butes such as fat\, water and protein\, in-line near-infrared spectroscopy
(NIRS) is an alternative to grab sampling and which provides richer infor
mation about the process. \n\nIn this ENBIS abstract\, we present benefits
of in-line measurements at the industrial scale. Specifically\, we demons
trate the advantages of in-line NIRS\, including improved precision of bat
ch estimates and enhanced process understanding\, through the analysis of
power spectral density (PSD) which served as a diagnostic tool. With the P
SD it was possible to attribute and quantify likely sources of variations.
\n\nThe results are based on a case regarding the large-scale production
of Gouda-type cheese\, where in-line NIRS was implemented to replace tradi
tional laboratory measurements. In conclusion\, the PSD of in-line NIRS pr
edictions revealed unknown sources of variation in the process that could
not have been discovered using grab sampling. Moreover\, the dairy industr
y benefited from more reliable data on key quality attributes\, providing
a strong foundation for future improvements.\n\nWhile our study focused on
a single industrial case\, the advantages of in-line NIRS and the applica
tion of PSD analysis are expected to have broader applicability in the foo
d industry.\n\n[1] Solberg\, L.E. *et al.*. In-Line Near-Infrared Spectros
copy Gives Rapid and Precise Assessment of Product Quality and Reveals Unk
nown Sources of Variation—A Case Study from Commercial Cheese Production
. Foods 2023.\n\nhttps://conferences.enbis.org/event/32/contributions/552/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/552/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Compound Poisson Process for Modeling of Aggregated Failures
DTSTART:20230913T080000Z
DTEND:20230913T082000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-555@conferences.enbis.org
DESCRIPTION:Speakers: Alessandro Di Bucchianico (Eindhoven University- of
Technology)\, Marek Skarupski (Eindhoven University of Technology)\n\nAs p
art of the Dutch national PrimaVera project (www.primavera-project.com)\,
an extensive case study with a leading high-tech company on predicting and
monitoring failure rates of components is being carried out. Following co
mmon practice from reliability engineers\, the engineers of the high-tech
company frequently use the Crow-AMSAA model for age-dependent reliability
problems. There are\, however\, two main assumptions that are not satisfie
d when the number of failures is aggregated by reports. First that we can
observe a large overdispersion in the data. The second is that the observe
d number of simultaneous events is greater than one. We propose a differen
t approach using a Compound Power Low Process. The discussion of the chose
n distribution functions and results of the fitted model simulations are p
resented. We compare our proposed model to the classical approach and comm
ent on practical issues related to the case study at hand.\n\nhttps://conf
erences.enbis.org/event/32/contributions/555/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/555/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Air Quality Monitoring: Combining Different Types of Concentration
Measures to Correct Physicochemical Model Outputs
DTSTART:20230912T161000Z
DTEND:20230912T163000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-525@conferences.enbis.org
DESCRIPTION:Speakers: Benjamin Auder (Université Paris-Saclay)\, Jean-Mic
hel Poggi (University of Paris-Saclay)\, Camille Coron (Université Paris-
Saclay)\, Emma Thulliez (INSA Rouen Normandie)\n\nOur work deals with air
quality monitoring\, by combining different types of data. More precisely\
, our aim is to produce (typically at the scale of a large given city)\, n
itrogen dioxide or fine particulate matter concentration maps\, at differe
nt moments. For this purpose\, we have at our disposal\, on the one hand\,
concentration maps produced by deterministic physicochemical models (such
as CHIMERE or SIRANE) at different spatiotemporal scales\, and on the oth
er hand\, concentration measures made at different points\, different mome
nts\, and by different devices. These measures are provided first by a sma
ll number of fixed stations\, which give reliable measurements of the conc
entration\, and second by a larger number of micro-sensors\, which give bi
ased and noisier measurements. Our approach consists in modeling the bias
of the physicochemical model (e.g. due to model assumptions that are not s
atisfied in practice\, such as constant altitude) and to estimate the para
meters of this bias using all concentration measures data. Our model relie
s on a division of space into different zones within which the bias is ass
umed to follow an affine transformation of the actual concentration. Our a
pproach allows us to improve the concentration maps provided by the determ
inistic models but also to understand the behavior of micro-sensors and th
eir contribution in improving air quality monitoring. The proposed approac
h is first introduced\, then implemented and applied numerically to a real
-world dataset collected in the Grenoble area (France).\n\nhttps://confere
nces.enbis.org/event/32/contributions/525/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/525/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Interpreting Turbulent Flows through Statistical Learning Methods
DTSTART:20230912T124500Z
DTEND:20230912T131500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-471@conferences.enbis.org
DESCRIPTION:Speakers: Ehsan Farzamnik\, Firoozeh Foroozan\, Stefano Discet
ti\, Nan Deng\, Andrea Ianiro\, Vanesa Guerrero (Universidad Carlos III de
Madrid)\, Kilian Oberleithner\, Bernd R. Noack\n\nTwo data-driven approac
hes for interpreting turbulent-flow states are discussed. On the one hand\
, multidimensional scaling and K-medoids clustering are applied to subdivi
de a flow domain in smaller regions and learn from the data the dynamics o
f the transition process. The proposed method is applied to a direct numer
ical simulation dataset of an incompressible boundary layer flow developin
g on a flat plate. On the other hand\, a novel nonlinear manifold learning
from snapshot data for shedding-dominated shear flows is proposed. Key en
ablers are isometric feature mapping\, Isomap\, as encoder and\, 𝐾-near
est neighbors algorithm as decoder. The proposed technique is applied to n
umerical and experimental datasets including the fluidic pinball\, a swirl
ing jet and the wake behind a couple of tandem cylinders.\n\nhttps://confe
rences.enbis.org/event/32/contributions/471/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/471/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Software Tool Implementation of Standard Guidelines in Technical D
ocumentation of In Vitro Diagnosis Medical Devices
DTSTART:20230913T090500Z
DTEND:20230913T092500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-556@conferences.enbis.org
DESCRIPTION:Speakers: Guillem Carretero (Werfen)\, Marina Vives-Mestres (D
atancia)\, Carmen Carbonero (Werfen)\, Alejandro Moreno (Werfen)\, Ignasi
Puig (Datancia)\n\nIn vitro diagnostics Medical Devices (IVDs) market has
had exponential growth in recent years\, IVDs are a crucial part of today
’s healthcare. Around the world\, IVDs needs to be approved for specific
regulations to market on different countries. To do so\, manufacturers ne
ed to submit the Technical Documentation to ensure safety and performance
for approval to U.S. Food and Drug Administration (FDA)\, In Vitro Diagnos
is Medical Devices Regulation (IVDR)\, Health Canada\, Japan Regulations a
mong others.\nTechnical Documentation includes the Analytical Performance
Report that describes the product accuracy\, specificity\, stability\, int
erferences\, limits of detection and quantitation among others. In all cas
es it should also include a description of the study design\, populations\
, statistical methods used\, acceptance criteria and rationale for sample
size. Guidelines as Clinical and Laboratory Standards Institute (CLSI) are
generally used to help describing\, designing\, and analyzing most of the
studies presented in the Technical Documentation. Those guidelines allow
organizations to improve their testing outcomes\, maintain accreditation\,
bring products faster to and navigate regulatory hurdles. \nGetting compl
iant reports including all relevant information through an organization an
d through all products is a time-consuming process for most device manufac
turers. To time-saving and automate the statistical methods used\, Datanci
a\, working with Werfen\, has developed and implemented a software tool to
facilitate the statistical analysis related to the execution of CLSI Guid
elines and to improve the report generation of those analysis in preparati
on of the submission of Technical Documentation. In this presentation we w
ill show the tool and share its use at Werfen.\n\nhttps://conferences.enbi
s.org/event/32/contributions/556/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/556/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Statistical Engineering: Strategy versus Tactics
DTSTART:20230912T131500Z
DTEND:20230912T134500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-570@conferences.enbis.org
DESCRIPTION:Speakers: Geoff Vining (Virginia Tech Statistics Department)\n
\nThe International Statistical Engineering Association on its webpage sta
tes\, “Our discipline provides guidance to develop appropriate strategie
s to produce sustainable solutions.” Clearly\, strategy should be an es
sential foundation the proper implementation of statistical engineering.
Yet\, virtually all of the materials on the website are more tactical than
strategic. This talk explores the issue\, offers an explanation why\, an
d outlines a pathway for improvement. This talk is the result of the auth
or’s experience as a Fulbright Scholar working with colleagues at the Fe
deral University of Rio Grande do Norte (UFRN) in Natal\, Brazil\, June to
August 2022 and May to July 2023.\n\nThe goal was to initiate a Statistic
al Engineering program over the two-year period 2022-23. Covid seriously
impacted the group’s efforts in 2022. However\, it did provide a start.
The focus was to train a cadre of faculty and students in the basics of s
tatistical engineering and work on projects with local companies with the
full support of the organizations’ senior leadership. The group establi
shed working relationships with two local companies in 2022\, but covid de
railed the proposed projects. The group enjoyed more success in 2023\, wo
rking with the university’s Institute for Tropical Medicine.\n\nThese in
teractions with local organizations to address their complex opportunities
provides an appropriate setting to distinguish the truly strategic elemen
ts of statistical engineering from the purely tactical. Understanding the
difference is essential for the future of the discipline. Basically\, th
e discipline of statistical engineering can address its own complex opport
unity by learning from our Brazilian experience.\n\nhttps://conferences.en
bis.org/event/32/contributions/570/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/570/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Challenges and Obstacles in Process Understanding and Monitoring w
ith Process Analytical Technologies
DTSTART:20230912T161000Z
DTEND:20230912T163000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-479@conferences.enbis.org
DESCRIPTION:Speakers: Alberto J. Ferrer-Riquelme (Universidad Politecnica
de Valencia)\, Barbara Giussani (University of Insubria)\, Giulia Gorla (U
niversity of Insubria)\n\nThe use of Process Analytical Technology (PAT) i
n dairy industries can enhance manufacturing processes efficiency and impr
ove final product quality by facilitating monitoring and understanding of
these processes. Currently\, near-infrared spectroscopy (NIR) is one of th
e most widely used optical technologies in PAT\, thanks to its ability to
fingerprint materials and simultaneously analyze various food-related phen
omena. Recently\, low-cost miniaturized NIR spectrometers\, coupled with m
ultivariate data analysis\, have been employed to solve classification\, d
iscrimination\, and quantification issues in various fields. However\, imp
lementing these technologies for online monitoring is still challenging.\n
In this study\, a lab-scale feasibility study has been conducted to invest
igate the potentialities and limits of a handheld spectrometer for kefir f
ermentation. Multivariate statistical tools were intended to consider time
dependency and dynamics over the process that happens through different p
hases. The possibilities offered by different statistical tools in gaining
information about process occurrence were examined on the one hand\, for
process understanding and\, on the other\, for process monitoring and endp
oint determination. \nExploiting data information showed great potential f
or miniaturized NIR in real-time monitoring and modeling of the fermentati
on process that could help close the loop for automated process management
.\n\nhttps://conferences.enbis.org/event/32/contributions/479/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/479/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Using Lean Practices to Overcome Challenges with Improving Warehou
se Operations
DTSTART:20230912T093500Z
DTEND:20230912T100500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-511@conferences.enbis.org
DESCRIPTION:Speakers: Teresa Cardoso-Grilo (Instituto Universitário de Li
sboa (ISCTE-IUL)\, Business Research Unit (BRU-IUL))\, Jamison Kovach (Uni
versity of Houston)\, Diogo Gomes (Iscte – Instituto Universitário de L
isboa)\n\nThis improvement project was conducted in a warehouse that provi
des repair services and storage for the equipment\, supplies/consumables\,
and repair parts needed to perform technical cleaning and hygiene service
s for clients such as in schools\, hospitals\, airports\, etc. While initi
ally organizing materials one section/area at a time using 5S (sort\, set-
in-order\, shine\, standardize\, and sustain)\, challenges encountered inc
luded space limitations with the existing layout\, co-location of repair a
nd storage operations\, inability to temporarily shut-down operations to o
rganize\, and resistance to disposing of unneeded items. To resolve the sp
ace/layout issues\, management invested in renovating the warehouse. While
cleaning prior to renovations\, many sorting activities took place. Post-
renovations\, to organize and streamline operations (set-in-order and shin
e)\, a new space utilization plan was developed and implemented that organ
ized items across the warehouse into separate storage zones\, designated s
pace only for performing repairs\, and stored all repair parts together in
the same zone. To standardize\, visual controls were implemented\, includ
ing floor markings and storage location labels\, as well as a shadow peg b
oard and a toolbox with foam cut-outs in the repair work area. In addition
\, standard operating procedures were developed. To sustain\, a bi-weekly
audit procedure was developed\, and a communication board was installed wh
ere\, among other things\, audit feedback would be posted. Finally\, inven
tory reorder points were established\, and a Kanban system was implemented
for material replenishment.\n\nhttps://conferences.enbis.org/event/32/con
tributions/511/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/511/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Self-Starting Bayesian Hotelling $T^2$ for Online Multivariate Out
lier Detection
DTSTART:20230912T163000Z
DTEND:20230912T165000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-594@conferences.enbis.org
DESCRIPTION:Speakers: Panagiotis Tsiamyrtzis (Politecnico di Milano)\, Kon
stantinos Bourazas (University of Ioannina)\, Apostolos Batsidis (Universi
ty of Ioannina)\n\nHotelling’s $T^2$ control chart is probably the most
widely used tool in detecting outliers in a multivariate normal distributi
on setting. Within its classical scheme\, the unknown process parameters (
i.e.\, mean vector and variance-covariance matrix) are estimated via a pha
se I (calibration) stage\, before online testing can be initiated in phase
II. In this work we develop the self-starting analogue of Hotelling’s $
T^2$\, within the Bayesian arena\, allowing online inference from the earl
y start of the process. Both mean and variance-covariance matrix will be a
ssumed unknown\, and a conjugate (power) prior will be adopted\, guarantee
ing a closed form mechanism. Theoretical properties\, including power calc
ulations of the proposed scheme\, along with root-cause related post-alarm
inference methods are studied. The performance is examined via a simulati
on study\, while some real multivariate data illustrate its use in practic
e.\n\nhttps://conferences.enbis.org/event/32/contributions/594/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/594/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Non-Linear Mixed Model Approach for Detecting Outlying Profiles
DTSTART:20230912T090500Z
DTEND:20230912T093500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-571@conferences.enbis.org
DESCRIPTION:Speakers: Valeria Quevedo (Universidad de Piura)\, Geoff Vinin
g (Virginia Tech Statistics Department)\n\nIn parametric non-linear profil
e modeling\, it is crucial to map the impact of model parameters to a sing
le metric. According to the profile monitoring literature\, using multivar
iate $T^2$ statistic to monitor the stability of the parameters simultaneo
usly is a common approach. However\, this approach only focuses on the est
imated parameters of the non-linear model and treats them as separate but
correlated quality characteristics of the process. Consequently\, they do
not take full advantage of the model structure. To address this limitation
\, we propose a procedure to monitor profiles based on a non-linear mixed
model that considers the proper variance-covariance\nstructure. Our propos
ed method is based on the concept of externally studentized residuals to t
est whether a given profile significantly deviates from the other profiles
in the non-linear mixed model. The results show that our control chart is
effective and appears to perform better than the $T^2$ chart. We applied
our approach in an aquaculture process to monitor the shrimp weight over 3
00 ponds.\n\nhttps://conferences.enbis.org/event/32/contributions/571/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/571/
END:VEVENT
BEGIN:VEVENT
SUMMARY:From Dashboards to Data Science Reactive Web Apps: Journey and Suc
cess Stories for Evidence-Based Decision Making in Industry and Business
DTSTART:20230911T131000Z
DTEND:20230911T134000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-539@conferences.enbis.org
DESCRIPTION:Speakers: Emilio L. Cano (Rey Juan Carlos University)\n\nData
science is getting closer and closer to the core of Business. Statistical
analysis is not anymore a task constrained to data analysts that end up in
a results report for making. On the one hand\, as Data Visualization and
Machine Learning models are spreading throughout all business areas\, it i
s needed something else than static reports. The deployment of Data Scienc
e products to be consumed by the stakeholders is a major area of developme
nt nowadays (MLOps). On the other hand\, not only statistical experts are
going to use the Data Science products. Decision making is carried out at
different levels all over the organization\, from process owners to execut
ive managers. Thus\, dynamic and interactive user interfaces that lead sta
keholders through the knowledge discovery path steamed from Data Science a
re needed. Last but not least\, well designed interfaces for cutting-edge
models allows to tackle another of the main concerns of Data Science: inte
rpretability.\n\nIn this work\, one of the most amazing workflows for depl
oying and using Data Science products is showcased: The Shiny web applicat
ions framework. Shiny surged as an R package to build reactive web applica
tions by using regular R code and nothing else. Shiny apps are more than a
dashboard for observing what happened\, but a sort of cockpit for anticip
ating what will happen and\, even better\, making decisions based on evide
nce to improve the future. The basics of the Shiny apps developement proce
ss will be shown\, and some success stories in industry and business will
be showcased.\n\nhttps://conferences.enbis.org/event/32/contributions/539/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/539/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Bayesian Estimation in Regression Models with Restricted Parameter
Spaces
DTSTART:20230911T115000Z
DTEND:20230911T121000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-488@conferences.enbis.org
DESCRIPTION:Speakers: Solmaz Seifollahi (University of Tabriz)\, Kaniav Ka
mary (Faculty member)\, Hossein Bevrani (University of Tabriz)\n\nRegressi
on models have become increasingly important in a range of scientific fiel
ds\, but accurate parameter estimation is crucial for their use. One issue
that has recently emerged in this area is the estimation of parameters in
linear or generalized linear models when additional information about the
parameters limits their possible values. One issue that has recently emer
ged in this area is the estimation of parameters in linear or generalized
linear models when additional information about the parameters limits thei
r possible values. Most studies have focused on parameter spaces limited b
y the information that can be expressed as Hβ = r. However\, in some fiel
ds\, such as applied economics or hyperspectral analysis\, parameters must
be non-negative\, which can be expressed as Hβ ≤ r. In such situations
\, classical inference methods may not be suitable\, and Bayesian inferenc
e can be a better alternative. In this paper\, we explore techniques that
have been developed to estimate parameters\, along with their drawbacks\,
including accuracy and time consumption. We then introduce new algorithms
that have been developed to address these issues\, and we present simulati
on studies demonstrating their efficacy. Finally\, we illustrate the perfo
rmance of these new algorithms with practical examples.\n\nhttps://confere
nces.enbis.org/event/32/contributions/488/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/488/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Process Optimization Using Bayesian Models for Bounded Data
DTSTART:20230912T152000Z
DTEND:20230912T154000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-529@conferences.enbis.org
DESCRIPTION:Speakers: Chellafe Ensoy-Musoro (Janssen Pharmaceutica)\n\nDes
ign space construction is a key step in the Quality by Design paradigm in
manufacturing process development. Construction typically follows the deve
lopment of a response surface model (RSM) that relates different process p
arameters with various product quality attributes and serves the purpose o
f finding the set of process conditions where acceptance criteria of the o
bjectives are met with required level of assurance. If a potentially large
number of process parameters is being looked at\, this RSM can be develop
ed from a screening plus augmentation study.\nAlthough normal RSM is typic
ally fitted for this investigation\, this is often no longer applicable fo
r bounded response. Using the incorrect model can lead to identification o
f the wrong parameters in the screening study\, thereby leading to a non-o
ptimal design space.\nIn this work\, we show the use of Beta-regression an
d Fractional-response generalized linear models as alternatives to the nor
mal RSM. All models are fitted in the Bayesian framework since the expecte
d posterior distribution is typically used in characterizing the design sp
ace. We compare the performance of the two models across different locatio
n and spread scenarios. We demonstrate this technique using simulated data
that was derived based on a real optimization study in chemical synthesis
.\n\nhttps://conferences.enbis.org/event/32/contributions/529/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/529/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Robust Bayesian Reliability Demonstration Testing
DTSTART:20230913T074000Z
DTEND:20230913T080000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-518@conferences.enbis.org
DESCRIPTION:Speakers: Hugalf Bernburg (Physikalisch-Technische Bundesansta
lt (PTB))\, Katy Klauenbberg (Physikalisch-Technische Bundesanstalt (PTB))
\, Clemens Elster (Physikalisch-Technische Bundesanstalt (PTB))\n\nTo demo
nstrate reliability at consecutive timepoints\, a sample at each current t
imepoint must prove that at least 100$p$% of the devices of a population f
unction until the next timepoint with probability of at least $1-\\omega$.
