Defect detection on rail lines is essential for ensuring safe and efficient transportation. Current image analysis methods with deep neural networks (DNNs) for defect detection often focus on the defects themselves while ignoring the related context. In this work, we propose a fusion model that combines both a targeted defect search and a context analysis, which is seen as a multimodal fusion task. Our model performs rule-based decision-level fusion, merging the confidence scores of multiple individual models to classify rail-line defects. We call the model “hybrid” in the sense that it is composed of supervised learning components and rule-based fusion. We first propose an improvement to existing vision-based defect detection methods by incorporating a convolutional block attention module (CBAM) in the you only look once (YOLO) versions 5 (YOLOv5) and 8 (YOLOv8) architectures for the detection of defects and contextual image elements. This attention module is applied at different detection scales. The domain-knowledge rules are applied to fuse the detection results. Our method demonstrates improvements over baseline models in vision-based defect detection. The model is open for the integration of modalities other than an image, e.g., sound and accelerometer data.
18 December 2025 in IEEE Transactions on Instrumentation and Measurement
The expansion and increasing complexity of railway infrastructure, combined with a growing demand for higher safety and maintenance standards, has driven important innovation in rail defect detection. This review examines recent methods for railway track defects identification, with a particular focus on their deployment on embedded computing architectures. Detection methods are categorised across multiple sensing modalities—vision, acoustics, vibration, and electromagnetic—while highlighting recent advances in deep learning. This study addresses the critical gap between the performance of algorithms and their potential to be deployed on hardware architectures to design reliable, real-time systems. The review evaluates these approaches in terms of suitability for real-time onboard deployment, identifies their limitations, and proposes a new multisensory, embedded system that will balance performance, energy efficiency, and scalability. A study of detection methods and their in-depth evaluation aims to bridge the gap between complex and high-accuracy detection algorithms and their integration into lighter railway monitoring systems.
1 December 2021 in Journal of Infrastructure Systems
Real-time train monitoring is becoming a major challenge for infrastructure managers (IMs). Infrastructure failures caused by the poor condition of trains significantly increase maintenance costs and cause numerous traffic disruptions. Currently, a conventional approach based on the total tonnage supported by the track over 1 year commonly is used to plan maintenance work. Tonnage measurement and wheel-condition analysis stations, offered by many companies, enable IMs to know the status and load of trains in real time and to alert railway companies in the event of a defective train. These systems provide IMs with the tools they need to move from passive to proactive maintenance. The main disadvantage of these systems is their high acquisition cost, which limit their use to main tracks. Monitoring the French network requires the acquisition of about 100 stations installed on the main and secondary tracks, representing an investment about €15 million. Because of this high cost, we sought to develop other, less expensive solutions to equip the secondary tracks for which the use of these systems is not economically justified, notably by transforming the rail bridges into dynamic scales. The acquisition cost of this system has been estimated at €20,000 to which must be added the installation costs (about €5,000). Thus the generalization of this device on the French network requires an investment of €3 million, a saving of about €10 million compared with traditional systems.
SNCF Réseau introduces a novel multi-modal embedded monitoring system, addressing challenges in railway infrastructure maintenance. The design incorporates visual, inertial, and sound sensors, enhancing adaptability, improving overall detection precision, and could reduce operational costs. This study addresses visual defects detection that can be integrated in a multi-modal monitoring system. The paper details the system’s architecture, synchronisation methods, and decision fusion process to improve the precision of limited mono-modal systems. A deep-learning visual based railway defects inspection was explored. Results show that small CNN (Yolov8 nano) can achieve similar (Yolov8 XL) high precision (mAP@0.5 ≥ 0.89) for a small number of objects (9) while improving implementation capability on embedded systems.
La tendance actuelle en matière de capteurs et d’équipements industriels déployés sur le réseau ferré nous amène à gérer des systèmes de plus en plus complexes qui obligent les agents à travailler dans un environnement incertain. Dans le même temps, le domaine du transport ferroviaire devient de plus en plus concurrentiel, ce qui contraint les gestionnaires d’infrastructures à une recherche permanente d’amélioration, d’optimisation et de productivité. Parallèlement, l’utilisation des techniques de surveillance classiques est de plus en plus coûteuse, tout en offrant des performances de moins en moins satisfaisantes. Les fonctions de surveillance du réseau sont souvent opérées hors ligne, empêchant ainsi le traitement en temps réel de l’information. L’intelligence artificielle offre des outils totalement dissociés de la structure de l’infrastructure, ne nécessitant pas la modélisation préalable de cette dernière et permettant un suivi en temps réel de son évolution. Cette nouvelle approche de la maintenance s’avère par conséquent évolutive et plus en adéquation avec les contraintes économiques auxquelles est soumise notre activité. Dans ce cadre, nous avons étudié une nouvelle architecture combinant l’emploi de plusieurs couches de réseaux neuronaux profonds et un modèle de fusion. Cette solution permet de garantir un taux de disponibilité optimale du réseau et de préserver l’infrastructure existante par une maintenance au juste à temps et au juste nécessaire. Pour le système destiné aux trains commerciaux, nous avons abordé les problématiques de programmation distribuée telles que la co-allocation des ressources. Une étude économique complète cette étude.