Transfer Learning Approach using Simulated Induction Motor Bearing Data: A Comparative Analysis of SE-ResNet and its Hybrid Variants

Abstract

Early detection of incipient bearing faults in induction motors has proven crucial in predictive maintenance, helping avoid machine downtime and costly repairs. The main challenge is collecting sufficient data for deep learning models since faults are a rare occurrence. This paper investigates the efficacy of a transfer learning approach for induction motor bearing fault diagnosis using simulated vibration data. Healthy and faulty bearings of different severities were simulated in MATLAB for various noise magnitudes. A Squeeze and Excitation Residual Network (SE-ResNet), previously trained on a large dataset for bearing faults of a Permanent Magnet Synchronous Motor (PMSM), is used as a feature extractor. By leveraging pre-trained knowledge, the model's weights were fine-tuned using Bayesian Optimization, aiming to mitigate the data scarcity issue while maintaining accurate fault classification. The model's performance was compared against three hybrid architectures incorporating Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM) layers. The study aims to assess the impact of adding recurrent layers to capture temporal dependencies within simulated vibration signals. Contrary to expectations, the hybrid models did not improve the classification accuracy compared to the standalone pre-trained SE-ResNet. The test accuracy remained the same for all the models at 97.297% whereas the computational cost increased for the hybrid models. This paper analyzes these findings, highlighting the challenges of transfer learning with simulated data.