\n\nFor testing that reliability\, we develop a failure time model which
is motivated by a Bayesian rolling window approach on the mean time to fai
lure. Based on this model and a Bayesian approach\, sampling plans to demo
nstrate reliability are derived. \n\nWe will apply these sampling plans on
data generated by power law processes\, that have a time dependent mean t
ime to failure\, to demonstrate the balance between the flexibility of the
developed model and the slightly increased costs due to not assuming a co
nstant mean time to failure. Good frequentist properties and the robustnes
s of the sampling plans are shown. \n\nWe apply these sampling plans to te
st if the verification validity period can be extended for e.g.\, a popula
tion of utility meters which are subject to section 35\, paragraph 1\, No.
1 of the Measures and Verification Ordinance in Germany [1]. Accordingly\
, the verification validity period may be extended if it can be assumed th
at at least 95% of the measuring instruments conform with specified requir
ements during the whole period of extension. \n\n[1] Mess- und Eichverordn
ung (MessEV)\, December 11th\, 2014 (Bundesgesetzblatt I\, p. 2010 - 73)\,
last amended by Article 1 of the Ordinance of October 26\, 2021 (Bundesge
setzblatt I\, p. 4742). \nRetrieved: May 15\, 2023\, from https://www.gese
tze-im-internet.de/messev/MessEV.pdf\n\nhttps://conferences.enbis.org/even
t/32/contributions/518/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/518/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Modelling of Multilayer Delamination
DTSTART:20230912T150000Z
DTEND:20230912T152000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-505@conferences.enbis.org
DESCRIPTION:Speakers: Horst Lewitschnig (Infineon Technologies Austria AG)
\, Renato Podvratnik (Infineon)\n\nNowadays\, die stacking is gaining a lo
t of attention in the semiconductor industry. Within this assembly techniq
ue\, two or more dies are vertically stacked and bonded in a single packag
e. Compared to single-die packages\, this leads to many benefits\, includi
ng more efficient use of space\, faster signal propagation\, reduced power
consumption\, etc.\nDelamination\, i.e.\, the separation of two intendedl
y connected layers\, is a common failure attribute of semiconductor dies.
Measured from 0 to 100 percent\, the delamination of a single die is typic
ally modeled by the beta distribution. Considering that the delamination l
evels of stacked dies correlate\, there is need for a model of the whole s
tack\, which is a probability distribution on the unit hypercube.\nContrar
y to\, e.g.\, the normal distribution\, there isn’t a standard extension
of the beta distribution to multiple dimensions. Thus\, we present and ex
tensively evaluate three different approaches how to obtain an appropriate
distribution on the unit cube. These are the construction of multivariate
beta distributions using ratios of gamma random variables\, the applicati
on of Gaussian copulas\, and the factorization of the joint distribution i
n conditional ones that are individually modeled via beta regression. The
model evaluation is based on simulated and real delamination data.\nFinall
y\, we extend the proposed models in a way that they are able to describe
delamination over time. Thus\, we provide an advanced framework for multiv
ariate delamination modeling\, which is of particular value for higher deg
rees of integration\, new package concepts\, and assessment of product qua
lifications.\n\nhttps://conferences.enbis.org/event/32/contributions/505/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/505/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Reliability Growth in the Context of Industry 4.0
DTSTART:20230912T152000Z
DTEND:20230912T154000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-512@conferences.enbis.org
DESCRIPTION:Speakers: Nikolaus Haselgruber (CIS Consulting in Industrial S
tatistics GmbH)\n\nOne of the last major steps in the development of compl
ex technical systems is reliability growth (RG) testing. According to [1]\
, RG is defined as […] improvement of the reliability of an item with ti
me\, through successful correction of design or product weaknesses. This m
eans that besides testing\, a qualified monitoring and inspection as well
as an effective corrective action mechanism is required. The simultaneousl
y running and interacting processes of testing\, inspection and correction
share some of their data sources and require input from different fields
of the development. Thus\, digitalisation of the RG process has high poten
tial in terms of effectivity in time\, costs\, data quality and longitudin
al comparability of results. \nThis talk summarizes the findings of implem
entations of the RG process in digital industrial environments. Establishe
d RG models are compared not only according to statistical properties but
also with regard to connectivity in machine-to-machine applications.\n[1]
Birolini\, A. (2004): Reliability Engineering. 4th ed.\, Springer\, Berlin
.\n\nhttps://conferences.enbis.org/event/32/contributions/512/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/512/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Optimal Design for Model Autocompletion
DTSTART:20230911T121000Z
DTEND:20230911T123000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-607@conferences.enbis.org
DESCRIPTION:Speakers: Arno Strouwen\n\nMost experimental design methodolog
y focuses on parameter precision\, where the model structure is assumed kn
own and fixed. But arguably\, finding the correct model structure is the p
art of the modelling process that takes the most effort.\n\nExperimental d
esign methodology for model discrimination usually focuses on discriminati
ng between two or more known model structures. But often part of the model
structure is entirely unknown\, and then these techniques cannot be appli
ed.\n\nRecently\, techniques such as sparse identification of nonlinear dy
namics and symbolic regression have been used to complete models where a p
art of the model structure is missing. However\, this work focussed on rec
overing the true model from a fixed dataset.\n\nIn this talk\, I propose a
n adaptive data gathering strategy which aims to perform model autocomplet
ion with as little data as possible. Specifically\, symbolic regression is
used to suggest plausible model structures\, and then a variant of the T-
optimal design criterion is used to find a design point that optimally dis
criminates between these structures. A new measurement is then gathered\,
and new regression models are constructed. This loop continues until only
one model structure remains plausible.\n\nhttps://conferences.enbis.org/ev
ent/32/contributions/607/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/607/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Monitoring Resistance Spot Welding Profiles via Robust Control Cha
rts
DTSTART:20230912T155000Z
DTEND:20230912T161000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-541@conferences.enbis.org
DESCRIPTION:Speakers: Christian Capezza (Department of Industrial Engineer
ing\, University of Naples "Federico II")\, Biagio Palumbo (Università di
Napoli Federico II)\, Antonio Lepore (Università degli Studi di Napoli F
ederico II - Dept. of Industrial Engineering)\, Fabio Centofanti (Universi
ty of Naples)\n\nMonitoring the stability of manufacturing processes in In
dustry 4.0 applications is crucial for ensuring product quality. However\,
the presence of anomalous observations can significantly impact the perfo
rmance of control charting procedures\, especially in complex and high-dim
ensional settings.\nIn this work\, we propose a new robust control chart t
o address these challenges in monitoring multivariate functional data whil
e being robust to functional casewise and cellwise outliers.\nThe proposed
control charting framework consists of a functional univariate filter for
identifying and replacing functional cellwise outliers\, a robust imputat
ion method for missing values\, a casewise robust dimensionality reduction
technique\, and a monitoring strategy for the multivariate functional qua
lity characteristic.\nWe conduct extensive Monte Carlo simulations to comp
are the performance of the proposed control chart with existing approaches
.\nAdditionally\, we present a real-case study in the automotive industry\
, where the proposed control chart is applied to monitor a resistance spot
welding process and to demonstrate its effectiveness and practical applic
ability.\n\nhttps://conferences.enbis.org/event/32/contributions/541/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/541/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Bayesian Multilevel Time-Varying Framework for Joint Modelling o
f Hospitalization and Survival in Patients on Dialysis
DTSTART:20230912T124500Z
DTEND:20230912T131500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-520@conferences.enbis.org
DESCRIPTION:Speakers: Esra Kurum (University of California\, Riverside)\n\
nOver 782\,000 individuals in the U.S. have end-stage kidney disease with
about 72% of patients on dialysis\, a life-sustaining treatment. Dialysis
patients experience high mortality and frequent hospitalizations\, at abou
t twice per year. These poor outcomes are exacerbated at key time periods\
, such as the fragile period after the transition to dialysis. In order to
study the time-varying effects of modifiable patient and dialysis facilit
y risk factors on hospitalization and mortality\, we propose a novel Bayes
ian multilevel time-varying joint model. Efficient estimation and inferenc
e are achieved within the Bayesian framework using Markov Chain Monte Carl
o\, where multilevel (patient- and dialysis facility-level) varying coeffi
cient functions are targeted via Bayesian P-splines. Applications to the U
nited States Renal Data System\, a national database which contains data o
n nearly all patients on dialysis in the U.S.\, highlight significant time
-varying effects of patient- and facility-level risk factors on hospitaliz
ation risk and mortality. Finite sample performance of the proposed method
ology is studied through simulations.\n\nhttps://conferences.enbis.org/eve
nt/32/contributions/520/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/520/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Resistance Spot Welding Process Monitoring Through Mixture Functio
n-On-Scalar Regression Analysis
DTSTART:20230912T161000Z
DTEND:20230912T163000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-542@conferences.enbis.org
DESCRIPTION:Speakers: Christian Capezza (Department of Industrial Engineer
ing\, University of Naples "Federico II")\, Davide Forcina (Department of
Industrial Engineering\, University of Naples "Federico II")\, Antonio Lep
ore (Department of Industrial Engineering\, University of Naples "Federico
II")\, Fabio Centofanti (Department of Industrial Engineering\, Universit
y of Naples "Federico II")\, Biagio Palumbo (Department of Industrial Engi
neering\, University of Naples "Federico II")\n\nThe advancement in data a
cquisition technologies has made possible the collection of quality charac
teristics that are apt to be modeled as functional data or profiles\, as w
ell as of collateral process variables\, known as covariates\, that are po
ssibly influencing the latter and can be in the form of scalar or function
al data themselves. In this setting\, the functional regression control ch
art is known to be capable of monitoring a functional quality characterist
ic adjusted by the influence of multiple functional covariates through a s
uitable functional linear model (FLM)\, even though\, in many applications
\, this influence is not adequately captured by a single FLM. In this pape
r\, a new profile monitoring control chart is proposed to let the regressi
on structure vary across groups of subjects by means of a mixture of regre
ssion models\, after a multivariate functional principal component decompo
sition step is performed to represent the functional data. The performance
of the proposed method is compared through a Monte Carlo simulation study
with other methods already presented in the literature. Furthermore\, to
demonstrate the flexibility of the proposed to handle FLMs with different
types of response and/or predictors\, a real-case study in the automotive
industry is presented in the function-on-scalar regression setting.\n\nhtt
ps://conferences.enbis.org/event/32/contributions/542/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/542/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Some Notes on Determining the Minimal Sample Size in Balanced 3-wa
y ANOVA Models where no Exact F-Test Exists
DTSTART:20230912T071000Z
DTEND:20230912T073000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-524@conferences.enbis.org
DESCRIPTION:Speakers: Bernhard Spangl (University of Natural Resources and
Life Sciences\, Vienna)\, Norbert Kaiblinger (University of Natural Resou
rces and Life Sciences\, Vienna)\n\nFor the two three-way ANOVA models $A
\\times BB \\times CC$ and $(A \\succ BB) \\times CC$ (doubled letters ind
icate random factors) an exact $F$-test does not exist\, for testing the h
ypothesis that the fixed factor $A$ has no effect. Approximate $F$-tests c
an be obtained by Satterthwaite's approximation. The approximate $F$-test
involves mean squares to be simulated. To approximate the power of the tes
t\, we simulate data such that the null hypothesis is false and we compute
the rate of rejections. The rate then approximates the power of the test.
\n\nIn this talk we aim to determine the minimal sample size of the two mo
dels mentioned above given a prespecified power and we\n\n(i) give a heuri
stic that the number of replicates $n$ should be kept small ($n=2$). This
suggestion is backed by all simulation results.\n\n(ii) determine the acti
ve and inactive variance components for both ANOVA models using a surrogat
e fractional factorial model with variance components as factors.\n\n(iii)
determine the worst combination of active variance components for both mo
dels using a surrogate response surface model based on a Box-Behnken desig
n. The special structure of the Box-Behnken design ensures that the used m
odels have similar total variance.\n\nAdditionally we propose three practi
cal methods that help reducing the number of simulations required to deter
mine the minimal sample size.\n\nWe compare the proposed methods\, present
some examples and\, finally\, we give recommendations about which method
to choose.\n\nhttps://conferences.enbis.org/event/32/contributions/524/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/524/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Towards Markets for Data and Analytics
DTSTART:20230911T073000Z
DTEND:20230911T083000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-482@conferences.enbis.org
DESCRIPTION:Speakers: Pierre Pinson (Imperial College London)\n\nWith all
the data being collected today\, there is a strong focus on how to generat
e value from that data for various stakeholders and society as a whole. Wh
ile many analytics tasks can be solved efficiently using local data only\,
typically\, their solution can be substantially improved by using data of
others. Obvious examples would include (i) supply chains where stakeholde
rs can highly benefit from data upstream (production side) and downstream
(consumption side)\, as well as (ii) tourism\, where for instance the hosp
itality industry may find value in data coming from transportation. Anothe
r important application area is that of energy systems\, where many stakeh
olders collect and own data\, would benefit from each other’s data\, but
are reluctant to share. Sharing limitations are often motivated by privac
y concerns (individuals)\, or by the potential loss of a competitive advan
tage (firms).\nWe explore various approaches to support collaborative anal
ytics to incentivise data sharing. Eventually\, this leads to discussing m
onetisation of data and of the contribution of features and data streams t
o the solving of common analytics tasks. We will zoom into the specific ex
amples of regression and prediction markets\, with application to energy s
ystem operation problems.\n\nhttps://conferences.enbis.org/event/32/contri
butions/482/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/482/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Multi-Objective Optimisation Under Uncertainty
DTSTART:20230912T152000Z
DTEND:20230912T154000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-523@conferences.enbis.org
DESCRIPTION:Speakers: Semochkina Dasha (Southampton Statistical Sciences R
esearch Institute (S3RI))\n\nBroadly speaking\, Bayesian optimisation meth
ods for a single objective function (without constraints) proceed by (i) a
ssuming a prior for the unknown function f (ii) selecting new points x at
which to evaluate f according to some infill criterion that maximises an a
cquisition function\; and (iii) updating an estimate of the function optim
um\, and its location\, using the updated posterior for f. The most common
prior for f is a Gaussian process (GP).\n \nOptimisation under uncertaint
y is important in many areas of research. Uncertainty can come from variou
s sources\, including uncertain inputs\, model uncertainty\, code uncertai
nty and others. Multi-objective optimisation under uncertainty is a power
ful tool and a big area of research. \n \nIn this talk\, I will give an ov
erview of Bayesian optimisation and talk about a few extensions to the emu
lation-based optimisation methodology called expected quantile improvement
(EQI) to a two-objective optimisation case. We demonstrate how this multi
-objective optimisation technique handles uncertainty and finds optimal so
lutions under high levels of uncertainty.\n\nhttps://conferences.enbis.org
/event/32/contributions/523/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/523/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Deep Neural Network-Based Parameter Estimation of the Fractional O
rnstein-Uhlenbeck Process
DTSTART:20230912T155000Z
DTEND:20230912T161000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-611@conferences.enbis.org
DESCRIPTION:Speakers: László Márkus (Dept. Probability Th. and Statisti
cs\, Eötvös Loránd University)\, Iván Ivkovic (Eőtvős Loránd Univer
sity)\, Dávid Kovács (Eőtvős Loránd University)\, Dániel Boros (Eöt
vös Loráns University)\n\nWe present a novel deep neural network-based a
pproach for the parameter estimation of the fractional Ornstein-Uhlenbeck
(fOU) process. The accurate estimation of the parameters is of paramount i
mportance in various scientific fields\, including finance\, physics\, and
engineering. We utilize a new\, efficient\, and general Python package fo
r generating fractional Ornstein-Uhlenbeck processes in order to provide a
large amount of high-quality synthetic training data. The resulting neura
l models significantly surpass the performance of state-of-the-art estimat
ion methods for fOU realizations. The consistency and robustness of the es
timators are supported by experiments. We believe that our work will inspi
re further research in the application of deep learning techniques for sto
chastic process modeling and parameter estimation.\n\nhttps://conferences.
