Research Of Nonlinear Parameter Identification For Geoelectric Model Based On DC Resistivity Method

Distinguishing the parameter of the DC source model is a complex nonlinear mapping problem. Conventional linear distinguishing method of nonlinear problem are linearized processing. And this will lead to an inevitable ending – local extremum. This problem will result in a low distinguishing accuracy model which can’t meet the requirements of the prospecting. The research shows that the nonlinear distinguishing algorithm based on the BP neural network can improve the distinguishing accuracy of the model and can carry out the experiment of deep geological exploration with its nonlinear approximation ability and learning ability.But in the DC parameter distinguishing process of the BP neural network, there are many disadvantaged: 1) Sensitive to the initial value, easy to fall into local minimum. 2) Convergence rate is slow, training period is long. 3) No regularity in determination of hidden layer nodes. 4) Exist over learning or under learning problem, especially when facing the high dimensional complex geoelectric model, there will be a rapid attenuation in distinguishing quality with the parameter expansion. All above factors limited the development and application of the BP neural network.To solve the problem in application of the DC BP neural network distinguishing algorithm, this paper improved the algorithm performance from two different angles. Firstly, combine the particle swarm optimization and genetic algorithm separately which have global search advantage. Taking the advantage of each algorithm to improve the distinguish quality. Secondly, propose BP neural network like model based on the mathematics and statistics, which well supported the vector machine distinguishing. Well solve the high dimension model distinguishing nonlinear mapping problem, overcome the shortcoming. Embody the advantage in processing the DC parameter distinguishing and improve the distinguishing accuracy.This paper focus on five parts: theory, numerical simulation, method, simulation model and engineering application. The nonlinear algorithm based on the neural network has been discussed. The main research contents are as follows:1) In the theory part, elaborate the modeling process and identification principle of the BP neural network. Starting with the working principle of the direct current resistivity method, deduced the distribution of electromagnetic field of the direct current point source and the electric resistivity formula under the inhomogeneous medium. Elaborate the common working methods, kinds of device and the arrangement of different device which will influence the process of the current distinguishing. Analyze the multilayer feed forward structure, transfer function, preforming of the internal training algorithm of the BP neural network. Analyses the main process influencing the BP neural network parameter distinguishing: sample partitioning method, input/output node selection, determination of training algorithm and the set of hidden layer nodes. Finally study each process influencing the BP neural network parameter distinguishing.2) For the numerical simulation, the basic theory of forward numerical simulation was mainly introduced, the characteristic of finite element method and finite difference numerical calculation method was compared and analyzed. Focus on the DC method model based on the 2D forward numerical simulation of finite differential method. Establish the geological response of DC geoelectric model by finite difference numerical model, that means establish the training data base of the nonlinear parameter distinguishing algorithm. This is the basic of the nonlinear parameter distinguishing algorithm. It is also the key of this paper.3) In method aspect, combined with the working principle of the BP neural network parameter identification, from the point of view to optimize the BP neural network parameters, put forward the suitable for dc hybrid optimization identification method of parameter identification: Neural network parameter identification method based on quantum particle swarm optimization(QPSO-BP), parameter identification method of neural network based on immune genetic algorithm(IGA-BP). In the traditional particle swarm optimization algorithm introduced in quantum theory, overcomes the drawback of local convergence of particle swarm algorithm, speed up the search speed of particles inside the space, to facilitate rapid parameter optimum search and identification of BP algorithm, to obtain the best identification model. Immune genetic algorithm to keep the search feature of genetic algorithm, and at the same time introducing the immune mechanism of adaptive characteristics of multiple objective solving, overcome the "premature" of genetic algorithm, combined with BP neural network to form a higher identification precision of recognition algorithm. From the angle of structure defects of BP neural network, puts forward a kind of BP neural network of least squares support vector machine(LSSVM) identification recognition algorithm, least squares support is an improved algorithm based on support vector machine, has better ability of identification of nonlinear and high dimensional recognition.4)In model simulation, the identification effect of horizontal double resistance model and vertical double resistance models are studied, aiming at the QPSO-BP algorithm, compared with the traditional linear identification algorithm and the BP identification algorithm, the geoelectric models to improve the precision of identification are realized, and the validity and feasibility of the QPSO optimization algorithm are proved. According to the IGA-BP algorithm, the identification of the vertical double resistance model and the complex model of the three abnormal body are studied to achieve the high accuracy, fast and stable identification results through quantitative analysis of data and identification of results into a graph. According to the LSSVM algorithm, the horizontal double resistance anomaly model and high resistivity layers with low resistivity body model are studied, the errors of LSSVM method in parameter identification are analyzed and compared, and improve the time and accuracy of DC identification by analysing the LSSVM algorithm quantitatively.5)In engineering application, the parameter identification method of DC electrical method is verified by the instrument of distributing high density resistivity method in the field. Improve the identification accuracy and quality of DC models effectively by reconstructing the identification model by linear identification algorithm for training samples and setting up reasonable algorithm parameters, the results are satisfactory, and the feasibility and effectiveness of the nonlinear identification algorithm are proved by the practical point of view. In conclusion, in this paper, the practice and application of nonlinear identification algorithm based on BP neural network are compacted by the theory and method research, it is verified by the simulation and experiment that parameter identification of DC electric model of nonlinear identification algorithm can effectively improve the resolution and improves the identification effect.The main innovations of this dissertation are as follows:1) A new model of neural network identification model is proposed, which is reasonable to construct the partition method of input and output samples. The model solves the problem of low identification accuracy due to the too much dimension of input samples. Quantum particle swarm algorithm and neural network algorithm intergrate for the first time. The fusion algorithm is applied to the electrical parameter identification, and the introduction of quantum theory can effectively accelerate the particles in the particle swarm search speed, to overcome the shortcomings of neural network to the initial value, effectively improving the identification results of the resolution.2) In the genetic neural identification algorithm in immune system is introduced and applied to identify parameters of DC geoelectric model, introducing immune mechanism to solve the problems of premature convergence of genetic algorithm. Overcoming the neural network is easy to fall into local maxima, significantly improving the accuracy of the identification results.3) Proposed DC parameter identification algorithm based on LSSVM, and using its nonlinear and high dimensional recognition ability, effective solving the BP neural network identification algorithm for structural defects in electrical parameter identification identification accuracy, identify the defects of long time.