A method and system for weld creep-fatigue failure prediction based on machine learning

Abstract

This invention discloses a machine learning-based method and system for predicting weld creep-fatigue failure, relating to the field of weld creep-fatigue failure prediction technology. The method includes: acquiring a peak-shaving response profile of a target weld under deep peak-shaving conditions; inputting the peak-shaving response profile into a weld failure prediction model to obtain the initial life level and suspected failure indicators of the target weld; determining the damage verification area of ​​the target weld and collecting damage evidence packages; generating weld verification samples; inputting the weld verification samples into the weld failure prediction model; and outputting the creep-fatigue failure prediction result of the target weld. This invention solves the technical problem of existing technologies relying solely on model data for prediction and lacking physical flaw detection data for verification, leading to large deviations in weld creep-fatigue life prediction under deep peak-shaving conditions. It achieves accurate prediction of weld creep-fatigue failure and effectively improves the accuracy of weld remaining life prediction under high-temperature peak-shaving conditions.

Description

Technical Field

[0001] This invention relates to the field of weld creep fatigue failure prediction technology, specifically to a machine learning-based method and system for predicting weld creep-fatigue failure. Background Technology

[0002] Under the background of normalized deep peak shaving operation of thermal power units, frequent load changes and long-term high-temperature and high-pressure stable operation of the units can easily cause the welds of pressure equipment to be simultaneously subjected to alternating thermal loads and continuous high-temperature static loads, resulting in creep and fatigue coupled damage. Existing predictions of weld creep fatigue life mostly rely on simulation parameters or unit operating data to directly calculate through machine learning models, without combining the actual physical and chemical flaw detection data of the welds to verify and correct the model prediction results. The models rely solely on historical operating data for prediction, which is easily affected by material dispersion and fluctuations in on-site operating conditions. This results in large deviations between the prediction results and the actual aging damage of the welds, and insufficient accuracy in predicting the remaining life, making it difficult to provide reliable data support for the formulation of unit weld maintenance cycles.

[0003] Existing technologies rely solely on model data for prediction and lack verification with physical flaw detection data, resulting in significant deviations in the prediction of weld creep fatigue life under deep peak shaving conditions. Summary of the Invention

[0004] This application provides a machine learning-based method and system for predicting weld creep-fatigue failure, which addresses the technical problem that existing technologies rely solely on model data for prediction and lack physical flaw detection data for verification, resulting in large deviations in the prediction of weld creep fatigue life under deep peak-shaving conditions.

[0005] In view of the above problems, this application provides a method and system for predicting weld creep-fatigue failure based on machine learning.

[0006] The first aspect of this application provides a machine learning-based method for predicting weld creep-fatigue failure, the method comprising:

[0007] A peak-shaving response profile of the target weld under deep peak-shaving conditions is obtained, wherein the peak-shaving response profile includes weld area identification, temperature fluctuation data, pressure fluctuation data, load holding data, and creep-fatigue test life data. The peak-shaving response profile is input into a weld failure prediction model to obtain the initial life level and suspected failure identifier of the target weld. When the suspected failure identifier meets the verification conditions, the damage verification area of ​​the target weld is determined according to the weld area identification, and the damage evidence package of the damage verification area is collected. The damage evidence package is associated with the peak-shaving response profile to generate a weld verification sample. The weld verification sample is input into the weld failure prediction model to obtain the failure master control identifier and life correction level of the target weld. The initial life level is corrected according to the life correction level, and the creep-fatigue failure prediction result of the target weld is output.

[0008] A second aspect of this application provides a machine learning-based weld creep-fatigue failure prediction system, the system comprising:

[0009] The system includes the following modules: a peak-shaving response profile acquisition module, used to acquire a peak-shaving response profile of the target weld under deep peak-shaving conditions, wherein the peak-shaving response profile includes weld area identification, temperature fluctuation data, pressure fluctuation data, load holding data, and creep-fatigue test life data; a suspected failure identifier acquisition module, used to input the peak-shaving response profile into a weld failure prediction model to obtain the initial life level and suspected failure identifier of the target weld; a damage evidence package acquisition module, used to determine the damage verification area of ​​the target weld based on the weld area identification when the suspected failure identifier meets the verification conditions, and to collect the damage evidence package of the damage verification area; a weld verification sample generation module, used to associate the damage evidence package with the peak-shaving response profile to generate a weld verification sample; a life correction level acquisition module, used to input the weld verification sample into the weld failure prediction model to obtain the failure master control identifier and life correction level of the target weld; and a prediction result output module, used to correct the initial life level according to the life correction level and output the creep-fatigue failure prediction result of the target weld.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] A peak-shaving response profile of the target weld under deep peak-shaving conditions is obtained; the peak-shaving response profile is input into a weld failure prediction model to obtain the initial life level and suspected failure indicator of the target weld; when the suspected failure indicator meets the verification conditions, the damage verification area of ​​the target weld is determined according to the weld area indicator, and the damage evidence package of the damage verification area is collected; a weld verification sample is generated; the weld verification sample is input into the weld failure prediction model to obtain the failure master control indicator and life correction level of the target weld; the initial life level is corrected according to the life correction level, and the creep-fatigue failure prediction result of the target weld is output. This achieves the technical effect of accurately predicting weld creep fatigue failure and effectively improving the accuracy of weld remaining life prediction under high-temperature peak-shaving conditions. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic diagram of a machine learning-based method for predicting weld creep-fatigue failure is provided for an embodiment of this application.

