Decision support system to evaluate a vandalized and deteriorated oil pipeline transportation system using artificial intelligence techniques. Part 1: modeling

1 Introduction and related works

Global economic growth in 2020 slumped to −3.7% from −3.4% in the previous month after a stronger impact from COVID-19 triggered an unprecedented global shock in oil demand and sales (OPEC 2020). The oil industry in Mexico has suffered economic flaws due to price fluctuations in recent years and reached historic sales of USD −2.27/barrel in April 2020. However, the problem of the volume of transport by pipelines increased considerably in 2019, generating variations of −25.2% thousands of barrels (TB), Within the foreign trade of hydrocarbons, for example, Itsmo generated variations due to the transport of 86.7% thousands of barrels per day (TBD), which represents lost in millions of dollars (MD) of –87.5%. Similarly, the volume of crude oil exports by geographical destination decreased by −6.8% compared to 2018, this is due to a declining trend in all indicators with previous years. Table 1 represents the aforementioned information as well as the graphic behavior [E] of the foreign trade of hydrocarbons from 2015 to 2019 (Pemex 2019). Low profits limit the maintenance of facilities, some of which have a critical impact on the mechanical integrity due to corrosion and vandalism. Pipeline transport systems (gas pipelines, multiple pipelines, and oil pipelines), the latter with an operating life of over 60 years, running from January to September 2019 Report 10,200 clandestine intakes, with a recurrence every nationwide every 38 min and 39 s, which generates high risks for all living beings in the vicinity (Infobae 2018; Leifer et al. 2012; Observatorios 2019).

Table 1:

Volume and export of pipeline transportation from 2015 to 2019 (Pemex 2019).

Volume of transport by pipeline (TB)
Product	2015	2016	2017	2018	2019	Variation 2019/2018 (%)
Crude oil	387,674	383,344	298,238	255,418	245,977	−3.7
Gasoline	350,694	324,197	257,982	238,964	160,392	−32.9
Jet fuel	20,627	19,816	13,329	11,595	8599	−25.8
Diesel	120,602	111,620	83,450	78,783	37,920	−51.9
Fuel oil	39,074	40,308	34,622	31,495	8161	−74.1
Others	4,078	27,767	18,079	20,085	14,731	−26.7
Total annual	922,749	907,052	705,700	636,340	475,781	−25.2

Foreign trade of hydrocarbons (TBD) [A] Net exports (MD) [B]
	[A]	[B]	[A]	[B]	[A]	[B]	[A]	[B]	[A]	[B]	[A]	[B]
Crude oil	1,172	18,471	1,194	15,582	1,174	20,047	1,184	26,541	1,103	22,403	−6.8	−15.6
Olmeca	124	2333	108	1569	19	358	−	−	−	−	−	−
Istmo	194	3489	153	2108	86	1589	31	722	4	90	−86.7	−87.5
Maya	854	12,629	934	11,905	1,069	18,100	1,153	25,818	1,099	22,313	−4.7	−13.6
Exportsa		21,117		17,499		22,489		29,333		24,500		−16.5
Total annuala		263		−2377		−3987		−3355		−568		−83.1

Volume of crude oil exports by geographic destination (TBD) [C] Value of crude oil exports by geographic destination (MD) [D]
	[C]	[D]	[C]	[D]	[C]	[D]	[C]	[D]	[C]	[D]	[C]	[D]
America	705	11,599	604	8026	638	10,918	673	15,360	609	12,696	−9.5	−17.3
Europe	248	3733	272	3437	219	3656	199	4329	182	3527	−8.7	−18.5
Far east	219	3119	318	4119	317	5474	311	6851	312	6180	0.3	−9.8
Total	1.172	18,451	1,194	15,582	1,174	20,047	1,184	26,541	1,103	22,403	−6.8	−15.6
[E]

1. ^aIncludes condensates, dry natural gas, oil, and petrochemicals. TB, thousands of barrels; TBD, thousands of barrels per day; MD, millions of dollars.

The pipeline network has an operating length of over 17,000 km for the transport of hydrocarbons, oil, and petrochemical products. The case study is being developed at the Mendoza Distribution Center (MDC) in the states of Veracruz and Puebla with a length of 572 km, pipes of 24/30 “Ø, and a variable operating program of 90,000 to 140,000 barrels of oil transported per day (BTD) low demand and high system instability (Pemex Refinación 2011). The process is carried out with a sequential hydrocarbon pumping system that enables the flow of product, powered by 6 stations, supported by telemetry and supervision, data acquisition and control system (SCADA), 96 isolation valves on the line, 17 check valves, 28 devil traps, and remote-controlled instrumentation kit. This study evaluates the effects of corrosion on the mechanical integrity of the PTS using three techniques that can be used to reconfigure and guarantee the operational reliability of its process. A specific contribution to knowledge is made by developing models that facilitate the assessment and management of a complex system, through internationally validated techniques. It is known that AI is used to solving problems in the oil and gas industry, as well as to assess and predict the behavior of the various corrosion phases in PTS (Din et al. 2015; Kaduková et al. 2014; Lu et al. 2019), where international standards are frequently discussed (Kaewpradap et al. 2017; Lee et al. 2015). Various authors have cited AI’s adaptability to traditional techniques to evaluate and predict its various classes of corrosion behavior (Sweet corrosion “CO2”; Acid corrosion “H2S”; Corrosion to oxygen “O2”; Microbiological corrosion “MIC”; Galvanic corrosion “GC”) for calculating pH values in soils, effects of organic acids, protective films, wetting, hydrogen sulfide, and flow properties in pipelines. The assessment and results of various corrosive prediction models versus field data have shown variability, making it impossible to explain which model is better than another (Nyborg 2010), However, the trend of these supported AI models can be indicated. The authors presented a systematic literature review (SLR) that deepens the current state-of-the-art with 245 works published between 1994 and 2018 that show an increasing interest in AI in related areas of the oil and gas industry (Cid-Galiot et al. 2021), As shown in Figure 1.

