Prediction method of adsorber weld fatigue life under hydrogen-rich alternating load based on digital twin

Abstract

The present invention belongs to the field of welding, and specifically discloses a method for predicting fatigue life of adsorber welds under hydrogen-rich alternating loads based on digital twins. The method obtains the hardness, defects, and residual stress distribution of weld joints through non-destructive testing; conducts high-cycle fatigue performance tests in a pure hydrogen atmosphere to establish a performance database; utilizes deep generative adversarial networks (GANs) to enhance data, and combines support vector regression (SVR) to establish a machine learning model to predict weld fatigue life by considering material strength degradation, residual stress, and defect characteristics; uses multiple indicators to evaluate the performance of the weld fatigue life model, and conducts correlation analysis to improve prediction accuracy. The corresponding steps include non-destructive testing of adsorber welds, weld fatigue performance testing, and high-cycle fatigue performance testing of defective weld joints. The present invention combines digital twin technology and machine learning to improve the accuracy and reliability of weld fatigue life prediction, which is of great significance for optimizing equipment maintenance and improving operational efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of welding, and specifically discloses a method for predicting the fatigue life of adsorber welds under hydrogen-rich alternating loads based on digital twins. Background Art

[0002] As core equipment in key fields such as coal chemical industry and petroleum, the performance stability and health of pressure swing adsorbers are crucial to ensuring the continuity and safety of the production process. Pressure swing adsorption technology achieves the effective separation of single gases from mixed gases through the principles of high-pressure adsorption and low-pressure desorption. However, in actual operation, because the equipment must withstand complex working conditions such as high temperature, high pressure and alternating loads, the weld area often becomes a weak link in equipment safety. Especially in hydrogen-rich environments, welds may suffer from serious failure problems such as cracking and bulging due to the combined effects of multiple factors such as alternating stress, mechanical vibration, temperature changes and hydrogen corrosion.

[0003] Currently, the primary method for monitoring the safety of pressure swing adsorption (PSA) welds is regular spot checks. This method is not only time-consuming and costly, but also lacks intelligence, making it difficult to monitor and accurately predict the equipment's status in real time. Therefore, a scientific, accurate, and efficient weld fatigue life prediction method is urgently needed to improve the operating efficiency of PSA units, ensure production safety, and reduce downtime and production costs caused by equipment failures. Summary of the Invention

[0004] In order to solve the above problems, the present invention discloses a method and device for predicting the fatigue life of adsorber welds under hydrogen-rich alternating loads based on digital twins. This method combines multidisciplinary technical means such as non-destructive testing, high-cycle fatigue performance testing, defect simulation and testing, database construction, deep learning adversarial network (GAN) data enhancement, support vector regression (SVR) machine learning model and digital twin technology, aiming to achieve accurate prediction and scientific management of the fatigue life of pressure swing adsorber welds. Through the present invention, the accuracy and reliability of weld fatigue life prediction can be significantly improved, providing a scientific basis for equipment maintenance plans and improvement measures, thereby improving the overall operating efficiency and safety of the pressure swing adsorption device.

[0005] The present invention includes the following technical solutions:

[0006] A method for predicting fatigue life of adsorber welds under hydrogen-rich alternating loads based on digital twins, characterized by comprising the following steps:

[0007] (a) Perform nondestructive testing (NDT) on the pressure swing adsorber, including ultrasonic testing, magnetic particle testing, and X-ray testing, to obtain the spatial distribution of defects and residual stresses in the heat-affected zone (HAZ) and weld seams of the adsorber weld joints. Defects include geometric features such as pores, microcracks, undercuts, and reinforcements. The hardness of the HAZ and weld seams shall be recorded.

[0008] (b) Perform high-cycle fatigue tests on welded joints of adsorber scaled-down components made of Q345R steel or other steel grades in a pure hydrogen atmosphere to obtain SN curves of the heat-affected zone and welds. The maximum stress range of the test is 200-400 MPa, and the stress ratio is -1 to 0.8.

[0009] (c) Prepare defective Q345R steel or other steel adsorber welded scale parts, count the number and size of defects, and obtain the SN curve of the defective weld joint in a pure hydrogen atmosphere;

[0010] (d) establishing a high cycle fatigue performance database of the adsorber weld joint weld and heat affected zone in a hydrogen atmosphere based on the SN curves and defect feature distribution data obtained in steps (a)-(c);

[0011] (e) Data augmentation is performed based on the adversarial network GAN, using the generator network Generator,G and the discriminator network Discriminator,D to generate data consistent with the real data distribution to obtain a sufficient number of labeled samples;

