Spatial Residual Stress Gradient‐Aware Railway Wheel–Axle Crack Classification and Risk Prioritization Using SPIRiT‐LUTanh‐DNN

ABSTRACT Railway wheels and axles are prone to cracks, abrasion, and stripping due to wheel–rail contact, operational loading, material properties, and environmental conditions. Existing works neglected the spatial residual stress gradient during early crack prediction, leading to reduced classification accuracy. Therefore, an enhanced railway wheel–axle crack classification and severity‐based risk prioritization model through spatial residual stress gradient assessment using SPIRiT–linear unit hyperbolic tangent–deep neural network (SPIRiT‐LUTanh‐DNN) is proposed. Initially, railway wheel and axle images undergo preprocessing, resolution enhancement, background removal, segmentation, crack detection, and crack initiation analysis. Patch feature vectors are then extracted, and the residual stress gradient is estimated using the sparse Xavier hidden Markov model (SX‐HMM) to construct a stress map. Finally, stress, crack, and initiation features are fused for defect classification with explanations, followed by severity‐based prioritization. Experimental results demonstrate that the proposed framework achieves a classification accuracy of 99.871%, outperforming existing methods.