Colloquium

At the research group Ambient Intelligence, we perform applied research: based on real-life use cases we collaborate with external partners to work towards a solution and gain insights from a technical and domain perspective. Surely, many people would agree that such projects do not traverse a paved road. In this talk, I give a few examples of projects we carried out to analyse data with various partners in my research line ‘applied data science’. Besides highlighting some techniques we used and adopted, I also emphasize the methodological lessons we learned thus far as well as the insights we would like to achieve in our current and future projects.

Semiconductor lithography system issues are challenging to diagnose from predefined models and historic data alone. That’s because such systems are characterized by high-dimensionality, non-stationarity and nonlinear behavior across multiple time and spatial scales. One line of research at ASML is to investigate model-independent approaches, which can help to find previously unknown causal relations from system diagnostic data. Therefore, we combine information theory based transfer entropy with eigencentrality in time series analysis. However, this approach must be computationally efficient for ASML system diagnostics (and prognostics) in real time. Moreover, it must be able to locate unobserved variables and capture unique as well as joint causal influences.

This presentation will discuss the merits and challenges of causal inference using information-theoretical concepts within ASML context and demonstrate a number of cases.

The Royal Netherlands Navy is planning to introduce several new ship classes within the next decade.  The new ships will be technologically much more advanced than the ships currently in service.  At the same time the operational availability needs to increase and the crew size needs to get smaller.  This combination leads to the necessity not only to automate many functions on board, but also to increase the level of support to the maintenance organization.  Sensor data can be used to better predict failures of machinery in order to allow the maintenance organization to perform timely repair actions and prevent catastrophic failures.  The developments include the reuse of available data used for Monitoring & Control purposes, the introduction of extra sensor technology to better detect failure modes that cannot be detected with current systems, and the introduction of AI and Machine Learning.  The Royal Netherlands Navy is cooperating with Industry to ensure the timely availability of these new methods and technologies.

Many aircraft are designed according to the deterministic damage tolerance philosophy to predict the crack growth life of the structure. Alternatively, a probabilistic damage tolerance analysis can be performed, called a structural risk analysis (SRA), taking into account all important scatter sources, such as, the initial flaw size, the inspection quality, the inspection scheme, the variability in loads and crack growth material properties, instead of using scatter and safety factors. For new military aircraft, SRA is mandatory. For current military aircraft, it already has become a valuable tool for fleet management, since it offers a risk (probability of failure) development over time, which cannot be obtained from the traditional deterministic damage tolerance analysis. By this, it better signals fleet management when to take corrective (maintenance) actions to prevent (critical) failures of the aircraft.

This presentation will address the general concept of probabilistic analyses and its more specific implementation in the field of probabilistic fracture mechanics (SRA) using available fleet data. The approach will be supported by a number of examples.

Risk Assessments performed on systems across many industrial sectors employ techniques such as Fault Tree Analysis and Event tree analysis which have their foundations back in the 1960/1970s.  Since that time technology has advanced and system designs, their operating practice and maintenance strategies are now significantly different to those of the 1970s.  Some of the restrictive assumptions such as: constant failure and repair rates for components, component failures being independent and the limited account of maintenance and renewal options in the component failure models employed, reduce the effectiveness of these methodologies to represent modern day systems.

In addition, research into the risk prediction techniques has made considerable advances in their capabilities since the 1970s but these advances tend to have addressed each deficiency in isolation.  Examples of significant advances are the Binary Decision Diagram (BDD) method of solving Fault Tree structures, improving both accuracy and efficiency, and the Petri net method which has been proven to be an effective means of predict system performance when complex maintenance and renewal strategies are employed.

This presentation will describe the motivation, progress and methodologies used on a project to develop the next generation of risk assessment methods which is funded by Lloyd’s Register Foundation.  The project aims to update the current risk assessment capabilities using a hybrid approach of methods including BDDs and Petri nets.   This research addresses the deficiencies in the current approaches and extends their capabilities to better represent systems employed across the industries.  The approaches developed will use the familiar causality structures of fault tree and event tree analysis, removing their traditional assumptions by changing the analysis methodologies employed.

Industrial partners from the nuclear, aerospace and railway industries are collaborating on the project to ensure it meets their requirements.

