Driving Risk Intelligence Using Sequence-Aware Deep Learning

New challenges are emerging as the number of connected cars and sensing technologies has grown rapidly, resulting in vast amounts of data, which are heterogeneous and heterogeneous streams in telematics.The high volume of connected vehicles and sensing technologies has produced large-scale and heterogeneous telematics streams and resulted in new problems to transform raw cyber-physical data into reliable and actionable driving intelligence. Previous methods of driving behavior prediction typically make use of features extracted at the trip level and thereby ignore potentially significant temporal patterns in the sequence of vehicle observations. This study presents a data engineering framework that is repeatable and can combine data quality management, temporal feature transformation, variable-length sequence construction, and deep learning-based behavioral analyses to support risk intelligence-driven data. The proposed framework leverages 12,994 Vehicle trips collected using the Vehicle Energy Dataset (VED) and 162,566 of those observations are timestamped and represented in a structured temporal way using heterogeneous data from the vehicles, surrounding traffic, environment and context. A sequence-aware bidirectional long short-term memory (BiLSTM) model with masked temporal pooling is developed to capture dynamic driving patterns while maintaining irregular trip lengths. Class-weighted optimization is used in model training to enhance the reliability in imbalanced behavioral categories, and macro-averaged F1-score is used as the main reliability metric. The experimental results show that the proposed framework achieved a macro-F1 score of 0.601 and accuracy of 61.9% on the unseen test set, effectively detecting the critical aggressive driving patterns, yet it was difficult to differentiate ambiguous driving patterns. The rest of the error analysis emphasizes temporal representation, data quality control and richer behavioural features in order to better enhance intelligent transportation analytics. The proposed framework offers a repeatable process for leveraging raw vehicle telematics data into reliable predictive services and offers a glimpse of the potential of data engineering methods for cyber-physical transportation systems.