Physics-Informed Machine Learning for Ultrasonic Guided Wave Full-Field Reconstruction and Damage Diagnosis

Abstract

Ultrasonic guided waves have shown significant efficacy in structural health monitoring due to their increased sensitivity to structural changes in mechanical properties. Particularly, they are advantageous for the detection, localization, and quantification of damage in thin-walled structures. Traditional diagnostic methods that utilize ultrasonic guided waves often depend on signal pre-processing to extract pertinent features, which may inadvertently lead to the loss of crucial information. Recent advancements have seen the incorporation of neural networks to circumvent pre-processing and allow for the direct analysis of complex signals. However, the blackbox nature of neural networks, their dependence on extensive labeled data sets, and limitations in generalizing beyond their training domain present considerable challenges. This study introduces a novel unsupervised physics-informed machine learning methodology aimed at overcoming these obstacles. The damage diagnosis problem is framed as an inverse problem, specifically a full-waveform inversion task, with the objective of reconstructing the material distribution throughout the structure to identify potential discontinuities caused by damage. The methodology combines neural networks with a custom finite difference solver capable of efficiently simulating wave propagation, embedding the physics of the problem into the solution. The proposed methodology was validated against a case study of a metal plate affected by a single damage. Despite the computational challenges posed by the three-dimensional nature of the problem, the findings indicate that the approach effectively localizes and quantifies damage within the investigated structure.