【摘要】 为实现生产噪声环境下小样本、多型号的轴承套圈磨削烧伤现象的有效识别，避免不合格零件流入装配环节，提出一种多元模态分解(multivariate variational modal decomposition, MVMD)和残差神经网络相结合的轴承套圈磨削烧伤识别方法。首先，利用灰狼算法对MVMD进行参数寻优，筛选本征模态函数进行信号重构，实现多元信号联合去噪；其次，将去噪后的信号利用格拉姆角场转换为二维图像并进行多通道融合，获得红绿蓝融合特征图；然后，将其作为输入构建融合多注意力的识别模型GMVMD-ECA-ResNet-MA进行磨削烧伤特征提取及分类，再使用不同型号轴承套圈数据并微调基础模型权重参数，进行迁移学习，实现轴承套圈多型号烧伤识别。最后，试验结果表明：GMVMD-ECA-ResNet-MA在仅有少量训练样本的情况下，烧伤识别率依然可达90%以上。与其他模型进行对比，两组迁移任务中所得模型的平均识别准确率分别为94.44%与95.83%,因此，本文所提方法得到的模型在生产噪声环境下具有更高的识别准确率和更强的泛化能力。更多还原

【Abstract】 Here, to effectively recognize grinding burn phenomenon of small samples and multiple types of bearing rings under production noise environment, and to avoid unqualified parts flowing into assembly process, a bearing ring grinding burns recognition method combining multivariate variant mode decomposition(MVMD) and residual neural network was proposed. Firstly, grey wolf optimized algorithm was used to optimize parameters of MVMD, select intrinsic mode functions for signal reconstruction, and realize joint denoising of multivariate signals. Secondly, denoised signals were converted into 2-D images using Gramm angle field for multi-channel fusion to obtain red green blue fusion feature maps. Then, these maps were used as input to construct a multi-attention fusion recognition model GMVMD-ECA-ResNet-MA for grinding burns feature extraction and classification. Using different types of bearing ring data and finely tuning weight parameters of basic model, transfer learning was performed to realize multi-type bearings ring burn recognition. Finally, the experimental results showed that GMVMD-ECA-ResNet-MA model can realize a burn recognition rate of over 90% with only a small number of training samples; compared with other models, average recognition correctness rates of this model obtained in 2 groups of transfer tasks are 94.44% and 95.83%, respectively, so GMVMD-ECA-ResNet-MA model can have higher recognition correctness rate and stronger generalization ability under production noise environment.更多还原