| 摘要: |
| 针对现有端到端神经网络通信系统的泛化能力改进及自编码器优化等问题，提出了一种基于卷积神经网络（Convolutional Neural Network，CNN）的端到端通信系统改进方案。该方案在自编码器结构中引入一维卷积层（Conv1D），通过对参数的重新设计，优化了网络性能。发送端采用多层Conv1D对输入序列进行特征提取，通过训练获得发送信号的最佳调制和编码方案；接收端同样采用多层Conv1D，来恢复受到噪声污染的符号。仿真实验表明，在不同输入比特长度及编码速率条件下，所提系统表现出了良好的泛化能力。并且，在加性高斯白噪声信道和瑞利衰落信道下，训练模型的误码性能与传统的调制方式性能相吻合，验证了系统方案的可行性和有效性。此外，对于数字传输常见的突发信道，所提方案具有良好的适应性，可获得1 dB左右误码性能的改善。 |
| 关键词: 通信系统 卷积神经网络 端到端学习 自编码器 |
| DOI： |
|
| 基金项目:国家自然科学基金资助项目（61371091） |
|
| An improved CNN end-to-end learning-based autoencoder communication system |
| WANG Xudong,LIN Bin,ZHANG Kaiyao,WU Nan |
| (College of Information Science and Technology，Dalian Maritime University，Dalian 116026，China) |
| Abstract: |
| In order to enhance the generalization ability of the autoencoder based end-to-end communication systems,an improved convolutional neural network(CNN)based autoencoder communication system is proposed.This scheme introduces 1-dimensional convolution layers(Conv1D)in the autoencoder and optimizes the network performance by redesigning the parameters.The transmitter uses multiple layers of Conv1D to extract the feature of the input sequence,and obtains the optimal modulation and coding scheme during training.The receiver also uses the multiple layers of Conv1D to recover the symbols contaminated by noise.Simulation experiments show that the proposed system exhibits fascinating generalization ability with different input bit lengths and code rates.Moreover,for the additive white Gaussian noise(AWGN)channel and the Rayleigh fading channel,the performance of the training model is consistent with that of the traditional modulation method,which verifies the feasibility and effectiveness of the system scheme.In addition,for the common burst channel of digital transmission,the proposed scheme exhibits good adaptability and improves the error performance by about 1 dB. |
| Key words: communication system convolutional neural network end-to-end learning autoencoder |
