Integrated Fuzzy-Logic Diagnosis and Fault-Signature Analysis of Operational, Corrosion, and Structural Failure Scenarios in an Aged Pipeline Network

1. Introduction and Related Works

The oil industry faces significant challenges from the deterioration of its infrastructure, especially pipeline transport systems (PTS). Their failure would cause considerable economic losses and environmental and safety risks, with catastrophic consequences for nearby communities and ecosystems. PTSs often operate under unstable conditions caused by operational variations, water hammer effects, cavitation, and corrosive phenomena. If not controlled promptly, these can cause system malfunctions. In this context, precisely assessing the failure type — leak or rupture — is critical because their social, environmental, and economic impacts differ significantly [1].

Inline inspection is a fundamental tool for pipeline integrity management, using magnetic flux leakage (MFL) and ultrasound (UT) technologies. However, data quality depends critically on statistical calibration and measurement error quantification [2]. Although “pig” devices are essential in pigging protocols, their frequent use is limited by high cost and operational risks. Therefore, developing low-cost tools, such as instrumented caliper foam pigs, offers a viable alternative to increase inspection frequency [3].

Modern control systems need strategies that keep them stable, safe, and working well even when faults happen. Passive systems are built to handle certain faults and disturbances from the start, so they can respond quickly, but they cannot always tell what kind of fault has happened or how serious it is. Active systems, on the other hand, can detect, locate, and estimate faults. They use a fault detection and diagnostic module to adjust control actions, such as reassigning actuators, replacing sensors, adapting parameters, or predicting faults to find the best solution. Hybrid strategies use both approaches by combining a passive robust controller with active diagnostic and reconfiguration features. The passive part keeps the system stable for a short time, while the active part finds the fault and decides how to fix it. This approach is often used in critical systems because it lowers the risk between when a fault happens and when it is diagnosed. Figure 1 shows the main types of fault-tolerant control strategies.

A system can operate in four different states, depending on how severe a fault is: normal, acceptable, degraded, and critical. These states reflect the system’s performance, stability, safety, and the seriousness of any faults. When faults occur, the system can respond with actions like normal control, compensation, reconfiguration, isolation, or shutting down safely [A]. Fault-tolerant control starts by analyzing the system to find out its current state, including its parts and structure. During the diagnostic stage, faults are detected, measured, and isolated to see how serious they are and how they affect the process. Based on this, the system can reconfigure itself using sensors and actuators that follow a control law suited to the situation. Throughout this process, supervision ensures the system stays stable and prevents the fault from spreading [B] (See Figure 2).

In active control systems the application of AI predominates, ANNs, fuzzy logic, and evolutionary algorithms have established themselves as key tools for reservoir simulation, production optimization, and data mining [6], achieving success rates above 99% in predicting corrosion percentage; while backpropagation (BP) networks have proven effective in predicting stray current density in buried pipelines by modeling nonlinear relationships between soil resistivity, burial depth, and pipe-to-soil voltage [7]. Bayesian networks have gained ground in the quantitative risk assessment of offshore facilities by integrating multiple sources of uncertainty within a unified probabilistic framework [8], and Bayesian probabilistic approaches have demonstrated the ability to discriminate between failure types in corroded pipelines — leak versus rupture — outperforming traditional deterministic criteria by incorporating the random nature of corrosion parameters [1].

In the analysis of failures in land-based pipelines, expert systems based on ANNs have shown promising results for the automated identification of failure mechanisms. The construction of knowledge bases from hundreds of real cases, combined with multilayer perceptron inductive learning algorithms with backpropagation, allows for a significant reduction in classification errors and preserves expert knowledge in the face of specialized personnel turnover [9].

The collapse of oil prices in March 2020, with the OPEC basket at $33.92 per barrel, exposed the structural fragility of hydrocarbon-dependent economies and the need to manage future energy demand precisely [10]. In Mexico, this situation worsened with prices dropping to 10.61 dollars per barrel in April 2020 and production declining from 1.6 to 1.4 million barrels per day between 2019 and 2024. This decline compromised the maintenance of facilities, many of which have operated for over 60 years and whose infrastructure represents a potential investment exceeding 10 billion dollars [11,12]. Operational reliability analysis at Pemex processing centers shows that implementing maintenance information systems reduces costs and increases reliability [13]. Meanwhile, the systematic theft of hydrocarbons, with 10,591 clandestine pipeline taps recorded by the end of 2025 and losses of 7,650 million pesos in the first half of that year, requires robust statistical frameworks for defect assessment [12] and updated information systems capable of supporting real-time decision-making [13].

Supplementary Table S1 presents the operational statistics reported by Pemex’s Corporate Directorate of Planning, Coordination and Performance for 2019–2025. The information is divided into three main sections:

– Production and Process: includes the production and processing volumes of hydrocarbons, such as crude oil, natural gas, liquids, condensates, and petrochemicals. These data allow for the identification of trends in the company’s productive and transformation capacity.

– Value (MM$): reflects the economic value in millions of dollars of products derived from oil and natural gas (gasoline, diesel, jet fuel, liquefied gas, among others). This section is key to evaluating the financial impact of operations, as well as PEMEX’s dependence on certain products in the market.

– Prices (US$/b) – Export: shows the export prices of Mexican crude oil, broken down by grade (Istmo, Maya, Olmeca, and Zapoteco). These values illustrate the industry’s sensitivity to the volatility of international energy markets.

Integrity management in PTS is a multidimensional problem affected by multiple factors. Corrosion is a silent agent that can degrade the system’s operational capacity through various failure scenarios. Several authors have presented hybrid control strategies to mitigate its impact. One developed a BBN-GIS framework using a Bayesian neural network to predict depths of external corrosion defects, achieving an R² = 0.90, applied to 316,079 segments of experimental data. However, its scope is limited to risk prediction without any operational response mechanism [15].

