A Model Identification Forensics Approach for Signal-Based Condition Monitoring

Abstract

Condition monitoring (CM) of machines and robots is vital to improve operational reliability and to avoid occupational incidents. Recently, deep learning (DL) has become popular in CM literature for its outstanding ability of learning fault patterns. However, due to the black box and non-intuitive nature of its layers, the logic behind its decisions is hard to understand. This shortcoming hinders the DL implementation in many critical applications where the user needs to ensure the reliability of the classifier. Hence, in this paper, a new framework for DL-based CM systems is proposed, which consists of four steps (1) Feature extraction (2) Fault diagnosis (3) eXplainable Artificial Intelligence (XAI)-based model optimization (4) Interpretation system. The experimental evaluations on two real-world datasets demonstrate that the proposed XAI interpreter was able to visualize the contributing patterns to fault types. The feature engineering block not only makes it easier for the operator to only observe the contributing features, but also it helps the model optimizer to speed up the runtime. The results show that the proposed model achieved a slightly better accuracy than the other state-of-the-art models.