Acoustic Temperature Tomography using a UNet based Deep Learning Approach.

The authors of this paper propose a deep learning-based tomographic reconstruction method. Time-of-flight measurements of acoustic waves passing through a region are utilized to reconstruct the temperature distribution within. While established methods based on a least-squares approach require a detailed and precise system model to achieve good results, the proposed method requires no modelling. It is created by customizing the UNet architecture, specifically its expanding path that turns the encoded information into an image representation. Furthermore, a synthetic dataset is created using finite element-based software for training and validating the deep learning models. The dataset contains acoustic time-of-flight measurements and the corresponding temperature distribution inside a finite area of interest. Our method can be easily adopted to a specific application. Compared to the popular Tikhonov least-squares method, the proposed method achieves lower average root-mean-square error when tested on our synthetic dataset, thus providing an alternative and improved way of approaching temperature tomography and inverse problems in general.