GLNet: a lightweight convolutional neural network for bearing fault diagnosis

Abstract

In recent years, with the advancement of deep learning technology, significant success has been achieved in the field of bearing fault diagnosis. However, the success of these techniques has also brought about issues such as large model parameter counts, high computational complexity, and low diagnostic efficiency. To address these problems, this study proposes a lightweight convolutional neural network called GhostLowRankNet. Firstly, a fixed-weight mixed pooling layer is adopted to effectively enhance feature extraction capability. Secondly, an improved GhostConv module is introduced, incorporating low-rank decomposition technology and a dual batch normalization-ReLU structure. This module reduces the parameter count while preserving the core advantages of the ghost feature generation mechanism and effectively mitigates the issue of insufficient feature representation in deep convolutional layers. Finally, fault classification is accomplished using a low-rank fully connected layer based on matrix decomposition. Experiments are conducted on the XJTU-SY rolling bearing dataset and the SEU dataset. The results demonstrate that the proposed method achieves superior diagnostic accuracy compared to some existing lightweight and traditional models, while also exhibiting lower floating-point operations, lower parameter counts, and shorter inference time.