A method for predicting failure propagation of high-temperature steam pipeline in thermal power plant

Abstract

The present disclosure provides a thermal power plant high-temperature steam pipeline failure propagation prediction method, which obtains a temperature field, a stress field and material degradation characteristics by constructing a three-dimensional digital twin model of the pipeline system and performing multi-physical field coupling analysis, simultaneously converts the pipeline system into graph structure data and utilizes a graph neural network model to fuse topological relationship and time sequence dynamic information, and realizes spatio-temporal joint prediction of failure propagation path, propagation speed and influence range; the method can significantly improve prediction accuracy, greatly extend the early warning time window, and has global pipeline network vulnerability analysis and hierarchical early warning capability, thereby providing intelligent and precise technical support for preventive maintenance and emergency decision-making of thermal power plant high-temperature steam pipelines.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, and in particular to a method for predicting the propagation of failures in high-temperature steam pipelines of thermal power plants. Background Technology

[0002] High-temperature steam pipelines in thermal power plants are core pressure-bearing equipment in the power system, typically operating at temperatures above 540℃ and pressures exceeding 17MPa. Operating under prolonged conditions of high temperature, high pressure, and complex loads, these pipeline systems experience irreversible damage to the materials, such as creep, fatigue, and thermal stress relaxation, leading to a gradual decrease in pipeline strength and eventual failure.

[0003] Currently, the following technologies are mainly used for monitoring high-temperature steam pipelines both domestically and internationally: Traditional non-destructive testing techniques, including ultrasonic testing, radiographic testing, magnetic particle testing, and penetrant testing, require system shutdown for maintenance and partial disassembly, making it impossible to achieve real-time assessment of the overall condition of the pipeline system; furthermore, they can only detect existing defects and cannot predict their development trends.

[0004] Monitoring systems based on fixed sensors deploy temperature, pressure, and strain sensors at key locations to monitor changes in pipeline operating parameters. However, traditional sensors have a limited number of measuring points (typically only 3-5 measuring points per 100 meters of pipeline), making it difficult to fully cover the spatial distribution characteristics of complex piping systems.

[0005] Existing pipeline health assessment methods mainly include life assessment based on empirical formulas, stress assessment based on finite element analysis, and reliability assessment based on statistical analysis. These methods typically use static models, which are difficult to reflect the dynamic degradation process of pipeline material properties over time, and also cannot take into account the interactions between various components within the pipeline system.

[0006] The main problems with existing technologies are: (1) Lack of global failure propagation analysis capability Existing technologies primarily focus on the condition assessment of individual pipe components, without establishing a correlation model between the overall pipe system topology and failure propagation. When a pipe segment fails, it is impossible to predict which adjacent pipe segments the failure will propagate to, the propagation speed, and the manner of propagation.

[0007] (2) Insufficient consideration of multiphysics coupling effects The failure of high-temperature steam pipelines is the result of the coupling effect of multiple physical fields, such as thermal field, stress field, and flow field. Existing evaluation methods usually consider each physical field separately, ignoring the interaction between the temperature field and the stress field (such as thermal stress generated by thermal expansion constraints, and non-uniform distribution of material properties caused by temperature gradients).

[0008] (3) Lack of intelligent and precise forecasting methods Existing methods rely on expert experience and human judgment, resulting in highly subjective predictions. They also lack the ability to identify and predict failure propagation patterns based on big data and artificial intelligence, making it impossible to accurately predict the failure propagation paths of complex piping systems.

[0009] (4) Insufficient early warning response time The early warning mechanism of existing monitoring systems is relatively passive, usually only triggering alarms when abnormal parameters exceed the threshold. It cannot predict the occurrence and development trend of failure in advance, and does not allow enough time for emergency response.

[0010] The technical problem to be solved by this invention is to provide a method for predicting the failure propagation of high-temperature steam pipelines in thermal power plants based on multi-physics coupling, so as to achieve accurate prediction of the failure propagation path, propagation speed and impact range of high-temperature steam pipeline systems, and provide early warning and auxiliary decision support for power plant operation and maintenance personnel. Summary of the Invention

[0011] To address the aforementioned technical problems, this disclosure provides a method for predicting the failure propagation of high-temperature steam pipelines in thermal power plants, including: S1: Construct a digital twin model of a high-temperature steam pipeline system, wherein the digital twin model includes pipeline geometric information, pipeline topology information, material property information, and operating condition information; S2: Based on the digital twin model, perform thermo-mechanical coupling analysis to calculate the temperature field distribution and stress field distribution of the pipeline system, and evaluate the creep damage and fatigue damage of the pipeline material; S3: Convert the pipeline system into graph structure data, define each pipe segment as a node in the graph, define the connection relationship between pipe segments as the edge of the graph, and assign each node a feature vector containing material degradation index, stress concentration factor and temperature distribution characteristics; S4 constructs a failure propagation prediction model based on graph neural networks. The failure propagation prediction model includes a node feature encoding layer, a graph convolutional propagation layer, a temporal feature extraction layer, and a failure probability prediction layer. S5 uses historical failure case data, simulation data, and real-time monitoring data to train and optimize the failure propagation prediction model; S6 inputs real-time sensor data into the trained failure propagation prediction model and outputs the failure probability, failure propagation path, and propagation time prediction results for each pipe segment.

[0012] In conjunction with the first aspect, the pipeline geometry information in step S1 includes pipe diameter, wall thickness, elbow angle, tee position, and valve position; The pipeline topology information includes node numbers, connection relationships, pipe segment lengths, and support point locations; The material property information includes the pipe material and its mechanical property parameters at different temperatures; The operating condition information includes design temperature, design pressure, and the actual operating temperature and pressure fluctuation range.

[0013] In conjunction with the first aspect, the thermo-coupling analysis based on the digital twin model described in step S2 includes: Calculate the temperature field distribution of the piping system under design conditions; The dynamic response of the temperature field during start-up and shutdown processes and load changes; Calculate the thermal stress distribution by combining the temperature field; Assess material creep damage and fatigue damage based on temperature and stress fields.

[0014] In conjunction with the first aspect, the feature vector in step S3 includes the temperature difference, pressure difference, and constraint stiffness between the connecting pipe segments.

[0015] In conjunction with the first aspect, the node feature encoding layer in step S4 uses a multi-head attention mechanism to encode the multidimensional features of the pipeline nodes and extract the local state representation of each pipe segment. The graph convolutional propagation layer uses a graph convolutional neural network or a multi-hop graph attention network for information propagation; The temporal feature extraction layer uses a temporal convolutional network or a Transformer encoder to extract the temporal evolution features of the pipeline state and capture the temporal dynamic characteristics of failure propagation. The failure probability prediction layer predicts the failure probability of each pipe segment, the propagation path of the failure in the piping system, the time it takes for the failure to reach each pipe segment, and the probability distribution of different propagation paths.

[0016] In conjunction with the first aspect, it also includes step S7: graded early warning, setting multiple levels of early warning based on the prediction results, and triggering the corresponding level of early warning when the failure probability exceeds the preset threshold.

[0017] In conjunction with the first aspect, the tiered early warning system includes: Blue alert: The failure probability exceeds the first threshold; enhanced monitoring is recommended. Yellow alert: The probability of failure exceeds the second threshold, or it is predicted that the virus may spread within 48 hours. It is recommended to prepare for maintenance. Red Alert: If the probability of failure exceeds the third threshold, or if it is predicted that the virus may spread within 24 hours, an emergency response should be initiated immediately.

[0018] In conjunction with the first aspect, when there are heterogeneous areas of different materials welded in the pipeline system, a heterogeneous edge type is set for the heterogeneous material welded joint in the diagram structure, and a higher failure propagation weight is assigned.

[0019] A third aspect of this disclosure provides an electronic device comprising: One or more processors; A storage unit for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the method.

[0020] A fourth aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program implements the method when executed by a processor.

[0021] Beneficial Effects: This disclosure provides a method for predicting the failure propagation of high-temperature steam pipelines in thermal power plants. By constructing a three-dimensional digital twin model of the pipeline system and performing multi-physics coupling analysis, the method obtains temperature field, stress field, and material degradation characteristics. Simultaneously, it converts the pipeline system into graph structure data and uses a graph neural network model to fuse topological relationships and temporal dynamic information, achieving spatiotemporal joint prediction of failure propagation paths, propagation speed, and impact range. This method can significantly improve prediction accuracy, greatly extend the early warning time window, and has the capability for global pipeline network vulnerability analysis and hierarchical early warning, thereby providing intelligent and precise technical support for preventive maintenance and emergency decision-making of high-temperature steam pipelines in thermal power plants. Attached Figure Description

[0022] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0023] Figure 1 A flowchart of a method for predicting the failure propagation of high-temperature steam pipelines in thermal power plants provided in this disclosure; Figure 2 The electronic device provided in this disclosure. Detailed Implementation

[0024] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0025] like Figure 1 As shown, the flowchart of a method for predicting the failure propagation of high-temperature steam pipelines in thermal power plants provided in this disclosure includes: S1: Construct a digital twin model of a high-temperature steam pipeline system, wherein the digital twin model includes pipeline geometric information, pipeline topology information, material property information, and operating condition information; The pipeline geometry information in step S1 includes pipe diameter, wall thickness, elbow angle, tee position, and valve position; The pipeline topology information includes node numbers, connection relationships, pipe segment lengths, and support point locations; The material property information includes the pipe material and its mechanical property parameters at different temperatures; The operating condition information includes design temperature, design pressure, and the actual operating temperature and pressure fluctuation range.

[0026] This embodiment takes the main steam pipeline system of a 600MW ultra-supercritical thermal power unit as an example to describe in detail the specific implementation of constructing a digital twin model of the high-temperature steam pipeline system in step S1 of this disclosure. The design parameters of the unit are: main steam temperature 605℃, main steam pressure 27MPa, P92 heat-resistant steel pipe material, total pipeline length of approximately 126m, and approximately 36 weld joints.

[0027] First, the design drawings and as-built data of the unit's main steam piping system were collected to obtain complete piping geometric parameters. Specifically, this included: Pipe diameter information: The nominal diameter of the main pipe is DN300 (outer diameter approximately 355.6mm), the nominal diameter of the branch pipe is DN200 (outer diameter approximately 219.1mm), and the wall thickness of each pipe section ranges from 25mm to 45mm, with the wall thickness of the elbow section being 35mm, the wall thickness of the straight pipe section being 28mm, and the wall thickness of the tee section being 45mm.

