Interpretable Motor Sound Classification for Enhanced Fault Detection Leveraging Explainable AI

In industries, machines communicate through sounds, decoded by predictive maintenance to prevent issues. Understanding motor sounds is crucial for seamless industrial operations. This research undertakes a comprehensive explo-ration of machine learning models, specifically Artificial Neural Network (ANN), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Random Forest, applied to motor sound data for classifying instances as either healthy or faulty. The ANN, boasting an 81.22 % accuracy, reveals commendable precision and recall values for both classes, indicating its robust predictive capabilities. However, there is room for improvement, particu-larly in accurately classifying healthy motors. SVM marginally outperforms the ANN with an accuracy of 81.32%, showcasing balanced precision and recall for both classes. Notably, KNN, while exhibiting a slightly lower accuracy of 80.22 %, excels in recall for the healthy class, emphasizing its efficacy in correctly identifying healthy motor sounds. Random Forest attains an accuracy of 81.32 %, featuring notably high recall for the healthy class (0.91), underscoring its proficiency in capturing instances of healthy motor sounds. In-depth metrics provide nuanced insights into the strengths and specificities of each model, offering a foundation for informed decisions based on application priorities and requirements. The study contributes not only quantitative metrics but also interpretability tools, including LIME and SHAP, to enhance transparency and elucidate the intricate patterns within motor sound data.