| Time-domain airborne electromagnetic detection is a geophysical exploration techniquewhich uses airplane as the delivery. It is precisely because the means of delivery of aircraft inthe exploration process can exploration to ground exploration is difficult to enter the forest,deserts, swamps, lakes, and residential areas and so on, especially for large-scale survey work.Combined with China’s complex terrain, more mountainous, poor working conditions andfactors, time domain airborne electromagnetic detection method has been widely applied inChina.Based on the basic theory of the time-domain airborne electromagnetic method, combinedwith the scientific research fact of the Natural Scientific Foundation project " This article relieson the National Natural Science Foundation of the key technical studies magnifying liked toinverse the time domain airborne electromagnetic" and the topic “Time-domainhelicopter-borne airborne electromagnetic survey theory research and system design” of the863planned major projects “Airborne geophysical prospecting techniques system”. With theartificial neural network and principal component analysis method to analyze the characteristicsof airborne electromagnetic data, and complete time domain airborne electromagnetic datainversion method, The main research work conclusions are as follows:Based on time-domain airborne electromagnetic method, according to the theory ofelectromagnetic induction, analysis of the decay curve of the secondary field emission current,which is characterized by the early data is changing rapidly, and late data changes slowly. Asthe secondary field decay very quickly, if all sampling data inversion, making the inversion oflong time, and therefore its inversion the windowed pumping channel processing. Significantlyreduced by the quantity of data after pumping channel processing, but there is still someredundancy and correlation between these data, half-space earth model, for example, tocalculate the correlation coefficient between the half-space model of21-channel, the correlationcoefficient is about0.99, and there is considerable redundant information, and these redundantinformation has some influence on the inversion results. The principal component analysis is given by the mathematical transformation of a groupof related variables through the transformation to turn into another set of irrelevant variables,the amount of information contained in these variables are different, the study of thecontribution rate in the principal component can be seen that the greatest contribution rate ofthe first principal component, the contribution rate of the second principal component followed,and so on. The contribution rate of the principal components in the inversion is different, byprincipal component reconstruction method to study the number of principal componentsrequired in the inversion, with the root-mean-square relative error and the cumulative variancecontribution rate, these indicators considered select the number of principal components.Half-space earth model, for example, to prove the case with noise, the relative error of the meansquare signal to noise ratio with to participate in reconstruction of the changes of the number ofprincipal components, a trend can be seen through the principal component data reconstructiondenoising capability, and can improve the signal to noise ratio of3-5dB.Real inversion, the exploration of ground conductivity distribution is unknown, majorproblems for a sample set of the neural network, so targeting for neural network training, withthe inversion based on neural network time-domain airborne electromagnetic data toconductivity depth imaging (CDI),21-channel time domain airborne data neural network as aneural network input variables,20pairs of pods depending on the height and the earth, as theconductivity as a neural network output, respectively, to train20networks, by testing samplesfor testing and evaluation of the network,20network as the network of imaging required toselect the test output and the desired output, the maximum relative error is less than6%, theinversion CDI imaging data by the20networks. The effect of the mapping as the initial modelof the earth, then the initial model parameters needed to build inversion sample set, prepare forthe inversion.Due to the neural network input variables will lead to a bigger difference in the neuralnetwork inversion. Therefore, in the neural network inversion, respectively,21-channelairborne electromagnetic data and a number of principal components as input of neural network,based on data and based on principal component two neural network of the inversion. In thestudy, a single two-story earth model and a single there-story earth model, for example,inversion of the initial earth model with CDI imaging technology, the initial model as abenchmark to establish the sample set. Construction of the sample set of three-layer model, theselection and extraction of the number of principal components and neural network training andtesting to understand based on the principal component of the neural network training, not onlyreduces the neural network input variables and to a certain extent simplify the structure of thetraining network, reduces the computation of the neural network, making the network more vulnerable to convergence. In order to simulate the actual avionics data, plus5percent of thewhite noise in the forward data on research. A single two-story earth model, for example,inversion based on the data and based on principal component two neural network methods,respectively, with no noise and5%noise data inversion. In the noiseless case, the principalcomponent-based inversion relative based on the inversion of the data on the number of steps ofthe training network has a great advantage, the advantage is not particularly evident on theinversion results. Containing5%noise case, the six principal components to reconstruct dataafter adding noise, the calculation of the reconstructed data is relatively noise-free data the rootmean square relative error of2.97%, but increases the noise data relatively free of theroot-mean-square relative error of the noise data was11.74%, which can be seen that theprincipal component analysis to a certain extent to reduce the impact of noise on the originaldata. Inversion based on the inversion of the principal component in the training steps stilloccupy a great advantage, and the inversion effect was significantly better than data-basedinversion.In order to further simulate real inversion, a series of3-layer models to connect thesimulation to a low resistance laminated quasi-two-dimensional earth model with a tilt toachieve the variety of this neural network model inversion. Earth model inversion with a singletwo-story, with5%noise case, relatively noise-free data, based on principal componentinversion reconstruction data to improve the root mean square relative error of7.5%. And in thecase of noise presence or absence of the inversion effect based on principal component are bestthan the based on data inversion, and the noisy case is more obvious. In order to confirm thepractical significance based on principal component analysis of the neural network inversion,inversion through field simulation experiment on the measured data, and effectively proved thepractical significance of the principal component analysis.In this paper, principal component analysis, based on principal component neural networkinversion and inversion to prove a principal component analysis, dimensionality reduction anddenoising. By contrast inversion with neural network-based data to prove the significance ofprincipal component analysis neural network inversion of these studies will be measuredairborne electromagnetic data inversion to provide a theoretical basis, the airborneelectromagnetic survey has some practical significance and academic value. |
PDF Full Text Request