Predicting Vehicle Engine Performance: Assessment of Machine Learning Techniques and Data Imputation

The accurate prediction of engine performance maps can guide data-driven optimization of engine technologies to control fuel use and associated emissions. However, engine operational maps are scarcely reported in literature and often have missing data. Assessment of missing-data resilient algorithms in the context of engine data prediction could enable better processing of real-world driving cycles, where missing data is a more pervasive phenomenon. The goal of this study is, therefore, to determine the most effective technique to deal with missing data and employ it in prediction of engine performance characteristics. We assess the performance of two machine learning approaches, namely Artificial Neural Networks (ANNs) and the extreme tree boosting algorithm (XGBoost), in handling missing data. Zero imputation and three statistical methods, specifically K-Nearest Neighbors (KNN), Expectation Maximization (EM) and Multivariate Imputation Chained Equations (MICE), were employed as a pre-processing step with ANN model. The efficiency of each technique was evaluated at various degrees of missing data in the input (25%, 50 and 75%). In addition, the imputation quality of missing data was estimated using Kernel Density Estimation (KDE) and Raw Bias. The ANN model’s ability to predict fuel consumption and thermal efficiency was limited by missing data input with the zero-imputation technique. Among statistical methods coupled with ANN, EM exhibited the least satisfying performance while KNN performed the best. XGBoost with sparsity aware algorithm outperformed other methods by demonstrating the best adaptability to input’s missing information.