Data Preprocessing for Utilizing Simulation Models for ML-based Diagnosis

With increasing complexity in cyber-physical systems (CPS), fault detection with root cause analysis during application is essential. Thus, several approaches and methods have been introduced in the past. In this paper, we contribute to fault localization in the context of CPS. We present a data preprocessing method that enables the real-time diagnosis of a system's behavior by classifying the present conditions. The suggested data preprocessing pipeline utilizes simulation models comprising fault models to compute information used for root cause analysis. The applied diagnosis methods, trained on the preprocessed data, enable a fault behavior analysis during operation by analyzing the system observations. This paper presents the complete processing pipeline, comprising the CPS analysis with relevant behavior information extraction and a classification algorithm. In addition, we demonstrate the results obtained from a use case considering a simplified DC e-motor model. Based on the use case, different machine learning (ML) algorithms, such as nearest neighbor, multi-layer perception, decision tree, and random forest, are evaluated on performance and diagnosis accuracy.