改进的神经网络级联相关算法及其在初至拾取中的应用

针对传统BP神经网络存在的问题，引入一种神经网络构造算法——级联相关（CC）算法。该算法具有比BP算法更快的收敛速度，能根据待解决问题自行确定网络结构，即能随时扩展网络拓扑结构以学习新样本。常规CC算法的初始网络只包含输入层和输出层。改进的CC算法起始于适当的BP网络结构（存在隐含层）；且为了防止权值病态递增，在训练候选隐含神经元的目标函数中加入了正则化项，对权值进行衰减。仿真试验表明：改进的CC算法具有更快收敛速度、更强泛化能力；瞬时强度比、振幅、频率、曲线长度比、相邻道相关性等五种地震属性特征交会图显示对初至波具有稳定的区分能力。本文构建的神经网络初至拾取方法在实际资料应用中取得了良好效果。

An improved neural-network cascade-correlation algorithm and its application in seismic first break picking

To overcome existing problems of BP neural network,we introduce the cascade-correlation algorithm into the neural network construction.The proposed algorithm has a faster convergence than BP algorithm,and it can decide its own network architecture based on problems to be solved.That means it can expand network topology to learn new samples.The initial network of the standard cascade-correlation algorithm has only an input layer and an output layer,while the improved algorithm starts with an appropriate BP network architecture (including hidden layers).In addition,in order to prevent weight-ill growth,a regularization term is added to objective functions in candidate hidden units training to decay weights.Simulation experiments demonstrate that the improved cascade-correlation algorithm has faster convergence and stronger generalization ability.Cross-plots of five attributes such as instantaneous-intensity ratio,amplitude,frequency,curve-length ratio,and adjacent-trace correlation,show that first breaks peaked by the proposed algorithm can be easily and reliably discriminated.The proposed algorithm achieves good performance in the first break picking on real data.