Deep energy-based learning of timed automata representations for cyber-physical systems
Publication date
2026-08-12
Document type
Dissertation
Author
Advisor
Referee
Beyerer, Jürgen
Granting institution
Helmut-Schmidt-Universität/Universität der Bundeswehr Hamburg
Exam date
2026-04-27
Organisational unit
Publisher
Universitätsbibliothek der HSU/UniBw H
Part of the university bibliography
✅
File(s)
Language
English
DDC Class
004 Informatik
Keyword
CPS
CPPS
Hybrid Dynamical Systems
Machine Learning
Model Learning
AI
Deep Learning
Production
Manufacturing
Anomaly Detection
System Reconfiguration
Process Optimization
Energy-Based Models
Timed Automata
dtec.bw
Abstract
The role of artificial intelligence (AI) in contemporary production systems, often referred to as Cyber-Physical Production Systems (CPPS), is gaining increasing importance. To improve robustness, safety, predictability, and efficiency, dynamical models of system components are learned from historical data using machine learning (ML) and applied in intelligent services such as process optimization, system diagnosis, and reconfiguration.
This raises three key questions: (1) What is a suitable formalism for representing CPPSs? (2) How can such representations be learned from data? (3) How can the learned models be applied in AI tasks?
The complexity of production data, stemming from the number and diversity of components and their interdependencies, presents a major challenge. Problems of similar complexity in other domains are typically addressed using deep learning (DL).
However, no established DL methodology exists for the broader class of CPPSs. This research aims to fill this gap with respect to domain-specific requirements: (1) general applicability and minimal reliance on expert knowledge, (2) symbolic interpretability, and (3) support for hybrid (discrete-event and continuous) dynamics.
Building on existing research and CPPS practice, this work combines subsymbolic deep energy-based models with symbolic timed automata to propose a methodology that satisfies these requirements.
The outcome is the Deep Energy-Based Timed Automata (DEBTA) framework, which addresses the following questions:
(1) How can a deep energy-based architecture be defined to support discretization and symbolization while extracting interpretable abstractions from complex data?
(2) How can such a model handle input data at different levels of abstraction?
(3) How can the learned models be applied to AI tasks, and how can their results be interpreted and formally verified?
The key idea lies in the hierarchical extraction of distributed binary system representations, which capture cross-variable and short-term dynamics, combined with a timed automaton model that captures high-level, long-term dynamics.
Finally, the thesis presents methods for anomaly detection, system reconfiguration, and process optimization based on system representations learned with DEBTA. The proposed techniques are evaluated using both synthetic and real-world case studies, with particular attention given to verifiability and explainability. The most important real-world applications demonstrating the advantages and limitations of the proposed methodology include (1) a conveyor system in a high-rack storage facility, (2) a smart meter assembly plant, and (3) the Environmental Control and Life Support System (ECLSS) of the Columbus module aboard the International Space Station (ISS).
This raises three key questions: (1) What is a suitable formalism for representing CPPSs? (2) How can such representations be learned from data? (3) How can the learned models be applied in AI tasks?
The complexity of production data, stemming from the number and diversity of components and their interdependencies, presents a major challenge. Problems of similar complexity in other domains are typically addressed using deep learning (DL).
However, no established DL methodology exists for the broader class of CPPSs. This research aims to fill this gap with respect to domain-specific requirements: (1) general applicability and minimal reliance on expert knowledge, (2) symbolic interpretability, and (3) support for hybrid (discrete-event and continuous) dynamics.
Building on existing research and CPPS practice, this work combines subsymbolic deep energy-based models with symbolic timed automata to propose a methodology that satisfies these requirements.
The outcome is the Deep Energy-Based Timed Automata (DEBTA) framework, which addresses the following questions:
(1) How can a deep energy-based architecture be defined to support discretization and symbolization while extracting interpretable abstractions from complex data?
(2) How can such a model handle input data at different levels of abstraction?
(3) How can the learned models be applied to AI tasks, and how can their results be interpreted and formally verified?
The key idea lies in the hierarchical extraction of distributed binary system representations, which capture cross-variable and short-term dynamics, combined with a timed automaton model that captures high-level, long-term dynamics.
Finally, the thesis presents methods for anomaly detection, system reconfiguration, and process optimization based on system representations learned with DEBTA. The proposed techniques are evaluated using both synthetic and real-world case studies, with particular attention given to verifiability and explainability. The most important real-world applications demonstrating the advantages and limitations of the proposed methodology include (1) a conveyor system in a high-rack storage facility, (2) a smart meter assembly plant, and (3) the Environmental Control and Life Support System (ECLSS) of the Columbus module aboard the International Space Station (ISS).
Version
Published version
Access right on openHSU
Open access