Session: Virtual Presentations in Acoustics, Vibration, and Phononics

Paper Number: 96700

96700 - Sparse Data Recovery Algorithm Based on BP Neural Network for Ultrasonic Guided Wave Imaging 

Ultrasonic guided wave detection technology plays an important role in non-destructive testing, widely used in the detection of pipelines, rails and plate-like structures. Compared with the traditional ultrasonic point-by-point scanning method, ultrasonic guided wave detection technology breaks through the limitations of point-by-point inspection, which is fast and comprehensive and suitable for the detection of large structures. Combined with a series of imaging methods, the spatial maps can be reconstructed, such as velocity maps, density maps, etc. The health condition and the service life of material can be evaluated according to the reconstructed spatial maps. Among them, the mainstream ultrasonic imaging methods include reverse-time migration method, wave-field-based imaging methods and some other tomography imaging methods. However, the speed of imaging is limited by the large amount of calculation and the imaging quality of current imaging methods is restricted by the number of testing transducers in service. In this topic, a sparse ultrasonic guided wave imaging method based on deep learning is proposed to quantitatively evaluate the damage degree of plate-like structure. The sparse ultrasonic guided wave imaging method contains data recovery and velocity map of damage reconstruction. In the process of data recovery, sparse guided wave-field data can be restored to dense guided wave-field data and then the dense guided wave-field data is used for deep learning inversion. The relationship between the recovered wave-field data and true velocity maps can be established by offline training. Afterwards, online testing is carried out, which means the relationship of wave-field/velocity map can be directly invoked for velocity map reconstruction. The velocity map should be converted to the remaining thickness according to the dispersion curve. The highlight of this imaging method and the significance lies in the fact that it breaks through the quantitative detection of the structure technology bottleneck which cannot be achieved by traditional detection technology, that is, the dispersion characteristics of guided wave is used to detect the velocity distribution on the surface of the plate-like structure effective reconstruction, and the dispersion curve is used to convert velocity distribution to thickness distribution so as to realize the quantitative detection of remaining plate-like structure. The guided wave imaging method based on deep learning not only has a great improvement in imaging quality compared with traditional imaging methods, but the calculation of Hessian matrix by traditional algorithm is avoided and the computational resources are greatly saved. This method is promising because the number of transducers is greatly reduced without sacrificing image quality in the hardware system, and the difficulty of data acquisition, data transmission, and data storage is also reduced.

Presenting Author: Xiaocen Wang Tianjin University

Presenting Author Biography: Xiaocen Wang is currently pursuing her PhD degree in mechanical specialty from Tianjin University, China. Her research interests include NDT based on ultrasonics and deep learning.

Authors:

Xiaocen Wang Tianjin University
Min Lin University of Wyoming
Jian Li Tianjin University
Dingpeng Wang Tianjin University
Yang Liu Tianjin University