enbis.org/event/32/contributions/611/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/611/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Case Studies of Statistical Process Control and Anomaly Detection
DTSTART:20230911T124000Z
DTEND:20230911T131000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-519@conferences.enbis.org
DESCRIPTION:Speakers: Salvador Naya (Universidade da Coruña)\, Miguel Flo
res (Escuela Politécnica Nacional)\, Javier Tarrío Saavedra (Universidad
e da Coruña)\, Luis Carral (Universidade da Coruña)\n\nStatistical proce
ss control (SPC)\, as part of quality control\, makes it possible to monit
or the quality levels of products and services\, detect possible anomalies
\, their assignable causes and\, consequently\, facilitate their continuou
s improvement. This work will present the application of various SPC tools
for the control of processes such as transit through the Expanded Panama
Canal or the energy efficiency and hygrothermal comfort in buildings. Depe
nding on the degree of complexity of data\, univariate\, multivariate or f
unctional data control charts will be used. Likewise\, other alternatives
for anomaly detection\, from the perspective of classification methods\, w
ill also be shown.\n\nReferences:\n\nCarral\, L.\, Tarrío-Saavedra\, J.\,
Sáenz\, A. V.\, Bogle\, J.\, Alemán\, G.\, & Naya\, S. (2021). Modellin
g operative and routine learning curves in manoeuvres in locks and in tran
sit in the expanded Panama Canal. The Journal of Navigation\, 74(3)\, 633-
655.\nFlores\, M.\, Naya\, S.\, Fernández-Casal\, R.\, Zaragoza\, S.\, Ra
ña\, P.\, & Tarrío-Saavedra\, J. (2020). Constructing a control chart us
ing functional data. Mathematics\, 8(1)\, 58.\nRemeseiro\, B.\, Tarrío-Sa
avedra\, J.\, Francisco-Fernández\, M.\, Penedo\, M. G.\, Naya\, S.\, & C
ao\, R. (2019). Automatic detection of defective crankshafts by image anal
ysis and supervised classification. The International Journal of Advanced
Manufacturing Technology\, 105\, 3761-3777.\nSosa Donoso\, J. R.\, Flores\
, M.\, Naya\, S.\, & Tarrío-Saavedra\, J. (2023). Local Correlation Integ
ral Approach for Anomaly Detection Using Functional Data. Mathematics\, 1
1(4)\, 815.\n\nhttps://conferences.enbis.org/event/32/contributions/519/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/519/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Predictive Maintenance Strategy Cost-Model
DTSTART:20230912T082000Z
DTEND:20230912T084000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-423@conferences.enbis.org
DESCRIPTION:Speakers: Florian Sobieczky (SCCH Software Competence Center H
agenberg)\, Ivo Bukovsky (University of South Bohemia)\, Ondřej Budík (
University of South Bohemia)\, Maqbool Khan (Pak-Austria Fachhoschule)\n\n
The benefit of predictive maintenance (PdM) as an enterprise strategy for
scheduling repairs compared to other maintenance strategies relies heavily
on the optimal use of resources\, especially for SMEs: Expertise in the p
roduction process\, Machine Learning Know-How\, Data Quality and Sufficien
cy\, and User Acceptance of the AI-Models have shown to be significant fac
tors in the profit calculation. Using a stochastic model for the productio
n cycle\, we show how all these factors determine reduction of revenue and
increase in maintenance cost\, providing quantitative conditions for the
beneficial use of PdM.\n\nhttps://conferences.enbis.org/event/32/contribut
ions/423/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/423/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Optimization of Imperfect Condition-Based Maintenance Based on Mat
rix Algebra
DTSTART:20230912T063000Z
DTEND:20230912T065000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-515@conferences.enbis.org
DESCRIPTION:Speakers: Bram de Jonge (University of Groningen)\n\nIndustria
l systems are in general subject to deterioration\, ultimately leading to
failure\, and therefore require maintenance. Due to increasing possibilit
ies to monitor\, store\, and analyze conditions\, condition-based maintena
nce policies are gaining popularity. We consider optimization of imperfect
condition-based maintenance for a single unit that deteriorates according
to a discrete-time Markov chain. The effect of a maintenance intervention
is stochastic\, and we provide reasonable properties for the transition m
atrix that models the effect of maintenance. Because maintenance does not
always bring us back to the as-good-as-new state\, we are dealing with a s
emi-regenerative process rather than a regenerative process. We provide di
fferent methods to determine the optimal maintenance policy and aim to pro
ve that the optimal policy is of the control-limit type.\n\nhttps://confer
ences.enbis.org/event/32/contributions/515/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/515/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Statistical Diagnostics of Turboprop Engines Condition
DTSTART:20230912T074000Z
DTEND:20230912T080000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-420@conferences.enbis.org
DESCRIPTION:Speakers: Zuzana Hübnerová (Brno University of Technology)\,
Jaroslav Juračka (Brno University of Technology)\n\nModern digital instr
uments and SW options are standardly used in various areas\, of course als
o in aviation. Today\, the pilot is shown a number of physical parameters
of the flight\, the state of the propulsion or the aircraft's systems. The
se instruments also automatically save the scanned data. \n\nAnalysis of c
ollected data allows simultaneous surveillance of several aircraft turbopr
op engines related variables during each flight. Data collection and subse
quent continuous evaluation promise early detection of incipient damage or
a fouling before the regularly planned inspections. This fact could prolo
ng the service intervals and extend the engine Time Between Overhauls (TBO
).\n\nDue to the complexity of the dependencies among the acquisited engin
e parameters\, various operation conditions (atmospheric pressure\, temper
ature\, humidity) and flight profiles\, conventional statistical process c
ontrol procedures are not suitable for the diagnostics. In the paper\, a m
ethodology for identification of the changes in engine condition based on
regression analysis methods is proposed. The results for a thousand flight
records are presented and discussed as well.\n\nhttps://conferences.enbis
.org/event/32/contributions/420/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/420/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Feature Selection Method Based on Shapley Values Robust to Conce
pt Shift in Regression
DTSTART:20230911T100000Z
DTEND:20230911T103000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-589@conferences.enbis.org
DESCRIPTION:Speakers: Carlos Sebastián Martínez-Cava (Fortia Energía -
Universidad Politécnica de Madrid)\, Carlos González Guillén (Universid
ad Politécnica de Madrid)\n\nFeature selection is one of the most relevan
t processes in any methodology for creating a statistical learning model.
Generally\, existing algorithms establish some criterion to select the mos
t influential variables\, discarding those that do not contribute any rele
vant information to the model. This methodology makes sense in a classical
static situation where the joint distribution of the data does not vary o
ver time. However\, when dealing with real data\, it is common to encounte
r the problem of the dataset shift and\, specifically\, changes in the rel
ationships between variables (concept shift). In this case\, the influence
of a variable cannot be the only indicator of its quality as a regressor
of the model\, since the relationship learned in the traning phase may not
correspond to the current situation. Thus\, we propose a new feature sele
ction methodology for regression problems that takes this fact into accoun
t\, using Shapley values to study the effect that each variable has on the
predictions. Five examples are analysed: four correspond to typical situa
tions where the method matches the state of the art and one example relate
d to electricity price forecasting where a concept shift phenomenon has oc
curred in the Iberian market. In this case the proposed algorithm improves
the results significantly.\n\nhttps://conferences.enbis.org/event/32/cont
ributions/589/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/589/
END:VEVENT
BEGIN:VEVENT
SUMMARY:How Fair is Machine Learning in Credit Scoring?
DTSTART:20230912T161000Z
DTEND:20230912T163000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-574@conferences.enbis.org
DESCRIPTION:Speakers: Golnoosh Babaei (University of Pavia)\, Paolo Giudic
i (Department of Economics and Management\, University of Pavia\, Pavia\,
Italy)\n\nMachine learning (ML) algorithms\, in credit scoring\, are emplo
yed to distinguish between borrowers classified as class zero\, including
borrowers who will fully pay back the loan\, and class one\, borrowers who
will default on their loan. However\, in doing so\, these algorithms are
complex and often introduce discrimination by differentiating between indi
viduals who share a protected attribute (such as gender and nationality) a
nd the rest of the population. Therefore\, to make users trust these metho
ds\, it is necessary to provide fair and explainable models. To solve this
issue\, this paper focuses on fairness and explainability in credit scori
ng using data from a P2P lending platform in the US. From a methodological
viewpoint\, we combine ensemble tree models with SHAP to achieve explaina
bility\, and we compare the resulting Shapley values with fairness metrics
based on the confusion matrix.\n\nhttps://conferences.enbis.org/event/32/
contributions/574/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/574/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Wind Speed Analysis and Re-Simulation for Long-Term Wind Farm Prod
uction Forecast
DTSTART:20230912T065000Z
DTEND:20230912T071000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-477@conferences.enbis.org
DESCRIPTION:Speakers: Merlin Keller (EDF)\, Nicolas Paul (EDF R&D)\n\nWe a
ddress the task of predicting amount of energy produced during the total d
uration of a wind-farm project\, typically spanning several decades. This
is a crucial step to assess the project's return rate and convince potenti
al investors.\nTo perform such an assessment\, onsite mast measures at dif
ferent heights often provide accurate data over a few years\, together wit
h so-called satellite proxies\, given by global climate models calibrated
using satellite data\, less accurrate\, but available on a much longer tim
e scale\, but. Based on both sources of data\, several methods exist to pr
edict the wind speeds at the different turbine locations\, together with t
he energy production.\nThe aim of this work is to quantify the uncertainti
es tainting such a forecast\, based on a parametric bootstrap approach\, w
hich consist in re-simulating the onsite mast measures and satellite proxi
es\, then propagating their uncertainties throughout the whole procedure.\
nWe show that the satellite time-series can be accurately reproduced using
a spectral factorisation approach. Then\, the onsite measures are simulat
ed thanks to the so-called shear model\, which assumes an exponential vert
ical extrapolation of average wind speeds\, together with a Gaussian proce
ss model of the residuals. \nOur results allowed to detect and correct a b
ias in the existing calculation method\, leading to more accurate predicti
ons\, and reduced uncertainties.\nWe illustrate the benefits of our approa
ch on an actual project\, and discuss possible extension\, such as optimal
wind farm design\, and accounting for climate change.\n\nhttps://conferen
ces.enbis.org/event/32/contributions/477/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/477/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Bayesian Calibration for the Quantification of Conditional Uncerta
inty of Input Parameters in Chained Numerical Models
DTSTART:20230911T121000Z
DTEND:20230911T123000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-562@conferences.enbis.org
DESCRIPTION:Speakers: Amandine MARREL (CEA)\, Antoine BOULORÉ (CEA)\, Gui
llaume DAMBLIN (CEA)\, Loïc GIRALDI (CEA)\, Oumar BALDÉ (CEA)\n\nNumeri
cal models have become essential tools to study complex physical systems.
The accuracy and robustness of their predictions is generally affected by
different sources of uncertainty (numerical\, epistemic). In this work\,
we deal with parameter uncertainty of multiphysics simulation consisting
of several numerical models from different physics which are coupled with
one another. Our motivating application comes from the nuclear field where
we have a fission gas behavior model of the fuel inside a reactor core de
pending on a thermal model. As each of the two models has its own uncertai
n parameters\, our objective is to estimate the possible dependence betwee
n the uncertainty of input parameters $\\theta\\in \\mathbb{R}^p\\\, (p\\g
eq 1)$ of the gas model conditionally on the uncertainty of the fuel condu
ctivity $\\lambda\\in \\mathbb{R}$ of the thermal model. To do so\, we set
out a nonparametric Bayesian method\, based on several assumptions that a
re consistent with both the physical and numerical models. First\, the fun
ctional dependence $\\theta(\\lambda)$\, is assumed to be a realization
of Gaussian process prior whose hyperparameters are estimated on a set of
experimental data of the gas model. Then\, assuming that the gas model is
a linear function of $\\theta(\\lambda)$\, the Bayesian machinery allows u
s to compute analytically the posterior predictive distribution of $\\thet
a(\\lambda)$ for any set of realizations of the conductivity $\\lambda$. T
he shape of $\\theta(\\lambda)$ obtained shows the necessity of such a con
ditional parameter calibration approach in multiphysics simulation.\n\nhtt
ps://conferences.enbis.org/event/32/contributions/562/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/562/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Robust Multivariate Control Charts Based on Convex Hulls
DTSTART:20230912T163000Z
DTEND:20230912T165000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-531@conferences.enbis.org
DESCRIPTION:Speakers: Polychronis Economou (University of Patras)\, Frank
Bersimis (University of Piraeus)\, Sotiris Bersimis (University of Piraeus
\, Greece)\, Subha Chakraborti (University of Alabama)\n\nRobust multivari
ate control charts are statistical tools used to monitor and control multi
ple correlated process variables simultaneously. Multivariate control char
ts are designed to detect and signal when the joint distribution of the pr
ocess variables deviates from in-control levels\, indicating a potential o
ut-of-control case. The main goal of robust multivariate control charts is
to provide a comprehensive on-line assessment of the overall process perf
ormance. They are particularly useful in industries while their use is exp
anded today in other domains such as public health monitoring. Various sta
tistical techniques are applied to develop robust multivariate control cha
rts\, such as multivariate extensions of Shewhart\, EWMA and CUSUM control
charts. In this paper\, we propose a robust multivariate control chart ba
sed on the notion of convex hull. The notion of convex hull comes from the
domain of computational geometry\, and it is used to describe the smalles
t convex polygon or polyhedron that contains all the points in a data set.
