Delayed service evaluation method and system for pressure vessel exceeding design age limit

Abstract

The invention provides a method and a system for evaluating the delayed service of a pressure vessel beyond the design age limit, and relates to the technical field of safety evaluation of the pressure vessel beyond the design age limit, and the method comprises the steps of data acquisition, model construction, model training, damage mode recognition, damage judgment, first inspection, second inspection, safety condition grade evaluation and the like. The system comprises a data acquisition module, a model construction module, a model training module, a damage mode recognition module, a damage judgment module, a first inspection module, a second inspection module and a safety condition grade evaluation module. The damage mode of the pressure vessel under the complex working condition is more accurately judged, a more persuasive and credible evaluation conclusion is provided for whether the pressure vessel exceeding the designed service life can be continuously and safely used or not, and the probability of safety accidents is reduced.

Description

Technical Field

[0001] This application relates to the field of safety assessment technology for pressure vessels exceeding their design service life, and in particular to a method and system for assessing the extended service life of pressure vessels exceeding their design service life. Background Technology

[0002] Pressure vessels, as widely used special equipment, undertake various critical tasks such as storage, transportation, and reaction. In industrial production, pressure vessels are typically large, complex, and expensive pieces of equipment. Over time, many pressure vessels approach or exceed their design lifespan, requiring significant investment for companies to purchase and install new ones. If a safety assessment can be conducted on pressure vessels exceeding their design lifespan and it can be determined that they can continue to be used, the cost of equipment replacement can be greatly reduced, which is crucial for a company's economic efficiency.

[0003] Chinese invention patent application CN118350268A, with a publication date of July 16, 2024, provides a safety assessment method for continuing to use in-service steel pressure vessels beyond their design service life. The method involves obtaining the original data of the equipment to be assessed, using fault tree analysis to obtain the main failure modes and their basic influencing factors, calculating the structural importance of the basic influencing factors, selecting the important influencing factors leading to equipment failure for fuzzy comprehensive evaluation based on the analytic hierarchy process, obtaining the risk level of the important influencing factors of the equipment to be assessed, and combining this with the equipment's previous periodic inspection data to evaluate the effectiveness of the inspections, ultimately determining whether the equipment can continue to be used safely.

[0004] The above-mentioned technical solutions did not consider the impact of different operating conditions on the damage modes of pressure vessels throughout the analysis process. Relying solely on the collected equipment data, it is impossible to accurately identify the damage modes that the vessel may experience under specific operating conditions. As a result, the pressure vessel safety assessment results are difficult to accurately reflect the true risk status of the pressure vessel, affecting the accuracy of the safety assessment of pressure vessels beyond their design service life. Summary of the Invention

[0005] To improve the accuracy of safety assessments of pressure vessels exceeding their design service life, this application provides a method and system for assessing the extended service life of pressure vessels exceeding their design service life.

[0006] Firstly, this application provides a method for evaluating the extended service life of pressure vessels beyond their design life, employing the following technical solution: A method for evaluating the extended service life of pressure vessels beyond their design life, the method comprising: Data acquisition includes collecting real-time data and collecting historical data; Real-time data acquisition: Acquire real-time data from the pressure vessel; Collect historical data: Collect historical data of pressure vessels and corresponding damage type labels; Model Construction: A pressure vessel damage identification model is constructed, which includes a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model. Model training includes damage pattern identification, primary processing, interaction extraction, secondary processing, and damage pattern ranking. Damage pattern identification: The historical data of the collected pressure vessel and the corresponding damage type labels are used as input to the random forest sub-model, and the first data is output. First processing: Match the first data with the historical data of the pressure vessel to obtain the second data; Extracting interaction relationships: The second data is used as the input to the graph neural network sub-model, and the third data is output. The second process involves matching the third data with historical data and corresponding damage type labels to obtain the fourth data. Damage pattern ranking: The fourth data point is used as the input to the regression prediction sub-model, and the output is the ranking of the impact of damage patterns in the historical data. Damage pattern recognition: Input the real-time data of the pressure vessel into the trained pressure vessel damage recognition model, output the damage pattern impact ranking of the real-time data, and take the damage pattern with the highest impact ranking as the main damage pattern of the real-time data. Damage assessment: Determine whether the main damage mode of real-time data is time-dependent. If so, proceed with the first check step; If not, proceed to the second test step; First inspection: Several special inspection schemes are pre-set, select the special inspection scheme corresponding to the main damage mode of the real-time data, inspect the current pressure vessel according to the selected special inspection scheme, output the inspection results, and perform the steps of safety status level assessment. Second inspection: A general inspection plan is set in advance, the current pressure vessel is inspected according to the general inspection plan, the inspection results are output, and the steps of safety status level assessment are performed; Safety Status Assessment: Based on the inspection results, assess the current safety status of the pressure vessel and evaluate the outcome of extending its service life.

[0007] By adopting the above technical solution, a pressure vessel damage identification model is constructed, comprising a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model. The random forest sub-model identifies pressure vessel damage patterns, the graph neural network sub-model provides the interaction relationships between damage patterns, and the regression prediction model analyzes the impact of damage patterns on the equipment. This helps to capture the damage characteristics of pressure vessels under different operating conditions, identify pressure vessel damage patterns, and rank these patterns by their impact, achieving the technical effect of quickly locating the main damage patterns. This improves the accuracy and reliability of pressure vessel safety assessment and provides strong technical support for the continued use of pressure vessels exceeding their design service life.

[0008] Optionally, after performing the model building step and before performing the damage pattern identification step, the following steps are also included: Operating condition extraction: Extract operating condition data from historical data to obtain operating condition category labels; Label hierarchy division: The damage types corresponding to historical data are hierarchically divided into top-level labels and corresponding sub-level labels, and the corresponding working condition category labels are matched for the top-level labels and corresponding sub-level labels. Tag encoding: The top-level tag, the corresponding sub-tags, and the corresponding working condition category tag of each damage type are taken as a group of tags, denoted as damage type tag group. The damage type tag group is encoded to obtain the top-level encoded tag and the corresponding sub-encoded tag of each damage type tag, denoted as damage type tag encoding group of damage type tag. Build classifiers: Build a multi-classifier for each top-level label and a binary classifier for each child label of the top-level label; First training: Input historical data into the multi-classifier for training, and predict the top-level encoding label of the damage type of the current historical data, which is denoted as the first label; Second training: Input the first label into the binary classifier for training, and predict the sub-label of the damage type of the current historical data, which is denoted as the second label; Output merging: The first label and the second label are concatenated, and the concatenated data is used as the output of the random forest sub-model.

[0009] By employing the aforementioned technical solutions, operating condition data is extracted, damage types are refined and classified, and corresponding operating condition category labels are matched to top-level and sub-level labels, further enhancing the model's sensitivity to changes in operating conditions. Furthermore, multi-classifiers and binary classifiers are constructed and trained, with predictions progressing from top-level labels to sub-level labels in a progressive manner, improving the accuracy and stability of model predictions. This helps the model learn the characteristics of damage patterns under different operating conditions, more accurately identifying potential damage patterns that pressure vessels may experience under specific conditions. Simultaneously, this contributes to improving the accuracy of pressure vessel safety assessments, ensuring that assessment results more accurately reflect the vessel's risk status, thus providing a more reliable basis for vessel maintenance and management.

[0010] Optionally, after performing the first processing step and before performing the step of extracting interaction relationships, the method further includes: Define the graph structure: Take the collection time of each set of historical data in the historical data as a node, construct a node feature matrix, construct a time relationship adjacency matrix according to the collection order and collection time interval of each set of historical data in the historical data, and denote it as the first adjacency matrix. Connect the nodes with the first adjacency matrix and denote it as the first graph structure. Third training: Input historical data into the graph neural network sub-model, perform graph convolution propagation based on the first graph structure to obtain the first node representation, which is used as the output of the graph neural network sub-model; Model optimization: Define a first loss function, minimize the first loss function to optimize the graph neural network sub-model, and obtain the optimized graph neural network sub-model, which is denoted as the new graph neural network sub-model.

