自组织特征映射网络的改进及在储层预测中的应用
|
摘要
自组织特征映射网络（SOM）虽然具有自组织、自学习和侧向联想能力，但在实际应用中仍存在一些缺陷。如网络在学习时收敛速度较慢，不同的初始条件和样本输入顺序对网络的学习过程和学习结果都很敏感，网络在无监督学习时不能充分利用可靠的先验知识。为此，本文对SOM算法进行了多项改进，主要有：①在网络初始化时将已知样本矢量作为典型样本，再对特定的输出节点的权向量进行初始化，然后对每次迭代后的网络进行非强制性的修正，从而提高了分类精度；②采用自适应方法调节学习速率，加快了网络的力J练速度；③给出判断迭代是否收敛的准则，提高了运算速度；④以输出层各书点权向量之间的欧氏距离为依据对每个输出节点的类别号进行重排，实现了对样本的有序分类。本文应用这种改进的SOM算法，对准噶尔盆地东部地区的一条穿越3口探井的地震剖面进行了储层预测，获得了良好的效果。
Although the self-organizing mapping neural network is capable of self organization, self learning and side associative thinking, it has some defects:. Slow convergence in learning course . The variations of initial condition and sample input sequence are very sensi-tive to the learning course and the learning result . No full use of reliable priori knowledge during its learning without supervi-sionTo remove the above defects, we improve here the neural network in the fol-lowing aspects:. In initializing the neural network, we take the known sample vectors as typi-cal samples, initialize the weight vectors of Riven output nodes, and revise unforce-dly the network after each iteration, thus improving the classification accuracy..Learning rate is adaptively regulated to speed up the training of the neuralnet work..The criterion for judging if the iteration converges is offered to quicken com-putation.. According to Euclidean distance between the weight vectors of the corre-sponding nodes in an output layer, the classification number. of output nodes arerearranged to achieve desirable sequential classiftcation.The improved neural network algorithm has been used to predict the reservoirof a seismic section that passes three boreholes in eastern Zhungeer basin, and theprediction brought good effect.
Although the self-organizing mapping neural network is capable of self organization, self learning and side associative thinking, it has some defects:. Slow convergence in learning course . The variations of initial condition and sample input sequence are very sensi-tive to the learning course and the learning result . No full use of reliable priori knowledge during its learning without supervi-sionTo remove the above defects, we improve here the neural network in the fol-lowing aspects:. In initializing the neural network, we take the known sample vectors as typi-cal samples, initialize the weight vectors of Riven output nodes, and revise unforce-dly the network after each iteration, thus improving the classification accuracy..Learning rate is adaptively regulated to speed up the training of the neuralnet work..The criterion for judging if the iteration converges is offered to quicken com-putation.. According to Euclidean distance between the weight vectors of the corre-sponding nodes in an output layer, the classification number. of output nodes arerearranged to achieve desirable sequential classiftcation.The improved neural network algorithm has been used to predict the reservoirof a seismic section that passes three boreholes in eastern Zhungeer basin, and theprediction brought good effect.
引文
1靳善,范俊波.神经网络与神经计算机原理·应用,西安交通大学出版社,1991
2焦李成.神经网络计算,西安电子科技大学出版社,1993
3Nguyen D and Widrow B. Neural networks for selflearning control systems· IEEE vontrol Systems Mag- azine.199o,10(3):18～23
4Wang Z Z and Hu D W. Self-organizing neural network adaptive control for non-linear systems. IEEE Industry Electronics Condrence, 1992
2焦李成.神经网络计算,西安电子科技大学出版社,1993
3Nguyen D and Widrow B. Neural networks for selflearning control systems· IEEE vontrol Systems Mag- azine.199o,10(3):18～23
4Wang Z Z and Hu D W. Self-organizing neural network adaptive control for non-linear systems. IEEE Industry Electronics Condrence, 1992
版权所有：© 2023 中国地质图书馆 中国地质调查局地学文献中心