Health assessment method of traction motor bearing based on transfer learning and convolutional neural network

To address the difficulties in acquiring labeled life-cycle vibration data and constructing health indicators that can reflect the bearing performance degradation trends in the health assessment of traction motor bearing, a health assessment method based on a transfer learning and convolutional neural network model was proposed for implementing the health assessment of traction motor bearing. The labeled bearing life-cycle data set was used as the source domain data by using the transfer learning, and the comprehensive test bench data were used as the target domain data to construct the data set. Undersampling and synthetic minority over-sampling techniques were used to expand and balance the life-cycle data set, and the abundant sample for convolutional neural networks training was obtained. The features describing the bearing degradation process were extracted in time domain and frequency domain. By using convolutional neural network and following the bathtub curve of bearing performance degradation, the health indicator was constructed by fusing the basic characteristics. Analysis results show that in the health assessment of traction motor bearing shaft current damage, the accuracy of the proposed health assessment method based on transfer learning and convolutional neural network is 98.17%. The accuracies of the methods constructing the health indicator according to linear, quadratic function and parabolic degradation law are 86.61%, 89.56% and 91.30%, respectively. Therefore, the proposed health assessment method has higher accuracy and better evaluation effect in the application. Moreover, the combination of expert knowledge and neural network learning knowledge reduces the fault characteristic dimension, solves the difficulty of health indicator construction, and realizes the health assessment of traction motor bearings through cross-device transfer learning.