基于选择性加权适配网络的多域新故障识别方法

Selective weighted adaptive network for multi-domain emerging fault identification

针对深度迁移学习诊断方法要求机械设备训练数据与测试数据具有相同类别空间,同时难以有效识别新故障的问题,提出了一种基于选择性加权适配网络的多域新故障识别方法。所提方法利用一维卷积神经网络提取源域与目标域深度判别特征,并集成领域判别器与多分类器结构,构建源域与目标域权重函数,自适应度量源域与目标域类别的相似程度;从而利用对抗学习策略来有效减少源域与目标域共享类数据的分布差异;最后利用高斯分布拟合方法自动判别权重阈值,实现对目标域已知故障和新故障的有效诊断。在齿轮箱变工况迁移诊断任务上对所提方法进行分析与应用验证,并与现有的其它方法进行比较,所提方法在所有任务上的调和平均值(E-score)达到0.8以上,验证了所提方法的有效性与优越性。

The existing deep transfer learning-based diagnosis methods usually require that the same fault class space is shared by training and test data,which fail to effectively identify new faults.Thus,a multi-domain emerging fault identification method based on selective weighted adaptive network is proposed.Firstly,a one-dimensional convolutional neural network is adopted to extract depth discriminative features across domains.Then,a domain discriminator and multi-classifier structures are integrated to construct weight functions of source and target domains to adaptively measure the similarity across different categories.The adversarial learning strategy is utilized to effectively reduce the distribution differences of shared classes across domains.Finally,the Gaussian distribution-based fitting method is adopted to automatically discriminate weight thresholds to realize effective fault diagnosis of known faults and emerging faults in the target domain.Experiments are conducted on a gearbox transmission test rig,where the transfer diagnosis tasks under variable operation conditions are designed.The proposed method obtains 0.8 E-score in various tasks.The effectiveness and the superiority of the proposed method are fully validated in comparison with other existing methods.