The application of black box models to combustion processes in the internal combustion engine

The internal combustion engine has been under considerable pressure during the last few years. The publics growing sensitivity for emissions and resource wastage have led to increasingly stringent legislation. Engine manufacturers need to invest significant monetary funds and engineering resources in order to meet the designated regulations. In recent years, reductions in emissions and fuel consumption could be achieved with advanced engine technologies such as exhaust gas recirculation (EGR), variable geometry turbines (VGT), variable valve trains (VVT), variable compression ratios (VCR) or extended aftertreatment systems such as diesel particulate filters (DPF) or NOx traps or selective catalytic reduction (SCR) implementations. These approaches are characterised by a highly non-linear behaviour with an increasing demand for close-loop control. In consequence, successful controller design becomes an important part of meeting legislation requirements and acceptable standards. At the same time, the close-loop control requires additional monitoring information and, especially in the field of combustion control, this is a challenging task. Existing sensors in heavy-duty diesel applications for incylinder pressure detection enable the feedback of combustion conditions. However, high maintenance costs and reliability issues currently cancel this method out for mass-production vehicles. Methods of in-cylinder condition reconstruction for real-time applications have been presented over the last few decades. The methodical restrictions of these approaches are proving problematic. Hence, this work presents a method utilising artificial neural networks for the prediction of combustion-related engine parameters. The application of networks for the prediction of parameters such as emission formations of NOx and Particulate Matters will be shown initially. This thesis shows the importance of correct training and validation data choice together with a comprehensive network input set. In addition, an application of an efficient and accurate plant model as a support tool for an engine fuel-path controller is presented together with an efficient test data generation method. From these findings, an artificial neural network structure is developed for the prediction of in-cylinder combustion conditions. In-cylinder pressure and temperature provide valuable information about the combustion efficiency and quality. This work presents a structure that can predict these parameters from other more simple measurable variables within the engine auxiliaries. The structure is tested on data generated from a GT-Power simulation model and with a Caterpillar C6.6 heavy-duty diesel engine.