Designing of two-dimensional acoustic beamforming array using machine learning

Abstract

Acoustic beamforming arrays are used for locating sound sources and their performance is measured by their directivity pattern. Designing an acoustic beamforming array to meet specific performance requirements can be challenging and requires numerical optimization of various factors such as array width, microphone positions, and the number of microphones. This study employed a Deep Neural Network (DNN) approach to predict the arrangement of a two-dimensional acoustic beamforming array giving the desired beam pattern. The DNN was trained on 90,000 microphone array designs and beam patterns of 1-kHz sound. The prediction accuracy was evaluated based on the mean absolute errors (MAE) of beam patterns, frequency-averaged main lobe width (MLW), and maximum sidelobe level (MSL). The results for commonly designed arrays datasets such as Brüel & Kjær style array had a 26.54% error in beam patterns, 2.34 degree/m error in MLW, and 4.21 dB error in MSL. However, for predicted arrays generated by the random array designs, the results were more accurate, with deviations of 22.08% in beam patterns, 1.82 degree/m error in MLW, and 5.07 dB in MSL.