Figure 1:

Systematic literature review.

FES unifies the knowledge of corrosion specialists and MIS, which can be used to identify the evolution patterns of corrosion. Fuzzy logic (FL) has proven results for the analysis of corrosion behavior (Guzman Urbina and Aoyama 2018; Jamshidi et al. 2013; Shahriar et al. 2012), leaks in gas pipelines, and fluid dynamics, which demonstrate their ability to derive uncertainty indices collectively or individually within a decision-making process (Bertuccio and Biezma 2012; Biezma et al. 2018; Kabir et al. 2016). An ANN is used to predicting the corrosion inhibition capacity of Mexican hydrocarbon mixtures based on their properties towards carbon steel. Its use in predicting corrosion behavior patterns in PTS is very broad and diverse in the selection of variables (Chamkalani et al. 2013; Hernández et al. 2002; Hernández et al. 2006; Ifezue and Tobins 2016; Liao et al. 2011; Supriyatman et al. 2012), which maximizes results when working with different techniques (Altabey 2016; Liao et al. 2012; Liu et al. 2012; Mohamadi Baghmolaei et al. 2014). The assessment of the PTS is developed individually, considering field data, laboratory, and cognitive knowledge, to achieve the best results and to be replicated worldwide under similar conditions. Below the authors mention the purpose, design/methodology/approach, findings, practical implications, and originality/value, which generated the research.

2 Methodology and results

The methodology underpins its success by considering the eight key issues for the discipline of decision support systems shown in Figure 2 (Arnott and Pervan 2008).

1. The relevance of DSS research, by connecting the practical part (Kaewpradap et al. 2017; Lee et al. 2015) with the theory (Cid-Galiot et al. 2021).

2. DSS research methods and paradigms, with case study research and design sciences (Din et al. 2015; Kaduková et al. 2014; Lu et al. 2019).

3. The theoretical foundations of DSS research, considering explicit bases in the decision-making judgment through theoretical and practical foundations applied in a case study in Mexico.

4. The role of the information technology (IT) artifact in DSS research, fundamental in all the processes to analyze, interpret, and connect the information.

5. Funding for DSS research, no implicit support from any organization. Compliance with keys 6, 7, and 8 is generated by considering the largest number of interpretive case studies and the academic rigor to support the research designs. Fuzzy logic (FL) has proven results for the analysis of corrosion behavior (Guzman Urbina and Aoyama 2018; Jamshidi et al. 2013; Shahriar et al. 2012) and ANN use in predicting corrosion behavior patterns in PTS is very broad and diverse in the selection of variables (Chamkalani et al. 2013; Hernández et al. 2002, 2006; Ifezue and Tobins 2016; Liao et al. 2011; Supriyatman et al. 2012). A three-stage structure is proposed for the DSS: the structure is based on the following three modules:

   1. Module 1: evaluation of the mechanical integrity of the PTS with an FES (FES, evaluates, analyzes, and predicts the status of PTS, using field data and cognitive knowledge)

   2. Module 2: ANN predicts the inhibition capacity of Mexican hydrocarbons in pipeline corrosion (Mexican mixtures of crude oil Istmo, Olmeca, and Maya)

   3. Module 3: operational and economic analysis with MCS (The MCS determines the analysis of the operational and economic risks of PTS under the effects of corrosives to inflation, oil supply, and demand is presented in Part 2 of the article).

Figure 2:

Methodology of DSS.

The technicians use the Matlab toolbox for FES development, for the design of ANN, and Palisade’s @Risk, for the design of MCS. Each procedure was developed following international regulations. The integration of the three modules to evaluate the PTS generates a comprehensive qualitative perspective of the current operational reliability of the system and a quantitative one when analyzing the desired module to determine decisions. The SSD approach, as an intelligent evaluation tool for operational reliability in pipelines, generates notable advantages over traditional procedures, as well as disadvantages by generating changes in the organization’s culture, as shown in Table 2 below.

3 Development of the DSS

3.1 Evaluation of the mechanical integrity of the PTS with an FES

3.1.1 Development of the architecture of the FES

The MIS, developed in the MDC PTS, from S-1 to S-7 (572 km), in the 24”–30” Ø Pipeline, through international procedures (Anon 1983) includes the internal inspection severity analysis results using straight beam ultrasonic technology inspection equipment. The study revealed 4651 indications of metal loss, 151 geometry defects, 22 mid-wall defects, and 837 registered anomalies (Kulikov et al. 2011). The methodology consists of the statistical analysis of the data and the development of FES, the classification of the steps is presented in Figure 3.

1. Step 1: Preliminary analysis of the data (In the preliminary analysis, the databases of the mechanical integrity study were examined to find obvious patterns that might exist within the data. In this case, the knowledge of the specialists in the area was also interpreted).

2. Step 2: Statistical analysis using commercially available statistical software (The statistical analysis of the data was performed with MINITAB^®. Goodness-of-fit tests were performed to verify if the data observed in a random sample fit with some level of significance to a given probability distribution).

3. Step 3: FES model construction (Various architectures were considered as well as adjustable distributions to the data of the variables, under Mandani and Sugeno approaches in MATLAB^®).

4. Step 4: Several tests are executed on the selected model (Several executions of the FES, randomizing data, to ensure correct interpretation and estimation of the data).

5. Step 5: Reliability analysis sis to identify variability between real and virtual system data (The reliability of the FES was validated using data obtained from the real system and those estimated by the Fuzzy expert system, in order to identify the variability between the normalized outputs).

Figure 3:

Development of the FES.

The MIS is very important and difficult to replicate because MIS is very costly to the company’s operating costs.

