Vitale, Francesco (2024) Anomaly Detection by Process Mining and Machine Learning for Industrial Cyber-Physical Systems. [Tesi di dottorato]

Documento PDF Vitale_Francesco_37.pdf Visibile a [TBR] Amministratori dell'archivio Download (11MB) | Richiedi una copia

Abstract

The ever-increasing spread and complexity of Industrial Cyber-Physical Systems (ICPSs) allow the seamless connection of the physical and cyber worlds through advanced physical plants, sensor networks and software, improving the quality of service of several industrial applications. Unfortunately, such complexity makes ICPSs vulnerable to faults and troubles the development of accurate models capturing their real behavior. Owing to the dependability issues of ICPSs, the scientific literature proposed many data-driven solutions based on Machine Learning (ML) for characterizing ICPSs and detecting disruptive anomalies. However, ML algorithms are typically process-agnostic and may be affected by explainability issues, especially those employing deep learning. Besides, the specific ML algorithm should be tailored to the data at hand, the types of anomalies to discover, and the learning paradigm to use. To provide process-based characterizations of ICPSs while maintaining a high level of explainability, this dissertation investigates PM for anomaly detection due to its ability to discover reference process models from ICPS data and check for deviations from such models. To develop PM-based anomaly detection in ICPSs while accounting for the shortcomings of existing approaches, this thesis puts forward a methodology that guides ICPS development toward implementing PM-based anomaly detection. The methodological steps provide insight into the ICPS application type, requirements, and data. These insights are used to drive the development of a specific PM-based anomaly detection technique through a flexible and explainable framework that integrates ML to enhance data pre-processing and improve the classification accuracy of anomalous behaviors. During the evaluation, several framework techniques were developed and applied to different industrial case studies, including railways, manufacturing, and healthcare. The results demonstrate that these techniques achieve high levels of explainability and detection effectiveness. These findings underscore the potential for integrating PM techniques in modern industries to enable interpretable, data-driven modeling of ICPS behavior and effectively detect deviations from nominal behavior.

Downloads

Actions (login required)

Modifica documento