Logic-guided hierarchical attentive taxonomy network for event-level ultrasonic rail defect detection

Machine learning methods for Real-world Railway ultrasonic testing (UT) face a critical bottleneck rooted in extreme few-shot data scarcity and severe long-tail class distribution, compounded by high signal similarity between defective and healthy regions (especially around bolt holes). These issues trigger model overfitting, unstable convergence and unreliable predictions, resulting in logically contradictory per-frame outputs violating physical domain rules and fragmented predictions that hinder maintenance decision-making. Unlike previous works, we propose the first event-level Railway UT defect detection framework, termed Hierarchical Attentive Taxonomy Network (HAT-Net), which addresses this fundamental issue through two complementary mechanisms. A Taxonomic Consistency Loss enforces ontology-derived mutual-exclusion constraints as a soft training regulariser, eliminating logical violations without inference-time overhead. A Hierarchical Attentive Perception Backbone—dilated residual 1D convolutions with Squeeze-and-Excitation channel attention followed by a BiLSTM smoothing layer – captures long-range temporal context while adaptively reweighting noisy probe channels and outputting event-level predictions. Experiments on continuous multi-channel ultrasonic data show that HAT-Net achieves Event F2 of 0.976, mean Average Precision (mAP) of 0.942 (±0.023) and Constraint Violation Rate (CVR) of 0.000%, outperforming prevalent baselines. Improve detection efficiency while making maintenance work more effective. Robustness studies also demonstrate its relatively stable performance against noise and few-shot settings, indicating improved generalisation.