基于自编码器的水声非相干恒重映射优化

Optimization of non-coherent underwater acoustic constant weight mapping based on auto-encoder

针对水声多径衰落非相干信道,提出一种灵活的恒重自编码方案以解决传统恒重映射(如Hadamard映射)权重单一,频谱利用率低的问题.将水声多径信道建模为泛化性更高的Rice衰落,利用深度神经网络(DNN)合理构建自编码器(AE),在多个衰落因子下进行输入到输出的端到端训练,在发射平均功率不变的条件下,得到权重分布更优的映射集.发射符号幅度不局限于‘0’和‘1’的分布集合,但分组仍保持恒重.接收端仍可采用能量检测,不依赖信道状态信息(CSI)、鲁棒性强.为扩展到高阶情况,提出一种低复杂度的网络输入模式使AE复杂度大大降低,加速训练收敛.仿真证明,在接收端CSI未知的情况下,所提方案在标准衰落信道及Bellhop模拟水声信道上均取得相较传统方案5–6 dB的性能增益.最后,采用通信带宽4 kHz,通信距离2500 m,海底深度1750 m的南海海试数据对结果进行了验证.

For underwater acoustic multipath-fading noncoherent channels, a fiexible constant weight automatic coding scheme is proposed to solve the problems of limited weight-distribution and low spectral-efficiency of traditional constant weight mappings(e.g., Hadamard mappings). The underwater acoustic multipath channel is modeled as the Rice fading,which is more generalized than the Rayleigh one. The auto-encoder(AE) is constructed by using the deep neural networks(DNN) and the end-to-end input-to-output training is carried out under several fading factors. Under the fixed average transmission power, the mapping set with better weight-distribution is obtained. The transmitted symbol amplitude is not limited to the ‘0’ or ‘1’, but the groups still maintain constant weight. The energy detection can still be used at the receiver side, which is robust without the channel state information(CSI). To extend to the high-order cases, a low complexity network-input mode is proposed to reduce the complexity of the proposed AE and accelerate the training convergence.With unknown CSI on the receiver side, the simulation results show that the proposed scheme achieves a performance gain of 5–6 dB over the standard fading channels and the Bellhop simulated underwater acoustic channel. The results are also verified by sea trial data collected in the South China Sea with seafioor depth of 1750 m, the communication bandwidth is 4 kHz and the communication distance is 2500 m.