| Natural gas,as one of the clean and low-carbon energy varieties,is becoming an increasingly important energy source worldwide.The dehydration process is a very critical step in the production process of natural gas.Equipment maintenance of natural gas dehydration unit is crucial for the efficient gas production in petrochemical companies.Anomaly detection in the condition monitoring data collected from the natural gas dehydration unit can assist equipment maintenance,improve maintenance efficiency,ensure normal equipment operation,and reduce losses due to equipment failure.As part of the 2018 Annual Technology Project of Chongqing Qi Kuang Company,this thesis use data-driven anomaly detection methods to build an anomaly detection model for the condition monitoring data of the natural gas dehydration unit with over one million standard cubic meter of natural gas per day of Qiqiao Central Station of Chongqing Qi Kuang Company from December 2015 to November 2018.The work is mainly composed of data pre-processing and anomaly detection model building.The data preprocessing include three researches carried out: data cleaning for the condition monitoring data collected from the natural gas dehydration unit,conditions recognition for the cleaning data,parametric clustering for data under a single working condition using correlation analysis and cluster analysis.The anomaly detection model building include two researches carried out: the designing of an improved multi-dimensional dynamic time warping algorithm,using the improved multi-dimensional dynamic time warping algorithm to detect anomalies in condition monitoring data collected from the natural gas dehydration unit.The major contents are summarized as follows:(1)Summarize the anomalies in the original operating data of the natural gas dehydration unit with over one million standard cubic meter of natural gas per day into three types: anomalies caused by missing data,anomalies beyond the normal range,and anomalies caused by data jumps.The anomalies caused by missing data can be divided into missing data with few variables and missing data with many variables.For each type of abnormal situation,a suitable cleaning method is designed,and a complete cleaning solution is finally provided for the cleaning of all data.(2)Working condition changes are found over the condition monitoring data,and the GGS algorithm is used to complete the conditions recognition work.Then for the data under a single working condition,this thesis uses Pearson coefficient correlation analysis and gray correlation analysis to analyze the correlation between the parameters from two perspectives respectively,linear and nonlinear.Besides,a hierarchical clustering analysis is taken.Finally,combining the results of the correlation analysis and the cluster analysis,a subsystem composed of 4 parameters K100?TI?302,K100?TI?307,K100?TI?309,K100?TI?310 is extracted from the system composed of 33 parameters.(3)A similarity measurement algorithm,named Multi-Dimensional Contextual Dynamic Time Warping(MDC-DTW),which enhances DTW by combining it with shape feature and local structural information,is proposed.To test the rationality of the MDCDTW's design,the algorithm is qualitatively and quantitatively analyzed,and the results show that the MDC-DTW's design is reasonable.In an effort to understand the benefits of MDC-DTW,the thesis compare it with four state-of-the-art algorithms,the results show that MDC-DTW has comparable accuracy and speed.(4)Divide the monitoring parameter data into pseudo-period data and non-pseudoperiod data,for each type of data,a suitable data segmentation method is provided.After the data segmentation work is done,first take a data segment under normal condition as sample,then use the MDC-DTW algorithm proposed in(3)to calculate the distance between the sample and the other data segments,next draw the line chart of these distances,finally anomaly detection is performed with the help of the line chart. |
PDF Full Text Request