A Lightweight Rolling Bearing Fault Diagnosis Method Based on Multiscale Depth-Wise Separable Convolutions and Network Pruning

Fault diagnosis in rolling bearings is critical important in preventing machinery damage. Current deep learning-based approaches for rolling bearing fault diagnosis mainly rely on complex models that require significant hardware storage and computing power. In this paper, we introduce a multiscale Depth-wise Separable Convolutions and network pruning (MS-DWSC-PN) approach for lightweight rolling bearing fault diagnosis. Initially, the original one-dimensional vibration signals are transformed into two-dimensional time-frequency images using continuous wavelet transform (CWT), rendering them suitable for MS-DWSC-PN. Secondly, the datasets and an adaptive learning rate reduction algorithm are utilized to train the model, and the lightweight network model is achieved through network pruning. Subsequently, the results obtained from extensive experiments on identical datasets using AlexNet, VGG16, and LeNet confirm that the proposed model exhibits lower FLOPs and reduced iteration time. Finally, the generalizability of the model under variable operating conditions is discussed. Compared with other intelligent diagnosis methods, the presented strategy can achieve smaller model size and higher accuracy under both constant and variable working conditions.