Technological institute East-Ukraine National University of the name of Vladimir Dal (the place of Severodonetsk)
Abstract: The present article focuses on the researches’ results ofartificial neural networks architecturing using in automated control systems for acetic acid synthesis unit by means of GUI, i.e. Neutral Network Toolbox interface of the software simulator Matlab. The structural schemes of such systems are attached.
Keywords: neural network with feed-forward back propagation; radial-basic neural network; activation function.
To run complex systems, it is necessary to build a model adequately reflecting the properties of the object to be controlled. In the majority of cases such model parameters are determined directly within the process of object is in operation, i.e. the identification is performed on the basis of occasional input and output signals. Nowadays the way of automated control systems architecturing on the basis of artificial intelligence technologies usage (e.g. neural networks, fuzzy logic, genetic algorithms) (Karimov R.N., Galushkin A.I., 2000; Kruglov V.V., Borisov V.V., 2001; Kruglov V.V., Borisov V.V., Ossovsky S., Medvedev V.S., Potemkin V.G. , 2002; Haykin S. , 2006, Ulshin V., Yurkov D. 2010, Kusz A., Maksym P., Marciniak A.W.2011) is being rapidly developed. Those factors which are badly formalized using common mathematical methods may be subjected to generalization (e.g. one’s professional experience or intuition, etc.). Only few attempts are known to use artificial intelligence technologies in chemical industry. They are used to interpret sensors’ readings, to run temperature mode of the technological processes, to monitor chemical and technological processes (Gardner J.W., Hines E.L., Wilkinson M., 1990; Baskin I.I., Skvortsova M.I., Palyulin V.A., Zefirov N.S., 1997; Topolski N.G., Vatagin V.S., H.-K. Hong, C.N. Kwon, S.-R. Kim, 2000; Zhang H., Balaban M.O., Principe J.C., 2003; Baskin I.I., Palyulin V.A., Zefirov N.S., Vatagin V.S., Nevsky A.V., 2005; Baskin I.I., Palyulin V.A., Zefirov N.S., 2006; Sysoyev V.V., Musatov V.J., Silayev A.V., Zalialov T.R., Maschenko A.A., 2007; Afanasenko A.G., Verevkin A.P., 2009, Park, J.-K., 1993, Nagy, Z., American Institute of Chemical Engineers., et al., 2000).
Architecturing and researching of artificial neural networks performance can be carried out via software-based simulators. The most commonly used packages to model neural networks characteristics are as follows: Neural Works Pro Plus, Neuro Solution, Matlab (Neural Network Toolbox), Neuro Wisard, ANsim, Neural Ware and others. The software is differed by its complexity, quantity of neurons types and algorithms of studying maintained at the system.
MATERIALS AND METHODS
The purpose of the present work was the building up and researching of the neural network properties which might be used in running of acetic acid synthesis unit at start-up at PJSC “SEVERODONETSK AZOT ASSOSIATION”. For this purpose the statistical data of the acetic acid synthesis unit while the plant starting up process were used. The environment of software-based simulator MATLAB 7.1.0 (GUI - Neural Network Toolbox interface) was used for building up process. This pack is recommended for neural network with different type of activation function architecturing.
The following parameters were used as the input data:
1. Inlet methanol consumption.
2. Inlet carbon oxide consumption.
3. Inlet methanol temperature.
4. Inlet carbon oxide pressure.
5. Inlet carbon oxide temperature.
The following parameters were used as the output data:
1. Reaction mass level at the unit.
2. Reaction mass pressure at the unit.
3. Reaction mass temperature at the unit.
50% of the main observations were involved in the process of the neural network learning (the other half was used for verification purposes).
The structure and parameters of the neural network with feed-forward back propagation shown at Fig.1 were determined as a result of GUI - Neural Network Toolbox interface MATLAB 7.1.0.
Fig.1. Neural network with feed-forward back propagation structure.
The network was built up on the basis of five neurons at the network input, five sigmoid (TANSIG) neurons of a buried layer and three linear (PURELIN) neurons of an output layer. The functions implementing learning algorithm as well as training and error functions ensuring minimum relative accuracy of data approximation were determined to build up the network. The respective parameters of three networks for the acetic acid synthesis units are given in Tab. 1.
While making a comparison of the networks respective parameters, the following observations were found: the minimum relative accuracy is insured by the network based on gradient descent method with disturbance as a learning function, function of the gradient descent with account of the moments as a training function, mean-square error as an error function.
Tab. 1 Neural networks parameters with feed-forward back propogation
Learning algorithm implementing function
Average relative accuracy
Gradient descent method
Learning function of gradient descent with account of the moments
Gradient descent method with account of the moments
Learning function of gradient descent
Learning function of gradient descent
In addition, a radial-basic neural network which structure is shown at Fig.2 was built up as a result of the research.
The parameters of the input and target values arrays, as well as GOAL (network tolerated mean-square error) and SPREAD parameters (the parameter of interference) were used as the input data for the radial-basic networks, while radial-basic network parameters were used as output data. SPREAD parameter of interference was taken to be bigger than a partition step of the learning sequence interval, but smaller than the interval itself, that is equal to 0.1. GOAL parameter was chosen to be equal to 0. While architecturing of the radial-basic network with a zero error, the number of neurons of a radial-basic layer is equal to the number of input values. Weight and bias of the radial-basic network are set in such a way that its outputs are accurately equal to the targets. Relative accuracies of the data fitting were determined as a result of networks forecasting by means of a testing data selection. The respective parameters of the network with a minimum relative accuracy for acetic acid synthesis unit are given in Tab. 2.
