Session: Virtual Presentations in Acoustics, Vibration, and Phononics

Paper Number: 96652

96652 - Fast and Robust Damage Imaging With a Cascaded Deep Learning Technique 

Quantitatively measuring the health status of mechanical structures is a long-standing challenge in industrial fields. For plate structure inspection in tank and pressure vessels, traditional techniques using point by point scan with probes are tedious and time-consuming. Although recent progress using lamb waves provides an effective alternative, traditional algorithms like full waveform inversion (FWI) and diffraction tomography (DT) suffer from slow speed and convergence problems. In this article, we provide an effective damage imaging technique using dispersive A0 mode lamb guided wave and a deep learning inversion method combined with supervision (DLIS). The proposed algorithm uses convolutional neural network (CNN) as first iteration to provide a fast and low-resolution background, and further optimizes the inversion results with descent direction matrix. In this way, the nonlinearity of the problem is effectively decomposed and yields better results. As a first test of the method, we generated 1000 training samples consists of corrosion defects with various sizes and shapes using 2D acoustic wave modeling. The output signals are converted into frequency domain and tagged with true velocity maps as labels for training of CNN model and descent direction matrix. In the testing process, first iteration of CNN model outputs a background model with the size, thickness distributions and locations of defects roughly match those of the true model. Then, in the second iteration descent direction matrix aids the algorithm to optimize the starting model and produce the finer details of the defect. Inversion results prove that DLIS have the full capability to handle single or multiple corrosion defects, and even using 2D acoustic equations for forward modeling, a small flat bottom circular pit with its diameter being 24 mm (λ/1.41) and depth of 2 mm is successfully reconstructed. In the second test, we varied the number of training samples and test the inversion accuracy of DLIS with small training sets. The results show that with 400 training samples the shape of defects can be roughly formed, and using a minimum of 800 training samples accurate inversion results can be effectively obtained. In the discussion part, the robustness of DLIS against noise is investigated. From the aforementioned perspective, we completely prove that DLIS method has the full capability to be applied in real life conditions. In short, the proposed method is a promising way not only in inspecting plate-like structures, but also possesses tremendous speed advantage over traditional methods and has the full potential to be used in the fast inspection of plates made of composite materials and pipes, geophysical prospecting and medical imaging because all these inverse problems share similar physics.

Presenting Author: Junkai Tong Tianjin University

Presenting Author Biography: Junkai Tong earned his Bachelor’s degree and Master’s degree in applied physics and condensed matter physics at School of Science, Tianjin University. He is currently continuing his research as a PhD student at State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University. His interests include development of quantitative NDT imaging algorithms with deep learning. He also has interests in quantitative ultrasonic diagnosis in biomedical imaging.

Authors:

Junkai Tong Tianjin University
Min Lin Department of Mechanical Engineering, University of Wyoming
Jian Li State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University
Xiaocen Wang State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University
Guoan Chu State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University
Yang Liu State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University