Answers are grounded in this paper and cite their sources.
Data-driven Structural Health Monitoring (SHM) methods are increasingly adopted for railway infrastructure monitoring, but their performance may deteriorate when operational conditions differ from those used during model development. This work investigates the generalization of a data-driven SHM methodology for bonded insulated rail joints (IRJs) under heterogeneous railway traffic conditions, focusing on domain shifts induced by variations in train speed and axle configuration. Starting from a Mahalanobis distance-based damage detector developed under homogeneous operating conditions, the study evaluates whether statistical alignment techniques can improve transferability without retraining it. The study assesses three lightweight alignment strategies (Mean Shift, Mean and Scale, and CORAL) using a long-term field dataset from fiber Bragg grating (FBG) sensors installed on IRJs operating under real traffic conditions. The analysis considers transfer scenarios involving changes in axle configuration, train speed, and their combined effects. Results show that the effect of domain shift depends on the specific operating condition and may affect both detector specificity and sensitivity. At the same time, simple mean-based alignment strategies substantially recover detection performance, often approaching that of a fully retrained model. Covariance-based alignment through CORAL generally provides the closest statistical alignment of the healthy distributions, but in some conditions this is accompanied by reduced damage sensitivity. The results demonstrate the potential of lightweight statistical alignment for extending existing SHM methodologies to heterogeneous operating conditions while preserving the original detector and reducing the need for complete baseline recalibration.
No takes yet. Share an insight, caveat, or question.
Cavaciuti et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: