Ultrasonic guided wave imaging of pipelines based on physics embedded inversion neural network

Abstract Pipeline corrosion quantification plays a vital role in guaranteeing the safety of critical industrial structures and thus significant work has been carried out to address such an issue. Although quantitative imaging is crucial for non-destructive testing, research in guided wave pipeline testing has primarily centered on qualitative approaches. Here, we propose a deep neural network built upon physical model to reconstruct pipe wall thickness from ultrasonic guided wave (UGW) signals. The workflow of reconstruction contains three layers, where each layer consists of a fixed forward network and a residual inversion network. The forward model is represented by an agent convolutional neural network which would be embedded into the entire inversion network. The residuals between data from the forward model and real signals are then mapped into velocity profile differences through sub-inversion network. Numerical experiments were conducted to verify the inversion performance of the deep neural network using thickness maps obtained from guided wave frequency domain information. Results show that inversion images are capable to reveal the positions, shapes, and depths of corrosion with high resolution and precision, yielding an average inversion of 87.37% in the test set. In addition, by utilizing the periodicity of the pipeline, the inversion accuracy of eight pairs of transducers were improved from 67.7% to 89.43% with high-order helical guided wave. Compared with traditional high-precision inversion methods such as full waveform inversion, the proposed method achieved approximately 300 times faster inversion speed at the cost of some accuracy. The research demonstrates that real-time quantitative imaging of defects on pipes can be achieved accurately by physics embedded network. Furthermore, an experimental verification of the method was carried out through UGW pipeline testing, demonstrating its feasibility. The mean squared error of wall thickness reconstruction was 0.0070, achieving a high level of precision.