Optimal Feature Selection for Distributed Bearing Fault Detection Using Cost-Effective, High-Performance Machine Learning Algorithms

This study presents a highly accurate and cost-efficient machine learning framework for diagnosing distributed bearing faults by identifying an optimal and compact set of frequency-domain features. The analysis is based on a comprehensive dataset collected from an experimental testbed comprising a coupled induction motor operating under 50 distinct combinations of speed and load. The dataset encompasses both healthy bearings and those exhibiting various distributed faults-such as lubrication degradation, contamination, electrical erosion, and flaking-as well as a localized fault in the form of a single-point defect on the outer race. Data were acquired using a multi-sensor module integrating three-axis vibration, stray magnetic flux, and two phase current signals. The primary objective is to reduce an initial set of 102 frequency-based features to a minimal subset suitable for deployment on resource-constrained microcontrollers without sacrificing classification performance. Employing Mutual Information and Random Forest techniques for feature selection, the study successfully identifies a critical subset of five features. A Random Forest classifier trained on these features achieved an impressive test accuracy of 99.58%. Furthermore, the combination of top-ranked features from both MI and RF methods led to improved classification performance over using either method independently, thereby enhancing the robustness and reliability of industrial bearing fault detection.