[0011] By adopting the above technical solution, the technical problem of not considering the impact of different working conditions and time changes on the damage mode of pressure vessels is effectively solved. Moreover, by using the acquisition time of historical data as a node, a temporal relationship adjacency matrix is ​​constructed based on time sequence and interval time. This helps the model to more accurately consider the impact of time factors on damage development when identifying damage modes, improves the accuracy of the model in capturing the interaction of damage modes and the spatiotemporal correlation, optimizes the model performance and enhances the interpretability of the model, thereby helping to improve the accuracy of pressure vessel safety assessment.

[0012] Optionally, after performing the third training step and before performing the model optimization step, the following steps are also included: Add a time decay factor: Adjust the weight of the relationship between two nodes based on the historical data collection time interval. ; in, Represents a node and Relationship weights Indicates the attenuation factor. Represents a node The collection time, Represents a node The collection time.

[0013] By adopting the above technical solution and introducing the time decay factor, the model can more accurately consider the correlation between different time points in historical data, enabling the model to capture the correlation of damage patterns changing over time, thereby more accurately predicting future damage patterns, reducing the interference of noise data, and improving the robustness of the model.

[0014] Optionally, after performing the step of defining the graph structure and before performing the third training step, the following steps are also included: Second adjacency matrix: Calculate the feature similarity of each group of historical data in the historical data, and construct a feature adjacency matrix using the feature similarity, denoted as the second adjacency matrix; The third adjacency matrix: Traverse the historical data and the corresponding damage type labels, count the frequency of co-occurrence of the same damage type label among each group of historical data, and denote it as the first frequency. Construct a damage pattern co-occurrence adjacency matrix based on the first frequency, and denote it as the third adjacency matrix.

[0015] By adopting the above technical solution, and comprehensively considering the similarity between each set of data and the co-occurrence between the damage modes of each set of data, a feature similarity adjacency matrix and a damage mode co-occurrence adjacency matrix are constructed. This enables the model to capture the correlation and interaction between different damage modes, which helps the model to comprehensively consider the common influence of multiple damage modes when identifying damage modes under specific working conditions, thereby improving the comprehensiveness and accuracy of the assessment.

[0016] Optionally, after performing the third training step and before performing the model optimization step, the following steps are also included: Fourth training: Denote the node and the second adjacency matrix as the second graph structure, input historical data into the graph neural network sub-model, perform graph convolution propagation based on the second graph structure, and obtain the second node representation; Fifth training: Denote the node and the third adjacency matrix as the third graph structure, input historical data into the graph neural network sub-model, perform graph convolution propagation based on the third graph structure, and obtain the third node representation; Relationship fusion: The first node representation, the second node representation, and the third node representation are weighted and accumulated to obtain the fourth node representation, which is used as the new first node representation.

[0017] By employing the above technical solutions and constructing different graph structures, the model can learn the impact of changing operating conditions on damage patterns during training, thereby improving the model's sensitivity to changes in operating conditions and helping it to more comprehensively capture the damage characteristics of pressure vessels under different operating conditions. These graph structures provide the model with additional information input.

[0018] Optionally, after performing the third training step and before performing the model optimization step, the following steps are also included: Weight model construction: Based on multilayer perceptron, a weight learning network is constructed; Weight learning: Extract working condition data from historical data to obtain working condition category labels, then import the working condition data into the weight learning network for model inference, and output the working condition weight corresponding to each working condition category label; Weight matching: Match the working condition weights with the first node features, and record them as the new first node features.

[0019] By adopting the above technical solution, the operating condition information is integrated into the node features, which makes the node features more comprehensively reflect the state of the pressure vessel under different operating conditions. At the same time, the model can more accurately identify the damage modes that the pressure vessel may produce under different operating conditions, thus improving the accuracy of safety assessment.

[0020] Optionally, after performing the step of extracting interaction relationships and before performing the step of the second processing, the method further includes: Calculate representation similarity: Calculate the similarity between the representations of each node in the third data, denoted as the first similarity; Construct the interaction matrix: Based on the first similarity, construct the interaction matrix between each node, and use the interaction matrix and the third data as the new third data.

[0021] By employing the aforementioned technical solutions, calculating representation similarity, and constructing an interaction matrix, the model can more accurately capture the interaction relationships between different nodes in the pressure vessel data. Considering the interaction relationships under different operating conditions, the model reveals the interactive connections between various nodes, enabling it to grasp the overall interaction between different damage modes and data under different operating conditions. This allows the model to more accurately capture the interaction relationships between different nodes in the pressure vessel data, fully considering the characteristics of interaction relationships under different operating conditions. Consequently, the safety assessment results derived from this model are more reliable and more accurately reflect the true risk status of the pressure vessel.

[0022] Secondly, this application provides an evaluation system for extended service life of pressure vessels beyond their design life. The evaluation system is applicable to the evaluation method described in any one of the first aspects above, and the evaluation system includes the following technical solution: The system for assessing the extended service life of pressure vessels beyond their design life includes: The data acquisition module includes a real-time data acquisition module and a historical data acquisition module; The real-time data acquisition module is used to acquire real-time data from the pressure vessel. The historical data acquisition module is used to collect historical data of pressure vessels and the corresponding damage type labels for the historical data. The model building module is used to build a pressure vessel damage identification model, which includes a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model. The model training module communicates with the data acquisition module and the model building module, and includes a damage pattern identification unit, a first processing unit, an interaction relationship extraction unit, a second processing unit, and a damage pattern sorting unit. The damage pattern identification unit is used to take the collected historical data of the pressure vessel and the corresponding damage type labels as input to the random forest sub-model and output the first data. The first processing unit is used to match the first data with the historical data of the pressure vessel to obtain the second data; Extract the interaction relationship unit, which is used to take the second data as input to the graph neural network sub-model and output the third data.

[0023] The second processing unit is used to match the third data with historical data and corresponding damage type labels to obtain the fourth data; The damage pattern ranking unit is used to take the fourth data as input to the regression prediction sub-model and output the damage pattern impact ranking results of historical data. The damage pattern recognition unit communicates with the data acquisition module and the model building module, inputs the real-time data of the pressure vessel into the trained pressure vessel damage recognition model, outputs the damage pattern impact ranking result of the real-time data, and takes the damage pattern with the first impact ranking as the main damage pattern of the real-time data. The damage discrimination unit is used to determine whether the main damage mode of real-time data is time-dependent. If the main damage pattern in the real-time data is time-dependent, then the first inspection unit is triggered; If the main damage pattern in the real-time data is independent of time, then the second test unit is triggered; The first inspection unit is connected to the damage discrimination unit. It is used to pre-set several special inspection schemes, select the special inspection scheme corresponding to the main damage mode of the real-time data, inspect the current pressure vessel according to the selected special inspection scheme, output the inspection results, and perform the steps of safety status level assessment. The second inspection unit is connected in communication with the damage judgment unit. It is used to pre-set a general inspection plan, inspect the current pressure vessel according to the general inspection plan, output the inspection results, and perform the steps of safety status level assessment. The safety status assessment unit is connected in communication with the first inspection unit and the second inspection unit. It is used to assess the current safety status level of the pressure vessel based on the inspection results and evaluate the results of the current pressure vessel's extended service life.

[0024] By adopting the above technical solutions, the model building module constructs a pressure vessel damage identification model that includes a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model. This helps to capture the damage characteristics of pressure vessels under different working conditions, identify the damage patterns of pressure vessels, and rank these patterns by impact. This achieves the technical effect of quickly locating the main damage patterns, improving the accuracy and reliability of pressure vessel safety assessment, and providing strong technical support for the continued use of pressure vessels beyond their design service life.

