Abstract
As the operating speed of high-speed trains continues to rise, the implementation of fault diagnosis under different working conditions remains challenging due to the scarcity of fault data. The issue of how to use existing labeled data to accurately and efficiently diagnose unlabeled data under even higher-speed conditions is an urgent problem to be solved. Addressing the aforementioned challenge, a novel transfer learning framework is proposed to establish a mapping relationship between feature distribution discrepancies and diagnosis performance. This framework evaluates the feasibility of effective transfer learning from lower-speed to higher-speed condition data. A transferability predictor is obtained through domain traversal transfer learning, feature distribution discrepancy computation, and ensemble regression learning. This predictor can select reliable source domain data for transfer learning without labels, ensuring high-precision diagnosis under unknown conditions. In order to achieve adequate validation of the proposed model, a variable-speed condition fault bearing test is conducted on a real high-speed train using a rolling test bench. The model is trained using the data collected from the real high-speed train, and the results show that the proposed method is able to select the source domain data adaptively and ensure high accuracy of higher speed condition monitoring data diagnosis.
Funding Information
  • Youth Science Fund of National Natural Science Foundation of China (52302446)
  • National Natural Science Foundation of China (U2268205)
  • Natural Science Foundation of Hunan Province China (2024JJ6500)

This publication has 32 references indexed in Scilit: