Machine learning-based recognition for recognizing operating conditions of multi-product pipelines

doi:10.16265/j.cnki.issn1003-3033.2024.06.1410

Abstract

Abstract:

In order to solve the problems that some operating conditions could not be automatically identified and the accuracy of abnormal operating condition recognition was low in the process of monitoring the production and operation of multi-product pipeline system, the intelligent operating condition recognition method was applied to construct a multi-product pipeline operating condition recognition model with real-time monitoring capability. First, logic rule discrimination methods and event logs in the multi-product pipeline system were used to supplement the data labels. Second, the data were segmented according to the start and end time of the operating conditions, and the subsequence of different operating conditions were extracted by using the sliding window. Third, the features of subsequence were extracted to construct the model for operating condition recognition of multi-product pipelines, and the recognition effects of six classification models, namely, random forest (RF), adaptive boosting (AdaBoost), support vector machine (SVM), time series forest (TSF), random interval spectral forest (RISF) and sequence learner (SEQL), were compared and analyzed. Finally, a real multi-product pipeline was used as an example for model validation. The results show that the TSF model has the highest recognition accuracy for the four operating conditions of valve switching, valve internal leakage, pigging and sling pump, and is more suitable for the recognition of short-term operating conditions. In contrast, the recognition precision of the AdaBoost model has a higher probability of including the true value in the 95% confidence interval.

Key words: machine learning, multi-product pipeline, operating conditions, operating condition recognition, classification model

CLC Number:

X944.4

LI Miao, LI Lingbo, ZUO Zhiheng, ZHANG Li, JIANG Luxin, SU Huai. Machine learning-based recognition for recognizing operating conditions of multi-product pipelines[J]. China Safety Science Journal, 2024, 34(6): 127-135.

Figures/Tables 9

Fig.1

Table 1

Fig.2

Table 2

Fig.3

Fig.4

Table 3

Table 4

Table 5

References 16

[1]	国家发展改革委, 国家能源局. 中长期油气管网规划[Z]. 2017.
[2]	张丽. 基于机器学习的成品油管道系统运行工况识别方法研究[D]. 北京: 中国石油大学, 2023.5.
	ZHANG Li. Research on machine learning based method for operational states recognition of product oil pipeline systems[D]. Beijing: China University of Petroleum, 2023.5.
[3]	秦程. 基于负压波与流量平衡法的管道泄漏监测系统研究[D]. 大连: 大连理工大学, 2021.
	QIN Cheng. Research on pipeline leakage monitoring system based on negative pressure wave and flow balance method[D]. Dalian: Dalian University of Technology, 2021.
[4]	ABHULIMEN K E, SUSU A A. Liquid pipeline leak detection system: model development and numerical simulation[J]. Chemical Engineering Journal, 2004, 97(1):47-67.
[5]	戚元华, 林伟国, 吴海燕. 基于时域统计特征的天然气管道泄漏检测方法[J]. 石油学报, 2013, 34(6):1195-1199. doi: 10.7623/syxb201306022
	QI Yuanhua, LIN Weiguo, WU Haiyan. Natural gas pipeline leakage detection method based on time-domain statistical characteristics[J]. Acta Petrolei Sinica, 2013, 34(6):1195-1199. doi: 10.7623/syxb201306022
[6]	ZUO Zhonglin, MA Li, LIANG Shan, et al. A semi-supervised leakage detection method driven by multivariate time series for natural gas gathering pipeline[J]. Process Safety and Environmental Protection, 2022, 164:468-478.
[7]	ZADKARAMI M, SHAHBAZIAN M, SALAHSHOOR K. Pipeline leak diagnosis based on wavelet and statistical features using Dempster-Shafer classifier fusion technique[J]. Process Safety and Environmental Protection, 2017, 105:156-163.
[8]	ZANG Dong, LIU Jinhai, WANG Huaizhen. Markov chain-based feature extraction for anomaly detection in time series and its industrial application[C]. 2018 Chinese Control and Decision Conference (CCDC). IEEE, 2018:1059-1063.
[9]	NING Fangli, CHENG Zhanghong, MENG Di, et al. A framework combining acoustic features extraction method and random forest algorithm for gas pipeline leak detection and classification[J]. Applied Acoustics, 2021, 182: DOI: 10.1016/j.apacoust.2021.108255.
[10]	赵利强, 王建林, 于涛, 等. 基于二代小波变换和多级假设检验的输油管道缓泄漏检测方法[J]. 石油学报, 2012, 33(5):898-903. doi: 10.7623/syxb201205024
	ZHAO Liqiang, WANG Jianlin, YU Tao, et al. Slow leakage detection method of oil pipeline based on second-generation wavelet transform and multi-stage hypothesis test[J]. Acta Petrolei Sinica, 2012, 33(5):898-903.
[11]	郎宪明, 王佳政, 曹江涛, 等. 基于改进VMD和TWSVM的多点泄漏检测方法[J]. 振动与冲击, 2021, 40(17):271-278.
	LANG Xianming, WANG Jiazheng, CAO Jiangtao, et al. Multi-point leak detection method based on improved VMD and TWSVM[J]. Journal of Vibration and Shock, 2021, 40(17):271-278.
[12]	张丽, 苏怀, 范霖, 等. 基于时序片段的油气管道运行工况识别方法[J]. 中国安全生产科学技术, 2022, 18(11):99-104.
	ZHANG Li, SU Huai, FAN Lin, et al. Identification method of oil and gas pipeline operating conditions based on time series fragments[J]. Journal of Safety Science and Technology, 2022, 18(11):99-104.
[13]	KHASHA R, SEPEHRI M M, TAHERKHANI N. Detecting asthma control level using feature-based time series classification[J]. Applied Soft Computing, 2021, 111: DOI: 10.1016/j.asoc.2021.107694.
[14]	LINES J, TAYLOR S, BAGNALL A. Time series classification with HIVE-COTE: the hierarchical vote collective of transformation-based ensembles[J]. ACM Transactions on Knowledge Discovery from Data, 2018, 12(5):52.1-52.35.
[15]	NGUYEN T L, GSPONER S, IFRIM G. Time series classification by sequence learning in all-subsequence space[C]. IEEE International Conference on Data Engineering. ICDE, 2017:947-958.
[16]	张江石, 李泳暾, 冒香凝, 等. 基于NLP的煤矿事故原因分类研究[J]. 中国安全科学学报, 2023, 33(6):20-26. doi: 10.16265/j.cnki.issn1003-3033.2023.06.1416
	ZHANG Jiangshi, LI Yongtun, MAO Xiangning, et al. Research on classification of causes of coal mine accidents based on NLP[J]. China Safety Science Journal, 2023, 33(6):20-26. doi: 10.16265/j.cnki.issn1003-3033.2023.06.1416

