Surrogate Model-Driven Deep Reinforcement Learning for Sensor Placement Optimization in Guided Wave Damage Localization

Abstract

Guided wave-based structural health monitoring (SHM) relies critically on sensor placement, as sensor layouts must provide sufficient information for subsequent damage inference. However, practical constraints on sensor quantity often degrade the spatial resolution of damage localization. Existing sensor placement optimization (SPO) methods typically maximize proxy metrics (e.g., coverage or information gain) without addressing the mismatch between sparse sensing and high spatial resolution inference of damage localization. To overcome this limitation, we propose a surrogate model-driven SPO framework that directly maximizes recoverable damage information from sparse measurements. A recoverability criterion is introduced to assess whether damage-induced reflections lie within the sensor observation window. A multiscale Fourier neural operator (FNO) is then developed to reconstruct full-field guided wave responses from multifrequency sparse signals. Based on this surrogate model, a reinforcement learning (RL) framework with an LSTM-based agent is employed to optimize sensor layouts, using reconstruction error as the reward. Experiments on the open guided wave (OGW) dataset demonstrate that the proposed method achieves high-resolution damage localization under significantly fewer sensor configurations. The results highlight the importance of information recoverability in sensor placement and provide a new paradigm for sparse sensing in guided wave SHM.