Metallographic studies demonstrate that the acicular ferrite microstructure obtained by TMCP in X65MS steel exhibits superior resistance to HIC and SSCC, with CLR, CTR, and CSR parameters equal to zero. However, their contribution is strictly metallurgical and does not include monitoring or control systems [16]. The inclusion of MnS and iron carbides as primary crack initiation sites in X70 steel highlights that oxygenated environments produce corrosion rates comparable to those in saturated H₂S (Icorr of 3.76 × 10⁻³ and 4.65 × 10⁻³ A/cm², respectively), but without control or fault-tolerance systems [17]. More complex BBN models (more than 100 nodes and 3 million Monte Carlo simulations) validated against ILI inspections predicted wall loss of 25% compared to the measured 24%. However, they function as a deferred assessment tool without real-time response capability [18].

This research builds on previous approaches by developing an intelligent, remote, and non-intrusive fault-tolerant monitoring system for PTS subject to structural fatigue. The proposal integrates analysis of the system’s main operational components with structural and environmental information from its service environment. It combines mechanical integrity studies using straight-beam ultrasonic inspection, physicochemical and microbiological characterization of soils, anodic potential measurements, and static and dynamic tests on specimens from pipeline segments showing structural degradation. This integration provides a systemic analysis framework to understand mechanisms associated with the generation and evolution of main failure scenarios under corrosive conditions.

The methodology is evaluated through a case study in Mexico, where each subsystem is represented by fuzzy logic models built from experimental databases and validated expert knowledge. These models are integrated into a predictive neuro-fuzzy control architecture designed to detect, quantify, locate, and isolate system failures. Identified anomalous conditions are monitored and mitigated using a genetic algorithm that determines optimal reconfiguration actions to preserve reliability and operational continuity. Failure scenarios are modeled using Monte Carlo simulations to estimate their cumulative economic and financial impact on company profits. Finally, monitoring results and access to intelligent models are integrated into a decision-support interface to assist the human operator with risk assessment, intervention prioritization, and selection of maintenance and reconfiguration strategies.

2. Methodology

The research structure presents a hybrid fault-tolerant control architecture to develop a neuro-fuzzy predictive control model. In this model, fuzzy logic transforms operational data and expert knowledge into interpretable failure scenarios. A neural network learns the nonlinear relationships needed to diagnose, quantify, and isolate these failures through a specific architecture. A genetic algorithm determines reconfiguration strategies to minimize their effects. Monte Carlo simulation estimates the economic consequences under uncertainty. The results from the four modules are integrated into a decision support interface. This interface is validated by comparing the real and virtual behavior of the PTS. This process links technical diagnosis, fault tolerance, operational optimization, and economic evaluation within a single intelligent monitoring framework (See Figure 3).

The research began with a systematic literature review of 287 scientific articles published between 2000 and 2021. This analysis established the potential of AI techniques for developing fault-tolerant control systems and identified main gaps in their application to PTS [19]. Based on these findings, Stage 1 focused on characterizing the pipeline’s components and functional structure, organized into four critical subsystems. According to the reviewed literature, these subsystems had been studied mostly independently, without integration into a unified architecture for fault-tolerant diagnostics, monitoring, and control.

Subsystem 1, related to operational capacity, enabled the collection, integration, and interpretation of about 2.5 million records from six sequential hydrocarbon pumping stations. The infrastructure crosses a complex topographic profile, has been in service for over 75 years, and has a history of vandalism and structural fatigue. Operational information came from telemetry, SCADA systems, and management and maintenance records stored in SAP. Integrating these sources reconstructed the system’s historical behavior and identified patterns linked to performance losses, anomalous conditions, and potential failure scenarios [20].

Subsystem 2 addressed the cathodic protection system’s performance by characterizing the soil along 572 km of the right-of-way and analyzing potential pipe-soil interactions. The assessment showed that about 34% of the route had highly corrosive conditions or critical protection deficiencies. Measurements were taken at 89 evaluation points using the CIPS technique, within the Comprehensive Rehabilitation Project for the Anti-Corrosion Protection System of the 30”-24”-20” Ø (PXR-OP-SCC-SUD-GCM-L-13-14) [21]. This information identified segments most exposed to external corrosion and established their relationship to soil conditions and cathodic protection performance.

Subsystem 3 focused on the pipeline’s mechanical integrity. The ultrasonic inspection by Weatherford Corporation identified 5,661 anomalies, about 12% of which required priority attention due to their severity and potential impact on operational continuity [22]. These results provided an objective basis for locating critical segments, prioritizing interventions, and linking structural degradation to recorded operating and environmental conditions.

Finally, Subsystem 4 involved metallographic and mechanical characterization of API L X65 steel samples showing structural fatigue. Destructive and non-destructive tests followed procedures standardized by ASME and API. The analyses identified a predominantly ferritic-pearlitic microstructure and assessed changes in mechanical properties related to service life, cyclic loading, and degradation processes [23]. Together, the integration of the four subsystems provided a multidimensional representation of the pipeline, linking operational behavior, cathodic protection, mechanical integrity, and microstructural evolution within a single intelligent analysis architecture.

Stage 3 involves parallel modeling of the virtual system using four complementary approaches. Module A implements four Mamdani-type fuzzy inference systems in MATLAB, one for each subsystem. The membership functions and rule bases were defined using experimental data and expert knowledge, then adjusted and validated statistically. Together, these models represent 28 failure scenarios related to operating conditions, cathodic protection performance, mechanical integrity, and metallurgical degradation of the pipeline.

Module B integrates a neuro-fuzzy control architecture into Simulink that uses scenarios generated by the Mamdani models as input variables. The combination of the fuzzy inference mechanism and the neural network’s learning capacity enables estimation of the system’s response and classification of its behavior into global, critical, reconfigurable, and reliable regions. The model identifies the evolution of anomalous conditions and determines whether the system can remain operational, requires reconfiguration, or is approaching a critical condition.

Module C applies a genetic algorithm to minimize the consequences of detected failures. Objective functions were defined to reduce diagnostic signature breadth, preserve operational continuity, and guide the system from critical regions toward recovery or reliable operation. The output of this module is an optimized set of reconfiguration actions, selected according to safety, availability, and operational performance criteria.

Module D evaluates the economic consequences of failure scenarios using MCS. The model incorporates deterministic parameters such as costs, revenues, customs duties, inflation, downtime, and maintenance expenses, as well as uncertain parameters related to equipment reliability, station availability, the type of mixture transported, and corrosion risk. This integration enables probabilistic estimation of the temporal evolution of losses and their effect on the company’s profitability.