The focus of your data processing is to regain the research database, critical areas due to vandalism, maintenance plans and inspection sheets for the points not visited, as well as the cognitive knowledge of the professional staff to determine the behavior and growth of the various phases of corrosion that are internal or external to the walls of the pipeline, using an FES. The system was divided into two categories: deterministic variables (DVs) referring to the inputs and uncertain variables (UVs) related to the outputs. The DVs are described below: TIW, thickness in the depth of the walls (inside and outside) of the pipe due to metallic discontinuity. MLP of the TIW, percentage of pipe wall thinning due to mechanical, manufacturing, or corrosion origins. ILW, dimensions of defects in the walls of the pipeline associated with debris or cracks. WIW, length of the corrosive inclusions on the pipe walls. WALL, pipeline participation is internal (1) or external (−1). REF determines the resistance process in the pipeline to repair it following the international standards defined by Eqs. (1)–(6) (Zhou and Huang 2012).

(1)REF=MAOPP safe

The maximum safe pressure P_safe for a duct with a metal loss indicator is defined as:

(2)P safe=P yield∗(1−Q1−QM)

where:

(3)P yield=2tD (SMYS+68.99.8067)∗Fd,

(4)for A=0.85×L×d y A0=L×t Q=AAO,

(5)for L2d×t≤50 M=1+0.6275L2D×t−0.003375L4D2×t2,

(6)and to L2d×t≥50 M=0.032×L2d×t+3.3

If the maximum allowable pressure of the defective pipeline is less than MAOP, the defect must be repaired. Uncertain variables (UV); based on the conditions of use time, PTS operation, and structural reliability. The architecture of DV and UV is based on the acquisition of parameters extracted from the MIS of the Mexican oil and gas industry PTS, the parameters of FES, and their linguistic and mathematical classification of the Mamdani model. It is described in Table 3.

Table 3:

FES parameters.

Deterministic variables inputs (unit)	Linguistic label	Membership function	Interval
TIW (mm)	Low	1-Gama	(0, 0.1, 5, 12)
	Regular	Triangular	(5.1, 12, 15)
	Good	Gama	(12, 15, 20, 20.1)
MLP of the TIW (%)	Low	1-Gama	(0, 0.1, 20, 50)
	Medium	Triangular	(20.1, 50, 60)
	High	Gama	(50.1, 60, 99.9, 100)
ILW (mm)	Good	1-Gama	(0, 0.1, 1000, 3000)
	Regular	Triangular	(1000.1,3000, 4000)
	Bad	Gama	(3000.1, 4000, 7999.9, 8000)
WIW (mm)	Good	1-Gama	(0, 0.1, 1000, 1500)
	Regular	Triangular	(1000.1, 1500, 2000)
	Bad	Gama	(1500.1, 2000, 2999.9, 3000)
WALL (+/−)	Internal	1-Gama	(0, 0.1, 0.5, 1)
	External	Gama	(0, −0,1, −0.5, −1)
REF (1:1)	Less than one	1-Gama	(0, 0.1, 0.5, 1)
	Greater than one	Gama	(1, 1.01, 1.5, 3)
Uncertain variable (output) The PTS operational and structural reliability (%)	Out of operation	1-Gama	(0, 0.01, 0.29, 0.30)
	Deficient	Trapezoidal	(0.30, 0.31, 0.49, 0.50)
	Stable	Trapezoidal	(0.50, 0.51, 0.79, 0.80)
	Desirable	Gama	(0.80, 0.81, 0.99, 1)

The handling of 1126 MIS data for 137 km of pipes from (E‐5A) to (E‐6), which are regarded as a critical area due to their high vandalism rates and robust databases, is exemplified to model the FES appropriately. The cognitive knowledge of the specialists, the basis of the IF-THEN rules of FES, a rule set is used to relate the Fuzzy variables to the classification of the results. A common Fuzzy rule relates m precedence variables X₁,… X_m to n consequential variables Y₁,… Y_n (Zadeh 1994). The interaction between DVs and UV was determined by 162 Fuzzy rules. The defuzzification process enables the interaction of Fuzzy sets (DVs, UV, and rules) to be translated into clear and defined values using the area centroid (COA) method, where Z^*COA is the crisp value for the “z” output and μ_A (z) is the aggregated output membership function, Eq. (7) (Daftaribesheli et al. 2011).

(7)Z∗COA=∫ZμA(z)zdz∫zμA (Z)Zdz

3.1.2 Tests, validation and results of the FES

The Fuzzy interaction between DVs in relation to UV is shown in Figure 3 [A], Test No. 1: TIW = 0.20 mm, MLP of TIW = 0.20 mm, ILW = 0.7 mm, with a WIW = 1000 mm, WALL = +1 and a REF <1. Displays the result of the UV radiation, establishes the desired conditions for the operational and structural reliability of the PTS, and estimates a numerical value of 90.1%. A total of 1070 tests were developed to evaluate the effectiveness of the FES and to optimize the results of the MIS. Considering the reliability of 95% of the data, Figure [B] reflects a positive linear correlation and a coefficient of determination (R²) of 0.9029. Calculated by the quotient between the covariance and the product of the standard deviations of both output variables (FES results and MIS results), which is used to determine the operational and structural reliability of the PTS (%).

The reliability of the FES was validated using data obtained from the real system “y (t) = MIS” and those estimated by the Fuzzy expert system “ỹ (t) = FES”, in order to identify the variability between the normalized outputs “σ (t) = y(t) − ỹ(t)”. Table 4 presents a representative sample of the population of real and estimated deterministic variables, with 600 data previously defined by a factorial design of experiments. Figure [C] represents an amplified view of the behavior of the variables, which facilitates the visual identification of their variance, representing low biases between the data not greater than 7% as shown by the variable ILW. Figure [D] shows a general and structural evaluation of the FES under risk scenarios that identify the evolutionary impact of the corrosion patterns with respect to the operating conditions of the system, which allows standardizing the current working conditions of the PTS and obtaining a maximum operating confidence interval of 66.7%. Determining that the PTS will work stably to ensure the transmission of 306 and 820 BTD. However, due to the global downturn in demand, in recent years, PTS performance indicators are declining, which is not satisfactory but currently incapable of reliably delivering large volumes of oil due to the effects of vandalism, mechanical integrity, corrosion, and low maintenance coverage, supplemented by an inadequate cathodic protection system and low level of corrosion inhibitor injection into the system, limiting compliance with historical operating programs, such as a plan established in 2008 with 460,000 BTD (Pemex Refinación 2011), At that time, the price of the Mexican blend reached an all-time high of US$132 per barrel (BANXICO 2020). The FES shows an evolutionary representation of the corrosion patterns with respect to the operating parameters over time in a general way in a certain section of the pipeline, directly contributing to new maintenance planning strategies and generating new practices that determine more reliable operating conditions for the company without generating losses. At present, the authors are developing an ANN that will estimate and classify in detail the evolutionary behavior of corrosion in pipes using X-rays of the system provided by straight beam ultrasound inspection equipment through image processing, which will serve as a complete tool to optimize results when intelligent module presented.

