Engine Knock Analysis And Feature Extraction Based On Wavelet Packet Transform

The combustion processes in cylinders are important for each engine working cycle, because the power performance, economy and emissions of the automobile are determined by it. Therefore, how to improve the combustion efficiency, reduce energy consumption and pollution have always been the focuses of the academe and industrial fields. A number of approaches are introduced to improve the performance of the engine, such as increasing the ignition advance angle and the compression ratio. But some serious problems appear, one of which is the engine knock. When the engine knock occurs, the combustion and the flame spread rate are very high. Due to the rapidly release of the combustion energy, high frequency shock wave, pressure oscillation in the cylinder, unexpected mechanical vibration and noise are caused. Serious engine knock can damage mechanic parts, reduce the output power, increase fuel consumption and pollution. However, the light knock can improve engine power and economy performance, so accurate detection of the light or critical knock is very important for the internal-combustion engine..The most common and practical knock sensor is the accelerometer, which can be mounted on the engine block head, and the vibration of the engine block is measured. Compared with signals measured with pressure sensor or fiber optical sensor, the vibration signal includes more background noise. Knock detection can be carried out with some signal processing methods; the usually used method is with the help of the bandpass filter. But performance of the filtering methods decrease when the knock is light or the signal-to-noise ratio is low. In order to detect light knock, this paper uses various modern signal processing techniques-wavelet packet transform and higher order cumulants to evaluate knock intensity accurately.Wavelet transform and wavelet packet transform, as very important time- frequency analysis methods for non-stationary signals, are suitable for engine vibration signal analysis, because knock causes abnormal change of signals, while the wavelet transform and wavelet packet transform are powerful tools for detecting abnormally change signals. Traditional Fourier transform doesn't have quality of local analysis. Short-time Fourier transform, wavelet transform and wavelet packet transform are compared in this paper. The multi-resolution analysis of wavelet transform is with high time resolution for high-frequency components, and high frequency resolution for low-frequency components. Therefore, wavelet transform and wavelet packet transform are more suitable to extrac engine knock feature.Wavelet packet is the improvement of wavelet transform, compared to the general wavelet transform method, the detail components can be further decomposed, which is more flexible with time-frequency resolution. This paper analyses the advantages of wavelet packet transform from both theory and practical signals processing. Wavelet transform is rough for knock signals analysis in frequency domain, while the vibration signals always contains a lot of complex structure noise, so wavelet transform is not suitable for the light knock detection. The subspace decomposition of wavelet packet transform indicates that wavelet packet transform is with more detailed analysis and higher time-frequency resolution than wavelet transform. The results of real vibration signals processing also indicate that wavelet packet transform is better than wavelet transform for knock feature extraction.This paper proposes a knock feature extraction method with the use of wavelet packet transform. Firstly, the effects on vibration signals caused by knock in time domain and frequency domain are analyzed. The results show that the knock characteristic frequency varies with different working environment, even for a given engine. Therefore, parameterized power spectrum density (PSD) estimation is proposed to estimate the knock characteristic frequency accurately.Since the knock frequency corresponds to the peak of power spectrum density, autoregressive (AR) model is the all-pole model which can easily estimate the peak of PSD, so an AR model is the best chiose. The model order is difficult to decide when modeling the typical knock signal. A low order will lead to a smooth spectrum estimation, which makes some of the peaks lost. While a high order will lead to false peaks, although it can improve the resolution of the spectrum. This paper proposes an approach to determine the model order with the prediction error, the analysis of the experiment signals proves that this method is able to describe the knock frequency accurately while computation can be saved.After the knock frequency range is estimated, adjusting the decomposition tree structure of wavelet packet transform, so that the knock component can be analyzed in detail. A denoising method for vibration signals is proposed, noise coefficients are removed, and the wavelet packet transform coefficients with engine knock information are reserved for wavelet packet reconstruction, so that valid knock feature is extracted while a lot of background noise is removed.In order to recognize light knock accurately, a knock feature extraction method combined wavelet packet transform with higher-order cumulants is proposed in this paper. The fourth-order cumulant diagonal slice of vibration signals with different knock intensity is analysed, the result shows that the vibration signal has super-Gaussian distribution, and the kurtosis of it varies with the current knock intensity. Then a knock feature extraction method combined wavelet packet transform with kurtosis is proposed based this conclusion, calculating the kurtosis of wavelet packet transform coefficients, which could find the coefficients contained knock information, wavelet packet restruction is carried out with those coefficients, so knock feature extraction method is improved further more.In order to compare the proposed method with other knock detection methods, this paper also introduces other methods used in the current knock control system, such as filtering method, describes the principle of it and designs IIR digital band-pass filter for the testing engine.Through extracting knock features from the real vibration signals, three methods, filtering methods, discrete wavelet transform method and the proposed method in this paper is compared. The results indicate that the proposed method can extracte valid knock feature while a lot of background noise is removed. Compared with the other two methods, the proposed method improves signal to noise ratio greatly, which is more suitable for the case of low SNR or light knock detection. An experiment system is set up, which includes a gasoline engine, an accelerometer and other measure and control equipments. In order to analyze the actual performance of the method proposed in this paper, and how the mother wavelet and engine rotation speed affect it, vibration signals with a variety of engine rotation speed and ignition advance angle are collected for off-line processing.How to choose the optimal mother wavelet is difficult for both of wavelet transform and wavelet packet transform. There is no uniform guidelines or common principle. Some important characters of three series of orthogonal wavelet functions are described in detail, and then 19 kinds of mother wavelet which usually used in engineering, are applied in the method to extract knock feature. Based on the experiment results, a principle of choosing mother wavelet for knock feature extraction is proposed in this paper.Considering the engine rotation speed has an effect on the cylinder head vibration signals, this paper analyzes the performance of the proposed method when with different speeds. Although the SNR of the raw vibration signal has obvious changes when the speed changes, this method can still extract effective information of engine knock, even for low SNR and light knock.Knock intensity evaluation is researched based on knock features extracted with wavelet packet transform. Since wavelet packet transform method proposed in this paper makes amplitude peak notable, a fast knock intensity evaluation can be achieved with the peak value. After analyzing the trends of peak and average of the knock components with a variety of ignition advance angle, a series of thresholds were set up to evaluate the current knock intensity, and determine a range for light or critical knock. When the adjusting strategy varies with different kinds of knock intensity, the method is helpful to improve the efficiency of the knock control system.