Deep Transfer Learning Approach Using Filtered Time-Frequency Representations of Current Signals for Bearing Fault Detection in Induction Machines

This paper addresses the challenge of limited labelled data in induction machine fault diagnosis by applying deep transfer learning with convolutional neural networks to classify ball bearing health conditions. Specifically, the objective is to classify ring and cage failures in ball bearings using a limited dataset acquired from an experimental test bench. Unlike traditional approaches that rely on vibration sensors, this study uses noninvasive current signals. Moreover, this study introduces a novel preprocessing approach that filters out the fundamental frequency of the current signal to enhance fault-related harmonics in time–frequency representations generated via continuous wavelet transform and short-time Fourier transform. Five pre-trained convolutional neural networks—ResNet18, ResNet50, VGG16, AlexNet and GoogLeNet—are fine-tuned on these representations, demonstrating up to a 47% improvement in classification accuracy. Furthermore, the approach maintains high accuracy even with only 10% of the original dataset, showcasing its sample efficiency. This work contributes to a scalable and data-efficient solution for reliable condition monitoring in industrial settings, further advancing the use of current signals for fault diagnosis.