Virtual Sensor Design Using Convolutional Neural Networks and Image Processing Techniques

Abstract

To achieve rapid decarbonisation and improve the performance of future internal combustion engines (ICEs), virtual sensors have emerged as a promising alternative to physical sensors. This study proposes a novel methodology for developing virtual sensors using advanced machine learning techniques, including image-based classification and generative adversarial networks (GANs), to predict key engine performance metrics and emissions. Real-time engine parameters such as in-cylinder pressure, engine speed, fuel injection rate, and oxygen concentration were used as input to train multiple machine learning models including ANN, Random Forest, SVM, XGBoost, and Decision Trees. Among these, the XGBoost regressor demonstrated the highest prediction accuracy with minimal computational cost. Furthermore, combustion data were transformed into grayscale images and used to train GANs, enabling the reconstruction of the rate of heat release (R.H.R) profiles. The results confirm that virtual sensors can achieve over 97% accuracy in predicting combustion characteristics and emissions, making them a viable tool for robust feedback control in ICEs operating under transient conditions.