[0014] Figure 2 This is a schematic diagram of a machine learning-based weld creep-fatigue failure prediction system provided in an embodiment of this application.

[0015] Figure 3 A line graph showing the relationship between the daily average number of peak load adjustments and the weld hardness decay in a machine learning-based weld creep-fatigue failure prediction method provided in the application embodiment.

[0016] Figure labeling: Peak shaving response profile acquisition module 10, suspected failure identifier acquisition module 20, damage evidence package acquisition module 30, weld verification sample generation module 40, life correction level acquisition module 50, prediction result output module 60. Detailed Implementation

[0017] This application provides a machine learning-based method and system for predicting weld creep-fatigue failure, which addresses the technical problem that existing technologies rely solely on model data for prediction and lack physical inspection data for verification, resulting in large deviations in weld creep fatigue life prediction under deep peak-shaving conditions.

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0019] Example 1, as Figure 1 As shown, this application provides a machine learning-based method for predicting weld creep-fatigue failure, the method comprising:

[0020] Step S100: Obtain the peak shaving response profile of the target weld under deep peak shaving conditions, wherein the peak shaving response profile includes weld area identification, temperature fluctuation data, pressure fluctuation data, load holding data and creep-fatigue test life data.

[0021] Specifically, the process first distinguishes the zone to which the target weld belongs, generating weld area identifiers corresponding to the base metal zone, weld metal zone, or heat-affected zone. It then collects the real-time temperature fluctuation sequence of the weld, pressure fluctuation parameters caused by start-up and shutdown load changes, and load holding time data during the constant load phase throughout the entire cycle of deep peak shaving operation of the unit. Simultaneously, it retrieves three types of test data for the weld: creep endurance test life, low-cycle fatigue test life, and creep-fatigue interaction test life. The process integrates the aforementioned zone identifiers, operating parameters, and multi-type life test data, and uniformly structures and encapsulates them to generate a peak shaving response profile of the target weld under deep peak shaving conditions.

[0022] Step S200: Input the peak-shaving response profile into the weld failure prediction model to obtain the initial life level and suspected failure indicator of the target weld.

[0023] Specifically, a massive amount of sample weld peak response profiles, corresponding sample lifetime levels, and sample failure labels are pre-collected. An XGBoost integrated machine learning algorithm is used to build a weld failure prediction model. During the training phase, multi-dimensional features such as temperature, pressure, load capacity, and test lifetime within the profile are first normalized and preprocessed. Using a decision tree as the base learner, the loss function is optimized through gradient iteration, and feature splitting and pruning constraints are automatically completed. The optimal model parameters are obtained through iterative convergence, completing model training. Subsequently, the peak response profile feature data of the target weld is input into the trained XGBoost weld failure prediction model. Through parallel feature deduction of multiple decision trees and weighted summarization of results, the initial lifetime level corresponding to the target weld and a suspected failure label indicating whether the weld has a failure tendency are finally output.

[0024] Step S300: When the suspected failure identifier meets the verification conditions, the damage verification area of ​​the target weld is determined according to the weld area identifier, and the damage evidence package of the damage verification area is collected.

[0025] Specifically, after the suspected failure marker indicates that the target weld has a tendency for creep-fatigue interactive failure and meets the conditions for review initiation, the damage review area is accurately delineated based on the three types of weld area markers stored in the image storage: base material area, weld metal area, and heat-affected zone. Hardness testing, metallographic structure testing, oxidation state testing, and crack flaw detection are carried out on the review area respectively, and four measured data points are obtained in sequence: hardness change, structure deterioration, oxidation damage, and crack distribution. All test data are integrated and assembled into a damage evidence package.

[0026] Step S400: Associate the damage evidence package with the peak-shaving response profile to generate a weld verification sample.

[0027] Specifically, the damage verification area information corresponding to the damage evidence package is extracted, and the temperature fluctuation, pressure fluctuation, load holding and creep-fatigue test life data matching the area are screened from the peak response profile. The measured hardness, microstructure, oxidation, crack-related damage data and corresponding working condition test data are bound and associated one by one to generate weld verification samples.

[0028] Step S500: Input the weld verification sample into the weld failure prediction model to obtain the failure control identifier and life correction level of the target weld.

