Neural network model of internal combustion engine
- PDF / 891,620 Bytes
- 10 Pages / 595.276 x 793.701 pts Page_size
- 35 Downloads / 273 Views
NEURAL NETWORK MODEL OF INTERNAL COMBUSTION ENGINE UDC 621.436-55
S. A. Serikov
Abstract. A mathematical model of internal combustion engine is obtained based on the approximation of its speed characteristics, efficiency, and toxicity of exhaust gases with the use of a feedforward artificial neural network. The dependence of the approximation error of engine characteristics on the structure and parameters of the model is investigated. Keywords: artificial neural network, internal combustion engine, identification, fuel efficiency, toxicity of exhaust gases INTRODUCTION Increasing the efficiency and ecological safety are the main trends in vehicle perfection. The features of working processes in internal combustion engine (ICE), which remains the most popular power unit for vehicles moving without external supply of energy, have a strong effect on the fuel efficiency and exhaust gases (EG) toxicity. In solving problems of the analysis and optimization of engine processes for specified fraction-speed modes of vehicles, identifying the ICE mathematical model is a key problem. ICE mathematical models as systems of differential equations based on the analysis of thermodynamic processes in the combustion chamber, intake and exhaust manifolds, features of the interaction of mixture and ignition systems, kinematic relations among various nodes and units, etc. appear extremely complex and unwieldy [1, 2]. The linearization of these equations significantly increases the residual of the outputs of the model and the object if the model should be used in a wide range of ICE operation modes. Mathematical models in the form of Volterra series have not become popular in the analysis of ICE modes since determining Volterra kernels involves a great volume of expensive experimental studies. Mathematical models obtained by approximation of experimental characteristics of the ICE taken during bench tests appear to be more preferable. Polynomial approximation of static characteristics is in most common use. Meanwhile, different sections of the domain of permissible power settings are approximated with individual polynomials, which makes the mathematical model rather unwieldy [3]. Considering the high cost of experimental data, their limited amount, high noise level, incompleteness, and frequent inconsistency, models based on artificial neural networks (ANN) appear the best. They are selectively sensitive in data concentration domains and smoothly interpolate in other domains [4, 5]. PROBLEM STATEMENT By an ICE mathematical model we will mean an operator F: U a Y that associates each input vector éw min £ w £ w max ù éwù & & £w & max ú êw &ú êw £w u = ê ú , u Î U Ì Ñ4 , "t; U = ê min ú b 0 £ b £1 ê& ú ê& ú & & êë b úû êë b min £ b £ b max úû
(1)
Kharkov National Automobile and Highway University, Kharkov, Ukraine, [email protected]. Translated from Kibernetika i Sistemnyi Analiz, No. 6, pp. 155–165, November–December 2010. Original article submitted May 5, 2009. 998
1060-0396/10/4606-0998
©
2010 Springer Science+Business Media, Inc.
Data Loading...
