基于遗传神经网络的大地电磁非线性反演

Inversion plays an important role in the interpretation of geophysical data. As there are still problems needed to be solved, such as the result of linear inversion depends on initial model and easily falls into local optimum in the iteration, single nonlinear inversion has premature convergence，slowly convergence speed,unstable results and so on, the applicant puts forward a hybrid optimizing algorithm combining two nonlinear inversion methods and applies into the 1D/2D magnetotelluric inversion. The basic framework of neural network is built for the 1D/2D magnetotelluric model and the network learning process based on magnetotelluric is set up，then the learning process of neural network is optimized by making use of genetic algorithm to get the optimal solution of network connection weights, finally the trained genetic neural network is verified through inversion ,and compared with other classical inversion methods. The algorithm uses genetic algorithm to optimize neural network weights, making full use of the parallel computing ability and global searching capability of genetic algorithm and the local optimization of neural network to improve the operation efficiency and calculation accuracy. Applying the trained genetic neural network to nonlinear inversion of magnetotelluric could make the inversion has the real-time property and improve the efficiency of field work. This algorithm aims at supplying a new method to improve the nonlinear geophysical inversion method.

反演是影响地球物理数据资料解释的重要因素，由于线性化反演结果依赖于初始模型、易陷入局部最优解，而单一的非线性反演往往存在早熟收敛、收敛速度较慢且结果不稳定等问题，申请者提出结合两种非线性方法的混合优化算法，并应用于大地电磁一二维反演中。本项目首先针对大地电磁一二维地电模型建立神经网络基本框架，研究基于大地电磁的网络学习过程，然后利用遗传算法对神经网络学习过程进行优化，得出网络连接权值的最优解，最后对训练好的遗传神经网络进行反演验证，并与其它经典反演方法进行比较。该算法采用遗传算法优化神经网络权值，充分利用遗传算法的并行计算能力和全局搜索性以及神经网络的局部寻优性，提高运算效率和计算精度。经过学习训练后的遗传神经网络应用于大地电磁非线性反演中，可使反演具有实时性，提高野外实践的工作效率。本算法将为改进和完善地球物理非线性反演提供新的方法，推进非线性反演的进一步发展。

实现了遗传算法优化神经网络的大地电磁反演。由于大地电磁神经网络学习训练过程一般采用随机值进行网络权值的初始化，而网络使用的学习算法是一种梯度下降法，有可能产生局部最优解而不是全局最优解，为进一步提高大地电磁非线性反演的稳定性、运算效率及准确度，将遗传神经网络算法引入大地电磁反演。首先针对大地电磁地电模型建立BP神经网络基本框架进行学习训练，网络输入为已知地电模型的视电阻率参数，输出为该地电模型参数，比较不同隐含层数及节点数的计算误差，设计了合理的神经网络结构；再利用遗传算法对神经网络学习训练过程进行优化，计算出多种地电模型网络连接权值和阈值的最优解；最后将学习训练好的网络用于未知模型进行反演测试，网络输入为未知地电模型的视电阻率参数，输出为该地电模型参数。实验结果表明：遗传神经网络算法充分结合了遗传算法的并行计算性和全局寻优性，以及神经网络的局部寻优性，相比单一神经网络算法，在网络学习训练中提高了解的收敛成功率和计算速度，在反演测试中能更准确地逼近真实模型，验证了遗传神经网络算法在大地电磁反演中的可行性和有效性。