TFF-CNN: time-frequency fusion-based CNN for DOA estimation of underwater acoustic array signal

Abstract

In the Marine environment, underwater acoustic positioning signals are susceptible to complex underwater acoustic channels and multiple types of underwater noise, resulting in inaccurate DOA estimation, which will significantly reduce the accuracy and reliability of underwater positioning, and the real-time performance of underwater positioning cannot be guaranteed. In this paper, a Time-Frequency fusion based convolutional neural network (TFF-CNN) method for underwater DOA estimation is proposed. In this method, fusion feature maps are constructed based on Time-Frequency characteristics of signals, and a novel convolutional neural network with the deconvolution layer is designed to enhance feature representation and noise resistance. The experimental results show that compared with traditional methods and other networks, the proposed algorithm has great advantages in the estimation accuracy and speed, especially under non-Gaussian noise.