A Model Based Design Approach For Knock Control in Internal Combustion Engines Using Machine Learning

Knock is the byproduct of an abnormal combustion taking place in the combustion chamber and is associated to high vibration levels, undesirable noise, increased emissions and degraded fuel economy. Conventional control strategies address knock in a corrective fashion by tuning the spark timing in order to decrease in-cylinder pressure and temperature. Although effective in avoiding engine’s prolonged exposure to the harmful vibratory forces, it may induce excessive spark retard and is implemented through a considerable number of calibrations. Recent breakthroughs on knock investigation show that knock intensity inherits random behavior and this statement plays a crucial role on how to appropriately design knock control laws. By leveraging this assumption using machine learning concepts and treating knock occurrence as a classification problem, a logistic regression model dependent on in-cylinder pressure, compressed mixture temperature and engine RPM is defined. The Likelihood Ratio Test is performed to validate the suggested hypothesis and McFadden’s pseudo-R2 is calculated to quantify its accuracy. By leveraging the hypothesis stated, a proportional gain feedback control is designed to account for steady state operation of combustion engines around a target knock rate. Simulation results show the performance of both logistic regression model and proportional gain knock control against a conventionally defined knock control. Keywords— Engine Knock, Logistic Regression, Machine Learning, Model based Design