Transformer Incremental Fault Diagnosis Method Using Lossless Estimation and Balanced Training

Abstract

Incremental fault diagnosis is an effective way to address the discrepancies between the actual diagnosis data distribution and the trained model. However, the traditional incremental learning method has the problem of forgetting the feature of historical faults and reducing the diagnostic accuracy when applied. A transformer incremental fault diagnosis method using lossless estimation and balanced training is proposed in this study. Firstly, this method achieves lossless estimation of historical task gradients in incremental updates based on Gaussian–Hermite transformation, thereby preserving historical task information as much as possible. Then, to enhance the “anti-forgetting” capability of historical features while ensuring the ability to learn new features, the method balances the weights between the loss functions of new and historical tasks to optimize the update direction of the network training. Finally, the verification results using real data collected from multiple locations indicate that the proposed method has both “anti-forgetting” ability for historical tasks and the ability to learn new tasks effectively. Moreover, the lowest accuracy rate is 97.04% in the test, and the incremental training efficiency is relatively high.