人工神经网络非线性地震波形反演
|
摘要
地震波形反演是非线性问题,其目标函数都是多极值的函数,若用线性化的反演方法求解,常会遇到迭代收敛于目标函数局部最优等困难。本文研究能求得目标函数全局最优解的遗传算法训练人工神经网络的地震波形反演方法。考虑到遗传算法训练神经网络地震波形反演的未知参数量大,而通常的二进制编码遗传算法占用计算机内存量大,不能在较小内存的计算机上实现,故以可节省内存的0-1编码遗传算法训练神经网络,提出了加速网络收敛的方法。
The seismic waveform inversion is a nonlinear problem. Its objective functions arc functions with several extreme values. When solving the problem using the linear inversion method,the difficulty of thc iteration converging to tho local optimum of the objoctive function is often met. The paper proposes a seismie waveform inversion method using the attifical neural network trained by the genetic algorithm which can get thc overall optimum solution of the objective function. Since the unknown paramcters of the seismie waveform inversion using the neural network trained by the genetic algorithm are many,the usual binary-coded genetic algorithm can not be implemented on the computer with a little memory space, so the 0-1 coded genetic algorithm which can save memory space is used to train the neural network. Also,the method to specd the network convergence is prcsented.
The seismic waveform inversion is a nonlinear problem. Its objective functions arc functions with several extreme values. When solving the problem using the linear inversion method,the difficulty of thc iteration converging to tho local optimum of the objoctive function is often met. The paper proposes a seismie waveform inversion method using the attifical neural network trained by the genetic algorithm which can get thc overall optimum solution of the objective function. Since the unknown paramcters of the seismie waveform inversion using the neural network trained by the genetic algorithm are many,the usual binary-coded genetic algorithm can not be implemented on the computer with a little memory space, so the 0-1 coded genetic algorithm which can save memory space is used to train the neural network. Also,the method to specd the network convergence is prcsented.
引文
[1]靳蕃、范俊波、谭永东,《神经网络与神经计算机原理应用》,西南交通大学出版社,1991年。
[2] Huang K,Liu W H,and Chang I C,Hopfield model of neural networks for detecticn ofbright spots,SEG59 Expanded Abstracts ,444～446,1989｝
[3] Liu X,Xue P,and Li P,Neural network method for tracing seismic events,SEG59 ExpandedAbstracts,716～718,1989.
[4] McCormack M D,Seismic trace editing and fitst picking using neural networks,SEG60 Ex.panded Abstracts ,321～324,1990.
[5] Veezhinathan J and Wagnre D,A neural oetwork approach to flrst-break picking,Proceed.ings of the Internahonal Joint Conference on Neural Networks ,San Digeo,1,235～240,1990
[6] Raiche A,A pattern recognition approach to geophysical Invorsion using neural nets,Gco-yhys.J.Int,1991,105,629-648.
[7] Murat M E and Rudman A J,Automated first arrival nlcklns:A neural network annroach,Geophysical Prospecting,1992,40,587～604.
[8]陈遵德、朱广生,Kohonen网络在油气横向预测中的应用,《石油物探》,1995,Vol.34,No2,53～56。
[9] Calderon-Macias C and Sen M K,Gcophysical interpretation by artificial ncural systems:Afeasibility study , SEG63 Expanded Abstracts ,254～257,1993
[10] Michaels P and R B,Recurrent neural network representation of inelastic wave cquation andfull-waveform inversion without local minima, SEG64 Expanded Abstracts, 22～25,1994.
[11] Roth G and Tarantola A,Inversion of scismic waveformsu sngneura lnctworks,SEG64Ex-panded Abstracts,788～791,1994.
[12]Rothman D H,Nonlinear inversion,statistical mechanics,and residual statics estimation,Geophysics,1985,50,2784～2796
[13] Sambitdge M S,Tarantola A and Kennett,An alternative strategy for non-linear inversionof seismic waveforms,Geophysical Prospecting ,1991,39,723～736
[14] 周辉、何樵登,利用Hartley变换模拟各向异性介质地震波场,《石油地球物理勘探》,1995,Vol.30,No.4,593～601。
[2] Huang K,Liu W H,and Chang I C,Hopfield model of neural networks for detecticn ofbright spots,SEG59 Expanded Abstracts ,444～446,1989｝
[3] Liu X,Xue P,and Li P,Neural network method for tracing seismic events,SEG59 ExpandedAbstracts,716～718,1989.
[4] McCormack M D,Seismic trace editing and fitst picking using neural networks,SEG60 Ex.panded Abstracts ,321～324,1990.
[5] Veezhinathan J and Wagnre D,A neural oetwork approach to flrst-break picking,Proceed.ings of the Internahonal Joint Conference on Neural Networks ,San Digeo,1,235～240,1990
[6] Raiche A,A pattern recognition approach to geophysical Invorsion using neural nets,Gco-yhys.J.Int,1991,105,629-648.
[7] Murat M E and Rudman A J,Automated first arrival nlcklns:A neural network annroach,Geophysical Prospecting,1992,40,587～604.
[8]陈遵德、朱广生,Kohonen网络在油气横向预测中的应用,《石油物探》,1995,Vol.34,No2,53～56。
[9] Calderon-Macias C and Sen M K,Gcophysical interpretation by artificial ncural systems:Afeasibility study , SEG63 Expanded Abstracts ,254～257,1993
[10] Michaels P and R B,Recurrent neural network representation of inelastic wave cquation andfull-waveform inversion without local minima, SEG64 Expanded Abstracts, 22～25,1994.
[11] Roth G and Tarantola A,Inversion of scismic waveformsu sngneura lnctworks,SEG64Ex-panded Abstracts,788～791,1994.
[12]Rothman D H,Nonlinear inversion,statistical mechanics,and residual statics estimation,Geophysics,1985,50,2784～2796
[13] Sambitdge M S,Tarantola A and Kennett,An alternative strategy for non-linear inversionof seismic waveforms,Geophysical Prospecting ,1991,39,723～736
[14] 周辉、何樵登,利用Hartley变换模拟各向异性介质地震波场,《石油地球物理勘探》,1995,Vol.30,No.4,593～601。
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