[0028] Elbow parameters: The piping system contains a total of 8 elbows, all with an angle of 90° and a bending radius of 1.5 times the pipe diameter (approximately 533 mm). The elbow type is hot-pressed elbow.

[0029] Three-way locations: There are three three-way structures in the system, located at the connection between the main steam pipeline and the first-level bypass, the connection between the main steam pipeline and the second-level bypass, and the distribution point from the main steam pipeline to the high-pressure cylinder steam inlet.

[0030] Valve locations: The system includes two main steam isolation valves (located on the furnace side and the turbine side respectively), one bypass regulating valve, and several drain valves and exhaust valves. All valves are equipped with clear location coordinates.

[0031] Flange specifications: All interfaces other than welded connections use butt-welded flanges with a pressure rating of Class 2500 and a tongue and groove sealing surface.

[0032] Based on the above geometric parameters, a full-size three-dimensional geometric model of the pipeline system was established using three-dimensional modeling software, with the model accuracy reaching the millimeter level.

[0033] Based on the three-dimensional geometric model, the topological information of the pipeline system is extracted: Node Numbering: The piping system is discretized, and each weld location, elbow start and end point, tee branch point, valve installation location, support and hanger installation point, and pipe diameter change point is defined as a topological node. In this embodiment, a total of 86 topological nodes are defined, numbered sequentially from boiler outlet to high-pressure cylinder inlet according to the steam flow direction, from N001 to N086.

[0034] Connection Relationships: Record the connection relationships between each node to form a pipeline system topology connection table. Connection relationships include series connections (connection of nodes at both ends of a straight pipe section), branch connections (connection of nodes in three directions at a tee), and end connections (connection of nodes on both sides of a valve). For example, N012—N013—N014 represent three adjacent nodes connected in series, and N025—N026 (main pipe direction) and N025—N027 (branch pipe direction) represent tee branching.

[0035] Pipe segment length: The pipe segment length between each node is obtained by precise measurement based on the 3D model. The shortest pipe segment is the short pipe between adjacent weld joints (about 0.8m), and the longest pipe segment is the straight pipe segment between two bends (about 6.5m). All pipe segment lengths are recorded in the topology connection table.

[0036] Support point locations: The piping system has a total of 17 sliding supports, 4 fixed supports, 6 guide supports, and 8 spring hangers. The location coordinates, support type, and constraint direction of each support point are recorded in the topology information. The fixed supports are located at three fixed points in the piping system, respectively located at the furnace-side outlet, the machine-side inlet, and the intermediate transition section.

[0037] Assign mechanical and thermophysical property parameters to the pipe material in the digital twin model: Pipeline materials: The main steam pipeline is made of P92 martensitic heat-resistant steel (9Cr-0.5Mo-1.8W-VNb), the elbows are also made of P92, the tees are forged of F92, and the welding rods are made of MTS-3. The grades and chemical compositions of the different materials are all recorded.

[0038] Mechanical property parameters at different temperatures: Collect and input the following performance parameters of P92 steel at multiple temperature points (20℃, 200℃, 300℃, 400℃, 500℃, 550℃, 600℃, 650℃): elastic modulus (approximately 212 GPa at 20℃, approximately 155 GPa at 600℃), Poisson's ratio (0.29~0.31), coefficient of thermal expansion (approximately 11.5×10⁻⁶ in the range of 20℃~600℃).-6 / ℃~12.5×10 -6 / ℃), yield strength (approximately 560MPa at 20℃ and approximately 320MPa at 600℃), tensile strength, creep curve (creep fracture time under different temperatures and stresses is given in the form of Larson-Miller parameters), and fatigue curve (given in the relationship between strain amplitude and number of cycles).

[0039] For F92 forging materials and MTS-3 weld materials, the complete mechanical property parameters at each of the above temperature points are also recorded, and the material property attenuation coefficient of the heat-affected zone of the weld is specially marked.

[0040] Determine the design operating parameters and actual operating conditions of the piping system: The design temperature is 605℃, the design pressure is 27MPa, the design service life is 30 years (based on 200,000 hours), and the design start-stop cycle is 500 times.

[0041] According to the power plant's operation statistics, the unit operates on an average of 7,500 hours per year, with about 12 start-ups and shutdowns per year. In actual operation, the main steam temperature fluctuates between 595℃ and 610℃, and the pressure fluctuates between 25MPa and 27.5MPa.

[0042] Determine the convective heat transfer coefficient between the outer wall of the pipe and the ambient air (take 10~25W / (m²·K)), the convective heat transfer coefficient between the inner wall and the high-temperature steam (calculated based on the steam flow rate and physical property parameters), the thickness of the pipe insulation layer (the main insulation layer is 200mm thick aluminum silicate fiber felt, and the outer protective layer is 0.7mm thick galvanized iron sheet) and its thermal conductivity.

[0043] Collect temperature-time and pressure-time curves for typical start-up and shutdown processes, including complete parameter change data for cold start (approximately 8 hours), warm start (approximately 4 hours), hot start (approximately 2 hours), and normal shutdown (approximately 2 hours).

[0044] The information from the four levels mentioned above is integrated into a unified three-dimensional digital twin model platform, establishing the correlation between the various parameters. By comparing historical operating data (including infrared temperature measurement data, wall temperature measurement point data, and stress monitoring data) with the initial values ​​calculated by the model, the boundary conditions and initial parameters of the model are verified and corrected to ensure that the digital twin model accurately reflects the geometric characteristics, topology, material properties, and operating conditions of the actual pipeline.

[0045] The final digital twin model is stored in a standardized format and can be exported for use in the subsequent multiphysics coupling analysis in step S2. In this embodiment, the total file size of the digital twin model is approximately 850MB, containing a complete geometric mesh (approximately 1.65 million meshes) and an associated attribute database.

[0046] S2: Based on the digital twin model, perform thermo-mechanical coupling analysis to calculate the temperature field distribution and stress field distribution of the pipeline system, and evaluate the creep damage and fatigue damage of the pipeline material; The thermal coupling analysis based on the digital twin model in step S2 includes: Calculate the temperature field distribution of the piping system under design conditions; The dynamic response of the temperature field during start-up and shutdown processes and load changes; Calculate the thermal stress distribution by combining the temperature field; Assess material creep damage and fatigue damage based on temperature and stress fields.

[0047] Next, the digital twin model is imported into the finite element analysis software for steady-state thermal analysis calculations. The specific implementation process is as follows: Meshing: Finite element meshing was performed based on the digital twin model. Hexahedral elements were used in the pipe wall thickness direction, divided into 4 layers to accurately capture the temperature gradient in the wall thickness direction. Elbows, tees, and the heat-affected zone of welds underwent localized mesh refinement, with a minimum element size of 2 mm. Straight pipe sections used a coarser mesh with an element size of 15 mm. After meshing, the finite element model contained approximately 780,000 elements and 920,000 nodes.

[0048] Boundary condition settings: Thermal analysis boundary conditions are set according to the operating parameters determined in Example 1. A steam temperature boundary (605℃) and a convective heat transfer boundary are applied to the inner wall of the pipe (the convective heat transfer coefficient is calculated based on a steam flow velocity of 3.5m / s, and is approximately 12,500W / (m²·K)); an insulation layer boundary condition is applied to the outer wall of the pipe, and the insulation layer is equivalent to a third type of thermal boundary (the equivalent heat transfer coefficient is calculated based on the thermal conductivity of the insulation layer of 0.045W / (m·K) and a thickness of 200mm, and is approximately 0.225W / (m²·K), with an ambient temperature of 25℃).

[0049] Solution calculation: The steady-state heat conduction equation is used to calculate the temperature field distribution of the pipeline system when it reaches thermal equilibrium. The nonlinear change of the material's thermal conductivity with temperature is considered in the solution process.

[0050] Results Extraction and Analysis: Calculation results show that the inner wall temperature of the main pipe is 603.5℃~605.0℃, and the outer wall temperature is 590.2℃~592.8℃, with an average temperature gradient of approximately 12℃~15℃ along the wall thickness. Specifically, at elbows, due to the centrifugal effect of steam flow, the outer wall temperature is slightly higher than the inner wall temperature (temperature difference approximately 3℃). At tees, due to the steam mixing effect, the temperature at the branch pipe is approximately 5℃ lower than the main pipe. The temperature distribution in the heat-affected zone of the weld exhibits localized nonlinear variations due to changes in the material's thermal conductivity. The temperature field data in the analysis results are derived as nodal temperature values, serving as the load input for subsequent stress field analysis.

[0051] Based on steady-state thermal analysis, transient thermal analysis is performed to simulate the dynamic response of the temperature field in the pipeline during unit start-up and shutdown and load changes: Transient loading: Based on the typical start-up and shutdown temperature-time curves collected in Example 1, the load step for transient analysis was set. Taking the cold start condition as an example, starting from an ambient temperature of 25°C, the temperature was gradually increased to 605°C over 8 hours. The heating process was divided into a preheating stage (0-2 hours, temperature rise to 200°C), a ramp-up stage (2-5 hours, temperature rise to 540°C), and a grid connection stage (5-8 hours, temperature rise to 605°C) according to the actual operating curve. The shutdown condition was carried out according to the normal shutdown curve, cooling from 605°C to 150°C over 2 hours.

[0052] Transient solution settings: The initial condition is set as the initial temperature of the piping system equal to the ambient temperature of 25℃; the time step is adaptively adjusted according to the rate of temperature change, with a minimum time step of 5s and a maximum time step of 60s; an implicit time integration algorithm is used for the solution. The nonlinear changes in material density, specific heat capacity, and thermal conductivity with temperature are considered.

[0053] Transient response results: Calculations show that during cold start-up, the inner wall temperature rise rate is approximately 1.5℃ / min (preheating stage) to 2.5℃ / min (acceleration stage), while the outer wall temperature rise rate lags behind the inner wall by about 30–45 minutes. The maximum transient temperature difference along the wall thickness during start-up is 52℃ (occurring in the middle of the acceleration stage, approximately 3.5 hours after start-up). During normal shutdown, the inner wall cooling rate is approximately 3.5℃ / min, and the maximum transient temperature difference along the wall thickness is 38℃ (occurring approximately 45 minutes after shutdown). During load changes (from 100% to 75% load), the steam temperature drops by approximately 25℃, and the transient temperature difference along the wall thickness is approximately 15℃.

