79 / 2023-09-09 17:59:14

A small sample gear wear diagnostic method based on a MWCJM domain adaptation transfer learning network

Gear wear, small sample problem, domain adaptation transfer learning,domain measurement methods

Gears, as a pivotal component, have widespread applications in the industrial domain. They are susceptible to wear and failure, especially under complex operating conditions and prolonged meshing movements. Diagnosing and monitoring gear wear holds significant importance. However, practical engineering applications often feature an abundance of normal samples under low-load conditions, resulting in a scarcity of fault samples. Furthermore, fault samples are even rarer under high-load conditions, leading to an issue where the model encounters an imbalance of large low-load samples and limited high-load samples, hampering its ability to effectively carry out fault diagnosis tasks. In response to this challenge, this paper establishes a transfer diagnostic model based on three measurement methods: Maximum Mean Discrepancy, Wasserstein, and CORAL. It utilizes domain-invariant features between different operational conditions for transfer learning. Based on this, a MMD-Wasserstein-CORAL joint measurement method (MWCJM) was proposed. By training the target domain model with a small amount of gear data from four different wear levels under high load conditions, the accuracy on the target domain test set can reach 94.5%, completing the migration diagnosis task of gear wear level under high load conditions. After comparison, it can be concluded that the migration model based on MWCJM has the highest classification accuracy and the best migration effect in the target domain test set after training.