[0025] In summary, this application includes at least one of the following beneficial technical effects: 1. A pressure vessel damage identification model is constructed, comprising a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model. The random forest sub-model identifies pressure vessel damage patterns, the graph neural network sub-model provides the interaction relationships between damage patterns, and the regression prediction model analyzes the impact of damage patterns on the equipment. This model helps capture the damage characteristics of pressure vessels under different operating conditions, identify pressure vessel damage patterns, and rank these patterns by their impact. It achieves the technical effect of quickly locating the main damage patterns, improving the accuracy and reliability of pressure vessel safety assessment, and providing strong technical support for the continued use of pressure vessels beyond their design service life.

[0026] 2. By extracting operating condition data and refining and classifying damage types, corresponding operating condition category labels are matched to top-level and sub-level labels, further enhancing the model's sensitivity to changes in operating conditions. Furthermore, multi-classifiers and binary classifiers are constructed and trained, progressively predicting from top-level labels to sub-level labels, improving the accuracy and stability of model predictions. This helps the model learn the characteristics of damage patterns under different operating conditions, more accurately identifying potential damage patterns that may occur in pressure vessels under specific conditions. This also contributes to improving the accuracy of pressure vessel safety assessments, ensuring that the assessment results more accurately reflect the vessel's risk status, thus providing a more reliable basis for vessel maintenance and management.

[0027] 3. By comprehensively considering the similarity between each set of data and the co-occurrence between the damage modes of each set of data, a feature similarity adjacency matrix and a damage mode co-occurrence adjacency matrix are constructed. This enables the model to capture the correlation and interaction between different damage modes, which helps the model to comprehensively consider the common influence of multiple damage modes when identifying damage modes under specific working conditions, thereby improving the comprehensiveness and accuracy of the assessment. Attached Figure Description

[0028] Figure 1 This is a flowchart of Embodiment 1 of this application; Figure 2 This is a flowchart of the S3 model training in Embodiment 1 of this application; Figure 3 This is a flowchart of the S31 random forest sub-model training in Embodiment 1 of this application; Figure 4 This is a flowchart of the training of the S34 graph neural network sub-model in Embodiment 1 of this application. Detailed Implementation

[0029] The following combination Figures 1 to 4 This application will be described in further detail.

[0030] Example 1: This example discloses a method for evaluating the extended service life of pressure vessels beyond their design life, such as... Figure 1 As shown, the method includes: collecting real-time data of the pressure vessel, historical data of the pressure vessel, and damage type labels corresponding to the historical data; constructing a pressure vessel damage identification model, which includes a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model; inputting the historical data of the pressure vessel and the corresponding damage type labels into the pressure vessel damage identification model for model training to obtain a trained pressure vessel damage identification model; inputting the real-time data of the pressure vessel into the trained pressure vessel damage identification model; outputting the damage pattern impact ranking result of the real-time data; taking the damage pattern with the highest impact ranking as the main damage pattern of the real-time data; then conducting pressure vessel inspection based on the main damage pattern; and assessing the safety status level of the pressure vessel and the result of the pressure vessel's delayed service based on the inspection results. This embodiment includes the following steps: S1 data acquisition: includes S11 for acquiring real-time data and S12 for acquiring historical data.

[0031] S11 Real-time Data Acquisition: Acquires real-time data from the pressure vessel, including equipment status data and environmental data. Equipment status data includes internal pressure, internal temperature, vibration signals, flow rate, internal stress, acoustic signals from various parts of the pressure vessel, dissolved oxygen concentration, pH value, corrosion morphology, number of loading cycles, and crack detection data. Operating condition data includes ambient temperature, humidity, pH value, and vibration signals.

[0032] In this embodiment, the internal pressure of the pressure vessel is collected by pressure sensors installed at different locations within the pressure vessel, and the internal temperature is collected by temperature sensors (such as thermocouples or platinum resistance temperature sensors (RTD)). Vibration signals of the pressure vessel are collected by acceleration sensors or vibration sensors (such as piezoelectric accelerometers or MEMS sensors). The flow rate within the pressure vessel is collected by flow meters (such as turbine flow meters, mass flow meters, or electromagnetic flow meters). The internal stress of the pressure vessel is collected by strain gauges or fiber optic grating (FBG) sensors. Acoustic signals from various parts of the pressure vessel are collected by acoustic sensors (such as ultrasonic sensors, microphone arrays, or piezoelectric sensors), the dissolved oxygen concentration inside the pressure vessel is collected by a dissolved oxygen sensor, the pH value of the medium inside the pressure vessel is collected by a pH sensor, and the corrosion morphology of the pressure vessel is collected by corrosion sensors (such as electrochemical corrosion sensors). The fatigue state of the equipment is estimated by recording the number of loading and unloading cycles using strain gauges or pressure sensors installed on the equipment, and crack detection data of the pressure vessel is collected using crack detection techniques (such as ultrasonic testing, X-ray, or acoustic emission techniques). Ambient temperature is collected through an ambient temperature sensor (such as a temperature probe or thermocouple), ambient humidity is collected through a humidity sensor (such as a capacitive or resistive humidity sensor), ambient pH value is collected through a pH sensor, and ambient vibration signals are collected through a vibration sensor (such as an accelerometer or MEMS sensor).

[0033] In this embodiment, when collecting real-time data of the pressure vessel, all real-time data of the pressure vessel collected at the same time point are taken as a set of real-time data of the pressure vessel.

[0034] S12 Historical Data Acquisition: Collects historical data of the pressure vessel and corresponding damage type labels. The historical data package for the pressure vessel includes real-time data on equipment status and environmental data. Equipment status data includes internal pressure, internal temperature, vibration signals, flow rate, internal stress, acoustic signals at various locations, dissolved oxygen concentration, pH value, corrosion morphology, number of loading cycles, and crack detection data. Environmental data includes ambient temperature, humidity, pH value, and vibration signals. Damage type labels include corrosion thinning, environmental cracking, material degradation, mechanical damage, and other damage. Corrosion thinning includes hydrochloric acid corrosion, sulfuric acid corrosion, hydrofluoric acid corrosion, phosphoric acid corrosion, carbon dioxide corrosion, naphthenic acid corrosion, phenol corrosion, low molecular weight organic acid corrosion, high temperature oxidation corrosion, atmospheric corrosion (without insulation layer), atmospheric corrosion (with insulation layer), cooling water corrosion, soil corrosion, microbial corrosion, boiler condensate corrosion, alkali corrosion, ash corrosion, smoke leak corrosion, ammonium chloride corrosion, amine corrosion, high temperature sulfide corrosion (hydrogen-free environment), high temperature sulfide corrosion (hydrogen environment), acidic water corrosion (alkaline acidic water), acidic water corrosion (acidic acidic water), methylammonium corrosion, galvanic corrosion, salt water corrosion, oxygen-containing process water corrosion, and concentration cell corrosion. Environmental cracking includes chloride stress corrosion cracking, carbonate stress corrosion cracking, nitrate stress corrosion cracking, alkali stress corrosion cracking, ammonia stress corrosion cracking, amine stress corrosion cracking, wet hydrogen sulfide damage, hydrofluoric acid-induced hydrogen stress cracking, hydrocyanic acid-induced hydrogen stress cracking, hydrogen embrittlement, high-temperature water stress corrosion cracking, polythionine stress corrosion cracking, liquid metal brittle fracture, ethanol stress corrosion cracking, sulfate stress corrosion cracking, and hydrofluoric acid stress corrosion cracking. Material degradation includes grain growth, nitriding, spheroidization, graphitization, carburizing, decarburization, and metal pulverization. Phase embrittlement, 475°C embrittlement, tempering embrittlement, irradiation embrittlement, titanium hydrogenation, reheat cracking, demetallization corrosion, sensitization-intergranular corrosion, and metal thermal aging. Mechanical damage includes mechanical fatigue, thermal fatigue (including ratcheting effect), vibration fatigue, contact fatigue, mechanical wear, erosion, cavitation, overload, thermal shock, creep, and strain aging. Other damage includes high-temperature hydrogen corrosion, corrosion fatigue, erosion, vapor trapping, low-temperature brittle fracture, overheating, refractory degradation, graphitization corrosion of cast iron, fretting corrosion, and combustion and explosion induced by high oxygen gas content.

