A lightweight convolutional neural network method for bearing fault diagnosis based on dynamic weighting and variable-scale feature extraction

Abstract

Fault diagnosis of rolling element bearings is crucial for ensuring the safety, reliability, and economic efficiency of modern industrial systems. However, conventional deep learning models often suffer from high computational costs and fixed receptive fields, which limit their deployment on resource-constrained edge devices. To address these issues, an adaptive variable-scale lightweight convolutional neural network (AVS-LCNN) is proposed. First, Gramian angular difference field coding is employed to transform one-dimensional vibration signals into time-frequency dual-channel images. Subsequently, depth-separable convolution is utilized to reconstruct the backbone network of AVS-LCNN, significantly reducing the number of model’s parameters. To better capture the multi-scale characteristics of fault signals, a variable-scale feature extraction mechanism is developed based on the dilated convolution and the atrous spatial pyramid pooling module. Additionally, a sample-aware dynamic weighting strategy is introduced, in which a soft gating mechanism generates a weights α, enabling the model to automatically optimize its structure according to the complexity of the input samples. Experiments conducted on the high-speed train bogies fault and Paderborn University datasets show that AVS-LCNN achieves a diagnostic accuracy rate of over 99% with only 0.17 M parameters, demonstrating a favorable balance among computational accuracy, robustness and inference efficiency. These results indicate that the proposed method provide an effective solution for industrial edge applications.