Neural Network Applications for Model Based Fault Detection with Parity Equations

The rising complexity of modern automotive engines with an increasing number of actuators and sensors to minimise emissions and fuel consumption and to maximise engine driveability require a detailed supervision for fault detection and on-board diagnosis. The European Community Directive 98/69/EC requires on-board diagnosis for spark ignition engines and will require it for diesel engines as of January 2003, mainly to prevent excessive emissions. Beside this regulation it is also in the interest of the automobile manufactures to establish capable diagnosis systems for maintenance, repair and the benefit of their customers. This paper will describe applications of neural networks for modelling complex fluidand thermodynamics with unknown physical model structure. Reference models, which describe the fault free process, are set up and identified with the special neural network LOLIMOT (Local-Linear-Model-Tree). Fault detection algorithms, which employ the method of parity equations, were successfully implemented and tested in real time with a 2 litre diesel engine and a Rapid Control Prototyping System. Measurements of online fault detection are shown for several built-in faults in the intake system of this diesel engine. Copyright © 2002 IFAC