[0035] In this embodiment, historical data collected at the same time point and its corresponding damage type label are considered as a set of historical data for pressure vessels, and one set of historical data contains one or more damage type labels.

[0036] S2 Model Construction: Construct a pressure vessel damage identification model, which includes a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model.

[0037] In this embodiment, CART decision trees, ID3 decision trees, C4.5 decision trees, and Gini decision trees can be used as the base models for the random forest sub-model. GCN (Graph Convolutional Network), GAT (Graph Attention Network), GraphSAGE (Graph Sample and Aggregation), and ChebNet (Chebyshev Spectral Graph Convolutional Network) can be used as the base models for the graph neural network sub-model. Linear regression, support vector regression, decision tree regression, random forest regression, gradient boosting tree regression, and neural network regression models can be used as the base models for the regression prediction sub-model. In this embodiment, CART decision trees are selected as the base model for the random forest sub-model, GAT is selected as the base model for the graph neural network sub-model, and gradient boosting tree regression is selected as the base model for the regression prediction sub-model.

[0038] S3 Model Training: Includes S31 Random Forest Sub-model Training, S32 Damage Pattern Identification, S33 First Processing, S34 Graph Neural Network Sub-model Training, S35 Interaction Relationship Extraction, S36 Interaction Matrix Construction, S37 Second Processing, and S38 Damage Pattern Ranking, such as... Figure 2 As shown.

[0039] S31 Random Forest Sub-model Training: Includes S311 Job Condition Extraction, S312 Label Hierarchy Division, S313 Label Encoding, S314 Classifier Construction, S315 First Training, S316 Second Training, and S317 Output Merging, as follows Figure 3 As shown.

[0040] S311 Operating Condition Extraction: Extract operating condition data from historical data, define the working environment of the pressure vessel, and obtain operating condition category labels. The operating condition category labels include single operating condition category labels such as low temperature environment, high temperature environment, normal temperature environment, high humidity environment, low humidity environment, medium humidity environment, acidic environment, alkaline environment, neutral environment, high vibration environment, and low vibration environment, as well as multi-dimensional operating condition category labels such as high temperature and high humidity environment, low temperature and low humidity environment, acidic and high humidity environment, and normal temperature, normal humidity, and neutral environment.

[0041] S312 Tag Hierarchy: Historical data corresponding to damage types are hierarchically divided into top-level tags and corresponding sub-level tags. The top-level tags and corresponding sub-level tags are then matched with corresponding operating condition category tags. Top-level tags include corrosion thinning, environmental cracking, material degradation, mechanical damage, and other damage. Sub-level tags corresponding to corrosion thinning include hydrochloric acid corrosion, sulfuric acid corrosion, hydrofluoric acid corrosion, phosphoric acid corrosion, carbon dioxide corrosion, naphthenic acid corrosion, phenol corrosion, low-molecular-weight organic acid corrosion, high-temperature oxidation corrosion, atmospheric corrosion (without insulation), atmospheric corrosion (with insulation), cooling water corrosion, soil corrosion, microbial corrosion, boiler condensate corrosion, alkali corrosion, ash corrosion, smoke leak corrosion, ammonium chloride corrosion, amine corrosion, high-temperature sulfide corrosion (hydrogen-free environment), high-temperature sulfide corrosion (hydrogen environment), acidic water corrosion (alkaline acidic water), acidic water corrosion (acidic acidic water), methylammonium corrosion, galvanic corrosion, salt water corrosion, oxygen-containing process water corrosion, and concentration cell corrosion. Sub-labels for environmental cracking include chloride stress corrosion cracking, carbonate stress corrosion cracking, nitrate stress corrosion cracking, alkali stress corrosion cracking, ammonia stress corrosion cracking, amine stress corrosion cracking, wet hydrogen sulfide damage, hydrofluoric acid-induced hydrogen stress cracking, hydrocyanic acid-induced hydrogen stress cracking, hydrogen embrittlement, high-temperature water stress corrosion cracking, polythionine stress corrosion cracking, liquid metal brittle fracture, ethanol stress corrosion cracking, sulfate stress corrosion cracking, and hydrofluoric acid stress corrosion cracking. Sub-labels for material degradation include grain growth, nitriding, spheroidization, graphitization, carburizing, decarburization, and metal pulverization. Phase embrittlement, 475°C embrittlement, tempering embrittlement, irradiation embrittlement, titanium hydrogenation, reheat cracking, demetallization corrosion, sensitization-intergranular corrosion, and metal thermal aging. Sub-labels for mechanical damage include mechanical fatigue, thermal fatigue (including ratcheting effect), vibration fatigue, contact fatigue, mechanical wear, erosion, cavitation, overload, thermal shock, creep, and strain aging. Sub-labels for other types of damage include high-temperature hydrogen corrosion, corrosion fatigue, erosion, vapor trapping, low-temperature brittle fracture, overheating, refractory degradation, cast iron graphitization corrosion, fretting corrosion, and combustion and explosion induced by high-oxygen gas.

[0042] S313 Tag Encoding: The top-level tag, the corresponding sub-tags, and the corresponding working condition category tag of each damage type tag are taken as a group of tags, denoted as the damage type tag group. The damage type tag group is represented by hot encoding to obtain the top-level encoded tag and the corresponding sub-level encoded tag of each damage type tag, denoted as the damage type tag encoding group of the damage type tag.

[0043] S314 classifier construction: Build a multi-classifier for each top-level label and a binary classifier for each top-level label's child labels.

[0044] S315 First Training: Input historical data into the multi-classifier for training, and predict the top-level encoding label of the damage type of the current historical data, which is denoted as the first label.

[0045] S316 Second Training: Input the first label into the binary classifier for training, and predict the sub-label of the damage type of the current historical data, which is denoted as the second label.

[0046] S317 Output Merging: The first label and the second label are concatenated, and the concatenated data is used as the output of the random forest sub-model.

[0047] In this embodiment, operating condition data is extracted, damage types are refined and classified, and corresponding operating condition category labels are matched to top-level and sub-level labels, further enhancing the model's sensitivity to changes in operating conditions. Furthermore, multi-classifiers and binary classifiers are constructed and trained, with predictions progressing from top-level labels to sub-level labels in a progressive manner, improving the accuracy and stability of model predictions. This helps the model learn the characteristics of damage patterns under different operating conditions, more accurately identifying the damage patterns that pressure vessels may experience under specific operating conditions, and also contributing to improving the accuracy of pressure vessel safety assessments.

[0048] S32 Damage Pattern Identification: The collected historical data of the pressure vessel and the corresponding damage type labels are used as input to the random forest sub-model, and the first data is output.

[0049] S33 First Processing: Align the first data with the historical data of the pressure vessel in time series, and then directly concatenate the first data with the historical data of the pressure vessel to obtain the second data.

[0050] S34 Graph Neural Network Sub-model Training: This includes S341 defining the graph structure, S342 third training, S343 adding a time decay factor, S344 introducing operating condition weights, and S345 optimizing the model, such as... Figure 4 As shown.

[0051] S341 defines the graph structure as follows: Each set of historical data is represented by its collection time as a node. For each node, its data features are used to construct a node feature matrix. Based on the collection order and time interval of each set of historical data, a temporal adjacency matrix is ​​constructed, denoted as the first adjacency matrix. That is, if two nodes are temporally adjacent or have a specific time interval relationship, a non-zero element is set at the corresponding position in the adjacency matrix to indicate a connection between them.

