Abstract: The ranges of convolutional neural network feature extraction were constrained by the size of the receptive field of the convolutional kernel, and it was difficult for a single-scale convolutional kernel to adequately capture the frequency components of different shocks. A dynamic wide kernels residual network(DWResNet) bearing fault diagnosis method was proposed to solve the above problems which made it difficult to improve the accuracy of bearing fault diagnosis. A wide residual kernel structure to the one-dimensional deep residual network framework was introduced. A parallel two-channel network structure for the feature of bearing faults was constructed. Then, the convolutional kernel was dynamically weighted by the network through an attention mechanism, which was adapted to extract feature information at different scales fully. As a result, effective identification of bearing faults was achieved. Experiments show that the proposed method achieves an accuracy of over 98% for bearing fault diagnosis tasks with different noise levels. Therefore, the dynamic scale weighting mechanism may effectively improve the bearing fault diagnosis, especially in the presence of high noise levels.

Key words: fault diagnosis, rolling bearing, residual network, noise interference

摘要： 卷积神经网络特征提取范围受其卷积核感受野大小的制约，单一尺度的卷积核难以充分捕捉不同冲击的频率成分。针对上述原因导致轴承故障诊断准确率难以进一步提高的问题，提出一种动态宽卷积残差网络轴承故障诊断方法。该方法在一维深度残差网络框架基础上引入宽残差核结构，构建轴承故障特征的双通道并行网络结构；然后通过网络注意力机制对卷积核进行动态加权，自适应地充分提取不同尺度特征信息，实现轴承故障的有效识别。验证实验结果表明，所提出的方法在不同噪声水平的轴承故障诊断任务中均能达到98%以上的准确率，网络动态尺度加权机制改进能有效提高轴承故障诊断效果，尤其在强背景噪声下仍能保持高精度故障诊断。

关键词: 故障诊断, 滚动轴承, 残差网络, 噪声干扰