Experimental Analysis of Multi-Sensor Data for Motor Condition Monitoring

Abstract

Electric motor bearing failures constitute one of the most frequent causes of unplanned industrial downtime, yet how can industrial operators maintain absolute confidence in AI-driven diagnostic systems when electric motors constantly shift speeds and evolve over years of service? While Deep learning has demonstrated significant potential in laboratory settings, most mod-els remain “black boxes” validated on narrow, single-session datasets that fail to reflect the unpre-dictable nature of real-world industrial environments.

This thesis bridges the gap between academic theory and industrial reliability by presenting a ro-bust machine learning framework for bearing fault diagnosis, validated against a four-year longitudinal dataset (2022-2025), spanning from 0-1500 RPM. Moving beyond traditional scalar thresholds, this study evaluates four supervised classification models and three anomaly detectors across 57 unique motor speeds using a rigorous Leave-One-Speed-Out (LOSO) protocol to ensure that every evaluation speed was unseen during training phase of AI models.

The experimental results show that traditional fault indicators like kurtosis and RMS fail to generalize across speed folds as a multi-scale Residual Raw CNN achieves a 99.81% binary and 98.47% three-class accuracy. Furthermore, this work utilizes DeepSHAP attribution to “look inside” these models providing the physical evidence that X-radial axis remains the primary diagnostic indicator except at high-speed exceeding 1200 RPM. Among the unsupervised detectors, CNN au-toencoder delivered the most operationally balanced performance with a 92.89% detection rate and only a 5.25% false alarm rate. Together these find-ings demonstrate that trustworthy, interpretable, longitudinally stable and speed robust bearing health monitoring is achievable within the ABB Detect-Predict-Recommend operational frame-work.