This review examines the integration of artificial intelligence (AI) with fibre optic sensing (FOS) for automated structural health monitoring (SHM) of railway infrastructure, encompassing embankments, bridges, tracks, tunnels, and turnouts. Railways are among the most sustainable and strategically critical modes of transportation; however, their distributed nature, increasing axle loads, and harsh environmental conditions present significant monitoring challenges. Distributed fibre optic sensing (DFOS) systems, including Fibre Bragg Grating (FBG), Distributed Acoustic Sensing (DAS), and Brillouin- and Rayleigh-based scattering sensors, are identified as uniquely capable platforms for real-time, long-range, and high-resolution measurement. Sensor deployment configuration, including gauge length, spatial resolution, and signal-to-noise ratio, critically determines whether a system supports train localisation or enables detection of track component defects such as fastener loosening, rail cracks, and ballast degradation. The convergence of AI with FOS has redefined SHM as a data-driven and predictive discipline. Machine learning algorithms including Q-Learning and Deep Q-Networks (DQN), deep learning architecture such as Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU), and reinforcement learning (RL)-based frameworks enable automated defect detection, localisation, and maintenance optimisation. Emerging frameworks incorporating Large Language Models (LLMs) and Digital Twin (DT)-integrated RL offer new opportunities for adaptive, closed-loop decision-making in railway maintenance. This review systematically identifies principal deployment challenges, including sensor cross-sensitivity, signal attenuation, massive data volumes from DAS systems, environmental degradation of protective coatings, labour-intensive installation, high interrogator costs, and power constraints at remote trackside locations. Railway-specific resolutions and alternatives are presented for each challenge, grounded in current literature.
Siahkouhi et al. (Fri,) studied this question.