A numerically-enhanced machine learning approach to damage diagnosis using a Lamb wave sensing network

Abstract

Abstract This paper describes a methodology for the design of a model-based diagnostic unit. The objective of the work is to define a suitable procedure for the design and verification of diagnostic performance in a simulated environment, trying to maximise the generalisation capability of pattern recognition algorithms when tested with real experimental signals. The system is designed and experimentally verified to solve the fatigue crack damage localisation and assessment problems in a realistic, though rather idealised, Structural Health Monitoring (SHM) framework. The study is applied to a piezoelectric Lamb wave sensor network and is validated experimentally on a simple aluminium skin. The analytically-derived dispersion curves for Lamb wave propagation in aluminium are used in order to determine the wave velocities and thus their arrival time at given sensors. The Local Interaction Simulation Approach (LISA) is used to simulate the entire waveform propagation. Once the agreement between analytical, numerical and experimental data is verified on a baseline undamaged condition, the parametric LISA model has been iteratively run, varying the position and the length of a crack on an aluminium skin panel, generating the virtual experience necessary to train a supervised learning regressor based on Artificial Neural Networks (ANNs). After the algorithm structure has been statistically optimised, the network sensitivity to input variations has been evaluated on simulated signals through a technique inspired by information gap theory. Real Lamb wave signals are then processed into the algorithm, providing feasible real-time indication of damage characteristics.