The integration of modules A, B, C, and D supports a predictive control strategy first applied to offline processes in a virtual environment. Results are presented through a decision-support interface that lets the operator graphically and statistically compare the model’s behavior with conditions in the real system. This paper presents the proposed architecture for integrating diagnostics, reconfiguration, economic evaluation, and operator support within a single intelligent monitoring framework.

Figure 4 presents the methodological structure of this article. Stages 1 and 2 cover analysis, diagnosis, and modeling of a pipeline transport network to develop an intelligent monitoring system for structural fatigue that is remote, non-intrusive, and fault-tolerant.

This study proposes an integrated and interpretable framework for diagnosing failure propagation in an aged pipeline network by combining operational, cathodic-protection, mechanical-integrity, and metallurgical information. Unlike conventional approaches that analyze corrosion, structural defects, or operating anomalies independently, the methodology represents 28 failure scenarios through Mamdani fuzzy models and integrates them into a fault-signature matrix. Its originality lies in combining large-scale historical SCADA, SAP, soil, CIPS, ultrasonic-inspection, and material-characterization data with validated expert knowledge to identify localized and system-wide degradation patterns. The framework preserves the traceability of each diagnosis through explicit variables, rules, and residual signatures. This provides greater interpretability than purely data-driven black-box models. Its practical value is the establishment of a structured basis for fault isolation, diagnostic signature breadth, maintenance prioritization, and the future development of neuro-fuzzy reconfiguration and fault-tolerant control strategies in PTS.

3. Results – System Analysis

3.1. Components and Structure

Fault-tolerant control in PTS is a fundamental tool for managing operational integrity in the oil industry. It involves detecting, isolating, and quantifying faults to determine recovery actions and achieve optimal performance within the transportation system. The Intelligent Monitoring System (IMS) evaluates the system through four main subsystems. Subsystem 1 analyzes the operational variables of the PTS linked to SCADA and SAP. Subsystem 2 evaluates the corrosion protection system to determine soil aggressiveness along the right-of-way. Subsystem 3 describes the intelligent ultrasonic inspection used to evaluate the mechanical integrity of the pipeline. Subsystem 4 characterizes the mechanical and microstructural properties of API L X65 steel through metallographic analysis. The graphical structure of the IMS is shown in Figure 5.

3.1.1. Subsystem 1 – Operational Capacity

The Mexican oil and gas industry operates a pipeline transport system (PTS) of more than 17,000 km in length, comprising 48 oil pipelines, 78 gas pipelines, 11 polyethylene pipelines, and four mixed pipelines, ensuring the timely supply of hydrocarbons and petrochemical products [11]. The diversity of Mexican crude oil blends, with their particular characteristics detailed in Table 2 Section A, significantly increases the complexity of the transport process.

The process begins with the reception of hydrocarbon at pumping station E-1 in Nuevo Teapa, Veracruz (Km 0). The flow continues sequentially through E-2 in Loma Bonita, Oaxaca (Km 166), E-3 in Arroyo Moreno, Veracruz (Km 277), E-4 in Zapoapita, Veracruz (Km 317), and E-5A in Ciudad Mendoza, Veracruz (Km 351). The latter is considered a power station due to the area’s topographic profile. The process continues through E-5 in Maltrata, Veracruz (Km 362), E-6 in San Martín Texmelucan, Puebla (Km 488), and ends at E-7 in Venta de Carpio, State of Mexico (Km 572).

The Mendoza Distribution Center (MDC) faces the challenge of stabilizing the operational reliability of the PTS to prevent hydraulic hammer and cavitation effects in a mechanically deteriorated system with a complex topographic profile. Table 1, Section A shows the Physical-Chemical Properties of Mexican Crude Oils. Section B shows the topographic and hydraulic profile for a pumping rate of 150 MBD on a 24” Ø line transporting MAYA crude. This crude has a constant viscosity of 21°API and a specific gravity of 0.88 Km/m³, considering the operational limits and the location of the main facilities above sea level.

The case study is developed at the MDC, Veracruz, Mexico, which operates with a program of 140 to 220 thousand barrels per day (MBD) [11]. Hydrocarbons are dispatched through 24” Ø and 30” Ø pipelines spanning 572 km, using a sequential pumping system driven by six stations assisted by SAP, telemetry, and a SCADA (Supervisory Control and Data Acquisition) system.

The distribution center has infrastructure comprising 96 line-sectioning valves, 17 check valves, 28 pig traps, and remote instrumentation. Table 2 Section A presents the detailed operational capacity, Section B describes the functional strategy of the re-pumping stations, which include metering, filtering, and recovery equipment.

From January to September 2024, PEMEX recorded 8,038 clandestine taps on petroleum and hydrocarbon pipelines nationwide, a 10.34% decrease from the same period in 2023. The frequency of taps was one every 49 minutes and 3 seconds for hydrocarbons and one every 8 hours and 37 minutes for liquefied petroleum gas. Hidalgo had the highest number of cases with 1,911 taps on hydrocarbon pipelines. Puebla led in liquefied petroleum gas with 354 cases. This article focuses on a 90 km section of the PTS, the E-5 to E-6 segment in Veracruz and Puebla. This area is critical due to its topography and high incidence of hydrocarbon theft. Some of these data are shown in Figure 6.

3.1.2. Design of the Tolerance Mechanisms of Subsystem 1

To develop fault-tolerant control, it is necessary to identify variables associated with faults that disrupt the system. These variables support the diagnostic framework and fault signature matrix, enabling the adaptive control law to generate appropriate regulations. The developed neuro-fuzzy control model is based on modeling fault scenarios generated by the subsystems’ input variables, allowing a quantitative representation of the associated factors.

For subsystem 1, the data were divided, normalized, and the initial model architecture defined. Goodness-of-fit tests and systematic graphical analysis helped understand the behavior of input variables under adverse conditions, based on simulations of possible combinations of associated variables. This facilitates interpreting their combined influence and the system’s sensitivity to fault scenarios. It allows isolating and examining the combined effect of factors in each simulation to improve control. This approach is useful when input parameters lack strict limits and expert knowledge is included through linguistic terms. Fifteen normalized variables within discrete ranges were represented using Gaussian, trapezoidal, triangular, or sigmoid membership functions. The choice depended on their operational behavior (see Table 3A) and failure scenarios (see Table 3B).

