UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Machine learning modeling of a direct-injected dual-fuel engine based on low density experimental data Karpinski-Leydier, Michael


Automotive systems are constantly increasing in complexity, requiring advanced modeling methods with large data sets to analyze these systems. This work proposes a machine learning approach to rapidly developing, steady state, control oriented, engine models that use optimization methods and engineering knowledge to reduce the burden of data collection and improve model performance and reliability. Data is collected from a pilot ignited direct injection natural gas engine using a full factorial approach for a high density data set and a design of experiments approach for a low density training data set with randomized validation data. An optimization approach for selecting hyperparameters for neural network and Gaussian process regression models is proposed. Models for emissions and performance metrics are created and compared to response surface models. The hyperparameter optimized models show an improvement in robustness and model performance, reducing the normalized root mean square error by 26% compared to other hyperparameter configurations. Gaussian process regression hyperparameter optimization shows the lowest error, 46% lower than response surface models. The Gaussian process regression hyperparameter optimized models are further improved using multi-region modeling, sensitivity analysis based input reduction, layered modeling, and hybrid layered modeling. The sensitivity based input reduction reduces the normalized root mean square error for all models by an average of 8% and up to 19%. The layered models reduce the normalized root mean square error for the CO by 52%, NOₓ by 30%, and particulate matter by 33%. The multi-region models reduce the normalized root mean square error for the O₂ by 40% and thermal efficiency by 16%. Using the best techniques for each output, the error is reduced by 19%, compared to hyperparameter optimization alone and 45% compared to typical Gaussian process regression models. These results show that hyperparameter optimization combined with the other techniques presented here significantly reduce model error. Using these techniques, it is possible to reduce the reliance on data for engine modeling. Future research in energy conversion technologies can use these techniques to rapidly develop new technologies without the cost in time and funding typically reserved for extensive data collection.

Item Citations and Data


Attribution-NonCommercial-NoDerivatives 4.0 International