改进的神经网络并行算法及其在地震初至拾取中的应用

针对经典BP神经网络易于陷入局部极小点、易于产生振荡等缺点,提出了神经网络初始权值的二分法,改进了一种网络结构自动确定算法,并将随机算子和遗忘因子引入BP神经网络中.在提高全局寻优能力的同时,加快了网络的收敛速度.在分析了神经网络内在并行性的基础上,基于MPI实现了改进算法的并行化,将算法应用于地震资料的初至拾取,并取得了良好的应用效果,验证了算法的有效性.

An improved neural network papallel arithmetic and its application in picking seismic first break

Based on analyzing the shortcoming of BP neural network,some methods are developed to improve the classic BP neural network,including the dichotomy to determine initialization of weights,an improved auto-determination method of network structure,random operator and forgetting factor introduced to BP neural network.With all these methods,the stronger capability of global optimization and quicker network's convergence speed have been obtained.Then after analyzing the internal parallel characteristic of neural network,the authors design and realize the parallel arithmetic of the improved BP neural network based on the MPI.The result of application in picking seismic first break is satisfactory and shows that the parallel arithmetic is effective.