基于联邦学习和改进多尺度ResNet的故障诊断

Fault diagnosis based on federal learning and improved multiscale ResNet

鉴于目前滚动轴承故障数据特征不足导致的模型泛化能力有限,且工业上利用深度学习进行故障诊断存在数据隐私泄露和数据孤岛的问题,提出了一种结合联邦学习和改进多尺度ResNet的模型。改进多尺度ResNet在传统残差块基础上增加了一条恒等映射连接和Uout层,增强了信息流动和模型泛化能力。利用联邦学习保护数据隐私的同时,结合改进多尺度ResNet可以提高模型提取故障特征的能力。不同齿轮的故障诊断模型通过第三方聚合参数,在不泄露数据的情况下实现多方联合训练故障特征。试验结果表明,相较于其他方法,所提方法具有更高的准确率和更好的鲁棒性。

In view of the limited generalization ability of the model caused by insufficient bearing fault data features and the problems of data privacy leakage and data island in deep learning fault diagnosis in industry,a model that combines federal learning and improved multiscale ResNet was proposed.The improved multiscale ResNet adds an identity mapping connection and Uout layer to the basis of the traditional residual block,which enhances the information flow and generalization ability of the model.While using federal learning to protect data privacy,combining improved multiscale ResNet can improve the ability of the model to extract fault features.The fault diagnosis model of different gears realizes multi-party joint training of fault features through third-party aggregation parameters without leaking data.Experimental results show that compared with the other methods,the proposed method has higher accuracy and better robustness.