Research On Fast Inversing Imaging For Transient Electromagnetic Method Based On Neural Network

Being a method of time-domain electrical magnetic surveying, the TransientElectromagnetic Method (TEM) that is used to observe and investigate the secondaryfield can greatly simplify investigation of anomaly of buried geologic bodies oncondition that the primary field does not exist. people can not get the picture timely forguiding the job being done, because there is a large quantity of data and the processingis quite complex in latter time.The thesis study modeling of TEM fast-imaging based on ANN, throughresearching methods of TEM fast-imaging, highlights great effect of ANN on TEMfast-imaging. And the main tasks of the paper are presented as followed:1) Briefly introducing basic theory of TEM, and response characteristics incentral-loop. Discuss solution theory of TEM apparent resistivity fast calculation andANN mapping, briefly state application of TEM to grounding grid.2) Calculation method of apparent resistivity based on the nonlinear BP neuralnetwork has been studied. establishing a neural network of some time with the magneticinduction intensity as input and apparent resistivity as output. the effects of trainingalgorithm like Levin Berg Marquardt method, one step secant method on the neuralnetwork have been compared, on the basis of it,comparing the effects of differentneurons contained in hidden layer on the network convergence speed and precision.finally, the optimal BP neural network structure with Levin Berg Marquardt method andten neurons in hidden layer has been obtained.3）Calculation method of apparent resistivity based on BP neural network of thecurve fitting has been raised. Calculation method of apparent resistivity based on thenonlinear BP neural network can achieve imaging purpose by training multi groupneural network. in order to improve the vertical resolution, more training groups have tobe increased, which make training process complex and time-consuming. Calculationmethod of apparent resistivity based on BP neural network of the curve fitting whichmakes the kernel function as input and the transient field parameters as output canovercome the shortcoming. the imaging purpose can easily be achieved just by trainingone network. different neurons in hidden layer and training methods in training havebeen selected, which obtained a optimal neural network with Levin Berg Marquardtmethod and ten neurons in hidden layer. it is simple and high efficient. 4）An optimizational genetic algorithm of BP neural network, namely GABP hasbeen proposed. the BP neural network has low convergence speed and easily fall intolocal minimal value, so the genetic algorithm has been used to optimize the networkweights. the results comparison of the two kinds of neural network dealing withsimulation data of high resistance shows that the convergence speed, accuracy andimaging quality of GABP have been improved. it is demonstrated that the geneticalgorithm can optimize the BP network.4）The GABP neural network and numerical calculation method is applied to thegrounding grid in the experiment, comparing the convergence speed and accuracy. It isverified that the GABP has superiority of rapid imaging on transient electromagnetic.the application of BP and GABP to dealing with simulation datum of grounding gridrespectively, the conclusion of GABP is better than BP in convergence speed,accuracyand imaging quality. the application of BP and GABP to dealing with experimentaldatum of grounding grid shows that the two methods consistently reflect the groundinganomalies, but imaging time for GABP is far less than that for numerical calculationmethod. it has showed the superiority of neural network.