[0054] Temperature time history recording of key parts: Temperature monitoring points are set up at stress-sensitive locations such as the inside of elbows, the root of tees, and the heat-affected zone of welds. The temperature-time curves of each monitoring point are recorded during the complete transient process. These time-series temperature data are stored for subsequent material damage assessment.

[0055] Based on the steady-state and transient thermal analysis results, the temperature field data is applied as a thermal load to the pipe structure model for stress field analysis: Structural analysis model setup: The mesh model used in the thermal analysis is converted into a structural analysis model, and the element type is converted to structural element. Material mechanical property parameters are nonlinearly interpolated based on the elastic modulus, Poisson's ratio, and coefficient of thermal expansion entered in Example 1 at different temperatures. Temperature-related material properties are automatically updated at each integration point according to the current temperature.

[0056] Boundary conditions are applied as follows: Displacement constraints are applied based on the support point information recorded in Example 1—fixed supports constrain the translational degrees of freedom in the X, Y, and Z directions; guide supports constrain the translational degrees of freedom in the radial directions while allowing free axial movement; sliding supports only constrain displacement perpendicular to the pipe axis, allowing free axial displacement; spring supports are given elastic support boundaries according to the actual spring stiffness. Pressure boundaries (design pressure 27 MPa) and axial force boundaries are applied to the cross-sections at both ends of the pipe.

[0057] Steady-state thermal stress calculation: The temperature field obtained from steady-state thermal analysis is applied as a volume load to the pipe structure to calculate the steady-state thermal stress distribution under design conditions. A thermo-mechanical sequential coupling method is used for the solution. First, the thermal strain (ε_th = α(T) × ΔT) is calculated from the temperature field. Then, the thermal strain is substituted as the initial strain into the structural equilibrium equations for solving. Mechanical stress generated by internal pressure is also considered in the calculation.

[0058] Stress Result Analysis: Steady-state thermal stress calculations show that the maximum stress in the pipeline system occurs at the following key locations: ① Inside the elbow (stress concentration factor approximately 2.1), with a peak stress of 342 MPa, about 1.07 times the yield strength of P92 steel at 600℃ (approximately 320 MPa), but below the high-temperature creep limit; ② At the root of the tee branch (stress concentration factor approximately 2.5), with a peak stress of 388 MPa; ③ In the heat-affected zone of the weld (stress concentration factor approximately 2.8), the peak stress reaches 412 MPa, with some areas exceeding the material's yield strength. The stress level in the straight pipe section is relatively low, with principal stresses ranging from 65 to 120 MPa. The stress distribution in the pipeline system exhibits significant non-uniformity, and the stress concentration areas are also regions of local temperature abrupt changes, reflecting a thermo-mechanical coupling effect.

[0059] Transient thermal stress calculation: The temperature field of each time step of the transient thermal analysis was sequentially applied to the structural model to calculate the dynamic changes of transient thermal stress during start-up and shutdown. The calculation results show that in the middle of the cold start-up acceleration phase, due to the maximum transient temperature difference (52℃) in the wall thickness direction, a peak thermal stress of 590MPa appeared on the inner side of the elbow (1.05 times the material's room temperature yield strength of 560MPa). Although the duration was short (approximately 40 minutes), it constituted a significant source of fatigue damage. During shutdown, the maximum transient stress in the weld heat-affected zone was 420MPa. In a typical start-up and shutdown cycle, the stress amplitude experienced by critical components was approximately 180–350MPa.

[0060] Based on the analysis results of the temperature and stress fields, multiple damage mechanics models are used to comprehensively assess the degree of degradation of the pipeline material: Creep Damage Assessment: The Larson-Miller parametric method is used to assess the creep damage of the material. This method is based on the following relationship: P_LM=T×(logt_r+C), where T is the absolute temperature, t_r is the creep rupture time, and C is a material constant (C=36 for P92 steel). Based on the temperature and stress field distributions, the equivalent stress and equivalent temperature of each pipe section are calculated and substituted into the Larson-Miller curve to determine the creep rupture life under that condition. Then, the cumulative creep damage factor is calculated using the time fraction method (Robinson criterion): D_c=Σ(Δt_i / t_{ri}), where Δt_i is the running time of the i-th time period, and t_{ri} is the creep rupture life under the corresponding condition.

[0061] Calculation results show that after 100,000 hours of operation, the cumulative creep damage factor is 0.15 in the elbow area, 0.21 at the tee root, 0.28 in the weld heat-affected zone, and 0.03–0.05 in the straight pipe section. Based on the critical value of creep damage for P92 steel (creep fracture occurs when D_c=1), the weld heat-affected zone is the highest priority area for creep failure. The creep damage factor of each pipe section is stored as an important feature in the damage database.

[0062] The Coffin-Manson relation was used to assess low-cycle fatigue damage caused by temperature fluctuations. Δε_p / 2 = ε_f' × (2N_f)^c, where Δε_p is the plastic strain amplitude, ε_f' is the fatigue ductility coefficient, N_f is the fatigue life, and c is the fatigue ductility exponent. Based on the stress-strain time history data of each key part extracted from the transient thermal stress analysis, the number of stress cycles and amplitude are counted using the rainflow counting method. Combined with the fatigue curve of P92 steel, the fatigue damage factor for each cycle is calculated: D_f = Σ(n_i / N_{fi}), where n_i is the number of cycles of the i-th stress amplitude, and N_{fi} is the fatigue life under the corresponding stress amplitude.

[0063] Calculation results show that the fatigue damage factor caused by each start-stop cycle (cold start-run-shutdown) is approximately 0.0012 to 0.0018 (varying slightly depending on the location), with 0.0015 on the inside of elbows, 0.0018 in the heat-affected zone of welds, and 0.0016 at the root of tees. Based on a design life of 500 start-stop cycles, the cumulative fatigue damage factor is 0.60 to 0.90, indicating that fatigue is a significant contributing factor to pipeline failure.

[0064] Considering that pipelines simultaneously experience creep and fatigue damage during actual operation, a linear cumulative damage criterion is used to assess the overall damage: D_total = D_c + D_f. The creep-fatigue interaction is checked according to the design specifications for P92 steel (EN 13445 or ASME BPVC Section III Division 5). For the critical parts in this embodiment, after 100,000 hours of operation combined with 500 start-stop cycles, the overall damage factor D_total of the weld heat-affected zone is 0.28 + 0.90 = 1.18, exceeding the theoretical critical value of 1.0, indicating a high risk of failure in this area. This aligns with engineering experience that the weld heat-affected zone is a common failure site in actual pipeline operation. The overall damage factor on the inner side of the elbow is 0.15 + 0.75 = 0.90, which is critical. The damage factor at the root of the tee is 0.21 + 0.80 = 1.01, also close to or exceeding the critical value.

[0065] The calculation results of the creep damage factor, fatigue damage factor, and comprehensive damage factor are mapped to various pipe segments and locations in the pipeline system to form a damage distribution database. Each pipe segment node records the following damage characteristic parameters: current cumulative creep damage factor D_c, current cumulative fatigue damage factor D_f, comprehensive damage factor D_total, remaining creep fracture life (hours), remaining fatigue life (cycles), annual damage increment rate, etc. These damage characteristic parameters will serve as the core components of the feature vector of the graph structure nodes in step S3.

[0066] The damage distribution obtained from the analysis was compared with the non-destructive testing results of the actual pipeline system (defect distribution in the ultrasonic testing report, material degradation areas in the hardness testing report, etc.) to verify the accuracy of the multiphysics coupling analysis results. In this embodiment, the agreement rate between the predicted high-damage areas and the areas with defects found in the actual ultrasonic testing reached 82%, verifying the effectiveness of the analysis model. For areas where the analysis results deviated significantly from the measured results, the model was corrected by adjusting the boundary conditions or material parameters.

[0067] All results from the multiphysics coupling analysis described above—including steady-state temperature field distribution, transient temperature field time history data, stress field distribution, and various damage factors—are output in a structured data format, providing a data foundation for the subsequent graph structure construction and feature assignment in step S3. The entire analysis process is completed in batches using automated scripts, ensuring computational efficiency and data consistency when dealing with complex piping systems.

[0068] S3: Convert the pipeline system into graph structure data, define each pipe segment as a node in the graph, define the connection relationship between pipe segments as the edge of the graph, and assign each node a feature vector containing material degradation index, stress concentration factor and temperature distribution characteristics; The feature vector in step S3 includes the temperature difference, pressure difference, and constraint stiffness between the connecting pipe segments.

[0069] Specifically, the division of pipe segment units follows these principles: based on the natural segmentation of the pipeline system, each pipe segment unit is defined as the area between two adjacent weld joints; for long straight pipe segments, they are further subdivided into units of 3 to 5 meters in length to ensure the accuracy of the diagram; each elbow, tee, reducer, and valve is divided into a separate pipe segment unit; adjacent pipe segment units are separated by weld joints or connection surfaces. Following these principles, the main steam pipeline system in this embodiment is approximately 126 meters long and is divided into 126 pipe segment units. Each pipe segment unit corresponds to a node in the diagram, and the node numbering follows the topological node numbering system established in S1, sequentially denoted as v1 to v... 126 Each node is labeled with its type attribute, including 84 straight pipe nodes, 8 elbow nodes, 9 tee nodes (divided into 3 nodes based on their three ports), 4 reducer nodes, 2 valve nodes, 16 weld nodes, and 3 other connection nodes. Different types of nodes will receive different feature weights in subsequent feature encoding. Simultaneously, the spatial coordinates and orientation vector of each node are recorded to characterize the spatial configuration of the pipeline system. The spatial coordinates are taken as the three-dimensional coordinates of the center point of the pipe segment unit, and the orientation vector is taken as the direction vector of the central axis of the pipe segment unit. This spatial information is used to assist in calculating the spatial distance and relative orientation between nodes.

[0070] Next, the edges are defined and the connection relationships are constructed. When two pipe segment units are physically directly connected (through welding or flange connection), an undirected edge is defined between their corresponding nodes. In this embodiment, a total of 132 edges are formed between 126 nodes. For a tee node, three edges are established between the node and its three adjacent nodes in each direction; for a valve node, the pipe segments on both sides of the valve are connected to the valve node through flanges to establish edges. Each edge is labeled with its connection type, including 108 welded connection edges, 16 flanged connection edges, and 8 elbow-to-straight pipe segment connections treated as welds. Although edges in the graph structure are usually considered undirected mathematically, considering the influence of the directionality of steam flow on failure propagation, this embodiment adds a direction weight parameter to each edge. The direction weight is determined according to the steam flow direction: if steam flows from node i to node j, the propagation weight of edge i→j is higher than that of edge j→i; for tee nodes with equal branching, the direction weight is allocated according to the proportion of steam flow in each branch pipe.

