Deep Transfer Learning for Bearing Fault Diagnosis using CWT Time–Frequency Images and Convolutional Neural Networks

Deep transfer learning has evolved into a powerful method for defect identification, particularly in mechanical systems that lack sufficient training data. Nonetheless, domain divergence and absence of overlap between the source and target domains might result in negative transfer. This study examines the partial knowledge transfer, for bearing fault diagnosis, by freezing layers in varying proportions to take advantage of both freezing and fine-tuning strategies. To assess the proposed strategy, three distinct pre-trained models are used, namely ResNet-50, GoogLeNet, and SqueezeNet. Each network is trained using three different optimizers: root mean square propagation, adaptive moment estimation, and stochastic gradient descent with momentum. The suggested technique performance is evaluated in terms of fault classification accuracy, specificity, precision, and training time. The classification results obtained using the CWRU datasets show that the proposed technique reduces training time while enhancing diagnostic accuracy, hence improving bearing defect diagnosis performance.