基于GLAE-MLP的炼化装置管道腐蚀健康状态评估方法

doi:10.11949/0438-1157.20251192

摘要/Abstract

摘要：

炼化装置进出料管道长期处于高温、高压与强腐蚀介质耦合作用下，冲刷腐蚀、化学腐蚀并存，传统基于阈值与规则库的健康分级在多源时变工况与类不均衡场景下存在应用局限性，依靠单一指标得到的管道腐蚀健康状态易失稳。为解决这一问题，本文以硫酸烷基化装置精制单元进出料管道为研究对象，提出了一种针对炼化装置管道的腐蚀健康状态评估方法。首先依据工程规范与企业管理规定设定四项指标阈值，将样本划分为健康/亚健康/退化/失效4个等级，接着提出基于门控线性单元增强自编码器（GLAE）和多层感知机（MLP）的健康状态评估分类模型，训练后对测试集样本进行健康状态评估。结果表明，基于GLAE-MLP健康状态评估模型在测试集中实现了97.2%的样本保持原判，仅2.7%发生相邻等级漂移，在与MLP、AE-MLP的对比中，GLAE-MLP模型分别在在准确率、精确率、召回率和F1分数4个指标上均表现更优，分别达到了97.24%、96.83%、95.51%、96.12%，能有效完成对管道腐蚀健康状态的评估。

关键词: 腐蚀, 过程控制, 健康状态, 硫酸烷基化, MLP, 过程系统

Abstract:

The inlet and outlet pipelines of the refining equipment are subjected to the coupling effect of high temperature, high pressure, and strong corrosive media for a long time, and erosion corrosion and chemical corrosion coexist. Traditional health classification based on threshold and rule library has application limitations in multi-source time-varying working conditions and class imbalance scenarios, and the pipeline corrosion health status obtained by relying on a single indicator is prone to instability. To solve this problem, this article takes the inlet and outlet pipelines of the refining unit in the sulfuric acid alkylation unit as the research object, and proposes a corrosion health status assessment method for the refining unit pipelines. Firstly, based on engineering specifications and enterprise management regulations, four threshold indicators are set to divide the samples into four levels: healthy/sub healthy/degraded/failed. Then, a health status assessment classification model based on gated linear unit enhanced autoencoder (GLAE) and multilayer perceptron(MLP) is proposed, and the health status of the test set samples is evaluated after training. The results showed that the GLAE-MLP health status assessment model achieved 97.2% of sample retention in the test set, with only 2.7% experiencing adjacent level drift. Compared with MLP and AE-MLP, the GLAE-MLP model performed better in accuracy, precision, recall, and F1 score, reaching 97.24%, 96.83%, 95.51%, and 96.12%, respectively, effectively completing the assessment of pipeline corrosion health status.

Key words: corrosion, process control, health status, Sulfuric acid alkylation, MLP, process systems

中图分类号:

TE624

胡瑾秋, 马铭骏, 张来斌, 董家延. 基于GLAE-MLP的炼化装置管道腐蚀健康状态评估方法[J]. 化工学报, DOI: 10.11949/0438-1157.20251192.

Jinqiu HU, Mingjun MA, Laibin ZHANG, Jiayan DONG. Corrosion health status assessment method for refining unit pipeline based on GLAE-MLP[J]. CIESC Journal, DOI: 10.11949/0438-1157.20251192.

