Machine learning-based orthotropic stiffness identification using guided wavefield data

The characterization of the full set of elastic parameters for an orthotropic material is a complex non-linear inversion problem that requires sophisticated optimization algorithms and forward models with thousands of iterations. The intricacy of this type of inversion procedure limits the possibility of using these algorithms for large-scale automation and real-time structural health monitoring. At this point, the introduction of machine learning-based inversion strategies might become helpful to overcome the existing limitations of conventional inversion algorithms. In the present study, a multilayer perceptron algorithm is used to identify elastic stiffness parameters of orthotropic plates using guided wavefield data. A large and diverse training dataset is created by using a semi-analytical finite element model, and the effect of both the training dataset size and the signal-to-noise ratio on the inference outcome are examined. The performance of the multilayer perceptron-based inversion method is first validated on a numerical dataset, and the method is then further applied on experimental data obtained from a multilayered glass-fiber reinforced polyamide 6 composite plate. Finally, the multilayer perceptron-based inference results are compared with the outcome of a traditional inversion algorithm, showing a difference of less than 0.5%.