Data-Driven Shield Tunnel Leakage Detection via Context-Aware Representation Learning and PPO-Lagrangian Reinforcement Learning

Abstract

Accurate and reliable leakage detection is essential for the intelligent maintenance of shield tunnels, yet conventional image-based methods often suffer from false alarms caused by illumination variation, surface moisture, structural shadows, and insufficient contextual information. This study proposes a context-aware representation-enhanced PPO-Lagrangian model for shield tunnel leakage detection, aiming to improve detection accuracy and operational reliability under complex inspection conditions. The proposed framework integrates visual features from tunnel inspection images with structural and operational context information, including tunnel ring location, joint adjacency, illumination condition, drainage status, and historical leakage records. A context-aware representation module is developed to generate enriched state features, while a PPO-Lagrangian policy optimization strategy is introduced to balance leakage recognition performance and false-alarm constraints. Experiments were conducted on a simulated shield tunnel inspection dataset containing 18,920 image windows covering crown, sidewall, invert, joint, drainage, and equipment-adjacent regions. The proposed model achieved an accuracy of 95.8%, a macro-F1 score of 94.2%, and a confirmed-leakage recall of 93.7%, outperforming Swin Transformer (93.1% accuracy) and DeepLabV3+ (92.4% accuracy). The average false alarm rate was reduced to 4.6%, compared with 7.9% for Swin Transformer and 9.1% for the unconstrained PPO model. Ablation experiments showed that removing context tokens decreased macro-F1 by 3.8 %, while removing the Lagrangian penalty increased false alarms by 2.6 %. The proposed framework provides an interpretable and risk-aware solution for intelligent shield tunnel inspection and supports practical maintenance decision-making.