基于分数阶导数衍生特征和小波散射网络的交通碰撞声学信号识别  

Traffic collision coustic signal recognition based on fractional derivative derived features and wavelet scattering network

作　　者：赵国荣 祁晋峰 武建军 熊威 徐秀 ZHAO Guorong;QI Jinfeng;WU Jianjun;XIONG Wei;XU Xiu(Taijiu Expressway Management(Taiyuan)Co.,Ltd.,Taiyuan 030000,China;Shanxi Intelligent Transportation Research Institute Co.,Ltd.,Taiyuan 300180,China;Puniu(Shanghai)Technology Co.,Ltd.,Shanghai 200241,China)

机构地区：[1]太旧高速公路管理(太原)有限公司,山西太原030000 [2]山西省智慧交通研究院有限公司,山西太原300180 [3]朴牛(上海)科技有限公司,上海200241

出　　处：《电子设计工程》2024年第20期181-186,共6页Electronic Design Engineering

摘　　要：针对现阶段交通碰撞事故报警系统实时性差、识别不准等问题,提出了一种联合分数阶导数衍生特征和小波散射网络系数的交通事故声学碰撞信号识别方法。针对车辆碰撞信号具备非平稳性和长程相关性,引入分数阶导数,从中提取可以表征车辆行驶状态的时频特征作为第一个特征数据集。采用小波散射网络提取具有多尺度、稳定性和局部不变性的散射系数作为第二个特征数据集。将这两个特征数据集进行融合,基于树模型构造交通碰撞事故声学信号识别模型。实验结果表明,在已有碰撞事故样本上的检出率为100%,误报次数控制在每天每公里0.5次以内。In order to solve the problems of poor real-time performance and inaccurate recognition of the current traffic collision accident alarm system,a traffic accident acoustic collision signal recognition method combining fractional derivatives and wavelet scattering networks is proposed.In view of the non-stationarity and long-range correlation of vehicle collision signals,fractional derivatives are introduced to extract time-frequency features that can characterize the vehicle driving state as the first feature data set.The wavelet scattering network is then used to extract the scattering coefficients with multi-scale,stability and local invariance as the second feature data set.These two feature data sets are fused,and a traffic collision accident acoustic signal recognition model is constructed based on the tree model.The experimental results show that the detection rate on existing collision accident samples is 100%,and the number of false alarms is controlled within 0.5 times per kilometer per day.

关 键 词：分数阶导数 车辆碰撞 分布式声学传感 小波散射网络 

分 类 号：TN06[电子电信—物理电子学]