基于非对称性对抗训练的多源域自适应智能故障诊断方法

Multi-source Domain Adaptation Intelligent Fault Diagnosis Method Based on Asymmetric Adversarial Training

使用传统域自适应方法对轴流风机进行跨工况故障诊断时,源域和目标域特征会相向靠拢从而改变训练好的源域特征分布;且当源域故障特征聚集在决策边界时,在域自适应后目标域故障特征同样聚集在决策边界,容易造成部分目标样本错分;此外单源域自适应会影响模型的泛化能力。针对上述问题,提出一种基于非对称性对抗训练的多源域自适应智能故障诊断方法,该方法使用三元中心损失减小源域故障特征的类内距离,增大其类间距离以提高目标样本的区分度;采用非对称性对抗训练方法实现目标域故障特征向源域单向移动;提取不同源域和目标域的域不变特征并输入各自故障分类器,使用余弦相似度对齐各分类器输出的同时施加对齐权重以提高模型的跨域诊断能力。通过试验证明,该方法在解决相关实际工业问题上成效显著。

When using the traditional domain adaptation method for cross-condition fault diagnosis of axial flow fan,the source domain and target domain features will move closer to each other,thus changing the trained source domain feature distribution.And when the source domain fault features are gathered at the decision boundary,the target domain fault features are also gathered at the decision boundary after domain adaptation,which is easy to cause misclassification of some target samples.In addition,single source domain adaptation will affect the generalization ability of the model.For the above problems,a multi-source domain adaptation intelligent fault diagnosis method based on asymmetric adversarial training(TC-MAADA)is proposed.The method first uses triplet-center loss to improve the discrimination of target samples by reducing the intra-class distance and increasing the inter-class distance of fault features in the source domain.Then adopts the asymmetric adversarial training to realize the one-way movement of the target domain fault features to the source domain.Finally,the domain-invariant features of different source and target domains are extracted and input to their respective fault classifiers,using the cosine similarity to align the outputs of each classifier while applying alignment weights to improve the cross-domain diagnostic ability of the model.Experiments show that the method is effective in solving relevant practical industrial problems.