Multi-sensory Fault Diagnosis of Broken Rotor Bars Using Transfer Learning

Abstract

Induction motors play a crucial role in industrial operations, constituting a significant portion of the workforce. The occurrence of faults in these motors can substantially impact their operational performance. Approximately 20% of these faults stem from broken rotor bars, underscoring the importance of early detection to prevent more severe issues. This study aims to leverage data from multiple sensors to detect broken rotor bar faults. Rather than evaluating sensor data independently, as commonly done in existing literature, this study adopts a novel approach. It employs wavelet transform to convert signals from two different sensors into time-frequency images, which are then concatenated to form a single dataset. Subsequently, a convolutional neural network based on MobileNetv2 architecture is developed to classify rotor faults using these images. Compared to single-sensor approaches, the proposed methodology yields a 2% enhancement in performance. Moreover, it demonstrates the capability to detect faults involving up to 4 broken rotor bars with an impressive accuracy rate of around 97.2%.