[0052] S342 Third Training: Configure the structural parameters of the graph neural network, including determining the number of graph convolutional layers, the output dimension of each layer, and the activation function. Input the first graph structure into the graph neural network sub-model. Each node performs information propagation and feature updates based on its own features and the feature information of its neighboring nodes. After propagation through multiple layers of graph convolutions, the representation of the first node is obtained, which serves as the output of the graph neural network sub-model.

[0053] S343 adds a time decay factor: adjusting the weight of the relationship between two nodes based on the historical data acquisition time interval. ; in, Represents a node and Relationship weights Indicates the attenuation factor. Represents a node The collection time, Represents a node The collection time.

[0054] In this embodiment, during graph convolution, the original connection weights between nodes are multiplied by the corresponding time decay factor to adjust the degree of dependence of nodes on information at different time points during information propagation. This allows the model to more reasonably consider the correlation between different time points in historical data, thereby more accurately capturing the pattern of damage patterns changing over time, reducing excessive interference from long-term data on the current analysis, and improving the accuracy and stability of the model's damage pattern prediction.

[0055] In this embodiment, by setting an appropriate attenuation factor This allows control over the degree to which time intervals affect the weights of node relationships. For example, when When the time interval is large, a slightly longer time interval will cause the node relationship weights to decay rapidly, indicating that the model pays more attention to the correlation of recent data; conversely, when the time interval is small... When the time interval is relatively short, the impact of the time interval on the weight is relatively small, and the model will consider longer-term data relationships.

[0056] In this embodiment, a grid search method is used to select a series of attenuation factors. Candidate values ​​are used to evaluate the graph neural network sub-model under different conditions using cross-validation or hold-out methods. Choose the value that best optimizes the model's performance metrics. The value serves as the final decay factor.

[0057] S344 Working Condition Weight Introduction: This includes S3441 Weight Model Construction, S3442 Weight Learning, and S3443 Weight Matching.

[0058] S3441 Weight Model Construction: Based on a multilayer perceptron, the number of neurons is determined according to the data type of the working condition data, and a weight learning network is constructed. S3442 Weight Learning: Extract working condition data from historical data, perform data processing and normalization preprocessing on the extracted data to obtain working condition category labels, and then import the working condition data into the weight learning network. The data passes through the hidden layers in sequence, and each neuron calculates the output based on the input data, connection weights and activation function, outputting the predicted working condition weight corresponding to each working condition category.

[0059] S3443 Weight Matching: Extract information about the working condition data from the first node feature, use the working condition weight as a weighting coefficient, multiply it with each element in the first node feature and concatenate them, and the resulting data is recorded as the new first node feature.

[0060] S345 Optimization Model: Define a first loss function, use an optimization algorithm (such as stochastic gradient descent, Adagrad, Adadelta, Adam, etc.) to minimize the first loss function, optimize the graph neural network sub-model, and obtain the optimized graph neural network sub-model, denoted as the new graph neural network sub-model.

[0061] The first loss function is: ; ; ; in, Indicates the weighting coefficient. This represents the graph structure prediction loss function. Represents the cross-entropy loss function. This represents the number of nodes in the first adjacency matrix. Indicates the first 1 node Indicates the first 1 node Represents a node and nodes The true weight, Represents a node and nodes Prediction weights, Indicates the number of damage pattern classification labels. Indicates the first Damage pattern classification labels, Damage pattern classification labels The true label, Damage pattern classification labels Predicted labels.

[0062] In this embodiment, a grid search method is used to select a series of... Candidate values ​​are used to evaluate the graph neural network sub-model under different conditions using cross-validation or hold-out methods. Choose the value that best optimizes the model's performance metrics. The value is used as the final weighting coefficient.

[0063] In other embodiments, after performing S341 to define the graph structure and before performing S343 to add the time decay factor, the method further includes: The S3421 second adjacency matrix: The feature similarity of each group of historical data is calculated using Euclidean distance, cosine similarity, or Pearson correlation coefficient. A pre-set similarity threshold is used; when the feature similarity is greater than the threshold, the two groups of historical data are considered related, and vice versa. A feature adjacency matrix, denoted as the second adjacency matrix, is constructed using these feature similarities.

[0064] S3422 Third Adjacency Matrix: Traverse all historical data and their corresponding damage type labels, and check for cases where damage type labels are the same. During the traversal, count the frequency of co-occurrence of the same damage type label between each group of historical data, denoted as the first frequency. Use the first frequency as the assignment between two groups of historical data, and construct a damage pattern co-occurrence adjacency matrix based on the first frequency, denoted as the third adjacency matrix.

[0065] S3423 Fourth Training: The nodes and the second adjacency matrix are denoted as the second graph structure. The second graph structure is input into the graph neural network sub-model. Each node performs information propagation and feature update based on its own features and the feature information of its neighboring nodes. After multi-layer graph convolution propagation, the second node representation is output.

[0066] S3424 Fifth Training: The nodes and the third adjacency matrix are denoted as the third graph structure. The third graph structure is input into the graph neural network sub-model. Each node performs information propagation and feature update based on its own features and the feature information of its neighboring nodes. After multi-layer graph convolution propagation, the third node representation is output.

[0067] S3425 Relation Fusion: The first node representation, the second node representation, and the third node representation are weighted and accumulated to obtain the fourth node representation, which is used as the new first node representation.

[0068] In this embodiment, initial weights are assigned to the first node representation, the second node representation, and the third node representation. Cross-validation or hold-out methods are used to evaluate the performance of the graph neural network sub-model under different weight assignments for the first, second, and third node representations. A set of weight assignments that optimizes the model's performance metrics is selected as the final weight coefficients. The sum of the weights for the first, second, and third node representations is 1.

[0069] S35 Extracts Interaction Relationships: The second data is used as the input to the graph neural network sub-model, and the third data is output.

[0070] S36 Interaction Matrix Construction: This includes S351 calculating representation similarity and S352 constructing the interaction matrix.

[0071] S351 calculates representation similarity: Extract the representation of each node from the third data output by the graph neural network sub-model, and calculate the similarity between the representations of each node in the third data using methods such as cosine similarity, Euclidean distance, and Manhattan distance. This similarity is denoted as the first similarity.

[0072] S352 Constructing the Interaction Matrix: An interaction threshold is pre-set. When the first similarity is greater than the interaction threshold, it indicates that the representations of the two nodes are interactive; otherwise, they are not. Based on the judgment results of the first similarity and the interaction threshold, an interaction matrix is ​​constructed between each node, and the interaction matrix and the third data are used as new third data.

[0073] S37 Second Processing: The third data is time-aligned with the historical data and the corresponding damage type labels. Then, the third data is used to perform a weighted summation of the corresponding features of the historical data to obtain the first weighted data. The first weighted data is classified according to the damage type labels corresponding to the historical data. The first weighted data with the same damage type are matched to obtain the fourth data.

[0074] S38 Damage Pattern Ranking: Using the fourth data point as input to the regression prediction sub-model, the model inferences and outputs the damage pattern impact ranking results of historical data.

[0075] S4 Damage Pattern Recognition: Real-time data of the pressure vessel is input into the trained pressure vessel damage recognition model. The model inference is sequentially passed through the random forest sub-model, graph neural network sub-model and regression prediction sub-model. The damage pattern impact ranking of the real-time data is output, and the damage pattern with the highest impact ranking is taken as the main damage pattern of the real-time data. S5 Damage Assessment: Determines whether the main damage mode of real-time data is time-dependent. If so, then perform the first check in S6.

[0076] If not, then perform the second test (S7).

