Study On Conductivity Depth Imaging And Inversion Of The Time-domain Airborne Electromagnetic Data Based On Neural Network

Airborne Electromagnetic (AEM) is a geophysical exploration method which use airplane as the delivery too. It has been widely used in lots of domains such as geological survey, mineral exploration, environment inspection and so on. It can be used in many areas where are hard to enter in ground such as forest, desert, swamp, lake and so forth. For the complicated landform and abundant mountains areas, it will be hard to explore on the ground. Therefore AEM as a clipping and economical explorative instrument has a promising application prospect. The academic study of AEM will provide an effective way of large area exploration for multi-metal mine and underground water resources.Based on the existent time-domain electromagnetic method, this disquisition is combined with the scientific research fact of the Natural Scientific Foundation project"The study of Inversion Technology based on nearly two-dimensional time-domain air electromagnetic method with models of integration"and the topic"Time-domain helicopter-borne airborne electromagnetic survey theory research and system design"of the 863 planned major projects"Airborne geophysical prospecting techniques system". The main research work conclusions are as follows:According to the expression of the frequency-domain electromagnetic field, we can compute the transient voltage of the receiving loop, that is computing the time-domain electromagnetic field in the domain of Laplacian first, and then computing the contrary Laplacian by Gaver-Stehfest transform. The time domain airborne electromagnetic responses are computed and analyzed for typical models. We have computed the time-domain airborne electromagnetic responses of the typical geoelectric model using this method and analyzed the outcome. Conclusions are: the time-domain airborne electromagnetic responses declined with time exponentially, in the lower conductivity segment of the half space model, the bigger the conductivity is, the higher the electromagnetic responses will be at the same sample time; in the higher conductivity segment of the half space model, the bigger the conductivity is, the lower the electromagnetic responses will be at the same sample time and the slower the attenuation will be.The time-domain electromagnetic responses of the ambipolar periodic pulse are computed as samples, since the conductivity can not be determined uniquely from the single channel EM data, the EM data that come from the two adjacent channels are then transformed into time constant and amplitude. At the same time, the idea of pseudo-layer makes it possible to minimize the errors introduced by the flight height and topography. Consequently the airborne electromagnetic data are converted into Conductivity Depth Imaging intelligently using fifteen artificial neural networks. We compared the CDI results between the methods of the Neural Network and the lookup-table, conclusions are as follows: the two methods have the same capability of identifying the lower resistance objects, the method of the Neural Network is better at identifying the higher resistance objects, but the curve of the Neural Network is not as slick as the lookup-table. Since the apparent thickness is computed by the experiential rule of the Huang 2008, exact account of the apparent thickness should be researched deeply.Conductivity Depth Imaging is a approximate inversion method, the malpractice of CDI is that it can not form the model approximation of the measured data and imagine under the constraint condition, in order to realize actual inversion, we have studied the half space model inversion using artificial neural networks in this paper, the input of the net are fifteen sets of airborne electromagnetic responses and the output is the apparent height and conductivity, the maximum relative error of the apparent height and conductivity between the test result and the true result is less than 2.5%. Actually, the mother earth is hardly to be approximated by half space model, if it is approximated by layered model, there will be more inversion parameters and more complicated relations, if the layered model inversion is taken using artificial neural network, there will be too many input and output variables that will result in the jam even the non-convergence of the net and will influence the current capability of the net. In order to reduce the numbers of the input and to predigest the neural network's structure, the technology of extracting the eigenvalues from the EM data based on the half space model is studied, which will establish research foundation for extracting the eigenvalues of the layered model.In order to reduce the numbers of the input and predigest the neural network's structures, the main components analysis methods is taken to extract the eigenvalues from the time-domain airborne electromagnetic data in this paper, The eigenvalues, as the input of the neural network, is used in half space inversion which achieves a good performance with the maximum relative error of the apparent height and conductivity between the test result and the true result less than 2.8%.Since the actual measure data would include some noise, we often use some methods such as strong interference elimination, metrical channels smoothing, metrical lines smoothing and so on to wipe off the noise included in the electromagnetic responses data. The noise-removing ability of the technology of extracting eigenvalues is researched in this paper, that is using the eigenvalues of the airborne electromagnetic data which is combined with noise about 5% measured in relative error to train and test the neural network. Conclusions are: the main components analysis methods has a good performance in wiping off the single frequency noise with the maximum relative error of the apparent height and conductivity between the test result and the true result less than 2.8%; the main components analysis methods has the ability of wiping off the multi-frequency noise with the maximum relative error less than 4.2%; the main components analysis methods can wipe off the white noise with the maximum relative error less than 10.4%;The research of Conductivity Depth Imaging of the time-domain airborne electromagnetic data based on Neural Network, the research of airborne electromagnetic data inversion based on the half space model and the research of technology of extracting eigenvalues are completed in this paper. These researches will establish academic foundation for the conductivity depth imaging of the actual metrical airborne electromagnetic data and for extracting the eigenvalues of the layered model, and has practical significance for native airborne electromagnetic research.