【摘要】 轨道交通关键零部件面临复杂工况下异常磨损问题，隧道掘进机(Tunnel Boring Machine,TBM)滚刀与列车制动闸片作为典型强摩擦部件，频发滚刀多边形磨损、崩刃、偏磨和制动摩擦块偏磨等异常磨损，导致系统服役性能退化并威胁运行安全.异常磨损后部件界面呈现复杂非线性接触-振动状态，且受工况干扰，传统状态识别方法的适应性和鲁棒性亟待提升.文中围绕TBM滚刀与制动摩擦块的异常磨损，开展破岩试验与制动试验，揭示了接触-振动行为演化规律，有针对性地构建了深度学习识别模型.破岩试验发现，偏磨滚刀滚动方向的振动均方根值达正常状态的15倍以上，峰态值低，表现为能量集中、模式单一的剧烈振动；多边形磨损峰态值最高，频率扰动和周期性显著；崩刃滚刀的状态与正常磨损接近.针对滚刀状态识别提出了融合多谱特征提取、对抗学习与物理纠偏机制的半监督网络(Physics-guided Feature-Spectral Manifold Adversarial Network,PFSMAN)，在仅使用100个标签样本时准确率达85.04%；制动试验中，偏磨摩擦块在8 kHz频段幅值显著增强，表明其模态耦合诱发系统局部不稳定.针对闸片偏磨识别构建了引入二次卷积神经元与权衡对比学习机制的TCLN(Text-based Contrastive Learning Network)模型，在样本不平衡率达100:1时仍保持93.85%的准确率.研究成果为轨道交通装备中典型强摩擦部件的异常磨损监测与智能诊断提供了理论支持与工程参考.更多还原

【Abstract】 In complex service environments, abnormal wear of critical components poses a significant challenge for rail transit equipment. As typical high-friction components, tunnel boring machine(TBM) disc cutters and train brake pads are frequently subjected to abnormal damage, such as polygon wear, chipping, and partial wear, under harsh operational conditions. These wear phenomena severely degrade system performance and compromise operational safety. After abnormal wear occurs, the friction interfaces of such components exhibit complex nonlinear contact-vibration behavior and are further affected by strong environmental disturbances, limiting the adaptability and robustness of traditional state recognition methods. This study focuses on typical abnormal wear scenarios of TBM disc cutters and train brake pads.Multi-cutter rock cutting experiments and train braking tests were conducted to systematically investigate the evolution of contact-vibration behavior induced by abnormal wear. Targeted deep learning-based recognition models were then developed to enable accurate identification of these wear conditions. Experimental results from the rock cutting tests show that the RMS value of vibration in the rolling direction of partial worn cutters exceeds that of normal cutters by more than 15 times, with the lowest kurtosis, indicating intense, energy-concentrated vibrations with a single-mode pattern. Polygon wear exhibited the highest kurtosis and the strongest periodic frequency disturbances, while chipping produced vibration patterns similar to normal wear. To address cutter wear state recognition, a semi-supervised network named PFSMAN(Physics-guided Feature-Spectral Manifold Adversarial Network) was proposed, which integrates multi-spectral feature extraction, adversarial learning, and a physics-guided correction mechanism. With only 100 labeled samples, it achieved an accuracy of 85.04%. In the braking tests, partial wear of the brake pads caused a significant amplitude increase around 8 kHz, suggesting local system instability triggered by modal coupling. For identifying brake pad wear, a model named TCLN(Text-based Contrastive Learning Network), featuring second-order convolution neurons and a contrastive learning balance mechanism, was developed. Even under extreme class imbalance(100:1), it achieved a recognition accuracy of 93.85%. The findings offer theoretical insights and practical references for the monitoring and intelligent diagnosis of abnormal wear in key high-friction components of rail transit systems.更多还原