图/表 20

图1 自编码器结构

Fig.1 Autoencoder structure

图2 GLU过程示意图

Fig.2 Gated Linear Unit process

图3 MLP结构示意图

Fig.3 Multi-Layer Perceptron structure

图4 基于GLAE-MLP的评估方法流程图

Fig.4 Process of evaluation method based on GLAE-MLP

表1 健康状态指标设置以及对应关系

Table 1 Health status indicator settings and corresponding relationships

指标	定义和公式	工程依据	评估维度
安全裕度 $Δ t$	$Δ t = t a c t u a l - t m i n$	ASME B31.3；API 579-1/FFS-1；API 570	绝对安全边界： $Δ t > 0$ 有裕度； $Δ t → 0$ 风险升高； $Δ t ≤ 0$ 触及/越过FFS判据。
相对厚度比 $t r w$	$t r w = t a c t u a l / t n o m i n a l$	API 574	相对退化：跨管径和材质可比，便于可视化与横向比较。
腐蚀速率 $r c$	$r c = m a x r c L T, r c S T$	API 570；API 574； API 581	腐蚀动力学
年检维次数 $f I M (τ)$	$f I M τ = N I M τ τ$	API 580；API 581	外部扰动：事件频度越高通常越不利；必要时用有效覆盖时长 $T e f f$ 替代 $τ$ 。

表1 健康状态指标设置以及对应关系

Table 1 Health status indicator settings and corresponding relationships

指标	定义和公式	工程依据	评估维度
安全裕度 $Δ t$	$Δ t = t a c t u a l - t m i n$	ASME B31.3；API 579-1/FFS-1；API 570	绝对安全边界： $Δ t > 0$ 有裕度； $Δ t → 0$ 风险升高； $Δ t ≤ 0$ 触及/越过FFS判据。
相对厚度比 $t r w$	$t r w = t a c t u a l / t n o m i n a l$	API 574	相对退化：跨管径和材质可比，便于可视化与横向比较。
腐蚀速率 $r c$	$r c = m a x r c L T, r c S T$	API 570；API 574； API 581	腐蚀动力学
年检维次数 $f I M (τ)$	$f I M τ = N I M τ τ$	API 580；API 581	外部扰动：事件频度越高通常越不利；必要时用有效覆盖时长 $T e f f$ 替代 $τ$ 。

图5 进出料管段示意图

Fig.5 Feeding and discharging pipelines

表2 基于三个判定原则的划分结果

Table 2 Classification results based on three decision principles

判别原则	健康	亚健康	退化	失效
最不利原则	339	291	86	4
加权评分原则	364	302	52	2
中位数原则	379	313	27	1

表3 基于最不利原则的P-4101-B21Y管段腐蚀健康状态划分规则

Table 3 Classification rules of corrosion health status of P-4101-B21Y pipe section based on the most unfavorable principle

等级	$Δ t$ /mm	$t r w$ $/ %$	$r c$ /(mm/a)	$f I M τ$ /(次/a)	划分规则
Ⅰ健康	[3,+∞)	[0.95,+∞)	[0,0.10)	[0,1)	4个指标同时满足Ⅰ级
Ⅱ亚健康	[2,3)	[0.85,0.95)	[0.10,0.30)	[1,2)	全部≥Ⅱ级，且至少一项在Ⅱ级
Ⅲ退化	[0,2)	[0.75,0.85)	[0.30,0.50)	[2,5)	任何一项在Ⅲ级，则≥Ⅲ级
Ⅳ失效	(-∞,0]	(-∞,0.75)	[0.50,+∞)	[5,+∞)	任何一项满足即为Ⅳ级

表3 基于最不利原则的P-4101-B21Y管段腐蚀健康状态划分规则

Table 3 Classification rules of corrosion health status of P-4101-B21Y pipe section based on the most unfavorable principle

等级	$Δ t$ /mm	$t r w$ $/ %$	$r c$ /(mm/a)	$f I M τ$ /(次/a)	划分规则
Ⅰ健康	[3,+∞)	[0.95,+∞)	[0,0.10)	[0,1)	4个指标同时满足Ⅰ级
Ⅱ亚健康	[2,3)	[0.85,0.95)	[0.10,0.30)	[1,2)	全部≥Ⅱ级，且至少一项在Ⅱ级
Ⅲ退化	[0,2)	[0.75,0.85)	[0.30,0.50)	[2,5)	任何一项在Ⅲ级，则≥Ⅲ级
Ⅳ失效	(-∞,0]	(-∞,0.75)	[0.50,+∞)	[5,+∞)	任何一项满足即为Ⅳ级

