A Lightweight Network With Adaptive Input and Adaptive Channel Pruning Strategy for Bearing Fault Diagnosis

Bearing fault diagnosis networks based on convolutional neural networks (CNNs) have received increasing attention from scholars. However, the above works broadly suffer from the following three shortcomings: 1) most of the networks are complex in structure and not lightweight enough, introducing a large number of computations and parameters; 2) batch normalization (BN) is used in most lightweight networks, which causes the networks to perform poorly after training with a small batch size (BS); and 3) the input length (IL) of the above CNN-based network is fixed, which reduces the generalization performance of the network in different application scenarios. Therefore, a lightweight network with adaptive input and adaptive channel pruning strategy is proposed in this study to solve these shortcomings. First, a lightweight network is designed to enable the network to be trained with small BS. Second, the proposed adaptive input module (AIM) enables the network to automatically adjust the IL for guaranteeing the generalization ability without tuning network parameters. Finally, an adaptive channel pruning strategy is proposed to further reduce the network parameters and computational workload. The superior performance of the proposed network is validated and confirmed on three bearing datasets accompanied by comparative analysis with typical networks.