The progressive reduction in permissible tailpipe emissions levels from automobiles has been achieved through the adoption of ever more complex engine control systems and aftertreatment components. This, in turn, has resulted in the development of increasingly sophisticated monitoring systems that can detect the failure or gradual degradation of any of these components and thereby fulfill the requirements of the stringent On-Board Diagnostic (OBD) legislation.
Traditional monitoring techniques involve a physical model approach, which describes the system under investigation. This approach has limitations, such as available knowledge base and computational load. Neural networks, on the other hand, have been recognized as a powerful tool for modeling systems which exhibit nonlinear relationships between measured variables, such as internal combustion engines. A particular type of neural network, the auto-associative neural network, has provided high accuracy in modeling non linear systems as well as with fault detection in such systems. However, neural networks have an unavoidable drawback because of the computational load required when the model is developed.
The research described in this paper proposes the use of physical models to pre-process the measured engine variables. These physical models transform the original measured variables into other calculated variables which have physical meanings. This reduced number of variables are then used to develop a neural network model of the system under investigation. This preprocessing improves the accuracy of the overall model and reduces the burden of building a neural network model significantly. The theory and use of the technique is shown in this paper with the application to the detection and diagnosis of air leaks within the inlet manifold system of a modern gasoline engine.