无线光通信下极化码DNN-NOMS译码方法研究

Research on DNN-NOMS decoding method of polarization code in wireless optical communication

针对无线光通信中大气湍流引起极化码置信度传播译码性能不佳的问题,提出了一种无线光通信下极化码DNN-NOMS(Deep Neural Networks-Normalized and Offset Min-Sum)译码方法。首先,把传统的极化码置信传播译码算法因子图转化为类似于低密度奇偶校验(Low-density Parity Check,LDPC)码的Tanner图,在Tanner图展开并转化为深度神经网络(DNN)图形表示的基础上,将MS(Min-Sum)译码方法同时添加归一化因子和偏移因子来给Tanner图的边赋予权重,简化极化码对数似然比的计算方法,通过限制训练参数的数量,选取在损失函数最小的条件下的因子参数,训练得到最优归一化因子和偏移因子的译码模型。仿真结果表明,在不同的大气湍流强度下,该译码方法以牺牲较小的存储空间为前提的情况下能选取更优的归一化因子和偏移因子参数,从而获得更好的误码率性能,且大幅度降低译码复杂度;在误码率为10^(-4)时,DNN-NOMS译码方法能产生0.21~3.56 dB的性能增益,且将迭代次数的运算次数降低87.5%。

Aiming at the problem of poor confidence propagation decoding performance of polarization codes caused by atmospheric turbulence in wireless optical communication,a Deep Neural Networks-Normalized and Offset Min-Sum(DNN-NOMS)decoding method under wireless optical communication was proposed.First,the factor graph of the traditional belief propagation decoding algorithm for polarized codes had been transformed into Tanner graphs which similar to Low-density Parity Check(LDPC)codes.The Tanner graphs were expanded and transformed into Deep Neural Network(DNN)graphical representations.The Min-Sum(MS)decoding method added the normalization factor and the offset factor,at the same time to the edge weights of the Tanner graph were given,which simplified the calculation method of the log likelihood ratio of the polarization code.By limiting the number of training parameters,the factor parameters were selected under the condition of the minimum loss function,and trained to obtain the optimal normalization factor and offset factor of the decoding model.The simulation results show that under different atmospheric turbulence intensities,the decoding method can select better normalization factor and offset factor parameters under the premise of sacrificing smaller storage space,so as to obtain better error codes.The DNN-NOMS decoding method can produce a performance gain of0.21-3.56 dB and reduce the number of iterations by 87.5%when the error rate is 10^(-4).