谱分析的神经网络方法──局部零残差法

The Neural Network Method for Spectrum Analysis──Local Zero Residue Method

提出了一种神经网络ＡＲ（自适应递归）参量估计的新方法──局部零残差法。利用所建立的神经网络的明确收敛性，在每一个数据到来后获得一组ＡＲ参量的零残差解，这时ＡＲ参量的特性是与时间序列的噪声相关的随机过程，最终的ＡＲ参量值为全部数据通过神经网络预测误差滤波器后，对各点获得的ＡＲ参量解的求和平均。由此，ＡＲ参量值确定的最大熵谱具有很好的稳定性，且谱的分辨率和精度是与滤波器的长度成正比，这给最佳滤波器的选择保留了充分的灵活性。另外，利用神经网络并行快速的特性，可用该系统进行实时谱分析。

new method for evaluating parameters based on the neural network AR(Adaptive Recursion ),i.e.the local zero residue method is proposed. Bymeans of the determinable convergence of established neural network, a set of zero residue solution of AR parameters is obtained when the every set of data has coming. For this time,the feature of AR parameters is a random process relating to the noise of the time series, the final value of AR parameters is the summation of average value which is obtained from NN (neural network) predicted filter and resolved by the Kalman filter using the all of obtained AR parameters. Therefore, the spectrum of maximum entroopy determined by the value of AR parameters has considerable stability, and its resolution and accuracy of signal are direct proportional to the length of this fliter,it gives the sufficient flexibility for choice of the optimal filter. Beside this,by using the characteristics of NN' s rapidly parallel computing, the spectrum analysis can be performed by the system in real time.