DECENTRALIZED MULTI-AGENT REINFORCEMENT LEARNING ARCHITECTURE FOR RAILWAY TRACK DAMAGE DETECTION IN TRAIN-BASED MONITORING SYSTEMS

Abstract

 Reliable monitoring of railway track conditions is essential to ensure operational safety and support predictive maintenance. This study proposes the first decentralized Multi Agent Deep Reinforcement Learning framework for real time rail damage detection in a Train Based Monitoring System. The architecture integrates three agents: DQN for front bogie vibration analysis, PPO for rear bogie vibration signals, and TD Learning for IMU based motion estimation. The models were trained using synthetic Root Mean Square vibration and IMU datasets within a Gymnasium based simulation environment representing eleven rail surface conditions. Individually, the TD Learning model achieved 49.0 percent accuracy, while the DQN and PPO models achieved 87.8 percent and 86.14 percent accuracy, respectively. A greedy ensemble fusion strategy was applied to combine the predictions of the three agents. The ensemble model achieved the best performance with 88.6 percent classification accuracy and a weighted F1 score of 0.887, demonstrating improved classification stability compared with standalone models. Several defects including corrugated rail, defect, and severe transverse crack were detected with perfect F1 scores. These results indicate that multi sensor reinforcement learning fusion improves robustness and provides a scalable solution for intelligent railway condition monitoring and automated defect detection.