Deep Generalized Canonical Correlation Analysis for Motor Fault Diagnosis

While motor current signature analysis (MCSA) is widely used for motor fault detection, it has shown limitations as a single modality sensor-based method in diagnosing some types of faults such as bearing roughness. This paper introduces a novel framework for motor fault diagnosis using multi-modal sensor data. The framework addresses the challenges of multi-modal sensor fusion in motor fault diagnosis by first aligning features from various sensors in a shared latent space using deep generalized canonical analysis (DGCCA) and then incorporating attention-based fusion to weigh the contribution of feature channels. Validated on datasets collected on motors with different bearing friction levels, the approach demonstrates significant performance gains over baseline methods, achieving high accuracy and robustness across varying operating conditions. The proposed method is positioned as a scalable and effective solution for industrial motor fault diagnosis applications.