To prevent the combinatorial rule explosion caused by evaluating fifteen operational inputs simultaneously, the original monolithic Mamdani system was decomposed into a hierarchical fuzzy inference structure. The first layer included five local inference systems representing upstream-station condition, downstream-station condition, maintenance and human-resource readiness, instrumentation and isolation capability, and emergency and communication availability. Their outputs were normalized risk indices from 0 to 100, classified as low, acceptable, degraded, or critical. A second diagnostic layer combined only the indices directly linked to each operational failure scenario, generating FS-1 through FS-9. Pressure-related and structural scenarios also received the safe-pressure ratio and structural-fatigue index from the mechanical-integrity subsystem. The hierarchical decomposition reduced the maximum rule base from 201,326,592 Cartesian combinations to about 736 interpretable rules while preserving traceability of the original input variables and failure scenarios. This architecture also allows independent calibration and validation of each local module before integration into the fault-signature matrix.

Because the membership functions overlap, a given input vector can activate several rules with varying intensities. Sensitivity, consistency, and redundancy analyses identified rules with negligible activation and equivalent consequents. These rules were developed by consensus with six certified technicians (ECM0301 Risk and Mechanical Integrity Assessment of Pipelines for the Transportation/Collection of Hydrocarbons, Petroleum Products, and Chemicals) from Pemex Logistic. Figure 7 shows the architecture of fuzzy system one and the interaction between the input variables and the modeled failure scenarios. Figure 10 presents the architecture of fuzzy system one and the interaction between the input variables and the modeled failure scenarios.

$$N_{rules}{\displaystyle \prod _{i=1}^n}{m}_i$$

(1)

The reliability of subsystem 1 was validated using real-system data and fuzzy model estimates to identify variability in the normalized outputs. Table 4 presents a sample of six real and estimated failure scenarios, with 600 data points defined by a factorial experimental design. Figure [A] shows a magnified view of the variable behavior, allowing visual identification of its variance, which ranged from 2.16 to 9.21. The response surface diagrams represent the graphical behavior of the system under fuzzy inference rules. Figure [B] shows that when the permissible operating limits of a pipeline with metal loss indicators (MAOP) are exceeded, the failure scenario (FS-4) increases because the average discharge pressure capacity (55.50 kg/cm²), established by the ANSI/ASME B31G standard for a downstream station (DSDP), is exceeded and the open communication and SCADA (OCE) system for controlling and monitoring operating conditions is unavailable. Indicators below 1.0 increase the likelihood of system failure. Figure [C] shows the activation of failure scenarios (FS-8), triggered by leaks at a pipeline location due to hydrocarbon. This is represented by the variables (EE) and (OCE). These relate to the availability of emergency equipment and auxiliary power generators, as well as open communication and remote-control equipment in case of a product spill emergency. The indicators decrease to 100% and below 1.0, respectively, for the variables that increase the incidence of scenario failures, while indicators above 1.0 indicate the incidence of product leaks at a pipeline location.

3.2. Section Title

3.2.1. Subsystem 2 – Cathodic Protection

This corresponds to the comprehensive rehabilitation project of the anticorrosive protection system for the 24”–30” Ø pipeline, Nuevo Teapa–Venta de Carpio section (572 km). The evaluation includes CIPS, DCVG studies, localization of electrical shunts using PCM equipment, and analysis of soil pH and resistivity along the right-of-way (PXR-OP-SCC-SUD-GCM-L-13-14). Soil is the most complex electrolyte here, with pH and resistivity as the key parameters to quantify its corrosive aggressiveness. pH values below 5.5 indicate acidic conditions that accelerate metallic corrosion, showing a direct correlation between higher acidity and greater aggressiveness.

Resistivity (ρ), for its part, measures the opposition of the soil to the flow of electric charge and is expressed through the generalized Ohm’s Law:

ρ = R · (A/L)

(2)

R is the electrical resistance (Ω), A is the cross-sectional area (cm²), and L is the distance between electrodes (cm). The electrochemical current causing corrosion is estimated as:

I = V · A / (ρ · L)

(3)

This expression shows soil corrosivity is inversely proportional to resistivity (Corrosivity ∝ 1/ ρ): The lower the ρ, the greater the ionic conductivity and the higher the galvanic corrosion rate. Based on experimental correlations between measured ρ values and corrosion rates in buried pipelines, empirical classification ranges are established and shown in Table 5.

The CIPS study is an indirect inspection technique that determines the level of cathodic protection of a buried or submerged pipeline using a metallic structure detector, voltage rectifiers, and GPS. Pipe-to-soil potentials are recorded with the current switched on and off at every meter, following established regulations [21].

The results of the soil resistivity and pH study along the right-of-way (ROW) are divided into sections called segments. Segment No. 9 covers soils from E-5 to E-6. Figure 8 shows the pH profile along the segment, with pH values on the “y” axis at each distance point on the “x” axis [A]. A total of 911 soil readings were conducted from Km 408 to Km 498. The classification of results [B] determines an average soil pH of 5.8. The soil resistivity survey shows the calculated resistivity value on the “y” axis at each distance point on the “x” axis [C]. With 907 soil readings, the classification of results [D] determines an average soil resistivity of 2,506.1 Ω-cm.

According to the above data and the general study of the PTS, the percentage of recorded data is illustrated in Figure 9, showing that acidic and slightly acidic pH values are present in equal proportion [A], alongside highly corrosive and corrosive soil [B]. This must be considered when assessing the level of cathodic protection and the condition of the mechanical coating, since the aggressive nature of the soil affects the coating’s state.

The CIPS study of Segment No. 9 is shown in graph [C], influenced by 5 rectifiers (RPC) with significant differences in potential and resistance. Several fall below the reference threshold of -850 millivolts (mV) set by regulations [11], NACE-SP0169-2013 section 6.2.1.3 [27]. Based on the difference between On and Off (interruption) potentials, 82.90% of data points are above the -100 mV interruption threshold.