表4 基于最不利原则的AL-4101-B41K管段腐蚀健康状态划分规则

Table 4 Classification rules of corrosion health status of AL-4101-B41K pipe section based on the most unfavorable principle

等级	$Δ t$ /mm	$t r w$ $/ %$	$r c$ /(mm/a)	$f I M τ$ /(次/a)	说明
Ⅰ健康	[4,+∞)	[0.96,+∞)	[0,0.05)	[0,1)	4个指标同时满足Ⅰ级
Ⅱ亚健康	[3,4)	[0.88,0.96)	[0.10,0.20)	[1,2)	全部≥Ⅱ级，且至少一项在Ⅱ级
Ⅲ退化	[1,3)	[0.78,0.88)	[0.30,0.40)	[2,5)	任何一项在Ⅲ级，则≥Ⅲ级
Ⅳ失效	(-∞,1]	(-∞,0.78)	[0.40,+∞)	[5,+∞)	任何一项满足即为Ⅳ级

表4 基于最不利原则的AL-4101-B41K管段腐蚀健康状态划分规则

Table 4 Classification rules of corrosion health status of AL-4101-B41K pipe section based on the most unfavorable principle

等级	$Δ t$ /mm	$t r w$ $/ %$	$r c$ /(mm/a)	$f I M τ$ /(次/a)	说明
Ⅰ健康	[4,+∞)	[0.96,+∞)	[0,0.05)	[0,1)	4个指标同时满足Ⅰ级
Ⅱ亚健康	[3,4)	[0.88,0.96)	[0.10,0.20)	[1,2)	全部≥Ⅱ级，且至少一项在Ⅱ级
Ⅲ退化	[1,3)	[0.78,0.88)	[0.30,0.40)	[2,5)	任何一项在Ⅲ级，则≥Ⅲ级
Ⅳ失效	(-∞,1]	(-∞,0.78)	[0.40,+∞)	[5,+∞)	任何一项满足即为Ⅳ级

表5 硫酸烷基化精制单元进出料管道工况参数数据集

Table 5 Operating parameter dataset of inlet and outlet pipelines in the sulfuric acid alkylation refining unit

T	X₁/℃	X₂	X₃/ MPa	X₄/(m³/h)	X₅/(m/s)	X₆/%	X₇/ppm	X₈/ppm	X₉/ppm	Y₁	Y₂	Y₃	Y₄
1	9.84	2.14	1.05	0.37	11.27	98.11	9.92	1.77	97.49	3.84	0.95	0.14	0
2	9.97	2.10	1.09	0.37	11.23	98.05	5.72	1.57	91.13	4.84	0.97	0.08	0
3	9.66	2.11	1.06	0.36	11.58	98.59	6.96	1.51	53.93	3.39	0.94	0.08	0
4	9.94	2.04	1.04	0.41	11.21	98.22	7.42	1.87	55.60	4.88	0.96	0.08	0
5	9.49	2.07	1.09	0.41	11.10	98.70	9.57	1.31	95.33	4.48	0.94	0.08	0
⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮
1549	8.72	2.11	1.03	0.37	11.14	98.52	9.35	1.86	61.35	2.84	0.87	0.32	1
⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮
3600	8.77	2.02	1.10	0.41	11.25	98.50	8.90	1.07	53.18	4.05	0.84	0.21	0

图6 数据集中不同健康状态对应的数据采集日期和采集管道分布

Fig.6 The distribution of data collection dates and collection pipelines corresponding to different health states in the dataset

图7 9种工况与4个指标之间的spearman相关系数图

Fig.7 Spearman correlation coefficient plot between 9 operating conditions and 4 indicators

图8 方差膨胀因子检验结果图

Fig.8 Variance expansion factor test results

表6 特征筛选后的硫酸烷基化精制单元进出料管道工况参数及状态标注

Table 6 Operating condition parameters and status annotation of inlet and outlet pipelines in the sulfuric acid alkylation refining unit after feature selection

