Research On Neural Network Algorithm For Inversion Of Fault Parameters

Geodetic inversion is a marginal subject that uses geodetic observation technology and nonlinear inversion algorithms to study the laws and characteristics of the earth's surface and its internal changes,and promote the interpretation of the energy process within the earth.Geodetic inversion includes many aspects,and seismic fault parameter inversion is an important part of it,and it is also the focus of this thesis.Using geodetic inversion algorithms to invert seismic fault parameters is conducive to analysis,structure and reconstruction of seismic fault structures.And it is helpful to further study the mechanism and damage degree of earthquakes.Geodetic inversion methods are divided into two categories according to whether to use derivative information,linearization algorithms and nonlinear algorithms.Due to the complicated process of linearization algorithms,nonlinear algorithms are commonly used to invert seismic fault parameters.Non-linear algorithms include genetic algorithm,simulated annealing algorithm,and neural network algorithm and so on.This article focuses on the application research of NNA in seismic fault parameter inversion.In order to obtain more accurate fault parameter inversion results,this paper studies the NNA of fault parameter inversion.The main research content and work of this paper are as follows:1)The theoretical advantages and disadvantages of NNA inversion of fault parameters are studied.The global positioning system(GPS)data can be used to obtain more reliable parameter solutions when NNA inversion fault parameters is analyzed.However,it has a narrow range of application for different types of seismic faults.In order to improve the accuracy of fault parameter inversion,this paper analyzes and improves NNA.Considering the special structure of NNA,its inner part is connected by the single layer connection weight.Since the single layer weight cannot completely transmit information in the information transmission,this paper considers to add one layer weight on this basis to form a two-layer neural network to make it more complete in the process of internal information transmission.Nelder-Mead simplex algorithm has strong local search ability,but its dependence on the initial solution is strong.However,the target solution generated each time in the NNA is a better parametric solution.Therefore,the target solution in the NM optimized NNA is considered as the second layer weight of NNA to optimize the NNA.Considering that the starting point plays an important role in the optimization algorithm,the starting point strategy is adopted in this paper to generate the initial starting point closer to the actual situation,which is beneficial to the algorithm to find the optimal parameter solution faster.The Nelder-Mead neural network algorithm proposed in this paper is verified by simulating different GPS monitoring points and experiments of different fault types.The experimental results show that compared with the existing algorithms,NMNNA has stronger applicability,more stable inversion results,higher inversion accuracy and faster computational efficiency.The precision advantage of the multi-starting Nelder-Mead neural network algorithm is further verified by the Bodrum-Kos earthquake.2)Through the study of GA,NM and NNA algorithms,we found that the genetic characteristics of GA were mainly derived from selective crossover operation,and the genetic characteristics could retain excellent genes and not be lost easily.Since NM has been used to optimize the local search ability of NNA in previous studies,this paper considers the use of GA's selective crossing-over operation to optimize the ability of NM-NNA to retain the better solution,so that the algorithm is not easy to lose the optimal parametric solution.In order to verify the rationality of GNMNNA proposed in this paper,three combination schemes are designed in this paper.The second is genetic simplex algorithm,the third is the genetic simple form of spiritual network algorithm.Through simulation experiments that the GA,NNA and three improved algorithms are compared,and the experimental results show that the stability of GNNA and GNMA is poor.On the contrary,the inversion stability of GNMNNA is better than other algorithms(GA,NNA,GNNA and GNMA),and GNMNNA has high precision parameter solution and strong stability.Moreover,the inversion precision of GNMNNA was tested by Bodrum-Kos earthquake and Monte Cristo Range earthquake,and the experimental results show that the deviation of GNMNNA obtained is small and the accuracy is high.3)The precision evaluation method of nonlinear inversion algorithm is studied.It is found that the precision evaluation strategies of the current nonlinear inversion algorithms are all similar,and the parameter deviation and standard deviation are solved by some sampling strategy,and the high and low precision of the algorithm is evaluated according to the parameter deviation and standard deviation.However,the deviation or mean-square error of the separate comparison parameters is not comprehensive enough,and there are certain unreasonable problems.Because the deviation calculation method is calculated from the difference between the estimated value and the mean value,the result of the deviation calculation cannot always reflect the precision of the parameter.Because the mean value is a general concept,a set of similar results in the parameter estimation and the mean value of a group of parameters with a large difference may be the same,which will cause an imbalance in the precision evaluation.In addition,considering the parameter standard deviation alone is also somewhat one-sided.In the parameters of different algorithms,the precision of one algorithm's parameter one is higher than the corresponding parameter precision of algorithm two,but the parameter two is lower than the corresponding parameter precision of algorithm two.Therefore,in this case,it is obviously impossible to conclude which algorithm is more accurate.Therefore,this paper gives a dimensionless formula for precision evaluation that takes into account the information of parameter deviation and standard deviation reasonably.Based on the volcano CDM and the earthquake Okada model,the precision of the two improved NNAs is evaluated.The experimental results show that the precision of G-NM-NNA is higher than that of NM-NNA.