Explainable machine learning for motor fault diagnosis

Industrial motors have been widely used in various fields such as power generation, mining, and manufacturing. Various motor faults and time-consuming motor maintenance processes will lead to serious economic losses in this context. Different sensing technologies, including acceleration, acoustic, and current sensing can be useful in motor condition monitoring, defect detection, and diagnosis. Regarding sensing data analytics, Machine Learning (ML) and Deep Learning (DL) techniques have been increasingly investigated, because of their promising capabilities in complex data characterization and pattern recognition. However, the explainability of ML and DL models and their decision-making remains a challenge, because of their black-box modeling by nature. Shapley Additive Explanations (SHAP), as a game theoretic approach, provides a way to explain ML and DL modeling results, by allocating credits (known as SHAP values) through local connections to quantify the contributions of input features to model outputs. In this paper, three commonly seen ML techniques, including Support Vector Machine (SVM), Random Forest (RF), and Neural Network (NN) are investigated for vibration-based motor fault diagnosis. Corresponding SHAP explanation methods are applied to the three ML techniques to discover the most important vibration features in detecting motor conditions and differentiating faults. Explanation results from the three ML techniques demonstrate great consensus: average vibration frequency contributes most to motor fault diagnosis. This explanation conclusion matches the physical understanding that fault occurrences would bring in additional frequency components to the spectrum. Improving the physical explainability of ML and DL techniques would significantly improve their credibility and generalizability.