基于在线惯序极限学习机的瞬变电磁非线性反演

基于梯度下降法的传统人工神经网络瞬变电磁反演方法计算效率低,不能保证全局收敛。为了解决上述问题,提出一种在线惯序极限学习机(online sequential extreme learning machine, OSELM)的瞬变电磁反演方法。该方法针对瞬变电磁法所获取的高维勘探数据进行建模反演,首先,通过随机设定隐层参数(输入权值和偏差)来简化模型的学习过程;然后,将测试得到的预测样本加入训练样本中,作为下一次的更新信息,建立在线贯序极限学习机预测模型,从而最大限度提高反演精度;最后,设计了两个经典的瞬变电磁层状地电模型并进行了拟二维地电模型的反演。反演结果表明,该方法能够较好地解决瞬变电磁法高维数据非线性建模的反演问题,同时相较极限学习机(extreme learning machine, ELM),非线性反演方法具有更加准确的反演结果、更好的泛化能力以及更高的计算效率,为神经网络在地球物理反演中的应用提供了新思路。

Online sequential extreme learning machine for transient electromagnetic nonlinear inversion

The traditional transient electromagnetic inversion method using artificial neural network based on gradient descent method is inefficient and can not guarantee global convergence. In order to solve these problems, this paper proposes a transient electromagnetic inversion method based on on online sequential extreme learning machine (OSELM). This approaches is used for inversion of high-dimensional exploration data obtained by transient electromagnetic method. Firstly, the hidden layer parameters (input weight and deviation) are randomly set to simplify the learning process of the model. Then, the prediction samples obtained from the test are added to the training samples as the next update information, and the online sequential extreme learning machine prediction model is established to maximize the inverse accuracy. Finally, the inversion results of two classical TEM layered geoelectric models and a quasi two-dimensional geoelectric model show that the proposed method can solve the problem of nonlinear modeling and high-dimensional data for TEM inversion, and a comparison with extreme learning machine (ELM) shows that this method has more accurate inversion, better generalization ability and higher calculation efficiency, which provides a new idea for the application of neural network in geophysical inversion.