Abstract: Engine downsizing and boosting have been some of the most widely used strategies for improving engine efficiency in recent years. Several studies have offered significant departures from on-road pre-ignition to steady-state engine laboratory studies, necessitating more robust data-driven diagnostic tools that can identify pre-ignition events in real world environments. The goal of this study is to apply deep neural networks for pre-ignition (PI) detection, based on scientific data obtained from less expensive sensors (like lambda and low-resolution exhaust back pressure (EBP) data), as a replacement for high resolution in-cylinder pressure measurements. Two deep neural network (DNN) models are proposed and applied for classification of 221,728 combustion cycles from 18 experiments with varying EBP. DNNs combined convolutional neural networks (CNNs) for detection of repetitive patterns in array-structured data, and recurrent neural networks (RNNs) for modelling in a temporal domain. The first model was fed data from the principal component analysis (PCA); the second model eliminated this step and was focused on time series input. As a performance metric, the area under the curve (AUC) of the receiving operating curve (ROC) was used for comparison of the two models. The model’s accuracy was tested on 44,305 cycles. Based on the AUC-ROC metric, the former model was better able to differentiate between pre-ignition and normal combustion cycles.
|Date made available
|KAUST Research Repository