基于改进ConvNeXt的轴承故障诊断研究
Research on Bearing Fault Diagnosis Based on Improved ConvNeXt

根据现有的网络模型(ConvNeXt)在处理小样本任务过程中，无法完全提供故障信息、所需样本数据量过大、泛化性能不足和鲁棒性能较低的特点，提出一种改进型ConvNeXt网络模型的轴承故障诊断方法。首先使用格拉姆角差场图像编码技术将故障样本进行解码得到相应的故障特征图；然后通过随机裁剪、旋转等方法，对数据增强模块进行改进；其次利用非对称卷积思想对ConvNeXt模型的大卷积核进行重构，增强模型对小样本任务的处理能力；最后融入CBAM注意力机制，提高模型对信号特征的通道和空间方面的提取能力。实验表明，改进型ConvNeXt网络模型对滚动轴承不同故障直径的识别准确率达到了98.3%，相比较GADF VggNet，GADF ResNet，GADF ConvNeXt等网络模型，分别提高了16.7%，1.4%和4.05%。结果表明，所改进模型提升了原始模型在处理小样本条件下故障诊断效果，并且在不同故障直径滚动轴承条件下，故障诊断的准确率优于其他模型，且具有较强的泛化性和鲁棒性。
Based on the characteristics of existing network models (ConvNeXt) that cannot fully provide fault information, require a large amount of sample data, have insufficient generalization performance, and low robustness in processing small sample tasks, an improved ConvNeXt network model for bearing fault diagnosis is proposed. Firstly, the Gram angle difference field image encoding technique is used to decode the fault samples and obtain the corresponding fault feature maps; Then, by using methods such as random cropping and rotation, the data augmentation module is improved; Secondly, the ConvNeXt model is reconstructed using asymmetric convolution to enhance its ability to handle small sample tasks; Finally, incorporating the CBAM attention mechanism enhances the model's ability to extract signal features in both channel and spatial aspects. The experiment showed that the improved ConvNeXt network model achieved a recognition accuracy of 98.3% for different fault diameters of rolling bearings. Compared with other network models such as GADF VggNet, GADF ResNet, and GADF ConvNeXt, it improved by 16.7%, 1.4%, and 4.05%, respectively. The results showed that the improved model improved the fault diagnosis performance of the original model under small sample conditions, and the accuracy of fault diagnosis was better than other models under different fault diameter rolling bearing conditions, with strong generalization and robustness.