Table 4:

Informative sample of the evaluation of the DVs of the MIS versus FES and operational behavior of corrosion in oil pipelines.

Normalized deterministic variables of the measures & estimated process: σ (t) = y(t) – ỹ(t).
MIS = y(t) & FES = ỹ(t)
TIW	ỹ(t) = TIW	MLP of TIW	ỹ(t) = MLP of TIW	ILW	ỹ(t) = ILW	WIW	ỹ(t) = WIW	WALL	ỹ(t) = WALL	REF	ỹ(t) = REF
0.1097	0.1227	1.9102	1.9177	1.1431	1.143	1.5958	1.5951	1.5382	1.538	0.3883	0.3888
0.7799	0.7000	0.5857	0.5802	1.597	1.502	0.5427	0.5467	1.8537	1.8598	0.2621	0.2754
0.8549	0.8629	1.4112	1.4178	1.5142	1.5187	0.4535	0.4638	1.1122	1.1267	0.7063	0.7061
0.1126	0.1100	0.8317	0.8576	0.1217	0.1365	0.1911	0.1965	1.8781	1.8712	1.6546	1.6577
0.7371	0.6811	0.1004	0.1001	0.5153	0.5106	0.7174	0.7191	0.5532	0.5743	1.2859	1.2802
0.4259	0.4051	0.7725	0.7973	0.3666	0.369	1.8735	1.8534	0.2981	0.2919	1.8424	1.8496
0.1739	0.1842	1.6956	1.6958	0.1616	0.167	0.2136	0.2189	1.2784	1.2703	1.087	1.0833
0.2505	0.2732	0.8187	0.8385	1.5254	1.5287	1.021	1.0213	1.2595	1.2643	1.8364	1.8487
0.1368	0.1302	1.8389	1.8576	0.3962	0.3961	0.6904	0.6634	0.1614	0.1611	0.4803	0.4868
0.5172	0.5322	1.7994	1.7583	1.8474	1.8472	1.2705	1.2619	1.9902	1.9919	1.4013	1.4087
[C]
[D]

3.2 ANN predicts the inhibition capacity of Mexican hydrocarbons in pipeline corrosion

3.2.1 Development of the architecture of the ANN

The measurement methods commonly used to measure corrosion in pipelines contain various sources of error that cast doubt on the reliability of the data. However, there are general practices in considering linear corrosion growths at any point, the reasoning making two assumptions that are difficult to verify, the distribution of corrosion data in a particular area of the pipe that maintains normal or Gaussian behavior, and the growth of defects over time are linear (Fernando Tomati 2011). For all the above methods, an alternative method is proposed which assumes that the measurement error is inherent to the process (noise), and then processes the results based on ANN, metallography, and statistics to extract valid conclusions from the available data. Usually the typical behavior of using log-normal, triangular, or Weibull distribution (Mohd et al. 2014).

Table 5 shows the equations of the physicochemical reactions that produce corrosion in pipelines caused by carbon dioxide (8–9) and shows that the reduction of hydrogen ions and the direct reduction of carbonic acid generally increase the rate of corrosion of carbon steel (10–11). When the water phases are tied to API gravity, the velocity of the fluid and the position of the pipeline can create changes to its surface by increasing the stress (12–13) on which the oil is exposed and water cuts for their separation create friction in the entrainment phase, degrade diameters and change the droplet size on the duct walls (14–23). However, this means that different hydrocarbon mixtures affect corrosion differently, even if they are tested under the same hydrodynamic conditions.

Table 5:

Oil and pipeline corrosion.

(8)	Fe + H2O + CO2 → FeCO3 + H2
(9)	Fe → Fe2+ 2e−
(10)	2H+ + 2e− → H2
	2H2CO3 + 2e− → H2 + 2HCO−3
(11)	log ( V corr ) = 5.8 − 1710 273 + t + 0.67   log (pCO2)
(12)	F oil = 4.3 WU liq + 0.545 α90 + 4.3 WUliqα90 (Foil ≤ 1)
(13)	W break = −0.0166 × API + 0.83
	50 > API > 20
(14)	F oil = 0.071 WWbreakUliqW < Wbreak,Foil ≤ 1
(14.1)	F oil = 0.071 W W break U liq 0.545   α 90   0.071 α 90   W W break U liq
	W ≥ Wbreak,Foil ≤ 1
(15)	F oil ≥ W
(15.1)	When Uliq<α90,Foil=1
(16)	( d m a x D ) dilute = 1.88 [Po(1−ϵw)Pm]−0.4 Weo−0.6 Reo0.08
	Where: Reo = PoDU2cηo  Weo = PoDU2cσ
(17)	( 1 − ϵ w ) P o P m ≅ 1
(18)	( d max D ) dense = 2.22C 0.6H(PoU2cDσ)−0.6(ϵw1−ϵw)0.6[PmPo(1−ϵw)F]-0.4F=0.046/Re0.2o
(19)	d max D = Max { ( d max D ) dilute  , ( d max D ) dense }
(20)	( d c b D ) = 3 8 P o | Δ p | F U 2 c D g cos ( θ ) = 3 8 F P o Δ p F r 0 Where Froude number is: Fr0 = U2cDgcos(θ) and Δp=|P0−Pw|
(21)	( d c σ D ) = [ 0.4 σ | P 0 − P w |   g D 2 cos ( β ) ] 0.5
	β = { | θ |                         | θ | < 45 ° 90 − | θ |     | θ | > 45 ° }
(22)	( d crit D ) = Min { ( d c b D ) . ( d c σ D ) }
(23)	d max ≤ d crit

1. Based on: Barnea 1987; Brauner 2001; Brodkey and Talbot 1969; Cai et al. 2004; Craig 1998; De Waard and Lotz 1993; Hernández et al. 2006.

