On the Issue of Simulating Dynamics of a Pneumo-Mechanical Device
Authors: Zelenov M.S., Atamasov N.V., Chernyshev A.V. | Published: 07.12.2018 |
Published in issue: #6(123)/2018 | |
Category: Mechanical Engineering and Machine Science | Chapter: Machine Science | |
Keywords: pneumatic systems, dynamics prediction, neural network, neural emulator, training dataset, mathematical simulation |
The paper considers the stages of developing an artificial neural network (a neural emulator) that predicts variation in parameters of state for a pneumatic spring operating under variable external loads. We present a mathematical model for predicting work cycles that is used to develop neural network algorithms for controlling pneumatic drives. In order to investigate the efficiency of this approach, we developed a program of artificial neural network training based on training datasets derived from numerical experiments using the work cycle model built on the basis of first-order differential equations. We describe the training sample structure and the principles behind generating training datasets. We used the Levenberg --- Marquardt algorithm to train our network and tested the trained neural emulator on the results of series of experiments with random initial parameters taken from a specific bounded domain. We estimated prediction efficiency by an integral criterion: the average coefficient of determination, computed separately for each prediction distance and every model output parameter, such as piston position, piston velocity and pressure difference between the pneumatic spring chambers. Comparative test experiment results show that the data
References
[1] Zelenov M.S., Chernyshev A.V., Krylov V.I. In reference to the neural control system development for position pneumatic drive. Kompressornaya tekhnika i pnevmatika, 2017, no. 4, pp. 28–33 (in Russ.).
[2] Eliseev V.L. Razrabotka i issledovanie neyrosetevykh algoritmov upravleniya statsionarnymi i nestatsionarnymi obektami. Diss. kand. tekh. nauk [Development and study on control algorithms for stationary and non-stationary objects. Cand. tech. sc. diss.]. Moscow, MEI Publ., 2012. 208 p.
[3] Sigeru Omatu, Marzuki Khalid, Rubiyah Yusof. Neuro-control and its applications. Springer, 1996. 255 p. DOI: 10.1007/978-1-4471-3058-1
[4] Gorban A.N. Generalized approximation theorem and computational capabilities of neural networks. Sib. zhurn. vychisl. matem., 1998, vol. 1, no. 1, pp. 11–24 (in Russ.).
[5] Mohajerin N. Identification and predictive control using recurrent neural networks: International Masters Thesis. Sweden, Orebro, 2012. 103 p.
[6] Mandziuk J., Mikolajczak R. Chaotic time series prediction with feed-forward and recurrent neural nets. Control and Cybernetics, 2002, vol. 31, no. 2, pp. 383–406.
[7] Druki A.A. Algoritmy neyrosetevogo detektirovaniya i raspoznavaniya simvolov na slozhnom fone. Diss. kand. tekh. nauk [Neural network detection and recognition algorithms for symbols on complex ground. Cand. tech. sc. diss.]. Tomsk, TPU Publ., 2015. 216 p.
[8] Hao Yu. Marquardt training. In: Industrial electronics handbook. Intelligent systems. Vol. 5. CRC Press, 2011. P. 12-1–12-15.
[9] Zelenov M.S., Chernyshev A.V. [Mathematical model development for positioning electropneumatic drive]. Sb. dokl. Vosmoy Vseross. konf. molodykh uchenykh i spetsialistov [Proc. 8th Russ. Conf. of Young Scientists and specialists]. 2015, pp. 609–610 (in Russ.).
[10] Burakov M.V. Neyronnye seti i neyrokontrollery [Neural networks and controllers]. Sankt-Petersburg, GUAP Publ., 2013. 284 p.