增强水声网络数据收发稳健性的统计模型与深度学习方法研究

To enhance the robustness of the data transmitting and receiving of the underwater acoustic network nodes which is often disturbed by the widespread non Gaussian distribution channel and noises in the complicated marine environment. Based on the practical needs for the channel equalizing and denoising, this application presents a soft time-frequency mask denoising statistical model, a channel equalizing statistical model and deep learning methodology that are applicable to underwater acoustic communication. The denoising and channel equalizing are considered as a soft classification where a statistical model is used to estimate the probability of whether the current measurement belong to noises or some transmitting symbol. Considering the limited power consumption and computation resource of the nodes of the underwater acoustic network, this paper attempts to explore deep learning methods for the above statistical models and network node treatment, and figure out a soft time-frequency mask denoising and channel equalizing solution based on deep learning. Since the classic Gaussian distribution assumption can lead to performance degradation in complicated marine environment due to the model mismatch, the application aims to solve the problem with deep learning, which has the advantage of fitting any complicated probability distribution model, and to reinforce the robustness of data receiving and transmitting of underwater acoustic networks in the complicated marine environment. This application is a cross-discipline frontier research combining deep learning and underwater acoustic signal processing, which has significance for the development of China’s underwater acoustic networks and underwater acoustic communication both theoretically and practically.

针对复杂海洋环境下广泛存在的非高斯分布信道与噪声影响水声网络节点数据收发稳健性的问题，密切结合水声网络对信道均衡及降噪的实际需求，提出适用于水声通信的软时频遮罩降噪统计模型、信道均衡统计模型及相应的深度学习方法。将降噪与信道均衡问题视为一个基于训练数据拟合某一未知概率分布、并使用拟合后的分布估计当前观测样本属于噪声或某一发送符号概率的软分类问题。针对水声网络节点功耗及计算资源受限的特点，探索适合于前述统计模型与节点处理的深度学习方法，构建基于深度学习的软时频遮罩降噪与信道均衡方法。旨在以深度学习可拟合任意概率分布的优势，突破经典高斯分布假设在复杂海洋环境下因模型失配而导致的性能下降问题，增强水声网络在复杂海洋环境下的数据收发稳健性。本项目属于深度学习与水声信号处理的交叉前沿研究，其研究成果对我国水声网络及水声通信技术的发展有着重要的理论意义与应用价值。

针对复杂海洋环境下水声网络节点数据收发不稳定的问题，项目组建立了适用于非相干水声通信的软时频遮罩降噪系统框架，基于稀疏信号的W-分离正交性，提出了基于软时频遮罩的水声MFSK 通信信号增强方法，构建了用于水声通信的软时频遮罩降噪统计模型，完成任务书规定研究内容的基础上，延伸开展了基于软时频遮罩降噪的MFSK 带内全双工（IBDF）通信方法研究，给出了系统结构，并通过仿真给出了系统在最优时频遮罩估计条件下的最优性能。项目组完成了基于解码误码率指标的降噪性能评估方法—评估了调制阶数、时频分析参数、用户数等参数对软时频遮罩降噪的影响，给出了结果趋势。.针对相干水声通信的高速率要求，我们在前馈神经网络均衡器的基础上引入了判决反馈结构，以Tanh 函数作为神经元节点激活函数，构建了一种判决反馈神经网络自适应均衡器，并基于实测典型多径传播水声信道对该均衡器性能进行了仿真，提高了均衡器收敛精度。为了有效应对水声信道的时变、空变特性，项目组在引入反馈机制的基础上，使用具有时间序列记忆能力的Elman神经网络构建了深度Elman 神经网络自适应均衡器，采用了15-5-3-2-1结构，并将神经元激活函数替换为更优的ReLu函数，大大降低了均衡器的迭代次数，减少网络训练时间，更适于功耗及计算资源受限的水下声网络节点使用。.最终，项目组，结合本单位型号任务研制的海上实验机会，于2019 年11 月赴海南三亚进行了海上试验，得到了海上实录数据，验证了算法有效性，达到了预期目标。