A variational mode decomposition based statistical approach for knock detection in spark ignition engines using engine block vibrations

Abstract Variational Mode Decomposition (VMD) is used to decompose the engine block vibration signals into distinct Intrinsic Mode Functions (IMFs) for separating the signals produced during the knock. However, the optimal number of modes required to extract the signals produced during knock must be carefully chosen, as an inappropriate choice may result in the loss of critical information or mode mixing. The number of decomposition modes is varied from 3 to 6, and the resulting IMFs are evaluated using a dual-feature selection criterion based on approximate entropy and kurtosis to determine the optimal number of modes. This method is also employed to identify the IMFs that may contain knock-related signals. The results reveal that the accuracy of the signal extraction depends on the number of decomposition modes. Comparison of results shows that VMD can effectively extract knock-induced signals when the signal is decomposed into three or four modes, as confirmed by the statistical method. A Design of Experiments (DoE) approach is used to generate synthetic signals containing knock-related signals in one or two cylinders, with varying amplitude and frequency, at engine speeds ranging from 2500 to 4500 rpm. This enables the evaluation of selection criteria across the entire operating range. Incorporating synthetic signals enables comprehensive validation of the desired methodology under repeatable and controlled circumstances with a broad scope of knock scenarios. High coefficient of determination (R 2) obtained for kurtosis (0.978) and approximate entropy (0.89) indicate a strong correlation between the synthetic and reference signals. Adjusted R 2 and predicted R 2 values also fall well within the acceptable range. The study also reconfirms that the optimum number of decomposition modes is three or four. Additionally, the accuracy of the selection criteria in identifying the correct IMF is found to be 94%. This study establishes a statistically validated, non-intrusive method for knock detection in SI engines.