基于自适应神经网络的信号处理

【作者】 范怀玉；

【导师】 申金媛；

【作者基本信息】 郑州大学 ， 光学， 2007， 硕士

【摘要】 人工神经网络在信号处理方面得到了广泛应用。本文主要研究了人工神经网络的结构优化和样本信息的选择问题，并将优化后的神经网络应用到非线性时间序列信号预测和荧光光谱的识别中。在构建人工神经网络模型过程中，选择什么样的学习算法以及如何寻找最佳的网络结构一直是决定网络性能优劣的关键所在。特别是网络拓扑结构的选择非常重要，不恰当的结构或导致推广能力很差，对于没有训练学习过的模式没有好的预测精度或高的正确识别率，或导致训练学习过程中收敛速度很慢甚至可能无法收敛。样本中的冗余信息会影响人工神经网络较精确表示网络输入和输出间的映射关系。在学习算法上，我们选择了递归最小方差（RLS：Kecursive Least Square），与其他算法相比，该算法中需要调节的参数较少，使得它的学习更为容易，而且，由于RLS算法能够自动估算出互连权重调节的最佳步长值，所以神经网络的收敛速度和效率都有了较大提高。在网络规模大小的选择方面，我们引入了互连权重删剪算法，在网络学习训练后，分别对网络中间层单元和输入层单元进行删剪优化。对输入层单元的删剪，即是对网络输入样本中的冗余信息分量进行删除。这种在网络学习训练过程中即刻画网络输入和输出之间的函数关系过程中，由与输出方差有关的能量函数来删除输入样本的冗余信息的方法，与在网络学习训练之前通过其他的数据压缩方法或特征提取方法相比，该方法在预测精度和正确识别率上更有效。本文将这种基于自适应训练及删剪算法的神经网络分别应用在非线性时间序列信号（如：股票价格指数、可变比特率视频量）的预测和非线性荧光光谱识别中。通过实验结果可以看出，删剪优化后网络的预测均方误差的数量级为10^-5，而且比未删剪时的预测性能还要好；对于信号识别，删剪优化后的正确识别率依然是100％。；实验表明基于删剪算法的优化网络不但降低了网络的计算复杂度，而且提高了预测精度或正确识别率。该方法因其快的收敛速度可以用来实时信号处理。更多还原

【Abstract】 Artificial neural network is used widely in the signal processing. This thesis will have mainly studied the questions of the artificial neural network structure optimization and the sample information choice. The optimized neural networks are applied to forecast the non-linear time series signal and to recognize the fluorescence spectrum. In the process of constructing the artificial neural network model, how to choose the study algorithm as well as to seek the best network architecture is always a key to decide the network’s performance. It is very important to choice of the network topology architecture. The inappropriate structure may cause the poor generalization ability and slow the convergence rate or non-convergence. The redundant information of the sample, which affects the correct input-output mapping of the neural network, is deleted. The Recursive Least Square （RLS） is chosen to adjust the interconnection weights. Compared with other algorithms, the RLS algorithm not only is easier to train a network because it has less parameters but also has faster convergence rate because it automatically has the optimal training stride values. In order to optimize the network’s topology architecture, we introduce the pruning algorithm into the network. It can prune the neurons not only in the middle layer but also in the input layer by proposed a new energy function. Pruning the neurons of the input layer means that the redundant component of the input sample is moved. Comparing other data compression or feature extraction methods, this method is more effective for improving the prediction accuracy or correct recognition rate because the neural pruning is relative to the output error. The neural network models based on the adaptive RLS learning and pruning algorithm are used to forecast nonlinear time series signals, such as stock indexes and variable bit rate video traffic, and recognize nonlinear fluorescence spectrum of ethylene. The computer simulations have done. The mean square error of the optimized network is up to about 10^-5 for forecasting nonlinear time series signals. The correct recognition rate of the proposed model is up to 100% for recognizing nonlinear fluorescence spectrum. The experimental results show that optimal network based on the pruning algorithm not only not only reduce the calculating complexity greatly, but also improve the prediction accuracy or correct recognition rate. The presented method can be used to process real-time the signals because of its faster convergence rate.更多还原