Abstract:

The high-quality measurement of temperature distribution is of great importance for industrial production. As a non-invasive measurement method, acoustic tomography (AT) is considered as a promising technique for the visualization of temperature distribution. To enhance the reconstruction quality, a two-stage high-resolution reconstruction algorithm was proposed for temperature field based on virtual observation (VO) and cascaded dense residual network (CDRNN). Firstly, the temperature field of coarse grid was obtained by the virtual observation algorithm, and then the CDRNN was built to predict the fine grid temperature distribution. The VO algorithm was used to achieve the overall least squares solution of the AT inverse problem, thereby reducing the reconstruction error caused by the bending of the acoustic path. Additionally, a dual-input compensation strategy was introduced to increase the utilization of the original measurement information by the CDRNN, and to improve the network stability. The network structure was streamlined by setting up sub-networks, dense connections and residual connections were also employed to improve network information flow. Sub-pixel convolutional layers were introduced to decrease network computing dimensions and boost reconstruction accuracy. Finally, the effectiveness of the algorithm was verified using a variety of simulated temperature field models. Through the numerical simulation of a typical temperature field model and comparison with the Landweber iterative method, ART algorithm, ART-NN algorithm, and VO algorithm, it was found that the average relative error and root mean square error of the VO-CDRNN algorithm were 0.44% and 0.68%, respectively, thus achieving better reconstruction results than other algorithms.

Key words: acoustic tomography, high-resolution reconstruction, virtual observation, residual connection, dense connection, sub-pixel convolutional layer