AVEC ‘06 Comparison of linear and nonlinear system identification

This paper presents simple and practical methodologies for early engine misfire detection. Two diagnostics models, one based on standard linear system identification approaches and a second using a novel nonlinear extension of the linear approaches, involving a multilayer perceptron neural network, were investigated. The models were validated using crankshaft angular velocity measurements taken from a V8 cylinder, 4.2 litre spark ignition engine. The models ’ performance was compared and analysed. This paper shows that standard linear approaches are capable of detecting engine misfire with sufficient accuracy, though they require high model order to accommodate the nonlinear nature of the IC process. Therefore, a nonlinear extension to the model is presented. It is shown that this model extension provides a more compact and simple model structure with better capabilities to handle the nonlinearities. Topics: vehicle diagnostics, modeling and simulation technology, sensors/actuators black-box technolog