An Underwater Array Localization Method Using Two-Stage Learning Model

  Objectives  In the localization method based on underwater acoustic arrays, noise affects the localization accuracy mainly. A localization method using a two-stage learning model is proposed to minimize noise's influence on the localization results.

  Methods  Firstly, a learning model based on a multilayer convolutional neural network is built for noise reduction by training the real and imaginary features of the received signal separately. Secondly, an improved weighted delay summation beamformer group model is constructed, and the gradient descent criterion adjusts the weights of each channel to obtain the optimal relative time delay and the best angle estimation. Finally, more accurate localization information is received by geometric solving.

  Results  Simulation experimental results show that the two-stage noise reduction model has awe-inspiring noise reduction performance compared to conventional underwater acoustic array processing for angular comparisons of 30°, 100°, and 130° in a SNR (signal-to-noise ratio) environment in the range of -25 dB to 10 dB. Lake tests show the same advantage of the proposed method in terms of RMSE (root mean square error) of positioning results.

  Conclusions  The two-stage learning model has excellent robustness in coping with SNR variations, and the received signal after processing by noise reduction has higher localization accuracy after processing based on an improved time-delay and angle estimation model, which can also be applied in practical applications such as indoor acoustic localization and radio radar arrays.