Semester

Spring

Date of Graduation

1999

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Mechanical and Aerospace Engineering

Committee Chair

Christopher M. Atkinson.

Abstract

The development of a real-time, on-board measurement of exhaust emissions from heavy-duty engines would offer tremendous advantages in on-board diagnostics and engine control. In the absence of suitable measurement hardware, an alternative approach is the development of software-based predictive approaches. This study demonstrates the feasibility of using in-cylinder pressure-based variables as the inputs to predictive neural networks that are then used to predict engine-out exhaust gas emissions. Specifically, a large steady-state engine operation data matrix provides the necessary information for training a successful predictive network while at the same time eliminating errors produced by the dispersive and time-delay effects of the emissions measurement system which includes the exhaust system, the dilution tunnel, and the emissions analyzers. The steady-state training conditions allow for the correlation of time-averaged in-cylinder combustion variables to the engine-out gaseous emissions. A back-propagation neural network is then capable of learning the relationships between these variables and the measured gaseous emissions with the ability to interpolate between steady-state points in the matrix. The networks were then validated using the transient Federal Test Procedure cycle and in-cylinder combustion parameters gathered in real time through the use of an acquisition system based on a digital signal processor. The predictive networks for NOx and CO 2 proved highly successful while those for HC and CO were not as effective. Problems with the HC and CO networks included very low measured levels and validation data that fell beyond the training matrix boundary during transient engine operation.

Download

Share

COinS