[0071] Next, node feature vectors are constructed. An initial feature vector is built for each node, which comprehensively reflects the multi-dimensional state information of the pipe segment unit. In this embodiment, the feature vector of each node has 64 dimensions. In terms of geometric feature dimensions, it includes the pipe segment outer diameter, wall thickness, pipe segment length, elbow angle (0 if it is not an elbow), bending radius, and node type code (one-hot code for straight pipe, elbow, tee / other), totaling 8 dimensions. In terms of temperature field feature dimensions, it includes the average temperature, inner wall temperature, outer wall temperature, maximum temperature difference in the wall thickness direction, maximum temperature difference in the circumferential direction, and deviation between the highest temperature and the design temperature of the pipe segment extracted from the steady-state thermal analysis in step S2, as well as the maximum temperature rise rate, maximum temperature fall rate, temperature fluctuation amplitude, temperature fluctuation frequency, peak temperature and cumulative temperature change value in the most recent start-stop cycle extracted from the transient thermal analysis, totaling 12 dimensions. In terms of stress field characteristics, it includes the peak equivalent stress, average equivalent stress, maximum principal stress, stress concentration factor, proportion of thermal stress to total stress, and proportion of internal pressure stress to total stress extracted from the stress field analysis in step S2, as well as the stress amplitude, peak thermal stress, stress triaxiality, maximum shear stress, maximum cyclic stress amplitude extracted by the rainflow counting method, and the ratio of stress to yield strength extracted from the transient stress analysis, for a total of 12 dimensions. In terms of material degradation characteristics, the data includes the cumulative creep damage factor D_c, cumulative fatigue damage factor D_f, comprehensive damage factor D_total, remaining creep fracture life (in hours, logarithmically normalized), remaining fatigue life (in cycles, logarithmically normalized), annual damage increment rate, and damage growth rate (damage increment in the most recent 1000 hours) extracted from the material degradation assessment in step S2. In addition, the data also includes the material type code (one-hot code of P92, F92, and MTS-3, totaling 3 dimensions), the operating time of the pipe section (in hours, normalized), the number of start-stop cycles, the time interval since the last non-destructive testing, the defect level found in the last testing (0 if none), the initial hardness value of the material, and the current estimated hardness value, totaling 16 dimensions. In terms of operational condition characteristics, the dimensions include the current pressure of the pipe section, pressure fluctuation amplitude, pressure fluctuation frequency, steam flow rate, mass flow rate, pressure difference between the inlet and outlet of the pipe section, insulation layer integrity (estimated based on infrared thermography inversion), and support constraint status (estimated based on vibration monitoring inversion), totaling 8 dimensions. In terms of historical failure-related characteristics, the dimensions include the number of historical defect records for the pipe section, the number of historical maintenance operations, 0 / 1 codes indicating whether a leak has occurred, the time interval since the most recent historical failure or defect repair, the ranking of the pipe section in previous risk assessments, the historical average failure probability of similar pipe sections, the risk level (1 to 5 codes) marked by expert experience, and reserved expansion dimensions, totaling 8 dimensions.After the 64-dimensional node feature vectors are constructed, they are Z-score standardized to make the mean of each dimension feature 0 and the standard deviation 1, so as to avoid the influence of features of different dimensions on the subsequent training of graph neural networks. The standardization parameters are obtained and saved by statistics from the training dataset for online standardization processing of new data.

[0072] Next, feature vectors are constructed for each edge. These feature vectors reflect the interaction between the connecting pipe segments. In this embodiment, each edge's feature vector has 32 dimensions. Regarding the temperature difference feature dimension, it includes the absolute value of the average temperature difference between the two connecting pipe segments, the inner wall temperature difference, the outer wall temperature difference, and the temperature difference difference in the wall thickness direction (i.e., the absolute value of the difference in wall thickness temperature between the two pipe segments), extracted from the steady-state thermal analysis in step S2. This comprises four dimensions. Temperature difference is a key driving factor for failure propagation; the larger the temperature difference, the stronger the tendency for heat conduction-driven failure propagation. Regarding the pressure difference feature dimension, it includes the steady-state pressure difference and the transient pressure difference (the maximum dynamic pressure difference during start-up and shutdown), comprising two dimensions. Pressure difference drives steam flow and affects the direction of stress transfer and failure propagation. In terms of constraint stiffness characteristics, there are four dimensions: constraint stiffness value between the two pipe segments and constraint type code (one-hot code for fixed, guided, sliding, and spring). The constraint stiffness is determined by the arrangement of supports and hangers at the connection of the two pipe segments. The higher the constraint stiffness, the higher the force transmission efficiency during failure propagation, and the higher the propagation probability. In terms of connection structure characteristics, there are six dimensions: connection type code (one-hot code for weld, flange, and integrated joint, totaling three dimensions), welded joint type code (one-hot code for butt joint, corner joint, and plug joint, totaling three dimensions; non-welded connections are set to 0), 0 / 1 code for whether there is a reinforcing plate at the connection, 0 / 1 code for whether heat treatment has been performed, design fatigue enhancement coefficient, and non-destructive testing history result level (levels 1 to 5). In terms of geometric and topological characteristics, the dimensions include the spatial distance between the center points of the two pipe segments, the angle between the axes of the two pipe segments, the radius of curvature at the connection (the larger value is taken for straight pipe connections), the 0 / 1 code for whether there is a valve isolation between the two pipe segments, the 0 / 1 code for whether there is an expansion joint, and the 0 / 1 code for whether they belong to the same support system, totaling 6 dimensions. In terms of material matching characteristics, the dimensions include the 0 / 1 code for whether the two pipe segments are made of the same material, the 0 / 1 code for dissimilar steel welding markings, the difference in hardness between the two pipe segments, and the difference in thermal expansion coefficients between the two pipe segments, totaling 4 dimensions. Among these, the failure propagation tendency of dissimilar steel welded joints (such as the welding of P92 and F92) is significantly higher than that of same material connections, therefore, dissimilar steel weld edges are given a higher propagation weight. In terms of operational correlation characteristics, the dimensions include the correlation coefficient of operating temperature between the two pipe segments (calculated based on historical temperature monitoring data), the correlation coefficient of pressure fluctuation, the steam flow ratio (for tee branch connections), and the 0 / 1 code for whether they share the same support system, totaling 4 dimensions. In terms of the historical characteristics of failure propagation, there are two dimensions: 0 / 1 codes indicating whether failure propagation has occurred between the two pipe sections in the past and the known number of failure propagations (with an upper limit of 10). These historical data provide direct prior experience for failure propagation prediction.

[0073] Finally, the graph structure data is stored and managed. The graph structure data is stored in adjacency list format and contains three core files: a node feature matrix file (126 rows x 64 columns, CSV format), an edge feature matrix file (132 rows x 32 columns, CSV format), and an adjacency list file (listing the node numbers of the two nodes connected by each edge, CSV format). It also includes auxiliary information such as a mapping table between node IDs and pipe segment names, and descriptions of edge IDs and connection locations. Because the state of the pipeline system changes dynamically over time, material damage accumulates continuously, and operating conditions are constantly updated, the graph structure data is version-managed by timestamps and updated monthly. Each update appends the newly added operating data and damage accumulation data for that month to the new version, while retaining historical versions to support time-series analysis. After the graph structure data is constructed, data quality verification is performed, including node connectivity verification (ensuring the graph structure is fully connected and there are no isolated nodes), feature integrity verification (ensuring all 64-dimensional features of all nodes have no missing values), and feature rationality verification (ensuring that the values ​​of each feature are within a reasonable range, such as damage factors between 0 and 1). For data anomalies found during verification, corrections are made by tracing back the original data or marking them as missing values ​​to be filled later. To facilitate understanding and verification of the graph structure data by engineering technicians, the graph structure is visualized. The size of the nodes in the graph reflects their overall damage level (the greater the damage, the larger the node), the color of the nodes reflects their temperature (red represents high temperature, blue represents low temperature), and the thickness of the edges reflects their connection strength or failure propagation probability weight. The visualization results are overlaid with the pipeline system layout diagram to facilitate a one-to-one correspondence between the graph structure and the actual pipeline layout. Specifically, in this embodiment, the key node data, such as node v corresponding to elbow 3, is shown. 12 Its comprehensive damage factor is 0.90, peak equivalent stress is 342 MPa, and average temperature is 603.2 degrees Celsius; the node v corresponding to the main tee pipe side. 38 Its comprehensive damage factor is 1.01, peak equivalent stress is 388 MPa, and average temperature is 604.8 degrees Celsius; the node v corresponding to the heat-affected zone of the weld. 41 Its comprehensive damage factor is 1.18, peak equivalent stress is 412 MPa, and average temperature is 602.7 degrees Celsius. Regarding side data, for example, the T-junction main pipe side node v... 38 The edge between the node and its adjacent node has a temperature difference of 4.2 degrees Celsius, a pressure difference of 0.15 MPa, a constraint stiffness of 1.20, and a welded connection with dissimilar steel components; another example is the bypass branch pipe section node v. 63 The edges between adjacent nodes have a temperature difference of 6.5 degrees Celsius, a pressure difference of 0.45 MPa, a constraint stiffness of 0.60, and are flange connections with no dissimilar steel welding. The above graph structure data is output in a standardized format for training and prediction of the graph neural network model in step S4. S4 constructs a failure propagation prediction model based on graph neural networks. The failure propagation prediction model includes a node feature encoding layer, a graph convolutional propagation layer, a temporal feature extraction layer, and a failure probability prediction layer. The node feature encoding layer described in step S4 uses a multi-head attention mechanism to encode the multidimensional features of the pipeline nodes and extract the local state representation of each pipe segment. The graph convolutional propagation layer uses a graph convolutional neural network or a multi-hop graph attention network for information propagation; The temporal feature extraction layer uses a temporal convolutional network or a Transformer encoder to extract the temporal evolution features of the pipeline state and capture the temporal dynamic characteristics of failure propagation. The failure probability prediction layer predicts the failure probability of each pipe segment, the propagation path of the failure in the piping system, the time it takes for the failure to reach each pipe segment, and the probability distribution of different propagation paths.