[0012] (f) using the data set in step (d) and the data generated in step (e), establishing an adaptive machine learning model based on support vector regression (SVR), wherein the input variables include stress and defect characteristics, and the output variable is the corresponding fatigue life; the model uses the fatigue crack growth rate formula da / dN f =C(ΔK) m To evaluate the fatigue crack growth life, where C and m are material constants; ΔK is the stress intensity factor, and ΔK is given by the formula Calculated; A is the projected defect area on the plane perpendicular to the maximum principal stress; Y is the position constant, which is 0.5 for internal defects and 0.65 for surface and subsurface defects; Δσ is the applied stress range;

[0013] (g) Using the coefficient of determination R 2, mean absolute percentage error (MAPE) and root mean square error (RMSE) are used to evaluate the performance of the established machine learning model to ensure prediction accuracy; wherein, the adaptive machine learning model in step (f) also includes correlation analysis between input variables, and the Pearson correlation coefficient is used to evaluate the multicollinearity between input variables to improve the prediction accuracy of the model; the predicted fatigue strength σw is calculated using the formula σw=Y1×(HV / σb)×Δσ; Y1 is the position constant, internal defects are 1.56, and surface and subsurface defects are 1.43; HV is the Vickers hardness, and σb is the tensile strength.

[0014] In the above step (a), first, non-destructive testing is performed on the pressure swing adsorber, including ultrasonic testing, magnetic particle testing, and X-ray testing, to obtain the spatial distribution of defects (such as pores, microcracks, undercuts, and reinforcement) in the heat-affected zone and weld of the weld joint, and the residual stress, and hardness information is recorded to provide basic data for subsequent analysis; (b) Then, high-cycle fatigue performance testing is performed on the weld joint of the Q345R steel or other steel adsorber scale part in a pure hydrogen atmosphere to obtain the SN curve of the heat-affected zone and weld, which reflects the fatigue performance of the material within a specific stress range. (c) At the same time, a Q345R steel or other steel adsorber scale part with defects is prepared and its SN curve is tested in a pure hydrogen atmosphere to consider the impact of defects on fatigue performance. (d) Based on the data obtained in steps (a) and (c), a high-cycle fatigue performance database of adsorber weld joints in a hydrogen atmosphere is established to provide rich training data for the machine learning model. (e) Then, a deep generative adversarial network (GAN) is used for data enhancement to generate data consistent with the real data distribution, so as to expand the sample size and improve the generalization ability of the model. (f) Subsequently, an adaptive machine learning model based on support vector regression (SVR) is established. The input variables include stress and defect characteristics, and the output variable is the corresponding fatigue life. The model uses the fatigue crack growth rate formula to evaluate the fatigue crack growth life, and considers the correlation analysis between the input variables. The Pearson correlation coefficient is used to evaluate the multicollinearity between the input variables to improve the prediction accuracy of the model. (g) Finally, the determination coefficient R is used to evaluate the multicollinearity between the input variables. 2 , mean absolute percentage error (MAPE) and root mean square error (RMSE) are used to evaluate the performance of the established machine learning model to ensure the accuracy and reliability of the prediction results.

[0015] Furthermore, in the above-mentioned digital twin-based method for predicting the fatigue life of adsorber welds under hydrogen-rich alternating loads, step (a) also includes residual stress distribution detection along the parent material-heat-affected zone-weld direction distribution curve, and recording the axial and circumferential residual stress values ​​of the weld.

[0016] Furthermore, in the above-mentioned digital twin-based method for predicting fatigue life of adsorber welds under hydrogen-rich alternating loads, the defect type and size distribution of the defective Q345R steel or other steel adsorber welded scale parts in step (c) match the defects actually detected in step (a).

[0017] Furthermore, in the above-mentioned digital twin-based method for predicting the fatigue life of adsorber welds under hydrogen-rich alternating loads, the GAN data enhancement process in step (e) includes: the generator G collects random variables z from the Gaussian distribution, obtains generated data G(z) after nonlinear transformation, and inputs the real data x and the generated data G(z) into the discriminator D at the same time. The discriminator D judges the authenticity of the data by calculating the probability that the input data comes from the real data x, thereby training the GAN model until the generator G can generate samples that are sufficiently close to the real data.

[0018] Furthermore, the above-mentioned digital twin-based method for predicting the fatigue life of adsorber welds under hydrogen-rich alternating loads also includes the application of the prediction results, that is, formulating maintenance plans and improvement measures for the pressure swing adsorber based on the predicted weld fatigue life to improve the operating efficiency and safety of the equipment.

[0019] Furthermore, the above-mentioned digital twin-based method for predicting the fatigue life of adsorber welds under hydrogen-rich alternating loads uses digital twin technology to digitally map, monitor, diagnose, predict, simulate and optimize the pressure swing adsorber, thereby realizing data management and equipment performance optimization throughout the entire life cycle.

[0020] Furthermore, in the above-mentioned digital twin-based method for predicting the fatigue life of adsorber welds under hydrogen-rich alternating loads, the machine learning model in step (f) also uses a data set of fatigue damage SN curves and defect characteristics, and combines deep generative adversarial networks (GANs) for data enhancement to obtain labeled samples for training the machine learning model to improve the accuracy and reliability of fatigue life prediction.