[0029] Specifically, the weld review samples bound to the working conditions and physical damage data are input into the trained XGBoost weld failure prediction model. The model combines historical peak-shaving parameters and field-measured damage characteristics to perform feature analysis, identify the dominant cause of damage, and output the failure master control identifier corresponding to fatigue master control, creep master control, or creep-fatigue interactive master control. Based on the severity of damage, the life correction level of the target weld is further calculated.

[0030] Step S600: Correct the initial life level according to the life correction level, and output the creep-fatigue failure prediction result of the target weld.

[0031] Specifically, the life correction level is compared with the initial life level and an upgrade or maintenance correction is performed. Based on the corrected life level, the remaining life range of the weld is defined. Combined with the previously identified fatigue, creep, or interactive failure control indicators, the key monitoring points of the weld are locked. Finally, the life range, failure control indicators, and key monitoring location information are integrated to output a complete weld creep-fatigue failure prediction result.

[0032] In one possible implementation, step S100 further includes:

[0033] Step S110: Obtain the weld area identifier of the target weld, wherein the weld area identifier is one of the base material area identifier, weld metal area identifier, and heat-affected zone identifier.

[0034] Step S120: Obtain temperature fluctuation data, pressure fluctuation data, and load data of the target weld during the depth peaking process.

[0035] Step S130: Obtain the creep rupture test life, low-cycle fatigue test life, and creep-fatigue interaction test life corresponding to the target weld.

[0036] Step S140: Generate the peak-shaving response profile based on the weld area identifier, temperature fluctuation data, pressure fluctuation data, load holding data, creep rupture test life, low-cycle fatigue test life, and creep-fatigue interaction test life.

[0037] Specifically, the weld structure is disassembled by combining metallographic macroscopic morphology observation with weld drawing marking. The base material area, weld metal area and heat-affected zone are distinguished by macroscopic sampling and visual inspection. A unique code is assigned according to the actual zone to generate the corresponding type of weld area identifier.

[0038] Relying on thermocouple temperature sensors and pipeline pressure transmitters embedded around the weld, the unit collects temperature and pressure time-series data in real time throughout the deep peak shaving process. Combined with the unit's DCS control system, the operating records during the constant load phase are retrieved, and three types of measured data are obtained: temperature fluctuation, pressure fluctuation, and load holding time.

[0039] Base material and welded specimens from the same batch as the target weld were collected and subjected to high-temperature creep rupture test, room temperature / high temperature low cycle fatigue test, and creep-fatigue alternating loading interactive test. The corresponding test life values ​​were recorded according to the fracture time of the specimens, thereby obtaining three types of life test data.

[0040] Standardized preprocessing was performed on weld area identification, on-site temperature fluctuation, pressure fluctuation, load holding operation data, and three types of test life data: creep endurance, low cycle fatigue, and creep-fatigue interaction. After unifying the data format, the data was integrated according to the corresponding fields to form a multi-dimensional structured data set, thus completing the construction of the peak-shaving response profile.

[0041] In one possible implementation, step S200 further includes:

[0042] Step S210: Obtain multiple sample peak-shaving response profiles, multiple sample lifetime levels, and multiple sample failure identifiers for multiple sample welds.

[0043] Step S220: Train a machine learning model based on the multiple sample peak response profiles, multiple sample lifetime levels, and multiple sample failure identifiers to obtain the weld failure prediction model.

[0044] Step S230: Input the peak response profile of the target weld into the weld failure prediction model, and output the initial life level and suspected failure indicator of the target weld.

[0045] Specifically, sample welds from different zones of multiple deep peak-shaving units were collected in batches. The corresponding area identifier, temperature and pressure load condition data, and three types of test life data for each weld were collected one by one to construct a sample peak-shaving response profile. Combined with offline physicochemical testing and long-term in-service tracking results, the life level and failure identifier of each sample weld were manually labeled to complete the collection of the model training dataset.

[0046] The XGBoost algorithm was selected to build a weld failure prediction model. The model consists of multi-layered serial CART decision trees. Each decision tree divides the sample space by feature splitting, and the overall model optimizes the prediction bias by stacking subtrees one by one using the gradient boosting approach. Before training, outlier removal, normalization scaling, and missing value completion were performed on the multi-dimensional features in the sample peak response profile, such as weld partition coding, temperature fluctuation, pressure fluctuation, holding time, and three types of test life. A multi-task training dataset was constructed by using the sample life level as the ordered classification label and the sample failure label as the binary classification label. The model was designed... The maximum tree depth is set to 6, the learning rate to 0.1, and L1 and L2 regularization coefficients are used to constrain the leaf weights. The loss function is solved using second-order Taylor expansion. In each iteration, a new decision tree is generated based on the residual of the previous model, and the prediction error is continuously corrected. The training is terminated when the loss converges and no longer decreases. After the training converges and all tree node splitting thresholds and weight parameters are solidified, a formed weld failure prediction model is obtained. Subsequently, only the peak response profile features of the target weld need to be input. The model can simultaneously output the initial life level and suspected failure indicators by performing feature discrimination and result accumulation calculation through multiple decision trees.

