Data-driven multi-dimension driving safety evaluation for real-world electric vehicles

While the number of electric vehicle traffic accidents has increased consistently over recent years, this has resulted in significant property losses and personal risks for drivers and passengers. To alleviate the above issues, this paper proposes a novel hybrid gated recurrent unit and temporal convolutional (GRUTC) neural network to evaluate driving safety. First, an evaluation framework consisting of ten indicators is constructed, including evaluating component safety and risky driving behaviors. Second, a long-short sample scoring rule is proposed to obtain a comprehensive evaluation of driving safety. The safety score is first evaluated based on the short samples, consisting of eight actual sampling points. Then, a comprehensive safety score is evaluated based on a long sample, which consists of nine consecutive short samples. Third, all four datasets are evaluated to establish the evaluation model. Verified by the yearlong operation data of dataset #1, the proposed method shows higher evaluating performance than commonly used methods. More importantly, a novel transfer learning method based on accumulated training with progressive datasets is proposed to improve the generalization. A stable and remarkable evaluating accuracy is obtained with the MAPE of 2.68% when the model is directly tested with dataset #4 after two transfers. This paper aims to assess the driving safety of real vehicles employing "driver-vehicle-road" multidimensional indicators and to improve the accuracy of the assessment by using the GRUTCN model.