Building Trust in Artificial Intelligence: A Necessity for Railway Applications
为提高铁路应用中对AI的信任，本文从鲁棒性、操作设计域和可解释性三方面提出解决方案，以满足安全标准。
0 citationsRead paper
Selected work
Representative Papers
CVCM Track Circuits Pre-emptive Failure Diagnostics for Predictive Maintenance Using Deep Neural Networks
CVCM track circuits constitute a critical signaling subsystem for railway train positioning; however, their latent degradation faults evade early detection by conventional methods relying on prominent signal changes, often triggering cascading operational disruptions. This paper proposes a deep learning–driven predictive maintenance framework tailored to CVCM track circuits, integrating temporal modeling, anomaly detection, and fault classification within an ISO 17359–compliant pre-failure diagnostic paradigm. To ensure robust uncertainty quantification, we incorporate conformal prediction, yielding stable, high-confidence confidence intervals across fault classes. The framework further supports scalable transfer deployment. Evaluated on ten real-world failure cases, it achieves an overall classification accuracy of 99.31% and detects anomalies within the first 1% of the fault evolution timeline—substantially enhancing early-warning timeliness and operational reliability.
0 citationsRead paper
Track Component Failure Detection Using Data Analytics over existing STDS Track Circuit data
To address the challenge of precise component-level fault localization in STDS track circuits, this paper proposes a data-driven diagnostic method leveraging multi-band AC current time-series data. For the first time, high- and low-frequency AC current signals from the STDS system are jointly utilized to extract time–frequency domain features, enabling the construction of an SVM-based classifier for 15 representative fault types. The method is validated on ten real-world track circuits, achieving 100% fault-type identification accuracy—confirmed independently by both domain experts and maintenance personnel. This work overcomes the limitations of conventional diagnostic approaches reliant on single-frequency analysis or empirical judgment, significantly enhancing fault localization accuracy and operational response efficiency. It establishes a practical, deployable technical pathway toward intelligent maintenance of railway signaling systems.
0 citationsRead paper
Scalable, Technology-Agnostic Diagnosis and Predictive Maintenance for Point Machine using Deep Learning
To address the challenges of multi-signal dependency, labor-intensive manual feature engineering, and poor cross-device generalizability in point machine (PM) fault diagnosis, this paper proposes an end-to-end deep learning–based predictive maintenance method leveraging solely single-channel power supply current signals. The approach eliminates handcrafted feature extraction and multi-source signal fusion, directly modeling temporal patterns in actuation current waveforms. Conformal prediction is integrated to rigorously quantify classification confidence, ensuring compliance with ISO 17359. Evaluated across diverse electromechanical point machines, the method achieves >99.99% precision, <0.01% false alarm rate, and negligible missed detection rate. It demonstrates strong robustness, broad cross-device applicability, and high interpretability—significantly enhancing operational reliability and practical deployability in railway signaling systems.
0 citationsRead paper
Recent publications
Latest Papers
Building Trust in Artificial Intelligence: A Necessity for Railway Applications
为提高铁路应用中对AI的信任，本文从鲁棒性、操作设计域和可解释性三方面提出解决方案，以满足安全标准。
0 citationsRead paper
CVCM Track Circuits Pre-emptive Failure Diagnostics for Predictive Maintenance Using Deep Neural Networks
CVCM track circuits constitute a critical signaling subsystem for railway train positioning; however, their latent degradation faults evade early detection by conventional methods relying on prominent signal changes, often triggering cascading operational disruptions. This paper proposes a deep learning–driven predictive maintenance framework tailored to CVCM track circuits, integrating temporal modeling, anomaly detection, and fault classification within an ISO 17359–compliant pre-failure diagnostic paradigm. To ensure robust uncertainty quantification, we incorporate conformal prediction, yielding stable, high-confidence confidence intervals across fault classes. The framework further supports scalable transfer deployment. Evaluated on ten real-world failure cases, it achieves an overall classification accuracy of 99.31% and detects anomalies within the first 1% of the fault evolution timeline—substantially enhancing early-warning timeliness and operational reliability.
0 citationsRead paper
Track Component Failure Detection Using Data Analytics over existing STDS Track Circuit data
To address the challenge of precise component-level fault localization in STDS track circuits, this paper proposes a data-driven diagnostic method leveraging multi-band AC current time-series data. For the first time, high- and low-frequency AC current signals from the STDS system are jointly utilized to extract time–frequency domain features, enabling the construction of an SVM-based classifier for 15 representative fault types. The method is validated on ten real-world track circuits, achieving 100% fault-type identification accuracy—confirmed independently by both domain experts and maintenance personnel. This work overcomes the limitations of conventional diagnostic approaches reliant on single-frequency analysis or empirical judgment, significantly enhancing fault localization accuracy and operational response efficiency. It establishes a practical, deployable technical pathway toward intelligent maintenance of railway signaling systems.
0 citationsRead paper
Scalable, Technology-Agnostic Diagnosis and Predictive Maintenance for Point Machine using Deep Learning
To address the challenges of multi-signal dependency, labor-intensive manual feature engineering, and poor cross-device generalizability in point machine (PM) fault diagnosis, this paper proposes an end-to-end deep learning–based predictive maintenance method leveraging solely single-channel power supply current signals. The approach eliminates handcrafted feature extraction and multi-source signal fusion, directly modeling temporal patterns in actuation current waveforms. Conformal prediction is integrated to rigorously quantify classification confidence, ensuring compliance with ISO 17359. Evaluated across diverse electromechanical point machines, the method achieves >99.99% precision, <0.01% false alarm rate, and negligible missed detection rate. It demonstrates strong robustness, broad cross-device applicability, and high interpretability—significantly enhancing operational reliability and practical deployability in railway signaling systems.
0 citationsRead paper