[0074] The node feature encoding layer employs a multi-head attention mechanism to encode the 64-dimensional multi-dimensional features of pipeline nodes, extracting the local state representation of each pipe segment. The core design idea of ​​this layer is that the features of different pipe segments in the pipeline system have different importance for failure propagation prediction. By automatically learning the correlation weights between each feature dimension through a self-attention mechanism, the original features are mapped into a more representative hidden feature vector.

[0075] In the specific implementation, the node feature matrix X∈R^(126×64) constructed in step S3 is first input into the encoding layer, where 126 is the number of nodes and 64 is the node feature dimension. To enhance the model's ability to capture the interaction relationships between features, this embodiment uses an 8-head self-attention mechanism for feature encoding. For each attention head h (h=1,2,...,8), the query vector, key vector, and value vector are calculated respectively. Then, the weighted results calculated by each attention head are concatenated to obtain the multi-head attention output with a dimension of 512×128. The outputs of the 8 heads (each head output dimension 64) are concatenated (total dimension 512) and mapped back to 128 dimensions. The output is then processed through residual connections and layer normalization, and then undergoes a nonlinear transformation through a feedforward neural network layer (FFN). The FFN contains two linear layers with a ReLU activation function in between, and the hidden layer dimension is 256. The final output is the encoded node feature vector Z∈R^(126×128), which increases the feature dimension from the original 64 dimensions to 128 dimensions, thus enhancing the feature representation capability.

[0076] Under the multi-head attention mechanism, the model can automatically learn the correlation importance between features of different types of pipe segments. For example, in actual training, the model learns that for elbow nodes, the attention weight between the stress concentration factor and the creep damage factor is relatively high (approximately 0.32), reflecting the driving effect of stress concentration on elbow creep damage; for weld nodes, the attention weight between the material type encoding and the fatigue damage factor is relatively high (approximately 0.28), reflecting the influence of different welding materials on fatigue performance; and for straight pipe nodes, the attention weight between temperature distribution features and the thermal stress ratio is relatively high (approximately 0.25). These automatically learned attention weight distributions are highly consistent with engineering physical mechanisms, verifying the effectiveness of the encoding layer.

[0077] The graph convolutional propagation layer uses a graph attention network (GAT) for information propagation to simulate the diffusion process of failures in a pipeline network. The function of this layer is to enable each node to absorb information from its neighbors through message passing between adjacent nodes within the graph topology of the pipeline system, thereby simulating the physical process of failures gradually propagating from the failure source segment to surrounding segments.

[0078] In specific implementation, this embodiment uses two layers of graph attention convolution (GATConv) stacked to form a propagation layer. The propagation stride of each layer is 1, that is, the node can receive information from its direct neighbors. After the two layers are stacked, the receptive field of each node is expanded to 2 hop neighbors, that is, it can receive information from all nodes within two pipe segments of the node.

[0079] The attention coefficient calculation considers both node features and edge features. For the edge connecting node i and node j, its edge feature vector is introduced into the attention mechanism. The introduction of edge features enables the propagation layer to perceive information such as temperature difference, pressure difference, constraint stiffness, and connection type between the two pipe segments, thereby more accurately modeling the propagation tendency of failure under specific connection relationships.

[0080] In failure propagation computation, the failure propagation process considers the following four key factors: First, spatial adjacency, meaning that failure propagation is more likely to occur between adjacent pipe segments. The GAT network structure naturally ensures that information is transmitted only between nodes connected by edges, and the self-attention weights further reinforce the importance of spatial proximity. Second, temperature gradient, where high-temperature pipe segments tend to propagate more strongly to low-temperature pipe segments. This physical law is reflected in the weight of temperature difference features in the attention coefficient—when the temperature difference in the edge features is large, the attention coefficient increases accordingly. Third, stress transmission, where pipe segments bearing greater stress are more likely to become propagation mediators. The peak equivalent stress and stress concentration coefficient in the node features affect the information transmission weight of the node when it acts as a propagation mediator. Fourth, material continuity, where the probability of propagation is higher between pipe segments of the same material. When the dissimilar steel welding identifier in the edge features is 1, the attention coefficient is automatically reduced, reflecting that the fracture toughness of dissimilar steel welded joints is reduced in reality, but their hindering effect on failure propagation is actually increased—it should be noted that the model learns the propagation probability rather than the fracture probability here, and the model will automatically adjust the weight direction based on the training data.

[0081] In this embodiment, the graph convolutional propagation layer has a hidden layer dimension of 256. The two GATConv layers propagate the node features sequentially from 128 dimensions to 256 dimensions, respectively, and output the updated node embedding matrix H∈R^(126×256). The node embeddings output by the propagation layer not only encode the state features of the node itself, but also aggregate the feature information of its neighboring nodes, thus enabling the modeling of failure propagation in pipeline networks.

[0082] The temporal feature extraction layer employs a Transformer encoder to extract the temporal evolution features of the pipeline state, capturing the temporal dynamics of failure propagation. Since pipeline failure is a slow, evolving process over time, relying solely on graph structure data from a single time step cannot accurately predict future failure trends. The temporal feature extraction layer takes graph state sequences from multiple historical time steps as input, extracts evolutionary patterns along the temporal dimension, and thus predicts future failure states.

[0083] In this specific implementation, a historical graph state sequence with a time window length of T=12 is constructed. That is, based on the current moment, the graph structure data is traced back 12 time steps, with each time step spaced 30 days apart, covering a total of one year of historical data. For each time step t (t=1,2,...,12), the node feature matrix X_t∈R^(126×64) constructed in step S3 is processed through a node encoding layer with shared parameters and a graph convolutional propagation layer to obtain the node embedding matrix H_t∈R^(126×256) for that time step. Then, the node embeddings of the 12 time steps are stacked in chronological order to form a temporal embedding tensor T_seq∈R^(12×126×256).

[0084] The temporal feature extraction layer consists of four stacked Transformer encoder layers, each containing a multi-head self-attention sublayer and a feedforward neural network sublayer. In the self-attention sublayer, a four-head attention mechanism is employed, with attention calculation performed along the temporal dimension: for each node i, its embedding vectors at 12 historical time steps are fed into the Transformer encoder, and the self-attention mechanism calculates the temporal dependencies between different time steps. The self-attention weights reflect the contribution of historical states to current and future predictions. For example, the model learns that state changes in the three most recent time steps (within the last 90 days) have the greatest impact on short-term predictions (the next 72 hours), with an attention weight of approximately 0.65; while earlier time steps (more than 90 days ago) mainly influence medium- to long-term trend predictions, with a weight of approximately 0.35.

[0085] The feedforward neural network sublayer of the Transformer encoder consists of two linear layers with a hidden dimension of 512, using the GELU activation function. Each sublayer is followed by residual connections and layer normalization to prevent gradient vanishing and accelerate training convergence. After processing by four Transformer encoder layers, the output is a temporally enhanced node embedding matrix T_out∈R^(12×126×256). Subsequently, a time-weighted pooling operation aggregates the information from 12 time steps into a comprehensive embedding P∈R^(126×256) for the current time step. The pooling weights are automatically learned by the model, with larger weights for time steps closer to the current time step.

[0086] The failure probability prediction layer, based on time-enhanced node embedding, predicts the failure probability of each pipe segment, the failure propagation path in the piping system, the time it takes for the failure to reach each pipe segment, and the probability distribution of different propagation paths. This layer is the output layer of the entire prediction model, mapping the node representations extracted from the preceding layers into specific prediction results.

[0087] The failure probability prediction layer consists of three parallel prediction heads, sharing the node embedding P∈R^(126×256) output from the preceding layer. The first prediction head is used to predict the failure probability of each pipe segment within the next 72 hours. This prediction head contains two fully connected layers. The first layer maps the 256-dimensional node embedding to 128 dimensions, using the ReLU activation function and Dropout regularization (dropout rate 0.2). The second layer maps the 128 dimensions to 1 dimension, outputting a failure probability value between 0 and 1 through the Sigmoid activation function. The final output p_fail∈R^126 represents the failure probability of each of the 126 pipe segments.

[0088] When predicting the propagation path, this layer models the failure propagation path prediction as a sequence generation problem, employing an autoregressive decoder architecture. It uses the predicted high-probability failure pipe segments as starting nodes and predicts the propagation direction of the failure hop-by-hop. Specifically, it first selects the top K pipe segments with the highest probabilities as potential failure sources (K=5 in this embodiment) based on the failure probability prediction results. For each failure source node s, the decoder iteratively performs the following operations: at the end node of the currently predicted propagation path, based on the embedding of the node and the characteristics of its neighboring nodes, it calculates the conditional probability of propagation to each neighboring node, and samples according to physical constraints such as temperature gradient direction and pressure difference direction, selecting the neighboring node with the highest probability to join the propagation path; this process is repeated until the termination condition is reached (propagation probability is below the threshold of 0.05 or the propagation steps reach the maximum number of steps, 10). Finally, it outputs the node sequence of each propagation path and its corresponding conditional probability product as the path probability. For the prediction of propagation arrival time, this layer models the propagation process as a time prediction regression problem: based on the initial state characteristics of the failure source node and the temperature and pressure difference characteristics of adjacent nodes, it predicts the expected time required for the failure to propagate to each downstream node. The prediction layer employs a separate regression sub-network containing three fully connected layers (dimensions 256→128, 128→64, and 64→1, respectively). It maps node embeddings and time step encodings to logarithmic values ​​of propagation arrival times, then uses exponential operations to restore the actual time values ​​in hours. During training, the regression task uses the mean absolute error loss function, and the prediction results are accompanied by 95% confidence intervals, calculated through multiple random deactivation (Monte Carlo Dropout) samplings in the model ensemble.

[0089] In this embodiment, all prediction heads employ a multi-task joint learning strategy during training. The total loss function consists of a weighted sum of three parts: failure probability prediction loss (using binary cross-entropy loss with a weight coefficient of 1.0), propagation path prediction loss (using negative log-likelihood loss with a weight coefficient of 0.8), and arrival time prediction loss (using Huber loss with a weight coefficient of 0.6). The complete graph neural network model constructed in this step has approximately 6.85 million parameters, all of which are optimized through the training process in step S5.

