A Multi-channel Transfer Learning Framework for Fault Diagnosis of Axial Piston Pump

Abstract

To deal with the problem that the traditional intelligent fault diagnosis models invalid when classifying data with different probability distributions, the Multi-channel Deep Transfer Learning Network (MDTLN) is proposed in this paper. The network structure is divided into three modules: domain adaptation module, condition recognition module and feature extraction module. Firstly, the target domain data are pretrained to obtain its inherent features. Secondly, the model is optimized by maximizing the domain recognition error and minimizing the classification. Finally, the target domain data are accurately classified through the trained model. Multilinear Principal Component Analysis (MPCA) is employed to reduce the dimension of data obtained from multiple sensors. The effect of this method is verified in axial piston pump dataset, and this method has obvious advantages compared with the advanced transfer learning method.