Engine knock recognition based on wavelet domains denoising and convolutional neural network

Based on the method of wavelet domain denoising， the noise signals from in-cylinder pressure were extracted， at crank angle of 0°—45°， fast Fourier transform was used for simultaneous analysis of the noise signal in the time and frequency domains，then the feature map was outputted. The map was inputted into convolutional neural network (CNN) for identifying different features in order to distinguish non-knock and knock. The knock test was conducted on a direct injection engine fueled with aviation kerosene. The result revealed that: the time-frequency map was significantly different between knock and non-knock，because slight knocking and severe knocking both produced large-amplitude noise signals within crank angle of 10°—30°. Wavelet denoising was better than bandpass filtering in knocking feature extraction， while CNN was better than Support Vector Machine (SVM) in knocking feature recognition; under four different operating conditions，the knock recognition accuracy was all over 91% by wavelet domain denoising combining with CNN method; the precision and recall of the knock were 83.16% and 98.79%，respectively. © 2024 Beijing University of Aeronautics and Astronautics (BUAA). All rights reserved.