Data Science Applied to Vehicle Telemetry Data to Identify Driving Behavior Profiles

Abstract

Currently, vehicles have a vast number of sensors, which are important for their operation. To have greater control over the actions of vehicles, the sensors can be combined with telemetry, so that the data recorded by the sensors are sent to a database. From the registration of this base, one of the biggest difficulties encountered is to use these data obtained and extract relevant information, so that you can take advantage of them and use them to improve the performance of the vehicle or the driver himself. Along with obtaining these results, there are, on the other hand, numerous trips, and drivers, from where this information is obtained, each with their way of driving and behavior during each journey. Currently, there are several types of classification among drivers. You can find some of the most common types of profiles and behavior through scientific articles, such as Driver Behavior Profiling, and in expert reviews. The project developed considers the scenario mentioned above and, thus, uses current methodologies and technologies, such as machine learning, to classify the profile of drivers based on the vehicle telemetry data on trips from a company in the sector. After pre-processing the data obtained, the most relevant ones were defined, relating the driving profile with a direct impact on final fuel consumption. Then, several driver profiles were created, considering information such as efficiency and driving style. With the end of the list of these profiles, an analysis was made by comparing them with the final consumption of each trip. Once the profiles were defined, and the KNN classification algorithm was proposed, a study of the KNN algorithm hyperparameters was conducted to identify the most promising configuration in the classification of driving profiles. Definitive validation was performed using the K-Fold Cross Validation method, with the validation result being greater than 92%.