基于多主分量神经网络的同步DS-CDMA伪码盲估计

PN code sequence blind estimate of synchronous DS-CDMA based on multi-principal component neural network

针对批处理方法在实现非等功率同步直接序列码分多址(direct sequence code-division multiple access,DS-CDMA)信号伪码序列盲估计时存在的复杂度高、收敛速度慢的问题,引入了3种多主分量神经网络(Sanger NN、LEAP NN和APEX NN)。首先将已分段的一周期DS-CDMA信号作为神经网络的输入信号,用神经网络各权值向量的符号函数代表DS-CDMA信号各用户的伪码序列,然后通过不断输入信号来反复训练权值向量直至收敛,最终DS-CDMA信号各用户的伪码序列就可以通过各权值向量的符号函数重建出来。此外,本文提出了一种在递归最小二乘(recursive least square,RLS)意义下的最优变步长收敛模型,极大地提高了网络的收敛速度。理论分析与仿真实验表明:将3种神经网络用于同步非等功率DS-CDMA信号伪码盲估计时的复杂度均明显降低,且LEAP NN与Sanger NN均可有效地实现-20dB信噪比、10个用户下的同步非等功率DS-CDMA伪码盲估计,APEX NN则相对较差,此外,LEAP NN消耗内存较大、收敛速度快,APEX NN相反,Sanger NN则介于两者之间。

Aiming at the problem of the batch processing method with high complexity and slow convergence speed for the pseudo-noise （PN） code sequence blind estimate of synchronous direct sequence code-division multiple access （DS-CDMA） signals under different power level, three multi-principal component neural net works （NNs） are introduced--Sanger NN, LEAP NN and APEX NN. Firstly, the period segmented DS-CD- MA signals are chosen as the neural network input and the symbol function of each weight vector is used to re- present the PN code sequence of each user. Then through the continuous input signal, the weight vectors of the NN are trained repeatedly until convergence. Finally, the PN code sequence of each user can be rebuilt by the symbolic function of each weight vector. Furthermore, an optimal variable step convergence model is put for- ward via the recursive least square （RLS）, which improves the convergence speed of the network greatly. Theory analysis and simulation results show that the complexity of three kinds of NNs when used to the PN code sequence blind estimate of synchronous DS-CDMA signals under different power level is reduced significantly, and when the signal to noise ratio （SNR） is -20 dB and the number of users is 10, the PN code sequence of synchronous DS-CDMA signals under different power level can still be estimated by LEAP NN and Sanger NN efficiently. Compared with LEAP NN and Sanger NN, APEX NN is poor relatively. In addition, LEAP NN consumes lager memory but it has fast convergence speed. APEX NN is contrary to LEAP NN; Sanger NN is between the LEAP NN and the APEX NN.