Development of neural network model of the multiparametric technological object

Abstract

At present, there are a large number of methods for identifying the technological objects on the basis of data of their industrial operation [1-3]. The most promising direction is the construction of a model, which will allow to take into account the multifactorial nature of the object, and the nonlinearity of interrelation between variables. This will make it possible to control the object, taking into account the change in its states, and based on the current data,to predict the change in the output value with different input characteristics[4-6]. All this will provide the opportunity to create an operating system, based on the currently measured technological indicators. In order to implement this approach, a comparative study of the regression analysis models, using polynomials of various types and neural network algorithms, for the synthesis of a complex technological unit model, was carried out in the work. In the regression analysis, the following models were investigated: polynomials, linear, fractional and exponential functions, Kolmogorov-Gabor polynomial. In the process of the research of neural networks to solve this problem, their structure was varied, with subsequent learning according to the Levenberg-Marcardt algorithm. In the process of simulation of the object models in the Matlab package, the degree of similarity of the outputs for each of the obtained models and the actual output of the object were estimated. Quadratic criterion and the coefficient of correlation were calculated, that made it possible to judge the accuracy of the constructed models. The best structure of the model was established for identifying a complex multiparameter object, using the example of statistics for the operation of a ball mill.It was a network with three hidden layers and 50, 35 and 25 neurons in them, with activation functions, respectively by layers - hyperbolic tangent, sigmoid function in 2 layers, and a linear activation function in the output layer. The vector, including 15 parameters, was supplied to the network input: the volume of ore supply to the mill, the volume of water supply to the mill and the mill’strommel, the signals with the first-, the second-, and the thirdorder lags, and the signal of current with the first-, the second-, and the third-order lags. This approach to identification has increased the accuracy of the object model, that ultimately will affect the quality of the developed control system of the unit as a whole, allowing to improve the quality of the ball millcontrol.