T	X₁/℃	X₂	X₅/(m/s)	X₆/%	X₇/ppm	X₈/ppm	X₉/ppm	Y₁	Y₂	Y₃	Y₄	S
1	9.84	2.14	11.27	98.11	9.92	1.77	97.49	3.84	0.95	0.14	0	1
2	9.97	2.10	11.23	98.05	5.72	1.57	91.13	4.84	0.97	0.08	0	1
3	9.66	2.11	11.58	98.59	6.96	1.51	53.93	3.39	0.94	0.08	0	1
4	9.94	2.04	11.21	98.22	7.42	1.87	55.60	4.88	0.96	0.08	0	1
5	9.49	2.07	11.10	98.70	9.57	1.31	95.33	4.48	0.94	0.08	0	1
⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮
1549	8.72	2.11	11.14	98.52	9.35	1.86	61.35	2.84	0.87	0.32	1	2
⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮	⋮
3600	8.77	2.02	11.25	98.50	8.90	1.07	53.18	4.05	0.84	0.21	0	1

图9 基于评估指标的健康状态时序刻画示意图

Fig.9 Temporal characterization of health status based on evaluation indicators

图10 基于GLAE-MLP的健康状态时序刻画示意图

Fig.10 Temporal characterization of health status based on GLAE-MLP

图11 基于评估指标和GLAE-MLP的健康状态评估结果对比

Fig.11 Comparison of health status evaluation results based on evaluation indicators and GLAE-MLP

图12 健康状态评估结果混淆矩阵

Fig.12 Confusion matrix of health status evaluation results

表7 3种评估模型的性能指标对比

Table 7 Comparison of performance metrics of 3 evaluation models

模型	准确率	精确率	召回率	F1分数
MLP	91.85%	84.94%	84.45%	84.69%
AE-MLP	92.82%	86.46%	89.52%	87.60%
GLAE-MLP	97.24%	96.83%	95.51%	96.12%

图13 基于GLEA-MLP模型的健康状态跃迁以及报警事件的对应情况

Fig.13 Correspondence between health state transitions and alarm events based on the GLEA-MLP model

