S-FBLS驱动的城市道路交叉口事故严重程度预测

S-FBLS-driven accident severity prediction at urban road intersection

摘要:

为全面提升交通事故严重程度预测效果,针对现阶段传统机器学习与深度学习方法预测精度有限、网络收敛缓慢等问题,提出一种改进模糊宽度学习系统（fuzzy broad learning system,FBLS）的城市道路交叉口事故严重程度预测方法。模型采用Takagi-Sugeno模糊系统取代宽度学习系统（broad learning system,BLS）的特征节点层,在保留BLS快速收敛特性的前提下,更为广泛地提取高维事故数据所隐藏的特征；同时在模型的输入层融合SMOTE过采样算法,平衡事故数据类别,增强预测结果的可靠性。通过英国大曼彻斯特地区交通事故的历史数据,在横向维度上选择原始FBLS,在纵向维度上选择交通事故严重程度预测常用的随机森林（RF）、支持向量机（SVM）、BP神经网络（BPNN）、长短期记忆网络（LSTM）、卷积神经网络（CNN）,分别与S-FBLS进行预测性能对比,结果表明:S-FBLS在横向上相较原始FBLS将严重事故的预测准确率提升52.87%,在纵向上相较5种对比模型提升网络训练速度97%以上,整体准确率分别提升2.2%、8.95%、8.68%、6.47%、5.64%,特异度平均提升6.49%,灵敏度平均提升6.31%,精确度平均提升5.66%。基于S-FBLS的事故严重程度预测方法可为城市道路交叉口事故发生预警提供可靠的理论支撑。

Abstract:

In order to comprehensively improve the effect of traffic accident severity prediction, for the current stage of traditional machine learning and deep learning methods with limited prediction accuracy and slow convergence of the network, proposing an improved FBLS method for predicting accident severity at urban road intersection. The model replaces the feature node layer of BLS with Takagi-Sugeno fuzzy system to extract the hidden features of high-dimensional accident data more extensively and still retains the fast convergence characteristics of BLS; the SMOTE algorithm is also fused in the input layer of the FBLS to balance the accident data categories and enhance the reliability of the prediction results. Through the historical data of traffic accidents in Greater Manchester, UK, the original FBLS was selected in the horizontal dimension, and RF, SVM, BPNN, LSTM, CNN, which are commonly used for traffic accident severity prediction, were chosen in the vertical dimension, to compare the model performance with the S-FBLS. The results show that comparison with the original FBLS, S-FBLS improves the accuracy of severe accidents by 52.87%, comparison with five comparative models, S-FBLS improves the network training speed by more than 97%, improves the overall accuracy by 2.2%, 8.95%, 8.68%, 6.47%, 5.64%, improves the specificity by an average of 6.49%, improves the sensitivity by an average of 6.31%, and improves the precision by an average of 5.66%. The S-FBLS-driven accident severity prediction method can provide a reliable theoretical support for the early warning of the occurrence of accident at urban road intersection.