Abstract
<div class="section abstract"><div class="htmlview paragraph">Diesel engines with their embedded control systems are becoming increasingly complex as the emission regulations tighten, especially concerning NO<sub>x</sub> pollutants. The combustion and emission formation processes are closely correlated to the intake manifold O<sub>2</sub> concentration. Consequently, the performance of the engine controllers can be improved if a model-based or sensor-based estimation of the O<sub>2</sub> concentration is available. The paper addresses the modeling of the O<sub>2</sub> concentration in a turbocharged diesel engine. Dynamic models, compared to generally employed steady state maps, capture the dynamic effects occurring over transients, when the major deviations from the stationary maps are found. Dynamic models positively affect the control system making it more effective and, exploiting information coming from sensors, they provide a more robust prediction performance. Firstly, a Nonlinear Output Error model (NOE), with simulation focus, fed with four inputs is presented. The considered nonlinear function set is the one of neural networks. The inputs are engine BMEP, engine RPM and EGR and VGT valves position. Two distinct datasets are used for training and validation of the NOE model. These sets are generated using GT-Power simulation software implementing a fine model of the engine, previously validated on experimental measurements taken on the real engine. Besides the transient validation, the NOE model was tested against GT-Power outputs on step tests involving the EGR and VGT actuators. At last the network output is compared with an O<sub>2</sub> steady state map over a transient in normal and faulty conditions. The performance of the model is satisfactory in both conditions. Secondly, the potential benefits of installing an O<sub>2</sub> sensor in the intake manifold is presented: a Nonlinear Auto-Regressive with eXogenous input (NARX) model is considered and compared to the previously investigated NOE. The results prove that, exploiting the output coming from the O<sub>2</sub> sensor, the model prediction capability significantly improves.</div></div>
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have