Research on the multi-signal DOA estimation based on ResNet with the attention module combined with beamforming (RAB-DOA)

Abstract

Direction of Arrival (DOA) estimation based on deep neural networks has been extensively studied recently, but multi-signal DOA estimation has not been sufficiently investigated. The strong mutual interference between signals emitted by multiple sources in different directions in multiple DOA leads to the reduction of detection accuracy, which limits the application in multi-object scenarios. In multi-signal DOA estimation, a residual network (ResNet) incorporating efficient channel attention module could significantly enhance the signal separation and localisation capabilities of the system. Therefore, this paper presents a multi-signal DOA estimation system based on ResNet with the attention module and beamforming (RAB-DOA). The system receives spatial signals through an array of detectors and uses a linear constrained minimum variance (LCMV) beamforming algorithm to optimize signal directivity and suppress interference. Phase adjustment is then performed during the scanning process to enhance the signal in the scanning direction and suppress interfering signals in other directions. Finally, the signals are binary classified using ResNet with an efficient channel attention module to obtain multi-signal DOA estimation results. Experiment results show that the detection accuracy and precision of the proposed algorithm are excellent, especially at low SNRs in spite of multiple interfering signals.