Details
| Original language | English |
|---|---|
| Journal | SAE Technical Papers |
| Issue number | 2021 |
| Publication status | Published - 6 Apr 2021 |
| Externally published | Yes |
| Event | SAE 2021 WCX Digital Summit - Virtual, Online, United States Duration: 13 Apr 2021 → 15 Apr 2021 |
Abstract
The development and calibration of modern combustion engines is challenging in the area of continuously tightening emission limits and the necessity for meeting real driving emissions regulations. A focus is on the knowledge of the internal engine processes and the determination of pollutants formations in order to predict the engine emissions. A physical model-based development provides an insight into hardly measurable phenomena properties and is robust against changing input data. With increasing modeling depth the required computing capacities increase. As an alternative to physical modeling, data-driven machine learning methods can be used to enable high-performance modeling accuracy. However, these are dependent on the learned data. To combine the performance and robustness of both types of modeling a hybrid application of data-driven and physical models is developed in this paper as a grey box model for the exhaust emission prediction of a commercial vehicle diesel engine. Internal engine processes are physically investigated to determine combustion characteristic quantities influencing the formation of NOx, CO, HC and soot emissions. With the physically modeled inputs, models based on machine learning methods, including Support Vector Machine and Feedforward Neural Network, are developed for emission modeling. The models are trained using the data from a commercial vehicle engine, validated against different hyperparameters and network architectures and tested against each other at 772 different operating points. A comparison is made to black box models formed from the measured data. In general, feedforward neural networks and support vector machines were enhanced by selecting the physically modeled inputs. The feedforward neural networks for HC and soot modeling were improved by approximately 20% and 10% with respect to the RMSE of the test data. For the support vector machines, CO and soot modeling benefited the most by 30% and 20% respectively of the RMSE of the test data. For a trained NOx model based on low load data its coefficient of determination regarding test data by high load is increased from 0.807 to 0.908.
ASJC Scopus subject areas
- Engineering(all)
- Automotive Engineering
- Engineering(all)
- Safety, Risk, Reliability and Quality
- Environmental Science(all)
- Pollution
- Engineering(all)
- Industrial and Manufacturing Engineering
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
In: SAE Technical Papers, No. 2021, 06.04.2021.
Research output: Contribution to journal › Conference article › Research › peer review
}
TY - JOUR
T1 - Hybrid Physical and Machine Learning-Oriented Modeling Approach to Predict Emissions in a Diesel Compression Ignition Engine
AU - Mohammad, Aran
AU - Rezaei, Reza
AU - Hayduk, Christopher
AU - Delebinski, Thaddaeus O.
AU - Shahpouri, Saeid
AU - Shahbakhti, Mahdi
PY - 2021/4/6
Y1 - 2021/4/6
N2 - The development and calibration of modern combustion engines is challenging in the area of continuously tightening emission limits and the necessity for meeting real driving emissions regulations. A focus is on the knowledge of the internal engine processes and the determination of pollutants formations in order to predict the engine emissions. A physical model-based development provides an insight into hardly measurable phenomena properties and is robust against changing input data. With increasing modeling depth the required computing capacities increase. As an alternative to physical modeling, data-driven machine learning methods can be used to enable high-performance modeling accuracy. However, these are dependent on the learned data. To combine the performance and robustness of both types of modeling a hybrid application of data-driven and physical models is developed in this paper as a grey box model for the exhaust emission prediction of a commercial vehicle diesel engine. Internal engine processes are physically investigated to determine combustion characteristic quantities influencing the formation of NOx, CO, HC and soot emissions. With the physically modeled inputs, models based on machine learning methods, including Support Vector Machine and Feedforward Neural Network, are developed for emission modeling. The models are trained using the data from a commercial vehicle engine, validated against different hyperparameters and network architectures and tested against each other at 772 different operating points. A comparison is made to black box models formed from the measured data. In general, feedforward neural networks and support vector machines were enhanced by selecting the physically modeled inputs. The feedforward neural networks for HC and soot modeling were improved by approximately 20% and 10% with respect to the RMSE of the test data. For the support vector machines, CO and soot modeling benefited the most by 30% and 20% respectively of the RMSE of the test data. For a trained NOx model based on low load data its coefficient of determination regarding test data by high load is increased from 0.807 to 0.908.
AB - The development and calibration of modern combustion engines is challenging in the area of continuously tightening emission limits and the necessity for meeting real driving emissions regulations. A focus is on the knowledge of the internal engine processes and the determination of pollutants formations in order to predict the engine emissions. A physical model-based development provides an insight into hardly measurable phenomena properties and is robust against changing input data. With increasing modeling depth the required computing capacities increase. As an alternative to physical modeling, data-driven machine learning methods can be used to enable high-performance modeling accuracy. However, these are dependent on the learned data. To combine the performance and robustness of both types of modeling a hybrid application of data-driven and physical models is developed in this paper as a grey box model for the exhaust emission prediction of a commercial vehicle diesel engine. Internal engine processes are physically investigated to determine combustion characteristic quantities influencing the formation of NOx, CO, HC and soot emissions. With the physically modeled inputs, models based on machine learning methods, including Support Vector Machine and Feedforward Neural Network, are developed for emission modeling. The models are trained using the data from a commercial vehicle engine, validated against different hyperparameters and network architectures and tested against each other at 772 different operating points. A comparison is made to black box models formed from the measured data. In general, feedforward neural networks and support vector machines were enhanced by selecting the physically modeled inputs. The feedforward neural networks for HC and soot modeling were improved by approximately 20% and 10% with respect to the RMSE of the test data. For the support vector machines, CO and soot modeling benefited the most by 30% and 20% respectively of the RMSE of the test data. For a trained NOx model based on low load data its coefficient of determination regarding test data by high load is increased from 0.807 to 0.908.
UR - http://www.scopus.com/inward/record.url?scp=85104866590&partnerID=8YFLogxK
U2 - 10.4271/2021-01-0496
DO - 10.4271/2021-01-0496
M3 - Conference article
AN - SCOPUS:85104866590
JO - SAE Technical Papers
JF - SAE Technical Papers
SN - 0148-7191
IS - 2021
T2 - SAE 2021 WCX Digital Summit
Y2 - 13 April 2021 through 15 April 2021
ER -