基于变化特征离散化的多维时间序列关联分析

发布时间：2018-04-27 02:38

本文选题：多维时间序列 + 时间序列表示　；参考：《东华大学》2017年硕士论文

【摘要】：随着工程机械上不同类型传感器如温度传感器、压力传感器、湿度传感器等的大规模部署,这些传感器产生了大量的多维度时间序列数据。通过对这些传感器采集的时间序列数据关联分析,可用于发现工程机械各部件工作状况的潜在关联关系。这种潜在关系将为工程机械系统的故障预警和分析提供支持。但是,由于工程机械时间序列具有维度高和数据量大的特点,使得传统的Apriori关联算法无法直接运用。因此,本文从实际的工程机械时间序列数据出发,针对工程机械时间序列具体特点,提出一种适用于工程机械领域多维时间序列的关联分析方法。首先,本文设计了多维时间序列关联分析系统的整体框架,该框架分为预处理模块,离散化模块和关联规则提取模块。预处理模块负责数据清洗和规范化,实现每个时间序列的取值范围的规范化,并使每个时间序列符合高斯分布。另外,离散化模块对规范化的时间序列进行表示和离散化,也就是将数值型数据转化成字符型数据。此外,关联规则提取模块利用改进的时间特性关联方法从离散化的时间序列中提取有效的关联规则。接着,本文详细说明了离散化模块所用到的技术。本文先对流行的PAA+SAX离散化方法进行评估,发现PAA+SAX离散化方法有不足之处。PAA表示方法会改变时间序列规范化的结果。并且PAA表示方法会忽略原始时间序列的关键信息。另外,SAX离散化方法只能表示时间序列的大小特性,并且SAX的前提是时间序列符合高斯分布。然而,本文的数据集大部分不是高斯分布。所以本文对SAX离散化方法进行改进,使其也可以处理非高斯分布的时间序列。针对SAX不能保留原始时间序列关键信息的不足,本文推荐了变化特征离散化方法,它可以处理非高斯分布的时间序列并保留原始时间序列的关键信息。最后,本文详细阐述了关联规则提取模块所用到的时间特性关联方法。该方法改进了传统的Apriori算法。因为传统的Apriori算法不考虑项的时间特性,所以需要改进候选项的生成算法,支持度的计算方法和关联规则生成算法。使得改进的时间特性关联方法不仅可以处理单重时间序列,还可以处理多重时间序列。不仅如此,本文使用频率阈值和支持度阈值来获得异常模式。针对本文数据集的周期性特点,本文推荐了可以减少关联算法运行时间的关键时间片段关联方法。
[Abstract]:With the large-scale deployment of different types of sensors such as temperature sensor, pressure sensor and humidity sensor in construction machinery, these sensors produce a large number of multi-dimensional time series data. Through the correlation analysis of the time series data collected by these sensors, it can be used to find out the potential correlation of the working conditions of the components of construction machinery. This potential relationship will provide support for early warning and analysis of engineering machinery system failures. However, because of its high dimension and large amount of data, the traditional Apriori association algorithm can not be used directly. Therefore, based on the actual construction machinery time series data, according to the specific characteristics of construction machinery time series, this paper proposes a method of correlation analysis suitable for multi-dimensional time series in the field of construction machinery. Firstly, the whole framework of multi-dimension time series association analysis system is designed, which is divided into preprocessing module, discretization module and association rule extraction module. The preprocessing module is responsible for data cleaning and normalization, and realizes the normalization of the value range of each time series, and makes each time series conform to Gao Si distribution. In addition, the discretization module represents and discretizes the normalized time series, that is, converting numeric data into character data. In addition, the association rule extraction module extracts the effective association rules from the discrete time series using the improved time characteristic association method. Then, this paper explains the technology used in the discretization module in detail. In this paper, the popular PAA SAX discretization method is evaluated, and it is found that the PAA SAX discretization method has some shortcomings. The PAA SAX representation method will change the result of time series normalization. And the PAA representation method ignores the key information of the original time series. In addition, sax discretization method can only express the size of time series, and the premise of SAX is that the time series conform to Gao Si distribution. However, most of the data sets in this paper are not Gao Si distribution. So the SAX discretization method is improved to deal with the time series of non-Gao Si distribution. In view of the deficiency that SAX can not retain the key information of original time series, this paper recommends a method of discretization of variation features, which can deal with the time series of non- distribution and retain the key information of original time series. Finally, the time characteristic association method used in the module of extracting association rules is described in detail. This method improves the traditional Apriori algorithm. Because the traditional Apriori algorithm does not consider the time characteristic of the item, it is necessary to improve the candidate term generation algorithm, the support degree calculation method and the association rule generation algorithm. The improved time characteristic correlation method can deal not only with single time series, but also with multiple time series. Moreover, the frequency threshold and the support threshold are used to obtain the abnormal pattern. In view of the periodicity of the data set in this paper, we recommend a critical time segment association method which can reduce the running time of the association algorithm.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：O211.61

【相似文献】