Multi-defect reconstruction in nondestructive testing: an interpretable neural network approach

Guided wave tomography (GWT) methods for precise multi-defect reconstruction are crucial for structural health monitoring. In this work, an improved physics-informed wave tomography framework (PIWT) is proposed for the quantitative reconstruction of multiple defects in plates. A trunk-branch network is employed to reconstruct the wave travel time and velocity field by synergizing the waveguide governing equations and the real travel time data from sensors. This approach speeds up the network convergence of loss function which includes the travel time data, its first-order derivatives, and the physical principle of wave equations to constrain the space of parameters for accurate defect reconstruction. Based on simulation data, the results demonstrate that PIWT achieves the highly accurate defect with the errors of 4.25% in position and 5.5% in depth. Also, experimental validations are conducted to demonstrate the feasibility of PIWT with a defect position error of less than 1.7% and depth location error under 15%. Furthermore, uniform manifold approximation and projection is applied to enable a clear visualization of trajectories representing the defect reconstruction convergence, thereby revealing how incremental sensor data enhance the model’s capability to approximate the true solution. This interpretation provides useful insights into the latent dynamics to bridge the gap between the black-box nature of deep neural networks and the need for transparent and explainable AI, ultimately reinforcing confidence in the model's applicability for broader engineering applications.