[0090] To facilitate engineers' understanding of the physical basis of the prediction results, this embodiment incorporates an interpretability enhancement design into the model construction. While the prediction layer outputs the prediction results, the model calculates the contribution of each input feature to the prediction results through backpropagation of attention weights (i.e., feature attribution analysis), outputting a ranking of the main influencing factors for the failure probability prediction of each pipe segment. For example, if the increased failure probability of a pipe segment is mainly caused by an increase in the creep damage factor, the model will label it "creep damage dominant"; if it is caused by an increase in the amplitude of recent temperature fluctuations, it will be labeled "thermal fatigue dominant." This interpretable output helps maintenance personnel quickly locate the root cause of risks and formulate targeted remedial measures.

[0091] S5 uses historical failure case data, simulation data, and real-time monitoring data to train and optimize the failure propagation prediction model; The training data is constructed by fusing three sources: historical failure case data, Monte Carlo simulation data, and real-time monitoring data. These three sources complement each other and each has its own focus.

[0092] The training data is constructed by fusing three sources: historical failure case data, Monte Carlo simulation data, and real-time monitoring data. These three sources complement each other and each has its own focus.

[0093] Regarding historical failure case data, nearly 15 years of pipeline defect records, failure accident reports, and non-destructive testing historical archives from the target power plant were collected. Through data cleaning and standardization, a total of 23 valid failure cases were obtained. For each failure case, the following information was recorded in detail: the location of the pipe segment where the failure occurred (corresponding to the node number in the diagram structure), the unit operating conditions at the time of the failure (temperature, pressure, load), the operating history before the failure (cumulative operating time, number of start-ups and shutdowns, recent temperature fluctuations), the failure type (creep cracking, fatigue cracking, corrosion thinning, etc.), the post-failure handling records, and whether the failure propagated to adjacent pipe segments and its propagation range. Among the 23 failure cases, 8 cases involved failure propagation, with propagation path lengths ranging from 1 to 4 pipe segments. Due to the extremely limited number of real cases, especially the scarcity of cases involving multi-step propagation, historical failure case data was mainly used as core benchmark data for model fine-tuning and validation, and not as the primary source of training data.

[0094] Regarding Monte Carlo simulation data, based on the high-precision digital twin model established in Example 1 and the multiphysics coupling analysis method verified in Example 2, a large amount of failure propagation simulation data under different initial conditions is generated using the Monte Carlo stochastic simulation method to make up for the lack of real data. The specific implementation method for generating simulation data is as follows: First, 1 to 3 nodes are randomly selected from the 126 nodes of the pipeline system as initial failure sources. The selection of failure sources adopts a stratified sampling strategy to ensure coverage of different pipe segment types (elbows, tees, welds, and straight pipes). Second, the operating condition parameters of the pipeline system at the time of failure are randomly set, including uniform random sampling of steam temperature in the range of 595 to 610 degrees Celsius, uniform random sampling of steam pressure in the range of 25 to 27.5 MPa, and random sampling of the actual creep and fatigue parameters of the pipeline material within their specification tolerances according to a normal distribution, in order to simulate the parameter uncertainty in actual operation. Then, based on the multiphysics coupling analysis method in step S2, the temperature field, stress field, and damage distribution of the pipeline system under this condition are calculated, and the failure probability of each pipe segment in the current state is determined according to the damage accumulation criterion. Finally, a propagation model based on physical rules (using a weighted function of temperature gradient, stress transmission path, and material continuity as the propagation probability) is used to simulate the propagation process of failure from the initial source point to adjacent pipe segments, and the propagation path, the time of propagation to each node, and the final impact range are recorded. Using the methods described above, this embodiment generated a total of 10,000 sets of Monte Carlo simulation data. Each set of data contains a complete graph structure input and corresponding failure propagation output labels (including the failure probability of each node, propagation path sequence, and propagation arrival time). The generated simulation data covers various scenarios such as single-source failure (approximately 6,000 sets), dual-source failure (approximately 3,000 sets), and triple-source failure (approximately 1,000 sets), providing sufficient sample diversity for model training.

[0095] For real-time monitoring data, pipeline operation monitoring data from the past three years stored in the target power plant's SIS system were collected. The monitoring point layout included 86 temperature monitoring points (including thermocouples and infrared thermometers), 52 pressure monitoring points, 38 strain monitoring points, and 24 vibration monitoring points, with a sampling frequency of once per minute, totaling approximately 2.3TB of data. The raw monitoring data underwent data cleaning and quality control: Sliding median filtering was used to remove outliers significantly exceeding reasonable ranges (such as temperature fluctuations exceeding 50 degrees Celsius or pressure fluctuations exceeding 5 MPa); linear interpolation was used to fill in short-term data gaps caused by sensor malfunctions (missing time less than 2 hours); and long-term continuous missing data segments were directly deleted and marked as missing. The cleaned effective data covered the complete monitoring cycle from the current moment backwards for three years, with a time resolution of one data point every 10 minutes, forming a time-series dataset. For each time point in the monitoring data, interpolation was used to map it to the nodes of the graph structure constructed in step S3, forming a timestamped graph state sequence, providing a real-world data foundation for the model's time-series training.

[0096] To address the multi-task learning problem of failure propagation prediction, this embodiment designs a composite loss function, which consists of a weighted sum of classification loss, regression loss, and graph structure loss, to simultaneously optimize the three tasks of failure probability prediction, propagation time prediction, and propagation path prediction.

[0097] The classification loss is used to optimize the prediction of the failure probability of each pipe segment, and a weighted binary cross-entropy loss function is adopted. Since the vast majority of pipe segments in the pipeline system will not fail under normal operating conditions, the positive and negative samples are extremely imbalanced (the proportion of failed pipe segments to the total pipe segments is usually less than 5%). In this embodiment, positive samples are given a higher weight coefficient to enhance the model's sensitivity to failure events.

[0098] Regression loss is used to optimize the prediction of failure propagation arrival time. The Huber loss function is adopted to balance the smoothness of the mean squared error and the robustness of the mean absolute error to outliers. The Huber loss uses squared loss when the prediction error is less than the threshold δ, and linear loss when the error is greater than δ. In this embodiment, δ=1.0 (logarithmic unit of time prediction) is used.

[0099] Graph structure loss is used to maintain consistency between prediction results and physical mechanisms. It penalizes prediction patterns that do not conform to physical laws by imposing constraints on the spatial distribution of predicted failure probabilities. Specifically, it includes two aspects: first, spatial smoothness constraint, which penalizes drastic jumps in failure probabilities between adjacent nodes, because in actual physical processes, failure probabilities in a pipeline network should change continuously; second, consistency constraint in propagation direction, which penalizes failure propagation probabilities in the opposite direction of temperature gradient and pressure difference. In the early stages of training, the classification loss has a relatively large weight to prioritize the accuracy of failure detection; as training progresses, the contributions of regression loss and graph structure loss gradually increase to ensure the prediction accuracy of propagation time and direction.

[0100] To address the core challenge of scarce real failure samples, this embodiment employs a three-stage training strategy: "pre-training with simulated data + fine-tuning with real data + contrastive learning enhancement".

[0101] The first stage is pre-training with simulation data. 10,000 sets of Monte Carlo simulation data were divided into a pre-training training set (8,000 sets), a pre-training validation set (1,000 sets), and a pre-training test set (1,000 sets) in an 8:1:1 ratio. The graph neural network constructed in step S4 was initially trained using the simulation training set. The optimizer was AdamW, with an initial learning rate of 0.001, a batch size of 32, and 500 training epochs. A cosine annealing decay strategy was used, decreasing the learning rate every 50 epochs, with a minimum learning rate of 0.00001. During the pre-training stage, the complete composite loss function was used for optimization, and an early stopping strategy was employed to monitor the validation set loss. If the validation loss did not decrease for 30 consecutive epochs, training was terminated to prevent overfitting to the simulation data. After pre-training, the model achieved a failure probability prediction accuracy of 87.6% on the simulation test set, and a propagation path prediction hit rate (the proportion of predicted paths with a Jaccard similarity greater than 0.6 to the actual paths) of 79.3%.

[0102] The second stage involves fine-tuning with real data. Data from 23 historical failure cases were merged with monitoring data from 47 time periods selected from real-time monitoring data exhibiting significant anomalies, forming a fine-tuning dataset. This dataset was divided into a 7:3 ratio for a training set (49 samples) and a validation set (21 samples). During fine-tuning, the optimal model parameters saved from the pre-training stage were loaded as initialization, the learning rate was reduced to 0.0001, the batch size was set to 16, and the training run was 200 epochs. To address the extremely small sample size, the following data augmentation techniques were employed: random edge dropping (DropEdge, drop rate 0.1 to 0.2 random) on graph-structured data, node feature masking (randomly masking some feature dimensions, masking rate 0.1), and data augmentation based on physical perturbations (adding small-amplitude Gaussian noise conforming to physical constraints to temperature and stress field features). The fine-tuning training also employed an early stopping strategy; training was terminated if the loss did not decrease for 20 consecutive validation epochs. After fine-tuning, the model achieved a backtest accuracy of 91.3% on historical failure cases and a propagation path prediction accuracy of 82.6% on the validation set.

[0103] The third stage is contrastive learning enhancement. Considering that even after fine-tuning, the number of real failure samples is still insufficient for the model to fully learn the discriminative features of various failure modes, this embodiment introduces a contrastive learning strategy to further enhance feature representations. The basic idea of ​​contrastive learning is to construct positive sample pairs (different data-enhanced versions of the same failure case) and negative sample pairs (between different failure cases or normal samples), training the model so that the embedding representations of positive sample pairs are close to each other, and the embedding representations of negative sample pairs are far apart. In specific implementation, after the fine-tuning stage, the parameters of the first three layers of the model are fixed, and a contrastive learning projection head (containing two linear layers with dimensions of 256→128 and 128→64) is added to the last feature layer. An additional 50 rounds of training are performed using the InfoNCE contrastive loss, with a batch size of 32 (containing 16 positive sample pairs and 16 negative sample pairs), and a learning rate of 0.00005. The positive sample pairs used in contrastive learning are generated by applying different random data augmentations (node ​​feature mask ratios of 0.1 and 0.15, and edge dropout rates of 0.1 and 0.15) to the same fine-tuned sample. After the comparative learning phase is completed, the projection head is removed, and the optimized parameters of the backbone network are retained as the final model parameters.

