Bjornsson, Ola (Lund University, Faculty of Engineering), Tunestal, Per (Lund University, Faculty of Engineering)

Ion Current-Based Knock Detection Using Convolutional Neural Networks

Scheduled for presentation during the Regular session "Powertrain Control I" (FrA1), Friday, November 1, 2024, 09:50−10:10, Room T1

7th IFAC Conference on Engine and Powertrain Control, Simulation and Modeling, Oct 30 - Nov 1, 2024, Dalian, China

This information is tentative and subject to change. Compiled on January 2, 2025

Abstract

Engine knock, an abnormal combustion phenomenon, can lead to significant engine damage and reduced performance in spark-ignited internal combustion engines. Traditional knock detection methods using acoustic sensors are prone to errors and limited by their sensitivity to sensor placement and engine noise. This study employs a Convolutional Neural Network (CNN) to classify knocking events using ion current measurements from a 13-liter compressed natural gas (CNG) heavy-duty spark-ignited engine. The dataset comprises ion current and in-cylinder pressure measurements, with the ion current serving as the CNN input and the pressure trace used to calculate Maximum Amplitude Pressure Oscillations (MAPO) for labeling knock intensity. The data was balanced and split into training, validation, and test sets to ensure robust performance evaluation. The CNN model achieved a 74.5% accuracy on the test set, with high precision in distinguishing between no-knock and heavy-knock cycles but more difficulty with mid-knock events. ROC curve analysis using the one-vs-rest method was performed, showing Area Under the Curve (AUC) values of 0.96, 0.83, and 0.92 for no-knock, mid-knock, and heavy-knock classifications, respectively.