开放集下机床关键部件故障诊断迁移方法

Open set fault diagnosis transfer approach for CNC machine key components

为了解决在开放集下的未知故障迁移问题,提出了基于理论边界优化的渐进式迁移学习网络.首先,对迁移学习模型的误差边界进行了分析,为模型优化提供理论指导;其次,设计基于对抗网络的双辨别器识别未知故障数据,通过粗辨别器输出相似性排序为精细辨别器进一步划分决策边界;最后,通过渐进式辨别器的输出对目标域样本和源域样本进行加权分布匹配.为了全面评估所提出模型的鲁棒性,设计了4种不同程度域偏移的故障迁移案例进行验证,实验结果表明:模型的平均类内识别精度为96.3%,相比其他迁移学习模型在不同开放度差异下表现出更优的诊断效果.

To address the open-set transfer learning with unknown faults,a progressive transfer learning network with theoretical bound was proposed in this paper. Firstly,the error upper bound of the transfer learning model was analyzed,which provided a theoretical basis for model optimization. Secondly,a two-stage adversarial learning module was designed to separate unknown samples,where the similarity obtained from the coarse discriminator was fed into the fine discriminator to further divide the decision boundary. Finally,the target and source domain samples were matched by weighted distribution through the output of the progressive discriminator.In order to comprehensively evaluate the robustness of the proposed model,four diagnosis transfer cases with different domain shifts were designed.The experimental results show that the average intra-class classification accuracy of the proposed model is 96.3%,which shows better diagnostic performance under different openness variances compared with other transfer learning models.