[0104] All model training in this embodiment was performed on a deep learning workstation equipped with an NVIDIA A100 GPU (40GB VRAM), using PyTorch 1.13.1 and PyTorch Geometric 2.3.0. The total training time was approximately 36 hours (including approximately 28 hours of pre-training, approximately 6 hours of fine-tuning, and approximately 2 hours of comparative learning).

[0105] In addition to the hyperparameters mentioned above, this embodiment also systematically tuned the following hyperparameters: the number of attention heads in the node encoding layer was optimally set to 8 through grid search among 4, 8, and 16; the number of stacked layers in the graph convolutional propagation layer was optimally set to 2 through experimentation among 1, 2, and 3, as more layers would lead to oversmoothing (i.e., the features of each node tend to be consistent); the number of Transformer encoder layers was optimally set to 4 among 2, 4, and 6; the hidden layer dimension was set to 256 based on the balance between parameter quantity and performance among 128, 256, and 512; the Dropout ratio was optimally set to 0.2 through cross-validation within the range of 0.1 to 0.5; the batch size was optimally set to 32 through testing among 16, 32, and 64 to balance training stability and convergence speed; the initial learning rate was set to 0.001, and gradually reduced to 0.00001 over 500 training rounds using a cosine annealing strategy. The final values ​​of all the above hyperparameters were determined through systematic experimental comparison on the validation set, with minimizing the validation set loss as the selection criterion.

[0106] To prevent overfitting, in addition to Dropout and early stopping strategies, two regularization techniques were employed: weight decay (L2 regularization with a coefficient of 0.0005) and gradient clipping (maximum gradient norm limited to 1.0). During training, model checkpoints were saved every 50 epochs, and the checkpoint with the lowest validation set loss was selected as the optimal model for deployment.

[0107] After the model training was completed, the predictive performance of the model was comprehensively verified using multiple evaluation metrics. The evaluation was conducted on an independent test dataset, which included 5 real failure cases that did not participate in any training process and 1000 sets of test samples generated by simulation (which did not overlap with the pre-trained test set).

[0108] In terms of failure probability prediction, the model's AUC (Area Under the Receiver Operating Characteristic) reached 0.932, indicating that the model has a good ability to distinguish between failures and non-failures. When the failure probability threshold was set at 0.05 (corresponding to the blue warning threshold), the model's recall was 88.6%, precision was 79.2%, and F1 score was 83.6%; when the threshold was 0.20 (corresponding to the yellow warning threshold), the recall was 76.3%, precision was 87.1%, and F1 score was 81.4%.

[0109] For propagation path prediction, Jaccard similarity (the intersection-union ratio of the predicted propagation path node set to the actual propagation path node set) and hit rate (the proportion of actual propagation nodes correctly included in the predicted path) were used for evaluation. The average Jaccard similarity on the test set was 0.68, and the average hit rate was 0.82. For propagation arrival time prediction, the average absolute error between the predicted and actual times was 3.2 hours, representing a 62% improvement in accuracy compared to traditional time estimation methods based on empirical formulas (with an average absolute error of approximately 8.5 hours).

[0110] The model trained in this embodiment was compared with a directly trained model without transfer learning and a fine-tuned model without contrastive learning through ablation experiments. The results show that the transfer learning strategy improved the model's AUC for predicting failure probabilities on real data from 0.867 to 0.932 (an improvement of 7.5%), while contrastive learning further improved the AUC to 0.942. The complete training strategy improved the Jaccard similarity of the propagation path from 0.52 to 0.68 (an improvement of 30.8%), fully validating the effectiveness of the three-stage training strategy. Furthermore, using the training scheme in this embodiment, the model's false positive rate (the proportion of normal pipe sections predicted as failed) on real failure cases was 6.8%, significantly lower than the 21.3% false positive rate of the model trained only with simulation data, demonstrating the crucial role of fine-tuning with real data in improving model reliability.

[0111] The optimal model parameters after training are permanently saved as a model weight file (approximately 156MB in size) and encapsulated into a standardized inference interface for real-time prediction deployment in step S6. All hyperparameter configurations, training logs, and performance evaluation reports during model training are archived to support subsequent model traceability, reproduction, and continuous optimization.

[0112] Regarding historical failure case data, nearly 15 years of pipeline defect records, failure accident reports, and non-destructive testing historical archives from the target power plant were collected. Through data cleaning and standardization, a total of 23 valid failure cases were obtained. For each failure case, the following information was recorded in detail: the location of the pipe segment where the failure occurred (corresponding to the node number in the diagram structure), the unit operating conditions at the time of the failure (temperature, pressure, load), the operating history before the failure (cumulative operating time, number of start-ups and shutdowns, recent temperature fluctuations), the failure type (creep cracking, fatigue cracking, corrosion thinning, etc.), the post-failure handling records, and whether the failure propagated to adjacent pipe segments and its propagation range. Among the 23 failure cases, 8 cases involved failure propagation, with propagation path lengths ranging from 1 to 4 pipe segments. Due to the extremely limited number of real cases, especially the scarcity of cases involving multi-step propagation, historical failure case data was mainly used as core benchmark data for model fine-tuning and validation, and not as the primary source of training data.

[0113] Regarding Monte Carlo simulation data, based on the high-precision digital twin model established in Example 1 and the multiphysics coupling analysis method verified in Example 2, a large amount of failure propagation simulation data under different initial conditions is generated using the Monte Carlo stochastic simulation method to make up for the lack of real data. The specific implementation method for generating simulation data is as follows: First, 1 to 3 nodes are randomly selected from the 126 nodes of the pipeline system as initial failure sources. The selection of failure sources adopts a stratified sampling strategy to ensure coverage of different pipe segment types (elbows, tees, welds, and straight pipes). Second, the operating condition parameters of the pipeline system at the time of failure are randomly set, including uniform random sampling of steam temperature in the range of 595 to 610 degrees Celsius, uniform random sampling of steam pressure in the range of 25 to 27.5 MPa, and random sampling of the actual creep and fatigue parameters of the pipeline material within their specification tolerances according to a normal distribution, in order to simulate the parameter uncertainty in actual operation. Then, based on the multiphysics coupling analysis method in step S2, the temperature field, stress field, and damage distribution of the pipeline system under this condition are calculated, and the failure probability of each pipe segment in the current state is determined according to the damage accumulation criterion. Finally, a propagation model based on physical rules (using a weighted function of temperature gradient, stress transmission path, and material continuity as the propagation probability) is used to simulate the propagation process of failure from the initial source point to adjacent pipe segments, and the propagation path, the time of propagation to each node, and the final impact range are recorded. Using the methods described above, this embodiment generated a total of 10,000 sets of Monte Carlo simulation data. Each set of data contains a complete graph structure input and corresponding failure propagation output labels (including the failure probability of each node, propagation path sequence, and propagation arrival time). The generated simulation data covers various scenarios such as single-source failure (approximately 6,000 sets), dual-source failure (approximately 3,000 sets), and triple-source failure (approximately 1,000 sets), providing sufficient sample diversity for model training.

[0114] For real-time monitoring data, pipeline operation monitoring data from the past three years stored in the target power plant's SIS system were collected. The monitoring point layout included 86 temperature monitoring points (including thermocouples and infrared thermometers), 52 pressure monitoring points, 38 strain monitoring points, and 24 vibration monitoring points, with a sampling frequency of once per minute, totaling approximately 2.3TB of data. The raw monitoring data underwent data cleaning and quality control: Sliding median filtering was used to remove outliers significantly exceeding reasonable ranges (such as temperature fluctuations exceeding 50 degrees Celsius or pressure fluctuations exceeding 5 MPa); linear interpolation was used to fill in short-term data gaps caused by sensor malfunctions (missing time less than 2 hours); and long-term continuous missing data segments were directly deleted and marked as missing. The cleaned effective data covered the complete monitoring cycle from the current moment backwards for three years, with a time resolution of one data point every 10 minutes, forming a time-series dataset. For each time point in the monitoring data, interpolation was used to map it to the nodes of the graph structure constructed in step S3, forming a timestamped graph state sequence, providing a real-world data foundation for the model's time-series training.

[0115] S6 inputs real-time sensor data into the trained failure propagation prediction model and outputs the failure probability, failure propagation path, and propagation time prediction results for each pipe segment.

[0116] After model training, the optimal model parameters are solidified and encapsulated into a standardized inference interface, which is then deployed on the power plant's SIS system server. The inference task is set to execute automatically every 24 hours, using real-time operational data collected that day for prediction. The real-time data input consists of two parts: static graph data, including the node numbers, types, geometric parameters of 126 nodes, and the connection relationships of 132 edges; this data is loaded during initial deployment and only needs to be updated when physical modifications are made to the pipeline system; and dynamic operational data, including real-time readings of each sensor point at the current moment and historical monitoring data from the past 12 months. For situations where there are fewer sensor points than graph nodes, the operational parameters of pipe segments without directly deployed sensors are estimated using spatial interpolation methods. These are then compiled into complete node and edge feature matrices, which are input into the model for inference.

[0117] During model inference, the node feature encoding layer encodes 64-dimensional node features into 128-dimensional hidden layer features. The graph convolutional propagation layer performs information transfer on the pipeline network topology through two layers of GAT convolution, enabling each node to absorb the state information of its neighboring nodes. The temporal feature extraction layer processes the graph state sequence of 12 historical time steps through a Transformer encoder to extract the temporal evolution law of failure propagation. Finally, three parallel prediction heads output the failure probability, failure propagation path and its conditional probability for each pipe segment, as well as the estimated time for failure propagation to reach each pipe segment. The entire inference process runs on the CPU, with a single inference taking approximately 10 seconds, meeting the real-time requirements of the field.

[0118] The prediction results are visualized through the human-machine interface of the power plant's SIS system. Failure probabilities are overlaid on the pipeline system layout diagram as a heatmap, with different colors representing different risk levels: green indicates safe (probability below 0.01), orange indicates concern (probability 0.01 to 0.05), and red indicates a warning state (probability above 0.05; see step S7 for specific warning levels). The right panel of the interface lists the top 10 risky pipe segments in descending order of failure probability, including the segment name, failure probability value, and main risk contribution factors. Propagation paths are displayed as a directed graph. Starting from the node with the highest predicted failure probability, the system generates and displays five propagation paths with the highest probability. The line thickness represents the conditional probability of propagation for that segment, and the expected propagation arrival time is marked on the line segment. The failure propagation process is also demonstrated using animated pulses. The predicted propagation arrival time for each pipe segment is simultaneously displayed as a 95% confidence interval, calculated using the Monte Carlo Dropout sampling method.

