Neural networks for 3D temperature field reconstruction via acoustic signals

Abstract Reconstructed 3D temperature field will provide critical input for the control mechanisms to optimize the thermal fluids and combustion process. In this paper, a distributed optical fiber sensing system is used to generate acoustic signals for real-time monitoring and optimization of spatial and temporal distributions of high temperature profile in a boiler furnace in fossil power plants. A code division multiple access (CDMA) based acoustic signal modulation technique for improved signal to noise ratio (SNR) and simultaneous sending/receiving is developed. A kernel regression model which approximates the temperature field as a finite summation of products of space-dependent Gaussian Radial Basis Functions (GRBF) and time-dependent coefficients is established. The inversion problem to estimate the best parameters of Gaussian functions is solved by optimizing a cost function using gradient descent method. Guidance on how to tune design parameters is also given. And regularization is applied for solving the trade-off problem between bias and variance. The numerical simulations show an approximation error less than 5% in 3D temperature field reconstruction. Besides that, the performance of learned model with variation of some relevant design parameters is evaluated, and error analysis for temperature field reconstruction with measurement noise is also given. To validate the availability and efficiency of our proposed 3D temperature field reconstruction mechanism, a 2D temperature field distribution experiment test on microphone is carried out and satisfactory estimation accuracy is achieved.