Initial results of the proposed procedures give evidence of a very good p
erformance under different real-life cases.\n\nhttps://conferences.enbis.o
rg/event/32/contributions/531/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/531/
END:VEVENT
BEGIN:VEVENT
SUMMARY:The Effects of Large Round-Off Errors on the Performance of Contro
l Charts for the Mean
DTSTART:20230911T113000Z
DTEND:20230911T115000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-530@conferences.enbis.org
DESCRIPTION:Speakers: Diamanta Benson-Karhi (The Open University of Israel
)\n\nThis research discusses the effects of large round-off errors on the
performance of control charts for means when a process is normally distrib
uted with a known variance and a fixed sample size. Quality control in pra
ctice uses control charts for means as a process monitoring tool\, even wh
en the data is significantly rounded. The objective of this research is to
demonstrate how ignoring the round-off errors and using a standard Shewha
rt chart affects the quality control of a measured process.\nThe first par
t of the research includes theoretical calculations for estimating the val
ues of alpha\, beta\, ARL0\, and ARL1\, relating to the unrounded data and
the large-rounded data. For the rounded data\, normality can no longer be
assumed because the data is discrete\, therefore the multinomial distribu
tion is used. Results show that under the null hypothesis (H0)\, alpha and
ARL0 indicate that false alarms are more frequent. Under the alternative
hypothesis (H1)\, the influence on beta and ARL1 is minor and inconsistent
. For some rounding levels there is a decline in the control chart perform
ances and in others\, there is an improvement. In the second part\, a simu
lation study is used to evaluate the performances of the control chart bas
ed on a single sample\, checking whether the conclusion (reject or fail to
reject) for a sample is consistent for rounded and unrounded data. The re
sults of the simulation match the theoretical calculations.\n\nhttps://con
ferences.enbis.org/event/32/contributions/530/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/530/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Data-Driven Escalator Health Analytics and Monitoring
DTSTART:20230911T124000Z
DTEND:20230911T131000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-424@conferences.enbis.org
DESCRIPTION:Speakers: Inez Zwetsloot (City University of Hong Kong)\n\nMTR
\, the major Hong Kong public transport provider\, has been operating for
40 years with more than 1000 escalators in the railway network. These esca
lators are installed in various railway stations with different ages\, ver
tical rises and workload. An escalator’s refurbishment is usually linked
with its design life as recommended by the manufacturer. However\, the ac
tual useful life of an escalator should be determined by its operating con
dition which is affected by runtime\, workload\, maintenance quality\, vib
ration etc.\, rather than age only. \n\nThe objective of this project is t
o develop a comprehensive health condition model for escalators to support
refurbishment decisions. The analytic model consists of four parts: 1) on
line data gathering and processing\; 2) condition monitoring\; 3) health i
ndex model\; and 4) remaining useful life prediction. The results can be u
sed for 1) predicting the remaining useful life of the escalators\, in ord
er to support asset replacement planning and 2) monitoring the real-time c
ondition of escalators\; including signaling when vibration exceeds the th
reshold and signal diagnosis\, giving an indication of possible root cause
(components) of the signal. \n\nIn this talk\, we will provide a short ov
erview of this project and focus on the monitoring (part 3) of this projec
t where we use LSTM neural networks and PU (positive unlabeled) learning t
o set up a method that can deal with unstable vibration data that is unlab
eled.\n\nhttps://conferences.enbis.org/event/32/contributions/424/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/424/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Assessing Risk Indicators in Clinical Practice with Joint Models o
f Longitudinal and Time-to-Event Data
DTSTART:20230912T134500Z
DTEND:20230912T141500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-533@conferences.enbis.org
DESCRIPTION:Speakers: Eleni-Rosalina Andrinopoulou\n\nStudies in life cour
se epidemiology involve different outcomes and exposures being collected o
n individuals who are followed over time. These include longitudinally mea
sured responses and the time until an event of interest occurs. These outc
omes are usually separately analysed\, although studying their association
while including key exposures may be interesting. It is desirable to empl
oy methods that simultaneously examine all available information available
. This method is referred to as joint modelling of longitudinal and surviv
al data. The idea is to couple linear mixed effects models for longitudina
l measurement outcomes and Cox models for censored survival outcomes. \n\n
Joint modelling is an active area of statistics research that has received
much attention. These models can extract information from multiple marker
s objectively and employ them to update risk estimates dynamically. An adv
antage is that the predictions are updated as more measurements become ava
ilable\, reflecting clinical practice. The predictions can be combined wit
h the physician’s expertise to improve health outcomes. It is important
for physicians to have such a prognostic model to monitor trends over time
and plan their next intervention.\n\nSeveral challenges arise when obtain
ing predictions using the joint model. Different characteristics of the pa
tient's longitudinal profiles (underlying value\, slope\, area under the c
urve) could provide us with different predictions. Using a simulation stud
y\, we investigate the impact of misspecifying the association between the
outcomes. We present appropriate predictive performance measures for the
joint modelling framework to investigate the degree of bias. We present se
veral applications of real-world data in the clinical field.\n\nhttps://co
nferences.enbis.org/event/32/contributions/533/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/533/
END:VEVENT
BEGIN:VEVENT
SUMMARY:D-Optimal Experiment Design for Nested Sensor Placement
DTSTART:20230912T063000Z
DTEND:20230912T065000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-597@conferences.enbis.org
DESCRIPTION:Speakers: Andrew Titman (Lancaster University)\, Rebecca Killi
ck (Lancaster University)\, Louise Sugden (InTouch Ltd.)\, David Sudell (L
ancaster University)\n\nInternet of Things sensors placed in the environme
nt may be subject to a nested structure caused by local data relay devices
. We present an algorithm for D-optimal experiment design of the sensor pl
acement under these circumstances. This algorithm is an adaption of an exi
sting exchange algorithm sometimes called the Fedorov algorithm. The Fedor
ov exchange algorithm has been shown in the literature to perform well in
finding good designs with respect to the D-optimality criterion. Our adapt
ion of the algorithm is designed for the special case of a two-level nesti
ng structure imposed upon the potential design points of a linear model. T
he adapted algorithm shows effective identification of a known optimal des
ign in simulated cases and also appears to converge on a design(s) for fur
ther simulated datasets and an application dataset\, where the optimal des
ign(s) is unknown.\n\nhttps://conferences.enbis.org/event/32/contributions
/597/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/597/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Optimal Experimental Designs for Testing of LED Lighting
DTSTART:20230913T080000Z
DTEND:20230913T082000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-590@conferences.enbis.org
DESCRIPTION:Speakers: Alessandro Di Bucchianico (Eindhoven University of T
echnology)\n\nDue to the LED industry's rapid growth and the ease of manuf
acturing LED lights\, the LED market is highly competitive\, making good p
rice-quality ratio and being first-to-market crucial for manufacturers. To
that end\, accurate and fast lifetime testing is one of the key aspects f
or LED manufacturers. Lifetime testing of LED lighting typically follows e
xperimental and statistical techniques described in industry standards suc
h as LM80 and TM-21. \nIn this presentation we take a critical look at the
statistics behind these industry standards. We also critically examine th
e common practice of measuring LED lighting at equidistant points in time
during lifetime testing from the point of view of optimal experimental des
igns.\n\nhttps://conferences.enbis.org/event/32/contributions/590/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/590/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Bayesian Spatial Modeling for Misaligned Data Fusion
DTSTART:20230911T090000Z
DTEND:20230911T093000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-495@conferences.enbis.org
DESCRIPTION:Speakers: Paula Moraga (King Abdullah University of Science an
d Technology (KAUST))\n\nSpatially misaligned data are becoming increasing
ly common in fields such as epidemiology\, ecology and the environment due
to advances in data collection and management. Here\, we present a Bayesi
an geostatistical model for the combination of data obtained at different
spatial resolutions. The model assumes that underlying all observations\,
there is a spatially continuous variable that can be modeled using a Gauss
ian random field process. The model is fitted using the integrated nested
Laplace approximation (INLA) and the stochastic partial differential equat
ion (SPDE) approaches. In order to allow the combination of spatially misa
ligned data\, a new SPDE projection matrix for mapping the Gaussian Markov
random field from the observations to the triangulation nodes is proposed
. We show the performance of the new approach by means of simulation and a
n application of air pollution prediction in USA. The approach presented i
s fast and flexible\, can be extended to model spatio-temporal data and di
fferent sources of uncertainty\, and provides a useful tool in a wide rang
e of situations where information at different spatial scales needs to be
combined.\n\nhttps://conferences.enbis.org/event/32/contributions/495/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/495/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Changes and Trends in Mortalities in Relation to COVID-19
DTSTART:20230913T082000Z
DTEND:20230913T084000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-564@conferences.enbis.org
DESCRIPTION:Speakers: László Martinek (Eötvös Loránd University)\, Mi
klós Arató (Eötvös Loránd University)\n\nThe COVID-19 pandemic showed
that our mortality models need to be reviewed to adequately model the var
iability between years.\nOur presentation has the following objectives: (1
) We determine the time series of mortality changes in the European Union\
, United States\, United Kingdom\, Australia and Japan. Based on these tim
e series\, we estimate proximity measures between each pair of countries i
n terms of the excess mortality changes. (2) We examine the quality of som
e well-known stochastic mortality models (e.g. Lee-Carter) from the perspe
ctive of forecasting expected mortality and its variance over time. In add
ition\, we set up a ranking between the countries based on the excess mort
ality they suffered in 2020-2021. (3) We analyse the impact of COVID-19 al
ong the dimensions gender\, age group and country. Effects are modelled in
different ways.\nWe have used population mortality data from mortality.or
g and from Eurostat for our calculations.\n\nhttps://conferences.enbis.org
/event/32/contributions/564/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/564/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Local Linear Forests as a Solution for Online Process Control
DTSTART:20230912T071000Z
DTEND:20230912T073000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-596@conferences.enbis.org
DESCRIPTION:Speakers: Lucile Terras (EMSE (Ecole des Mines de Saint-Etienn
e))\, Cyril Alegret (STMicroelectronics)\, François Pasqualini (STMicroel
ectronics)\, Agnès Roussy (EMSE (Ecole des Mines de Saint-Etienne))\n\nIn
this study\, we propose to use the Local Linear Forest (R. Friedberg et a
l.\, 2020) to forecast the best equipment condition from complex and high-
dimensional semiconductor production data. In a static context\, the analy
sis performed on real production data shows that Local Linear Forests outp
erform the traditional Random Forest model and 3 other benchmarks. Each mo
del is finally integrated into an online advanced process control solution
\, where predictions made from continuous learning are used to automatical
ly adjust the recipe parameters of a production operation in real time. Th
rough the distribution of simulated process output\, we demonstrate how Lo
cal Linear Forests can effectively improve the quality of a mixed producti
on process in terms of variance reduction and process capability index imp
rovement. We compare the results with the control system in production and
demonstrate how this Machine Learning technique can be used as a support
for Industry 4.0.\nReference: Rina Friedberg\, Julie Tibshirani\, Susan At
hey\, and Stefan Wager. Local Linear Forests. Journal of Computational and
Graphical Statistics\, 30(2)\, 2020.\n\nhttps://conferences.enbis.org/eve
nt/32/contributions/596/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/596/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Model-Robust Subsampling Approach in Presence of Outliers
DTSTART:20230911T115000Z
DTEND:20230911T121000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-503@conferences.enbis.org
DESCRIPTION:Speakers: Chiara Tommasi (University of Milan)\, Laura Deldoss
i (Università Cattolica del Sacro Cuore)\n\nAbstract\nIn the era of big d
ata\, several sampling approaches are proposed to reduce costs (and time)
and to help in informed decision making. Most of these proposals require t
he specification of a model for the big data. This model assumption\, as w
ell as the possible presence of outliers in the big dataset\, represent a
limitation for the most commonly applied subsampling criterions.\nThe task
of avoiding outliers in a subsample of data was addressed by Deldossi et
al. (2023)\, who introduced non-informative and informative exchange algor
ithms to select “nearly” D-optimal subsets without outliers in a linea
r regression model. In this study\, we extend their proposal to account fo
r model uncertainty. More precisely\, we propose a model robust approach w
here a set of candidate models is considered\; the optimal subset is obtai
ned by merging the subsamples that would be selected by applying the appro
ach of Deldossi et al. (2023) if each model was considered as the true gen
erating process. \nThe approach is applied in a simulation study and some
comparisons with other subsampling procedures are provided.\n\nReferences\
nDeldossi\, L.\, Pesce\, E.\, Tommasi\, C. (2023) Accounting for outliers
in optimal subsampling methods\, Statistical Papers\, https://doi.org/10.1
007/s00362-023-01422-3.\n\nhttps://conferences.enbis.org/event/32/contribu
tions/503/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/503/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Statistics: Less Math and More Visual Thinking
DTSTART:20230911T153500Z
DTEND:20230911T155500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-592@conferences.enbis.org
DESCRIPTION:Speakers: Lourdes Pozueta (AVANCEX +I\, S.L.)\n\nThe exponenti
al integration of technologies in different disciplines\, the ease of acce
ss to data\, the proliferation of publications in Internet\, ... etc.\, ca
uses an increase in the number of new beliefs that try to explain the orig
in of the differences between behaviors with pseudoscientific discourses b
ased on data. People are not using Statistics well.\n\nStatistical profess
ionals can do good to society by sharing experiences that help to understa
nd Statistics as a multidisciplinary science of great VALUE.\n\nIn this ta
lk I will present some of my experiences that can inspire other:\n• Shar
ing success stories of applying statistics with children and young people.
I try to open their minds about Mathematics and also to break gender ster
eotypes. I motivate them with words like “discover what happen”\, what
type of patterns you see. We ended up talking about numbers\, mathematics
\, the importance of measuring well and the importance of Statistical Thin
king in all disciplines.\n• Participating in television programs that de
al with issues of great audience but misapplying Statistics (COVID)\n• P
resenting in generic Quality events applications of Statistics for Continu
ous Improvement and Innovation.\n• Disseminating in webinars for enginee
rs\, specific tactics\n\nStatistics deals with how to COLLECT data to EXTR
ACT useful information for decision making. We professionals have a missio
n to share what we do in a simple way. I recommend starting by training th
e look at the data\, the great VALUE in god visualizations\, and later\, w
e will talk about how to collect data\n\nhttps://conferences.enbis.org/eve
nt/32/contributions/592/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/592/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Innovations in Modelling Spectral Data
DTSTART:20230911T124000Z
DTEND:20230911T131000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-565@conferences.enbis.org
DESCRIPTION:Speakers: Phil Kay (SAS)\, Christopher Gotwalt (JMP Statistica
l Discovery LLC)\n\nSpectroscopy and chromatography data - from methods su
ch as FTIR\, NMR\, mass spectroscopy\, and HPLC - are ubiquitous in chemic
al\, pharmaceutical\, biotech and other process industries. Until now\, sc
ientists didn't have good ways to use this data as part of designed experi
ments or machine learning applications. They were required to ‘extract f
eatures’ such as the mean\, peak height\, or a threshold crossing point.
Summarising and approximating the spectral data in this way meant that mo
dels were less accurate and difficult to interpret.\n \nNow you can dire
ctly model these data types in designed experiments and machine learning a
pplications with Functional Data Explorer in JMP Pro. Wavelet analysis is
a new capability in the platform that make it easier than ever to build mo
dels that treat spectral data as first-class citizens in their own right.\
n\nhttps://conferences.enbis.org/event/32/contributions/565/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/565/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Multivariate Six Sigma: A Case Study in the Automotive Sector
DTSTART:20230912T065000Z
DTEND:20230912T071000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-481@conferences.enbis.org
DESCRIPTION:Speakers: Joan Borràs-Ferrís (Universitat Politècnica de Va
lència)\, Sergio García Carrión (Universitat Politècnica de València
(UPV))\, Lourdes Pozueta (AVANCEX +I\, S.L.)\, Alberto J. Ferrer-Riquelme
(Universidad Politecnica de Valencia)\n\nTraditional Six Sigma statistical
toolkit\, mainly composed of classical statistical techniques (e.g.\, sca
tter plots\, correlation coefficients\, hypothesis testing\, and linear re
gression models from experimental designs)\, is seriously handicapped for
problem solving in the Industry 4.0 era. The incorporation of latent varia
bles-based multivariate statistical techniques such as Principal Component
Analysis (PCA) [1] and Partial Least Squares (PLS) [2] into the Six Sigma
toolkit\, giving rise to the so-called Multivariate Six Sigma [3\, 4]\, c
an help to handle the complex data characteristics from this current conte
xt (e.g.\, high correlation\, rank deficiency\, low signal-to-noise ratio\
, and missing values).\nIn this work\, we present a multivariate Six Sigma
case study\, related to a lack of capability issue with vibration toleran
ces for a part of the car's brake system. We illustrate the benefits of th
e integration of latent variables-based multivariate statistical technique
s into the five-step DMAIC cycle\, achieving a more efficient methodology
for process improvement in Industry 4.0 environments.\n[1] S. Wold\, K. Es
bensen\, and P. Geladi\, “Principal component analysis\,” Chemometrics
and intelligent laboratory systems\, 2(1–3):37–52\, 1987.\n[2] S. Wol
d\, M. Sjöström\, and L. Eriksson\, “PLS-regression: a basic tool of c
hemometrics\,” Chemometrics and Intelligent Laboratory Systems\, 58(2):1
09–130\, 2001.\n[3] A. Ferrer\, “Multivariate six sigma: A key improve
ment strategy in industry 4.0\,” Quality Engineering\, 33(4):758–763\,
2021.\n[4] D. Palací-López\, J. Borràs-Ferrís\, L. T. da Silva de Oli
veria\, and A. Ferrer\, “Multivariate six sigma: A case study in industr
y 4.0\,” Processes\, 8(9):1119\, 2020.\n\nhttps://conferences.enbis.org/
event/32/contributions/481/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/481/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Multivariate Six Sigma: A Case Study in a Chemical Industry
DTSTART:20230912T063000Z
DTEND:20230912T065000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-480@conferences.enbis.org
DESCRIPTION:Speakers: Daniel Palací-López (IFF Benicarlo\, Benicarlo\, S
pain)\, Sergio García-Carrión (Universitat Politècnica de València)\,
Joan Borràs-Ferrís (Universitat Politècnica de València)\, Alberto J.
Ferrer-Riquelme (Universidad Politecnica de Valencia)\n\nThe large volume
of complex data being continuously generated in Industry 4.0 environments\
, usually coupled with significant restrictions on experimentation in prod
uction\, tends to hamper the application of the classical Six Sigma method
ology for continuous improvement\, for which most statistical tools are ba
sed in least squares techniques. Multivariate Six Sigma [1]\, on the other
hand\, incorporates latent variables-based techniques such as principal c
omponent analysis or partial leas squares\, overcoming such limitation.\nH
owever\, trying to optimize very tightly controlled processes\, for which
very small variability is allowed for the critical to quality characterist
ic of interest\, may still pose a challenge in this case. This is because\
, in absence of first-principles models\, data-based empirical models are
required for optimization\, but such models will perform poorly when the r
esponse variable barely varies. This is typically the case in a lot of che
mical processes where the selectivity of a reaction has remained mostly co
nstant in the past\, but then an improvement is required on it: since hist
orical data shows not enough excitement in this parameter\, no model can b
e built to optimize it.\nThis work presents the challenges in applying the
Multivariate Six Sigma methodology to a chemical reaction in a real indus
trial case study in order to optimize its selectivity\, for which a proper
predictive model could not be directly obtained.\n[1] A. Ferrer\, “Mult
ivariate six sigma: A key improvement strategy in industry 4.0\,” Qualit
y Engineering\, 33(4):758–763\, 2021.\n\nhttps://conferences.enbis.org/e
vent/32/contributions/480/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/480/
END:VEVENT
BEGIN:VEVENT
SUMMARY:bayespm: BAYESian Process Monitoring in R
DTSTART:20230913T092500Z
DTEND:20230913T094500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-599@conferences.enbis.org
DESCRIPTION:Speakers: Dimitrios Kiagias (School of Mathematics and Statist
ics\, University of Sheffield\, UK)\, Konstantinos Bourazas (Dept. of Math
ematics and Statistics & KIOS Research and Innovation Center of Excellence
\, University of Cyprus\, Cyprus)\, Panagiotis Tsiamyrtzis (Dept. of Mecha
nical Engineering\, Politecnico di Milano\, Italy & Dept. of Statistics\,
Athens University of Economics and Business\, Greece)\n\nThe univariate Ba
yesian approach to Statistical Process Control/Monitoring (BSPC/M) is know
n to provide control charts that are capable of monitoring efficiently the
process parameters\, in an online fashion from the start of the process i
.e.\, they can be considered as self-starting since they are free of a pha
se I calibration. Furthermore\, they provide a foundational framework that
utilizes available prior information for the unknown parameters\, along w
ith possible historical data (via power priors)\, leading to more powerful
tools when compared to the frequentist based self-starting analogs. Use o
f non-informative priors allow these charts to run even when no prior info
rmation exists at all. Two big families of such univariate BSPC/M control
charts are the Predictive Control Chart (PCC) and the Predictive Ratio Cus
um (PRC). PCCs are specialized in identifying transient parameter shifts (
i.e.\, outliers) of moderate/large size\, while PRCs are focused on detect
ing persistent parameter shifts of even small size. Both PCC and PRC are g
eneral\, closed form mechanisms\, capable of handling data from any discre
te or continuous distribution\, as long as it belongs to the regular expon
ential family (e.g.\, Normal\, Binomial\, Poisson\, etc.). In this work\,
we will present the R package bayespm which implements the PCC and/or PRC
control charts for any data set that comes from a discrete or a continuous
distribution and is a member of the regular exponential family. Real data
examples will illustrate the various options that include online monitori
ng along with inference for the unknown parameters of a univariate process
.\n\nhttps://conferences.enbis.org/event/32/contributions/599/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/599/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Examining the impact of critical attributes on hard drive failure
times: multi-state models for left-truncated and right-censored semi-compe
ting risks data
DTSTART:20230911T121000Z
DTEND:20230911T123000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-620@conferences.enbis.org
DESCRIPTION:Speakers: Pete Philipson (School of Mathematics\, Statistics &
Physics\, Newcastle University\, United Kingdom)\, Kevin J. Wilson (Scho
ol of Mathematics\, Statistics & Physics\, Newcastle University\, United K
ingdom)\, Matthew Forshaw (School of Computing\, Newcastle University\, U
nited Kingdom)\, Jordan Oakley (School of Mathematics\, Statistics & Physi
cs\, Newcastle University\, United Kingdom)\n\nA recent study based on dat
a from Microsoft reports that 76 − 95% of all failed components in data
centres are hard drives. HDDs are the main reason behind server failures.
