【摘要】 在水声通信中，传统的正交频分复用(Orthogonal Frequency Division Multiplexing, OFDM)系统需要大量导频数量和循环前缀以维持系统性能，严重浪费有限的频域资源。因此，利用深度学习辅助OFDM恢复失真的传输数据。具体地讲，利用改进的残差块提取和双向记忆接收信号。将自归一化网络(Self-Normalizing Network, SNN)与注意力机制结合，有效分配信道权重，以便系统更有效地利用信道资源，最小化信号失真。使用深度神经网络(Deep Neural Network, DNN)实现接收信号的分类，以准确地恢复接收信号。提出的ICANet(Improved Residual Network, Convolutional Neural Network and Attention Mechanism Self-Normalizing Network)模型可应用于由Bellhop软件生成的水声环境。仿真结果证明，与传统技术中的最小二乘法(Least Square, LS)以及现有的深度学习模型相比，所提出的模型在循环前缀受限的情况下可达到更低的误码率。更多还原

【Abstract】 In underwater acoustic communication, traditional orthogonal frequency division multiplexing(OFDM) systems require a large number of pilot symbols and cyclic prefixes to maintain system performance, resulting in significant wastage of limited frequency domain resources.Therefore, deep learning is utilized to assist in restoring distorted transmitted data in OFDM.Specifically, an improved residual block is used for feature extraction and bidirectional memory for receiving signals.Self-normalizing network(SNN) is combined with attention mechanism to effectively allocate channel weights, so that the system can more effectively utilize channel resources and minimize signal distortion.Deep neural network(DNN) is employed for the classification of received signals, ensuring accurate signal recovery.The proposed improved residual network, convolutional neural network and attention mechanism self-normalizing network(ICANet) model can be applied to underwater acoustic environments generated by Bellhop software.Simulation results demonstrate that compared to traditional techniques like least square(LS) and existing deep learning models, the proposed model achieves lower bit error rate, especially under limited cyclic prefixes.更多还原