3.2.2. Design of the Tolerance Mechanisms of Subsystem 2

For subsystem 2, the data were divided and normalized, and the initial model architecture was defined. Goodness-of-fit tests and graphical analysis helped to understand the behavior of input variables under adverse conditions. This was based on simulations of possible combinations of associated variables. Five normalized variables within discrete ranges were represented by membership functions according to their behavior, along with five failure scenarios (see Table 6).

A total of 324 inference rules were evaluated, corresponding to the Cartesian product of the linguistic categories assigned to the five input variables. Sensitivity, consistency, and redundancy analyses identified rules with negligible activation and equivalent consequents. These rules were developed by consensus with 18 Level 1 and 2 certified technicians from Pemex Logístic and DTsi México. They specialize in AMPP standards, direct internal corrosion assessment (NACE SP0206-2006, NACE SP0208-2008), cathodic protection systems, and evaluation of coatings and corrosion inhibitors (NRF-005-Pemex-2009). Figure 10 presents the architecture of fuzzy system two and the interaction between the input variables and the modeled failure scenarios.

The reliability of subsystem 2 was validated using real system data and fuzzy model estimates to identify variability in normalized outputs, similar to subsystem 1, with variance ranges between 0.98 and 4.73. Figure 11 shows the response surface diagrams. Graph [A] represents the activation of FS-10, which decreases to zero if MCP and CCP variables are at optimal levels; otherwise, failure recurs. [B] The SpH variable shows that soils with pH below 5.5 are acidic and rapidly cause corrosion in bare steel. This is amplified by a deficient mechanical coating on at least 40% of the pipe (SRE), causing exponential activation of FS-11. [C] Soil Corrosive Aggressiveness Failure (FS-13) can be controlled with alkaline indicators (pH above 5.5) and desirable SRE indicators; otherwise, its activation increases. [D] The level of cathodic protection (CCP) and soil acidity (SpH) are key variables for reducing the high corrosive risk caused by soil resistivity to electrical charge flow (FS-14).

3.3. Section Title

3.3.1. Subsystem 3 – Mechanical Integrity (Ultrasonic Intelligent Pigging Inspection)

The mechanical integrity study (MIS) segment in 10 sections and was conducted on the 24”–30” Ø pipeline, Nuevo Teapa–Venta de Carpio section (572 km), from E-1 to E-7, using a straight-beam ultrasonic inspection tool with radially directed sensors to detect pipe wall defects. The analysis of Segment 9 revealed 4,651 metal loss indications, 151 geometry defects, 22 mid-wall defects, and 837 anomalies. Data processing considers the variables TIW, PTIW, LEIW, DIW, WALL, and ERF, whose relationships are defined by the equations in Table 7.

The first principal component represents 39% of the total variance. The variables most highly correlated with the first principal component (PC1) are PTIW (0.461), LEIW (0.502), and DIW. The first three principal components account for 73.8% of the data variation. Table 8 shows the principal component analysis (PCA) of Segment 9.

The scree plot orders the eigenvalues from largest to smallest. The trend in the magnitude and direction of the coefficients shows that the TIW causes significant variability in calculating the ERF component, as illustrated in Figure 12. The influence plot shows the behavior of the first two components. The outlier plot identifies 27% of data points above the reference line.

The scatter plot allows evaluation of the correlation between the principal components. Figure 13 shows clustering of data with an estimated repair factor between 0.5 and 0.9. This demonstrates a high degree of deterioration in the PTS. The indicators with the highest ERF are generated on the external wall of the pipeline.

3.3.2. Design of the Tolerance Mechanisms of Subsystem 3

For subsystem 3, the data were divided and normalized, and the initial model architecture was defined. Goodness-of-fit tests and graphical analysis helped to understand the behavior of input variables under adverse conditions. This was based on simulations of possible combinations of associated variables. Seven normalized variables within discrete ranges were represented by membership functions according to their behavior, along with five failure scenarios (See Table 9).

A total of 972 inference rules were evaluated, corresponding to the Cartesian product of the linguistic categories assigned to the seven input variables. Sensitivity, consistency, and redundancy analyses identified rules with negligible activation and equivalent consequents. These rules were developed by consensus with 18 Level 1 and 2 certified technicians from Pemex Logístic and DTsi México. They specialize in AMPP standards, direct internal corrosion assessment (NACE SP0206-2006, NACE SP0208-2008), cathodic protection systems, and evaluation of coatings and corrosion inhibitors (NRF-005-Pemex-2009). Figure 14 presents the architecture of fuzzy system third and the interaction between the input variables and the modeled failure scenarios.

The reliability of subsystem 3 was validated using real system data and fuzzy model estimates to identify variability in the normalized outputs, similar to subsystem 1, with variance ranges between 0.93 and 4.77. Figure 15 shows the response surface plots. [A] At low Psafe values, the FS-15 output stays between low and intermediate levels and is sensitive to PTIW. As Psafe rises to a critical region, the output abruptly increases and reaches a high level near 80. Beyond this threshold, PTIW’s effect diminishes, and FS-15 remains in a critical state. This shows Psafe is the dominant variable activating the high state of FS-15. [B] The FS-16 output shows a low output region when both variables are at certain lower intervals, surrounded by intermediate response zones. The combination of high PTIW and/or Psafe values leads to an upward trend, showing no single variable acts completely independently. [C] The FS-18 output has three clearly differentiated activation levels that increase stepwise: low, intermediate, and high. For low LEIW values, FS-18 stays around 20–40; as LEIW increases, the response moves to an intermediate region near 50–60 and finally reaches a high plateau near 100. Although TIW shifts some transitions, LEIW has the greatest influence on the output. [D] The FS-20 surface shows binary behavior. At low ERF values, the output stays near 40 for much of the WALL interval. When ERF exceeds a threshold, FS-20 abruptly changes to a high region near 80. Thus, ERF is the main activation variable, while WALL conditions the response’s specific configuration.

3.4. Section Title

3.4.1. Subsystem 4 – Metallographic Analysis

Presents the metallographic analysis of an out-of-service underground pipeline with over 75 years of operation to manage generational knowledge of corrosive patterns in land-based pipelines. The module integrates metallographic analysis, chemical etching, roughness analysis, ultrasonic thickness measurement, and tensile testing. The physical and chemical characteristics were compared against current national and international regulations enabling a specific characterization of API L X65 steel [28,29,30,31,32].

