燃气管网作为城市生命线工程的重要组成部分，便利生活的同时也存在着泄漏隐患。为保障燃气管线的安全运行，燃气安全监测系统蓬勃发展。燃气物联网数据具有时序性、数据量大、产生频率高、价值密度低等特点，且窨井中采集的甲烷浓度受到沼气等外界环境的干扰。因此，从海量监测数据中挖掘普遍规律、构建普适模型具有重要意义。为此，本文开展基于时序特征提取的地下管网燃气泄漏识别方法研究。提出了甲烷浓度异常事件的定义。由于燃气物联网时序数据的低价值密度属性，其所包含的重要信息只集中在一小段时间内，因此结合专家经验，定义连续时间范围内浓度大于1%vol的甲烷浓度片段称为一个甲烷浓度异常事件；而窨井中检测到的甲烷气体有燃气泄漏和沼气堆积两个可能来源，因此甲烷浓度异常事件可以进一步划分为燃气泄漏事件和沼气堆积事件，从而将燃气泄漏识别转化为甲烷浓度异常事件分类问题。提出了甲烷浓度异常事件上升片段的截取方法。综合分析甲烷浓度曲线及各类事件的特点发现，在燃气泄漏的识别中，甲烷浓度上升片段是关注的重点所在。因此，通过数据平滑、数据压缩、分段线性表示、片段合并等步骤，实现了上升片段的截取。此外，针对正负样本不均衡问题，提出了聚类和欠采样相结合的办法。提出了基于特征的甲烷浓度异常事件分类模型。从甲烷浓度时序数据的基础时域特性、甲烷浓度时序数据的波动性、温度特性三个方面，建立了特征工程；比较了K近邻、决策树、支持向量机分类算法，其中决策树算法的效果最优，召回率可达92.9%，且由决策树模型得到特征重要度排序，最重要的两个特征依次为上升片段的最大值和反映整体上升趋势的拟合斜率。提出了基于距离的甲烷浓度异常事件相似度比较模型。由于甲烷浓度异常事件的数据长度不同、持续时长各异，因此采用了弹性度量方式；结合多项式拟合，采用动态时间规整算法，计算两个甲烷浓度异常事件之间的相似度，最终将距离最小值对应的训练集样本类别作为测试集的预测结果；召回率可达81.8%。本文针对燃气物联网时序数据开展研究，对于燃气泄漏识别具有一定的启示意义。

As an important part of urban lifeline engineering, gas pipeline network facilitates daily life, but also has potential leakage hazards. In order to ensure the safe operation of gas pipelines, gas safety monitoring systems have flourished. The data of the gas IoT has the characteristics of time sequence, large data volume, high generation frequency, and low value density. Besides, the methane detected in inspection wells is interfered by the external environment such as biogas. Therefore, it is of great significance to find universal laws and build universal models from massive monitoring data. Research on gas leakage identification of underground pipeline network based on feature extraction of time series is studied in this paper.The definition of methane concentration abnormal event is proposed. Due to the low value density of time series, important information is only concentrated in a short period. Therefore, based on expert experience, a segment in which methane concentration is continuously greater than 1% vol is defined as a methane concentration abnormal event. Gas leakage and biogas accumulation are two possible sources of detected methane in inspection wells. Hence, methane concentration abnormal event can be further divided into gas leakage events and biogas accumulation events, thereby the identification of gas leakage is turned into the classification of methane concentration abnormal events.A method for intercepting rising segments of methane concentration abnormal events is proposed. Through comprehensive analysis of methane concentration curves and events’ characteristics, it is found that rising segments of methane concentration abnormal events are important in the identification of gas leakage. Therefore, rising segments are obtained by data smoothing, data compression, piecewise linear representation, and segment merging. In addition, a method of clustering and under-sampling is proposed to deal with the imbalanced dataset.A feature-based classification model for methane concentration abnormal events is proposed. The feature engineering contains three aspects, including the basic time domain characteristics of methane concentration, the volatility of methane concentration and temperature characteristics. The result shows that, decision tree performs best among k-nearest neighbor, decision tree, and support vector machine, with a recall rate of 92.9%. And the top 2 important features in the decision tree are the maximum value of the rising segment and the fitting coefficient reflecting the overall upward trend.A distance-based similarity comparison model for methane concentration abnormal events is proposed. Due to different data length and duration time of events, a resilience measurement method is used to calculate similarity. The similarity between two methane concentration abnormal events is calculated by dynamic time warping combined with polynomial fitting. Finally, the training set sample category corresponding to the minimum distance is used as the prediction result of the test set, and the recall rate could reach 81.8%.This paper studies the time series data of gas IoT, which has a reference significance in gas leakage identification.