Corrosion occurs in a PTS due to various conditions and variables. The authors agree that the analysis of its spread must begin with the analysis of the effect produced by the product to be transported (chemical nature of the hydrocarbon mixtures) (Efird and Jasinski 1989; Hernández et al. 2002, 2006; Supriyatman et al. 2012). Because of this, this research evaluates the usefulness of ANNs in predicting the corrosion inhibition capacity of Mexican hydrocarbon mixtures based on their properties compared to carbon steel, which will define the foundations of a robust intelligent system in development. Because of their territorial diversity, Mexican hydrocarbons retain various classifications of naphthenic, paraffinic, and asphaltic types (Servicio Geológico Mexicano 2017; Tissot and Welte 1984), using field and laboratory experimental data to analyze the physical and chemical properties of the hydrocarbon to develop an effective neural model that can predict the ability of the hydrocarbon to protect against environmental corrosion in pipelines, a sensitivity analysis (SA) which was used to extract the variables with the greatest impact on the reaction. The methodology consists of the statistical analysis of the data and the development of the neural network model (Jose et al. 2001), the classification of the steps is presented in Figure 4.

1. Step 1: Preliminary analysis of the data (In the preliminary analysis various graphs and scatter plots of the data were examined to find out obvious patterns that might exist within the data. In this case no distinction was made between paraffinic and asphaltenic crude oils, hoping that the ANN would generate a model that could be indistinctively used for all types of crude oils).

2. Step 2: Statistical analysis using commercially available statistical software (Statistical analysis of the data was performed using MINITAB^®. Multiple regression analysis was performed to come up with a regression equation that would be able to explain how the model could be augmented by knowing any possible linear relationships among each of the input variables and the output).

3. Step 3: Neural network model construction (A number of different architectures, such as multilayer perceptron, generalized feed-forward network, modular neural network, and others, were considered for the ANN model was performed using MATLAB).

4. Step 4: Multiple test runs on the selected model (A number of runs on the selected ANN were performed by randomizing the data, to ensure that the network was able to understand, interpret and learn from the data).

5. Step 5: Sensitivity analysis to find how input variables (chemical constituents) affect the output (percentage inhibition) and their correlation with the other variables [Sensitivity analysis was performed for the chosen MLP network and for all the test runs for that particular model. The sensitivity was computed based on the corresponding difference (delta) in the output(s) as graphed using the max–min criteria of the output (inhibition)].

Figure 4:

The Fuzzy interaction between DVs in relation to UV. (A) 0.91 and (B) R²= 0.9029.

The variables used are numeric, with eleven describing the properties of the hydrocarbon and four being operational. The salt content, which is related to the excess chloride in the hydrocarbon, generates high corrosion rates (Drews 2008a). Water and sediment, their significant content, maximize the corrosion processes in carbon steel (Drews 2008c). Reid vapor pressure, hydrocarbon evaporation rates produce excellent corrosion that is related to the operating temperatures and pressures of the piping (Floyd 1992). Flow improvers, structured mainly of synthetic polymers with high viscoelastic properties to reduce friction in pipelines (Drews 2008b), pressure operative, the variable force exerted by the liquid on the surface of the pipeline, considered as a constant for Eq. (26), but outside the ANN modeled to analyze risk scenarios for the system. The components of the three main Mexican hydrocarbon mixtures Maya (heavy), Istmo (light), and Olmeca (extra light) are experimentally evaluated under the same conditions and separated by their properties (Table 6A), with their inhibitory capacity being evaluated byEqs. (24) and(25) which, through multiple linear regression analysis in MINITAB, identify a high variability between their components and determine a regression equation that makes it possible to explain the behavior of the predictor model by knowing the possible linear relationships between the output and input variables (Eq. (26)).

Table 6:

Physical–chemical properties of Mexican crudes (based on Pemex Refinación 2011) and multiple regression analysis results.

Physical–chemical properties of Mexican crudes (based on Pemex Refinación 2011)				Multiple regression analysis results
Component/predictor	Olmeca	Istmo	Maya	Coef	SE Coef	T	P
Aromatics (%weight)	34.8	41	32.6	−0.6913	0.3241	−2.13	0.032
Resins (%weight)	10.8	7.8	12.7	−0.8212	0.4606	−1.78	0.129
Saturates (%weight)	62.2	53.8	43.2	−0.6871	0.2914	−2.35	0.014
Asphaltenes (%weight)	0.58	1.2	10.6	−0.8637	0.4280	−2.01	0.047
API density @60 °F (°API)	38.0	32.0	21.0	0.018419	0.003820	4.82	0.000
Total of nitrogen (ppm)	920	1828	3368	0.00004618	0.00007514	0.61	0.476
Sulfur content (%weight)	1.0	1.6	3.6	−0.09330	0.04817	−1.93	0.114
Total acids (mg KOH/g)	0.10	0.21	0.28	0.09281	0.02291	4.05	0.000
Vanadium (ppm)	9	50	270	0.0001178	0.0001105	1.06	0.301
Nickel (ppm)	2	8	54	−0.004813	0.005407	−0.89	0.448
Crude oil (%weight)	89.2	89.2	72.0	0.0012593	0.0002201	5.72	0.000
Salt content (LB/MB)	50.0	50.0	50.0	−0.08510	0.04112	−2.06	0.052
Water and sediment (% volume)	0.5	0.5	0.5	−0.9318	0.4170	−2.23	0.002
Reid vapor pressure (LB/PULG2)	6.50	6.50	6.50	−0.00003312	0.00009582	−0.34	0.503
Flow improvers (% volume)	−	−	−	0.0010507	0.0002841	3.69	0.000
Pressure operative (constant Kgcm2 )	−	−	−	56.00	25.17	2.22	0.027

1. ppm, parts per million. The value for all parameters is considered maximum range, except API density considered minimum value. Other Mexican hydrocarbons; Naranjos, Alamo, Muro, Horcon, Marfoantares, Pozoleo, Papaloapan, Arenque, Altamira, and Panuco.

