Answers are grounded in this paper and cite their sources.
Steel rails are critical components of industrial dynamic transportation systems, and their in-service fault diagnosis demands reliable discrimination of multiple defect categories under multi-mode Lamb-wave dispersion and sample-level physical-parameter drift. The rail surface is modelled by a thin metal plate instrumented with two EMAT probes (the standard laboratory surrogate for in-service rail inspection), and the proposed architecture is evaluated on this rail-equivalent plate geometry. Purely data-driven one-dimensional classifiers plateau near 80% test accuracy on a 25,000-sample simulated EMAT A-scan benchmark, while conventional physics-informed neural networks (PINNs) that inject the physical prior only at the loss-function level fail to break this ceiling. We propose a hybrid data–physics residual network, the proposed EMAT-PINN, that couples a convolutional backbone with a logit-orthogonal four-head architecture tying each defect class—hole, crack, corrosion, weld—to one simulator-derived physical quantity (reflected energy, S0/A0 ratio, arrival time, or dispersion shift) via a bias-free additive projection of the class logit. The bias-free construction guarantees that deleting or zeroing any head collapses the affected class logit to the shared baseline, so the remaining heads cannot reroute around the missing head—a structural non-replaceability that supports explainable fault diagnosis. Combined with a four-term physics regression loss (λphys=2.0), the resulting proposed EMAT-PINN attains 95.05% test accuracy at only 0.195 M parameters, with every knockout ablation dropping the model below the 80% threshold commonly referenced as a practical acceptance benchmark. This per-class mapping provides an auditable link between the model’s internal representation and the physical scattering mechanism behind each decision, directly supporting explainable fault diagnosis in industrial deployment.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: