Combustion stability is a key contributor to engine shake at idle speed and can impact the overall perception of vehicle quality. The sub-firing harmonics of the combustion torque are used as a metric to assess idle shake and are, typically, measured at different levels of engine break mean effective pressure (BMEP). Due to the nature of the combustion phenomena at idle, it is clear that predicting the cycle-to-cycle and cylinder-to-cylinder combustion pressure variations, required to assess the combustion uniformity, cannot be achieved with the state of the art simulation technology.
Inspired by the advancement in the field of machine learning and artificial intelligence and by the availability of a large amount of measured combustion test data, this paper explores the performance of various machine learning algorithms in predicting the idle combustion uniformity. The algorithms that are explored include Neural Network (NN), Support Vector Machine (SVM), Ensembles of Trees (EOT) and Gaussian Process (GP). The variables selected as inputs to these algorithms include BMEP, indicated mean effective pressure (IMEP), pumping mean effective pressure (PMEP), spark timing, crank angle at 10% fuel burn (Burn0010), crank angle at 90% fuel burn (Burn1090), and crankshaft timing at 50% fuel burn (CA50). The algorithms output the amplitude of 0.5 order, 1.0 order and the 1.5 order harmonics of the combustion torque.
The results presented in this paper show the superiority of the Gaussian Process algorithm in predicting the combustion torque harmonics. Using this algorithm, this paper further investigates the sensitivity of the engine torque harmonics to the parameters, used as inputs to the algorithm, in order to establish potential design guidelines for upfront combustion system development.
