DBSCAN和GRU算法在桥梁监测系统的研究  

Research on DBSCAN and GRU algorithm in bridge monitoring system

作　　者：刘欢[1] 李富年[1] 颜永逸 宋晓东[2] 杨国静[2] 林俊平 LIU Huan;LI Funian;YAN Yongyi;SONG Xiaodong;YANG Guojing;LIN Junping(College of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,China;China Railway Second Engineering Group Co.,Ltd.,Wuhan 430061,China;School of Civil Engineering and Mechanics,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]武汉科技大学信息科学与工程学院,湖北武汉430081 [2]中铁二院工程集团有限责任公司,湖北武汉430061 [3]华中科技大学土木工程与力学学院,湖北武汉430074

出　　处：《现代电子技术》2022年第20期114-118,共5页Modern Electronics Technique

基　　金：国家自然科学基金项目(51778258);国家自然科学基金项目(51922046);中铁二院工程集团有限责任公司科研项目(KYY2019029(19-21));中铁第四勘察设计院集团有限公司科研项目(2020K006,2019D001,2020D002)。

摘　　要：桥梁监测系统通过实时监测桥梁的各项指标来保证桥梁的安全运行,但监测数据在传输的过程中,不可避免地会产生噪声,从而对后续的数据预测产生较大干扰。通常利用聚类找出离散点来去除噪声,传统的K-means算法聚类前需要指定聚类簇数,以空间中K个点为中心进行聚类,对最靠近的对象归类,但海量的桥梁数据易受环境因素影响,因而无法预先指定数据簇。DBSCAN无需事先知道要形成的簇类的数量,可以自动确定簇个数。另外,桥梁数据是以时间戳存储的时序数据,在时序数据预测算法中,LSTM和GRU算法能够解决RNN算法的梯度爆炸问题,而GRU比LSTM参数量少,可以减少过拟合风险。基于此,文中以郑万高铁巫山大宁河双线大桥为研究对象,提出一种DBSCAN和GRU神经网络相结合的数据预测算法,以DBSCAN剔除噪声数据,并利用GRU神经网络对桥梁的压力进行深度学习,预测下一时刻的数据,然后进行异常检测。实践结果表明:所提算法可以准确地预测桥梁下一时刻的压力值,与LSTM算法相比,该算法的决定系数提高5.2%,均方根误差和平均绝对误差分别降低8.3%和11.6%;同时系统还能及时发送预警短信,为桥梁的安全提供保障。The bridge monitoring system can ensure the safe operation of the bridge through real-time monitoring of various indicators of the bridge. However,noise will be generated inevitably during the transmission of monitoring data,which will greatly interfere with the subsequent data prediction. Clustering is usually used to find discrete points to remove noise. In the traditional K-means algorithm,the number of clusters are need to be specified before clustering,clustering is conducted around k points in the space, and the closest objects are classified. However, the massive bridge data is easily affected by environmental factors,so it is impossible to specify the data clusters in advance. DBSCAN can automatically determine the number of clusters without knowing the number of clusters to be formed in advance. Bridge data is time-series data stored in time stamps. In the time-series data prediction algorithms,LSTM and GRU algorithms can solve the gradient explosion problem of RNN algorithm,while GRU has fewer parameters than LSTM,which can reduce the risk of overfitting. On this basis,the Zhengwan high-speed railway Wushan Daning river double-line bridge is taken as the research object,a data prediction algorithm combining DBSCAN and GRU neural network is proposed. DBSCAN is used to eliminate noise data,and GRU neural network is used to conduct deep learning for the bridge pressure,predict data at the next moment,and perform anomaly detection. The practice results show that the proposed algorithm can accurately predict the pressure value of the bridge at the next moment. In comparison with the LSTM algorithm,the determination coefficient of the model is increased by 5.2%,and the root mean square error and average absolute error are reduced by 8.3% and 11.6%,respectively. The system can send early warning short messages in time,which provides a guarantee for the safety of the bridge.

关 键 词：桥梁监测 时序数据 噪声数据 K-MEANS DBSCAN RNN LSTM GRU 异常检测 

分 类 号：TN931.3-34[电子电信—信号与信息处理]