Hybrid Nonlinear Inversion For Electrical Resistivity Imaging Based On Neural Network

Abstract:Electrical Resistivity Imaging is an important geophysical prospecting method which is widely used in hydrogeology, environment, archaeology, mineral exploration and oil-gas exploration. Considerable economic profits have been made by it. With the intensive studies of theory and the development of engineering application in recent years, the request for exploration scope and interpretation precision is also improving. So the traditional linear inversion method will face new challenge. The application of neural network nonlinear inversion in Electrical Resistivity Imaging is studied in the dissertation, and the optimization algorithm, modeling approach and inversion flow of BP neural network as well as RBF neural network are analyzed.BP neural network has the disadvantages of slow convergence and easily getting into local minimum. BP neural network is combined with Particle Swarm Optimization and Differential Evolution algorithm to optimize the inversion process and to improve the inversion result from two different angles, both of which have global searching performance. Firstly, by combining BP neural network with Particle Swarm Optimization algorithm, a PSO-BP algorithm based on chaotic oscillation is proposed. The inertia weight is adjusted adaptively to improve the global optimum capability of PSO by using the Chaos oscillation curve in this method. The weights and threshold values of the BP neural network are trained and optimized. This algorithm is independent of the initial model and it enhances the stability and accuracy of the neural network inversion. Secondly, by combining BP neural network with Differential Evolution algorithm, a chaotic constraint DE-BP algorithm is presented. In this algorithm, Tent equation is applied to set parameters F and CR automatically, and the restricted parameter is used to enhance the ability of converging to global optimum. The results of numerical computation and model simulation show that the proposed method has better convergence performance and higher imaging quality.The theory and method of neural network inversion are improved by the combination of intelligence algorithm and BP neural network, though optimization technique to network parameters can only improve the learning performance of neural network, it can not solve the problem of local optimum. Compared with BP neural network, Radial basis function neural network has many advantages such as simple structure, fast learning rate and improved global search ability. By combining RBF neural network with information criterion, an improved HQOLS learning algorithm based on Hannan-Quinn Criterion for hidden layer structure selection is presented. The inversion performances of k-means clustering algorithm, gradient algorithm, OLS algorithm and the proposed algorithm are compared, the feasibility of HQOLS algorithm and RBFNN inversion are proved.HQOLS algorithm can select the hidden layer structure of RBF neural network adaptively and solve the problem of hidden layer structure design, but the selected RBF neural network structure is not always the best. Deeper research has been done on the application of RBF neural network and Information Criterion in Electrical Resistivity Imaging inversion. A two-stage learning algorithm framework is presented. In the proposed framework, aftering calculating the optimal IC value automatically, the hidden layer structure is obtained, and the global searching algorithm is used to optimize the centers and widths of the radial basis functions in the hidden layer. AIC and PSO are used to realized the framework in this dissertation.The conclusion of this dissertation indicates BP neural network and RBF neural network have both advantages and disadvantages, and the research results can provide advantageous experience for the neural network inversion of electrical method in the future.