The “RUL loss rate”, or time derivative or the RUL (remaining useful life), measures the speed at which an asset’s condition degrades and it therefore is getting closer to failure in the absence of any preventive maintenance action.
The average RUL loss rate is the derivative of the “mean residual life” (MRL) ; therefore understanding the latter’s properties is of potential interest for maintenance policy optimization.
First we study a special class of time-to-failure distributions: those for which the MRL is a linear function of time, i.e. the average RUL loss rate is constant. This class contains very well known special cases: the exponential distribution, for which the MRL is constant (equal to the MTTF), and therefore the RUL loss rate is equal to 0; at, at the other extreme, what we call the Dirac distribution, for which the asset’s lifetime is deterministic, and the loss rate is equal to 1: for every hour that passes, the remaining useful life decreases by one hour.
In general, for that family of distributions, which we characterize explicitly, the average RUL loss rate takes a value between 0 and 1. For instance, the uniform distribution is characterized by an average RUL loss rate of one half.
It is shown that, for that special family, the average RUL loss rate can be obtained explicitly as a (decreasing) function of the coefficient of variation of the time to failure. A closed-form expression for the confidence interval for the RUL is obtained.
Then those results are generalised in two directions:
◼ 1) by introducing a nonlinear time transformation that allows for transposing to some classes of time-to-failure distributions ( such as Weibull or gamma) the results obtained in the special case.
◼ 2) by considering concurrent degradation modes; for instance, the combination of a mode with constant failure rate (exponential distribution) and a mode with deterministic life time ( Dirac distribution);
Implications for maintenance policy are discussed.

The amount of measured and collected condition monitoring data for complex industrial assets has been recently increasing significantly due to falling costs, improved technology, and increased reliability of sensors and data transmission. However, faults in safety critical systems are rare. The diversity of the fault types and operating conditions makes it often impossible to extract and learn the fault patterns of all the possible fault types affecting a system. Consequently, faulty conditions cannot be used to learn patterns from. Even collecting a representative dataset with all possible operating conditions can be a challenging task since the systems experience a high variability of operating conditions. Therefore, training samples captured over limited time periods may not be representative for the entire operating profile. The collection of a representative dataset may delay the implementation of data-​driven fault detection and isolation systems. Furthermore, domain experts require an interpretability of the obtained results.

The talk will give some insights into potential solutions that enable to 1) transfer models and operational experience between different units of a fleet and between different operating conditions also in unsupervised setups where data on faulty conditions is not available; and 2) fuse physical performance models and deep neural networks, thereby not only improving the performance but also improving the interpretability of the developed models.

SPC (Statistical Process Control)  is the part of industrial statistics that deals with monitoring data streams in order to timely detect changes (the word control in the name is a historical misnomer because it leads to confusion with feedback control; a better name should have been Statistical Process Monitoring). The field originated in the manufacturing industry.

In this talk, we give a brief overview of the standard procedures in SPC, applications fields and position SPC in the wider data science context. We will discuss an industrial case study with wind turbines, that illustrates the methodological challenges in the field.

The wind turbine case study arose in a maintenance context. We will discuss benefits of including SPC in predictive maintenance strategies and highlight some methodological challenges and how they could fit within the PrimaVera context.

With the growth of ICT and the real-time flow of data, it becomes increasingly important to be able to measure complex systems, quantify their level of health and resilience, and be able to detect changes with great speed and precision. This would inherently facilitate the transition from curative, reactive actions to preventive, prescriptive actions. In this talk, I will demonstrate through several examples how data analytic techniques can be used to transform data into knowledge and actions.

Tom Heskes, Professor of Artificial Intelligence, Institute for Computing and Information Sciences (iCIS), Radboud University Nijmegen

Discovering causal relations from data lies at the heart of most scientific research today. In apparent contradiction with the adagio “correlation does not imply causation”, recent theoretical insights indicate that such causal knowledge can also be derived from purely observational data, instead of only from controlled experimentation. In the “big data” era, such observational data is abundant and being able to actually derive causal relationships would open up a wealth of opportunities for improving science, healthcare, and possibly predictive maintenance. In this talk, I will sketch how insights from statistics and machine learning may lead to novel approaches for robust discovery of relevant causal relationships.

Dr. Jan Braaksma (Director WCM Summer School/University of Twente)

Jan Braaksma is an associate professor in the chair of Maintenance Engineering at the University of Twente and the director of the WCM Summer School part of World Class Maintenance. He has worked for the University of Groningen (RuG) and the Dutch Defense Academy (NLDA). He holds a Master’s degree in Business and ICT and a PhD degree in Economics and Business. Jan’s research focuses on Asset Management with a special attention for Asset Life Cycle Planning, Maintenance Engineering and Design for Maintenance.

A significant part of his research is in cooperation with companies and organisations such as Liander, Strukton Rail, Netherlands Railways (NS), Prorail, Heineken, AkzoNobel, Sitech, Huntsman, Heijmans, Sabic, TataSteel and the Ministry of Defense. Jan is responsible for the Master Class on Maintenance Engineering & Management provided by the University of Twente.

Jan is involved in WP5 in PRIMAVERA.
https://www.utwente.nl/en/et/dpm/me/staff/braaksma/
https://primavera-project.com/

Predictive maintenance can be seen as the most rewarding maintenance strategy. Therefore currently there is much focus is on the use of this strategy, however there can be debated that predictive maintenance is not always the most feasible strategy.

