STCD-Net: A Novel Change Detection Architecture for Anomaly Detection on High-Speed Train Body in Railway Maintenance
Shuang Mei, Hebao Wang, Zhaolei Diao, Jiaqing Zhang, Huabo Shen, Guojun Wen · IEEE Transactions on Instrumentation and Measurement · 2025
High-speed train system is a preferred choice for long-distance transportation due to its speed and stability. To assure its safety and uninterrupted operational reliability, stringent quality inspection and operational maintenance are imperative. Nonetheless, conventional techniques like manual inspection and manual video review suffer from notable constraints. They not only require a high workload and skilled operators but also pose a risk of missed or erroneous detections due to individual factors, leading to safety incidents. This article proposes a novel change detection approach to address the aforementioned challenges. The core concept is to intelligently “compare” the train body images captured at time T2 with the non-abnormal images captured at time T1 when the train just rolled off the production line, thereby facilitating the effective identification of anomalies, including those that have not been previously encountered. Specifically, a novel siamese-transformer based change detection network (STCD-Net) framework is proposed in this manuscript for the inspection task. The core structure of this model is a multiscale residual Siamese transformer encoder (MSRSTE) architecture, specifically tailored for efficiently extracting multiscale differentiated features of images in the compared sets. Furthermore, spatial group enhanced attention gates (SAGs) and a multiscale feature fusion (MSFF) module are also developed to facilitate accurate localization of anomalies. Extensive real-world experimental evaluations have demonstrated that the proposed framework significantly outperforms traditional anomaly detection algorithms, achieving remarkable mean intersection over union (mIoU) scores of 73.57% and 86.29%, respectively. Furthermore, the framework exhibits compatibility with various types of defects while requiring only a modest number of annotated samples for training, showcasing its potential for practical applications, particularly in high-speed train body inspections.