综合动量法和可变学习速度的BP神经网络地震初至拾取

为了提高初至拾取精度和效率，研究了BP神经网络初至拾取方法，提出综合动量法和可变学习速度的BP神经网络地震初至拾取方法，其主要原理是对网络权值的更新过程进行改进，当均方根误差在权值更新后超过设定的误差范围，则权值更新取消；在既定的误差范围内，权值更新则被接受，且学习速度发生变化。分析不同地震属性对初至波识别的可行性，选取均方根振幅比、曲线长度比、振幅、频率等4种特征属性进行模型测试，结果表明改进方法的初至拾取效果优于常规BP神经网络方法。实际资料测试验证，改进方法构建的网络结构简单，参数少，收敛速度快，具有较强稳定性和抗噪能力，初至拾取精度高。

Seismic first-break picking based on BP neural network integrated with momentum method and adaptive learning rate method

A seismic first-break picking method based on BP neural network integrated with momentum method and adaptive learning rate method was proposed in this paper.It improves the network weight updating process.If mean square error is not within the given error range,the weight update is cancelled.Otherwise,the weight is updated,and the learning rate changes accordingly.Through the analysis on the feasibility of first-break identification using different seismic attributes,four typical attributes,including RMS amplitude ratio,curve length ratio,amplitude and frequency,were chosen for model test.Model test results indicated that the improved method performs better than conventional BP neural network method.The application in real data demonstrated that the improved BP neural network algorithm has simple network structure,few parameters,fast convergence speed,good performance in stability and anti-noising and high first-break picking precision.