V. I. Vasilyev, I. F. Nugaev

One of the major problems solved at building of oil and gas chinks, the continuous control of a condition of a face of a chink is. For the decision of the given problem continuous measurement of some geophysical parametres, such as azimuthal and antiaircraft corners of a tangent to a trajectory of a trunk of a chink in a face point is carried out, level of a natural radiating background, specific electric resistance разбуриваемой breeds, etc. Feature of measurement of the specified parametres is presence of distortions and noise of the measurements caused by intensive with great dispatch-vibrating loadings which test deep measuring converters in the course of drilling. In this connection, one of actual problems is the estimation of real geophysical parametres on the data received as a result of gaugings. Approaches put into practice to the decision of the given problem [1], as a rule, are connected with averaging of the re-peated gaugings made in separate points of a trajectory of a chink. However the given way often appears inefficient owing to limitation of number of dot gaugings and high level of a dispersion of the received data. In article the way of an estimation of the geophysical parametres, based on a principle of smoothing of the measured data is considered. The principle is based on the analysis of some the measurements made consistently on length of a trajectory of a chink. The algorithm multilevel регуляризирующего synthesis нейросетевой is offered smoothing model with application of vector criterion of the smoothing, providing consecutive definition of a class and model parametres. Advantage of algorithm is increase of efficiency of use of generalising properties of a neural network by stage introduction регуляризации at level of a class of model, and also an exception of a problem of a choice of parametres регуляризации the generalised criterion. The developed way of realisation of the given algorithm on the basis of a RBF-network, includes hierarchical ranging of parametres of a neural network and a way of an estimation of their choice on the basis of the offered vector criterion. The spent experimental researches have shown high efficiency of the developed algorithms and their advantage before traditional single-level approaches from the point of view of increase of adequacy of smoothing model.