Machine Learning Techniques for High Performance Engine Calibration

Abstract

Ever since the advent of electronic fuel injection, auto manufacturers have been able to increase fuel efficiency and power production, and to meet stricter emission standards. Most of these systems use engine sensors (Speed, Throttle Position, etc.) in concert with look-up tables to determine the correct amount of fuel to inject. While these systems work well, it is time and labor intensive to fine tune the parameters for these look-up tables. In general, automobile manufacturers are able to absorb the cost of this calibration since the variation between engines in a new model line is often small enough as to be inconsequential for a specific calibration. However, a growing number of drivers are interested in modifying their vehicles with the intent of improving performance. While some aftermarket performance upgrades can be accounted for by the original manufacturer equipped (OEM) electronic control unit (ECU), other more significant changes, such as adding a turbocharger or installing larger fuel injectors, require more drastic accommodations. These modifications often require an entirely new ECU calibration or an aftermarket ECU to properly control the upgraded engine. The problem is now that the driver becomes responsible for the calibration of the ECU for this "new" engine. However, most drivers are unable to devote the resources required to achieve a calibration of the same quality as the original manufacturers. At best, this results in reduced fuel economy and performance, and at worst, unsafe and possibly destructive operation of the engine. The purpose of this thesis is to design and develop--using machine learning techniques--an approximate predictive model from current engine data logs, which can be used to rapidly and incrementally improve the calibration of the engine. While there has been research into novel control methods for engine air-fuel ratio control, these methods are inaccessible to the majority of end users, either due to cost or the required expertise with engine calibration. This study shows that there is a great deal of promise in applying machine learning techniques to engine calibration and that the process of engine calibration can be expedited by the application of these techniques.