ON NONPARAMETRIC MODELING ALGORITHMS OF BOF PROCESS

Abstract

The problem of data pre-processing in the identification of multidimensional discrete-continuous processes is considered. The main content of the paper is devoted to the method of generating working training sample from the initial one, represented by the data of the object normal operation. This step is very important in the non-parametric identification of discrete-continuous processes. Non-parametric identification algorithms belong to the class of local approximations of unknown stochastic dependencies. In nonparametric identification the step of selecting an object model to the accuracy up to the parameter vector is absent. This approach takes place in the variety of real problems, because the priori existing information is not enough to determine the reasonable parametric model structure. The procedure presented below is similar to butsrtap based on the initial training sample, which reflects the characteristics of the identified object. Numerous computational experiments carried out by statistical modeling have showed high efficiency of generation techniques discussed below which is laid into the foundation of the adaptive system modeling. In addition, it can automatically solve the problem of restoration an unknown stochastic dependence on the definition boundary of the relevant input-output object variables. The following technics and algorithms of nonparametric recovery stochastic dependencies were used to study the oxygen-converter process. A sample of observations made from passports of 176 low carbon oxygen steel melted by the contract at JSC “EVRAZ ZSMK” oxygen-converter workshop No. 2. New working sample which contains both the measurements and the generated data was formed according to the proposed methodology. Using the working sample makes it possible to increase the accuracy of the training simulation in 2–3 times.