Hagenberg Prescriptive Analytics Days 2024

4th April 2024, 11 – 17:15 Uhr

It has become a standard method in the production and service industries to use trained Machine Learning models for the prediction of anomalous events such as breakdowns or deviations from the regular process mode. Prescriptiveness has emerged as a generalization of this procedure: It means “Reporting of the patterns of the favorable instances” – and refers to deducing and providing not only the prediction of process anomalies – but also the capability of controlling the process parameters for maximizing advantageous effects – such as long production cycles, or a small number of cyber-attacks.

Day 2 – Open Problem Session:

On day two of the HPAD, the workshop’s contributors joined for a discussion-breakfast to report about open problems that have emerged as eminent and related during the previous day’s presentations. The goal is to share these emerged thoughts with the goal to exchange expert-knowledge, references to prior work and research motivation among the contributors and the audience.
Read more here.

Book of Abstracts

Tom Heskes, Roel Bouman:
Anomaly detection in industry: What do we need for cutting-edge applications?

Anomaly detection is widely applied throughout all branches of industry. Predictive maintenance, fraud detection, and automated measurement labeling all make eager use of anomaly detection. Yet, while applications are numerous, algorithms are often chosen based on outdated, biased or "gut-feeling" information. We've gone back-to-basics, providing researchers and industry experts with guidelines on what you need for good performance in practice. We supplement this quantitative research with illustrations on how you can typically bridge the gap from theory to practice. We then illustrate the power of anomaly detection by looking at a use-case of anomaly detection in Dutch power grid planning, where we use it to automatically label data. We present methodology for segmenting and labeling time-series using explainable machine learning. Lastly, we briefly discuss several other applications of anomaly detection.

Tom Heskes, Professor of Artificial Intelligence, Data Science, Institute for Computing and Information Sciences (iCIS), Faculty of Science, Radboud University Nijmegen, tom.heskes@ru.nl, https://www.cs.ru.nl/~tomh/
Roel Bouman, PhD student Data Science/Machine Learning at Radboud University, roelbouman@gmail.com, https://www.ru.nl/personen/bouman-r

Vagan Terziyan:
Adversarial Analytics — Adversarial Maintenance —Adversarial AI:
A New Bermuda Triangle?

This presentation invites you into the mysterious kingdom of adversarial technologies, including analytics, maintenance, and AI, where we explore the intricate interplay between security, reliability, robustness, resilience, and responsible autonomy. We briefly touch the challenges posed by adversaries in both the digital and physical worlds and discuss the hidden potential of proactive defense, risk awareness, and robustness in the era of AI. Going beyond traditional prescriptive maintenance principles, we navigate the uncharted waters of adversarial intelligence through the metaphorical "Bermuda Triangle" towards a safer and more resilient future.

Vagan Terziyan, Professor (Distributed Systems /AI), Faculty of Information Technology, University of Jyväskylä, FINLAND. Head of Collective Intelligence research group. Email: vagan.terziyan@iyu.fi; URL: https://ai.it.jyu.fi/vagan; Slides of the talk: https://ai.it.jyu.fi/Adversarial_Analytics.pptx

Michael Affenzeller, Jan Zenisek, Florian Bachinger:
Domain Knowledge infused Industrial Machine Learning

The evolution of modern production industries towards integrated sensor-driven monitoring is facilitating the emergence of concepts like Online Production Schedule Optimization or Predictive Maintenance. Through these approaches, the status of machinery and workpiece on production lines is continually assessed with the aim of anticipating the utilization of production capacity, potential machinery breakdowns, etc., and initiating proactive optimizing measures.
Due to the high complexity of industrial machinery and the environment in which they are integrated, analysis of modern production lines requires methods that facilitate the transformation of monitoring data into reliable, trustworthy, and interpretable prediction and prescription models. For this purpose, we aim to tap the knowledge of domain experts, who have a deep understanding of the production process, and use it for the modeling process.
In this talk, we present modern software solutions which utilize domain knowledge from industrial applications to improve machine learning based modeling. We showcase the ability of our implemented domain specific language to describe complex production processes and its ability to integrate algorithms of scientific machine learning on real-world use-cases. This digital representation of systems will enable us to prescribe optimizations in systems such as production processes.