Fig.2. Radial-basic network structure.
Tab. 2. Radial-basic neural network parameters
Average relative accuracy
If we compare the relative accuracies of feed-forward neural networks to those of a radial-basic network, we can see that the minimum relative accuracy is insured by the radial-basic neural network, that is one of the networks has an advantage before another in solution of the management problem. There’re much more neurons in the radial-basic networks than a compared network with feed forward signal and sigmoid activation functions at a buried layer has.
Tocreate control system with the usage of the neuron networks at the acetic acid synthesis unit start-up process it’s necessary to determine neuron network structure, to hold a learning on the basis of technological specifications and to do the approbation of network performance with an application of acetic acid plant equipment. Neural network architecturing with a usage of GUI - Neural Network Toolbox interface MATLAB 7.1.0 has proved the success of the neuron network building up and learning process and its satisfactory quality. That will let use neural networks to manage the technological processes of the acetic acid synthesis and proves the urgency of the further researches of this area.
1.Afanasenko A.G., Verevkin A.P., 2009, Neuronet simulation of carbonation process quality rating // Vestnik U, - Vol. 13 ¹2 (35). p.222-225.
2. Baskin I.I., Palyulin V.A., Zefirov N.S., 2006, Multilayered perceptrons in “structure-property” relationship studies for organic compounds. // Rus. Chem. Journ. ( Journal of Russian chemical society after D. I. Mendeleyev).- T. L. ¹2. - p.86-96.
3. Baskin I.I., Palyulin V.A., Zefirov N.S., 2005, Using artificial neural networks for chemical compounds properties prognostication. // Neural computers: design, applications. – ¹ 1-2. - p.98-101.
4. Baskin I.I., Skvortsova M.I., Palyulin V.A., Zefirov N.S. ,1997, Quantitative chemical structure-property/activity relationship studies using artificial neural networks. // Foundations of computing and decision sciences. –Vol. 22, ¹2. - p.107-116.
5. Galushkin A.I., 2000,The book, Theory of neural networks. – M.: I, 416 p.
6. Gardner J.W., Hines E.L., Wilkinson M., 1990, Application of artificial neural networks to an electronic olfactory system. // Measurement science and technology. –Vol. 1, p.446-451.
7. Haykin S., 2006, The book, .Neural networks. Complete course. - M.: Williams 1104 p.
8. Hong H.-K., Kwon C.N., Kim S.-R., 2000, Portable electronic nose system with gas sensor array and artificial neural network. // Sensors and Actuators B. –Vol. 66, p.49-52.
9. Karimov R.N., 2000, Experimental information processing. P.3: Multivariate analysis: learning aid, Saratov: SSTU, , 108 p.
10. Kruglov V.V., Borisov V.V., 2002, The book Artificial neural networks. Theory and practice.-2nd edition, pattern, - M.: Hot line – Telecom, 382 p.
12. Kusz A., Maksym P., Marciniak A.W.2011. Bayesian networks as knowledge representation system in domain of reliability engineering. TEKA Commission of Motorization I Energ. Roln., 11 C, p. 173-180.
13. Medvedev V.S., Potemkin V.G., 2002, The book ,Neural networks. MATLAB 6 /Under the editorship of Cand. of Sc. Potemkin V.G - M.: DIALOG IEFI, 495 p.
14. Nagy, Z., American Institute of Chemical Engineers., et al., 2000, A Comparison of First Principles and Neural Network Model Based Nonlinear Predictive Control of a Distillation Column.// Distributed by American Institute of Chemical Engineers, New York, N.Y.
15. Park, J.-K., 1993, "Modeling of Distillation Column and Reactor Dynamics Using Artificial Neural Networks (Neural Networks)."// DAI/- 55, ¹. 01B, 6660.
16. Sysoyev V.V., Musatov V.J., Silayev A.V., Zalialov T.R., Maschenko A.A. ,2007, The application of neural networks method for response analysis of the gas identification multisensory system-on-chip. // Bulletin of Saratov State Technical University, - ¹1 (21) – issue 1. – p. 80-87.
17. Topolski N.G., Vatagin V.S. ,2000, Computer-aided fire safety systems in chemical industries // Mary Kay O’Connor Process safety center symposium. Proceeding.- October 24-25.- Reed Arena. – Texas A&M. –University, College Station, Texas. – p. 348-349.
18. Vatagin V.S., Nevsky A.V., 2005, Neurotechnologies of designing integrated automated control systems of industrial safety // State-of-the-art high technologies. - vol. 4. p.75-81.
19. Zhang H., Balaban M.O., Principe J.C., 2003, Improving pattern recognition of electronic nose data with time-delay neural networks. // Sensors and Actuators B. –Vol. 96, p.385-389.
20. Ossovsky S. 2002. Neural networks for data processing /Translation from Polish by Rudinsky. - M.: Finances and statistics, 304 p.
21.Ulshin V., Yurkov D. 2010.The adaptive system on the basis of artificial neuron networks. TEKA Commission of Motorization I Energ. Roln., 10 D, p. 15-24.