摘要
目前，用于地震道自动编辑的最好方法是张学工、李衍达提出的新奇滤波器方法。该方法是一个满足负Hebb学习规则的单层全连接反馈网络。网络的实现分为两个阶段，即训练阶段和识别阶段。在训练阶段，如果每步均输入一个不变的样本序列，则随迭代次数的增加，输出序列迅速地收敛，且系统对样本信息具有记忆特性。在识别阶段，所输入的信号与训练阶段的信号一致时，则网络输出一个小值，否则输出一个大值。但是，为了让学习样本具有统计特性，就要输入一大批样本序列，这就给系统带来困难。为了克服因输入样本序列增加所出现的困难，本文将原网络改为单层前馈全连接网络，修改后的网络收敛性好、运算速度快、且容易实现。
In nowadays,the best method for automatic seistnic trace editing is Novelty filter method presented by Mr. Zhang Xuegong and Mr. Li Yanda. This method uses the neural network which is based on minus Hebb rule and has one layer of neurones connected each other by feedbacks. The neural network works in two steps:training and detecting. If a stationary sample sequence is inputted in each time in the training stage, the output sequence will converge rapidly with the increase of iteration;what is more,the network system remembers the signal series. If the inputted signals in detecting stage are consistent with those in training stage,the network gives a small output, otherwise a big output. In order to make learning sample have statistical properties,a batch of sample sequences must be inputted,thus bringing troubles to the network. To remove the troubles caused by the increase of inputted sample sequences,we use the network which has one layer of neurones connected each other by feed forward. This neural network results in good convergence, fast operation and easy realization.
出处
《石油地球物理勘探》
EI
CSCD
北大核心
1994年第2期147-155,共9页
Oil Geophysical Prospecting
关键词
人工神经网络
地震道
自动编辑
seismic data processing,trace editing,artificial neural network
