Neural network model for estimating the quality indicators of industrial fractionator products

For industrial processes, an operational assessment and forecast of the main difficult-to-measure quality indicators of output streams are required. Soft sensors have become widely used for their assessment and prediction in real time. The use of neural networks for their development allows us to take into account the nonlinear features of a technological object and allow us to obtain a more accurate forecast. However, the structure of neural networks is extensive and the choice of the optimal structure, as well as the question of the convergence of learning with a given structure is still beyond the reach of theoretical methods. In this regard, a new neural network learning algorithm and a new approach to determining the structure of a neural network are presented. The latter allows you to establish the structure of a neural network with optimal convergence of the resulting structure. The proposed algorithms are presented in detail on the example of a fractionator and have shown their effectiveness.