Consequently\, the ability to predict failures in hard disk drives (HDDs)
is a major objective of HDD manufacturers since avoiding unexpected failur
es may prevent data loss\, improve service reliability\, and reduce data c
entre downtime. Most HDDs are equipped with a threshold-based monitoring s
ystem named Self-Monitoring\, Analysis and Reporting Technology (SMART). T
he system collects performance metrics\, called SMART attributes\, and det
ects anomalies that may indicate incipient failures.\n\nIn this talk\, we
define critical attributes and critical states for hard drives using SMART
attributes and fit multi-state models to the resulting semi-competing ris
ks data. The multi-state models provide a coherent and novel way to model
the failure time of a hard drive and allow us to examine the impact of cri
tical attributes on the failure time of a hard drive. We derive prediction
s of conditional survival probabilities\, which are adaptive to the state
of the drive. Using a dataset of HDDs equipped with SMART\, we find that d
rives are more likely to fail after entering critical states. We evaluate
the predictive accuracy of the proposed models with a case study of HDDs e
quipped with SMART\, using the time-dependent area under the receiver oper
ating characteristic curve and the expected prediction error. The results
suggest that accounting for changes in the critical attributes improves th
e accuracy of predictions.\n\nhttps://conferences.enbis.org/event/32/contr
ibutions/620/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/620/
END:VEVENT
BEGIN:VEVENT
SUMMARY:New CUSUM Charts\, the GLR Procedure and the Parabolic Mask
DTSTART:20230912T155000Z
DTEND:20230912T161000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-603@conferences.enbis.org
DESCRIPTION:Speakers: Sven Knoth (Helmut Schmidt University Hamburg\, Germ
any)\n\nThe cumulative sum (CUSUM) control chart iterates sequential proba
bility ratio tests (SPRT) until the first SPRT ends with rejecting the nul
l hypothesis. Because the latter exhibits some deficiencies if the true me
an is substantially different to the one used in the underlying likelihood
ratio\, Abbas (2023) proposes to substitute the SPRT by a repeated signif
icance test (RST)\, cf. to Armitage et al. (1969). To fix the latter's mis
sing ability to renewal (core element of the CUSUM chart)\, Abbas (2023) c
ombines SPRT und RST. The resulting control chart\, labelled as "step CUSU
M"\, performs quite well for a wide range of potential shifts in the mean
of a normal random variable. However\, the older generalized likelihood ra
tio (GLR) procedure\, e. g. Reynolds & Lou (2010)\, deploys similar alarm
thresholds and performs even better. Both are more difficult to analyze th
an the plain CUSUM chart. Interestingly\, the GLR scheme is equivalent to
applying a parabolic mask (Wiklund 1997). The GLR procedure experienced qu
ite some up and downs during the last decades\, but it should be more used
in routine monitoring work. Eventually\, some reflections upon the cost-b
enefit relation are given.\n\nReferences\n\nAbbas (2023)\, On efficient ch
ange point detection using a step cumulative sum control chart\, QE\, http
s://doiorg/10.1080/08982112.2023.2193896\, 1--17\n\nArmitage\, McPherson\,
Rowe (1969)\, Repeated Significance Tests on Accumulating Data\, JRSSA 13
2(2)\, 235--244 \n\nReynolds Jr.\, Lou (2010)\, An Evaluation of a GLR Con
trol Chart for Monitoring the Process Mean\, JQT 42(3)\, 287--310\n\nWiklu
nd (1997)\, Parabolic cusum control charts\, CSSC 26(1)\, 107--123\n\nhttp
s://conferences.enbis.org/event/32/contributions/603/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/603/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Tremendous Impact of the Very New and Promising OMARS DOE in Pharm
a Industry for Quicker Access to New Vaccines
DTSTART:20230913T082000Z
DTEND:20230913T084000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-602@conferences.enbis.org
DESCRIPTION:Speakers: Emilie Ansel (GSK)\, Bernard Francq (GSK)\, Pierre-Y
ves Vitry (GSK)\, Laurent Ferrant (GSK)\, Pascal Gerkens (GSK)\n\nIn the p
ast\, screening (which process parameters are impactful) and optimisation
(optimise the response variable or the critical quality attribute\, CQA) w
ere 2 distinct phases performed by 2 designs of experiments (DoE). Then\,
the definitive screening designs (DSDs) published approximately 10 years a
go attracted a lot of attention from both statisticians and non-statistici
ans\, espcially in the pharma industry. The idea is to combine screening a
nd optimisation in a single step. This allows to reduce the total number o
f experiments and the research development time with a substantial gain in
the budget.\n\nRecently\, a new type of DoE called OMARS for orthogonal m
inimally aliased response surface has been published. These OMARS DoEs out
perform DSDs in many criteria. Firstly\, the orthogonality criteria where
the independence between main effects is fulfilled\, and also between main
effects and interaction terms. Secondly\, the projection property where O
MARS designs are able to estimate a response surface model (main effects\,
interactions\, quadratic terms) from the remaining significant parameters
. OMARSs also outperform DSDs in presence of categorical factors.\n\nIn th
is presentation\, we will assess the impact of the new OMARS DoEs in the p
harma industry. A case study will be used with fermentations on Ambr syste
m for vaccine development (with 6 process parameters and multiple response
variables) where the OMARS DoE allows to cut at least by 2 the total numb
er of experiments. Finally\, it will be shown that the OMARS substantially
accelerates the R&D process and shortens the time-to-market of future dru
gs and vacccines.\n\nhttps://conferences.enbis.org/event/32/contributions/
602/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/602/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Sensitivity Analysis in the Presence of Hierarchical Variables
DTSTART:20230912T080000Z
DTEND:20230912T082000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-545@conferences.enbis.org
DESCRIPTION:Speakers: Julien Pelamatti (EDF R&D)\, Vincent Chabridon (EDF
R&D)\n\nIn the context of sensitivity analysis\, the main objective is to
assess the influence of various input variables on a given output of inter
est\, and if possible to rank the influential inputs according to their re
lative importance. In many industrial applications\, it can occur that th
e input variables present a certain type of hierarchical dependence struct
ure. For instance\, depending on some architectural choices (e.g.\, combus
tion or electric motor technology)\, which can be seen as parents variable
s\, some of the children variables (e.g.\, engine battery weight) may or
may not have an effect on the output of interest. When dealing with given-
data sensitivity analysis\, this may result in missing tabular data\, as t
he inactive children variables may not make physical sense\, or may not be
measurable (e.g.\, number of pistons for an electric motor). In this wor
k\, we focus on a hierarchical and functional type of relation between the
inputs for the purpose of performing sensitivity analysis. The aim of thi
s work is to propose an adaptation of existing sensitivity indices to accu
rately quantify the influence of all inputs on the output while taking int
o consideration their hierarchical dependencies. An adaptation of Sobol’
sensitivity indices is studied and two given-data estimators are suggeste
d. The theoretical analysis and numerical tests on different toy-cases\,
as well as on a real-world industrial data set\, show promising results in
terms of interpretability\, but also highlight some limitations regarding
the indices estimation with limited amounts of data and in the presence o
f statistical dependence between inputs.\n\nhttps://conferences.enbis.org/
event/32/contributions/545/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/545/
END:VEVENT
BEGIN:VEVENT
SUMMARY:It’s About Time – the Impact of Time Delay and Time Dynamics o
n Soft Sensing in Industrial Data
DTSTART:20230913T082000Z
DTEND:20230913T084000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-433@conferences.enbis.org
DESCRIPTION:Speakers: Marco Cattaldo (Nofima)\, Alberto J. Ferrer-Riquelme
(Universidad Politecnica de Valencia)\, Ingrid Måge (Nofima AS)\n\nThe i
ncreasing affordability of physical and digital sensors has led to the ava
ilability of considerable process data from a range of production processe
s. This trend\, in turn\, has enabled researchers and industrial practitio
ners to employ these large amounts of data to improve process efficiency a
t most levels\, thereby facilitating the operation of the process. A funda
mental step in some of these applications is to obtain a frequent and reli
able prediction of a quantity that is either impractical\, impossible\, or
time-consuming to measure to use as a surrogate in further modelling or c
ontrol steps. These surrogate measurements are usually derived by utilisin
g models that link easy-to-measure process variables to the quantities of
interest\; these models are frequently called “soft sensors”.\nIn deve
loping soft sensors for continuous processes\, it is common to have time d
elays and dynamics in the data\, as both are intrinsic to continuous produ
ction processes and how they are operated. It is essential to consider the
se aspects when developing the soft sensor\, as they can be detrimental to
the soft sensor’s performance.\nIn this contribution\, we illustrate an
d compare different techniques to account for time delay[1–5] and dynami
cs[6–9] in the pre-processing and modelling steps of soft sensor develop
ment. On the time delay side\, these techniques vary from the classical co
rrelation coefficient to information-theoretic measurement and complex opt
imiser-based methods\, while on the time dynamics side\, the focus is main
ly on dynamic latent variable methods.\nThe work is based on a real case s
tudy from the food industry.\n\nhttps://conferences.enbis.org/event/32/con
tributions/433/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/433/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Predictive Models for the Family Life Cycle in the Banking Enviro
nment
DTSTART:20230911T134000Z
DTEND:20230911T141000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-616@conferences.enbis.org
DESCRIPTION:Speakers: Lidia López Fernández (ABANCA)\n\nThe family life
cycle is a theoretical model that describes the different stages that a fa
mily normally goes through during its life. These stages are associated wi
th changes in the family nucleus composition and with the relations betwee
n members. From a banking point of view\, it is important to note that the
financial needs of the family will also change throughout its life. There
fore\, the aim of this work is to build a model using statistical learning
techniques\, such as the supervised classification XGBoosting model\, tha
t provide information of the stage of the family life cycle corresponding
to each client\, to offer them the financial products that best suit their
needs. \n\nTherefore\, we collect the socio-demographic\, financial and t
ransactional internal bank information. These data allow bank personnel to
estimate the family stage of the bank adult clients\, by XGBoost. They ar
e calibrated by a training\, validation and test process. All the used mod
els are evaluated and compared using suitable metrics.\n \nThe information
provided by the proposed methodology will be included in the propensity m
odels used by the bank. It will be used to improve bank tasks such as the
developing of propensity models referred to the contracting of a life insu
rance. For example\, when a person has children\, they need to ensure cert
ain capital for them in case of death\, incapacity or other unpredictable
issue. Consequently\, we will be able to estimate which clients have a hig
h probability of having children\, and thus need this type of insurance.\n
\nhttps://conferences.enbis.org/event/32/contributions/616/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/616/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Novel Low-Dimensional Learning Approach for Automated Classifica
tion of 2-D Microstructure Data in Additive Manufacturing
DTSTART:20230912T090500Z
DTEND:20230912T093500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-605@conferences.enbis.org
DESCRIPTION:Speakers: Wei Yang (1H. Milton Stewart School of Industrial &
Systems Engineering\, Georgia Institute of Technology)\, Bianca Maria Colo
simo (2Department of Mechanical Engineering\, Politecnico di Milano)\, Kam
ran Paynabar (1H. Milton Stewart School of Industrial & Systems Engineerin
g\, Georgia Institute of Technology)\, Mohammad Najjartabar Bisheh (1H. M
ilton Stewart School of Industrial & Systems Engineering\, Georgia Institu
te of Technology)\, Marco Grasso (Politecnico di Milano\, Department of Me
chanical Engineering)\n\nNovel production paradigms like metal additive ma
nufacturing (AM) have opened many innovative opportunities to enhance and
customize product performances in a wide range of industrial applications.
In this framework\, high-value-added products are more and more character
ized by novel physical\, mechanical and geometrical properties. Innovative
material performances can be enabled by tuning microstructural properties
and keeping them stable and repeatable from part to part\, which makes mi
crostructural analysis of central importance in process monitoring and qua
lification procedures. The industrial practice for microstructural image d
ata analysis currently relies on human expert’s evaluations. In some cas
es\, grain size and morphology are quantified via synthetic metrics like t
he mean grain diameter\, but these features are not sufficient to capture
all the salient properties of the material. Indeed\, there is a lack of me
thods suited to automatically extracting informative features from complex
2-D microstructural data and utilizing them to classify the microstructur
es. Aiming to fill this gap\, this study presents a novel low-dimensional
learning approach\, where both the morphological grain properties and the
crystal orientation distribution features are extracted and used to cluste
r real microstructure data into different groups moving from complex 2D pa
tterns to a lower-dimensional data space. A case study in the field of met
al AM is proposed\, where the proposed methodology is tested and demonstra
ted on electron backscattered diffraction (EBSD) measurements. The propose
d methodology can be extended and generalized to other applications\, and
to a broader range of microstructures.\n\nhttps://conferences.enbis.org/ev
ent/32/contributions/605/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/605/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Large Batch Sampling for Boundary Estimation Using Active Learning
: A Case Study from Additive Manufacturing
DTSTART:20230911T121000Z
DTEND:20230911T123000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-608@conferences.enbis.org
DESCRIPTION:Speakers: Bianca Maria Colosimo (Politecnico di Milano)\, Stef
ania Cacace (Politecnico di Milano)\, Kamran Paynabar (1H. Milton Stewart
School of Industrial & Systems Engineering\, Georgia Institute of Technolo
gy)\n\nThis paper explores the problem of estimating the contour location
of a computationally expensive function using active learning. Active lear
ning has emerged as an efficient solution for exploring the parameter spac
e when minimizing the training set is necessary due to costly simulations
or experiments.\nThe active learning approach involves selecting the next
evaluation point sequentially to maximize the information obtained about t
he target function. To this aim\, we propose a new entropy-based acquisiti
on function specifically designed for efficient contour estimation. Additi
onally\, we address the scenario where a large batch of query points is ch
osen at each iteration. While batch-wise active learning offers efficiency
advantages\, it also presents challenges since the informativeness of the
query points depends on the accuracy of the estimated function\, particul
arly in the initial iterations.\nTo illustrate the significance of our wor
k\, we employ the estimation of processability window boundaries in Additi
ve Manufacturing as a motivating example. In experimental campaigns using
this technology\, a large number of specimens is printed simultaneously to
accommodate time and budget constraints. Our results demonstrate that the
proposed methodology outperform standard entropy-based acquisition functi
ons and space-filling design\, leading to potential savings in energy and
resource utilization.\n\nhttps://conferences.enbis.org/event/32/contributi
ons/608/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/608/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Data Science and Statistical Machine Learning in Industry 4.0: Per
sonal Reflections
DTSTART:20230911T113000Z
DTEND:20230911T115000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-572@conferences.enbis.org
DESCRIPTION:Speakers: Alberto J. Ferrer-Riquelme (Universidad Politecnica
de Valencia)\n\nData Science has emerged to deal with the so-called (big)
data tsunami. This has led to the Big Data environment\, characterized by
the four Vs: volume\, variety\, velocity\, and veracity. We live in a new
era of digitalization where there is a belief that due to the amount and s
peed of data production\, new technologies coming from artificial intellig
ence could now solve important scientific and industrial problems solely t
hrough the analysis of empirical data\, without the use of scientific mode
ls\, theory\, experience\, or domain knowledge. In this talk I will discus
s on the risk of this belief and on some insights about statistical machin
e learning\, that is\, the integration of machine learning with statistica
l thinking and methods (mainly latent variables-based) to succeed in probl
em solving\, and process improvement and optimization in Industry 4.0.\n\n
Ferrer\, A. (2020). Appl Stochastic Models Bus Ind.\, 36:23–29. \nDOI: 1
0.1002/asmb.2516\n\nhttps://conferences.enbis.org/event/32/contributions/5
72/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/572/
END:VEVENT
BEGIN:VEVENT
SUMMARY:SMB-PLS for Expanding Multivariate Raw Material Specifications in
Industry 4.0
DTSTART:20230911T131000Z
DTEND:20230911T134000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-487@conferences.enbis.org
DESCRIPTION:Speakers: Joan Borràs-Ferrís (Universitat Politècnica de Va
lència)\, Alberto Ferrer (Universitat Politècnica de València )\, Carl
Duchesne (Laval University )\n\nThe advantages of being able to define pre
cisely meaningful multivariate raw material specifications are enormous. T
hey allow increasing the number of potential suppliers\, by allowing a wid
er range of raw material properties\, without compromising the Critical Qu
ality Attributes (CQAs) of the final product. Despite their importance\, s
pecifications are usually defined in an arbitrary way based mostly on subj
ective past experience\, instead of using a quantitative objective descrip
tion of their impact on CQAs. Moreover\, in many cases\, univariate specif
ications on each property are designated\, with the implicit assumption th
at these properties are independent from one another. Nevertheless\, multi
variate specifications provide much insight into what constitutes acceptab
le raw material batches when their properties are correlated (as usually h
appens) [1]. To cope with this correlation several authors suggest using m
ultivariate approaches\, such as Partial Least Squares (PLS) [2]. \nBeside
s\, not only raw material properties influence the quality of the final pr
oduct\, but also process conditions. Thus\, we propose a novel methodology
\, based on the Sequential Multi-block PLS (SMB-PLS)\, to identify the var
iation in process conditions uncorrelated with raw material properties\, w
hich is crucial to implement an effective process control system attenuati
ng most raw material variations. This allows expanding the specification r
egion and\, hence\, one may potentially be able to accept lower cost raw m
aterials that will yield products with perfectly satisfactory quality prop
erties.\n\n[1] C. Duchesne and J. F. MacGregor\, J. Qual. Technol.\, 36\,
78–94\, 2004.\n[2] J. Borràs-Ferrís\, D. Palací-López\, C. Duchesne\
, and A. Ferrer\, Chemom. Intell. Lab. Syst.\, 225\, 2022.\n\nhttps://conf
erences.enbis.org/event/32/contributions/487/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/487/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Broadening the Spectrum of OMARS Designs
DTSTART:20230912T144000Z
DTEND:20230912T150000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-521@conferences.enbis.org
DESCRIPTION:Speakers: Peter Goos (University of Leuven)\, José Núñez Ar
es (University of Leuven)\n\nThe family of orthogonal minimally aliased re
sponse surface designs or OMARS designs bridges the gap between the small
definitive screening designs and classical response surface designs. The i
nitial OMARS designs involve three levels per factor and allow large numbe
rs of quantitative factors to be studied efficiently. Many of the OMARS de
signs possess good projection properties and offer better powers for quadr
atic effects than definitive screening designs with similar numbers of run
s. Therefore\, OMARS designs offer the possibility to perform a screening
experiment and a response surface experiment in a single step\, and the op
portunity to speed up innovation. The initial OMARS designs study every qu
antitative factor at its middle level the same number of times. As a resul
t\, every main effect can be estimated with the same precision\, the power
is the same for every main effect\, and the quadratic effect of every fac
tor has the same probability of being detected. We will show how to create
"non-uniform-precision OMARS designs" in which the main effects of some f
actors are emphasized at the expense of their quadratic effects\, or vice
versa. Relaxing the uniform-precision requirement opens a new large can of
useful three-level experimental designs. The new designs form a natural c
onnection between the initial OMARS design\, involving three levels for ev
ery factor and corresponding to one end of the OMARS spectrum\, and the mi
xed-level OMARS designs\, which involve three levels for some factors and
two levels for other factors and correspond to another end of the OMARS sp
ectrum.\n\nhttps://conferences.enbis.org/event/32/contributions/521/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/521/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Multi-Criteria Evaluation and Selection of Experimental Designs fr
om a Catalog
DTSTART:20230912T065000Z
DTEND:20230912T071000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-492@conferences.enbis.org
DESCRIPTION:Speakers: Jose Nunez Ares (KU Leuven)\, Peter Goos (KU Leuven)
\n\nIn recent years\, several researchers have published catalogs of exper
imental plans. First\, there are several catalogs of orthogonal arrays\, w
hich allow experimenting with two-level factors as well as multi-level fac
tors. The catalogs of orthogonal arrays with two-level factors include alt
ernatives to the well-known Plackett-Burman designs. Second\, recently\, a
catalog of orthogonal minimally aliased response surface designs (or OMAR
S designs) appeared. OMARS designs bridge the gap between the small defini
tive screening designs and the large central composite designs\, and they
are economical designs for response surface modeling. The catalogs contain
dozens\, thousands or millions of experimental designs\, depending on the
number of runs and the number of factors\, and choosing the best design f
or a particular problem is not a trivial matter. In this presentation\, we
introduce a multi-objective method based on graphical tools to select a d
esign. Our method analyzes the trade-offs between the different experiment
al quality criteria and the design size\, using techniques from multi-obje
ctive optimization. Our procedure presents an advantage compared to the op
timal design methodology\, which usually considers only one criterion for
generating an experimental design. Additionally\, we will show how our met