[0021] Furthermore, the above-mentioned digital twin-based prediction method for adsorber weld fatigue life under hydrogen-rich alternating load uses the Pearson correlation coefficient to perform correlation analysis on input variables when training the machine learning model to remove multicollinearity and improve the prediction performance of the model.

[0022] Furthermore, the above-mentioned digital twin-based method for predicting the fatigue life of adsorber welds under hydrogen-rich alternating loads also includes comparing the prediction results of the machine learning model with the actual situation, and optimizing the model parameters through continuous iteration to improve the accuracy and reliability of fatigue life prediction.

[0023] The present invention also discloses a digital twin-based adsorber weld fatigue life prediction system under hydrogen-rich alternating loads, which is characterized by comprising:

[0024] 1) Nondestructive testing module: used to perform nondestructive testing on pressure swing adsorbers, including ultrasonic testing, magnetic particle testing, and X-ray testing, to obtain defects and residual stress spatial distribution in the heat-affected zone and weld seam of adsorber weld joints. Defects include geometric features such as pores, microcracks, undercuts, and reinforcements.

[0025] 2) Fatigue performance test module: used to perform high-cycle fatigue performance tests on welded joints of adsorber scaled parts made of Q345R steel or other steel grades in a pure hydrogen atmosphere, and obtain the SN curves of the heat-affected zone and welds;

[0026] 3) Defect simulation and testing module: used to prepare defective Q345R steel or other steel adsorber welded scale parts, count the number and size of defects, and obtain the SN curve of defective weld joints in a pure hydrogen atmosphere;

[0027] 4) Database construction module: used to establish a high-cycle fatigue performance database of adsorber welded joints in a hydrogen atmosphere based on the SN curves and defect feature distribution data obtained by the nondestructive testing module and the defect simulation and testing module;

[0028] 5) Data Processing and Model Training Module: This module uses a deep generative adversarial network (GAN) for data augmentation, utilizing a generator network and a discriminator network to generate data consistent with the real-world data distribution. Furthermore, the module utilizes the dataset in the database construction module and the GAN-generated data to establish an adaptive machine learning model based on support vector regression (SVR) for predicting weld fatigue life. The model also utilizes the fatigue crack growth rate formula and fatigue damage SN curve for evaluation.

[0029] 6) Model evaluation and optimization module: used to use the determination coefficient R 2 , mean absolute percentage error (MAPE) and root mean square error (RMSE) are used to evaluate the performance of the machine learning model, and the prediction accuracy is improved by iteratively optimizing the model parameters. The module also uses the Pearson correlation coefficient to perform correlation analysis on the input variables to remove multicollinearity.

[0030] 7) Prediction result application module: used to formulate maintenance plans and improvement measures for the pressure swing adsorber based on the prediction results of the machine learning model to improve the operating efficiency and safety of the equipment;

[0031] 8) Digital Twin Module: Used to digitally map, monitor, diagnose, predict, simulate, and optimize pressure swing adsorbers, enabling full lifecycle data management and equipment performance optimization.

[0032] Compared with the prior art, the present invention has the following outstanding beneficial effects:

[0033] 1. Improved prediction accuracy: By combining multidisciplinary techniques such as nondestructive testing, high-cycle fatigue testing, and defect simulation and testing, this method can comprehensively and accurately obtain the spatial distribution of defects and residual stresses in adsorber weld joints, as well as their fatigue performance data in a hydrogen atmosphere. This data provides a solid foundation for building precise machine learning models, significantly improving the accuracy and reliability of weld fatigue life predictions.

[0034] 2. Intelligent Management: This invention utilizes deep generative adversarial networks (GANs) for data augmentation, expanding the sample size and improving the model's generalization capabilities. Furthermore, an adaptive machine learning model, built through support vector regression (SVR), enables intelligent prediction of weld fatigue life. This facilitates scientific management of pressure swing adsorbers, reduces manual intervention, and improves management efficiency.

[0035] 3. Optimize maintenance plans: Based on the predicted weld fatigue life, the present invention can develop targeted maintenance plans and improvement measures. This helps to identify potential safety hazards in advance and take timely measures for repair or replacement, thus avoiding downtime and production costs caused by equipment failure.

[0036] 4. Improved Equipment Performance: Using digital twin technology, this invention digitally maps, monitors, diagnoses, predicts, simulates, and optimizes the pressure swing adsorber. This helps provide a comprehensive understanding of the equipment's operating status and performance, enabling timely identification and resolution of issues, thereby improving overall operational efficiency and safety.