The study sample was extracted from a segment of a 24” Ø pipeline with a wall thickness between 8 and 11 mm, located in the municipality of Esperanza, Puebla, Mexico. The extraction took place in June 2023 after an emergency response to a hydrocarbon spill. Visually, the sample showed cracks, dents, generalized corrosion, and residuals of a protective coating designed for high-humidity and corrosive environments.

The longitudinal and transverse metallographic probes extracted from the segment were encapsulated in bakelite at 190 °C and 23 kN for 9 minutes. They were progressively polished with sandpapers of different grit sizes and finally polished with a 0.3 μm alumina solution at 300 RPM until a mirror-like surface was obtained. To reveal the steel microstructure, chemical etching with 1% Nital was applied for 25 seconds per probe.

Scanning electron microscopy (SEM) analysis at magnifications of up to 50 μm revealed in the API L X65 steel a ferritic-pearlitic microstructure with heterogeneous phase distribution, where pearlite acts as a cathode relative to ferrite, promoting localized dissolution of the material. The determined pearlite phase percentage was 1.039%, a condition that makes the material susceptible to stress corrosion cracking processes.

Grain size was calculated using the linear intercept method in accordance with ASTM E-112 standard, yielding an ASTM grain size No. 9 (Figure 16, Section A).

The chemical composition analysis (Figure 16, Section B) identified the steel as a high-carbon alloy (6.33 wt%) with iron as the main component (90.07 wt%), obtained at 20.0 kV. The chemical elements were uniformly distributed throughout the analyzed matrix, as detailed in Section [B1] for zones 1, 2, and 3. The chemical composition of the sample is presented in Table 10.

The tensile test was carried out to obtain the mechanical properties and determine the steel grade [34]. Illustration [A] in Figure 17 shows the dimensions under the ASTM E8/E8M standard of two probes used in these tests and the results obtained continuously [33]. Table 11 shows the length, width, and thickness measurements for each probe before testing, as well as their final characteristics. Probes 1 and 2 exhibit pitting or cracks from corrosion, indicating material fatigue. Six probes were used; only 2 representative probes are presented due to similar results and paper length constraints.

Probe 1: This probe shows a linear elongation-time relationship [B]. Over 83 seconds, it elongates up to 4.09 millimeters (mm). Illustration [C] shows the load-deformation relationship for the same probe. The tensile test results showed the sample has a yield strength of 394 megapascals (MPa) and a tensile strength of 501 MPa [D] on average.

Probe 2: The elongation-time graph for this probe is also linear, likely because it is an isotropic material [C]. Minimal variation in results was obtained, with a yield strength of 404 MPa and an ultimate tensile strength of 517 MPa [D]. The overall results for both probes are shown in Table 12.

Vickers and Rockwell hardness measurements were performed on API L X65 steel according to ASTM E-92 and ASTM E-18 standards, using a 1/16” ball indenter with a load of 30 Kp for 15 seconds [31]. The results showed negative percentage differences of 26.9% and 19.8% relative to the reference values in ASTM E-140, indicating mechanical degradation of the material, as illustrated in Figure 18.

Ultrasonic thickness measurement is done by timing how long an acoustic wave takes to travel the distance. The operating principle is illustrated in image [A] of Figure 19. The experiment used a QS5 device with various probes (7.87” & 1.96”), with a manufacturer-specified inter-layer thickness of 11 mm. The ultrasonic device maintains a nominal speed of 1.20 mm/second in longitudinal measurements.

With a perimeter of 39.37”, 800 measurements were obtained showing variability caused by diverse cavities in the material from advanced corrosion. Figure 19 illustrates ultrasonic results for Probe 1 [B] and metal loss indicators of 22% [C] compared to manufacturer specifications. Length measurements defined the microgeometry of surfaces for each probe after structural fatigue, using a Mitutoyo SJ-201 profilometer, identifying cavity variability from 905.51 to 1,023.62 micro-inches (μ in) caused by corrosive effects [D].

3.4.2. Design of the Tolerance Mechanisms of Subsystem 4

For subsystem 4, the data were divided and normalized, and the initial model architecture was defined. Goodness-of-fit tests and graphical analysis helped to understand the behavior of input variables under adverse conditions. This was based on simulations of possible combinations of associated variables. Seven normalized variables within discrete ranges were represented by membership functions according to their behavior, along with five failure scenarios (See Table 13).

A total of 2,187 inference rules were evaluated, corresponding to the Cartesian product of the linguistic categories assigned to the seven input variables. Sensitivity, consistency, and redundancy analyses identified rules with negligible activation and equivalent consequents. These rules were developed by consensus with 18 Level 1 and 2 certified technicians from Pemex Logístic and DTsi México. They specialize in AMPP standards, direct internal corrosion assessment (NACE SP0206-2006, NACE SP0208-2008), cathodic protection systems, and evaluation of coatings and corrosion inhibitors (NRF-005-Pemex-2009). Figure 20 presents the architecture of fourth fuzzy system and the interaction between the input variables and the modeled failure scenarios.

The reliability of subsystem 4 was validated using real-system data and fuzzy model estimates to identify variability in normalized outputs, similar to subsystem 1, with variance ranges between 1.02 and 3.90. Figure 21 shows the response surface diagrams. Graph [A] FS-22 remains nearly constant around 50 across almost the entire domain. Its variation is small, suggesting the fault is activated or deactivated gradually as the FPPS and GSSP indicators change, indicating both variables have limited influence. [B] The activation of FS-24 shows three clearly differentiated regions. MCC has the greatest influence on its activation, while FPPS determines the minimum response zone. [C] FS-25 shows MCC predominates in triggering the fault, while GSSP mainly modulates it, decreasing to about 40% without significant impact. [D] FS-28 exhibits complex behavior with a wide response range. It has a central region of intermediate values near 30–45, surrounded by upper zones near 50–55. The surface geometry shows a pronounced nonlinear interaction: the effect of GSSP changes according to the MCC level, and vice versa.

