System Level Diesel Engine Emission Modeling Using Neural NetworksOutlineBackgroundLiterature Review-ModelingProposed Neutral Networks ModelSlide 6Results and DiscussionFuture WorkDecember 14, 2010System Level Diesel Engine Emission Modeling Using Neural NetworksME 539 Project PresentationBy Jian GongInstructor: Prof. Yu Hen Hu2Outline•Background•Literature Review•Proposed Neutral Networks Model•Neutral Network Structure•Training and Testing•Results and Discussion•Future Work3Background•Diesel engine emission•Worldwide tightening of the emission regulations•Need accurate prediction of engine emissions•Challenges•Engine-out emission involves thermo-fluid physics and complex chemical mechanisms•Object---To build system level predictive emission models using Neutral Networks4Literature Review-Modeling •Zero-dimensional phenomenological models•Simple & Low accuracy: involves basic thermal-fluid physics•Fast•Multi-dimensional CFD (Computational Fluid Dynamics) models •Complicated & More accurate: involves detailed thermal-fluid physics and large reaction mechanism •Extremely high computational cost•System-level models•Physical phenomenological model coupled with Neutral Networks•Moderate accuracy & reasonable computation cost5Proposed Neutral Networks Model•Neutral Network Structure1st layer Input layerOutput layerEngine operating data (scalar)Crank-angle resolved in-cylinder data (vector)CO (carbon monoxide)6Proposed Neutral Networks Model•Notes on the Network•Physic laws are acted as activation functions in each layer•Currently, only three weights are used in the 2nd layer. More weights could be used in the 1st layer in the future.•Training and testing•Training method•Back propagation algorithm•Training & Testing data•Totally 16 data set from experimental measurement (currently)•8 data set for training and another 8 data set for testing7Results and Discussion•Challenges during training•Trade-off between # of weights, accuracy and computational load•Additional constrains and reduced degree of freedom due to physical laws eg. PV(w*x+w0) =Const (1.2<(w*x+w0)<1.7)Model W1W2W3Mean Square ErrorPhysical model (without NN)1 1 10.77552Trained model (with NN)0.02 2 10.05586341 2 3 4 5 6 7 8010002000300040005000Operating Mode Number [-]CO [ppm] Experimental dataTrained modelOriginal model8Future Work•Carefully integrated more weights into the physic model (like 1st layer)•Test the trained model using testing
View Full Document