[0037] 5. Reduced Inspection Costs: Compared to traditional periodic spot checks, this method uses machine learning models for prediction, reducing the number and cost of actual inspections. Furthermore, the high accuracy of the predictions can more effectively guide equipment maintenance and repairs, further reducing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a diagram of the residual stress detection position in Example 3;

[0039] Figure 2 is the residual stress distribution diagram in Example 3;

[0040] Figure 3 Schematic diagram of the GAN model in Example 3;

[0041] Figure 4 Specific formula diagrams for the generator network Generator, G and the discriminator network Discriminator, D in Example 3;

[0042] Figure 5 This is a graph showing the thermal fatigue performance test data of a defective unit that has not been put into service in Example 3;

[0043] Figure 6 This is a graph showing fatigue performance testing data of defective welds in service in Example 3;

[0044] Figure 7 The artificial neural network of the SVR-based adaptive machine learning model in Example 3;

[0045] Figure 8 This is the fatigue life prediction diagram of the heat-affected zone with defects that has not been put into service in Example 3;

[0046] Figure 9 This is the fatigue life prediction diagram of the unserviced weld in Example 3;

[0047] Figure 10 The figure is a flow chart of the overall method of the present invention. DETAILED DESCRIPTION

[0048] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0049] Example 1

[0050] Fatigue life prediction of adsorber welds under hydrogen-rich alternating loads based on digital twin.

[0051] Step (a) Nondestructive testing:

[0052] Perform non-destructive testing on Q345R steel pressure swing adsorbers, including ultrasonic testing, magnetic particle testing, and X-ray testing.

[0053] The geometric characteristics of defects (such as pores, microcracks, undercuts, and reinforcement) are recorded, and the residual stress distribution is detected along the direction of base material-heat-affected zone-weld.

[0054] Record the axial and circumferential residual stress values ​​of the weld, as well as the hardness of the heat-affected zone and weld.

[0055] Step (b) High cycle fatigue performance test:

[0056] In pure hydrogen atmosphere, high cycle fatigue test is carried out on welded joints of adsorber scaled parts made of Q345R steel or other steel grades.

[0057] Test stress range: 200-400MPa, stress ratio: -1~0.8.

[0058] Obtain SN curves of heat-affected zone and weld.

[0059] Step (c) Testing of defective welded joints:

[0060] A Q345R steel adsorber welded scale model with defects was prepared, and the defect types and size distributions matched the actual detection data envelope.

[0061] The SN curves of the weld and heat-affected zone of welded joints with different defect characteristics were tested in a pure hydrogen atmosphere.

[0062] Step (d) Database establishment:

[0063] The data from steps (a)-(c) are integrated to establish a high cycle fatigue performance database of adsorber welded joints in a hydrogen atmosphere.

[0064] Step (e) GAN data enhancement:

[0065] Generative Adversarial Network (GAN) is used for data enhancement.

[0066] The generator G collects random variables z from the Gaussian distribution and obtains the generated data G(z) through nonlinear transformation.

[0067] The discriminator D determines the authenticity of the data by calculating the probability that the input data comes from the real data.

[0068] The GAN model is trained iteratively until the generated data is close enough to the real data.

[0069] Step (f) Machine Learning Model Construction:

[0070] An adaptive machine learning model is established based on support vector regression (SVR).

[0071] Input variables: stress, defect characteristics (such as defect size, number, location), and residual stress value.

[0072] Output variable: fatigue life.

[0073] The fatigue crack growth life is estimated using the fatigue crack growth rate formula:

[0074] Formula: da / dN f =C(ΔK) m Where C and m are material constants; ΔK is the stress intensity factor.

[0075] ΔK is calculated by the formula It is calculated that A is the projection of the defect area on the vertical plane of the maximum principal stress, Y is the position constant (0.5 for internal defects and 0.65 for surface and subsurface defects), and Δσ is the applied stress range.

[0076] Step (g) Model evaluation and application:

[0077] The coefficient of determination R 2 , mean absolute percentage error (MAPE) and root mean square error (RMSE) were used to evaluate model performance.

[0078] The Pearson correlation coefficient was used to analyze the correlation between input variables and eliminate multicollinearity.

[0079] Develop maintenance plans based on prediction results to improve equipment operating efficiency and safety.

[0080] Example 2

[0081] Deepening application of digital twin technology in adsorber weld fatigue life prediction

[0082] Steps (a) to (f):

[0083] Nondestructive testing, performance testing, database establishment, data enhancement, and machine learning model construction were performed according to steps (a) to (f) of Example 1.

[0084] Step (e) Further optimization of GAN data enhancement:

[0085] Conditional GAN ​​is introduced to control specific properties of generated data.

[0086] Conditional variables (e.g., stress level, defect type) are added to the input of the generator G and the discriminator D to guide the data generation process.

[0087] Step (f) Deepening of the machine learning model:

[0088] Based on the SVR model, integrated learning methods (such as Bagging and Boosting) are introduced to improve prediction stability and accuracy.

[0089] Cross-validation techniques were used to evaluate model performance and select the optimal model parameters.