hodology can be used for both screening and optimization experimental desi
gn problems. Finally\, we will demonstrate a novel software solution\, ill
ustrating its application for a few industrial experiments.\n\nhttps://con
ferences.enbis.org/event/32/contributions/492/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/492/
END:VEVENT
BEGIN:VEVENT
SUMMARY:On the Opportunities and Limitations of Deep Artificial Intelligen
ce Methods for Industrial Process Analytics
DTSTART:20230912T074000Z
DTEND:20230912T080000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-470@conferences.enbis.org
DESCRIPTION:Speakers: Marco P. Seabra dos Reis (Department of Chemical Eng
ineering\, University of Coimbra)\n\nThe use of data for supporting induct
ive reasoning\, operational management\, and process improvement\, has bee
n a driver for progress in modern industry. Many success stories have been
shared on the successful application of data-driven methods to address di
fferent open challenges\, across different industrial sectors. The recent
advances in AI/ML technology in the fields of image & video analysis and n
atural language have spiked the interest of the research community to expl
ore their application outside these domains\, namely in the chemical\, foo
d\, biotechnological\, semiconductor\, and pharmaceutical industries\, amo
ng others. \nBut this boost in activity has also increased the difficulty
of understanding the multiple underlying rationales for applying them\, ot
her than the mere curiosity of “to see what comes out” (still valid\,
but arguably inefficient). Furthermore\, it is often difficult to assess t
he added value of using these new methods\, as many times they are not rig
orously compared with conventional solutions presenting state-of-the-art p
erformances. \nTherefore\, it is now opportune to discuss the role of the
new wave of AI in solving industrial problems\, supported by a fair and un
passionate assessment of their added value. Also\, looking at a wider pict
ure of the approaches that operate by induction from data\, another aspect
to bring to the table regards how to find the best balance and take the m
ost of the possible synergies between statistics\, machine learning\, and
deep AI. These questions will be addressed in the talk\, and examples will
be presented and discussed.\n\nhttps://conferences.enbis.org/event/32/con
tributions/470/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/470/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Degradation Process Monitoring in Agro-Food Industry Using Multiva
riate Image Analysis
DTSTART:20230912T071000Z
DTEND:20230912T073000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-568@conferences.enbis.org
DESCRIPTION:Speakers: Lourdes Pozueta (AVANCEX +I\, S.L.)\, Alberto J. Fer
rer-Riquelme (Universidad Politecnica de Valencia)\, Alberto Ferrer-Hermen
egildo (Universitat Politècnica de València (UPV))\, José Manuel Prats-
Montalbán (Universtitat Politècnica de València)\n\nIn this talk we int
roduced a multivariate image analysis (MIA)-based quality monitoring syste
m for the detection of batches of a vegetable fresh product (Iceberg type
lettuce) that do not meet the established quality requirements. This tool
was developed in the Control stage of the DMAIC cycle of a Six Sigma Multi
variate project undertaken in a company of the agri-food sector.\n\nAn exp
erimental design was carried out by taking RGB pictures of lettuce trays s
tored at two temperatures (room and fridge) every 12 hours for 5 days. By
using RGB images obtained only from fresh lettuce trays\, a MIA-based prin
cipal component analysis (MIA-PCA) model extracting color and textural inf
ormation was built. By exploring the PCA loadings we discovered that a fou
r-component MIA-PCA model was able to provide information about degradatio
n in terms of loss of color intensity\, dehydration and appearance of brow
n areas. Afterwards\, the RGB data obtained from the experimental design w
ere projected onto this model and Hotelling-T2 and SPE values obtained and
plotted: the degradation process was clearly shown in the lettuce trays s
tored at room temperature.\n\nFinally\, a Shewhart individual control char
t was built from the Hotelling-T2 values obtained from fresh lettuce trays
. Applying the graph to experimental data\, the lettuce trays stored at fr
idge temperature were under control during the five days but those stored
at room temperature showed a progressive signal of out-of-control at 12 ho
urs onwards.\n\nThe propose control chart allows the online rejection of l
ow-quality lettuce at the reception stage from suppliers.\n\nhttps://confe
rences.enbis.org/event/32/contributions/568/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/568/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Measurement of Thermal Conductivity at a Nanoscale Using Bayesian
Inversion
DTSTART:20230911T121000Z
DTEND:20230911T123000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-509@conferences.enbis.org
DESCRIPTION:Speakers: Sarah Douri (LNE)\, Nolwenn Fleurence (LNE)\, Bruno
Hay (LNE)\, Séverine Demeyer (LNE)\n\nThermal management is a key issue f
or the miniaturization of electronic devices due to overheating and local
hot spots. To anticipate these failures\, manufacturers require knowledge
of the thermal properties of the used materials at the nanoscale (defined
as the length range from 1 nm to 100 nm)\, which is a challenging issue be
cause thermal properties of materials at nanoscale can be completely diffe
rent from those of the bulk materials (materials having their size above 1
00 nm in all dimensions). \nThe proposed approach aims at establishing a c
alibration curve (as part of a calibration protocol) to provide metrologic
ally traceable estimations of the thermal conductivity at nanoscale and it
s associated uncertainty (x-axis)\, using SThM (Scanning Thermal Microscop
y\, having a high spatial resolution of tens of nm) measurements and their
associated uncertainty (y-axis).\nThis contribution focuses on the develo
pment of a Bayesian approach to simultaneously estimate the calibration cu
rve with uncertainty on both axes and to predict the thermal conductivity
of unknown materials and their associated uncertainty.\nThe approach is ap
plied to 12 samples of bulk materials with traceable thermal conductivitie
s with 5% relative expanded uncertainty in the range 1-100 $Wm^{-1}K^{-1}$
. For these materials\, uncertainty on the y-axis ranges between 0.4% and
2% relative expanded uncertainty.\nWith this methodology\, a thermal condu
ctivity of 0.2 $Wm^{-1}K^{-1}$ is estimated with less that 4 % relative un
certainty.\nThe effect of uncertainty sources (in particular on the y-axis
) on the range of sensitivity of the SThM technique for quantitative therm
al conductivity measurements is investigated.\n\nhttps://conferences.enbis
.org/event/32/contributions/509/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/509/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Conformity Assessment of a Sample of Items
DTSTART:20230911T093000Z
DTEND:20230911T100000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-510@conferences.enbis.org
DESCRIPTION:Speakers: Francesca Pennecchi (Istituto Nazionale di Ricerca M
etrologica - INRIM)\n\nA document of the Joint Committee for Guides in Met
rology [JCGM 106:2012 - Evaluation of measurement data – The role of mea
surement uncertainty in conformity assessment] provides a Bayesian approac
h to perform conformity assessment (CA) of a scalar property of a single i
tem (a product\, material\, object\, etc.). It gives a methodology to calc
ulate specific and global risks of false decisions for both the consumer a
nd the producer. Specific risks\, which are conditional probabilities\, ar
e related to a specific item whose property has been measured. Global risk
s\, which are probabilities of joint events\, refer to an item that could
be randomly drawn from that population of items.\nThe JCGM 106 approach ca
n be extended to assess the properties of a sample of N items rather than
a single item at a time. In this work\, the probability of truly conformin
g items within a finite sample is modelled. This probability is a quality
index of the sample as a whole. Resorting to appropriate discrete random v
ariables\, two probabilistic models are developed\, employing the above-me
ntioned specific and global risks as the distributional parameters of thos
e variables. The first model is based on a Poisson binomial distribution t
hat can infer the number of items within the sample having a good (conform
ing) true property value. The second model\, based on a multinomial distri
bution\, allows evaluating probabilities of incorrect decisions on CA of t
he items within the sample (false positives and negatives)\, as well as pr
obabilities of correct decisions (true positives and negatives).\n\nhttps:
//conferences.enbis.org/event/32/contributions/510/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/510/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Dynamic Bayesian Network-Based Run-to-Run Control Scheme for Optim
al Quality Engineering in Semiconductor Manufacturing
DTSTART:20230912T144000Z
DTEND:20230912T150000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-598@conferences.enbis.org
DESCRIPTION:Speakers: Wei-Ting Yang (BI Norwegian Business School)\, Jakey
Blue (National Taiwan University)\n\nRun-to-Run (R2R) control has been us
ed for decades to control wafer quality in semiconductor manufacturing\, e
specially in critical processes. By adjusting controllable variables from
one run to another\, quality can be kept at desired levels even as the pro
cess conditions gradually change\, such as equipment degradation. The conv
entional R2R control scheme calculates the adjustment value for the next r
un primarily based on output quality measurement\, which may provide delay
ed information and fail to reflect real-time process shifts. Nowadays\, ad
vanced process equipment is equipped with numerous sensors to collect data
and monitor process conditions. Sensor data are also extensively utilized
for various process-related tasks\, including quality prediction or fault
diagnosis. In this research\, we propose a novel R2R control scheme that
incorporates more timely control by considering uncertainties and relation
ships among sensor data\, controllable variables\, and target variables to
enable online R2R control. Dynamic Bayesian Networks (DBN)\, which serves
as the core of the R2R control scheme\, graphically links all variables f
rom different time periods. Network connections can be learned from histor
ical data and also imposed based on known causal relationships. By leverag
ing the information from the previous run and the desired target value\, t
he particle-based method is employed to compute the optimal control settin
gs for the upcoming run using the trained DBN. Finally\, the performance o
f the proposed approach is evaluated using real-world data.\n\nhttps://con
ferences.enbis.org/event/32/contributions/598/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/598/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Maximum Covariance Unfolding Regression: A Novel Covariate-Based M
anifold Learning Approach for Point Cloud Data
DTSTART:20230912T093500Z
DTEND:20230912T100500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-588@conferences.enbis.org
DESCRIPTION:Speakers: Kamran Paynabar (School of Industrial and Systems En
gineering)\, Qian Wang (Wells Fargo)\n\nPoint cloud data are widely used i
n manufacturing applications for process inspection\, modeling\, monitorin
g and optimization. The state-of-art tensor regression techniques have eff
ectively been used for analysis of structured point cloud data\, where the
measurements on a uniform grid can be formed into a tensor. However\, the
se techniques are not capable of handling unstructured point cloud data th
at are often in the form of manifolds. In this paper\, we propose a nonlin
ear dimension reduction approach named Maximum Covariance Unfolding Regres
sion that is able to learn the low-dimensional (LD) manifold of point clou
ds with the highest correlation with explanatory covariates. This LD manif
old is then used for regression modeling and process optimization based on
process variables. The performance of the proposed method is subsequently
evaluated and compared with benchmark methods through simulations and a c
ase study of steel bracket manufacturing.\n\nhttps://conferences.enbis.org
/event/32/contributions/588/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/588/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Interpretable Property Prediction on Full Scale Paperboard Machine
DTSTART:20230912T074000Z
DTEND:20230912T080000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-431@conferences.enbis.org
DESCRIPTION:Speakers: David Runosson (Linköping University)\n\nIn paper &
paperboard making\, sampling of product properties can only be made by th
e end of each jumbo reel\, which occurs 1-2 times per hour. Product proper
ties can vary significantly faster and do so in both machine and cross mac
hine directions. The low sampling may result in significant consequences s
uch as the rejecting an entire jumbo reel\, weighing about 25 tons\, by cl
assifying it as defective and resolving it into pulp if a specific propert
y test fails. \nPredictive models have the potential to inform operators a
bout the expected value of product properties\, but often black box-models
are required due to the complex relationships among input variables. \nWh
ile black box-models can provide robust predictions\, they are not interpr
etable for the operator\, and thus their value is limited. Therefor the fi
eld of XAI (Explainable Artificial Intelligence) has evolved\, in which al
gorithms help users to interpret black box models.\nIn this paper\, we inv
estigate the possibility of using a Random Forest to predict the results f
rom the Scott-Bond test for z-directional strength. Scott-Bond is used sin
ce it exhibits a complex and multifactorial nature\, characterized by sign
ificant short-term and long-term variations\, as well as significant measu
rement variance. Hence\, a predictive model would be beneficial.\nWe evalu
ate the model's potential as operator support by utilizing the XAI algorit
hm LIME combined with feature engineering to provide interpretability. Our
approach aims to provide valuable insights into how to achieve desired st
ates while maintaining robust predictions\, ultimately improving product q
uality\, and minimizing the waste of resources.\n\nhttps://conferences.enb
is.org/event/32/contributions/431/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/431/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Dimension Reduction Methods Based on FINE Algorithm for Clustering
Patients from Flow Cytometry Data
DTSTART:20230913T094500Z
DTEND:20230913T100500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-489@conferences.enbis.org
DESCRIPTION:Speakers: Walid Laziri (INRIA)\, Sophie Mézières (INRIA BIGS
- IECL)\, Frédéric Allemand (EMOSIS)\, Anne Gégout-Petit (Université
de Lorraine)\n\nFlow cytometry is used in medicine to diagnose complex dis
orders using a multiparametric measurement (up to 20 parameters). This mea
surement is performed in a few seconds on tens of thousands of cells from
a blood sample. However\, clustering and analysis of this data is still do
ne manually\, which can impede the quality of diagnostic discrimination be
tween "disease" and "non-disease" patients. A computational algorithmic ap
proach that automates and deepens the search for differences or similariti
es between cell subpopulations could increase the quality of diagnosis.\n\
nThe approach considered in this study is information geometry\, which inv
olves lowering the dimensionality of multiparametric observations by consi
dering the subspace of the parameters of the statistical model describing
the observation. The points are probability density functions\, and the su
bspace is equipped with a special geometrical structure called a manifold.
The objective of the reported study is to explore an algorithm called Fis
her Information Non-parametric Embedding (FINE)\, by applying it to flow c
ytometry data in the context of a specific severe disorder\, heparin-induc
ed thrombocytopenia (HIT).\n\nThis exploration consisted in testing differ
ent alternatives of the FINE algorithm steps such as the use of the Kullba
ck Leibler divergence under a Gaussian assumption or the Wasserstein dista
nce as measures of dissimilarity between the multiparametric probability d
istributions of the flow cytometry data for HIT+ vs HIT-.\n\nhttps://confe
rences.enbis.org/event/32/contributions/489/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/489/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Functional Data Analysis in Reliability and Maintenance Engineerin
g: An Application to Aircraft Engines
DTSTART:20230912T131500Z
DTEND:20230912T134500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-576@conferences.enbis.org
DESCRIPTION:Speakers: Cevahir YILDIRIM (UC3M)\, Alba Maria FRANCO PEREIRA
(UCM)\, Rosa Elvira LILLO RODRIGUEZ (UC3M)\n\nIn this work\, a practical r
eliability analysis and engine health prognostic study is performed using
a Functional Data Analysis (FDA) approach. Multi-sensor data collected fro
m aircraft engines are processed in order to solve one of the most importa
nt reliability analysis problems\, which is estimating the health conditio
n and the Remaining Useful Life (RUL) of an aircraft engine. Time-variant
sensor data is converted to smooth sensor curves in the form of functional
data\, and the Multivariate Functional Principal Component Analysis (MFPC
A) approach is applied to predict the RUL and to develop a Predictive Main
tenance (PdM) policy. The distribution of the principal component scores a
llowed us to understand sensor behavior and suggests a classification of d
ifferent types of engines based on qualitative variables.\n\nhttps://confe
rences.enbis.org/event/32/contributions/576/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/576/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Practice Makes Perfect – Perfect Exercises for Perfect Practice
…
DTSTART:20230913T094500Z
DTEND:20230913T100500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-609@conferences.enbis.org
DESCRIPTION:Speakers: Oliver Mülken (FHNW School of Life Sciences)\, Juli
a Rausenberger (FHNW School of Life Sciences)\, Franziska Kramer (FHNW Sch
ool of Life Sciences)\, Stefanie Feiler (FHNW School of Life Sciences)\n\n
The aim of introductory mathematics courses at university level is to prov
ide students with the necessary tools for their studies. In terms of compe
tence levels\, the contents are still basic: the students should *know* an
d *understand* the underlying concepts\, but mainly should be able to *app
ly* the relevant methods correctly in typical situations (even if they hav
e not fully understood the concepts…).\nIn terms of constructive alignme
nt\, both teaching and final assessment should therefore assure that they
reach this goal. This however requires training\, so\, the lecturers shoul
d supply a sufficient number of exercises they can try their skills on. \n
In the introductory statistics courses\, we use online assessments on the
learning management system Moodle since 2020\, and the group of Applied Ma
thematics has ongoing projects on how to create digitally adjusted exam qu
estions. In this line\, we have started to employ the Moodle-plug-in STACK
which uses the underlying computer algebra systems Maxima and provides a
very flexible setting. The result: (almost) infinite exercises using rando
mized input. Even if you are not working with Moodle\, the question types
can serve as an inspiration for your own set-ups\, e.g.\, using R interfac
es.\nWe will present our statistics questions\, but also give an outlook o
n other typical STACK applications such as interactive graphs or tests whe
re students may actively use hints\; and discuss the current status of sha
ring data bases.\nThe talk is complementing the presentations of Sonja Kuh
nt and Jacqueline Asscher\, all focussing on optimal (digital) learning an
d assessment.\n\nhttps://conferences.enbis.org/event/32/contributions/609/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/609/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Decoupling Method for Analyzing Fold-Over Designs
DTSTART:20230913T065000Z
DTEND:20230913T071000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-493@conferences.enbis.org
DESCRIPTION:Speakers: John Tyssedal (NTNU)\, Yngvild Hamre (NTNU)\n\nFold-
over designs often have attractive properties. Among these is that the eff
ects can be divided into two orthogonal subspaces. In this talk\, we intro
duce a new method for analyzing fold-over designs called “the decoupling
method” that exploits this trait. The idea is to create two new respons
es\, where each of them is only affected by effects in one of the orthogon
al subspaces. Thereby the analysis of odd and even effects can be performe
d in two independent steps\, and standard statistical procedures can be ap
plied. This is an advantage compared to existing two-steps methods\, where
failing to identify active effects in one step may influence the variance
estimate in the other step. An additional advantage of obtaining two inde
pendent variance estimates in separate steps is the opportunity to test fo
r missing higher-order effects. In our paper\, the method is successfully
tested on two different types of designs\, a fold-over of a 12 run Placket
t-Burman design and a 17 run definitive screening design with one center r
un added. Furthermore\, it is evaluated through a simulation study in whic
h scenarios with different selection criteria and heredity conditions are
considered. In this talk\, the focus will be explaining the proposed metho
d and demonstrating it through an example.\n\nhttps://conferences.enbis.or
g/event/32/contributions/493/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/493/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Practical Applications of Multivariate Analytics in the Process In
dustry
DTSTART:20230911T134000Z
DTEND:20230911T141000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-614@conferences.enbis.org
DESCRIPTION:Speakers: Bernt Hiddema (AspenTech)\n\nOver the last 30 years
processing industries such as refining\, chemicals and life sciences have
been using data driven models to achieve economic and environmental goals
through optimization. Some of these applications include advanced process
control\, real-time optimization and univariate statistical process monit
oring. Although these methods are successful for many applications\, ther
e are a subset of use cases where the complexity of the data and the natur
e of the process require more advanced modelling techniques. \n\nOther ind
ustries like banking\, commerce and medicine have seen major breakthroughs
in recent years thanks to the application of artificial intelligence and
machine learning. Some of the applications include predicting consumer be
haviour\, classifying health conditions or improving user experiences. Ca
n these approaches also be applied in the process industry and what other
techniques are available to drive profitability and sustainability? \n\nIn
this presentation Aspen Technology\, the leader in industrial AI for over
40 years\, will unpack how and where artificial intelligence and multivar
iate techniques are applied to batch processing in life sciences and speci
ality chemicals. Learn from practical examples how multivariate analysis
and optimization enable decision-making by leveraging the causal relations
hips in systems and how this approach may translate to other industries.\n
\nhttps://conferences.enbis.org/event/32/contributions/614/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/614/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Latent Variables Multivariate Statistical Methods for Data Analyti
cs in Industry 4.0
DTSTART:20230914T070000Z
DTEND:20230914T110000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-621@conferences.enbis.org
DESCRIPTION:Speakers: Alberto J. Ferrer-Riquelme (Universidad Politecnica
de Valencia)\, Joan Borràs-Ferrís (Universitat Politècnica de València
)\n\nhttps://conferences.enbis.org/event/43/\n\nhttps://conferences.enbis.