If Coef, y SE Coef, is divided to get the t-value (T), which is compared to the distribution and determines whether the predictor is significant, if its value is higher, it is more significant (Table 6B).

(24)Inhibiting capacity (paraniffinic)=−0.230+0.0026x (%crude oil)+0.0576x(resin)+0.000844x(N)−0.0333x(asphaltenes)

(25)Inhibiting capacity (asphaltenic)=0.48+0.004x (%crude oil)+0.0048x(resin+asphaltenes)+0.0436x(S)

(26)% Inhibition=56−0.6913 Aromatics−0.8212 Resins−0.6871 Saturates−0.8637 Asphaltenes+0.018419 API Density+0.00004618 Total of nitrogen−0.09330 Sulfur content+0.09281 Total acids+0.0001178 Vanadium−0.004813 Nickel+0.0012593 Crude oil−0.08510 Salt content−0.00003312 Reid vapor pressure+0.0010507 Flow improvers

The p-values for the estimated coefficients of API density, total acids, crude oil, and flow improver are 0.000, indicating that they are significantly related to % inhibition. The p-values above zero indicate that they are not related to inhibition at different levels of significance. The R-square value obtained in various simulations was 27.8%, which is quite low. It was analyzed using other methods such as stepwise regression and there was no significant improvement, suggesting that the relationship between the predictor and the response variables are not linear.

To determine the best supervised or unsupervised adaptive learning architecture for ANN, the 18 available in MATLAB software were evaluated using normalized laboratory data, various transfer functions, weighted connections, and backpropagation training algorithms, which are divided into heuristics and numerical optimization of the latter ANN with the best results using a Levengerd–Marquardt (Trainlm) algorithm, applied to a Perceptron supervised learning architecture with a LOGSING transfer function, with a best performance of 0.92 see Figure 5 [A]. The ANN comprises 15 input variables for predicting an output variable (% inhibition) and comprises fifteen neurons in its input layer, seven neurons in the hidden layer, and one neuron for the output layer [B].

Figure 5:

Neural network development.

3.2.2 Tests, validation and results of the ANN

A total of 43 networks were trained, the 40 configuration was the best with a training acceptance rate of 0.9657, and the tests of 0.9586 using 80% of the data to train the network, 35% for cross-validation, and 30% for tests, using a performance function (MSE), the network showed optimal results in its output variables without showing over-training problems with a validation of 0.9556, see Figure 6

Figure 6:

ANN’s architecture and structure. Radial basis (fewer neurons) (UL) - 0.14; radial basis (exact fit) (UL) - 0.35; probabilistic (UL) - 0.57; perceptron (SL) - 0.92; better performance NARX series parallel (UL) - 0.37; NARX (UL) - 0.03; LVQ (UL) - 0.27; linear layer (train) (UL) - 0.18; linear layer (design) (UL) - 0.11; layer recurrent (SL) - 0.71; hopfield (UL) - 0.4; generalized regression (UL) - 0.29; feed‐forward time delay (UL) - 0.52; feed‐forward distributed time delay (UL) - 0.21 feed‐forward backpropagation (SL) - 0.89; Elman backpropagation (SL) - 0.83; competitive (UL) - 0.46; cascade‐forward backpropagation (SL) - 0.76.

Using a sample of 746 laboratory data, weight loss corrosion test results were obtained with samples and various hydrocarbon-saline mixtures (2.6% NaCl), with the average operating conditions of the PTS (56 psi constant and 60 °C) being simulated by phases of water cuts from 25, 50, 75, and 95%, data from the investigation of the mechanical integrity and the physicochemical properties of hydrocarbons in steel samples API L X65 clean and corroded (sample extracted from PTS for fatigue), according to ASTM standard (Anon 1985; Drews 2008a,b,c), 6 metallographic samples of the steel were used to calculate the corrosion rate and for surface analysis and corrosive characterization. The samples were polished using various 1% silicon carbide and nital sandpapers to determine their microstructure of 9 and 12, respectively, which should be weighed on an analytical balance to determine corrosion rates according to ASTM G1-90, see Figure 6 [A] The inhibition capacity is obtained by dividing the value of the hydrocarbon metallographic tests and the clean material tests (Equation 27). An SA was applied using hydrocarbon concentrations of 1%, 30%, 50%, and 90% to analyze behavioral patterns inhibition that showed similarities between low and high ranges. The Figures [B–E], represent the behavior of the SA regarding the input variables crude oil, API density @ 60 °F, Total of nitrogen demonstrating that at higher concentrations its corrosive inhibition capacity (output variable) increases unlike Nickel, which decreases as its concentration increases, generating corrosive susceptibility to the pipeline. A tornado chart shows the relative contribution of each input variable and how it affects the output variable. The Figure [F] shows that the flow improver (21.21%), the API density (18.82%), and the nickel content (13.04%) have a more significant influence on the output variable (% inhibition). The rest of the variables have a decreasing and insignificant influence (Figure 7).

(27)Inhibiting capacity=1−Corrosion ratewith hydrocarbonCorrosion rateclean

Figure 7:

ANN testing, training, and validation. Output = 0.95 * target ÷ 0.28, Training: R = 0.9657, Target (o: Data, -: Fit, --:y = T); Output = 1 * target + 0.27, Test: R = 0.9586, Target (o : Data, - : Fit, -- : y =T); Output = 0.69 * target ÷ 5, Validation: R = 0.9556, Target (o : Data, - : Fit, --- : y =T); Output = 0.87 * target ÷ 2.2,  All: R = 0.9629; Target (o : Data, - : Fit, --- : y =T).

