Online Learning of a Neural Fuel Control System for Gaseous Fueled SI Engines

Abstract

This dissertation presents a new type of fuel control algorithm for gaseous fuelled vehicles. Gaseous fuels such as hydrogen and natural gas have been shown to be less polluting than liquid fuels such as gasoline, both at the tailpipe and on a total cycle basis. Unfortunately, it can be expensive to convert vehicles to gaseous fuels, partially due to small production runs for these vehicles. One of major development costs for a new vehicle is the development and calibration of the fuel controller. The research presented here includes a fuel controller which does not require an expensive calibration phase.The controller is based upon a two-part model, separating steady state and dynamic effects. This model is then used to estimate the optimum fuelling for the measured operating condition. The steady state model is calculated using an artificial neural network with an online learning scheme, allowing the model to continually update to improve the controller's performance. This is important during both the initial learning of the characteristics of a new engine, as well as tracking changes due to wear or damage.The dynamic model of the system is concerned with the significant transport delay between the time the fuel is injected and when the exhaust gas oxygen sensor makes the reading. One significant result of this research is the realization that a previous commonly used model for this delay has become significantly less accurate due to the shift from carburettors or central point injection to port injection.In addition to a description of the control scheme used, this dissertation includes a new method of algebraically inverting a neural network, avoiding computationally expensive iterative methods of optimizing the model. This can greatly speed up the control loop (or allow for less expensive, slower hardware).An important feature of a fuel control scheme is that it produces a small, stable limit cycle between rich and lean fuel-air mixtures. This dissertation expands the currently available models for the limit cycle characteristics of a system with a linear controller as well as developing a similar model for the neural network controller by linearizing the learning scheme.One of the most important aspects of this research is an experimental test, in which the controller was installed on a truck fuelled by natural gas. The tailpipe emissions of the truck with the new controller showed better results than the OEM controller on both carbon monoxide and nitrogen oxides, and the controller required no calibration and very little information about the properties of the engine.The significant original contributions resulting from this research include: -collection and summarization of previous work, -development of a method of automatically determining the pure time delay between the fuel injection event and the feedback measurement, -development of a more accurate model for the variability of the transport delay in modern port injection engines, -developing a fuel-air controller requiring minimal knowledge of the engine's parameters, -development of a method of algebraically inverting a neural network which is much faster than previous iterative methods, -demonstrating how to initialize the neural model by taking advantage of some important characteristics of the system, -expansion of the models available for the limit cycle produced by a system with a binary sensor and delay to include integral controllers with asymmetrical gains, -development of a limit cycle model for the new neural controller, and -experimental verification of the controller's tailpipe emissions performance, which compares favourably to the OEM controller.