Dual-path domain adaptation network for cross-speed brake pads state monitoring in high-speed trains under long ramps

High-speed trains require prolonged continuous braking to control speed under long ramps, causing brake pads to suffer from uneven wear, exfoliation, and crack damage due to sustained heavy loads. However, the rotational speed varies continuously during braking under long ramps, which causes feature distribution differences in the vibration signals collected at different rotational speeds, while the target domain lacks label information. To solve the above issues, this paper proposes dual-path domain adaptation network to reduce the feature distribution differences across different rotational speed conditions, thereby transferring knowledge from the labeled source domain to the target domain. This achieves the cross-speed state monitoring of high-speed train brake pads under long ramps. Firstly, a multi-scale attention feature extraction module is constructed, which combines one-dimensional wide convolution with multi-scale convolution. The improved convolution block attention mechanism is utilized to enhance critical features. Secondly, a dual-path domain adaptation module is employed to reduce distribution differences. It integrates a domain discrepancy minimization path based on multi-kernel maximum mean discrepancy and a domain feature alignment path based on adversarial training with gradient reversal layer. And the dynamic adaptation weight is designed to adaptively adjust the importance of two paths. Finally, experimental results of cross-speed tasks show that the proposed model achieves an average accuracy of 99.33%, and it outperforms comparison methods with accuracy gains of 6.11 and 7.06% under limited single-source domain conditions. These results confirm that the proposed model provides a reliable solution for cross-speed brake pad state monitoring in high-speed trains under long ramps.