Search Results

Neural networks are useful tools for optimization studies since they are very fast, so that while capturing the accuracy of multi-dimensional CFD calculations or experimental data, they can be run numerous times as required by many optimization techniques. This paper describes how a set of neural networks trained on a multi-dimensional CFD code to predict pressure, temperature, heat flux, torque and emissions, have been used by a genetic algorithm in combination with a hill-climbing type algorithm to optimize operating parameters of a diesel engine over the entire speed-torque map of the engine. The optimized parameters are mass of fuel injected per cycle, shape of the injection profile for dual split injection, start of injection, EGR level and boost pressure. These have been optimized for minimum emissions. Another set of neural networks have been trained to predict the optimized parameters, based on the speed-torque point of the engine.

Neural networks have been used for engine computations in the recent past. One reason for using neural networks is to capture the accuracy of multi-dimensional CFD calculations or experimental data while saving computational time, so that system simulations can be performed within a reasonable time frame. This paper describes three methods to improve upon neural network predictions. Improvement is demonstrated for in-cylinder pressure predictions in particular. The first method incorporates a physical combustion model within the transfer function of the neural network, so that the network predictions incorporate physical relationships as well as mathematical models to fit the data. The second method shows how partitioning the data into different regimes based on different physical processes, and training different networks for different regimes, improves the accuracy of predictions.

A cylinder model was developed using artificial neural networks (ANN). The cylinder model utilized the trained ANN models to predict engine parameters including cylinder pressures, cylinder temperatures, cylinder wall heat transfer, NOx and soot emissions. The ANN models were trained to approximate CFD simulation results of an engine. The ANN cylinder model was then applied to predict engine performance and emissions over the standard heavy-duty FTP transient cycle. The engine responses varying over the engine speed and torque range were simulated in the course of the transient test cycle. It was demonstrated that the ANN cylinder model is capable of simulating the characteristics of the engine operating under transient conditions reasonably well.

Artificial neural networks (ANN) have been recognized as universal approximators for nonlinear continuous functions and actively applied in engine research in recent years [1, 2, 3, 4, 5, 6, 7 and 8]. This paper describes the methodology and results of using the ANN to model a turbocharged DI diesel engine. The engine was simulated using the CFD code (KIVA-ERC) over a wide range of operating conditions, and numerical simulation results were used to train the ANN. An efficient data collection methodology using the Design of Experiments (DOE) techniques was developed to select the most characteristic engine operating conditions and hence the most informative data to train the ANN. This approach minimizes the time and cost of collecting training data from either computational or experimental resources. The trained ANN was then used to predict engine parameters such as cylinder pressure, cylinder temperature, NOx and soot emissions, and cylinder heat transfer.

An overall diesel engine and aftertreatment system model has been created that integrates diesel engine, exhaust system, engine emissions, and diesel particulate filter (DPF) models using MATLAB Simulink. The 1-D engine and exhaust system models were developed using WAVE. The engine emissions model combines a phenomenological soot model with artificial neural networks to predict engine out soot emissions. Experimental data from a light-duty diesel engine was used to calibrate both the engine and engine emissions models. The DPF model predicts the behavior of a clean and particulate-loaded catalyzed wall-flow filter. Experimental data was used to validate this sub-model individually. Several model integration issues were identified and addressed. These included time-step selection, continuous vs. limited triggering of sub-models, and code structuring for simulation speed. Required time-steps for different sub models varied by orders of magnitude.

A procedure has been developed to build system level predictive models that incorporate physical laws as well as information derived from experimental data. In particular a soot model was developed, trained and tested using experimental data. It was seen that the model could fit available experimental data given sufficient training time. Future accuracy on data points not encountered during training was estimated and seen to be good. The approach relies on the physical phenomena predicted by an existing system level phenomenological soot model coupled with ‘weights’ which use experimental data to adjust the predicted physical sub-model parameters to fit the data. This approach has developed from attempts at incorporating physical phenomena into neural networks for predicting emissions. Model training uses neural network training concepts.