3D Temperature Field Reconstruction Using RBF Neural Network Incorporating Acoustic Physical Information

Abstract Temperature field measurement based on the acoustic method calculates temperature distributions of the measurement region using multi-path acoustic travel-time data. This paper presents an improved 3D temperature field acoustic reconstruction algorithm using a radial basis function neural network (RBFNN) that incorporates acoustic physical information. The reconstruction method adopts an RBFNN structure embedded with acoustic information, where the acoustic physical information includes the acoustic paths and acoustic velocity. By applying the actual measured sound wave propagation time as a data constraint and constructing a loss function, it achieves effective integration of acoustic physical information with experimental data (sound wave propagation time), thereby enhancing the accuracy of 3D temperature field reconstruction through acoustic tomography. In addition, the algorithm incorporates a particle swarm optimization (PSO) method with path-integral principal axis selection to account for the effects of acoustic refraction on temperature field reconstruction. Numerical results demonstrate that the proposed ray tracing algorithm reduces computation time and improves the accuracy in tracing acoustic rays. Furthermore, the reconstruction method significantly improves the accuracy of the temperature field reconstruction and enhances noise resistance compared to traditional acoustic reconstruction algorithms. Experimental results and uncertainty analysis show that this method has good reconstruction quality for the temperature field with an average relative error of less than 5% compared to the result measured by thermocouples, and further validates the practical feasibility of the proposed algorithm in engineering applications.