Reconstructing 2-D Gas Temperature Distribution With Deep Neural Networks

Abstract

Acoustic tomography (AT) is considered a promising visualization technique for gas temperature distribution (TD). Generally, a temperature map of a region of interest (ROI) is reconstructed with acoustic velocities of multiple routes inside the ROI using an AT reconstruction algorithm. To improve the accuracy of the reconstruction, increasing the number of iterations is inevitable, which may be time-consuming. Besides, existing reconstruction algorithms rarely consider utilizing practical measurements to improve their performance. In this study, a convolutional neural network (CNN) is proposed to train a machine-learning model that can reconstruct a gas TD from acoustic velocities. By adjusting the label and the loss function, a practical training approach was found. Six models were successfully trained to reconstruct a gas TD. Two were trained with the ideal TD, three were trained with a reconstructed TD, and one was trained with temperatures from specific locations. After that, a TD with peak temperatures located in 13 different positions was applied to test the models’ performance in tracking the hot spot. The results were compared with an existing reconstruction method. Besides, the acoustic velocities from three gas TDs were applied to evaluate the 2-D visualization performance. One of them was similar to the training data and the other two were different. The results indicated that the models trained with the ideal gas TD could track the hot spot more closely, and the models trained with reconstructed gas TDs had similar performance as the selected reconstruction algorithm. However, the 2-D visualization results using models trained with the ideal gas TD were poor compared with models trained with reconstructed TDs when the input acoustic velocities were different from the training data. This indicated that the proposed method could successfully learn the relationship between TD and acoustic velocities from an ordinary reconstruction algorithm. Furthermore, the execution time of the proposed model was 0.109 s, which is 96% less than the selected iterative reconstruction method. Consequently, the proposed neural networks should be considered a reliable and efficient 2-D gas TD reconstruction methodology.