基于卷积神经网络的多天线Polar码联合解调-解码方案

为了验证运用神经网络进行信道解码的可行性,利用其提高短码长Polar码的译码准确率,提出了一种基于卷积神经网络(convolutional neural network,CNN)的多天线(multiple-input multiple-output)Polar码新颖联合解调-解码方案.搭建了一种包括4个卷积层,2个全连接层和1个输出层的卷积神经网络,采用最小均方误差作为损失函数,通过计算机生成了Polar码编码的多天线数据并对网络进行训练,使训练得到的神经网络能很好地提取出Polar码比特间的关系特征,从而拟合出Polar码译码函数.仿真结果表明,在相同信噪比条件下,基于CNN的Po-lar码联合解调-解码方案的误码率优于已有的基于全连接神经网络方案;所提方案在不同码率的仿真实验中的误码率皆优于基于全连接神经网络方案,损失曲线的收敛速度更快,显示了基于CNN的Polar码联合解调-解码方案具有更好的泛化能力和学习能力.

Joint demodulation-decoding scheme of multi-antenna Polar codes based on the convolutional neural network

To verify the possibility of applying neural network to channel decoding and the ability of reducing the bit error rate (BER) of the short Polar Codes,a novel joint demodulation-decoding scheme for multi-antenna Polar Codes is pro-posed. A convolutional neural network ( CNN) is constructed, with six learned layers: four convolutional, two fully-con-nected and one output layer. The mean squared error ( MSE) loss function is adopted and polar-coded multi-antenna dataset is generated to train the network,so that the features between the bits of Polar Codes can be extracted and a well-fitted Polar Codes decoding function is achieved. Simulation results show that the proposed scheme has superior performance in BER over the existing fully connected network based scheme. Moreover,in all the simulations with different code rates, the pro-posal achieves better results than the original one, which proves superior generalization ability of the joint demodulation-decoding scheme for multi-antenna Polar codes.