A Novel Model-based Predictive Wavelet Network Strategy for Control of Spark Injection Engines
Javad Abdi, Hossein Ahmadi Noubari, Hossein Borhanifar · 2011
An application of wavelet neural network/wavenet in control problems of nonlinear systems is investigated in this study. A wavelet neural network is constructed as an alternative architecture to a neural network to approximate a nonlinear system. Based on approximation capability of wavelet network, a suitable adaptive control law and parameter updating algorithm as applied to nonlinear system uncertainty estimation are developed. It is shown that using wavelet neural networks, an effective uncertainty estimation and control strategy can be obtained. This method improves plant performance effectively and provides robustness against variations caused by changes in operating points of the system. Simulation results show the superiority of the method. Control of the A/F ratio using catalyst converters has been shown to be an effective alternative for the reduction of emission rate of noxious gases such as CO, HC and NOx in SI engines. However due to the presence of uncertainty in plant models as well as time delays and the parameter variations caused by changes in operating point of the engine or environment conditions, an accurate control of the A/F ratio is not always possible. Some of the solutions alternatives cope for this problem are based on using maps of engine bed tests derived for the entire range of operating points. For example, in fuzzy control methods, these maps are applied extensively (6). The solutions to this problem are based on the estimation theory that utilizes identification techniques to determine engine parameter values. However, because of the noise effect and the rapidly changing engine work point even in extended techniques, the estimated parameter values will be associated with considerable error (1, 2). While using nonlinear methods such as sliding mode, may improve the results, however due to the presence of delay and noise in the plant, unwanted oscillations may occur (7, 6). In this paper we will attempt to use wavelet neural network to model the dynamics of the system and to determine an optimum A/F ratio. The method is considered to be a popular control approach to control A/F ratio due to its capability to learn plant model by using, inverse plant maps and uncertain factors in plants (5). Wavelet neural networks or wavelet networks are the neural networks in which the activation functions are considered to be wavelet basic functions (e.g., sigmoid, Morlet, etc.). Nevertheless because of the delay and parameter variations in work point space, this method doesn't guarantee precise and rapid response (8, 9). In this paper, our analysis will start with the introduction of plant model in section 2 followed by the use of both wavelet neural network and controller using sigmoid activation functions. The structure of the controller will be discussed in section 3. Simulation results will be provided in section 4.