Development of reconstruction method of void distribution for ultrasonic tomography using a machine learning algorithm

Kojima, Yuya*; Murakawa, Hideki*; Sugimoto, Katsumi*; Kondo, Teppei*; Abe, Yuta  ; Aizawa, Kosuke 

This study proposes a novel void distribution reconstruction method utilizing a machine learning (ML) algorithm in ultrasonic UT) to improve the accuracy of void distribution reconstruction. A simple neural network trained with 1920 simulated single spherical bubble data with a diameter of 5 mm achieved high accuracy (correlation coefficients) when predicting the simulated waveforms. However, the direct application of the method to the experimental data yielded poor results owing to the inherent noise in the real experimental data. To address this challenge, two data techniques were introduced: applying the maximum value within a moving window and subtracting the received waveforms from those obtained without bubbles. These preprocessing methods enabled the prediction of the void distribution obtained from the experimental data.