FMEA/RCM is a structured method which can be used to determine a maintenance concept based on the ways in which an asset can possibly fail and the impacts these so-called “failure modes” can possibly have. I will discuss the development and improvement of maintenance concepts over time and the need for the right asset information. I will have special attention for its usage for the identification of possible predictive maintenance candidates and criticality driven asset information management. The organization and management of (future) data collection is advocated as a crucial element for making better maintenance decisions.

During the colloquium I will go specifically into insides gained from a multiple case study on the use of FMEA in industry and insights I gained from working together with industry.

Prof. dr. Mariëlle Stoelinga is a professor of risk management, both at the Radboud University Nijmegen, and the University of Twente, in the Netherlands. Stoelinga is the project coordinator on the PrimaVera, a large collaborative project on Predictive Maintenance in the Dutch National Science Agenda NWA. She also received a prestigious ERC consolidator grant Stoelinga is the scientific program leader of the Risk Management Master, a part-time MSc program for professionals.She holds an MSc and a PhD degree from Radboud University Nijmegen, and has spent several years as a post-doc at the University of California at Santa Cruz, USA.

Predictive maintenance is a promising technique that aims at predicting failures more accurately, so that just-in-time maintenance can be performed, doing maintenance exactly when and where needed. Thus, predictive maintenance promises higher availability, fewer failures at lower costs. In this talk, I will advocate a combination of model-driven (esp fault trees) and data analytical techniques to get more insight in the costs versus the system performance (in terms of availability, reliability, remaining useful lifetime) of maintenance strategies. I will show the results of three case studies from railroad engineering namely rail track (with Arcadis), the HVAC (heating, ventilation, airco; with NS).
I will also go into recent developments on learning fault trees and rare event simulation.

Tiedo Tinga is a professor in Life Cycle Management at the Netherlands Defence Academy (NLDA) and in Dynamics based Maintenance at the faculty of Engineering Technology of the University of Twente. He has a background in Materials Science and Mechanical Engineering. Before joining academia, he has been working as a scientist at the National Aerospace Laboratory NLR for almost 10 years. His research focuses on the detection and prediction of failures in systems, using combinations of the physics of failure, thorough understanding of the (dynamic) system behavior, advanced monitoring techniques and data analysis. All his research projects are in close collaboration with industrial partners.

In this webinar a general introduction to (predictive) maintenance will be given. First the basic motivation for maintenance will be presented and an overview of various maintenance policies will be given. Then the more advanced policies, based on condition monitoring and prognostics will be discussed in more detail. The various options (model-based, data-driven) will be shown and the current status and challenges will be sketched. Finally, some case studies from our research projects will be used to demonstrate the potential and limitations

Geert-Jan van Houtum is Professor of Maintenance, Reliability, and Quality at the Department Industrial Engineering and Innovation Sciences (IE&IS) of Eindhoven University of Technology since 2008. Prior to that he filled positions as assistant and associate professor at the same department (1999-2007) and the University of Twente (1994-1998) and as visiting professor at Carnegie Mellon University (2001). He obtained his M.Sc and Ph.D. degree in Applied Mathematics from Eindhoven University of Technology in 1990 and 1995, respectively.

His research is focused on the maintenance and reliability of capital goods, and in particular on: (i) Design and control of service supply chains; (ii) Maintenance concepts, in particular predictive maintenance; (iii) Design for availability. He has over 80 publications in international refereed journals such as Operations Research, Manufacturing and Service Operations Management, IIE Transactions, and European Journal of Operational Research. He is area editor at Service Science and associate editor at Manufacturing and Service Operations. Much of his research is in cooperation with the industry. He works with companies such as ASML, Canon, Dutch Railways, Philips, Marel, the Royal Dutch Airforce, the Royal Dutch Navy, Thales, and Vanderlande. He is vice-dean IE of the Department IE&IS since September 2017. Further, he is a board member of the Service Logistics Forum.
 
For a list of publications, see: https://research.tue.nl/en/persons/geert-jan-jan-van-houtum/publications/

About 10 years ago, we started with predictive maintenance research in the Netherlands. In my projects, we studied systems in the high-tech, maritime, and chemical industry.  In this presentation, I present a general predictive maintenance approach that works for systems where a limited set of components causes most of the failures. This approach builds on stochastic processes, data mining, Bayesian learning, machine learning, and Operations Research techniques. We will also discuss what we can investigate in the coming 10 years. An important learning point of the past 10 years is that data analysis methods often lead to predictions with a certain percentage of false positives. That is often not good enough for users of systems to replace a component or module preventively. But these predictions can still be useful to be better prepared when a failure occurs.