[0047] After standardizing the peak response profile data of the target weld to be consistent with the training set, the feature vector composed of partition identifier, temperature and pressure parameters, load data and various test lifetimes is imported into the trained XGBoost weld failure prediction model. Through the feature bifurcation discrimination of multiple CART decision trees in the model and the weighted accumulation of predicted values, the model completes the lifetime classification and failure tendency determination respectively. Finally, it outputs the initial lifetime level corresponding to the weld and the suspected failure label that indicates whether there is a creep-fatigue failure tendency.

[0048] In one possible implementation, step S300 further includes:

[0049] Step S310: When the suspected failure indicator indicates that the target weld has a tendency for creep-fatigue interactive failure, it is determined that the review conditions are met.

[0050] Step S320: Based on the weld area identifier corresponding to the suspected failure identifier, determine the damage verification area from the base metal area, weld metal area, and heat-affected zone.

[0051] Specifically, the model reads the suspected failure indicators. If the indicator determines that the weld is affected by alternating temperature and pressure and long-term load, and there is a trend of interactive failure induced by creep and fatigue coupling, then the system automatically determines that the conditions for on-site physical and chemical verification have been met.

[0052] Retrieve the suspected failure identifier and bind it to the exclusive weld area identifier stored in the peak response profile. First, use the area identifier to distinguish whether the target weld belongs to a single area among the three categories of base material area, weld metal area, and heat-affected zone. Based on the weld space range corresponding to the identifier, extend a fixed size range slightly outward along the weld thickness direction and weld length direction to accurately delineate the damage verification area used for subsequent physical and chemical testing.

[0053] In one possible implementation, step S300 further includes:

[0054] Step S330: Perform hardness testing on the damaged verification area to obtain hardness change data.

[0055] Step S340: Perform tissue testing on the damaged area to obtain tissue deterioration data.

[0056] Step S350: Perform oxidation state detection on the damage verification area to obtain oxidation damage data.

[0057] Step S360: Perform crack detection on the damage verification area to obtain crack distribution data.

[0058] Step S370: Generate the damage evidence package based on the hardness change data, microstructure deterioration data, oxidation damage data, and crack distribution data.

[0059] Specifically, a portable Vickers hardness tester is used to set up multiple test points in a grid with equal spacing within the designated damage verification area. Real-time measured hardness values ​​are collected at each point, and the difference is calculated and statistically analyzed by comparing them with the factory reference hardness value of the material. Finally, the hardness change data corresponding to the verification area is obtained.

[0060] A small amount of metal sample was taken from the damage verification area and then placed under a metallographic microscope for microscopic observation after being prepared by mounting, grinding, polishing and etching. The degradation types and degradation percentages, such as grain growth, carbide spheroidization and precipitation and grain boundary void initiation, were statistically analyzed by comparing with the standard metallographic spectrum of the material. The data were then compiled to form quantitative data on the degradation of the microstructure.

[0061] After sampling the damage verification area, the actual thickness of the oxide layer on the substrate surface was measured and the elemental composition and distribution range of the surface oxide were detected by scanning electron microscopy with energy dispersive spectroscopy. The oxide coverage area and oxidation corrosion depth were statistically analyzed by comparing with the original oxide-free baseline parameters of the parent material, and the quantitative oxidation damage data were obtained by sorting and summarizing.

[0062] An ultrasonic flaw detector was used to conduct a comprehensive scan along the weld direction in the damage review area to locate the depth and extension size of buried cracks. Then, a magnetic particle flaw detector was used to detect fine cracks on the weld surface and near the surface. The location, length, depth and extension direction of each crack were recorded one by one, and all the detection information was summarized to form complete crack distribution data.

[0063] The four types of measured data—hardness change, microstructure deterioration, oxidation damage, and crack distribution—are standardized in a unified data format. They are then archived according to the test items and bound to the corresponding weld area number and test time information. After data verification and removal of abnormal records, the data is structured, packaged, and assembled into a traceable and complete damage evidence package.

[0064] In one possible implementation, step S400 further includes:

[0065] Step S410: Extract the damage verification area corresponding to the damage evidence package.

[0066] Step S420: Extract temperature fluctuation data, pressure fluctuation data, load holding data, and creep-fatigue test life data corresponding to the damage verification area from the peak-shaving response profile.

[0067] Step S430: Bind the damage evidence package with the data corresponding to the damage verification area to generate the weld verification sample.

