Dynamic monitoring of railway bridges via coupling digital twins with deep reinforcement learning

Railway-bridge anomaly detection during retrofitting is a core structural health monitoring (SHM) task, requiring timely alarms for safety. It is challenging because responses are strongly nonstationary and distorted by environmental and operational variability, while sensing budgets constrain what can be observed. Classical feature-and-threshold pipelines are interpretable but fragile to operating-condition drift, yielding unstable alarms without compensation. Deep time-series detectors are effective, but data-only scores are hard to calibrate into risk cues and usually assume fixed sensing. Digital twin (DT) approaches provide uncertainty-aware belief tracking, yet twin-only alarming may respond slowly and may ignore sequential accuracy-cost optimization. We propose Residual Digital-Twin Monitoring with Branching Dueling Q-learning (RDTM-BDQ), coupling an event-triggered Bayesian DT, a mask-aware residual Mamba encoder, and a Branching Dueling Double Deep Q-Network (BDQ) policy. Hybrid evidence fuses DT innovation energy and a normalized safety-margin index with an inverse-projection inconsistency score from the Mamba embedding, forming a belief-augmented state representation for joint alarm and sensing decisions. Training is off-policy with prioritized replay, using a reward that penalizes delay, false alarms, misses, and sensing cost, plus a consistency loss for the inverse projection head. Experiments on the public KW51 and Z24 datasets demonstrate the effectiveness of RDTM-BDQ: on KW51 it achieves 0.95 F1-score (F1) and 0.98 precision, and on Z24 it reaches 0.92 F1-score, confirming robustness to both retrofit-induced and environmentally-induced nonstationarity. Operating-point analysis on KW51 further shows an explicit accuracy-burden-cost trade-off, e.g., reducing detection delay from 41.4 h to 3.9 h with controllable monitoring cost. Ablations confirm that DT cues, residual Mamba encoding, and branching Q-learning each improve accuracy and stability.