The ANN allows the user to plot the graph of the hydrocarbon versus its inhibitory capacity and shows the behavior of the Mayan, Isthmus, and Olmeca hydrocarbons with changes in concentration. At low concentrations (1 and 30%) the sensitivity of the hydrocarbon is high, with tests above 83% determining its low inhibitory capacity. However, at high concentrations (50 and 90%) the sensitivity is enhanced and the effects of the variables are not as relevant. The multiple regression analysis determined the correlation of the input variables with the knowledge of hydrocarbon chemistry, whereby the negative influence of saturated fatty acids, asphaltenes, and resins on the flow improver, the API density, and the sulfur strengthens nickel and vanadium. It is important to emphasize that the results presented here represent the field and experimental data relating to only three samples of Mexican hydrocarbon. The results should not be considered universal until the model is fully developed and additional data is used for calibration. Further research is needed to expand both the experimental database and modeling capacity, as well as generate the knowledge needed to determine the factors that govern the effect of hydrocarbons on the corrosion of carbon steel (Figure 8).

Figure 8:

Metallographic process, sensitivity analysis, and tornado diagram. [F] Flow improvers = 21.21, API Density @60 °F = 18.82, Nickel = 13.04, Crude oil = 10.26, Vanadium = 9.71, Resins = 8.19, Aromatics = 5.05, Saturates = 4.06, Total of nitrogen = 3.76, Asphaltenes = 2.81, Total acids = 1.68, Salt content = 1.36, Reíd vapor pressure =1.18, Sulfur content = 1.04, Water and sediment = 0.82.

4 Discussion

Corrosion is known as the deterioration or destruction of a material due to a chemical or electrochemical reaction with its environment, being an unbearable but predictable phenomenon in the face of international procedures and regulations capable of mitigating its effect for later control. The work presented does not stand out for generating a high immediate impact among the audience, but for establishing an integral vision that will help reduce the loss of properties due to corrosion in pipelines, due to its ability to model, identify, and adapt known corrosive parameters. obtained through field experimentation, laboratory, historical data, and cognitive knowledge, which established a new position of the corrosive behavior of the pipe before the physical–chemical characteristics of the transported product, which makes it possible to establish the use of new materials, alloys, heat treatments, coatings, cathodic protection, as well as inhibitors that prolong the mechanical integrity of a deteriorated and vandalized PTS, through the management and planning of best decision practices that, in collaboration with artificial intelligence, maximize the reliability of the system. The research supports its contribution to knowledge through a detailed exploration of previous works that generated a path to follow on the subject, as shown in Table 7 below.

The application of FL and MIS through nondestructive testing for probabilistic calculation of failure in corroded pipes using ANSI/ASME B31G codes in Mandani models shows results in oil and gas pipelines (Bertuccio and Biezma 2012; Dundulis et al. 2016; Zhou et al. 2016) where the presence of high concentrations of hydrogen sulfide (H₂S) in natural gas create corrosive environments, in which diffuse systems evaluate risk scenarios that ensure the reliability of the system and are reflected in millions of dollars (Guzman Urbina and Aoyama 2017). The authors give priority to developing a diffuse model that evaluates variables such as PH, resistivity, humidity, sulfates, chlorides, and redox potential in soils to determine external corrosion in pipelines (Biezma et al. 2018). The precision value of the ANN is reliable and comparable, but we can conclude that the corrosive prediction models of gas pipelines contain a broad field of investigation compared to oil pipelines (Chamkalani et al. 2013; Ifezue and Tobins 2016; Liao et al. 2011; Supriyatman et al. 2012), whereby the variables and their influence on the output impaired the studies presented (Efird and Jainski 1989; Hernández et al. 2002, 2006; Tissot and Welte 1984). Therefore, it would not be advisable to compare success rates between ANNs but to be homologous and persistent in the analysis regarding the importance of flow improver, API density, nickel in hydrocarbon, and its effect in maximizing inhibitory capacity, identified as tense active properties to lubricate the metal surface and facilitate the entrainment of water in the pipeline, as opposed to water and sediment, sulfur content, vapor pressure, which show a decrease in inhibiting capacity with increasing content. By adding variables such as temperature, age of the pipeline, and cathodic protection systems to the models, the knowledge is integrated into the ANN (Senouci et al. 2014; Shabarchin and Tesfamariam 2016). In the above work, assessments were developed using intelligent models to assess and predict risks within the PTS. This work is a pioneer for the Mexican oil and gas industry in the corrosion of pipelines. It is important to note that the main contribution of this research is to generate new knowledge to the current state-of-the-art through models that facilitate and support pipeline assessment procedures through AI in the world, which allows reconfiguring and generating new management methodologies in pipelines, guaranteeing operational reliability in its processes. This enables us to predict the impact of the most important factors, improve the mechanical integrity of the pipeline, and calculate the associated economic viability. To increase the effectiveness and reduce the errors of FES and ANN, it is recommended to include more variables that may arise because of new incidents in the system and to ensure the data guarantee representation and consistency.