[0068] Specifically, the area coding information embedded in the damage evidence package is read, and the spatial range corresponding to the previously defined base material area, weld metal area or heat-affected area is located by relying on the coding for tracing, so as to accurately extract the location information of the damage review area that uniquely corresponds to the damage evidence package.

[0069] Using the weld area identifier bound to the damage review area as a search index, the target profile is accurately located in the database storing all peak-shaving response profiles. The time series data of unit deep peak-shaving temperature fluctuation, pressure fluctuation, steady-state load parameters, and creep-fatigue interaction test life data corresponding to the region are then filtered, split, and extracted to complete the separate extraction and collection of specified feature data.

[0070] Using the damage verification area number as the primary key, a one-to-one index is established to associate the measured data of hardness change, microstructure deterioration, oxidation damage, and crack distribution within the damage evidence package with the temperature and pressure, holding load, and creep fatigue life data extracted from the peak response profile. The data field format is standardized and the data integrity is verified. After encapsulation and integration, a weld verification sample is generated for subsequent iterative optimization of the model.

[0071] In one possible implementation, step S500 further includes:

[0072] Step S510: Input the weld verification sample into the weld failure prediction model.

[0073] Step S520: The weld failure prediction model identifies the main failure control identifier of the target weld based on the correspondence between the peak-shaving response profile and the damage evidence package.

[0074] Step S530: Output the life correction level of the target weld based on the failure master control identifier and damage evidence package.

[0075] Specifically, the working condition characteristics and damage measurement data in the weld verification samples are first standardized and feature aligned according to the unified preprocessing rules during model training. After removing invalid and abnormal data, the processed feature data is imported into the trained XGBoost weld failure prediction model according to the model input format for incremental iterative learning and optimization of the model.

[0076] Based on the feature importance scoring mechanism within the XGBoost model, the working condition features such as temperature and pressure alternation, load holding time, and material test life in the peak response profile of the verification sample are correlated and mapped with the four types of measured damage indicators in the damage evidence package: hardness decay, microstructure deterioration, oxidation degree, and crack development. The model calculates the contribution weight of each feature to crack initiation and material aging, compares the feature threshold ranges corresponding to the three failure modes of creep-dominated, fatigue-dominated, and creep-fatigue coupled, and classifies the damage inducing factors with the highest weight to determine the final identification and output of the failure master control identifier corresponding to the target weld.

[0077] After acquiring the failure control identifier and complete damage evidence package data, the failure type is first determined based on the failure control identifier as creep-dominated failure, fatigue-dominated failure, or creep-fatigue interactive failure. The corresponding preset grading and evaluation benchmark table is retrieved. This benchmark table assigns differentiated scoring weights according to four types of detection indicators and divides them into clear deduction thresholds and score grading intervals. Then, the four measured and quantified data recorded in the damage evidence package—hardness change range, tissue degradation degree, oxide layer parameters, and crack size—are extracted. Each item is compared with the grading standards of each indicator in the corresponding benchmark table to complete the item deduction and scoring. The remaining total score after deducting the loss points is calculated. The corresponding basic life grade is matched according to the score interval to which the total score belongs. Then, the grade is checked and corrected in combination with the initial life grade output by the previous model. If the measured result of any detection item reaches the limit scrapping standard of full deduction for a single item, it is directly forced to be classified to the lowest life grade. Finally, the calculation is completed and the life correction grade corresponding to the target weld is output.

[0078] In one possible implementation, step S500 further includes:

[0079] The failure master control identifier includes at least one of the following: fatigue master control identifier, creep master control identifier, and creep-fatigue interactive master control identifier.

[0080] Specifically, the failure control identifier includes three independent judgment labels. Among them, the fatigue control identifier indicates that the weld damage is mainly caused by the continuous action of cyclic alternating stress due to frequent start-ups and shutdowns of the unit and load fluctuations. The alternating stress repeatedly stretches the weld matrix, gradually promoting surface slip deformation and fatigue crack initiation. Among the various test data, the degree of crack development has the highest contribution to the overall aging damage. The creep control identifier indicates that the weld is under high temperature and constant load conditions for a long time. Under the combined effect of continuous static load and high temperature environment, slow material creep aging occurs, which manifests as carbides at the microscopic level. Spheroidization, grain boundary void expansion, and microstructure deterioration-related detection indicators have the highest weight among all damage factors. The creep-fatigue interactive master control indicator represents the weld being subjected to the dual effects of long-term high-temperature static load and periodic alternating load. Grain boundary defects caused by creep will accelerate fatigue crack propagation, and the matrix damage caused by fatigue will exacerbate the accumulation of creep voids. The four damage indicators of hardness, microstructure, oxidation state, and cracks deteriorate synergistically and the damage promotes each other. This indicator can be selected and marked on the target weld according to the weight distribution of various damage data, and finally form a unique corresponding failure master control indicator.

