Label Self-Correction Intelligent Diagnosis Method and Embedded System for Axle Box Bearings of High-Speed Trains

Label noise is inevitable when manually annotating monitoring data for the axle box bearings of high-speed trains. This issue can cause deep learning-based diagnosis models to overfit on false-labeled samples, thereby reducing the models' diagnostic accuracy. To address this problem, this article presents an intelligent diagnosis method that is resistant to label noise for the axle box bearings of high-speed trains. The proposed method consists of two modules. Firstly, the module leverages the strengths of different networks to independently extract features from dual perspectives. Subsequently, it filters noise labels through model interaction and ingeniously employs a correction model to rectify incorrectly labeled samples. Then, the proposed method is embedded within a hardware-based system, primarily constructed with a microprocessor. Both the method and the system have been validated through diagnostic cases of axle box bearings. The results demonstrate that the proposed method can significantly improve the accuracy of the diagnostic models, and the hardware-based system is capable of successfully displaying signals from faulty bearings and providing real-time diagnostics, thereby further enhancing the reliability and efficiency of the diagnostic process.