基于时间序列分析的桥梁健康监测信息处理方法研究

发布时间：2018-07-09 13:18

本文选题：时间序列分析 + 桥梁健康监测信息　；参考：《重庆交通大学》2015年硕士论文

【摘要】：随着在役桥梁结构安全问题日益突出以及桥梁监测技术的不断发展,桥梁健康监测系统已得到广泛的应用,基于实时桥梁结构响应信息对桥梁结构状态进行快速准确地分析也已成为亟待解决的热点问题。本文以时间序列分析为基础,依次对桥梁健康监测信息的数据预处理过程、相似查询过程、异常检测过程进行分析,三个过程相互衔接配合构建了桥梁健康监测信息异常检测模型,并实现了基于时间序列分析的桥梁健康监测信息分析系统。本文首先提出了基于聚类分段的单变量时间序列孤立点识别方法,以局部标准差变化为量度对整个单变量时间序列在时间轴上进行分段,使得对于某个时间序列子段,其包含的每个样本对象的增加或删除都不会使这一子段的局部标准差产生明显变化,但相邻段之间的样本局部平均值存在明显差异。分段结束后对于段内样本对象不大于孤立点判定阈值(一般为1)的子段做异常处理。经实验分析此方法对时间序列中的孤立点具有精准的挖掘能力。在对时间序列中的空缺值进行处理的过程中,使用基于最近距离邻法的空缺值填补方法,以局部标准差变化为量度分析找出整个时间序列中相应空缺点近邻范围内最可能相似数据段的上下界,使用相似数据段中数据的加权平均值来作为空缺数据样本的最相似估计值。之后通过使用主成分分析法(PCA)对马桑溪大桥的桥梁健康监测数据集的进行特征提取以构建了马桑溪大桥桥梁健康监测数据集的CMTS数据库,验证了使用PCA法对桥梁健康监测信息进行数据压缩的可行性。针对原有的K-means聚类算法聚类数目必须预先赋值、初始质心随机选取以及处理海量数据时效率较低的三个缺点,提出基于网格划分和三角形三边定理的改进的K-means聚类算法。改进后的K-means算法可以根据样本数据集样本分布特征对样本维度空间进行网格划分,并对其中的密集网块进行统计以自行确定k值与初始质心位置,并通过三角形三边定理的引入,大大减少了聚类过程中的迭代次数与计算复杂度。通过对比改进的K-means算法与原有K-means算法对UCI数据库中的PAMAP2 Physical Activity Monitoring数据集进行聚类分析的结果准确性与处理效率,验证了改进后的K-means算法的准确性与高效性。引入索引树结构,构建了基于B+索引树的k近邻相似查询算法。利用马桑溪大桥的桥梁健康监测数据集的CMTS集对基于B+索引树的k近邻相似查询算法进行检验,验证了其高效性与准确性。以时间序列局部异常系数LOF作为检测样本数据是否异常的量度,对多变量时间序列的异常检测算法进行分析。之后对以多变量桥梁健康监测时间序列集为研究对象的桥梁健康监测信息异常检测过程进行分析,并对桥梁健康监测数据的异常检测模型进行建立。最后通过整合前文中所涉及的所有方法与算法,以Matlab为平台搭建了基于时间序列分析的桥梁健康监测信息分析系统,并使用牛棚特大桥实时监测数据集导入桥梁健康监测信息分析系统中对异常检测模型进行验证,结论与牛棚特大桥阶段性检测报告内容相符,验证了基于时间序列分析的桥梁健康监测信息异常检测模型的可行性。
[Abstract]:The bridge health monitoring system has been widely used with the increasing security of the bridge structure and the continuous development of bridge monitoring technology. The rapid and accurate analysis of bridge structure based on real-time bridge structural response information has become a hot point problem to be solved urgently. This paper is based on time series analysis. In turn, the data preprocessing process of bridge health monitoring information, similar query process, and abnormal detection process are analyzed. The bridge health monitoring information anomaly detection model is built up with the three processes, and the bridge health monitoring information analysis system based on time series analysis is realized. A piecewise single variable time series outlier recognition method is used to segment the whole single variable time series on the time axis with the local standard deviation change as the measure. The increase or deletion of each sample object in the subsection of a time series will not make a significant change in the local standard deviation of the subsection, but it is adjacent to the subsection. There are obvious differences in the local mean values between the segments. After the segment end, the sample object in the segment is not more than the subsection of the outlier decision threshold (usually 1). This method has the accurate mining ability for the outliers in the time series. Using the vacancy value filling method based on the nearest neighbor method, the upper and lower bounds of the most likely similar data segments in the nearest neighbor range of the corresponding space defects in the whole time series are found out by the variation of the local standard deviation, and the most similar estimation value of the vacant data samples is used as the weighted mean value of the data in the similar data segments. The principal component analysis (PCA) is used to extract the data set of the bridge health monitoring data of the Ma sang Xi Bridge to construct a CMTS database of the health monitoring data set of the bridge of Ma sang Xi Bridge. The feasibility of using the PCA method to compress the data of the bridge health monitoring information is verified. The number of clustering algorithms for the original K-means clustering algorithm is necessary. The improved K-means clustering algorithm based on the grid division and the triangle three edge theorem is proposed. The improved K-means algorithm can be used to mesh the sample principal dimension space according to the sample data set sample distribution characteristics, and then the improved K-means algorithm can be used to mesh the sample principal dimension space. The dense network blocks are used to determine the K value and the initial centroid position by themselves. By introducing the triangle three edge theorem, the number of iterations and computational complexity in the clustering process is greatly reduced. By comparing the improved K-means algorithm and the original K-means algorithm, the PAMAP2 Physical Activity Monitoring data set in the UCI database is carried out. The accuracy and efficiency of clustering analysis results verify the accuracy and efficiency of the improved K-means algorithm. The index tree structure is introduced to construct a k nearest neighbor query algorithm based on the B+ index tree. The CMTS set of the bridge health monitoring data set of the Ma sang Creek bridge is used for the k nearest neighbor query algorithm based on the B+ index tree. Test, verify its efficiency and accuracy. Take the time series local anomaly coefficient LOF as the measure of whether the sample data is abnormal, analyze the anomaly detection algorithm of the multivariable time series. After that, the bridge health monitoring information anomaly detection process with the multi variable bridge health monitoring time series is the research object. In the end, the bridge health monitoring information analysis system based on the time series analysis is set up by integrating all the methods and algorithms involved in the previous article, and the bridge health monitoring letter is introduced with the real-time monitoring data set of the bullpeng super large bridge to import the bridge health monitoring letter by integrating all the methods and algorithms involved in the previous article. In the interest analysis system, the anomaly detection model is verified, and the conclusion is consistent with the content of the phased detection report of the bullpeng bridge. The feasibility of the bridge health monitoring information anomaly detection model based on time series analysis is verified.
【学位授予单位】：重庆交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U446

【参考文献】