Uncertainty PERMEATED - explainable AI in a condition monitoring framework for industrial assets

Abstract

The first chapter introduces the topic of condition-based maintenance and contextualizes the importance of this technique, especially for critically important, complex and costly systems like machine tools. Condition-based maintenance can be seen as a special case of diagnosis and data analysis. Consequently, the second chapter introduces terms and definitions, which serve as foundation for the following discussion. The third chapter presents the state of research and a detailed review of publications in the context of data-driven diagnostics for condition-based maintenance. Different ideas behind and the purpose of model-based diagnostics, as well as signal-based diagnostics are outlined. Chapter 4 focuses on uncertainty, which is the main challenge in condition-based maintenance. Recommending a maintenance action has potentially costly real-world impacts. It is therefore necessary be aware of the risks of decisions. Uncertainty about the real state of the system seems to be inherent to the task of condition-monitoring. The lack of interpretability and auditability of decisions and the reasons for them are identified as main obstacles for a more widespread adoption of data-driven techniques. Subsequently chapter 5 introduces a diagnostics framework called PERMEATED, which embraces these results and is designed to deal with the existing uncertainties by incorporating them and emphasizing the importance of trust. The application of this framework to a real world application for machine tools is presented. Chapter 6 discusses some existing machine learning approaches for condition-monitoring applications and a applies them to a particular task regarding the dynamic behavior of a machine tool drive axis. Their performance is compared to an alternative, PERMEATED-compatible method, called SLIM. A different approach to satisfy the principles of the PERMEATED diagnostics process is given in the last section of the chapter. Instead of using inherently interpretable machine learning models, this chapter uses so-called explainers to retrieve explanations from opaque machine learning models. Chapter 7 summarizes the results of this thesis with regards to the task of condition monitoring for industrial assets and concludes with the identification of areas, where further research is necessary to make the application of techniques of machine learning more applicable for the task of condition-based maintenance.