An unsupervised machine learning technique to identify knock from a knock signal time-frequency analysis

Abstract

Knock is a characteristic phenomenon of spark ignition engines that can result in lower fuel efficiency and higher combustion temperatures. When it is not adequately identified and controlled, knock might lead to intensive engine damage. Knock detection methods are based on recognising the large in-cylinder pressure oscillations that resonate in the combustion chamber as a consequence of the rapid auto-ignition of the end-gas. However, in-cylinder pressure sensors may be invasive and expensive. In this work, a method for knock classification is proposed based on the knock sensor measurement, which is a low-cost and easy-to-implement vibration sensor. Unfortunately, the vibration of the engine block contains information from several phenomena, and only intensive knock can be identified by classical methods. To counteract the amount of noise, machine learning (ML) techniques are applied to the knock sensor signal. A training database composed of several operating conditions is used to extract the main characteristics of knock. The signal is analysed in the time-frequency domain by using the short-time Fourier transform (STFT), the features of the signal are obtained by singular value decomposition (SVD), and finally, support vector machine (SVM) techniques are used to identify outliers. Experimental data from a spark-ignited engine at 2000 rpm and 3000 rpm with spark advance steps was used to train and validate the method. Results are compared to state-of-the-art knock recognition methods based on pressure and knock sensors. The technique was able to detect 73% of the knock cycles, which was similar to the results achieved using the pressure sensor data (77%). Compared to another method using the knock sensor signal as well, the percentage of false negative cases was reduced by 20%.