A combined framework of Biplots and Machine Learning for real-world driving volatility and emissions data interpretation

Abstract

Advanced visualization techniques can be useful for a better understanding of driving behavior and vehicle emissions in real-time. This study used classic and sparse HJ-biplots to examine the relationship between driving behavior, vehicle engine, exhaust emissions, and route type variables. Different Machine Learning classifiers were applied. Second-by-second vehicle dynamic, engine, and emissions data were collected from three light-duty vehicles (hybrid, diesel, and gasoline) and along three different routes (urban, rural, and highway). The dataset included a sample of 12,150 s of speed, acceleration, vehicular jerk, engine speed, engine load, fuel flow rate, vehicular specific power mode, carbon dioxide and nitrogen oxides emissions. The proposed methodology not only enables the distinction of driving styles, road types, and emissions profiles but also allows for revealing the correlation of variables in a single plot. The Random Forest algorithm showed to present the highest accuracy. This study can be useful in the context of road traffic emissions monitoring since it identifies hidden relationships in input data, and it reduces the redundancy in input parameters without compromising information.