[0081] In one possible implementation, step S600 further includes:

[0082] Step S610: When the life correction level represents the damage degree of the target weld that is greater than the damage degree corresponding to the initial life level, the initial life level is upgraded.

[0083] Step S620: When the life correction level characterizes the damage degree of the target weld which is less than or equal to the damage degree corresponding to the initial life level, the initial life level is maintained and corrected.

[0084] Specifically, the damage benchmark corresponding to the lifespan correction level is compared with that corresponding to the initial lifespan level. If the damage level corresponding to the lifespan correction level is higher than that corresponding to the initial lifespan level, a lifespan level upgrade correction operation is performed. The level quantification assignment rules are set: Level I lifespan is recorded as value 1, Level II as value 2, Level III as value 3, and Level IV as value 4. The correction calculation formula is constructed as follows: Correction level difference ΔN = Ncorrection - Ninitial, where Ncorrection represents the quantified value corresponding to the lifespan correction level and Ninitial represents the quantified value corresponding to the initial lifespan level. When the calculated ΔN > 0, the difference is the level upgrade level. The level is upgraded step by step based on the initial lifespan level according to the calculated level value. Finally, the upgraded level replaces the original initial lifespan level, completing the upgrade correction of the initial lifespan level.

[0085] Actual calculations were performed on the circumferential weld of the superheater header of a 300MW unit (the heat-affected zone was used as the verification area). The unit frequently changes load due to deep peak shaving. First, operating parameters were collected to construct a peak shaving response profile. The model initially predicted an initial lifespan level of II. Suspected failure indicators indicated a tendency for creep-fatigue interactive failure. On-site testing was initiated, and the measured data of the four tests are shown in Table 1. After generating a damage evidence package, the weld verification sample was obtained.

[0086] Table 1. Measured and scoring data of the heat-affected zone of the target weld.

[0087] Hardness change Hardness decreased by 20.5% 25 14 Organizational deterioration Globalization of carbides + sporadic grain boundary voids 25 15 Oxidative damage Oxide layer thickness 72μm 25 15 Crack distribution Network of microcracks, no through cracks 25 15

[0088] The model determined the main failure indicator to be the creep-fatigue interactive main control indicator, and based on the scoring results, the life correction level was determined to be Level III. After an upgrade correction using ΔN = 3 - 2 > 0, the final weld life correction was also Level III. The model was plotted based on measured hardness data corresponding to different peak-shaving frequencies of the unit. Figure 3 The line graph visually demonstrates that the hardness degradation of the weld gradually intensifies with the increase in peak adjustment frequency, thus proving the technical advantage of this invention in using measured data to correct the prediction level and improve the accuracy of life prediction.

[0089] A standardized lifespan grade quantification assignment rule is uniformly adopted, with weld lifespan grades I, II, III, and IV corresponding to quantified values ​​1, 2, 3, and 4 respectively, representing progressively increasing weld damage levels. A grade difference calculation formula ΔN = Ncorrected - Ninitial is constructed to conduct grade comparison and discrimination, where Ncorrected is the lifespan grade quantification value obtained after verification by the damage evidence package, and Ninitial is the lifespan grade quantification value initially predicted by the model. When the calculated grade difference ΔN ≤ 0, the actual weld damage level corresponding to the lifespan grade is determined to be less than or equal to the theoretical damage level represented by the initial lifespan grade. This indicates that the prediction result of the initial lifespan grade is consistent with or conservatively reflects the actual weld damage state, and there is no need to upgrade or downgrade the initial lifespan grade. The initial lifespan grade judgment result is directly retained and solidified, completing the maintenance and correction operation of the initial lifespan grade, ensuring that the lifespan judgment result is stable and consistent with the actual working condition damage situation.

[0090] Example 2, based on the same inventive concept as the machine learning-based weld creep-fatigue failure prediction method in the foregoing examples, such as... Figure 2 As shown, this application provides a machine learning-based weld creep-fatigue failure prediction system. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0091] The peak shaving response profile acquisition module 10 is used to acquire the peak shaving response profile of the target weld under deep peak shaving conditions. The peak shaving response profile includes weld area identification, temperature fluctuation data, pressure fluctuation data, load holding data, and creep-fatigue test life data.

[0092] The suspected failure identification module 20 is used to input the peak-shaving response profile into the weld failure prediction model to obtain the initial life level and suspected failure identification of the target weld.

[0093] The damage evidence package acquisition module 30 is used to determine the damage verification area of ​​the target weld based on the weld area identifier when the suspected failure identifier meets the verification conditions, and to acquire the damage evidence package of the damage verification area.

[0094] The weld seam verification sample generation module 40 is used to associate the damage evidence package with the peak-shaving response profile to generate a weld seam verification sample.

[0095] The lifetime correction level acquisition module 50 is used to input the weld review sample into the weld failure prediction model to obtain the failure master control identifier and lifetime correction level of the target weld.

