RANDOM FOREST BASED MISFIRE DETECTION USING KONONENKO DISCRETISER

This paper evaluates the use of random forest (RF) as a tool for misfire detection using statistical features. The engine block vibration contains hidden information about the events occurring inside the engine. Misfire detection was achieved by processing the vibration signals acquired from the engine using a piezoelectric accelerometer. The hidden information regarding misfire was decoded using feature extraction techniques. The effect of Kononenko based discretiser as feature size reduction tool and Correlation-based Feature Selection (CFS) based feature subset selection is analysed for performance improvement in the RF model. The random forest based model is found to have a consistent high classification accuracy of around 90% when designed as a multi class ,ode and reaches 100% when the conditions are clubbed to simulate a two-class mode . From the results obtained the authors conclude that the combination of statistical features and RF algorithm is well suited for detection of misfire in spark ignition engines.