4. Design of the Diagnostic System

4.1. Detection, Isolation and Quantification of Failure Scenarios

Detecting faults in complex processes without available mathematical models is difficult; however, modeling and optimizing their behavior using AI is a prominent alternative today. Detecting fault scenarios allows for the diagnosis of nonlinear processes based on input/output data under normal operating conditions, enabling the quantification and isolation of faults in the system. This allows for the recovery of degraded operational phases through supervised reconfiguration.

The Fault Signature Matrix (FSM) is a tool used in the fault detection and isolation stage. Its function is to relate each possible system fault to the pattern of residuals, symptoms, or indicators expected when that fault occurs. Each row represents a residual, indicator, or symptom, and each column represents a possible fault.

$$FSM=\left[\begin{array}{ccc}{S}_{11}& \dots & S_{1m}\\ ⋮& \ddots & ⋮\\ S_{n1}& \dots & S_{nm}\end{array}\right]$$

(10)

where:

R_i = residual or diagnostic indicator

a_i = Residual activation

F_j = failure considered

S_ij = expected response of the residual R_i to the failure F_j

$$\mathbf{S}=\left|S_{ij}\right|\in {\left\{\mathrm{0,1}\right\}}^{mxn}$$

(11)

$$S_{ij}=\left\{\begin{array}{c}1,ifresidual{R}_{i}isactivatedbyfailure{F}_j\\ 0,Theresidualdoesnotrespondtothefailure\end{array}\right.$$

R_i (t) = y_i (t) – ŷ_i (t)

(12)

$$a=\left\{\begin{array}{c}1,\left|R_i\right|>T_i,\\ 0,\left|R_i\right|\le T_i,\end{array}\right.$$

(13)

where:

T_i = Diagnostic threshold

This methodology, based on a set of consistency indicators called analytical redundancy relationships (ARRs) and high-gain dynamic state observers, was applied to reconstruct the evolution of the detected fault. These relationships are derived from the connections between the system elements (Failure Scenario).

For example:

If r¹, r³, r⁵, r⁷ ≠ 0 Then = FS-1

If r², r³, r⁶, r⁷ ≠ 0 Then = FS-2

The matrix contains, in coded form, the dependency of a given failure scenario (matrix column) on each residual (matrix row). Fault isolation consists of finding which of the fault signatures in the matrix most closely approximates the signature found experimentally. The Diagnostic signature breadth used in this report is based on the count of residuals activated by each failure scenario out of a total of 28 possible, as shown in Table 14. The greater the number of activated residuals, the greater the propagation of the fault throughout the system and the more difficult its operational isolation becomes.

Table 15 presents the 28 scenarios ordered from highest to lowest number of activated residuals, with their assigned diagnostic signature breadth and the fault matrix allows us to visualize that scenario 3, 4, and 9 have a very close interaction with the entire system:

Diagnostic signature breadth represents the amplitude of the signature relative to interconnections with other subsystems, indicating diagnostic redundancy, potential propagation, and low isolation capability. However, greater residual activation does not necessarily mean greater physical damage. Conversely, a fault generating significant residuals can trigger multiple simultaneous faults, potentially causing the system to spiral out of control.

4.2. Key Findings

Scenario FS-3 is the only one that activates all 26 residuals, confirming the fault ma-trix observation that scenarios FS-3, FS-4, and FS-9 interact closely with the entire system. Both FS-3 and FS-4 belong to subsystem 1, which serves as a cross-cutting monitoring and control layer for all other subsystems. A failure in this subsystem compromises the overall diagnostic capability of the PTS.

The seven scenarios of subsystem 4 (FS-22 – FS-28) exhibit a systemic impact through 17 activated residuals (61% of the system). Pipeline material properties (yield strength, composition, grain size, and microstructure) are fundamental conditions governing the behavior of all other subsystems. A metallurgical alteration propagates as a critical failure signature through the residuals in the matrix.

Structural scenarios FS-15–FS-21 exhibited progressively narrower signatures, decreasing from 16 to 11 activated residuals. This pattern suggests increasing diagnostic specificity rather than broader systemic propagation. Structural defects in the pipeline wall (TIW, PTIW, DIW, LEIW, WALL, ERF, and P Safe) are interrelated with each other and with varia-bles in other subsystems as deterioration progresses. This makes late-stage failures significantly more difficult to isolate.

Localized failures are represented by low-impact scenarios such as FS-1, FS-2, FS-5, FS-6, and FS-7, which trigger between 4 and 7 residuals, less than 25% of the failure matrix. Their failure signature is unique and well-defined, facilitating diagnostic isolation and reducing operational response time.

Implementing the diagnostic system allows early identification of failure scenarios by comparing observed and theoretical signatures. This is essential for modeling neuro-fuzzy tolerant control, which facilitates operational decision-making and activates specific tolerance mechanisms for each detected failure. It provides a robust tool for continuous monitoring and proactive management of pipeline integrity.

4.3. Limitations and Scope

This work is a new, comprehensive methodological contribution to monitoring and managing corrosion in pipeline transport system through intelligent modeling of isolated parameters and variables that jointly influence the pipeline’s operational performance. However, its scope is limited by technical, operational, and contextual conditions to consider when interpreting results and planning implementation.

The system’s geographic and operational scope was developed and validated specifically for the Pemex Logística Nuevo Teapa-Tula pipeline (24” and 30” diameter, totaling 572 km), with emphasis on segments 9 and 8, covering 90 km along the Veracruz-Puebla border due to their high rates of vandalism.

The models were trained using historical data from 2010–2024, so their accuracy depends on the quality and continuous updating of databases in SAP and SCADA systems. Parameter recalibration is necessary when incorporating new data sources or extending to other sections of the national pipeline network.

The fuzzy logic-based tolerance mechanisms were built on the cognitive knowledge of specialists. While this adds to the system’s robustness, it creates dependence on the availability of expert personnel. The diagnostic system assumes a single failure, so scenarios with multiple simultaneous failures could exceed its current capacity. Ultrasonic inspection processed 5,661 anomalies, identifying 12% as requiring priority attention with an ERF > 0.8. Detecting incipient failures in topographically difficult-to-access areas remains an operational challenge.

