Non-destructive testing of steel-lined concrete structure using multiple agents with deep prior.

Non-destructive testing is a cornerstone of structural integrity assessment that enables internal evaluation of materials without inflicting damage. Among various imaging methods, Ultrasonic Model-Based Iterative Reconstruction (UMBIR) has gained attention for its ability to enhance ultrasonic imaging by incorporating physical and statistical priors. Notably, its extension, Multi-Frequency Ultrasonic Model-Based Iterative Reconstruction (MF-UMBIR) to process multi-frequency datasets have shown improved accuracy over traditional techniques. However, its performance has some limitations in highly complex structures. This paper presents Deep Multi-Agent Consensus Equilibrium (Deep-MACE) method, a novel reconstruction framework that integrates multi-frequency forward model agents with a deep learning prior using the consensus equilibrium formulation. By combining data consistency across different excitation frequencies with the expressive power of a learned U-Net prior, Deep-MACE enables high-fidelity imaging in acoustically heterogeneous environments. The primary objective is to reconstruct images that reveal structural damage, such as corroded rebars and delaminations, thereby supporting non-invasive Structural Health Monitoring (SHM) of steel-lined concrete structures. Experimental results demonstrate that both UMBIR and MF-UMBIR suffer from limitations in defect visibility and robustness when applied to steel-lined concrete structures. In contrast, Deep-MACE consistently produces clearer reconstructions, successfully identifying all internal rebars with fewer artifacts and improved spatial resolution. These results highlight the potential of integrating deep priors into multi-agent frameworks for advanced ultrasonic non-destructive testing.