[0077] In this embodiment, time-related damage modes include, but are not limited to, all damage modes in corrosion thinning (including hydrochloric acid corrosion, sulfuric acid corrosion, hydrofluoric acid corrosion, phosphoric acid corrosion, carbon dioxide corrosion, naphthenic acid corrosion, phenol corrosion, low molecular weight organic acid corrosion, high-temperature oxidative corrosion, atmospheric corrosion (without insulation layer), atmospheric corrosion (with insulation layer), cooling water corrosion, soil corrosion, microbial corrosion, boiler condensate corrosion, alkali corrosion, ash corrosion, smoke leak corrosion, ammonium chloride corrosion, amine corrosion, high-temperature sulfide corrosion (hydrogen-free environment), high-temperature sulfide corrosion (hydrogen environment), acidic water corrosion (alkaline acidic water), acidic water corrosion (acidic acidic water), methylammonium corrosion, galvanic corrosion, salt water corrosion, oxygen-containing process water corrosion, and concentration cell corrosion), spheroidization, graphitization, and temper embrittlement in material deterioration, mechanical fatigue, vibration fatigue, contact fatigue, thermal fatigue (including ratcheting effect), and creep in mechanical damage, as well as high-temperature hydrogen corrosion in other types of damage. If the primary damage mode of the real-time data belongs to one of the aforementioned damage modes, then the primary damage mode of the real-time data is determined to be time-dependent, and the expected lifetime correlation judgment (S51) is performed. If the primary damage mode of the real-time data does not belong to one of the aforementioned damage modes, then the primary damage mode of the real-time data is determined to be time-independent, and the second test (S7) is performed.

[0078] S51 Expected Life Related Judgment: Determine whether the main damage mode of real-time data is related to the expected life of the equipment, including the first judgment.

[0079] S511 First Judgment: Determine whether the main damage mode of the real-time data belongs to one of the following damage modes: corrosion thinning, mechanical fatigue, vibration fatigue, contact fatigue, or thermal fatigue (including ratcheting effect). If yes, proceed to S6 First Test; if no, proceed to S512 Second Judgment.

[0080] S512 Second Judgment: Determine whether the main damage mode of the real-time data belongs to high-temperature hydrogen corrosion. If not, proceed to S513 Third Judgment; if yes, continue to determine whether the pressure vessel material is carbon steel, molybdenum steel, or chromium-molybdenum steel. If the pressure vessel material is any one of carbon steel, molybdenum steel, or chromium-molybdenum steel, proceed to S6 First Inspection; otherwise, proceed to S7 Second Inspection.

[0081] S513 Third Judgment: Determine whether the main damage mode of the real-time data belongs to either spheroidization or graphitization damage mode. If not, proceed to S514 Fourth Judgment; if yes, proceed to S5131 to determine the service temperature.

[0082] S5131 Determine Service Temperature: Determine whether the material and service temperature of the pressure vessel meet any one of the following four requirements.

[0083] Carbon steel pressure vessels have an operating temperature below 375℃; The service temperature of chromium molybdenum steel pressure vessels is below 425℃; The service temperature of chromium molybdenum vanadium steel pressure vessels is below 475℃; The service temperature of austenitic stainless steel pressure vessels is below 525℃.

[0084] If satisfied, perform the second test (S7); if not satisfied, perform test (S5132) to determine the spheroidization and graphitization levels.

[0085] S5132 Determine the spheroidization and graphitization level: Determine whether the spheroidization and graphitization level of the pressure vessel meets any one of the following three requirements.

[0086] Carbon steel pressure vessels and low alloy steel pressure vessels with a original microstructure of pearlite have a spheroidization level of 3 or above. Chromium-molybdenum steel pressure vessels with original metallographic structure of bainite are spheroidized to level 4 or above; Pressure vessels are graphitized to level 3 or above.

[0087] If satisfied, proceed with S5133 to determine the inferiority trend; if not satisfied, proceed with S7, the second test.

[0088] S5132 Determine Deterioration Trends: Determine whether the past inspections of the current pressure vessel have shown obvious spheroidization or graphitization deterioration trends. If yes, proceed to S6 First Inspection; if no, proceed to S514 Fourth Determination.

[0089] S514 Fourth Judgment: Determine whether the main damage mode of the real-time data belongs to the tempering embrittlement damage mode. If not, proceed to S515 Fifth Judgment; if yes, continue to determine whether the material of the current pressure vessel is Cr-Mo steel and the actual operating temperature is 345℃-595℃. If yes, proceed to S6 First Inspection; if not, proceed to S515 Fifth Judgment.

[0090] S515 Fifth Judgment: Determine whether the main damage mode of the real-time data belongs to the creep damage mode. If so, calculate the current evaluation temperature of the pressure vessel. If the current evaluation temperature of the pressure vessel is less than the current critical control temperature of the pressure vessel, then execute S7 Second Test; otherwise, execute S6 First Test.

[0091] S6 First Inspection: Several special inspection schemes are pre-set. Select the special inspection scheme corresponding to the main damage mode of the real-time data. Based on the selected special inspection scheme, inspect the current pressure vessel, output the inspection results, and perform the steps of safety status level assessment.

[0092] According to Article 8.3 of the "Safety Technical Supervision Regulations for Fixed Pressure Vessels," a general inspection plan should be pre-established. This general inspection plan includes macroscopic inspection, inspection of insulation layers, linings, and weld overlays; inspection of vacuum-insulated pressure vessels; wall thickness measurement; surface defect detection; buried defect detection; material analysis; inspection of pressure vessels where internal inspection is not possible; stud inspection; strength verification; inspection of safety accessories; pressure resistance testing; and leakage testing.

[0093] Several special testing procedures were established, including the following: 1. For pressure vessels exhibiting corrosion thinning damage modes, it is necessary to determine the severely corroded areas and corrosion types based on previous wall thickness measurements of the pressure vessel: a) For uniform corrosion, the uniform wall thickness of each cylinder section and each plate of the end cap should be measured at no less than 5 points in severely corroded parts.

[0094] b) For localized corrosion, the shape and size of the severely corroded area should be determined, and a grid should be drawn within the severely corroded area. The grid spacing depends on the area of ​​the corroded area, generally 20mm-50mm, and the wall thickness of each intersection falling within the corroded area should be measured.

[0095] c) For dispersed pitting corrosion types, the pitting corrosion depth in severely corroded areas should be measured in detail.

[0096] 2. For pressure vessels with spheroidization damage mode, hardness testing and surface defect testing of no less than 20% should be carried out on high-temperature parts, stress concentration parts, deformation zones, etc. If the hardness value is significantly reduced, metallographic testing should be carried out, and if necessary, a serviceability evaluation should be conducted.

[0097] 3. For pressure vessels with graphitization damage mode, metallographic testing and surface defect testing of no less than 30% should be carried out on high-temperature parts, stress concentration parts, deformation zones, etc., and a serviceability evaluation should be conducted if necessary.

[0098] 4. For pressure vessels with temper embrittlement damage mode, historical inspection results should be reviewed, and surface defects and buried defects in welds should be detected, with a proportion of not less than 30% (the detection area should overlap with the area of ​​the last periodic inspection by not less than 10% of the total weld length). If necessary, a serviceability evaluation should be conducted.

[0099] 5. For pressure vessels with mechanical fatigue or vibration fatigue damage modes, 100% surface defect detection should be carried out at stress concentration points and suspected vibration points, and vibration monitoring or service evaluation should be conducted when necessary.

[0100] 6. For pressure vessels with thermal fatigue (including thermal ratchet) damage modes, macroscopic inspection should focus on whether the structure has undergone obvious deformation. If necessary, use technologies such as laser total station and laser automatic scanner 3D imaging to detect the amount of deformation. Also, 100% surface defect detection should be carried out at stress concentration points and areas with large local temperature gradients. If necessary, a suitability for use evaluation should be conducted.

[0101] 7. For pressure vessels with contact fatigue damage modes, macroscopic inspection should focus on whether there are pits or dents on the contact surfaces, and 100% surface defect detection should be carried out on stress concentration areas and contact surfaces. If necessary, a serviceability evaluation should be conducted.

