LEIPL: A novel fault diagnosis method for high-speed train axle box bearings under noisy labels

Reliable fault diagnosis of high-speed-train axle box bearings is challenged by noisy annotations in maintenance data. This paper proposes LEIPL, a noise-robust diagnostic method that expands the original label space with an auxiliary negative class and unifies positive and negative supervision. The original noisy label and top-K predictions are used to construct candidate positive labels, while the remaining labels provide negative supervision. A confidence-regularization term is further introduced to stabilize model optimization. Experiments on the HTBF and BJTU-RAO datasets show that LEIPL achieves accuracies of 99.16% and 99.20%, respectively, under -4 dB Gaussian white noise and 50% symmetric label noise. The results demonstrate strong robustness and low computational complexity under severe noisy-label conditions.