Search Results

Reliable estimation of pumping and friction losses in modern combustion engines allows better control strategies aiming at optimal fuel consumption and emissions. Sophisticated simulation tools enable detailed simulation of losses based as well on physical and thermodynamic laws as well as on design data. Models embedded in these tools however are not real-time capable and cannot be implemented into the programs of the electronic control units (ECU's). In this paper an approach is presented that estimates the pumping and friction losses of a combustion engine with variable valve train (VVT). Particularly the pumping losses strongly depend on the control of variable valve train by ECU. The model is based on a combination of a globally physical structure embedding data driven sub models based on test bed measurements. Losses are separated concerning different component groups (bearings, pistons, etc.).

The modern engine development process is characterized by shorter development cycles and a reduced number of prototypes. However, simultaneously exhaust after-treatment and emission testing is becoming increasingly more sophisticated. It is expected that predictive simulation tools that encompass the entire powertrain can potentially improve the efficiency of the calibration process. The testing of an ECU using a HiL system requires a real-time model. Additionally, if the initial parameters of the ECU are to be defined and tested, the model has to be more accurate than is typical for ECU functional testing. It is possible to enhance the generalization capability of the simulation, with neuronal network sub-models embedded into the architecture of a physical model, while still maintaining real-time execution. This paper emphasizes the experimental investigation and physical modeling of the port fuel injected SI engine.

The modern engine design process is characterized by shorter development cycles and a reduced number of prototypes. However, simultaneously exhaust after-treatment and emission testing is becoming increasingly more sophisticated. The introduction of predictive real-time simulation tools that represent the entire powertrain can likely contribute to improving the efficiency of the calibration process. Engine models, which are purely based on physical first principles, are usually not capable of real-time applications, especially if the simulation is focused on cold start and warm-up behavior. However, the initial data definition for the ECU using a Hardware-in-the-Loop (HiL)-Simulator requires a model with both real-time capability and sufficient accuracy. The use of artificial intelligence systems becomes necessary, e.g. neural networks. Methods, structures and the realization of a hybrid real-time model are presented in this paper, which combines physical and neural network models.

Future HD Diesel engine technology is facing a combination of both extremely low exhaust emission standards (US 2002/2004, EURO IV and later US 2007, EURO V) and new engine test procedures such as the European Transient Cycle (ETC) in Europe and the Not-to-Exceed Area (NTE) in the US). Customers furthermore require increased engine performance, improved efficiency, and long-term durability. In order to achieve all targets simultaneously, future HD Diesel engines must have improved fuel injection and combustion systems and utilize suitable technologies such as exhaust gas recirculation (EGR), variable geometry turbine turbocharger systems (VGT) and exhaust gas after-treatment systems. Future systems require precision controlled EGR in combination with a VGT-turbocharger during transient operation. This will require new strategies and calibration for the Electronic Engine Control Unit (ECU).

In the last two decades the abilities of neural networks as universal approximation tools of non linear functional relationships as well as identification tools for nonlinear dynamic systems have been recognized and used successfully in many applications areas like modelling, control and diagnosis of technical systems. At the same time an increasing interest in optimal design methods is observed. Design of experiment is used to cope with the growing amount of measurements needed for the calibration of engines due to the rising number of control variables to be considered and the need for more accuracy in the description of engine behaviour to derive the best control strategies. In this paper a strategy for the integration of the concept of D-optimality in the learning process of neural networks is proposed. This leads to an optimal selection of data to be presented to the training procedure of the neural network aiming to a generation of robust neural models using fewer training data.