[0102] 8. For pressure vessels exhibiting creep damage, 100% surface and buried defect inspection should be conducted on areas with localized high temperatures, stress concentrations, welded joints, areas with excessive defects or that have undergone repairs, and the heat-affected zones of unequal-thickness welded joints, dissimilar steel welded joints, and T-joints within high-temperature areas. Simultaneously, based on stress analysis, hardness and metallographic testing should be performed to comprehensively analyze the degree of creep damage, and a serviceability evaluation should be conducted if necessary.

[0103] 9. For pressure vessels with high-temperature hydrogen corrosion damage mode, historical inspection results should be reviewed, and surface defects and buried defects in welds on the inner surface should be detected, with a proportion of not less than 50% (the inspection area should overlap with the total weld length of the area inspected in the last periodic inspection). Buried defects in the equipment base material should be randomly checked, and hardness and metallographic examinations should be conducted in stress concentration areas. For lined pressure vessels, macroscopic inspection should include random checks for lining peeling and detachment. If peeling or detachment is found, the sampling proportion should be expanded, and underlayer crack detection should be carried out. If necessary, a serviceability evaluation should be conducted.

[0104] S7 Second Inspection: A general inspection plan is pre-set according to Section 8.3 of the "Safety Technical Supervision Regulations for Fixed Pressure Vessels." This general inspection plan includes macroscopic inspection, inspection of insulation layers, linings, and weld overlays; inspection of vacuum-insulated pressure vessels; wall thickness measurement; surface defect detection; buried defect detection; material analysis; inspection of pressure vessels where internal inspection is not possible; stud inspection; strength verification; safety accessory inspection; pressure resistance test; and leakage test. The pressure vessel is inspected according to the general inspection plan, the inspection results are output, and the S8 safety status level assessment is performed.

[0105] S8 Safety Status Level Assessment: Based on the inspection results, the current safety status level of the pressure vessel is assessed in accordance with Section 8.5 of the "Safety Technical Supervision Regulations for Fixed Pressure Vessels". For matters not specified in Section 8.5 of the "Safety Technical Supervision Regulations for Fixed Pressure Vessels", the inspection agency shall make a comprehensive determination of the safety status level. 1) It should not be rated as Level 1.

[0106] 2) If the overall safety status is assessed as level 2 or 3, the inspection conclusion is that it meets the requirements and can continue to be used.

[0107] 3) If the overall safety status is assessed as Level 4, the inspection conclusion is that it basically meets the requirements and can be used for monitoring under certain conditions.

[0108] 4) If the overall safety status is rated as level 5, the inspection conclusion is that it does not meet the requirements and should not be used.

[0109] In this embodiment, a pressure vessel damage identification model is constructed, comprising a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model. The random forest sub-model is used to identify pressure vessel damage patterns, the graph neural network sub-model provides the interaction relationships between damage patterns, and the regression prediction model analyzes the impact of damage patterns on the equipment. This helps to capture the damage characteristics of pressure vessels under different operating conditions, identify the damage patterns of pressure vessels, and rank these patterns by their impact, achieving the technical effect of quickly locating the main damage patterns and improving the accuracy and reliability of pressure vessel safety assessment.

[0110] Example 2: This example discloses a system for evaluating the extended service life of pressure vessels beyond their design life. The system includes: The data acquisition module includes a real-time data acquisition module and a historical data acquisition module.

[0111] The real-time data acquisition module is used to acquire real-time data from the pressure vessel.

[0112] The historical data acquisition module is used to collect historical data of pressure vessels and the corresponding damage type labels.

[0113] The model building module is used to build a pressure vessel damage identification model, which includes a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model.

[0114] The model training module communicates with the data acquisition module and the model building module, and includes a random forest sub-model training unit, a damage pattern identification unit, a first processing unit, an interaction relationship extraction unit, a second processing unit, and a damage pattern sorting unit.

[0115] The random forest sub-model training unit includes a working condition extraction sub-unit, a label hierarchy partitioning sub-unit, a label encoding sub-unit, a classifier construction sub-unit, a first training sub-unit, a second training sub-unit, and an output merging sub-unit.

[0116] The operating condition extraction subunit is used to extract operating condition data from historical data and obtain operating condition category labels.

[0117] The label hierarchy sub-unit is used to hierarchically divide the damage types corresponding to historical data into top-level labels and corresponding sub-level labels, and match the corresponding working condition category labels to the top-level labels and corresponding sub-level labels.

[0118] The tag encoding subunit is used to take the top-level tag, the corresponding sub-tag, and the corresponding working condition category tag of each damage type tag as a group of tags, denoted as the damage type tag group. The damage type tag group is encoded to obtain the top-level encoded tag and the corresponding sub-level encoded tag of each damage type tag, denoted as the damage type tag encoding group of the damage type tag.

[0119] Construct a classifier subunit to build a multi-classifier for each top-level label and a binary classifier for each top-level label's child labels.

[0120] The first training subunit is used to input historical data into the multi-classifier for training and predict the top-level encoding label of the damage type of the current historical data, denoted as the first label.

[0121] The second training subunit is used to input the first label into the binary classifier for training and predict the sub-label of the damage type of the current historical data, denoted as the second label.

[0122] The output merge sub-unit is used to concatenate the first label and the second label, and the concatenated data is used as the output of the random forest sub-model.

[0123] The damage pattern identification unit is communicatively connected to the random forest sub-model training unit and the first processing unit. It is used to take the collected historical data of the pressure vessel and the damage type labels corresponding to the historical data as input to the random forest sub-model and output the first data.

[0124] The first processing unit is used to match the first data with the historical data of the pressure vessel to obtain the second data.

[0125] Extract the interaction relationship unit, which is used to take the second data as input to the graph neural network sub-model and output the third data.

[0126] The second processing unit is used to match the third data with historical data and corresponding damage type labels to obtain the fourth data.

[0127] The damage pattern ranking unit is used to take the fourth data as input to the regression prediction sub-model and output the damage pattern impact ranking results of historical data. The damage pattern recognition unit communicates with the data acquisition module and the model building module, inputs the real-time data of the pressure vessel into the trained pressure vessel damage recognition model, outputs the damage pattern impact ranking results of the real-time data, and takes the damage pattern with the highest impact ranking as the main damage pattern of the real-time data.

[0128] The damage discrimination unit is used to determine whether the main damage mode of real-time data is time-dependent. If the main damage pattern in the real-time data is time-dependent, then the first inspection unit is triggered.

[0129] If the main damage pattern in the real-time data is independent of time, then the second test unit is triggered.

[0130] The first inspection unit is connected to the damage discrimination unit. It is used to pre-set several special inspection schemes, select the special inspection scheme corresponding to the main damage mode of the real-time data, inspect the current pressure vessel according to the selected special inspection scheme, output the inspection results, and trigger the safety status level assessment unit.

[0131] The second inspection unit is connected in communication with the damage assessment unit. It is used to pre-set a general inspection plan, inspect the current pressure vessel according to the general inspection plan, output the inspection results, and perform the steps of safety status level assessment.

[0132] The safety status assessment unit is connected in communication with the first inspection unit and the second inspection unit. It is used to assess the current safety status level of the pressure vessel based on the inspection results and evaluate the results of the current pressure vessel's extended service life.

[0133] In this embodiment, the model building module constructs a pressure vessel damage identification model that includes a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model. This helps to capture the damage characteristics of pressure vessels under different operating conditions, identify the damage patterns of pressure vessels, and rank these patterns by impact. This achieves the technical effect of quickly locating the main damage patterns, improves the accuracy and reliability of pressure vessel safety assessment, and provides strong technical support for the continued use of pressure vessels beyond their design service life.