org/event/32/contributions/621/
LOCATION:2.4
URL:https://conferences.enbis.org/event/32/contributions/621/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Practical Reinforcement Learning in Logistics
DTSTART:20230912T063000Z
DTEND:20230912T065000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-615@conferences.enbis.org
DESCRIPTION:Speakers: Frans de Ruiter (CQM)\, Jan-Willem Bikker (CQM)\n\nR
einforcement learning is a variant on optimization\, formulated as a Marko
v Decision Problem\, and is seen as a branch of machine learning. CQM\, a
consultancy company\, has decades of experience in Operations Research in
logistics and supply chain projects. CQM performed a study in which reinfo
rcement learning was applied to a logistics case on tank containers. Becau
se of inbalanced flows\, these containers need to be relocated all over th
e world between harbors. The challenge is about sending empty containers f
rom i to j to deal with trading imbalances\, such that random demand for c
ontainers at each of the harbors can be met as much as possible. Instead o
f reducing the problem to a deterministic one and subsequently optimize\,
reinforcement learning deals with the randomness inherently and considers
cumulative rewards and costs\, using a simulation model. A non-standard ch
allenge is the extremely large dimension of the action space\, which is no
t commonly addressed in literature on reinforcement learning. Employing se
veral visualizations of aggregations of the scheme\, comparing to benchmar
k methods\, and applying statistical principles to robustness checks were
performed as well. This study was carried out as part of the European proj
ect ASIMOV (https://www.asimov-project.eu/ ).\n\nhttps://conferences.enbis
.org/event/32/contributions/615/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/615/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Sharing Ideas for Formulating Easy to Write Exam Questions with a
Focus on Statistical Practice
DTSTART:20230913T092500Z
DTEND:20230913T094500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-604@conferences.enbis.org
DESCRIPTION:Speakers: Jacqueline Asscher (Kinneret College)\n\nI see final
exams as a necessary evil and a poor assessment tool\, and their preparat
ion as a daunting\, time consuming task\, but to my students the final exa
m is of prime importance. They invest hours in solving exam questions from
previous years\, so I treat the exam questions as a very important teachi
ng tool\, despite a personal preference for projects\, case studies and ex
ercises using simulators. Ironically\, many instructors who are proponents
of active learning have observed that the level of student collaboration
reached in the preparation of solutions to old exams is seldom reached in
project work\, where tasks are typically divvied up like pie.\n\nIn order
for the diligent solution of the exam questions from previous years to hel
p our students learn to be good statisticians\, some exam questions must g
o beyond straightforward testing of knowledge of the topics.\n\nIn this ta
lk I will share ideas I have developed to meet this challenge\, addressing
issues including: how to write questions that can be recycled\; where to
find ideas for applied questions\; how to identify underlying principles a
nd translate them into exam questions\; how to come up with creative ways
to incorporate statistical software (here JMP) in an exam solved without c
omputers\; how to deal with language challenges. The examples are from cou
rses in introductory statistics\, industrial statistics and DOE.\n\nI hope
that my ideas will help you to formulate your own exam questions\, and an
ticipate hearing your ideas.\n\nhttps://conferences.enbis.org/event/32/con
tributions/604/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/604/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Distribution-Free Joint Monitoring of Location and Scale for Moder
n Univariate Processes
DTSTART:20230912T090500Z
DTEND:20230912T093500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-538@conferences.enbis.org
DESCRIPTION:Speakers: Marcus Perry (University of Alabama)\n\nAutocorrelat
ed sequences of individual observations arise in many modern-day statistic
al process monitoring (SPM) applications. Often times\, interest involves
jointly monitoring both process location and scale. To jointly monitor aut
ocorrelated individuals data\, it is common to first fit a time series mod
el to the in-control process and subsequently use this model to de-correla
te the observations so that a traditional individuals and moving-range (I-
MR) chart can be applied. If the time series model is correctly specified
such that the resulting residuals are normal and independently distributed
\, then applying the I-MR chart to the residual process should work well.
However\, if the residual process deviates from normality and/or\, due to
time series model misspecification\, contains levels of autocorrelation\,
the false alarm rate of such a strategy can dramatically rise. In this pap
er we propose a joint monitoring strategy that can be designed so that its
in-control average run length is robust to non-normality and time series
model misspecification. We compare its performance to that of the I-MR con
trol chart applied to the residuals under different misspecification scena
rios. Our conclusions suggest that the proposed joint monitoring strategy
is a useful tool for today’s modern SPM practitioner\, especially when m
odel misspecification is a concern.\n\nhttps://conferences.enbis.org/event
/32/contributions/538/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/538/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Drivers of Sustainable Tourism in Europe: How to Design Efficient
Business Strategies
DTSTART:20230911T090000Z
DTEND:20230911T093000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-432@conferences.enbis.org
DESCRIPTION:Speakers: Francesca Bassi (University of Padova)\, Juan Antoni
o Marmolejo Martìn (University of Granada)\n\nThis article studies the wi
llingness of the citizens of the 27 EU countries to change their travel an
d tourism habits to assume a more sustainable behavior. The study wants to
contribute to the recent literature on the topic of interconnections betw
een tourism and sustainability. The data comes from the Flash Eurobaromete
r survey 499\, involving more than 25\,000 European citizens. The survey t
ook place in October 2021 and wanted to analyze travel behavior and the im
pact of the Covid-19 pandemic on it\, booking channels and information sou
rces for travel preparations\, reasons for selecting destinations\, option
s and information on sustainable tourism. The hierarchical structure of th
e data - citizens within countries - is assumed applying a multilevel appr
oach of analysis that considers heterogeneity between and within countries
. The estimation of the multilevel latent class model allowed to identify
seven groups of European citizens similar by their willingness to adopt to
urism-related sustainability practices\, and the association of these late
nt groups with the 27 European countries. Using sociodemographic variables
\, it was also possible to profile these groups as well as to describe the
typical citizen belonging to each cluster. Moreover\, drivers of sustaina
ble tourism are identified\, both at county and citizen level. The results
of the analyses give many useful information for strategic management in
the tourism sector.\n\nhttps://conferences.enbis.org/event/32/contribution
s/432/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/432/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Comparative Probability Metrics: Using Posterior Probabilities to
Account for Practical Equivalence in A/B Tests
DTSTART:20230912T115500Z
DTEND:20230912T122500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-561@conferences.enbis.org
DESCRIPTION:Speakers: Luke Hagar (University of Waterloo)\, Nathaniel Stev
ens (University of Waterloo)\n\nRecently\, online-controlled experiments (
i.e.\, A/B tests) have become an extremely valuable tool used by internet
and technology companies for purposes of advertising\, product development
\, product improvement\, customer acquisition\, and customer retention to
name a few. The data-driven decisions that result from these experiments h
ave traditionally been informed by null hypothesis significance tests and
analyses based on p-values. However\, recently attention has been drawn to
the shortcomings of hypothesis testing\, and an emphasis has been placed
on the development of new methodologies that overcome these shortcomings.
We propose the use of posterior probabilities to facilitate comparisons th
at account for practical equivalence and that quantify the likelihood that
a result is practically meaningful\, as opposed to statistically signific
ant. We call these posterior probabilities comparative probability metrics
(CPMs). This Bayesian methodology provides a flexible and intuitive means
of making meaningful comparisons by directly calculating\, for example\,
the probability that two groups are practically equivalent\, or the probab
ility that one group is practically superior to another. In this talk\, we
will describe a unified framework for constructing and estimating such pr
obabilities\, and we will illustrate a sample size determination methodolo
gy that may be used to determine how much data are required to calculate t
rustworthy CPMs.\n\nhttps://conferences.enbis.org/event/32/contributions/5
61/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/561/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Measurement Uncertainty: Introducing New Training Material and a E
uropean Teachers’ Community
DTSTART:20230911T113000Z
DTEND:20230911T115000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-514@conferences.enbis.org
DESCRIPTION:Speakers: Peter Harris (National Physical Laboratory NPL)\, Ka
ty Klauenberg (Physikalisch-Technische Bundesanstalt (PTB))\, Nicolas Fisc
her (Laboratoire national de métrologie et d'essais LNE)\, Francesca Penn
ecchi (Istituto Nazionale di Ricerca Metrologica - INRIM)\n\nMeasurement u
ncertainty is a key quality parameter to express the reliability of measur
ements. It is the basis for measurements that are trustworthy and traceabl
e to the SI. In addition to scientific research\, guidance documents and e
xamples on how to evaluate the uncertainty for measurements\, training is
an important cornerstone to convey an understanding of uncertainty. \n\nIn
Europe courses on measurement uncertainty are developed and provided by m
etrology institutes\, and also by universities\, research institutions\, n
ational accreditation bodies\, authorities in legal metrology\, service co
mpanies and many more. In 2021 a broad consortium was formed to jointly 1)
develop new material for measurement uncertainty training and to 2) estab
lish an active community for those involved in measurement uncertainty tra
ining. This project-like collaboration is called MU Training. It is an act
ivity hosted by Mathmet\, the European Metrology Network for Mathematics a
nd Statistics\, and aims to improve the quality\, efficiency and dissemina
tion of measurement uncertainty training.\n\nThis contribution will give a
n overview on how the activity MU Training advanced the teaching of measur
ement uncertainty in the past two years. We will describe how an active co
mmunity was established that supports the teachers of measurement uncertai
nty. In addition\, we will describe the freely available training material
\, that was developed for trainees and teachers\, and that includes videos
as well as overviews about courses\, software and examples.\n\nFinally\,
possibilities for future collaboration will be sketched to further increas
e the understanding of measurement uncertainty and thus to contributed to
more reliable measurements in Europe.\n\nhttps://conferences.enbis.org/eve
nt/32/contributions/514/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/514/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Can You Dig It? Using Machine Learning to Efficiently Audit Utili
ty Locator Tickets Prior to Excavation to Protect Underground Utilities
DTSTART:20230912T100500Z
DTEND:20230912T103500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-558@conferences.enbis.org
DESCRIPTION:Speakers: B. Scott Crawford (VA811)\, David Edwards (Virginia
Tech)\, Kenneth Spade (VA811)\, Ryan Christianson (Virginia Polytechnic In
stitute & State University)\, Jennifer Van Mullekom (Virginia Polytechnic
Institute & State University)\n\nOrdinary citizens rarely think about prot
ecting underground utilities\, until a water main has burst or internet se
rvice is interrupted by an excavation project. The project might be as sm
all as a fence installation or as large as burying fiber optic cable along
large sections of major highways. Many states and countries have a centr
al service provider that distributes notices to utility companies regardin
g impending excavations. When contacted by the central service with a req
uest\, each utility company that services a parcel of land will mark the l
ocation of utility lines alerting excavators and thereby preventing servic
e interruptions and protecting workers and citizens alike from serious inj
ury\, or even death. That provider is VA811.com in Virginia\, United Stat
es. \nAt VA811.com\, an increasing number of excavation tickets are enter
ed via web users\, which have a higher number of errors\, as opposed to th
ose entered by call agents. Until recently\, VA811 has performed random a
udits of their tickets. In 2020\, VA811.com approached the Virginia Tech
Statistical Applications and Innovations Group (VT SAIG) to build a predic
tive model that would screen for problematic tickets. Since then\, VT SAI
G has developed two predictive models. This talk will detail the case stud
y in the context of the phases of Cross Industry Standard Data Mining Prac
tice (CRISP-DM). Statistical methods include measurement systems analysis
and gradient boosted machines. Features were engineered using text minin
g and geographical information systems data. Practical aspects of project
implementation will also be discussed including data cleaning\, model imp
lementation\, and model monitoring.\n\nhttps://conferences.enbis.org/event
/32/contributions/558/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/558/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Nonparametric Control Charts for Change-Points Detection: A Compar
ative Study
DTSTART:20230912T155000Z
DTEND:20230912T161000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-421@conferences.enbis.org
DESCRIPTION:Speakers: Michele Scagliarini (University of Bologna)\n\nDistr
ibution-free control charts have received increasing attention in non-manu
facturing fields because they can be used without any assumption on the di
stribution of the data to be monitored. This feature makes them particular
ly suitable for monitoring environmental phenomena often characterized by
highly skewed distribution. In this work we compare\, using two Monte Carl
o studies\, the performance of several non-parametric change point control
charts for monitoring data distributed according the Generalised Inverse
Gaussian (GIG) distribution. The aim is to identify the most suitable moni
toring algorithm considering jointly the ability in detecting shifts in lo
cation and/or scale and the percentage of missed alarms. The choice of the
GIG distribution is motivated by the fact that on the one hand it is ofte
n used to describe environmental radioactivity data\, but on the other han
d it has never been considered in connection with non-parametric control c
harts. For our purposes\, aware of being non-exhaustive\, we consider a no
n-parametric change-point control chart based on the Mann-Whitney statisti
c\; a distribution-free control chart based on Recursive Segmentation and
Permutation (RS/P)\; a monitoring algorithm using the Kolmogorov-Smirnov t
est\; and a chart which relies on the Cramer-von-Mises statistics. The res
ults reveal that the monitoring algorithm based on recursive segmentation
and permutation has the best performance for detecting moderate shifts in
the location\, whereas for the other scenarios examined the Kolmogorov-Smi
rnov control chart provides the best results both in terms of out-of-contr
ol ARL and missed alarms.\n\nhttps://conferences.enbis.org/event/32/contri
butions/421/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/421/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A Monte Carlo EM for the Poisson Log-Normal Model
DTSTART:20230911T124000Z
DTEND:20230911T131000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-573@conferences.enbis.org
DESCRIPTION:Speakers: Julien STOEHR (Université Paris-Dauphine\, PSL)\, S
téphane ROBIN (Sorbonne Université)\n\nThe Poisson log-normal (PLN) mode
l is a generic model for the joint distribution of count data\, accounting
for covariates. It is also an incomplete data model. A classical way to a
chieve maximum likelihood inference for model parameters $\\theta$ is to r
esort to the EM algorithm\, which aims at maximizing\, with respect to $\\
theta$\, the conditional expectation\, given the observed data $Y$\, of th
e so-called complete log-likelihood $\\mathbb{E}[\\log p_\\theta(Y\, Z) \\
mid Y]$.\n\nUnfortunately\, the evaluation of $\\mathbb{E}[\\log p_\\theta
(Y\, Z) \\mid Y]$ is intractable in the case of the PLN model because the
conditional distribution of the latent vector conditionally on the corresp
onding observed count vector has no closed form and none of its moments ca
n be evaluated in an efficient manner.\n \nVariational approaches have bee
n studied to tackle this problem but lack from statistical guarantees. Ind
eed the resulting estimate $\\widehat{\\theta}_{VEM}$ does not enjoy the g
eneral properties of MLEs. In particular\, the (asymptotic) variance of $\
\widehat{\\theta}_{VEM}$ is unknown\, so no test nor confidence interval c
an be derived easily from the variational inference. Starting from already
available variational approximations\, we define a first Monte Carlo EM a
lgorithm to obtain maximum likelihood estimators of this model. We then ex
tend this first algorithm to the case of a composite likelihood in order t
o be able to handle higher dimensional count data. Both methods are static
ally grounded and provide confidence region for model parameters.\n\nhttps
://conferences.enbis.org/event/32/contributions/573/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/573/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Pareto Solutions Resilience
DTSTART:20230913T071000Z
DTEND:20230913T073000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-472@conferences.enbis.org
DESCRIPTION:Speakers: João Lourenço (IPS-ESTSetúbal)\, Nuno Costa (ESTS
etubal)\n\nThe simultaneous optimization of multiple objectives (or respo
nses) has been a popular research line because processes and products are\
, in nature\, multidimensional. Thus\, it is not surprising that the varie
ty and quantity of responses modelling techniques\, optimization algorithm
s\, and optimization methods or criteria put forward in the RSM literature
for solving multiresponse problems are large. The quality of Pareto front
iers has been also evaluated by various authors\, and there are several ap
proaches and metrics to rank those solutions. However\, no metric to asses
s the resilience of Pareto solutions was proposed so far. Thus\, assuming
that the experiments were well planned and conducted\, and their results a
ppropriately analysed\, a novel metric is proposed to assess and rank the
Pareto solutions in terms of their resilience (sensitivity to changes or p
erturbations in the variables setting when implemented in the production p
rocess (equipments) or during its operation). This metric is easy-to-imple
ment and its application is not limited to problems developed in the RSM f
ramework. To consider the solutions resilience in the solution selection p
