关键词: 滚动轴承, 故障诊断, 卷积神经网络, 双向时间卷积网络, 变分模态分解

Abstract: To address the limitations of traditional convolutional neural networks (CNN) in feature extraction due to their restricted receptive field and inability to capture temporal features, a model combining CNN and bidirectional temporal convolution network(BiTCN) was proposed. The proposed model can expand the receptive field and effectively capture temporal dependencies. The original vibration signals were decomposed into K intrinsic mode functions ( IMFs) using variational mode decomposition (VMD). The decomposed signals were fed into CNN layers for feature extraction and signal compression. The processed signals were input into BiTCN to extract bidirectional temporal features. Dilated convolutions are applied to maximize the receptive field. The signals were passed through pooling layers and fully connected layers for fault diagnosis of rolling bearings. Experimental results demonstrate that the proposed model excels in feature extraction and temporal feature perception, and achieves superior fault diagnosis performance and generalization capability across multiple datasets.

Key words: rolling bearings, fault diagnosis, convolutional neural network( CNN), bidirectional temporal convolution network( BiTCN), variational mode decomposition( VMD)