roach to Fuel asis Function Network

This paper proposes a Radial Basis Function (RBF) based approach for the fuel injection control problem. In the past neural controllers for this problem have centred on using a CMAC type neural network with some success. Here we show that an RBF network with a fraction of the size of the CMAC network is capable of delivering superior control performance on a mean value engine model simulation. The proposed approach requires no a priori knowledge of the engine subsystems, and on-line learning is achieved using LMS updates. It is well known that air-fuel ratio (AFR) control is a difficult problem in the automotive control field. It is necessary to maintain the AFR close to its stoichiometric value of 14.64 to ensure vehicle driveability and low pollutant levels, as stoichiometry represents the maximum point of catalytic converter efficiency. Variations of greater than 1% above stoichiometry result in a dramatic increase in CO and HC emission levels, while a decrease of more than 1% involves NO, emissions increasing by up to 50%. Recent and government regulations such as the 1988 California Clean Air Act require significant reductions in HC, CO and NO, emissions. There are also other benefits in accurate AFR control including improved fuel efficiency. The prohibitive cost of in-cylinder sensors capable of measuring the mass of air in the cylinder means that indirect control methods relying on 0thmeasurements must be utili&. The engine is a highly non-linear system, and this precludes the use of conventional linear controllers from achieving the desir of maintaining the AFR within a small of stoichiometry [ 161. Previously, engine control modules @CM> have k e n used to form extensive look-up tables that dictate the amount of fuel to be injected for any engine condition. This method has been shown to be unsatisfactory for the high level of accuracy desired [3] and a new generation of engine controllers are currently being proposed to overme many of the problems encountered by ECMs. Much of the previous research into AFR controllers has centred around model-based controllers (e.g. [2], [14]) where an accurate model of various subsystems of the engine is developed and this is used to determine the mass of fuel to be injected. Model based controllers have significant drawbacks limiting their implementation in a production line scenario. Engine wear and individual engine variances mean that the controller cannot perform at an optimal level over time and large numbers of vehicles. An adaptive technique that can make the necessary adjustments required for an individual engine is clearly preferable. Neural networks have been shown able to represent nonlinear mappings successfully. In particular, RBF networks have been proven to possess the "best approximation property" 1121, that is they can represent a continuous function arbitrarily well. In previous control applications, RBF networks have been successfully applied to a host of problems including adaptive control for bank-to-turn missiles [13] and automated car parking [15]. Thus they have significant potential for the nonlinear AFR control problem. Previously, the only neural control technique applied to the fuel injection problem has been the CMAC neural network [3], [7] with some success. This paper demonstrates that RBFs offer equal or greater promise in the field of fuel injection control.