A hybrid physics-assisted machine-learning-based damage detection using Lamb wave

Abstract

This research presents a hybrid physics-aided multi-layer feed forward neural network (MLFFNN) model to improve damage detection under Lamb wave responses. Here, a damage parameter database (DPD) is created from the complex responses of a thin aluminum plate generated using finite-element (FE) simulations. A double pulse-echo transducer configuration is implemented over the 1.6 mm thick aluminum plate with notch-like defect, which generates only A $$_{0}$$ mode in the plate structure and records damage-specific S $$_{0}$$ mode. Sixty-six FE simulations are conducted, each representing a distinct damage scenario in terms of damage location and Lamb wave frequency. Artificial noise is added to compensate environmental interference. Orthogonal matching pursuit was performed to improve the sparsity of the signal. Thereafter, the damage-specific features are extracted from the sparsed S $$_{0}$$ signal to construct DPD for all 66 FE simulations. The fully developed DPD is deployed to train an MLFFNN supervised by a robust Levenberg–Marquardt algorithm. A set of initial tests are conducted for higher damage-depth to plate-thickness ratio with 1.0 mm notch depth, and the fully trained MLFFNN predicts the damage location with 99.94% accuracy. The proposed algorithm achieves a good level of generalization, including the cases of overlapping echoes and cluttered responses due to multiple reflections for the given damage scenarios.