Explainable AI for Bearing Fault Detection Systems: Gaining Human Trust

In modern industrial processes, monitoring the health of rotating machinery is an important task. Many machine learning (ML) and deep learning (DL)-based models have shown good fault detection and diagnosis results. However, these models need to provide explanations and insights to users and experts in order to increase the adoption and spread of these technologies. Another common problem is the lack of labeled historical data, which makes it impossible to use supervised models. Therefore, we propose a new approach for fault detection and diagnosis (FDD) systems to overcome these challenges. In this approach, we used a vibration signal dataset and converted it into a statistical dataset by extracting the required features. An ML model, CatBoost, is used to detect and predict any faults in the machine and classify the fault type with an accuracy of 89.18%. To understand and explain the implemented model, we used an explainable AI method, SHapley Additive exPlanations (SHAP). Using SHAP, we obtained the dependency of the model on each of the features to predict the different classes of faults, thus making the model trustworthy for the users.