Empirical estimation of borderline knock onset

Abstract

Modern knock control strategies are very effective in regulating engine operation about some desired borderline knock limit condition, but the methods required to identify this condition are poorly defined and pose a significant challenge for calibration engineers. In this work, several new methods are developed to detect knock onset and its evolution from the statistical properties of a given dataset. The first method uses a dual log-normal model to characterize the mixed knocking/non-knocking population. The uncertainties of the estimates are investigated as a function of the model parameters, and it is also shown that the standard deviation of the estimates can be reduced significantly (in this case by a factor of 4.6 if the method is applied to the more informative ensemble of raw knock signals rather than processed knock intensity values. The second method uses a simpler, non-parametric approach to detect knock onset from the excess kurtosis of the data, taken across cycles and averaged over a 10° window. The method provides a statistically powerful test for detecting departures from normality that is associated with knock onset. The study explores the uncertainty of the kurtosis estimates in relation to the model parameters and uses this to establish a practical threshold above which knocking can be identified. It is shown that the technique identifies knock onset from cylinder pressure-based knock data across a broad range of engine speeds without the need to adjust or further calibrate this threshold. The method is also applied to accelerometer-based data, but in this case its performance degrades with engine speed due to increased noise levels.