[0090] Step (g) Model verification and optimization:

[0091] Compare the prediction results of the machine learning model with the actual situation, and improve the prediction accuracy by iteratively optimizing the model parameters.

[0092] Use strategies such as grid search or random search to find the optimal hyperparameter combination.

[0093] Deepening application of digital twin technology:

[0094] A digital twin model of the pressure swing adsorber was established to monitor the equipment's operating status and weld fatigue in real time. When the weld fatigue life is predicted to be approaching a critical value, an early warning mechanism is triggered and maintenance recommendations are automatically generated. The digital twin model is used to simulate and optimize equipment performance, guiding equipment design and improvement.

[0095] Example 3

[0096] This embodiment provides a practical application of a method for predicting fatigue life of adsorber welds under hydrogen-rich alternating loads based on digital twins. Figure 1-10 The embodiments of the present invention are described in detail.

[0097] The overall process flow of the present invention is as follows Figure 10 shown.

[0098] Step (a): Nondestructive testing

[0099] First, the pressure swing adsorber is subjected to non-destructive testing, including ultrasonic testing, magnetic particle testing and X-ray testing. Figure 1 Residual stress detection is performed at the position shown, and the defects and residual stress spatial distribution of the heat-affected zone and weld in the weld joint are obtained. Defects mainly include geometric features such as pores, microcracks, undercuts and reinforcements, and the hardness of the heat-affected zone and weld is recorded. In addition, the residual stress distribution test is also performed along the parent material-heat-affected zone-weld direction distribution curve, and the residual stress values ​​in the axial and circumferential directions of the weld are recorded, such as Figure 2 shown.

[0100] Step (b): High cycle fatigue performance test

[0101] High-cycle fatigue testing was performed on welded joints of Q345R steel adsorber scale components in a pure hydrogen atmosphere. The maximum stress range was 200-400 MPa, and the stress ratio range was -1 to 0.8. The SN curves for the heat-affected zone and weld were obtained, reflecting the fatigue performance of the material within a specific stress range.

[0102] Step (c): Defect simulation and testing

[0103] Prepare a Q345R steel adsorber welded scale model with defects. The defect types and size distribution match the defects actually detected in step (a). Count the number and size of defects, and obtain a S / N curve for the defective weld joint under a pure hydrogen atmosphere. Figure 5 and Figure 6 The heat impact diagram of the unserviced defective weld and the weld diagram of the service defective weld are shown respectively, and the formula Where Y is the position constant, which is 0.5 for internal defects and 0.65 for surface and subsurface defects; A is the projected defect area on the plane perpendicular to the maximum principal stress; Δσ is the applied stress range, and the specific value is set according to the test conditions.

[0104] Step (d): Database construction

[0105] Based on the SN curves and defect feature distribution data obtained in steps (a)-(c), a high-cycle fatigue performance database for adsorber welded joints in a hydrogen atmosphere was established. This database provided rich data support for subsequent machine learning model training.

[0106] Step (e): Data augmentation

[0107] Data enhancement based on deep generative adversarial network GAN. Figure 3 As shown, the GAN model consists of a generator network, Generator G, and a discriminator network, Discriminator D. Generator G samples a random variable z from a Gaussian distribution and, after nonlinear transformation, generates generated data G(z). Real data x and generated data G(z) are simultaneously fed into the discriminator D. Discriminator D determines the authenticity of the data by calculating the probability that the input data is derived from the real data x. This process trains the GAN model until Generator G can generate samples that are sufficiently close to real data. Figure 4 The specific formula diagrams of the generator network Generator, G and the discriminator network Discriminator, D are shown.

[0108] Step (f): Machine Learning Model Building

[0109] Using the data set in step (d) and the data generated in step (e), an adaptive machine learning model based on support vector regression (SVR) is established. The input variables include stress and defect characteristics, and the output variable is the corresponding fatigue life. The model uses the fatigue crack growth rate formula da / dN f =C(ΔK) m To evaluate the fatigue crack growth life, C and m are material constants, determined according to experimental data. ΔK is the stress intensity factor, which is given by the formula The model also considers the correlation analysis between input variables and uses the Pearson correlation coefficient to evaluate the multicollinearity between input variables to improve the prediction accuracy of the model. Figure 7 A schematic diagram of establishing an artificial neural network for an adaptive machine learning model based on SVR is shown.

[0110] Step (g): Model evaluation and optimization

[0111] The coefficient of determination R 2The performance of the established machine learning model is evaluated using the mean absolute percentage error (MAPE) and the root mean square error (RMSE) to ensure prediction accuracy. At the same time, the prediction results of the machine learning model are compared with the actual situation, and the model parameters are optimized through continuous iteration to improve the accuracy and reliability of fatigue life prediction.

[0112] Application of prediction results

[0113] Based on the predicted weld fatigue life, maintenance plans and improvement measures can be developed for pressure swing adsorbers to improve equipment efficiency and safety. Specifically, the prediction results can be used to manage welds in a hierarchical manner, prioritizing the repair or replacement of welds with shorter predicted lifespans.

