Optimization of risk-based and cost-minimized maintenance strategy for corroded pipeline using reinforcement learning

This study proposes an innovative reinforcement learning (RL)-based approach for long-term maintenance planning in multi-zone pipeline systems. The RL-based model aims to optimize maintenance decisions by minimizing total life-cycle cost. A modified Bayesian Neural Network (BNN) was first developed to predict corrosion depth and length. Failure probabilities before and after repair were then calculated. The Dueling Deep Q-Network (Dueling DQN) model was employed to find an optimal, zone-specific maintenance strategy. A comprehensive sensitivity analysis was also conducted to evaluate the impact of key parameters, including consequence cost, failure probability threshold, composite repair effectiveness and operating pressure. Case studies demonstrate the effectiveness of the RL model in providing an adaptive and efficient approach to pipeline maintenance under corrosion growth uncertainty. Higher consequence costs and stricter failure probability thresholds lead to earlier interventions, while weaker repair effectiveness causes the model to schedule more frequent maintenance actions. Higher operating pressure significantly increases failure risk and total cost. Under extremely high operating pressure, using replacement at the beginning provides greater long-term benefits than using composite wrap. This study provides a framework for data-driven, risk-based maintenance decision-making in complex pipeline systems.