Driving scenario recognition for advanced hybrid electric vehicle control

Abstract

Fuel consumption in a Hybrid Electric Vehicle (HEV) is typically impacted by powertrain operation modes, short- and long-term driving trend and style, and road type and traffic conditions. Typically, HEVs have rule-based supervisory control using heuristic logic. This approach works sub-optimally because it does not have knowledge of either road conditions or driving trends. We propose a machine learning approach to enhance the HEV controller performance. We create a Driving Scene Recognizer (DSR) that uses the contextual information available to recognize the current driving scenario. This information would be used by the supervisory controller to decide the optimal vehicle commands at each instant of the drive cycle. A hierarchical deep learning network is trained on videos of driving data and vehicle sensor data to classify typical driving scenarios. We demonstrate the performance of the DSR on real-world test data.