Improving Detection of Induction Motor Bearing Faults: A Study on Hyperparameter Tuning and ResNet-18 Layer Modification

Detecting incipient bearing faults of an induction motor is crucial for minimizing downtime, reducing maintenance costs, and ensuring safety. In this study, a novel approach to improve the detection accuracy of induction motor bearing faults using a combination of hyperparameter tuning and modifications to the Residual Network - 18 (ResNet-18) architecture is investigated. The effect of optimizing these aspects to enhance ResNet-18’s ability to classify various fault types within a dataset of bearing vibration signals is explored. This research focuses on ResNet-18 architecture, which has demonstrated remarkable performance in various image classification tasks, to evaluate the impact of specific modifications to its layers in order to improve further its suitability for bearing fault detection. The appropriate balance between model complexity and interpretability is achieved by altering the depth, width, and skip connections within ResNet-18. Next, parameters such as batch sizes, learning rates, L2 regularization, number of epochs, and optimizer are investigated by systematically tuning these hyperparameters and applying layer modifications, a new Bayesian Optimized Squeeze and Excitation ResNet model, which has a higher training accuracy of 98.44%, a validation accuracy of 99.48%, testing accuracy of 99.48%, and lower computational cost, as compared to the ResNet-18 model, is achieved. The proposed BOSE-ResNet contributes to the development of a more effective and precise bearing fault diagnosis model while enhancing machinery reliability in industrial applications and providing valuable insights for practitioners and researchers in the field of condition-based maintenance.