【作者】 田甜；

【导师】 李占超；

【作者基本信息】 扬州大学 ， 水利工程， 2024， 硕士

【摘要】 大坝是一个不断变化和发展的系统,对水库的安全运行起着决定性作用。混凝土坝在长期的运行过程中,受到多种因素的影响,其状态会发生变化,故人们长期以来持续对大坝进行监测分析,以提高对其结构行为的理解。随着信息技术的不断发展,大量的监测数据被采集并用于混凝土坝的状态评估。在混凝土坝监测数据分析中,相似性度量作为一种基础而又重要的技术手段,可以帮助识别数据之间的关联性和差异性,从而揭示混凝土坝状态的内在规律和特征,进一步了解混凝土坝的运行状态和变化趋势。鉴于此,本文旨在弥补混凝土坝监测数据挖掘中经典度量方法的短板,深入开展时间序列相似性度量方法的研究。通过验证这些方法在混凝土坝监测数据挖掘中的有效性和可行性,以填补该领域的研究空白,为混凝土坝监测数据挖掘提供更可靠、更精准的分析方法和工具。具体内容如下:(1)针对混凝土坝监测数据自身不仅蕴藏着坝体运行的丰富信息,同时也凸显出日益加剧的不稳定性与随机性,目前缺乏精确的相似性度量方法对其分析,提出基于发散因子度量指标的加入趋势特征的动态弯曲时间符号近似聚合(DTW-TMSAX)相似性度量方法。首先探讨混凝土坝监测数据所存在的基于距离和基于密度的相似性指标方面的不足,进而提出一种基于历史过程中自然邻居的偏斜分布方法,并在此基础上定义偏度和局部密度的概念。随后,进一步根据偏度和局部密度的比值,推导出时间序列数据的发散因子,将这一指标作为综合相似性度量的有效工具。再考虑到原位监测数据的时序性,实现对混凝土坝在待定时间段内监测序列相似程度的精确量化分析,采用滑动窗口与DTW-TMSAX相似性度量方法,以发散因子为基础进行了应用。工程实例表明,该相似性度量过程可以很好地发现混凝土坝监测数据的不相似段。(2)针对混凝土坝在复杂服役环境受多种动态因素影响下,如何深入研究动态驱动因素影响坝体的动态行为和准确度量此时混凝土坝监测数据的相似性问题,提出了基于模型系数的扩展Frobenius范数(Eros)和分形维数相似性度量方法。首先从时变状态的角度重新审视了混凝土坝变形的驱动因素,根据混凝土坝监测序列与外界因素的复杂动态关系,同时考虑混凝土坝每日气温变化的延迟响应,加入校正函数,构建混凝土坝时变系数统计模型,以更精确地捕捉坝体变形的内在规律;其次,为了有效应对系数的时变特性,获得更为准确的模型系数序列,引入基于遗忘因子的递归最小二乘算法,将其作为混凝土坝模型的时变系数估计方法;再者,根据时变系数的整体相似模式和每一项系数序列自身的波动情况,分别提出加权范数的主成分分析和分形维数表示的相似性度量方法,从而全面、准确地揭示混凝土坝结构在不同时期的变化特征。在工程实例中,通过系数的相似性度量方法反映的混凝土坝波动情况与实际相符合。(3)针对传统的相似性度量方法大都仅关注时间点之间的对应关系,随着时间序列数据规模的增加和数据复杂性的提高,已无法高效度量混凝土坝监测序列之间的相似性,提出一种卷积神经网络(CNN)的无监督相似性度量方法。首先,分析传统时间序列相似性度量算法的局限性,并强调了CNN在时间序列分类任务中的优势。随后,详细介绍CNN的基本概念与网络结构,并创新性地将其应用于混凝土坝监测序列的相似性度量中。具体地,通过单次输入一个混凝土坝监测序列作为训练数据,并设定标签为0,以此为基础度量与其他序列的相似度。本文构建的基于CNN的相似度算法模型在3个不同聚类状态下的UCR数据集中进行大规模的实验,并比较两个传统的相似性度量方法,从而证明所提出的算法模型的有效性。结合实际工程表明,本章提出的CNN模型在度量混凝土坝监测序列相似性方面展现出显著的有效性,为混凝土坝安全监测提供了新的技术手段。更多还原

