Improved deep transfer learning and transmission error based method for gearbox fault diagnosis with limited test samples

Abstract

As an important component of mechanical transmission system, gearbox state is critical to system safety and efficiency. The fault diagnosis of gearbox is of great significance for monitoring its operation states and identifying potential problems. Firstly, to improve the generalization ability of traditional fault diagnosis model and reduce the diagnostic loss for similar faults occurred in different conditions, an improved deep transfer learning network model is established based on deep subdomain adaptation method and residual feature extraction network. Then, taking a heavy commercial vehicle gearbox as research object, a dynamic simulation model considering its fault state is established, and the transmission error bench test is designed to verify the correctness of the model with different load torque. Finally, under the condition of limited test samples, a gearbox fault diagnosis method based on improved network model and simulation data is proposed and its effectiveness is verified through different comparative experimental tasks and evaluation metrics. The results show that, by using dynamic simulation data of gearbox transmission error, the established deep transfer learning model and proposed gearbox fault diagnosis method can obtain excellent diagnostic performance with high diagnosis precision and low training loss, and excessive test resource investment can be avoided effectively.