基于学习机制的广域电磁法激电信息提取研究

Extracting the polarizability of the weak induced polarization information from the wide field electromagnetic method is a complex nonlinear function optimization problem with the characteristic of non-convexity and multi-solution. Using the completely nonlinear inversion algorithm to inverse the polarizability has the bottleneck problems such as easy to fall into local extremum, low computational efficiency and difficulty in stable convergence. The learning mechanism is utilized in this project to learn the historical data generated in the process of the complete nonlinear inversion iteration as a priori information, so as to realize the intelligent inversion of optimizing at the same time of learning, and strive to properly solve the key issues such as multiple maximum, inefficiency and ill-posed existing in the inversion. Specifically, first, the statistical learning method is used to learn the parameter knowledge of the shuffled frog leaping inversion algorithm in the iterative process in order to realize the adaptive setting of the inversion algorithm parameters, coordinate the global and local search ability of the inversion algorithm, and improve the solution quality; Second, Bayesian neural network is adopted to learn the forward modeling knowledge of algorithm evaluation, and a proxy model that can quickly evaluate the solution is established to improve the efficiency of solving the inversion algorithm; Third, the reinforcement learning method is applied to learn the sensitivity difference of resistivity and polarizability in the minimum structural inversion process to realize the adaptive regularization of the objective function and improve the stability of the inversion solution. A practical intelligent extraction method of IP information will be formed in this project to obtain resistivity and polarizability parameters simultaneously, which is anticipated to provide the theoretical and application support for the realization of the “structural + attribute” exploration of wide field electromagnetic method under the complex geological conditions.

从广域电磁法中提取弱小的激电信息极化率，是一个复杂的非线性函数寻优问题，具有非凸性和多解性。采用完全非线性反演算法求解该问题存在容易陷入局部极值、计算效率低和难以稳定收敛等问题。本项目利用学习机制，将反演迭代过程中产生的历史数据作为先验信息进行学习，实现边学习边优化的智能反演，力求妥善解决反演中存在的多极值、低效率和不适定等关键问题。具体而言，采用统计学习方法学习蛙跳算法在迭代过程中的参数知识，实现反演算法参数的自适应设置，协调算法的全局和局部搜索能力，提高解的质量；采用贝叶斯神经网络学习算法评估时的正演知识，建立能够快速评估解的代理模型，提高反演的求解效率；采用强化学习方法学习电阻率和极化率在最小构造反演过程中的敏感度差异，实现目标函数的自适应正则化，提高反演的稳定性。本项目能同时反演获得电阻率和极化率参数，为复杂地质条件下广域电磁法实现“构造＋属性”勘探提供理论和应用支持。

地下介质的频率域电磁波响应为电磁感应和激电效应的综合反映，开展从广域电磁法中提取激电信息的研究，能够通过一次测量，同时获取反映地电构造的电阻率参数和表征地电属性的极化率参数，更多物性参数的获得有助于提高频率域电磁法反演与解释的精度，实现“构造+属性”勘探，具有重大的科学意义和实用价值。然而采用完全非线性反演算法求解该问题存在容易陷入局部极值、计算效率低和难以稳定收敛等问题。.本项目提出基于学习机制的激电信息提取方法，将完全非线性反演迭代过程中产生的历史数据作为先验信息进行学习，实现边学习边优化的智能化反演。具体而言包括以下三个方面：.（1）采用统计学习方法学习蛙跳算法反演迭代过程中的算法参数知识，实现自适应的反演参数设定，协调反演过程中的全局搜索和局部搜索，提高求解的质量。.（2）通过学习反演算法在评估解的质量时的正演知识，设计精度稍低但计算速度快的贝叶斯神经网络代理模型，并将代理模型与正演算法相结合评估解的质量，在不损失求解质量的同时提高求解的效率。.（3）采用强化学习方法学习最小构造反演过程中极化率参数和电阻率参数敏感度的变化规律，以实现最小构造目标函数中正则化参数和约束条件的自适应设定，提高求解的稳定性。.基于本项目的研究，实现了能同时获得电阻率和极化率参数的反演方法，为复杂地质条件下广域电磁法实现“构造＋属性”勘探提供了理论和应用支持，丰富了基于智能优化技术的非线性反演理论与方法，提高了广域电磁法的资料处理及解释水平。