Virtual sensors for spark ignition engines using neural networks

Abstract

The overall goal of this project is to design and develop an engine monitoring and control system for spark ignition engines that will help to reduce emissions and increase efficiency. Certain engine parameters are already measured by existing measurement sensors. Other parameters necessary or desirable for intelligent engine monitoring or control are not currently measured, either because those measurements would be too costly or too slow to be of use in real time. The approach is to use the suite of available sensor measurements along with neural networks with online learning capabilities to develop "virtual sensors" for the parameters that are needed but cannot be easily or rapidly measured. The data from these virtual sensors can then be used for performance monitoring and to make intelligent engine control decisions. A general aviation (GA) aircraft engine was used for data collection for this phase of the project. Three virtual sensors were developed in this project. These virtual sensors estimate parameters for pilot aid, diagnostics, and emission monitoring. High quality outputs were obtained for all parameters for normal operating conditions. The estimation errors ranged from /spl plusmn/3% to /spl plusmn/6%. This level of accuracy demonstrates feasibility of the virtual sensor concept for this application.