Underwater acoustic channel estimation method based on response generative network

To address the issue of low channel estimation accuracy in underwater acoustic communication environments, an underwater acoustic channel estimation method based on the response generative network (RGNet) was proposed. First, an underwater acoustic impulse response model based on MIMO-OTFS channel characteristics was established. On this basis, a time-delay-compensated localized analysis algorithm was proposed for 3D reconstruction of underwater signals to reduce feature errors caused by the dynamic changes of the channels. Considering OTFS time-delay-Doppler matrix’s susceptibility to noise pollution, which leads to matrix expansion and distortion, a self-iterative signal updating network was proposed. The input signal of the generative network was updated by a weight and bias update algorithm based on L1-regularized least squares. Finally, to address the instability of traditional deep learning models for underwater acoustic signals, a decomposed optimization algorithm based on Bures-Wasserstein objective function was proposed. The convergence process of the objective function was decomposed into several sub-problems to improve the model’s convergence speed and reduce errors. Experimental results show that the proposed method achieves a normalized mean square error of 0.04 for channel estimation at a signal-to-noise ratio of 5 dB. Moreover, compared with other methods, the proposed method achieves lower signal peak blur and less signal distortion in the time-delay and Doppler response heat maps.