Engine Knock Evaluation Using a Machine Learning Approach

Artificial Intelligence is becoming very important and useful in several scientific fields. Machine learning methods, such as neural networks and decision trees, are often proposed in applications for internal combustion engines as virtual sensors, faults diagnosis systems and engine performance optimization. The high pressure of the intake air coupled with the demand of lean conditions, in order to reduce emissions, have often close relationship with the knock events. Fuels autoignition characteristics and flame front speed have a significant impact on knock phenomenon, producing high internal cylinder pressures and engine faults. The limitations in using pressure sensors in the racing field and the challenge to reduce the costs of commercial cars, push the replacement of a hardware redundancy with a software redundancy. Therefore, it becomes strategically important to develop a robust predictive model that, using the physical properties such as air temperature and pressure, fuel consumption and engine speed, could increase the engine performance under a large range of operating conditions, without computational efforts. In this paper, three machine learning approaches were implemented to predict the knock onset and knock intensity of a SI engine. The tool is fed by several input variables coming from a CFD-1D engine model whose calibration has been performed by using experimental data. Input parameters influencing the knock phenomenon, such as engine speed, air-fuel ratio, max internal cylinder pressure, combustion timing, and physical air conditions in the plenum, have been used as dataset for training and test phases. Once trained, the machine learning models were tested on their ability to predict outputs based on samples not used during the training set. The outputs predicted were compared with the target ones and the accuracy of the model was evaluated in terms of RMS and R2.