Label self-correction intelligent diagnosis method and embedded system for axle box bearings of high-speed trains with noisy labels

Abstract

Due to annotation errors, delayed labeling, and noise interference, data label noise is a common issue in high-speed train datasets, leading to overfitting of existing intelligent diagnostic methods on noisy-label samples and a decline in the accuracy of fault diagnosis, which affects the correct assessment of high-speed train bearing health. To tackle this issue, this article presents an adaptive label self-correction intelligent diagnostic method. The method consists of three main parts: First, it employs dynamic thresholds and multi-network interactive training to separate clean from noisy labels. Second, it corrects noisy labels using classifiers trained on clean data, with two designed correction methods for high-accuracy label correction. Third, it retrains the model by reweighting loss to ensure that it fully captures information from noisy label data. Additionally, based on the proposed method, an AI microprocessor diagnosis system is developed for real-world health monitoring of axle box bearings. Both the method and the system have been validated through diagnostic cases of axle box bearings. Validation through diagnostic cases demonstrates that the method can train high-accuracy diagnostic models under label noise conditions and the system can rapidly diagnose data in real-time.