A Novel TBAFP-GroupConv for Intelligent Fault Recognition of Axlebox Bearings in Railway Vehicles

The operating environment of train axlebox bearings is highly complex due to track irregularities and the vibrations and loads transmitted from the train body, making fault detection and identification particularly challenging. Existing multi-scale methods often increase model complexity and computational demands while failing to establish effective cross-scale correlations. This limitation hinders the ability of the model to capture multi-scale information, especially in noisy and complex environments, ultimately compromising the model stability. To address this issue, this paper proposes a novel vision-based fault-identification architecture called TBAFP-GroupConv. This approach first converts raw vibration acceleration signals into time-frequency spectrograms using continuous wavelet transform (CWT). Top-down and bottom-up attention-based feature pyramids (TBAFPs) incorporate bidirectional skip connections along with specially designed local and global attention mechanisms to enable efficient multi-scale feature fusion. Additionally, a group convolution (GroupConv) strategy was employed to extract fine-grained fault features while minimizing the critical information loss. Furthermore, the loss function was optimized by introducing a temperature parameter T and enhancing the robustness and generalization capabilities of the model through knowledge distillation, transferring prior knowledge from the teacher model ResNet50 to TBAFP-GroupConv. Experimental validation using both open-source datasets and data collected from a rail vehicle traction drive system test bench demonstrates that the proposed algorithm significantly outperforms state-of-the-art classification models in terms of accuracy and adaptability. Experimental results on both open-source datasets and data collected from the rail vehicle traction drive system test bench demonstrate that the proposed algorithm achieves an accuracy of 96.48% and an MCC score of 0.9599 under noisy conditions (SNR = -4dB). This represents a 5.9% improvement in the detection accuracy compared with conventional fault detection models. Paired t-tests conducted on the MCC metric yielded p-values below 0.05, indicating statistically significant performance differences between each benchmark model and the proposed approach.