апрель 2019

Application of the functional neuro network system for the optimization of energy condensation systems technology

D. S. Yefremov , L. I. Korotka
Химия и современные технологии
Abstract / Full Text

At the present stage of mining technologies development, there is an intention to refuse the industrial use of TNT and TNT explosives. The given problem can be solved by the creation and application of safe technologies of nitrate energy-condensed systems (ECS) [1]. Such systems are characterized by high levels of safety and handling since they consist of non-explosive components.

The industrial introduction of the developed technology revealed the need for adjusting the technological parameters (amount and concentration of the porous solution), depending on the mechanical and strength characteristics of the input raw granules (agrarian ammonium nitrate).

The task of optimizing the ECS technology involves computer simulation of the process, which will allow predicting the parameter of the relative dynamic strength of the granules of the final product, depending on the characteristics of the input raw material and technological parameters. We want to mention, that the main input parameters are the dynamic strength of the granules of the raw material (x1), the amount of the porous solution (x2) and the mass concentration of the porous solution (x3). According to the results of the research, it was established that the target parameter of the strength of the modified granules (the initial parameter - y) changes with time (x4) and acquires optimal values (y = 97-98%) for 7-8 days after manufacturing.

The complexity of the solution of the problem is that the output parameter, in turn, depends on a number of factors whose definition in the conditions of the ECS technology is complicated or impossible.

To achieve the goal, we propose the use of computing intelligence technologies, in particular, artificial neural networks (NN). This approach is due to the fact that obtaining an analytical model, for example, using regression analysis, a full factor experiment, etc., is not necessary. It's enough to get an imitation model that allows you to approximate a function y of many variables. As you know, neural networks are universal approximation systems.

According to the completeness theorem, sigmoid activation functions (AF) are used in the work: logistic (1) and hyperbolic tangent (2).

It is known that the feature of the neural networks use is that the network is able to provide an adequate value of the function that is approximated but does not provide the function itself in analytical form. It is common knowledge that the memory of the designed and trained network is its synaptic weights (matrices of the coefficients of the equations for determining neurons) that are used in the further work with the network.

The simulation stages include the choice of NN architecture, filtering (for "noisy") and normalization of input parameters, determining the volume of the training sample and its formation, determining the parameters and training of NN, checking the trained network. The initial values of the weights and the NN coefficient of training are selected according to the recommendations [2, 3]. The work has been carried out with different models of NN, some results of numerical experiments are given in Table 1.

Table 1 − Results of numerical experiments

Neural network architecture   Acti-vation fun-ction Step of training Required amount of training sample Neural network specified error Relative error for the test sample Amount of training periods
[5-5-1] (1) 0,1 100 0,1 0,0990 3
[5-5-1] (1) 0,5 200 0,05 0,0464 2
[5-9-1] (1) 0,1 140 0,1 0,0968 2
[5-5-1] (2) 0,5 1000 0,01 0,0309 2
[5-9-1] (2) 0,5 100 0,1 0,0933 2
[5-5-5-1] (1) 0,5 300 0,05 0,0503 3
[5-5-5-1] (1) 0,5 1500 0,01 0,0112 3
[5-5-5-1] (2) 0,5 300 0,05 0,1152 3
[5-5-5-1] (2) 0,5 1500 0,01 0,0256 3

It should be noted that the careful preparation of input samples to a large extent contributed to the fact that various networks have been trained in a rather small number of periods, regardless of the activation function, the specified step of training and the specified error. The results of testing a designed and trained neural network have shown that its generalization capabilities fully meet expectations.

  1. Коваленко И.Л. Технология модифицирования аграрной аммиачной селитры в производстве энергоконденсированных систем /И.Л. Коваленко, Д.В. Киященко // Science and education a new dimension. Natural and technical sciences. 2015. III (8), Issue 73. P. 107 110.
  2. Коваленко І.Л. Застосування нейронних мереж для оптимізації технології енергоконденсованих систем /І.Л. Коваленко, Л.І. Коротка// Комп’ютерне моделювання та оптимізація складних систем. – Дніпро: ДВНЗ УДХТУ, 2017. – С. 60-63. (DOI: 10.32434/CMOCS-2017)
  3. Коротка Л.І. Аналіз нейромережевих моделей в задачах оптимізації технології енергоконденсованих систем /Л.І. Коротка// Математичне моделювання. – 2018. – № 1 (38). – С. 69-76.