[0114] Digital Twin Technology

[0115] This invention also uses digital twin technology to digitally map, monitor, diagnose, predict, simulate, and optimize the pressure swing adsorber, enabling full lifecycle data management and equipment performance optimization. Digital twin technology can reflect the operating status and performance of the equipment in real time, providing strong support for equipment maintenance and management.

[0116] Specific examples:

[0117] Taking a certain model of pressure swing adsorber as an example, the weld fatigue life prediction was performed according to the above steps. Nondestructive testing was used to obtain spatial distribution data of defects and residual stresses in the heat-affected zone and weld of the welded joint. High-cycle fatigue performance testing and defect simulation and testing were performed to establish a high-cycle fatigue performance database. After data enhancement using GAN, an SVR machine learning model was established. The model was then evaluated and optimized to obtain prediction results. Figure 8 and Figure 9 The predicted fatigue life diagrams of the unserviced defective heat-affected zone and weld are shown separately.

[0118] Verification and Comparison with Measured Data: The applicant compared the predicted results with fatigue life data from welds of pressure swing adsorbers in service. Through long-term monitoring and recording, we found that the error rate between the predicted and actual lifespans was extremely low, validating the high accuracy of the prediction model. In developing the SVR machine learning model, we conducted multiple iterations, continuously adjusting model parameters and implementing cross-validation to reduce the risk of overfitting. The resulting model demonstrated excellent predictive performance on both the training and validation sets, ensuring the stability and reliability of the prediction results.

[0119] For comparison, we also used other commonly used fatigue life prediction methods, such as those based on fracture mechanics and empirical formulas, to predict the same set of weld data. This comparative analysis showed that our method demonstrated higher prediction accuracy and applicability, especially when dealing with complex defects and residual stress distributions.

[0120] In summary, as described in Example 3, the present invention achieves accurate prediction and scientific management of the fatigue life of the pressure swing adsorber weld by combining multidisciplinary technical means such as non-destructive testing, high-cycle fatigue performance testing, defect simulation and testing, database construction, deep generative adversarial network (GAN) data enhancement, support vector regression (SVR) machine learning model, and digital twin technology.

[0121] Example 4

[0122] Fatigue life prediction system for adsorber welds under hydrogen-rich alternating loads based on digital twin

[0123] 1. System Overview

[0124] This system aims to build a comprehensive, digital twin-based system for predicting adsorber weld fatigue life under hydrogen-rich alternating loads. By integrating modules such as nondestructive testing, fatigue performance testing, defect simulation and testing, database construction, data processing and model training, model evaluation and optimization, application of prediction results, and digital twins, and combining the necessary hardware and technical support, the system enables accurate prediction of pressure swing adsorber weld fatigue life and full lifecycle management.

[0125] 2. System composition

[0126] Hardware device layer

[0127] Servers and storage devices: Provide high-performance computing and large-capacity storage to support data processing, model training, and the operation of digital twin models.

[0128] Sensors and data acquisition equipment: Real-time monitoring of the operating status of the pressure swing adsorber, including parameters such as temperature, pressure, and vibration, providing data support for digital twins and digital operations and maintenance.

[0129] Network communication equipment: ensures unimpeded data communication between modules within the system and supports remote access and collaboration.

[0130] Professional testing equipment: including ultrasonic flaw detectors, magnetic particle flaw detectors, X-ray testing equipment, etc., used for non-destructive testing of weld defects and residual stress distribution.

[0131] Fatigue testing machine: used to test the fatigue performance of welded joints of Q345R steel adsorber scale parts and obtain key data such as SN curves.

[0132] Software and technology layer

[0133] Database management system: stores and manages non-destructive testing data, fatigue performance test data, defect simulation data, digital twin models and other data to ensure data security and reliability.

[0134] Data analysis and processing software: Processes and analyzes the collected data, including defect identification, residual stress calculation, fatigue life prediction, etc., to provide data support for model training and prediction.

[0135] Machine learning framework and tools: A machine learning model is established based on algorithms such as support vector regression (SVR) to predict weld fatigue life, and data enhancement is performed through deep generative adversarial networks (GANs) to improve model generalization capabilities.

[0136] Digital twin platform: Builds a digital model of the pressure swing adsorber to implement digital mapping, monitoring, diagnosis, prediction, simulation, and optimization functions, providing technical support for full life cycle management.

[0137] Security protection software: ensures the security and stability of the system, prevents data leakage and illegal access, including firewalls, antivirus software, etc.

[0138] Functional module layer

[0139] Non-destructive testing module: Use professional testing equipment to perform non-destructive testing on welds to obtain defect and residual stress distribution data.

