基于神经网络和蛙跳算法的超高密度电法快速反演研究

Ultra-high density resistivity method inversion is a complicated non-linear function optimization process, which is high dimensional and non-convex. The traditional linear inversion method has the following problems such as slow convergence, poor computing efficiency and easily falls into the local minimum. This project focuses on the rapid nonlinear inversion method composed of neural network and shuffled frog leaping algorithm (SFLA). The inversion results of neural network are used to optimize the initial population of SFLA, the inertia weight with oscillatory and decline tendency is selected to enhance global search and accelerate convergence. In addition, the regularization technology is used to improve the stability of inversion. These measures can solve the problem properly of multiple extreme, ill-posedness and inefficiency in the process of inversion. The observation data optimization method based on the model resolution is studied, which contributes to the selection and classification for the observation data of ultra-high density resistivity method. The relationship between geoelectric model inversion and neural network is studied and the regularization neural network is used to model the inversion process. Finally, the hybrid inversion algorithm is accelerated by CPU+GPU parallel computation framework based on CUDA. The inversion efficiency will be improved from the three aspects including data optimization, efficient hybrid inversion design and parallel computing. A practical method of fast non-linear inversion will be formed in the project and the research results will provide theoretical support for large-scale and high-accuracy electrical resistivity imaging inversion.

超高密度电法反演是一个复杂的非线性函数综合寻优过程，具有高维和非凸的特性。传统的非线性反演方法易陷入局部极值、收敛缓慢和计算效率低。因此，本项目聚焦于神经网络和蛙跳算法混合的快速非线性反演方法，利用神经网络的反演结果优化蛙跳算法的初始种群，采用振荡下降形态的惯性权重增强蛙跳算法的全局搜索能力并加速收敛，结合正则化技术提高反演的稳定性，力求妥善解决反演中存在的多极值、不适定和低效率等关键问题。通过研究基于模型分辨率的观测数据优化方法，实现对超高密度电法观测数据的优选与分类；探讨地电模型反演和神经网络之间的关系，建立正则化神经网络解释观测数据的反演模型；最后采用CUDA并行框架的CPU+GPU算法加速混合算法的计算过程。通过从数据优化、高效混合反演设计和并行计算三个层面上提高反演的效率，最终形成一套实用的快速非线性反演方法，为大规模、高精度的超高密度电法反演提供理论支持。

超高密度电法是一种成本低、效率高、信息丰富和解释精度高的新型勘探方法，超高密度电法的反演是一个复杂的非线性函数综合寻优过程，具有高维和非凸的特性。传统的非线性反演方法存在易陷入局部极值、收敛缓慢和计算效率低等缺点。.本项目针对以上问题开展研究，通过研究基于核主成分分析的观测数据优化方法，实现对超高密度电法观测数据的优选与降维，降低观测数据间的信息冗余，减少90%以上的观测数据；研究神经网络和蛙跳算法混合的快速非线性反演方法，探讨地电模型反演和神经网络之间的关系，建立正则化神经网络解释观测数据的反演模型，并通过改进的蛙跳算法优化神经网络反演模型的全局搜索能力，提高反演成像的质量；研究反演算法的并行计算能力，通过采用CPU和GPU的并行计算框架从不同粒度上显著加速反演算法，提高反演速度50%以上。.本项目从数据优化、高效混合反演算法设计和并行计算三个层面上针对超高密度电法的非线性反演开展了深入研究，妥善解决了反演中存在的多极值、不适定和低效率等关键问题，并最终形成了一套快速非线性反演方法，为大规模、高精度的超高密度电法反演提供理论和应用支持。