[0119] In addition to the interface display, the system automatically generates a structured PDF report containing a complete list of failure probabilities for all 126 pipe segments, a ranking of the top 20 risky pipe segments, detailed diagrams of the top 5 predicted propagation paths, historical failure probability trend charts for each key pipe segment, and a risk analysis summary and maintenance recommendations automatically generated by the system based on the prediction results. The report is for archiving and distribution by the equipment management department.

[0120] To continuously evaluate model performance and support model optimization, the system has established an online prediction verification and feedback mechanism. Operations personnel fill in actual failure details (whether a failure occurred, the location of the failure, etc.) through the feedback interface. Based on this, the system marks the cases as positive or negative feedback samples and stores them in the iterative training library. When monthly performance indicators show a significant decline or the cumulative number of negative feedback samples reaches a certain amount, the model retraining process is triggered. After the updated model is verified and approved, it is redeployed and launched online, achieving continuous model optimization and adaptive updates.

[0121] The method further includes step S7: graded early warning, setting multiple levels of early warning based on the prediction results, and triggering the corresponding level of early warning when the failure probability exceeds a preset threshold.

[0122] The tiered early warning system includes: Blue alert: The failure probability exceeds the first threshold; enhanced monitoring is recommended. Yellow alert: The probability of failure exceeds the second threshold, or it is predicted that the virus may spread within 48 hours. It is recommended to prepare for maintenance. Red Alert: If the probability of failure exceeds the third threshold, or if it is predicted that the virus may spread within 24 hours, an emergency response should be initiated immediately.

[0123] When there are heterogeneous areas of different materials welded in the piping system, set the heterogeneous edge type for the heterogeneous material welded joints in the diagram structure and assign a higher failure propagation weight.

[0124] The tiered early warning mechanism transforms the quantitative prediction results output in step S6 into actionable operation and maintenance instructions, with three levels of early warning. A blue warning is triggered when the failure probability of any pipe section exceeds 5%, or the total probability of any predicted propagation path exceeds 10%. Upon triggering, the system notifies operation and maintenance personnel via an interface prompt, recommending that the monitoring frequency of that pipe section be increased from every 10 minutes to every 2 minutes, with manual inspection every 8 hours, and that non-destructive testing be scheduled for the next maintenance. A yellow warning is triggered when the failure probability of any pipe section exceeds 20%, or when a failure is predicted to propagate to adjacent pipe sections within 48 hours, or the total probability of any propagation path exceeds 25%. Upon triggering, the system automatically sends SMS notifications to the shift supervisor, equipment management engineer, and maintenance supervisor, requiring a risk assessment meeting to be held within 12 hours, a comprehensive non-destructive test of the pipe section to be conducted, spare parts to be prepared, and a maintenance plan to be developed. If necessary, the unit load should be reduced to below 85%. The red alert is triggered when the failure probability of any pipe section exceeds 40%, or when it is predicted that the failure may propagate to adjacent pipe sections within 24 hours, or when the total probability of any propagation path exceeds 40%. Upon triggering, the system automatically triggers a telephone voice notification in addition to interface alerts and SMS notifications, requiring immediate activation of the emergency response plan. Emergency non-destructive testing must be completed within 4 hours. If a serious defect is confirmed, planned shutdown and maintenance must be arranged immediately. If immediate shutdown is not possible, the load must be reduced to below 60% and the steam temperature to below 570℃, with continuous close monitoring until maintenance is completed. The determination of the three-level warning threshold is based on historical statistics from this embodiment: when the failure probability is below 5%, the actual failure rate is only 1.2%; between 5% and 20%, the actual failure rate is 8.7%; between 20% and 40%, the actual failure rate is 31.5%; and above 40%, the actual failure rate reaches 68.4%, verifying the rationality of the threshold division.

[0125] Dissimilar material welded joints are among the weakest links in a piping system, and their failure propagation characteristics differ significantly from those of pipe sections connected by the same material. In this embodiment, the dissimilar material welded areas are specially treated during the structure construction stage of step S3. When the material types of the two pipe section nodes corresponding to a certain edge are different, the system automatically marks that edge as a dissimilar edge type, adds a dissimilar edge type code and a dissimilar steel weld identifier to the edge feature vector, and records material property difference parameters such as the difference in thermal expansion coefficient, thermal conductivity, and hardness between the two base materials. In this embodiment, 8 dissimilar edges were identified out of 132 edges, corresponding to 8 locations of dissimilar steel welded joints.

[0126] During model training and inference, the graph attention network automatically incorporates dissimilar steel welding identifiers into the attention calculation. According to the training results of this embodiment, the attention weights learned by the model for dissimilar edges are on average 25% to 40% higher than those for homogeneous edges, reflecting the physical characteristic that dissimilar welded joints are more likely to become failure propagation channels. Simultaneously, due to the different creep strengths on both sides of the dissimilar steel welded joint, the probability of failure propagating from the weaker side to the stronger side is higher. The model automatically learns the directional bias through differences in hardness and thermal expansion coefficients (in this embodiment, P92 strength is higher than F92, and the bias factor for propagation from the F92 side to the P92 side is approximately 1.3). In the warning output, for pipe segment nodes near dissimilar edges, the system additionally labels them with a "dissimilar steel welded joint" risk tag. The maintenance recommendations prioritize phased array ultrasonic testing (with higher accuracy than conventional ultrasonic testing), and remind users that specialized welding process evaluation and transition layer welding measures are required during repair. In the typical prediction of this embodiment, the dissimilar edge e 41 Two-end nodes v 38 and v 39 The failure probabilities were 0.29 and 0.18, respectively. The system triggered a yellow warning and specifically marked the risk of dissimilar steel welded joints. It was recommended to reduce the operating temperature by 15°C before maintenance to reduce thermal stress.

[0127] Electronic device 200 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 200 may include, but is not limited to, processor 201 and memory 202. Those skilled in the art will understand that... Figure 2 This is merely an example of electronic device 200 and does not constitute a limitation on electronic device 200. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0128] The processor 201 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0129] The memory 202 can be an internal storage unit of the electronic device 200, such as a hard disk or RAM of the electronic device 200. The memory 202 can also be an external storage device of the electronic device 200, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 200. Furthermore, the memory 202 can include both internal and external storage units of the electronic device 200. The memory 202 is used to store the computer program 203 and other programs and data required by the electronic device. The memory 202 can also be used to temporarily store data that has been output or will be output.

[0130] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit it. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A method for predicting the failure propagation of high-temperature steam pipelines in thermal power plants, characterized in that, Includes the following steps: S1: Construct a digital twin model of a high-temperature steam pipeline system, wherein the digital twin model includes pipeline geometric information, pipeline topology information, material property information, and operating condition information; S2: Based on the digital twin model, perform thermo-mechanical coupling analysis to calculate the temperature field distribution and stress field distribution of the pipeline system, and evaluate the creep damage and fatigue damage of the pipeline material; S3: Convert the pipeline system into graph structure data, define each pipe segment as a node in the graph, define the connection relationship between pipe segments as the edge of the graph, and assign each node a feature vector containing material degradation index, stress concentration factor and temperature distribution characteristics; S4 constructs a failure propagation prediction model based on graph neural networks. The failure propagation prediction model includes a node feature encoding layer, a graph convolutional propagation layer, a temporal feature extraction layer, and a failure probability prediction layer. S5 uses historical failure case data, simulation data, and real-time monitoring data to train and optimize the failure propagation prediction model; S6 inputs real-time sensor data into the trained failure propagation prediction model and outputs the failure probability, failure propagation path, and propagation time prediction results for each pipe segment.

2. The method according to claim 1, characterized in that, The pipeline geometry information in step S1 includes pipe diameter, wall thickness, elbow angle, tee position, and valve position; The pipeline topology information includes node numbers, connection relationships, pipe segment lengths, and support point locations; The material property information includes the pipe material and its mechanical property parameters at different temperatures; The operating condition information includes design temperature, design pressure, and the actual operating temperature and pressure fluctuation range.

3. The method according to claim 1, characterized in that, The thermal coupling analysis based on the digital twin model in step S2 includes: Calculate the temperature field distribution of the piping system under design conditions; The dynamic response of the temperature field during simulated start-up and shutdown processes and load changes; Calculate the thermal stress distribution by combining the temperature field; Assess material creep damage and fatigue damage based on temperature and stress fields.

4. The method according to claim 1, characterized in that, The feature vector in step S3 includes the temperature difference, pressure difference, and constraint stiffness between the connecting pipe segments.

5. The method according to claim 1, characterized in that, The node feature encoding layer described in step S4 uses a multi-head attention mechanism to encode the multidimensional features of the pipeline nodes and extract the local state representation of each pipe segment. The graph convolutional propagation layer uses a graph convolutional neural network or a multi-hop graph attention network for information propagation; The temporal feature extraction layer uses a temporal convolutional network or a Transformer encoder to extract the temporal evolution features of the pipeline state and capture the temporal dynamic characteristics of failure propagation. The failure probability prediction layer predicts the failure probability of each pipe segment, the propagation path of the failure in the piping system, the time it takes for the failure to reach each pipe segment, and the probability distribution of different propagation paths.

6. The method according to claim 1, characterized in that, It also includes step S7: graded early warning, which sets up multiple levels of early warning based on the prediction results, and triggers the corresponding level of early warning when the failure probability exceeds the preset threshold.

7. The method according to claim 6, characterized in that, The tiered early warning system includes: Blue alert: The failure probability exceeds the first threshold; enhanced monitoring is recommended. Yellow alert: The probability of failure exceeds the second threshold, or it is predicted that the virus may spread within 48 hours. It is recommended to prepare for maintenance. Red Alert: If the probability of failure exceeds the third threshold, or if it is predicted that the virus may spread within 24 hours, an emergency response should be initiated immediately.

8. The method according to claim 1, characterized in that, When there are heterogeneous areas of different materials welded in the piping system, set the heterogeneous edge type for the heterogeneous material welded joints in the diagram structure and assign a higher failure propagation weight.

9. An electronic device, characterized in that, include: One or more processors; A storage unit for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the method of claim 1.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in claim 1.

Images