A Study On EFI Engine Fault Diagnosis Technology Based On Neural Networks

In this thesis, we studied the feasibility of using artificial neural network (ANN)models to develop model-based techniques for failure diagnosis in spark ignition (SI)engines. Due to the preliminary nature of this study, we just proposal an approachwhich is based on RBF networks to diagnosis the man-made malfunction. RBF neuralnetworks have recently drawn much attention due to their good generalization abilityand a simple network structure that avoids unnecessary and lengthy calculation ascompared to the Multilayer Feed forward Networks (MFNs). For an RBF classifier,especially with Gaussian function as its radial basis function, the network learns thepattern probability density instead of dividing up the pattern space as MFNs do.Therefore, when an out-of-category pattern is evaluated, such an RBF network is likelyto classify it as an unknown category. In addition, the widely used BP trainingalgorithm for MFNs is often too slow, particularly in the case of large size problems.Since RBF network can establish its parameters for hidden neurons directly from theinput data and train the network parameters by the way of linear optimization, it isgenerally much faster, compared to MFNs, to complete the training.The function between exhaust gas and engine failure are highly nonlinear, and as aresult are difficult to model. In this thesis, data from a DongFeng EQ6102engine, wereused to develop ANN based models for the above systems. The models were then usedin a decision-making process to diagnosis engine faults. This study revealed that theANN had great potential for developing techniques for fault diagnosis. We believe thatANN-based techniques are superior in general to other engine diagnostic approaches.Since the former have the potential for systematically diagnosing a variety of softincipient failures under different engine operating conditions. Other potentialadvantages of ANS-based techniques include low cost，reduced need for enginemapping and calibration，and potential transferability of the FDI concepts from oneengine class to another.