Beyond 10log10M array gain: A beamforming method under non-Gaussian noise and multi-sources

Abstract

Recently, deep learning techniques have been applied in the field of direction of arrival (DOA) to enhance beamforming in various practical applications, such as traffic positioning and hydro-acoustic detection. However, the effectiveness of these algorithms relies heavily on the availability of accurate training data that closely resembles real-world data, which can be challenging to obtain, particularly in hydro-acoustic environments. This is primarily due to the limited array gain (AG) and the constant fluctuations in ambient noise. Additionally, the presence of non-Gaussian noise in the background further complicates the process of identifying the desired signal, and it is worth noting that state-of-the-art deep learning methods have a limitation in detecting up to four sources. To address these challenges, we propose a novel method that employs a diagonal beam spectrum (DBS) feature to suppress non-Gaussian noise and enhance array gain. By incorporating DBS, our method surpasses the theoretical 10logM array gain and achieves an impressive 25logM array gain, while effectively detecting signals as low as -38dB (where M denotes the number of sensors). We also employ SA-U-NET, an image edge detection network based on DBS, to detect more than four sound sources and optimize angular positioning. The experiments focus on accuracy and array gain improvements of algorithms. Our proposed method achieves up to 91.67% correctness and 35dB array gain (i.e., 4 times better than the next best baseline) in experiments compared to other methods, demonstrating the potential of DBS and SA-U-NET.