Support Vector Approaches for Engine Knock Detection
Mr Matthias Rychetsky, Darmstadt University of Technology, Institute of Microelectronic Systems,Karlstr. 15, 64283 Darmstadt, Germany

Abstract: Engine knock is an undesired fast combustion in gasoline engines, which can destroy the combustion chamber. Since this explosions cause vibrations of the engine block, they can be detected by means of signal processing of an acceloremeter signal. As Support Vector Machines are a powerful approach to construct classifiers we applied them to this recognition task.

SVMs have been used in a wide range of problems and have shown there very strong performance. Nevertheless they have some drawbacks: The number of support vectors for a real world problem can be very large (unless you usecomputational expensive reduced set methods), which results in high computational costs for on-line calculations (large number of kernel products have to be evaluated). Furthermore they lag an interpretability. Only from the choice of the support vectors some information can be drawn about the data. A hierarchical approach may help in both cases. It can reduce the number of calculations needed and it extracts some information about the data. This is the motivation why we propose a system which consists of local experts. These experts are only responsible for a small area in the input domain and need therefore not to be very complex. In order to accelerate the system, experts are only placed directly at the class border (for a classification system). The remaining part of the data is classified by a SVM with large margin. Further research is in the moment conducted to use integer approximations to support vector solutions to speed up the recall phase.

This seminar was held at the Department of Computer Science, Royal Holloway, University of London on 8th November, 1999.

back