25 October 2024 in IEEE Transactions on Intelligent Transportation Systems
Real-time monitoring and analysis of sensitive areas in high-speed railway (HSR) are crucial for ensuring the safe and smooth operation of high-speed trains. To address the problem of frequent missing and false alarms caused by anomaly data in HSR monitoring system, this study proposes an innovative network framework: Intelligent detection network of HSR anomaly monitoring data (HSRA-Net). The framework comprises of two modules: the data augmentation module and the anomaly detection module. The data augmentation module designs multiple alternative generative adversarial networks for sample augmentation. To achieve the end-to-end classification, the anomaly detection module improves the residual network by creating a deep residual shrinkage network with self-attention (DRSN-SA). An online monitoring system was installed and operated continuously for several years on a high-speed turnout of a continuous beam bridge to validate the proposed framework. The collected data includes displacement, stress, and temperature. The proposed framework has demonstrated excellent performance, generalizability, and deployability through sufficient model comparison. It can accurately and efficiently diagnose anomalies in the operation of the monitoring system. This study is of great significance for improving the anomaly detection task of the HSR monitoring system.
26 August 2025 in Engineering Structures
9 April 2024 in International Journal of Rail Transportation
17 November 2025 in Reliability Engineering & System Safety
16 May 2026 in SSRN Electronic Journal
Subgrade differential settlement can compromise ride safety and operational stability by intensifying wheel–rail interaction and vehicle–track vibrations. Early-grade identification remains challenging due to weak onset response variations and the nonstationarity of acceleration signals. Conventional data-driven models often provide neither physics-consistent indicators nor well-calibrated confidence for threshold-based alarms and risk grading. Physics-Aware Time-Series Mixer with feature-wise linear modulation priors (PATS) is developed for ordinal identification of settlement amplitude A, wavelength L, and lateral coefficient K using carbody lateral and vertical accelerations. PATS integrates (1) improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) to separate low-frequency geometric trends from mid-to-high-frequency impacts, (2) physics-informed descriptors, namely sensitive intrinsic mode function (IMF) energy priors and statistics-based cumulative characteristic function (SCCF) features, injected into a lightweight Time-Series Mixer (TSMixer) backbone via feature-wise linear modulation (FiLM), and (3) a multi-group ordinal head based on consistent rank logits (CORAL) to output reliability-aware probabilities. Experiments show 97.41% accuracy with 0.0187 expected calibration error (ECE) on K at 1.050 ms/sample, 96.30% accuracy with 0.0350 ECE on A at 0.964 ms/sample, and 100.00% accuracy with 0.0011 ECE on L. The artificial intelligence contribution lies in an interpretable physics-guided prior injection mechanism for nonstationary multivariate time-series grading, coupled with a reliability-oriented ordinal output design. The engineering application contribution is a deployable onboard monitoring framework for early-grade identification of high-speed railway subgrade differential settlement from onboard accelerations, enabling threshold-based alarming and risk grading.
15 June 2026 in Engineering Applications of Artificial Intelligence