基于深度神经网络的水声信号恢复方法研究*

针对干扰或噪声环境下水声目标信号难以获取的问题,该文提出研究基于深度神经网络的自适应水声被动信号波形恢复方法。在单阵元情况下,该方法提取对数功率谱特征作为输入,采用深度神经网络回归模型自适应学习目标信号的自身特征,输出降噪后的对数功率谱特征并还原时域波形。在多阵元情况下,提出阵列深度神经网络降噪方法,将部分或全部阵元特征拼接为长向量作为输入,从而利用空域信息。为全面利用阵列丰富的时频域信息,该文提出一种两阶段特征融合深度神经网络,在第一阶段将阵列分为若干个子阵,将每个子阵分别用阵列深度神经网络进行处理,在第二阶段将第一阶段的各子阵处理结果与阵列接收信号同时输入一个深度神经网络进行融合学习。实验表明,所提出的单阵元和两阶段融合深度神经网络取得了显著优于常规波束形成的恢复结果,能够准确估计目标信号波形和功率并显著提高输出信噪比。

Underwater acoustic target waveform recovery based on deep neural networks

A deep neural network (DNN) based method is proposed to recover underwater acoustic target signal waveform under noise. In single sensor condition, the log-power spectral (LPS) feature is extracted as input, and a DNN regression model is employed to adaptively learn the inherent pattern of target signal and output the enhanced LPS to recover the waveform. In multi-sensor condition, an array-based DNN which uses the concatenated feature from partial or all sensors as input is proposed to exploit spatial information. To fully use the rich temporal and spatial information from the array, we propose a two-stage DNN. In the first stage, the array is split into sub-arrays and each sub-array is processed by an array-based DNN, while in the second stage, the sub-array enhanced features and noisy array features are input to a DNN for integration. Experiments show that our single-sensor and two-stage DNN achieved far better recovery results than conventional beamforming, can accurately recover the target waveform and power and significantly improve the output signal-to-noise ratio.