Deconvolution of acoustic beamforming maps in interference environments with mean-reverting stochastic differential equations

Accurate identification of noise sources is the key technology for low noise design of products. Acoustic measurements are not always performed in the anechoic chamber due to practical constraints, and the signals measured by the microphone arrays may be contaminated by background interference. The results of acoustic imaging are blurred by background interference, making it difficult to identify the noise sources. A grid-based high-resolution acoustic deconvolution method via the score-based generative model (SGM) is proposed. The multi-step forward and reverse processes in the SGM are expected to alleviate the difficulty of predicting the real sound source distribution underlying the contaminated data through the neural network in a single step. In this research, the forward and reverse processes of the SGM following mean-reverting stochastic differential equations are utilized to relate the real source distribution to the output of conventional beamforming. The proposed deconvolution method is capable of accurately identifying the target sound source and suppressing the interference in numerical simulations of Gaussian and reverberant interference, as well as the closed test section wind tunnel experiment.