rocess can avoid wasting resources and time in implementing theoretical so
lutions in production process (equipments) that do not produce the expecte
d product output(s) or equipment behaviour. A classical case study selecte
d from the literature is used to illustrate the applicability (usefulness)
of the proposed metric.\n\nhttps://conferences.enbis.org/event/32/contrib
utions/472/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/472/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Fast and Furious: Some Anonymous Quotations from 43 Years Working
as an Applied Statistician
DTSTART:20230913T065000Z
DTEND:20230913T071000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-417@conferences.enbis.org
DESCRIPTION:Speakers: Martin Gibson (AQUIST Consulting)\n\nOver the last 4
3 years I have been privileged to work across the UK and overseas as an ac
ademic\, industrial statistician\, quality leader\, quality executive\, ma
nagement consultant and external examiner and advisor to various UK Univer
sities. \nIn that time\, I have focussed on systemic improvement of all e
nd-to-end processes in research and development\, new product development\
, manufacturing\, supply chain operations and all back-office support func
tions (HR\, finance\, sales & marketing) using statistical thinking. \nTh
roughout my career I have met a wide range of professionals\, the majority
of whom have minimal knowledge of statistics\, statistical thinking or Sy
stems Thinking. \nIn my presentation I will Illustrate through anonymous
quotations from those professionals the lack of statistical thinking that
exists in those organisations\, and the systemic issues that prevail. \nI
will summarise those quotations into three main areas and ask key questio
ns that arise to open a discussion on what we (ENBIS) must do to broaden a
nd widen the education of statistical thinking in all disciplines to ensur
e the better design of products\, systems\, strategies\, and decision maki
ng.\n\nhttps://conferences.enbis.org/event/32/contributions/417/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/417/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Forecasting Electric Vehicle Charging Stations' Occupation: Smarte
r Mobility Data Challenge
DTSTART:20230912T071000Z
DTEND:20230912T073000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-422@conferences.enbis.org
DESCRIPTION:Speakers: Yvenn Amara-Ouali (Université Paris Saclay)\n\nIn t
his talk\, we propose to discuss the **Smarter Mobility Data Challenge** o
rganised by the AI Manifesto\, a French business network promoting AI in i
ndustry\, and TAILOR\, a European project aiming to provide the scientific
foundations for trustworthy AI. The challenge required participants to te
st statistical and machine learning prediction models to predict the statu
ses of a set of electric vehicle (EV) charging stations in the city of Par
is\, at different geographical resolutions. The competition attracted 165
unique registrations\, with 28 teams submitting a solution and 8 teams suc
cessfully reaching the final stage. After providing an overview of the con
text of electric mobility and the importance of predicting the occupancy o
f a charging station for smart charging applications\, we describe the str
ucture of the competition and the winning solutions.\n\nhttps://conference
s.enbis.org/event/32/contributions/422/
LOCATION:2.12
URL:https://conferences.enbis.org/event/32/contributions/422/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Partitioning Metric Space Data
DTSTART:20230913T065000Z
DTEND:20230913T071000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-425@conferences.enbis.org
DESCRIPTION:Speakers: Yariv Marmor (ORT Braude College of Engineering)\, E
mil Bashkansky (Braude College of Engineering)\n\nThe partitioning of the
data into clusters\, carried out by the researcher in accordance with a ce
rtain criterion\, is a necessary step in the study of a particular phenome
non. Subsequent research should confirm or refute the appropriateness of s
uch a division\, and in a positive case\, evaluate the discriminating powe
r of the criterion (or\, in other words\, the influencing power of the fac
tor according to the level of which the data was divided). If the data com
es from a metric space\, this means that for any pair of data\, a distance
is defined that characterizes the dissimilarity between them. Speaking of
data\, we are not necessarily talking about numbers\, it can be informati
on of any kind about the objects under study (such as spectrograms\, 3B fo
rms\, etc.) obtained as a result of measurement\, observation\, query\, et
c.\, however distance between data\, expressing how far apart the objects
of interest are represented by a scalar. The correct choice of the distanc
e metric is a fundamental problem in quality control\, pattern recognition
\, machine learning\, cluster analysis\, etc. We propose two universal dis
criminating statistics - SP (segregation power) based on the ratio and the
difference of inter to intra clusters’ correlated estimates of the dist
ance between objects and discuss their specificity and sensitivity as well
as their universalism and robustness in relation to the type of objects u
nder study.\n\nhttps://conferences.enbis.org/event/32/contributions/425/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/425/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Modeling in the Observable or Latent Space? A Comparison of Dynami
c Latent Variable based Monitoring for Sensor Fault Detection
DTSTART:20230912T134500Z
DTEND:20230912T141500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-508@conferences.enbis.org
DESCRIPTION:Speakers: Natércia Fernandes (University of Coimbra)\, Marco
P. Seabra dos Reis (Department of Chemical Engineering\, University of Coi
mbra)\, Tiago Rato (University of Coimbra)\n\nThe latent variable framewor
k is the base for the most widespread methods for monitoring large-scale i
ndustrial processes. Their prevalence arises from the robustness and stabi
lity of their algorithms and a well-established and mature body of knowled
ge. A critical aspect of these methods lies in the modeling of the dynamic
s of the system\, which can be incorporated in two distinct ways: explicit
ly\, in terms of the observed variables\, or implicitly\, in the latent va
riable’s domain. However\, there is a lack of conceptual and evidence-ba
sed information to support an informed decision about which modeling appro
ach to adopt.\nTo assess the impact of these opposing modeling approaches
in monitoring performance\, we test and compare two state-of-the-art metho
ds representative of each class: Dynamic Principal Component Analysis with
Decorrelated Residuals (DPCA-DR\; explicit modeling) [1] and Dynamic-Inne
r Canonical Correlation Analysis (DiCCA\; implicit modeling) [2]. For comp
leteness\, the standard Principal Component Analysis (PCA) and Dynamic Pri
ncipal Component Analysis (DPCA) monitoring methods were also considered.\
nThese monitoring methods were compared on a realistic simulator of a Biod
iesel production unit [3] over several sensor faults. Our results highligh
t limitations of state-of-the-art methods\, such as reduced sensitivity du
e to fault adaptation and inability to handle integrating systems. The res
ults also point to an advantage of using DPCA-DR for detecting sensor faul
ts.\n\nReferences:\n1. Rato\, et al.\, Chemometrics and Intelligent Labora
tory Systems\, 2013. 125(15): p. 101-108.\n2. Dong\, et al.\, IFAC-PapersO
nLine\, 2018. 51(18): p. 476-481.\n3. Fernandes\, et al.\, Industrial & En
gineering Chemistry Research\, 2019. 58(38): p. 17871-17884.\n\nhttps://co
nferences.enbis.org/event/32/contributions/508/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/508/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Cloud-Powered Spatial Analytics: Leveraging Cloud Scalability for
Advanced Data Insights
DTSTART:20230911T131000Z
DTEND:20230911T134000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-526@conferences.enbis.org
DESCRIPTION:Speakers: Miguel Alvarez Garcia (CARTO)\n\nCloud computing has
transformed the way businesses handle their data and extract insights fro
m it. In the geospatial domain\, the main cloud platforms such as BigQuery
\, AWS\, Snowflake\, and Databricks have recently introduced significant d
evelopments that allow users to work with geospatial data. Additionally\,
CARTO is developing a Spatial Extension - a set of products and functional
ities built on top of these main Cloud providers that enable users to run
geospatial analytics and build compelling visualizations.\n\nIn this prese
ntation\, we will highlight the advantages of cloud-based geospatial analy
sis\, including scalability and agility. We will demonstrate the potential
of CARTO’s Analytics Toolbox through real-life scenarios\, emphasizing
its technical details and statistical techniques to provide attendees with
a more in-depth understanding of its functionality.\n\nWe will also explo
re the application of cloud-powered geospatial analytics across various do
mains\, such as retail\, consumer packaged goods\, urban planning\, transp
ortation\, and natural resource management. Attendees will be shown how cl
oud-powered geospatial analytics has been used to solve complex problems a
nd improve decision-making processes in these domains.\n\nThe session aims
to provide a comprehensive overview of the latest advances in cloud-power
ed geospatial analytics and their potential applications. Attendees will g
ain insights into the latest tools and techniques for processing and analy
zing geospatial data on cloud platforms\, as well as the benefits and chal
lenges associated with scaling geospatial analytics. This session is ideal
for individuals involved in geospatial data analysis\, cloud computing\,
or data science in general.\n\nhttps://conferences.enbis.org/event/32/cont
ributions/526/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/526/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Towards Traceable and Trustworthy Digital Twins for Quality Contro
l
DTSTART:20230911T115000Z
DTEND:20230911T121000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-540@conferences.enbis.org
DESCRIPTION:Speakers: Giacomo Maculotti (Politecnico di Torino)\n\nDTs are
simulation models that replicate physical systems in a virtual environmen
t\, dynamically updating the virtual model according to the observed state
of its real counterpart to achieve physical control of the latter. DTs co
nsist of a Physical to Virtual (P2V) and a Virtual to Physical (V2P) conne
ction. DTs require complex modelling\, often resorting to data-driven appr
oaches. DTs allow for defects and systems fault prediction\, enabling reli
able predictive maintenance and process adjustment and control to be imple
mented: DTs are essential for sustainability and digitalization.\nThe crea
tion of DTs often neglects quality control measurements\, resulting in the
ir lack of traceability and inability to associate them with a confidence
level in the prediction. The evaluation of the measurement uncertainty wil
l allow DTs’ application in the industrial context for quality control\,
defects and system faults prediction\, statistical predictive defect corr
ection and system maintenance within a traceable application framework. \n
Available methods for DT’s uncertainty evaluation neglect coupling with
the different parts of the DT\, especially the closed-loop feedback contro
l and the V2P connection. Bayesian approaches will allow for rigorous mana
gement of such coupling effect also by non-parametric approaches. A rigoro
us definition of DT’s metrological characteristics is unavailable\, and
both accuracy and precision shall be defined\, catering for the V2P closed
-loop feedback control.\nThis is being developed by the Trustworthy virtua
l experiments and digital twins (ViDiT) project\, funded by the European P
artnership on Metrology\, tackling four complex applications: robot and ma
chine tools\, nanoindentation\, primary electrical and cylindricity measur
ements.\n\nhttps://conferences.enbis.org/event/32/contributions/540/
LOCATION:2.9/2.10
URL:https://conferences.enbis.org/event/32/contributions/540/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Design of Experiments (DoE)-Based Approach to Better Capture Uncer
tainty in Future Climate Projections
DTSTART:20230912T093500Z
DTEND:20230912T100500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-507@conferences.enbis.org
DESCRIPTION:Speakers: Priscilla Mooney (Norwegian Research Centre (NORCE))
\, Carla Vivacqua (Universidade Federal do Rio Grande do Norte)\, Alok Sam
antaray (Norwegian Research Centre (NORCE))\n\nWe are living in the big da
ta era. The amount of data created is enormous and we are still planning t
o generate even more data. We should stop and ask ourselves: Are we extrac
ting all the information from the available data? Which data do we really
need? The next frontier of climate modelling is not in producing more data
\, but in producing more information. The objective of this talk is to sha
re how to mitigate future challenges associated with the exponential incre
ase in climate data expected over the next decade. Our approach uses effic
ient design processes and methods to ensure effectiveness in data producti
on and data analysis. \nNumerical climate model simulations have become th
e largest and fastest growing source of climate data. This is largely due
to societal demands for climate information that is both relevant and usef
ul. To satisfy this demand\, numerical models need to run large ensembles
to quantify uncertainties. Traditionally\, the simulations that constitut
e members of an ensemble are chosen in an ad hoc way leading to what is ca
lled an ‘ensemble of opportunity’. The current ‘ensemble of opportun
ity’ approach is inefficient and incomplete\, since only part of the par
ameter space is covered by the framework.\nThe main scientific question is
: Can the ‘ensemble of opportunity’ be replaced by something better? S
tatistics is a useful tool in this regard. We provide an overview of a Des
ign of Experiments (DoE)-based-approach\, grounded on statistical theory\,
which makes it possible to fully sample the uncertainty space\, while sav
ing computation cost.\n\nhttps://conferences.enbis.org/event/32/contributi
ons/507/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/507/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Global Importance Measures for Machine Learning Model Interpretabi
lity\, an Overview
DTSTART:20230912T155000Z
DTEND:20230912T161000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-484@conferences.enbis.org
DESCRIPTION:Speakers: Bertrand Iooss (EDF R&D)\, Vincent CHABRIDON (EDF R&
D)\, Vincent Pelamatti (EDF R&D)\n\nMachine learning (ML) algorithms\, fit
ted on learning datasets\, are often considered as black-box models\, link
ing features (called inputs) to variables of interest (called outputs). In
deed\, they provide predictions which turn out to be difficult to explain
or interpret. To circumvent this issue\, importance measures (also called
sensitivity indices) are computed to provide a better interpretability of
ML models\, via the quantification of the influence of each input on the o
utput predictions. These importance measures also provide diagnostics rega
rding the correct behavior of the ML model (by comparing them to importanc
e measures directly evaluated on the data) and about the underlying comple
xity of the ML model. This communication provides a practical synthesis on
post-hoc global importance measures that allow to interpret the model gen
eric global behavior for any kind of ML model. A particular attention is p
aid to the constraints that are inherent to the training data and the cons
idered ML model: linear vs. nonlinear phenomenon of interest\, input dimen
sion and strength of the statistical dependencies between inputs.\n\nhttps
://conferences.enbis.org/event/32/contributions/484/
LOCATION:Auditorium
URL:https://conferences.enbis.org/event/32/contributions/484/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Multivariate Bayesian Mixed Model for Method Comparability
DTSTART:20230912T150000Z
DTEND:20230912T152000Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-430@conferences.enbis.org
DESCRIPTION:Speakers: Michael Van den Eynde (Janssen Pharmaceutica N.V.\,
Janssen Pharmaceutical Companies of Johnson & Johnson\, Belgium)\, Laurent
Natalis (Pharmalex S.A.\, Belgium)\, Yimer Wasihun Kifle (Janssen Pharma
ceutica N.V.\, Janssen Pharmaceutical Companies of Johnson & Johnson\, Bel
gium)\, Martin Otava (Janssen-Cilag s.r.o.\, Janssen Pharmaceutical Compan
ies of Johnson & Johnson\, Czechia)\, Olympia Tumolva (Janssen Pharmaceuti
ca N.V.\, Janssen Pharmaceutical Companies of Johnson & Johnson\, Belgium)
\n\nIn pharmaceutical manufacturing\, the analytical method to measure the
responses of interest is often changed during the lifetime of a product d
ue to new laboratory included\, new equipment\, or different source of sta
rting material. To evaluate an impact of such change\, method comparabilit
y assessment is needed. Method comparability is traditionally evaluated by
comparing summary measures such as mean and standard deviation to a certa
in acceptance criterion\, or by performing two one sided tests (TOST) appr
oach. In this work\, method comparability is applied in the context of two
Malvern Mastersizer laser diffraction instruments MS2000 (old platform) a
nd MS3000 (new platform) that are used to measure particle size distributi
on. A design of experiment is implemented\, followed by the formulation of
a multivariate Bayesian mixed model that was used to encompass a complex
scenario. A Bayesian approach allows for a posterior distribution-based ev
aluation of method comparability. Aside from traditionally used summary cr
iteria\, posterior predictive distributions were also computed and compare
d for the two platforms. Moreover\, a risk-based assessment of method tran
sition was done through computation of probability of success of passing c
ertain specification limits for the two platforms\, and through assessment
of the impact of changing the method on the performance of the overall pr
ocess. The workflow has been successfully applied to multiple drug substan
ces and drug products.\n\nhttps://conferences.enbis.org/event/32/contribut
ions/430/
LOCATION:2.13
URL:https://conferences.enbis.org/event/32/contributions/430/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Analytical problem solving based on causal\, correlational and ded
uctive models
DTSTART:20230912T124500Z
DTEND:20230912T131500Z
DTSTAMP:20240815T094200Z
UID:indico-contribution-500@conferences.enbis.org
DESCRIPTION:Speakers: Jeroen de Mast (University of Waterloo + JADS)\, Rog
er Hoerl\, Willis Jensen\n\nMany approaches for solving problems in busine
ss and industry are based on analytics and statistical modelling. Analytic
al problem solving is driven by the modelling of relationships between dep
endent (Y) and independent (X) variables\, and we discuss three frameworks
for modelling such relationships: cause-and-effect modelling\, popular in
applied statistics and beyond\, correlational predictive modelling\, popu
lar in machine learning\, and deductive (first-principles) modelling\, pop
ular in business analytics and operations research. We aim to explain the
differences between these types of models\, and flesh out the implications
of these differences for study design\, for discovering potential X/Y rel
ationships\, and for the types of solution patterns that each type of mode
lling could support. We use our account to clarify the popular descriptive
-diagnostic-predictive-prescriptive analytics framework\, but extend it to
offer a more complete model of the process of analytical problem solving\
, reflecting the essential differences between causal\, correlational and
deductive models.\n\nhttps://conferences.enbis.org/event/32/contributions/
500/
LOCATION:2.7/2.8
URL:https://conferences.enbis.org/event/32/contributions/500/
END:VEVENT
END:VCALENDAR