Selection of Intrinsic Mode Functions Using Statistical Indicators for Knock Detection in Spark-Ignition Engines

Spark-ignition (SI) engine knock remains a critical challenge that adversely affects engine performance, efficiency, and durability. Real-time knock detection using engine block vibration signals is complex due to its nonlinear and non-stationary nature. Advanced signal decomposition techniques such as Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD), Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), and Variational Mode Decomposition (VMD) are commonly employed to extract knock-related features. However, there is limited consensus on selecting the intrinsic mode functions (IMFs) that best represent knock characteristics. This work introduces a statistical selection criterion based on correlation coefficient (> 0.5), approximate entropy (ApEn > 0.5), and kurtosis (two IMFs with the highest kurtosis values) to identify knock-relevant IMFs. Vibration signals were denoised using Symlet wavelets and then decomposed using the aforementioned techniques. The effectiveness of the selection was evaluated using spectrogram analysis and validated with Maximum Amplitude of Pressure Oscillation (MAPO) values. Results demonstrate that the proposed selection criteria successfully isolated knock-related IMFs, particularly IMF1 and IMF2, when CEEMDAN was applied to signals at 4000 rpm - an improvement over previous works in which only IMF1 was considered for knock analysis. The selected IMFs corresponded to MAPO values exceeding 2 bar, confirming strong knock events. Higher ApEn values, approaching 1 for IMF1 of the vibration signal at 4000 rpm, may be attributed to two strong knock events (cylinders 2 and 3). In contrast, ApEn values around 0.6 for IMF1 at 2800 rpm corresponding to one strong knock (cylinder 1) and one weak knock (cylinder 3) suggest that further investigation is required to confirm the correlation between approximate entropy and knock intensity. This multi-metric approach enhances the robustness and reliability of knock detection and offers a low-complexity, high-accuracy framework suitable for real-time implementation using non-intrusive vibration sensors.