[0096] The prediction result output module 60 is used to correct the initial life level according to the life correction level and output the creep-fatigue failure prediction result of the target weld.

[0097] Furthermore, the system is also used to implement the following functions:

[0098] Obtain the weld area identifier of the target weld, wherein the weld area identifier is one of the base metal area identifier, weld metal area identifier, and heat-affected zone identifier; obtain temperature fluctuation data, pressure fluctuation data, and load holding data of the target weld during the deep peak shaving process; obtain the creep rupture test life, low-cycle fatigue test life, and creep-fatigue interaction test life corresponding to the target weld; generate the peak shaving response profile based on the weld area identifier, temperature fluctuation data, pressure fluctuation data, load holding data, creep rupture test life, low-cycle fatigue test life, and creep-fatigue interaction test life.

[0099] Furthermore, the system is also used to implement the following functions:

[0100] Obtain multiple sample peak-shaving response profiles, multiple sample lifetime levels, and multiple sample failure indicators for multiple sample welds; train a machine learning model based on the multiple sample peak-shaving response profiles, multiple sample lifetime levels, and multiple sample failure indicators to obtain the weld failure prediction model; input the peak-shaving response profile of the target weld into the weld failure prediction model, and output the initial lifetime level and suspected failure indicator of the target weld.

[0101] Furthermore, the system is also used to implement the following functions:

[0102] When the suspected failure indicator indicates that the target weld has a tendency for creep-fatigue interactive failure, it is determined that the review condition is met; according to the weld area indicator corresponding to the suspected failure indicator, the damage review area is determined from the base metal area, weld metal area and heat-affected zone.

[0103] Furthermore, the system is also used to implement the following functions:

[0104] Hardness testing is performed on the damage verification area to obtain hardness change data; tissue testing is performed on the damage verification area to obtain tissue degradation data; oxidation state testing is performed on the damage verification area to obtain oxidation damage data; crack testing is performed on the damage verification area to obtain crack distribution data; and the damage evidence package is generated based on the hardness change data, tissue degradation data, oxidation damage data, and crack distribution data.

[0105] Furthermore, the system is also used to implement the following functions:

[0106] Extract the damage verification area corresponding to the damage evidence package; extract the temperature fluctuation data, pressure fluctuation data, load holding data, and creep-fatigue test life data corresponding to the damage verification area from the peak-shaving response profile; bind the damage evidence package with the data corresponding to the damage verification area to generate the weld verification sample.

[0107] Furthermore, the system is also used to implement the following functions:

[0108] The weld seam verification sample is input into the weld seam failure prediction model; the weld seam failure prediction model identifies the failure master control identifier of the target weld seam based on the correspondence between the peak-shaving response profile and the damage evidence package; and outputs the life correction level of the target weld seam based on the failure master control identifier and the damage evidence package.

[0109] Furthermore, the system is also used to implement the following functions:

[0110] The failure master control identifier includes at least one of the following: fatigue master control identifier, creep master control identifier, and creep-fatigue interactive master control identifier.

[0111] Furthermore, the system is also used to implement the following functions:

[0112] When the lifetime correction level indicates that the damage level of the target weld is greater than the damage level corresponding to the initial lifetime level, the initial lifetime level is upgraded; when the lifetime correction level indicates that the damage level of the target weld is less than or equal to the damage level corresponding to the initial lifetime level, the initial lifetime level is maintained.

[0113] It should be noted that the order of the embodiments described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. Specific embodiments of this specification have been described above. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0114] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0115] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A machine learning-based method for predicting weld creep-fatigue failure, characterized in that, The method includes: Obtain a peak-shaving response profile of the target weld under deep peak-shaving conditions, wherein the peak-shaving response profile includes weld area identification, temperature fluctuation data, pressure fluctuation data, load holding data, and creep-fatigue test life data; The peak-shaving response profile is input into the weld failure prediction model to obtain the initial life level and suspected failure indicator of the target weld. When the suspected failure indicator meets the verification conditions, the damage verification area of ​​the target weld is determined according to the weld area indicator, and the damage evidence package of the damage verification area is collected. The damage evidence package is associated with the peak-shaving response profile to generate a weld re-verification sample; The weld review sample is input into the weld failure prediction model to obtain the failure control identifier and life correction level of the target weld. The initial life level is corrected according to the life correction level, and the creep-fatigue failure prediction result of the target weld is output.

2. The machine learning-based weld creep-fatigue failure prediction method as described in claim 1, characterized in that, Obtain a peak-shaving response profile of the target weld under deep peak-shaving conditions, including: Obtain the weld area identifier of the target weld, wherein the weld area identifier is one of the base metal area identifier, weld metal area identifier, and heat-affected zone identifier; Acquire temperature fluctuation data, pressure fluctuation data, and load data of the target weld during the depth peak shaving process; Obtain the creep rupture test life, low-cycle fatigue test life, and creep-fatigue interaction test life corresponding to the target weld. The peak-shaving response profile is generated based on the weld area identification, temperature fluctuation data, pressure fluctuation data, load holding data, creep rupture test life, low-cycle fatigue test life, and creep-fatigue interaction test life.