Michael Affenzeller, FH-Prof. PD DI Dr., HEAL, AIST, Center of Excellence for Smart Production, Focus: Digital Transformations; Upper Austrian University of Applied Sciences; michael.affenzeller@fh-hagenberg.at, https://pure.fh-ooe.at/de/persons/michael-affenzeller

Jan Zenisek, BSc MSc, ASiC, Center of Excellence for Smart Production, Focus Digital Transformations, HEAL, Produktion und Operations Management, Research Center Hagenberg, jan.zenisek@fh-hagenberg.at, https://pure.fh-ooe.at/de/persons/jan-zenisek

Florian Bachinger, BSc MSc, Center of Excellence for Smart Production, Focus Digital Transformations, HEAL, Research Center Hagenberg, florian.bachinger@fh-hagenberg.at, https://pure.fh-ooe.at/de/persons/florian-bachinger

Gejza Dohnal:
Change alone is unchanging

When we observe a process that is identifiable – meaning we can develop a mathematical model to correspond with this process – it's highly likely that there will be a change in the behavior of this process over time, necessitating a change in our model as well. Identifying this point of change, understanding how the process has altered, and potentially explaining this change constitute what is known as the "change-point problem."

The study of the change-point problem has roots in statistical literature dating back to the 1930s. One of the earliest practical applications of this problem appears in Shewhart's control charts (1932), which aimed to detect points of change in the mean or variation of a monitored process to prevent potential damages resulting from such changes. E.S. Page's seminal work in 1957 on problems involving parameter changes at unknown points provided a probabilistic foundation for this problem.

In recent decades, developments have been influenced by two key factors: the exponentially increasing volume of data and the advancing capabilities of computational and artificial intelligence methods. Naturally, this evolution has impacted the approaches to identifying change points, leading to the development of new methods, some of which will be discussed in this contribution.

Gejza Dohnal, Gejza Dohnal, RNDr., Full CSc.Professor at Czech Technical University in Prague, Czechia

Jaromir Antoch:
Off-Line change point analysis

The question of whether long-term observed time series are stationary has been a central point of interest for both statisticians and people applying statistics for a long time. An answer to this question became more pertinent when in many observed series an increasing trend appeared. In climatology, for example, not only scientists, but especially politicians and journalists, started to connect this increase to emissions of greenhouse gases such as carbon dioxide, methane, and nitrous oxide. With the goal of deciding whether the series nowadays behaves similarly to those from when the measurements began, the statisticians developed a whole batch of procedures for studying their stationarity. In such a situation, statisticians concentrated on more complicated models for description of (not only climatological) series that enable one to estimate both abrupt-shift and gradual changes, to detect changes not only in the mean but also in the variance, in occurrence of extremal events or changes in seasonal behavior, etc. All methods that aim to detect and estimate changes of any type in stochastic models are referred to as "change point methods”. In the lecture, we will focus on deciding on the stationarity in the mean. Although emphasis will be on theory, it will be illustrated by the analysis of the Prague-Klementinum temperature series, the longest series of Czech temperature measurements, using discussed change-point methods. We will also touch, if time allows, on the question of efficient estimation of the number of changes and/or seasonality.

Charles University, Faculty of Mathematics and Physics, Dept. of Probability and Mathematical Statistics, Sokolovská 83, CZ, 18675 Prague 8 Karlin, Czech Republic, antoch@karlin.mff.cuni.cz, https://www.mff.cuni.cz/en/faculty/organizational-structure/people?hdl=303

Ivo Bukovsky, Florian Sobieczky, Witold Kinsner, Noriyasu Homma: Learning Entropy for Machine-Learning-Based System Monitoring and Its Explainability

The paper recalls the machine learning based approach to anomalies in data, i.e. learning entropy, as a complementary concept to the traditional probabilistic approach to anomalies, i.e. the Shannonian concept of information and entropy. Moreover, learning entropy allows us to introduce a novel concept to explain real-time learning neural weights, which can be viewed as a novel concept of learning (entropy) attention to deal with the transient behavior of neural weights in the presence of instantaneous system anomalies.

It is shown that this novel system monitoring framework is practically efficient for implementing real-time learning and can instantly detect the frequency component changes of a down-sampled vibration signal (below the Nyquist frequency) that captures unobserved disturbances of an underlying dynamical system; it also demonstrates its intriguing ability to detect frequency changes below the noise level with LE. In industrial applications, the proposed method can help minimize false alarms in fast conditional real-time monitoring based on incremental learning and provide a quick explanation for the detected anomalies.

Ivo Bukovsky, Associate Professor at University of South Bohemia in České Budějovice, Czechia, ibuk@prf.jcu.cz, https://www.jcu.cz/cz/univerzita/lide/clovek?identita=Bukovsky_Ivo_107111

Florian Sobieczky, Senior Researcher at Software Competence Center Hagenberg, 4232 Hagenberg i. Mkr., Austria, florian.sobieczky@scch.at, https://www.scch.at/team/florian.sobieczky