Open AccessUSE OF A ROLLED DEEP NEURAL NETWORK TO DETERMINE THE CHANGE OF THE STRESSED AND DEFORMED STATE OF VERTICAL STEEL CYLINDERAL TANKS BY THE MOVEMENT OF THEIR SURFACE
Author(s) -
Yu. V. Pankiv,
Kh. V. Pankiv
Publication year - 2020
Publication title -
metodi ta priladi kontrolû âkostì
Language(s) - English
Resource type - Journals
eISSN - 2415-3575
pISSN - 1993-9981
DOI - 10.31471/1993-9981-2020-2(45)-5-12
Subject(s) - artificial neural network , strain gauge , structural engineering , calibration , position (finance) , stress (linguistics) , barkhausen stability criterion , computer science , mechanical engineering , engineering , artificial intelligence , mathematics , magnetic field , magnetization , linguistics , statistics , philosophy , physics , finance , quantum mechanics , economics
The main place in determining the mechanical characteristics of the RVS material is occupied by methods and means of VAT control, which include the method of coercive force; magnetic anisotropy method; Barkhausen method, magnetic metal memory (MPM) method, strain gauge method. The advantages of magnetic methods are the ability to control without decommissioning the RVS, safety. The disadvantages inherent in almost all magnetic control methods include the need to prepare the controlled surface, the difficulty of determining the position of the sensors in relation to the maximum loads, the dependence of control results on methods and conditions of measurement, the influence of the air layer between the sensor and the controlled surface.It's shown that to determine the further safe operation possibility of vertical steel cylindrical tank it is necessary to know their stress-strain state. The shortcomings of the existing experimental and mathematical methods of its estimation were highlighted. It's proposed to use a convolutional deep neural network to determine the stress state of a vertical steel cylindrical tank. As input data, it's proposed to use data on the movement of its wall obtained, for example, as a result of geometric calibration at two points in time. The input data was presented in the form of an array of dimensions 8x12, then used convolution and max-pulling. The last layer is fully connected. It's proposed to use cross-entropy as a cost function. To increase the amount of training data, it is proposed to use the values of displacements on stresses obtained by modeling different effects on a cylindrical tank with different shape defects using the SolidWorks package. Possible ways to improve the proposed method are proposed.For further research and improvement of the proposed method, you can try to use other hyper parameters in the neural network, in particular to change the number of feature maps, the size of the local receptive field, the size of the shift step of the receptive field and others.You can also try using the source layer with a different number of neurons and softmax as a function of cost.