5 Conclusions

Based on field data, laboratory, and cognitive knowledge, an intelligent assessment has been developed that allows us to analyze and predict the mechanical integrity of the PTS, the ability of a hydrocarbon to protect against internal corrosion in pipelines, and its economic impact on the profits of the company. The FES was able to interpret and predict general corrosive behaviors through field data to plan preventive maintenance and improve the mechanical integrity of the pipeline; however, this intelligent module is limited to identifying and classifying specific behaviors such as microbiological, galvanic corrosion, among others. The ANN was able to model and identify the inhibition capacity offered by the components of a Mexican hydrocarbon mixture through static procedures standardized by ASTM, limited to dynamic procedures or modeling of oil or gas products. The integration of the FES with the ANN does not generate mutual quantitative results; they only share field and laboratory databases that allow characterising the behavior of corrosion in carbon steels and acid soils, highlighting the management of new operating practices that for no reason they seek to replace the honorable work of the staff. More research is needed to expand both the experimental database and the modeling capabilities, as well as determine the knowledge governing the effect of hydrocarbons on corrosion. The calculation of the net benefit of the research is not established to determine whether the failure in the models is acceptable according to the current state of knowledge, but to interpret with variable flexibilities and results, since some non-quantifiable criteria may be important in assessing the overall risk, such as the cost of protecting life and the natural environment. The Part 2. – Analysis of the operational and economic risk of the PTS under corrosive effects, through Monte Carlo simulation (MCS), interprets information and knowledge of 564 km of soils, from the latest cathodic protection study, essential to determine corrosive profiles of the PTS, considering supply and demand effects, with 1095 data on the price of the Mexican mix from 2016 to 2019, as well as monthly inflation data for the same period, to generate financial estimates through the discounted cash flow (DCF) methodology, considering a wide range of risks and their correlation, incorporating changes over time under a corrosive impact approach, which is not a contribution fully comparable with the current works in the literature for its Innovation to model dynamic behaviors.

Nomenclature

Θ

Inclination of pipe

α

The angle of deviation (in degrees) of the tubing from the vertical.

ASTM

American Society for Testing and Materials

A

Area of indication of metal loss

A _o

Defect cross-sectional area with the same length equal to the length of the metal loss indication.

AI

Artificial Intelligence

ANN

Artificial neural network

API

American Petroleum Institute

BTD

Barrels transported per day

C

Corrosion

CK

Cognitive knowledge

C-PTS

Corrosion in pipeline transportation systems

C _H

Constant with the order of 0.

CO₂

Sweet corrosion

Coef

The regression coefficient for a given variable

COVID-19

Coronavirus diseases

COA

Area centroid method

D

External nominal length diameter

DCF

Discounted cash flows

d

Depth of metal loss in millimeters

DSS

Decision support system

d _max

Maximum droplet size between oil separation and water carry-over

d _crit

Critic droplet size between oil separation and water carry-over

d _cb

Drop size between oil separation and water carry-over

d _cσ

Critical drop diameter proposed by Brodkey

dilute

Oil-water dispersion

DVs

Deterministic variables

ɛ _w

Water cuts

F

Friction factor

F _oil

Water phases linked to API gravity in oil

Fd

Design factor

FL

Fuzzy logic

FES

Fuzzy expert system

GC

Galvanic corrosion

GDP

Gross domestic product

HD

Historical data

H₂S

Acid corrosion

ILW

Dimensions of defects in the walls of the pipeline, associated with scaling or cracking

MFL

Magnetic flux leakage

ρ

The density of the liquid in kg m⁻³

PIM

Management in the integrity of pipelines

Psi

Pounds-force per square inch

PTS

Pipeline transportation system

PSBC

Profit from the sale of a barrel of crude oil

L

Measured metal loss length

m

The oil phase, the oil–water mixture

M

Folias factor

Max

Maximum

MATLAB

MATrix LABoratory

MAOP

Maximum allowable operating pressure

MDC

Mendoza distribution center

MD

Millions of dollars

MIS

Mechanical integrity study

MSE

Mean square error

MI

Mechanical integrity

Min

Minimum

MIC

Microbiological corrosion

MLP of the TIW

Percentage of pipe wall thinning due to mechanical, manufacturing, or corrosion origins

η _o

The viscosity of oil phase, in Pa.s.

N

Number of values

NACE International

National Association of Corrosion Engineers

NaCI

Sodium chloride

o

The subscripts of the liquid density

O₂

Corrosion with oxygen

σ

Oil surface tension, in Nm⁻¹

OPEC =

Organization of the Petroleum Exporting Countries

p _Safe

Maximum safe pressure for a pipeline with an indication of metal loss

PCA

Principal components analysis

pH

Hydrogen potential

p-value (P)

Probability of getting a test statistic that is at least as extreme as the actual calculated value

S-1

Station 1 (KM 0) – Nuevo Teapa, Veracruz, Mexico

S-2

Station 2 (KM 166) – Loma Bonita, Oaxaca, Mexico

S-3

Station 3 (KM 277) – Arroyo Moreno, Veracruz, Mexico

S-4

Station 4 (KM 317) – Zapoapita, Veracruz, Mexico

S-5A

Station 5A (KM 351) – Ciudad Mendoza, Veracruz, Mexico

S-5

Station 5 (KM 362) – Maltrata, Veracruz, Mexico

S-6

Station 6 (KM 488) – San Martín Texmelucan, Puebla, Mexico

S-7

Station 7 (KM 572) – Venta de Carpio, State of Mexico, Mexico

SE Coef

Standard error of the coefficient

SA

Sensitivity analysis

Std Dev

Standard deviation

SMC

Simulation Monte Carlo

SCADA

Supervisory Control and Data Acquisition

SMYS

Specified minimum yield strength of the material of the pipeline

t

Nominal wall thickness

TB

Thousands of barrels

TBD

Thousands of barrels per day

TIW

Thickness at the depth of the walls (internal and external) of the pipe by metallic discontinuity

SLR

Systematic literary review

U _liq

Liquid velocity in m/s

U _c

The velocity of the continuous phase, in ms⁻¹

UVs

Uncertain variables

PTS

The pipeline transport system

R ²

Coefficient of determination

RS

Regulations and Standards

REF

Repair estimate factor

V _corr

Corrosion rate

W

Average water fraction of the liquid measured at the wellhead

WALL

Pipeline involvement is internal (1) or external (−1)

W _break

Indication of the interfacial tension between the crude oil and the water

WIW

Length of corrosive inclusions on the pipe walls

y (t)

Normalized deterministc variables of the measures MIS

ỹ (t)

Normalized deterministic variables of the estimated FES