A Cross-Domain Intelligent Fault Diagnosis Method Based on Multi-source Domain Feature Adaptation and Selection

Abstract

Although fault diagnosis methods integrating transfer learning are research hotspots, their ability to handle industrial fault diagnosis problems with large domain differences still needs to be improved. A multi-source domain feature adaptation and selection (MDFAS) method is presented to address the issues of domain mismatch and domain negative transfer. The method integrates the top-level network parameter transfer strategy with the 2D Convolutional Neural Network (2DCNN) backbone network to acquire the target domain feature extractor quickly. Multiple feature adaptive extractors (FAEs) are constructed using a multi-branch structure to align the source and target domain's feature distributions, respectively. The inter-domain distance computed by multi-kernel maximum mean discrepancy (MK-MMD) is embedded in the FAEs loss function to improve the inter-domain matching degree. Based on the information gain of the adaptively integrated features, the ensemble adaptive selection is performed on the extracted feature matrices to exclude the negative transfer feature. Finally, the effective feature matrix is input into the diagnosis classifier for classification. Cross-domain fault diagnosis experiments are developed based on the data set gathered from several types of rotating machinery operated under varied working conditions. The experimental results show that the proposed method outperforms the existing intelligent fault diagnosis methods in terms of fault detection accuracy, generalization, and stability.