Cohen’s class bilinear distributions and convolutional neural networks applied to broken rotor bar diagnosis

Time-frequency (t-f) signal processing techniques are particularly advantageous for induction motor (IM) fault diagnosis under dynamic and variable industrial operating conditions. Broken rotor bar (BRB) faults remain among the most challenging to detect because of their proximity to the fundamental frequency and significantly lower amplitude in comparison. Additionally, traditional approaches often result in false positives or negatives in scenarios involving load variation, power quality issues, and inverter-fed operations. To address these issues, this work proposes a comprehensive and objective methodology to evaluate eight Cohen-Class Bilinear Distributions (CCBD) to diagnose BRB. CCBDs offer high-resolution t-f representations, a crucial advantage for fault identification. However, their use is limited by cross-terms, nonlinear artifacts inherent to bilinear processing. To overcome this limitation, convolutional neural networks (CNNs) are applied to automatically classify t-f images and identify the CCBD methods that effectively minimize the cross-terms while preserving fault signature harmonics. This strategy also avoids subjective and time-consuming visual inspections. In addition, this work proposes a novel CNN architecture with an attention module (CNN-Attention), designed to enhance performance in this context. The evaluation considers challenging conditions, including 1) line-fed and 2) inverter-fed operation, 3) voltage unbalance, and 4) load oscillations, applied to a 2 HP, 60 Hz motor. Generalization capability is validated with data collected from a different laboratory, using an independent 1 HP, 50 Hz motor and five different inverter models. Experimental results show that combining CNN-Attention with CCBDs enables highly accurate and fast classification, achieving approximately 96% accuracy even when trained and tested on distinct laboratory datasets, demonstrating the effectiveness and adaptability of the proposed method.