拟牛顿优化BP网络盲判决反馈均衡器

针对水声通信严重多途效应导致的码间干扰，利用神经网络良好的非线性拟合能力，将盲判决反馈均衡器结构与神经网络相结合，同时通过拟牛顿算法提升神经网络的收敛速度，提出了一种拟牛顿优化神经网络的盲判决反馈均衡器。用两个单隐层误差反向传播（Back Propagation，BP）网络替换判决反馈均衡器前馈和反馈滤波器，利用拟牛顿迭代计算神经网络权值，在不计算二阶导数的前提下，使用近似矩阵来近似各层网络权值误差性能函数Hessian矩阵的逆矩阵，通过测量各层权值的梯度变化进行迭代计算。应用自动增益控制和锁相环进行幅度和相位修正。仿真结果表明，拟牛顿优化神经网络的盲判决反馈均衡器在水声信道均衡问题中具有更快的收敛速度及更低的误码率。

Blind decision feedback equalizer of quasi Newton optimized BP network

In view of at the inter symbol interference(ISI) caused by serious multipath effect in underwater acoustic communication, a blind decision feedback equalizer based on quasi Newton optimization neural network (named as B-QNBPDFE) is proposed, in which the structure of blind decision feedback equalizer (B-DFE) and back propagation (BP) neural network are combined, and the convergence speed of neural network is improved by quasi Newton algorithm. Two single hidden layer BP networks are used to complete the function of DFE feedforward and feedback filters. The weights of neural networks are calculated by quasi Newton iteration. Without calculating the second derivative, the inverse matrix of Hessian matrix is approximated by approximate matrix. The iterative calculation is carried out by measuring the gradient change of weights of each layer. Finally, phase correction is carried out by phase-locked loop. The simulation results show that the blind decision feedback equalizer based on Quasi Newton optimization neural network has faster convergence speed and lower bit error rate in underwater acoustic channel equalization.