[0140] Fatigue performance test module: Perform high-cycle fatigue performance tests on the welded joints of Q345R steel adsorber scale parts in a pure hydrogen atmosphere to obtain SN curves.

[0141] Defect simulation and testing module: Prepare defective welded scale parts, count the number and size of defects, and obtain the SN curve of defective weld joints.

[0142] Database construction module: Integrate data from non-destructive testing, fatigue performance testing, and defect simulation and testing to establish a high-cycle fatigue performance database.

[0143] Data processing and model training module: Data processing and model training are performed based on machine learning frameworks and tools to establish a weld fatigue life prediction model.

[0144] Model evaluation and optimization module: using the coefficient of determination R 2 , MAPE, and RMSE are used to evaluate the performance of the model, and the prediction accuracy is improved through iterative optimization.

[0145] Prediction result application module: formulate maintenance plans and improvement measures based on model prediction results to improve equipment operation efficiency and safety.

[0146] Digital twin module: realizes digital mapping, real-time monitoring, fault diagnosis, performance prediction, simulation optimization and other functions of the pressure swing adsorber, providing decision support for full life cycle management.

[0147] 3. System Implementation

[0148] Hardware deployment and configuration: Select appropriate hardware devices according to system requirements, and install, configure and debug them to ensure the stability and reliability of the system hardware environment.

[0149] Software development and integration: Develop various functional modules based on the selected software and technologies, and conduct integration testing to ensure the integrity and synergy of system functions.

[0150] Data preparation and import: Collect and organize data from non-destructive testing, fatigue performance testing, and defect simulation and testing, and import them into the database management system for storage and management.

[0151] Model training and optimization: Use data processing and model training modules to perform model training and optimization to ensure the accuracy and stability of the weld fatigue life prediction model.

[0152] System testing and acceptance: Conduct comprehensive testing on the system, including functional testing, performance testing, safety testing, etc., to ensure that the system meets the design requirements and has practical application capabilities.

[0153] Operation maintenance and optimization: Perform daily operation and maintenance of the system, regularly update data, optimize models, upgrade software, etc., to ensure the long-term stable operation and continuous optimization of the system.

[0154] From the above embodiments, it can be seen that the specific advantage of the present invention lies in the integration of multidisciplinary technical means such as non-destructive testing, fatigue performance testing, defect simulation and testing, database construction, deep generative adversarial network GAN data enhancement, support vector regression SVR machine learning model and digital twin technology, which realizes the accurate prediction and scientific management of the fatigue life of the pressure swing adsorber weld, effectively improves the operating efficiency and safety of the equipment, reduces maintenance costs and risks, and provides strong support for the full life cycle management of the equipment.

[0155] The above are only a few preferred embodiments of the present invention, and their description is relatively specific and detailed, but it should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art may make various modifications and improvements without departing from the scope of the present invention, and such modifications and improvements are within the scope of protection of the present invention.

Claims

1. A method for predicting fatigue life of adsorber welds under hydrogen-rich alternating loads based on digital twins, characterized in that: The following steps are involved: (a) Perform nondestructive testing (NDT) on the pressure swing adsorber, including ultrasonic testing, magnetic particle testing, and X-ray testing, to obtain the spatial distribution of defects and residual stresses in the heat-affected zone (HAZ) and weld seams of the adsorber weld joints. Defects include geometric features such as pores, microcracks, undercuts, and reinforcements. The hardness of the HAZ and weld seams shall be recorded. (b) High-cycle fatigue performance testing was performed on the welded joints of Q345R steel adsorber scaled components in a pure hydrogen atmosphere. SN curves of the heat-affected zone and weld were obtained. The maximum stress range of the test was 200-400 MPa, and the stress ratio was -1 to 0.

8. (c) Prepare defective Q345R steel adsorber welded scaled parts or defective high-cycle fatigue specimens, count the number and size of defects, and obtain the SN curves of the weld and heat-affected zone of the defective weld joint in a pure hydrogen atmosphere; (d) establishing a high cycle fatigue performance database of adsorber welded joints in a hydrogen atmosphere based on the SN curves and defect feature distribution data obtained in steps (a)-(c); (e) Data augmentation is performed based on the adversarial network GAN, using the generator network Generator,G and the discriminator network Discriminator,D to generate data consistent with the real data distribution to obtain a sufficient number of labeled samples; (f) using the data set in step (d) and the data generated in step (e), establishing an adaptive machine learning model based on support vector regression (SVR), wherein the input variables include stress and defect characteristics, and the output variable is the corresponding fatigue life; the model uses the fatigue crack growth rate formula da / dN f =C(ΔK) m To evaluate the fatigue crack growth life, where C and m are material constants; ΔK is the stress intensity factor, and ΔK is given by the formula Calculated; A is the projected defect area on the plane perpendicular to the maximum principal stress; Y is the position constant, which is 0.5 for internal defects and 0.65 for surface and subsurface defects; Δσ is the applied stress range; (g) Using the coefficient of determination R 2 , mean absolute percentage error (MAPE) and root mean square error (RMSE) are used to evaluate the performance of the established machine learning model to ensure prediction accuracy. The adaptive machine learning model in step (f) also includes correlation analysis between input variables, and the Pearson correlation coefficient is used to evaluate the multicollinearity between input variables to improve the prediction accuracy of the model. The fatigue strength σw is predicted using the formula σ w =Y1×(HV / σb)×Δσ; Y1 is the position constant, 1.56 for internal defects, and 1.43 for surface and subsurface defects; HV is the Vickers hardness, and σb is the tensile strength.