分类模型	优势	局限性
RF^[13]	准确率高;可很好处理缺失值;适用高维数据;可度量样本间的相似性^[13]	不可解释性强;在某些噪声较大的分类问题上会过拟合;准确度依赖数据体量^[13]
AdaBoost^[13]	具有很高的精度;充分考虑每个分类器的权重;不用对特征进行筛选^[13]	弱分类器数目不好设定;数据不平衡将导致分类精度下降;训练耗时较长^[13]
SVM^[13]	可避免维数灾难;泛化能力强;算法简单,具有较好鲁棒性;对小样本、非线性及高维模式识别具备独特优势^[13]	难以训练大规模数据集;对参数和核函数选择敏感;实际中需要组合多个二分类实现多分类^[13]
TSF^[14]	有可解释性,计算效率高,可提取重要时间特征;可克服间隔特征空间巨大的问题^[14]	适用于单变量时间序列,对于多维时间序列数据的分类问题,需要先将多维数据降维,再进行训练^[14]
RISF^[14]	算法简单;可准确处理噪声较大的数据^[14]	用于长序列分析时,此方法运行较为缓慢^[14]
SEQL^[15]	有可解释性,可将数字矢量转换为符号表示;可减少时间序列长度的长度矢量;更适合处理多分类问题^[15]	模型较为复杂,训练时间较长^[15]

站名	数据维度	采样时长/d	采样频率/s	信号类型
a	21	60	1	进出站压力、进出站流量、下载流量、主输泵(1号、2号、3号、4号)进出口压力、给油泵(1号、2号、3号、3a号)进出口压力
b	13			进出站流量、进出站压力、下载流量、主输泵(1号、2号、3号、4号)进出口压力
c	4			越站流量、进出站压力、下载流量
d	7			进出站流量、进出站压力、下载流量、主输泵进出口压力
e	5			进出站流量、进出站压力、下载流量
f	7			进出站流量、进出站压力、下载流量、主输泵进出口压力
g	5			进出站流量、进出站压力、下载流量
h	3			进站流量、进站压力、下载流量
i	2			下载流量、进站压力

应用场景	评价指标	模型对比
应用场景	评价指标	SVM	TSF
清管	P/%	85	89
	R/%	27	53
	A/%	99	99
	F₁值	0.41	0.66

应用场景	评价指标	模型对比
应用场景	评价指标	AdaBoost	TSF
阀门内漏工况	P/%	97	98
	R/%	73	79
	A/%	99	99
	F₁值	0.83	0.87
阀门开关工况	P/%	98.9	99.3
	R/%	75	77
	A/%	99	99
	F₁值	0.85	0.87
甩泵	P/%	86	87
	R/%	53	55
	A/%	99	99
	F₁值	0.66	0.67

场景	分类模型	测试数据长度/s	精确率 (95%置信区间)/%
阀门开关	AdaBoost	360	85.65 (±1.9)
阀门开关	TSF	60	86.64 (±0.83)
甩泵	AdaBoost	240	98.9 (±0.9)
甩泵	TSF	60	99.4(±0.2)
阀门内漏	AdaBoost	60	97.34 (±0.5)
阀门内漏	TSF	60	98.65 (±0.4)