Semi-supervised multi-indicator noise filtered optimization for fault diagnosis of high-speed train axle box bearings under noisy labels

Abstract

Due to lengthy data collection and manual annotation challenges, fault datasets for high-speed train axle box bearings often contain mislabeled samples. This study proposes a semi-supervised method that integrates multi-indicator noise filtering with robust training to counteract performance degradation caused by label noise. A classifier-distinguished adversarial autoencoder is employed, which utilizes reconstruction error, latent features, and discriminator scores in a voting mechanism to separate clean and noisy labels. A two-stage training strategy is then applied: pretraining a residual network on clean label samples, followed by semi-supervised joint training on all data. Evaluated on two public datasets with 50% symmetric noise, the method achieves an average accuracy of over 90%, outperforming comparable approaches.