Direction of arrival estimation using a horizontal array with beamforming domain feature fusion in shallow water

A convolutional neural network method for direction of arrival estimation based on beam domain feature fusion of the horizontal line array is proposed to solve the problem of performance degradation of direction of arrival in marine environment. Firstly, a data computational simulation model is built for generating the dataset, and beam-domain data are obtained using both conventional beamforming and minimum variance distortionless response beam formers. Then, the convolutional neural network models are trained separately using the two beam-domain data. Both simulation and sea trial data processing results confirm that the direction of arrival estimation accuracy of the trained models outperforms that of conventional beamforming, minimum variance distortionless response and matched beam processing, especially in the end-fire direction where the performance improvement is obvious. Finally, the two beam-domains features are used to train the convolutional neural network by early-fusion. The sea trial data test results confirm that the direction of arrival estimation performance of the feature-fused convolutional neural network is further improved. The probability of reliable direction estimation within 5° error range is about 4% higher than that the single-feature convolutional neural network estimation result, and the root mean square error is reduced by about 0.2°.