Full Waveform Inversion Guided Wave Tomography Based on Recurrent Neural Network

Abstract

Corrosion quantitative detection of plate or plate-like structures is a critical and challenging topic in industrial Non-Destructive Testing (NDT) research which determines the remaining life of material. Compared with other methods (X-ray, magnetic powder, eddy current), ultrasonic guided wave tomography has the advantages of non-invasiveness, high efficiency, high precision and low cost. Among various ultrasonic guided wave tomography algorithms, travel time or diffraction algorithms can be used to reconstruct defect or corrosion model, but the accuracy is low and heavily influenced by the noise. Full Waveform Inversion (FWI) can build accurate reconstructions of physical properties in plate structures, however, it requires a relatively accurate initial model, and there is still room for improvement in the convergence speed, imaging resolution and robustness. This thesis starting with the physical principle of ultrasonic guided waves, the dispersion characteristic curve of the guided wave propagating in the plate structure converts the change of the remaining thickness of the plate structure material into the wave velocity variation when the ultrasonic guided wave propagates in it, and provides a physical principle for obtaining the thickness distribution map from the velocity reconstruction. Secondly, a guided wave tomography method based on Recurrent Neural Network Full Waveform Inversion (RNN-FWI) is proposed. Finally, the efficiency of the above method is verified through practical experiments. The main work of the thesis includes:  The feasibility of conventional full waveform inversion for guided wave tomography is introduced and verified.  An FWI algorithm based on RNN is proposed. In the framework of RNN-FWI, the effects of different optimization algorithms on imaging performance and the effects of different sensor numbers and positions on imaging performance are analyzed.  The quadratic Wasserstein distance is used as the objective equation to further reduce the dependence on the initial model. The depth image prior (DIP) based on convolutional neural network (CNN) is used as the regularization method to further improve the conventional FWI algorithm, and the effectiveness of the improved algorithm is verified by simulation and actual experiments.