Improving the accuracy of neural network solution of the inverse problem of electrical prospecting by sequential determination of parameters: verification on model dataстатья
Аннотация:The inverse problem (IP) of electrical prospecting is a problem of construction of the distribution of electrical conductivity (EC) in an underground area by the components of electromagnetic fields measured on its surface. The sought distribution in its most general form is described by parameters – the values of EC at the nodes of a pre-defined spatial grid, with subsequent interpolation of values between nodes. To describe the distribution adequately, the number of such parameters must be sufficiently large, reaching several hundred even in the two-dimensional case. In their previous studies, the authors considered the solution of the IP of magnetotelluric sounding (MTS) using artificial neural networks (perceptrons). First, the solution of the multi-parameter MTS IP with N determined parameters was performed by its division into N single-parameter problems. Later it was shown that introducing information about EC of higher-lying blocks to the input of the neural network allows increasing the precision of the solution in a number of cases. In this study, the observed effect was verified in computational experiment on artificial model data specified explicitly as complex polynomial dependences of the “observed values” (dependent variables) on the “parameters” (independent variables).