Interpretable Machine Learning for Predictive Maintenance of Electric Motors: A Comparative Study of Subspace kNN and Symbolic Classification

Predictive Maintenance (PdM) is an important technique used to ensure the reliability and efficiency of industrial equipment. In recent years, Machine Learning (ML) techniques have become increasingly vital to PdM, offering powerful tools for analyzing complex data patterns and making accurate predictions about equipment health. Machine learning also uncovers relationships between input variables and system responses that are difficult to capture with traditional physical or mathematical models. This study compares the performance of two machine learning models - Subspace k-Nearest Neighbors (kNN) and symbolic classification -for predicting faults in rotating machines based on vibration data. The results demonstrate that Subspace k-Nearest Neighbors (kNN) handles different environments (e. g., orientation, position of the motor) with 98% accuracy. Symbolic classification, particularly in the two-stage hierarchical approach, provides interpretable models and easy deployment while maintaining a similarly high level of accuracy. The study outlines trade-ofs between model complexity, accuracy, and interpretability, and provides guidance for selecting appropriate PdM techniques. Future work includes expanding the dataset to improve generalization and incorporating high-frequency sensors for broader applicability. This research contributes to bridging the gap between theoretical ML models and practical industrial applications.