This content is not included in your SAE MOBILUS subscription, or you are not logged in.

THC Concentration Estimation Model using FTIR Spectrum

Journal Article
ISSN: 2641-9637, e-ISSN: 2641-9645
Published September 21, 2021 by SAE International in United States
THC Concentration Estimation Model using FTIR Spectrum
Citation: Yabushita, H., Nagaoka, M., Gyoten, Y., Yoshioka, M. et al., "THC Concentration Estimation Model using FTIR Spectrum," SAE Int. J. Adv. & Curr. Prac. in Mobility 4(2):583-591, 2022,
Language: English


A novel total hydrocarbon (THC) emission concentration estimation model is proposed for reduction of engine development cost and simplification of exhaust measurements. The proposed method uses the absorbance spectra of a Fourier transform infrared (FTIR) spectrometer, which contains the information on a wide variety of hydrocarbons, as input. The model is based on machine learning algorithms including the least absolute shrinkage and selection operator (LASSO) regression and bagging techniques. To train the model, we created a dataset containing pairs of a spectrum of engine exhaust gas and the THC concentration. In addition, we incorporate absorbance spectra of individual hydrocarbon components and several inorganic components so that the model learns the contribution of each hydrocarbon to THC concentration and to ignore interferences of irrelevant gas components. Since the intensity of absorbance spectrum is not entirely linear to all over the THC concentration, a piecewise linear model is constructed for THC concentration estimation. The model was evaluated on both exhaust gases before and after catalyst of a gasoline engine. The proposed approach reduces the error of THC estimations to less than 5% in pre-catalyst exhaust gas and less than 14 % in post-catalyst exhaust gas in both steady and transient operating conditions, in comparison to the 20% error for the regression model using only hydrocarbon concentration data. The results show that modeling directly using FTIR spectrum achieves a higher accuracy, moreover, training with individual gas component spectrum data provides further improvement.