重点营运车辆的异常驾驶行为识别研究

为加强对重点营运车辆异常驾驶行为的监督与检测，本文基于时间序列符号化算法(TSA) 与多尺度卷积神经网络模型(MCNN)提出一种组合模型TSA-MCNN，用于识别重点营运车辆异常驾驶行为。首先，对北斗数据进行预处理，并基于营运车辆存在多种车型、多种速度限制、多种异常驾驶行为的特点划分4种异常驾驶行为，构建异常样本数据集。其次，构建TSA-MCNN模型识别样本数据集，其过程分为两阶段，第1阶段，针对重点营运车辆的特点，引入能够粗粒化处理数据特征的时间序列符号化算法与能够多通道参数输入的多尺度卷积神经网络进行组合，并基于Keras库完成TSA-MCNN模型的搭建；第2阶段，利用样本数据集作为模型的输入变量，完成模型的训练、测试与识别。最后，以广河高速重点营运车辆北斗数据验证TSA-MCNN模型的性能， 同时，与异常识别传统算法的卷积神经网络(CNN)模型与动态时间扭曲-K最近邻(DTW-KNN)模型进行对比分析。验证结果表明：TSA-MCNN模型整体识别准确率为97.25%，相对于CNN模型与DTW-KNN模型提高了20.50%与5.63%。其中，TSA-MCNN模型对于正常驾驶行为、超速驾驶行为、紧急停车行为、临时停车行为、低速驾驶行为的识别精确率相对于CNN模型(DTW-KNN模 型)分别提高了26%(13%)、26%(6%)、23%(5%)、28%(3%)、0(0)，说明该模型对于重点营运车辆异常驾驶行为的识别具有良好的性能。

Recognition of Abnormal Driving Behavior of Key Commercial Vehicles

To strengthen the supervision and detection of key commercial vehicles' abnormal driving behaviors, we proposed a combined model for recognizing abnormal driving behavior of key commercial vehicles based on a time series symbolization algorithm (TSA) and a multi- scale convolutional neural network model (MCNN). Firstly, we pre- processed the Beidou data. The commercial vehicles have the characteristics of multiple models, different speed limits, and various abnormal driving behaviors, which can be used to define four abnormal driving behaviors. And a sample data set was constructed. Secondly, we constructed a TSA-MCNN model to identify the sample data set. The process can be divided into two stages. In the first stage, we introduced a time series symbolic algorithm that can coarsely process data features and a multi- scale convolutional neural network that is capable of multi- channel parameter input to build the TSA-MCNN model based on the Keras library. In the second stage, we used the sample data set as the input variable to complete the training, testing, and identification of the model. Finally, we verified the performance of the TSA- MCNN model by key commercial vehicles' BeiDou data of Guanghe Expressway and compared it with the traditional convolutional neural network (CNN) model and the DTW- KNN model. The results show that the recognition accuracy of the TSA-MCNN is 97.25%, which is 20.50% and 5.63% higher than that of the CNN model and DTW- KNN model. And the recognition accuracy of the TSA- MCNN model for different behaviors including normal driving, speeding driving, emergency stopping, temporary stopping, and low-speed driving is 26%, 26%, 23%, 28%, and 0 higher than the CNN model, and 13%, 6%, 5%, 3%, and 0 higher than the DTW-KNN model. In conclusion, the proposed model has good performance for the recognition of abnormal driving behavior of key commercial vehicles.