Adaptive Risk-Based Life-Cycle Management for Large-Scale Structures Using Deep Reinforcement Learning and Surrogate Modeling

Abstract

Optimal life-cycle management is a challenging task for large-scale structures. The complexity of structural states, represented by the numerous combinations of component conditions, and the vast number of inspection and maintenance options often prompt the decision-makers to adopt a simple time- or condition-based management method rather than a performance-based one. To improve this situation, this study proposes a novel method for adaptive risk-based life-cycle management of large-scale structures. The proposed method can yield bespoke inspection and maintenance plans at the individual component level based on their contribution to the overall structural performance. The obtained plan can also adapt itself to the unfolding information gained from inspection and maintenance actions. This advanced method, termed DeepLCM, is enabled by (1) efficient surrogate modeling based on deep neural networks for structural risk assessment; and (2) a deep reinforcement learning algorithm for adaptive life-cycle management. The method is applied to a steel girder bridge in Montgomery County, Pennsylvania. The inspection and maintenance plan obtained using DeepLCM is compared with those obtained using the conventional life-cycle management techniques including time-, condition-, and risk-based methods. The case study also investigates the effect of the spatial granularity of inspection and maintenance actions on the resulting life-cycle cost.