【Abstract】 This dam,continually shifting and transforming,is crucial for the secure management of reservoirs.During prolonged operations,multiple elements impact concrete dams,resulting in altered conditions.Consequently,consistent observation and analysis of dams have been conducted to improve comprehension of their structural characteristics.Owing to persistent advancements in information technology,an abundance of surveillance data has been amassed for evaluating the state of concrete dams.Analyzing concrete dam monitoring data involves using similarity measurement,a critical technical technique,to unravexpose data correlations and variances,unveiling the intrinsic laws and traits of concrete dam states and deepening our comprehension of its operational conditions and patterns.Consequently,this document seeks to tackle the limitations found in traditional measurement techniques for mining data in concrete dam monitoring and to conduct comprehensive analysis of time series similarity measurement approaches.Through confirming the practicality and success of these techniques in concrete dam data mining,our aim is to bridge the research void in this domain and offer more dependable and accurate analytical techniques for mining concrete dam monitoring data.The detailed elements are outlined below:(1)Concrete dam monitoring data not only contains rich information about dam operation but also highlights the increasing instability and randomness.Currently,there is a lack of accurate similarity measurement methods for its analysis.We propose a dynamic time warping with trend feature incorporation using the divergence factor measure(DTW-TMSAX)similarity measurement method.Firstly,we discuss the shortcomings of distance-based and density-based similarity indicators in concrete dam monitoring data and propose a method based on skewed distribution of natural neighbors in historical processes.On this basis,the concepts of skewness and local density are defined.Subsequently,based on the ratio of skewness to local density,the divergence factor of time series data is derived,which serves as an effective tool for comprehensive similarity measurement.Considering the temporal nature of in-situ monitoring data,we achieve precise quantification analysis of the similarity degree of monitoring sequences of concrete dams within a specified time period.We apply a sliding window and the DTW-TMSAX similarity measurement method based on the divergence factor.Engineering examples demonstrate that this similarity measurement process can effectively identify dissimilar segments in concrete dam monitoring data.(2)In response to the dynamic influences of various factors on concrete dams in complex service environments,and the need to accurately measure the similarity of monitoring data under such dynamic conditions,we propose an Extended Frobenius Norm and Fractal Dimension(Eros)similarity measurement method based on model coefficients.Firstly,we reexamine the driving factors of concrete dam deformation from the perspective of timevarying states.Taking into account the complex dynamic relationship between concrete dam monitoring sequences and external factors,as well as the delayed response to daily temperature changes,we introduce a correction function and construct a time-varying coefficient statistical model of concrete dams to more accurately capture the intrinsic patterns of dam deformation.Secondly,to effectively address the time-varying characteristics of coefficients and obtain more accurate coefficient sequences,we introduce a recursive least squares algorithm based on forgetting factors as the time-varying coefficient estimation method for concrete dam models.Furthermore,based on the overall similar patterns of timevarying coefficients and the fluctuations of each coefficient sequence,we propose similarity measurement methods using weighted norms of principal component analysis and fractal dimension representation,respectively.These methods aim to comprehensively and accurately reveal the changing characteristics of concrete dam structures over different periods.In engineering examples,the fluctuations of concrete dams reflected by the similarity measurement methods of coefficients are consistent with actual observations.(3)In response to the limitations inherent in conventional similarity measurement techniques,which primarily focus on temporal correspondences and struggle to efficiently assess the similarity of concrete dam monitoring sequences amidst escalating scale and complexity,we introduce an unsupervised similarity measurement approach grounded in Convolutional Neural Networks(CNNs).We commence by dissecting the constraints of traditional time series similarity algorithms and underscore the strengths of CNNs in time series classification endeavors.Subsequently,we offer a comprehensive exposition on the fundamental principles and network architecture of CNNs,culminating in their innovative application to the similarity assessment of concrete dam monitoring sequences.In a pioneering move,we leverage a single concrete dam monitoring sequence as training data,assigning labels set to 0,to gauge its similarity with other sequences within this framework.Extensive experimentation on the UCR dataset across three distinct clustering states is conducted using the CNN-based similarity algorithm model delineated herein.Comparative analyses are performed against two conventional similarity measurement methodologies,thereby elucidating the efficacy of the proposed algorithm model.By amalgamating theoretical insights with real-world engineering applications,the CNN model expounded in this chapter exhibits pronounced efficacy in appraising the similarity of concrete dam monitoring sequences.This breakthrough furnishes novel technical avenues for enhancing the safety monitoring protocols of concrete dams.更多还原