基于时间迁移模型的旋转机械实时故障诊断

A new time transfer model is proposed to enhance the real-time fault diagnosis performance of rotating machine when the working condition change occurs. Here the source domain is composed of historical data and the target domain is composed of current measurement data. Firstly, the data domains of the model are determined according to the varying working condition rules, and their time-domain feature vectors are extracted to construct the five-dimension spaces. Secondly, the source and target domains are mapped into a two-dimension sub-space using the maximum variance projection (MVP) and the manifold regularization projection (MRP), respectively. Meanwhile, the minimum mean difference (MMD) criterion is used to minimize the distance between source domain and target domain in two-dimension space. Finally, in the projection space, the BP neural network and support vector machine (SVM) classifiers are adopted to build the classification models of the source domain, which are applied in target domain. Also, the diagnostic model is updated through selecting the samples in source domains. Experiments on the gear drive-train system were conducted, the experiment results prove that the time transfer model can solve the real-time mechanical fault diagnosis problem when the working condition change occurs. Compared with traditional transfer component analysis (TCA) model, the proposed time transfer model can improve the diagnostic performance, the proposed model provides a valuable technical solution for the engineering application of mechanical fault diagnosis.