Machine learning for fault classification within the context of the energy graph-based visualisation technique

Abstract

Modern processing plants are complex and span multiple physical domains, posing challenges for fault detection and isolation (FDI). The advent of Industry 4.0 has facilitated the extensive collection of data, thereby accelerating the adoption of data-driven methods for fault detection and isolation. Techniques like energy graph-based visualisation (EGBV) address the challenges of multi-domain systems by using exergy and energy to characterise the system and analyse historical process data. The EGBV technique is a recent FDI technique that aims to exploit the energy interactions between system components to describe the state of health of the system. In this study, the potential of machine learning (ML) to detect and isolate faults from the energy interactions within the EGBV context was investigated. Several objectives were identified to successfully apply and evaluate the performance of ML techniques in the context of the EGBV technique. A Simulink implementation of the Tennessee Eastman Process was used to generate process data to implement and evaluate the use of ML techniques for FDI purposes within the EGBV context. Exploratory data analysis was used on the different types of EGBV data to assess their effectiveness in representing process data where system faults were present. This analysis also aided in explaining why some ML techniques had difficulty detecting and classifying certain faults when using specific EGBV data types. Subsequent objectives involved the application of appropriate ML techniques to the different types of EGBV data, taking their abstractness and structures into account. Furthermore, the fault detection and classification performances of the ML techniques were compared with each other to identify if a certain data type allowed for better fault detection and classification. Finally, the performance of these ML techniques was compared with benchmarks, including the original non-ML EGBV technique. The results showed that the non-ML EGBV technique surpassed both ML techniques and other benchmarks in fault detection, while ML exceeded in fault classification over the other mentioned methods. This study concluded that the use of energy and exergy data with the EGBV technique produced more accurate results compared to the raw data used by other benchmark techniques.