3. The machine learning-based weld creep-fatigue failure prediction method as described in claim 1, characterized in that, The peak-shaving response profile is input into the weld failure prediction model to obtain the initial life level and suspected failure indicators of the target weld, including: Obtain multiple sample peak-shaving response profiles, multiple sample lifetime levels, and multiple sample failure indicators for multiple sample welds; The weld failure prediction model is obtained by training a machine learning model based on the multiple sample peak-shaving response profiles, multiple sample lifetime levels, and multiple sample failure identifiers. Input the peak response profile of the target weld into the weld failure prediction model, and output the initial life level and suspected failure indicator of the target weld.

4. The machine learning-based weld creep-fatigue failure prediction method as described in claim 1, characterized in that, When the suspected failure indicator meets the review conditions, the damage review area of ​​the target weld is determined based on the weld area indicator, including: When the suspected failure indicator indicates that the target weld has a tendency to exhibit creep-fatigue interactive failure, the review condition is determined to be met. Based on the weld area identifier corresponding to the suspected failure identifier, the damage review area is determined from the base metal area, weld metal area, and heat-affected zone.

5. The machine learning-based weld creep-fatigue failure prediction method as described in claim 1, characterized in that, The damage evidence package collected for the damage verification area includes: Hardness testing was performed on the damaged area to obtain hardness change data; Tissue analysis was performed on the damaged area to obtain tissue degradation data; Oxidation state detection was performed on the damage verification area to obtain oxidation damage data; Crack detection was performed on the damage verification area to obtain crack distribution data; The damage evidence package is generated based on the hardness change data, microstructure deterioration data, oxidative damage data, and crack distribution data.

6. The machine learning-based weld creep-fatigue failure prediction method as described in claim 1, characterized in that, The damage evidence package is correlated with the peak-shaving response profile to generate a weld re-examination sample, including: Extract the damage verification area corresponding to the damage evidence package; Extract temperature fluctuation data, pressure fluctuation data, load holding data, and creep-fatigue test life data corresponding to the damage verification area from the peak-shaving response profile; The damage evidence package is bound to the data corresponding to the damage review area to generate the weld review sample.

7. The machine learning-based weld creep-fatigue failure prediction method as described in claim 1, characterized in that, The weld seam verification sample is input into the weld seam failure prediction model to obtain the main failure control identifier and life correction level of the target weld seam, including: Input the weld seam verification sample into the weld seam failure prediction model; The weld failure prediction model identifies the main failure controller of the target weld based on the correspondence between the peak-shaving response profile and the damage evidence package. Based on the failure master control identifier and damage evidence package, the life correction level of the target weld is output.

8. The machine learning-based weld creep-fatigue failure prediction method as described in claim 7, characterized in that, The failure master control identifier includes at least one of the following: fatigue master control identifier, creep master control identifier, and creep-fatigue interactive master control identifier.

9. The machine learning-based weld creep-fatigue failure prediction method as described in claim 1, characterized in that, The initial lifetime level is corrected according to the lifetime correction level, including: When the life correction level represents a greater degree of damage to the target weld than the degree of damage corresponding to the initial life level, the initial life level is upgraded. When the life correction level, which characterizes the damage level of the target weld, is less than or equal to the damage level corresponding to the initial life level, the initial life level is maintained and corrected.

10. A machine learning-based weld creep-fatigue failure prediction system, characterized in that, The system is used to implement the machine learning-based weld creep-fatigue failure prediction method according to any one of claims 1-9, the system comprising: The peak shaving response profile acquisition module is used to acquire the peak shaving response profile of the target weld under deep peak shaving conditions. The peak shaving response profile includes weld area identification, temperature fluctuation data, pressure fluctuation data, load holding data and creep-fatigue test life data. The suspected failure identifier acquisition module is used to input the peak-shaving response profile into the weld failure prediction model to obtain the initial life level and suspected failure identifier of the target weld. The damage evidence package acquisition module is used to determine the damage verification area of ​​the target weld based on the weld area identifier when the suspected failure identifier meets the verification conditions, and to acquire the damage evidence package of the damage verification area. The weld seam verification sample generation module is used to associate the damage evidence package with the peak-shaving response profile to generate a weld seam verification sample. The lifetime correction level acquisition module is used to input the weld review sample into the weld failure prediction model to obtain the failure control identifier and lifetime correction level of the target weld. The prediction result output module is used to correct the initial life level according to the life correction level and output the creep-fatigue failure prediction result of the target weld.

Images