Transformer Based Driving Behavior Safety Prediction for New Energy Vehicles

Hao Lin, Junjie Yao

Abstract: The classification of driving behavior, with a particular emphasis on discerning safe from unsafe practices, is a task of paramount importance in the appraisal of drivers, and its significance is escalating in the epoch of autonomous driving. Driving behavior classification typically employs an assortment of features, such as velocity, acceleration, pedal pressure, turn signal utilization, and Global Positioning System (GPS) signals, amongst others. Nonetheless, these features exhibit considerable heterogeneity and do not offer comprehensive coverage. The extant literature pertaining to time series classification grapples with efficaciously addressing the high-dimensional nature, voluminous data, and the complexity of scenarios within the safety classification of driving behavior, especially for new energy vehicles. In this study, we have amassed an extensive corpus of sensor data, generated during the operation of new energy vehicles. Our research focused on the classification of driving behaviors concerning safety within the context of new energy vehicles and was predicated upon self-supervised learning. We proffered a time series model that leverages the Transformer architecture, tailored specifically for the aforementioned scenario, and employed a pre-training framework. To ascertain the efficacy of the proposed model, it was subjected to rigorous validation against a dataset comprising driving data from new energy vehicles. The model exhibited commendable performance and was further assessed through a series of downstream tasks.

External IDs:dblp:conf/adma/LinY23