A Comparison of Virtual Sensors for Combustion Parameter Prediction of Gas Engines Based on Knock Sensor Signals

Precise prediction of combustion parameters such as peak firing pressure (PFP) or crank angle of 50% burned mass fraction (MFB50) is essential for optimal engine control. These quantities are commonly determined from in-cylinder pressure sensor signals and are crucial to reach high efficiencies and low emissions. Highly accurate in-cylinder pressure sensors are only applied to test rig engines due to their high cost, limited durability and special installation conditions. Therefore, alternative approaches which employ virtual sensing based on signals from non-intrusive sensors retrieved from common knock sensors are of great interest. This paper presents a comprehensive comparison of selected approaches from literature, as well as adjusted or further developed methods to determine engine combustion parameters based on knock sensor signals. All methods are evaluated on three different engines and two different sensor positions. The investigated approaches include a convolutional neural network, extreme gradient boosting regression models, non-linear feature regression models, a partial differential equation, as well as one method based on the analysis of structure-borne sound to derive an appropriate correlation. For evaluation of these implemented methods, data was acquired from extensive measurements of two spark-ignited single-cylinder large gas engines and one dual fuel single cylinder large engine under different operating conditions. The results show that the data-driven approaches achieved a root mean squared error (RMSE) of under 5.69 bar for the PFP and a RMSE of under 0.53 ° crank angle (CA) for the MFB50 across all investigated datasets. One method from the literature was adapted for the present study by applying the continuous wavelet transform and extracting certain features from the time-frequency spectrum to establish a suitable correlation for the desired combustion parameters. By achieving RMSE values for PFP of under 5.45 bar and for MFB50 of under 1.12 ° CA over all processed datasets, this adapted, novel method demonstrated high potential for the underlying regression tasks.