Simulation-derived semantic features enabling zero-shot domain adaptation for rail fastener failure detection in unseen scenarios

Abstract

Rail fasteners are crucial components in maintaining the structural integrity of railway infrastructure, and timely identification of fastener failures is essential for ensuring operational safety. Conventional deep learning methods for structural health monitoring rely heavily on extensive labelled failure data for effective model training. This becomes particularly challenging when dealing with early-stage failures (e.g., fastener loosening), where such labelled data may be scarce or unavailable. To address this issue, we propose a simulation-derived semantic features enabling zero-shot domain adaptation (SSFZSDA) method, designed to enhance the generalization capabilities by transferring learned semantic features from simulated datasets to unlabelled real-world datasets. Specifically, semantic features that implicitly represent various structural health conditions are extracted from the simulation dataset using vehicle-track coupled dynamics. By integrating these semantic features with real-world healthy data through domain adaptation and contrastive learning techniques, the proposed method is capable of effectively detecting failures in previously unseen practical scenarios. The results demonstrate that SSFZSDA achieves 99.7% accuracy in detecting three distinct levels of fastener loosening based on track acceleration data, while also demonstrating excellent performance based on vehicle-mounted acceleration data, achieving 92% identification accuracy. The proposed method’s effectiveness and generalization are validated through comparative analysis, outperforming other state-of-the-art zero-shot domain adaptation methods.