Investigation of Machine Learning-Based Acoustic 2D Gas Pyrometer

• Published in: 2022 IEEE International Ultrasonics Symposium (IUS) pp. 1 - 3
• Main Authors: Lin, Fu-Sung, Huang, Man-Ching, Wu, Chia-Hsi, Lin, Jia-Ling, Li, Chih-Ying, Lin, Hsiao-Chi, Pal, Soma, Huang, Chih-Hsien
• Format: Conference Proceeding
• Language: English
• Published: IEEE 10.10.2022
• Subjects: Acoustic Tomography; Algorithm; Convolution; Data visualization; Deep learning; Machine learning; Neural networks; Reconstruction algorithms; Temperature Distribution; Temperature measurement; Training
• ISSN: 1948-5727
• DOI: 10.1109/IUS54386.2022.9957880

Acoustic tomography (AT) is considered a promising visualization technique for gas temperature distribution (TD) in various industrial applications. The gas TD of the region of interest (ROI) can be reconstructed according to the sound velocities of multiple routes between transmitters and receivers. To improve the accuracy of the reconstruction, increasing the number of iterations is inevitable with the cost of being time-consuming. Recently, serval machine learning techniques are used to speed up the imaging reconstruction process, such as X-ray, CT, or MRI. However, to our knowledge, using machine learning on reconstructing gas TD from acoustic velocities is rarely reported. Therefore, in this research, a convolution neural network (CNN) is developed to examine the capability of reconstructing gas TD with a machine learning approach. Two models would be trained using two different datasets. One model was trained with ideal TD, the other one was trained with reconstructed TD. Besides, acoustic velocities from two gas TDs were applied to evaluate the 2D visualization performance. One TD was similar to the training data and the other were different. The results indicated that the model trained with ideal gas TD could track the hot spot more closely. However, the 2D visualization results using the model trained with ideal gas TD were poor compared to the other one trained with reconstructed TD when the input acoustic velocities of the TD with two hot spots. It indicated that the proposed method could successfully learn the relationship between TD and acoustic velocities from an ordinary reconstruction algorithm. Besides, the execution time of the proposed model was 0.109s, which is 96% less than the selected iterative reconstruction method. Consequently, the proposed neural networks should be considered as a reliable and efficient 2D gas TD reconstruction methodology.