2. The method according to claim 1, characterized in that Step (a) also includes detecting the residual stress distribution along the parent material-heat affected zone-weld direction distribution curve, and recording the residual stress values ​​in the axial and circumferential directions of the weld.

3. The method according to claim 1, characterized in that The defect type and size distribution of the defective Q345R steel adsorber weld joint in step (c) match the defect geometric envelope actually detected in step (a).

4. The method according to claim 1, wherein The GAN data enhancement process in step (e) includes: the generator G collects random variables z from the Gaussian distribution, obtains the generated data G(z) after nonlinear transformation, and inputs the real data x and the generated data G(z) into the discriminator D at the same time. The discriminator D judges the authenticity of the data by calculating the probability that the input data comes from the real data x, thereby training the GAN model until the generator G can generate samples that are close enough to the real data.

5. The method according to claim 1, wherein It also includes the application of prediction results, that is, formulating maintenance plans and improvement measures for pressure swing adsorbers based on the predicted weld fatigue life to improve the operating efficiency and safety of the equipment.

6. The method according to any one of claims 1 to 5, characterized in that The method uses digital twin technology to digitally map, monitor, diagnose, predict, simulate and optimize the pressure swing adsorber, thereby achieving full life cycle data management and equipment performance optimization.

7. The method according to claim 1, characterized in that The machine learning model in step (f) also uses the fatigue damage SN curve and its corresponding defect feature dataset, and combines it with the deep generative adversarial network (GAN) for data enhancement to obtain labeled samples for training the machine learning model to improve the accuracy and reliability of fatigue life prediction.

8. The method according to claim 7, characterized in that When training a machine learning model, the Pearson correlation coefficient is used to perform correlation analysis on the input variables to remove multicollinearity and improve the predictive performance of the model.

9. The method according to claim 8, characterized in that It also includes comparing the prediction results of the machine learning model with the actual situation, and optimizing the model parameters through continuous iteration to improve the accuracy and reliability of fatigue life prediction.

10. A digital twin-based adsorber weld fatigue life prediction system under hydrogen-rich alternating loads, characterized in that: include: 1) Nondestructive testing module: used to perform nondestructive testing on pressure swing adsorbers, including ultrasonic testing, magnetic particle testing, and X-ray testing, to obtain defects and residual stress spatial distribution in the heat-affected zone and weld seam of adsorber weld joints. Defects include geometric features such as pores, microcracks, undercuts, and reinforcements. 2) Fatigue performance test module: used to perform high-cycle fatigue performance tests on welded joints of Q345R steel adsorber scaled parts in a pure hydrogen atmosphere, and obtain the SN curves of the heat-affected zone and welds; 3) Defect Simulation and Testing Module: This module is used to prepare defective Q345R steel adsorber welded scale parts, count the number and size of defects, and obtain the SN curve of defective welded joints in a pure hydrogen atmosphere. 4) Database construction module: used to establish a high-cycle fatigue performance database of adsorber welded joints in a hydrogen atmosphere based on the SN curves and defect feature distribution data obtained by the nondestructive testing module and the defect simulation and testing module; 5) Data Processing and Model Training Module: This module uses a deep generative adversarial network (GAN) for data augmentation, utilizing a generator network and a discriminator network to generate data consistent with the real-world data distribution. Furthermore, the module utilizes the dataset in the database construction module and the GAN-generated data to establish an adaptive machine learning model based on support vector regression (SVR) for predicting weld fatigue life. The model also utilizes the fatigue crack growth rate formula and fatigue damage SN curve for evaluation. 6) Model evaluation and optimization module: used to use the determination coefficient R 2 , mean absolute percentage error (MAPE) and root mean square error (RMSE) are used to evaluate the performance of the machine learning model, and the prediction accuracy is improved by iteratively optimizing the model parameters. The module also uses the Pearson correlation coefficient to perform correlation analysis on the input variables to remove multicollinearity. 7) Prediction result application module: used to formulate maintenance plans and improvement measures for the pressure swing adsorber based on the prediction results of the machine learning model to improve the operating efficiency and safety of the equipment; 8) Digital Twin Module: Used to digitally map, monitor, diagnose, predict, simulate, and optimize pressure swing adsorbers, enabling full lifecycle data management and equipment performance optimization.

Images