参考文献 32

[1]	Soomro A A, Mokhtar A A, Kurnia J C, et al. Integrity assessment of corroded oil and gas pipelines using machine learning: a systematic review[J]. Engineering Failure Analysis, 2022, 131: 105810.
[2]	Yang J F, Li R, Chen L C, et al. Research on equipment corrosion diagnosis method and prediction model driven by data[J]. Process Safety and Environmental Protection, 2022, 158: 418-431.
[3]	张洪华, 曾凌, 张小珍. 石油储油罐健康状态评估系统的建设分析[J]. 机电技术, 2022, 45(6): 87-90.
	Zhang H H, Zeng L, Zhang X Z. Analysis on the construction of health assessment system for oil storage tanks[J]. Mechanical & Electrical Technology, 2022, 45(6): 87-90.
[4]	王宝轩, 沈功田, 闫河, 等. 大型石油储罐健康管理方法应用研究[J]. 机械工程学报, 2017, 53(16): 125-133.
	Wang B X, Shen G T, Yan H, et al. Study on the application of health management methods for large oil storage tanks[J]. Journal of Mechanical Engineering, 2017, 53(16): 125-133.
[5]	American Society of Mechanical Engineers (ASME). Process Piping [S]. New York, NY: American Society of Mechanical Engineers (ASME), 2022: 121-250.
[6]	American Petroleum Institute (API). Piping Inspection Code: In-Service Inspection, Rating, Repair, and Alteration of Piping Systems [S]. Washington, D.C.: American Petroleum Institute (API), 2023: 21-80
[7]	American Petroleum Institute (API). Risk-Based Inspection [S]. Washington, D.C.: American Petroleum Institute (API), 2023: 31-70.
[8]	练小晖. RBI技术在炼化装置中的应用[J]. 化工装备技术, 2012, 33(6): 37-40, 43.
	Lian X H. Application of RBI technology in refining unit[J]. Chemical Equipment Technology, 2012, 33(6): 37-40, 43.
[9]	Bhatia K, Khan F, Patel H, et al. Dynamic risk-based inspection methodology[J]. Journal of Loss Prevention in the Process Industries, 2019, 62: 103974.
[10]	Moore D A. Security Risk Assessment Methodology for the petroleum and petrochemical industries[J]. Journal of Loss Prevention in the Process Industries, 2013, 26(6): 1685-1689.
[11]	杨旭东, 周艳丽, 刘志娟, 等. 基于RS-PSO-SVM算法的腐蚀管道剩余强度预测技术研究[J]. 石油工程建设, 2020, 46(3): 8-12.
	Yang X D, Zhou Y L, Liu Z J, et al. Research on residual strength prediction technology of corroded pipeline based on RS-PSO-SVM algorithm[J]. Petroleum Engineering Construction, 2020, 46(3): 8-12.
[12]	Lu Y J, Wang C W, Wang C H. A hybrid diagnostic system for corrosion mechanism identification in petrochemical equipment: Integrating Bayesian probabilities and association rule[J]. Process Safety and Environmental Protection, 2025, 201: 107577.
[13]	孙同敏. 基于DBN-SVM的航空发动机健康状态评估方法[J]. 控制工程, 2021, 28(6): 1163-1170.
	Sun T M. Evaluation method of aero-engine health status based on DBN-SVM[J]. Control Engineering of China, 2021, 28(6): 1163-1170.
[14]	单增海, 李志远, 张旭, 等. 基于多传感器信息融合和多粒度级联森林模型的液压泵健康状态评估[J]. 中国机械工程, 2021, 32(19): 2374-2382.
	Shan Z H, Li Z Y, Zhang X, et al. Health evaluation of hydraulic pump based on multi-sensor information fusion and multi-granularity cascade forest model[J]. China Mechanical Engineering, 2021, 32(19): 2374-2382.
[15]	尹刚, 钱中友, 曹文琦, 等. 基于Adaboost-PSO-SVM的铝电解槽健康状态诊断方法研究[J]. 化工学报, 2024, 75(1): 354-365.
	Yin G, Qian Z Y, Cao W Q, et al. Research on health diagnosis method of aluminum reduction cell based on Adaboost-PSO-SVM[J]. CIESC Journal, 2024, 75(1): 354-365.
[16]	陈雄姿, 于劲松, 唐荻音, 等. 基于贝叶斯LS-SVR的锂电池剩余寿命概率性预测[J]. 航空学报, 2013, 34(9): 2219-2229.
	Chen X Z, Yu J S, Tang D Y, et al. Probabilistic prediction of lithium battery remaining life based on Bayesian LS-SVR[J]. Acta Aeronautica et Astronautica Sinica, 2013, 34(9): 2219-2229.
[17]	李超然, 肖飞, 樊亚翔, 等. 基于卷积神经网络的锂离子电池SOH估算[J]. 电工技术学报, 2020, 35(19): 4106-4119.
