DEEP LEARNING FOR INTERFACIAL DAMAGE ESTIMATION IN AN INVERSE ULTRASOUND SCATTERING ANALYSIS

Resumo

We formulate and solve a time-harmonic inverse scattering problem to estimate the interfacial defect
distribution at an adhesion interface of a composite plate. We use the incident field that mostly interact with such
defects and assume prior knowledge of the material properties of each layer of the laminate. We model the adhesion
interfaces using the Quasi-Static-Approximation, where approximates it by a set of tangential and normal springs,
and allow the interfacial stiffness to depend upon the position along the interface. To solve the direct problem, the
interfacial defect distribution is generated by a prior of smooth stochastic field. In addition, we develop a deep
learning neural network, using the reflected signal as input, to solve the formulated inverse problem. We validate
our implementation and evaluate the presented method’s performance for noisy input data and different defect
distribution scenarios with the aid of numerical simulations. From the obtained numerical results, we may say
that the proposed method is robust to the presence of noise and has the potential to detect and classify interfacial
defects.