Although the system’s modular architecture facilitates adaptation to other industrial contexts, direct transfer to pipelines with metallurgical characteristics, hydrocarbon compositions, or geographic conditions substantially different from the studied pipeline requires a specific validation phase. The corrosive behavior models were built from Mexican crude oil blends—naphthenic, paraffinic, and asphaltic—which limits their applicability to hydrocarbons with significantly different physicochemical profiles.

The IMS was designed based on a predictive fault-tolerant control system with offline data, providing the foundation for online adaptive control. Currently, it is a decision support tool that complements but does not replace the technical judgment of specialized personnel. Its gradual implementation through pilot sectors has trained new operating personnel in system reconfigurations during failures.

This article presents the first stage of the development and validation of the IMS; the second work will present the modeling, testing, and optimization of neuro-fuzzy control through the integration of genetic algorithms, as well as a cost-benefit analysis of the impact of failure scenarios through Monte Carlo simulation.

5. Discussion

The fault-signature matrix revealed a hierarchical pattern of failure propagation. Scenario FS-3 activated all 26 residuals. FS-4 and FS-9 also showed extensive interaction with other subsystems, confirming the transversal role of operational communication, pressure control, and monitoring functions. Metallurgical scenarios FS-22 – FS-28 activated 17 residuals, or 61% of the matrix. Structural scenarios FS-15 – FS-21 showed a decreasing from 16 to 11 activated residuals. In contrast, FS-1, FS-2, FS-5, FS-6, and FS-7 remained localized, activating fewer than 25% of the residuals and presenting more distinguishable signatures. These results indicate that diagnostic signature breadth depends not only on the initiating event but also on its capacity to propagate across operational, corrosion-protection, structural, and material domains.

The dominant signatures align with the physical mechanisms governing aged pipeline systems. Operational failures may reduce observability or cause pressure and flow deviations that promote hydraulic transients, cavitation, and cyclic stresses. At the same time, low soil resistivity, aggressive physicochemical conditions, coating degradation, and cathodic-protection deficiencies favor electrochemical dissolution and localized metal loss. This degradation appears in the progression of wall-thickness, corrosion-depth, defect-width, wall-loss, and estimated-repair-factor scenarios. At the material scale, changes in grain size, microstructure, mechanical properties, and defect population reduce the capacity of API L X65 steel to redistribute stresses, facilitating crack initiation and propagation. The matrix thus captures a coupled deterioration process where electrochemical damage alters structural resistance and operational loading accelerates defect evolution.

These findings complement international studies that have generally addressed corrosion prediction, structural integrity, or metallurgy as separate problems. BBN–GIS models have achieved high accuracy in predicting external corrosion depth. Probabilistic models validated against inline inspections have closely reproduced measured wall loss [15,18]. Likewise, metallurgical studies have shown the influence of microstructure, MnS inclusions, carbides, and corrosive environments on HIC, SSCC, and crack initiation in pipeline steels [16,17]. However, those approaches mainly provide condition or risk estimates without explicitly linking their outputs to system-wide fault propagation. The present results extend this knowledge by showing how operational, electrochemical, structural, and metallurgical anomalies interact within a unified diagnostic architecture. Nevertheless, direct performance superiority cannot yet be claimed because the present model and the cited approaches use different databases, outputs, and validation criteria.

A central contribution is the integration of historical SCADA, SAP, cathodic-protection, soil, ultrasonic-inspection, and metallurgical data with expert knowledge. Data-driven information anchors the rules in observed pipeline behavior. Specialist knowledge represents nonlinear interactions and rare but safety-critical events that may be underrepresented in historical databases. This hybrid construction improves interpretability because each diagnosis can be traced to activated variables, rules, and residuals, unlike purely black-box predictions. It also preserves operational knowledge accumulated by experienced personnel and converts heterogeneous information into comparable failure signatures. Consequently, the matrix provides more than a severity ranking. It establishes a structured basis for fault isolation, intervention prioritization, and subsequent fault-tolerant control.

From an operational perspective, widespread signatures such as FS-3, FS-4, FS-9, FS-8, and FS-10– FS-14 should receive priority because they may compromise both pipeline integrity and diagnostic capability. Localized signatures can support targeted maintenance. However, the results are limited to the studied 572 km network and depend on historical data quality, expert-defined membership functions, and the current single-fault assumption. Generalization requires recalibration and external validation in independent pipeline sections with different soils, coatings, operating regimes, and material histories. Future work should evaluate simultaneous and evolving failures, quantify diagnostic performance through sensitivity, specificity, F1-score, false-alarm rate, and detection delay, and prospectively validate the signatures using synchronized SCADA, CIPS, and inspection data. These steps will determine whether the proposed framework can reliably support online diagnosis and neuro-fuzzy reconfiguration under real operating conditions.

6. Conclusions

This study developed an integrated fuzzy-logic framework to diagnose failure propagation in an aged pipeline network using operational, cathodic-protection, mechanical-integrity, and metallurgical data. The four Mamdani systems represented 28 failure scenarios through a unified fault-signature matrix. Scenario FS-3 had the greatest systemic effect by activating all residuals, while FS-22 – FS-28 activated 61% of the matrix, confirming the critical influence of material degradation on pipeline integrity. Structural scenarios FS-15 – FS-21 showed progressively broader signatures as wall deterioration increased, whereas FS-1, FS-2, FS-5, FS-6, and FS-7 remained localized and were easier to isolate.

The main contribution is combining historical SCADA and SAP records, cathodic-protection and soil measurements, ultrasonic inspection, material characterization, and expert knowledge within an interpretable diagnostic structure. This integration classifies failures by diagnostic signature breadth and propagation capacity, providing a technical basis for early detection, maintenance prioritization, and operator decision support. Unlike isolated assessments of corrosion or structural damage, the proposed matrix captures the interdependence of operational, electrochemical, structural, and metallurgical deterioration mechanisms.

The conclusions are limited to the analyzed pipeline network, the quality of available historical data, expert-defined membership functions, and the assumption of mainly individual failure scenarios. This article presented the system analysis and design of diagnostic tools for fault-tolerant control. In Part Two, we will present the integration of the neuro-fuzzy model using an optimized predictive control law, analyzed through risk simulations in a case study within the Mexican oil industry.