	Li C R, Xiao F, Fan Y X, et al. SOH estimation of lithium-ion battery based on convolutional neural network[J]. Transactions of China Electrotechnical Society, 2020, 35(19): 4106-4119.
[18]	陈丰, 徐欢. 基于MLP算法的电力企业数据健康度模型算法研究[J/OL]. 自动化技术与应用, 1-5[2025-10-12].
	Chen F, Xu H. Research on the algorithm of data health model for power enterprises based on MLP algorithm [J/OL]. Techniques of Automation and Applications, 1-5[2025-10-12].
[19]	张驰, 郭媛, 黎明. 人工神经网络模型发展及应用综述[J]. 计算机工程与应用, 2021, 57(11): 57-69.
	Zhang C, Guo Y, Li M. Review on the development and application of artificial neural network model[J]. Computer Engineering and Applications, 2021, 57(11): 57-69.
[20]	周敏. 中国石油炼化企业腐蚀与控制现状[J]. 腐蚀与防护, 2012, 33(S2): 62-68.
	Zhou M. Current situation of corrosion and control in PetroChina refining and chemical enterprises [J]. Corrosion & Protection, 2012, 33 (S2): 62-68.
[21]	匡正, 袁志波, 徐振磊. 基于HMM-MLP的泵站监测健康诊断系统研究[J]. 中国农村水利水电, 2024(7): 255-261, 269.
	Kuang Z, Yuan Z B, Xu Z L. Research on monitoring and health diagnosis system of pumping station based on HMM-MLP[J]. China Rural Water and Hydropower, 2024(7): 255-261, 269.
[22]	Huang X H, Xiao H W, Liu X C, et al. Quantitative analysis of steel composition based on multivariate Genetic Algorithm-Optimized Multilayer Perceptron (GA-MLP) model combined with Laser-Induced Breakdown Spectroscopy (LIBS)[J]. Microchemical Journal, 2025, 218: 114805.
[23]	来杰, 王晓丹, 向前, 等. 自编码器及其应用综述[J]. 通信学报, 2021, 42(9): 218-230.
	Lai J, Wang X D, Xiang Q, et al. Overview of self-encoder and its application[J]. Journal on Communications, 2021, 42(9): 218-230.
[24]	黄立威, 江碧涛, 吕守业, 等. 基于深度学习的推荐系统研究综述[J]. 计算机学报, 2018, 41(7): 1619-1647.
	Huang L W, Jiang B T, ( Lü/lv/lu/lyu) S Y, et al. Review of recommendation system based on deep learning[J]. Chinese Journal of Computers, 2018, 41(7): 1619-1647.
[25]	Liu C, Zhen J T, Shan W. Time series classification based on convolutional network with a Gated Linear Units kernel[J]. Engineering Applications of Artificial Intelligence, 2023, 123: 106296.
[26]	Jahan I, Ahmed M F, Ali M O, et al. Self-gated rectified linear unit for performance improvement of deep neural networks[J]. ICT Express, 2023, 9(3): 320-325.
[27]	裴洪, 胡昌华, 司小胜, 等. 基于机器学习的设备剩余寿命预测方法综述[J]. 机械工程学报, 2019, 55(8): 1-13.
	Pei H, Hu C H, Si X S, et al. Overview of the prediction methods of equipment remaining life based on machine learning[J]. Journal of Mechanical Engineering, 2019, 55(8): 1-13.
[28]	贾科, 杨哲, 魏超, 等. 基于斯皮尔曼等级相关系数的新能源送出线路纵联保护[J]. 电力系统自动化, 2020, 44(15): 103-111.
	Jia K, Yang Z, Wei C, et al. Pilot protection of new energy transmission line based on Spearman's rank correlation coefficient[J]. Automation of Electric Power Systems, 2020, 44(15): 103-111.
[29]	Zhang L, Wang L. Optimization of site investigation program for reliability assessment of undrained slope using Spearman rank correlation coefficient[J]. Computers and Geotechnics, 2023, 155: 105208.
[30]	American Petroleum Institute (API), American Society of Mechanical Engineers (ASME). Fitness-For-Service [S]. Washington, D.C.: American Petroleum Institute (API); New York, NY: American Society of Mechanical Engineers (ASME), 2021: 301-600.
[31]	Luo L Y, Fan X Q, Wang M, et al. Two-stage deep complex gated convolution attention network for multi-channel noise suppression[J]. Signal, Image and Video Processing, 2025, 19(7): 512.
[32]	Liu S H, Wang Y S, Sun J K, et al. An efficient Spatial–Temporal model based on gated linear units for trajectory prediction[J]. Neurocomputing, 2022, 492: 593-600.