[0134] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for evaluating the extended service life of pressure vessels beyond their design life, characterized in that, The method includes: Data acquisition includes collecting real-time data and collecting historical data; Real-time data acquisition: Acquire real-time data from the pressure vessel; Collect historical data: Collect historical data of pressure vessels and corresponding damage type labels; Model Construction: A pressure vessel damage identification model is constructed, which includes a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model. Model training: Damage pattern identification, first processing, interaction extraction, second processing, and damage pattern ranking; Damage pattern identification: The historical data of the collected pressure vessel and the corresponding damage type labels are used as input to the random forest sub-model, and the first data is output. First processing: Match the first data with the historical data of the pressure vessel to obtain the second data; Extracting interaction relationships: The second data is used as the input to the graph neural network sub-model, and the third data is output. The second process involves matching the third data with historical data and corresponding damage type labels to obtain the fourth data. Damage pattern ranking: The fourth data point is used as the input to the regression prediction sub-model, and the output is the ranking of the impact of damage patterns in the historical data. Damage pattern recognition: Input the real-time data of the pressure vessel into the trained pressure vessel damage recognition model, output the damage pattern impact ranking of the real-time data, and take the damage pattern with the highest impact ranking as the main damage pattern of the real-time data. Damage assessment: Determine whether the main damage mode of real-time data is time-dependent. If so, proceed with the first check step; If not, proceed to the second test step; First inspection: Several special inspection schemes are pre-set, select the special inspection scheme corresponding to the main damage mode of the real-time data, inspect the current pressure vessel according to the selected special inspection scheme, output the inspection results, and perform the steps of safety status level assessment. Second inspection: A general inspection plan is set in advance, the current pressure vessel is inspected according to the general inspection plan, the inspection results are output, and the steps of safety status level assessment are performed; Safety Status Assessment: Based on the inspection results, assess the current safety status of the pressure vessel and evaluate the outcome of extending its service life.

2. The method for evaluating the extended service life of pressure vessels beyond their design life as described in claim 1, characterized in that, After performing the model building step and before performing the damage pattern identification step, the following steps are also included: Operating condition extraction: Extract operating condition data from historical data to obtain operating condition category labels; Label hierarchy division: The damage types corresponding to historical data are hierarchically divided into top-level labels and corresponding sub-level labels, and the corresponding working condition category labels are matched for the top-level labels and corresponding sub-level labels. Tag encoding: The top-level tag, the corresponding sub-tags, and the corresponding working condition category tag of each damage type tag are taken as a group of tags, denoted as damage type tag group. The damage type tag group is encoded to obtain the top-level encoded tag and the corresponding sub-encoded tag of each damage type tag, denoted as damage type tag encoding group of damage type tag. Build classifiers: Build a multi-classifier for each top-level label and a binary classifier for each child label of the top-level label; First training: Input historical data into the multi-classifier for training, and predict the top-level encoding label of the damage type of the current historical data, which is denoted as the first label; Second training: Input the first label into the binary classifier for training, and predict the sub-label of the damage type of the current historical data, which is denoted as the second label; Output merging: The first label and the second label are concatenated, and the concatenated data is used as the output of the random forest sub-model.

3. The method for evaluating the extended service life of pressure vessels beyond their design life as described in claim 1, characterized in that, After performing the first processing step and before performing the step of extracting interaction relationships, the process also includes: Define the graph structure: Take the collection time of each set of historical data in the historical data as a node, construct a node feature matrix, and construct a time relationship adjacency matrix based on the collection order and collection time interval of each set of historical data in the historical data, which is denoted as the first adjacency matrix. The nodes and the first adjacency matrix are denoted as the first graph structure. Third training: Input historical data into the graph neural network sub-model, perform graph convolution propagation based on the first graph structure to obtain the first node representation, which is used as the output of the graph neural network sub-model; Model optimization: Define a first loss function, minimize the first loss function to optimize the graph neural network sub-model, and obtain the optimized graph neural network sub-model, which is denoted as the new graph neural network sub-model.

4. The method for evaluating the extended service life of pressure vessels beyond their design life according to claim 3, characterized in that, After performing the third training step and before performing the model optimization step, the following steps are also included: Add a time decay factor: Adjust the weight of the relationship between two nodes based on the historical data collection time interval. ； in, Represents a node and Relationship weights Indicates the attenuation factor. Represents a node The collection time, Represents a node The collection time.

5. The method for evaluating the extended service life of pressure vessels beyond their design life according to claim 3, characterized in that, After the step of defining the graph structure and before the step of the third training, the following steps are also included: Second adjacency matrix: Calculate the feature similarity of each group of historical data in the historical data, and construct a feature adjacency matrix using the feature similarity, denoted as the second adjacency matrix; The third adjacency matrix: Traverse the historical data and the corresponding damage type labels, count the frequency of co-occurrence of the same damage type label among each group of historical data, and denote it as the first frequency. Construct a damage pattern co-occurrence adjacency matrix based on the first frequency, and denote it as the third adjacency matrix.

6. The method for evaluating the extended service life of pressure vessels beyond their design life according to claim 3, characterized in that, After performing the third training step and before performing the model optimization step, the following steps are also included: Fourth training: Denote the node and the second adjacency matrix as the second graph structure, input historical data into the graph neural network sub-model, perform graph convolution propagation based on the second graph structure, and obtain the second node representation; Fifth training: Denote the node and the third adjacency matrix as the third graph structure, input historical data into the graph neural network sub-model, perform graph convolution propagation based on the third graph structure, and obtain the third node representation; Relationship fusion: The first node representation, the second node representation, and the third node representation are weighted and accumulated to obtain the fourth node representation, which is used as the new first node representation.

7. The method for evaluating the extended service life of pressure vessels beyond their design life according to claim 3, characterized in that, After performing the third training step and before performing the model optimization step, the following steps are also included: Weight model construction: Based on multilayer perceptron, a weight learning network is constructed; Weight learning: Extract working condition data from historical data to obtain working condition category labels, then import the working condition data into the weight learning network for model inference, and output the working condition weight corresponding to each working condition category label; Weight matching: Match the working condition weights with the first node features, and record them as the new first node features.

8. The method for evaluating the extended service life of pressure vessels beyond their design life according to claim 3, characterized in that, After performing the step of extracting interaction relationships and before performing the second processing step, the process also includes: Calculate representation similarity: Calculate the similarity between the representations of each node in the third data, denoted as the first similarity; Construct the interaction matrix: Based on the first similarity, construct the interaction matrix between each node, and use the interaction matrix and the third data as the new third data.

9. A system for evaluating the extended service life of pressure vessels beyond their design life, characterized in that, The system is applicable to the method as described in any one of claims 1-8, the system comprising: The data acquisition module includes a real-time data acquisition module and a historical data acquisition module; The real-time data acquisition module is used to acquire real-time data from the pressure vessel. The historical data acquisition module is used to collect historical data of pressure vessels and the corresponding damage type labels for the historical data. The model building module is used to build a pressure vessel damage identification model, which includes a random forest sub-model, a graph neural network sub-model, and a regression prediction sub-model. The model training module communicates with the data acquisition module and the model building module, and includes a damage pattern identification unit, a first processing unit, an interaction relationship extraction unit, a second processing unit, and a damage pattern sorting unit. The damage pattern identification unit is used to take the collected historical data of the pressure vessel and the corresponding damage type labels as input to the random forest sub-model and output the first data. The first processing unit is used to match the first data with the historical data of the pressure vessel to obtain the second data; Extract the interaction relationship unit, which is used to take the second data as the input of the graph neural network sub-model and output the third data; The second processing unit is used to match the third data with historical data and corresponding damage type labels to obtain the fourth data; The damage pattern ranking unit is used to take the fourth data as input to the regression prediction sub-model and output the damage pattern impact ranking results of historical data. The damage pattern recognition unit communicates with the data acquisition module and the model building module, inputs the real-time data of the pressure vessel into the trained pressure vessel damage recognition model, outputs the damage pattern impact ranking result of the real-time data, and takes the damage pattern with the first impact ranking as the main damage pattern of the real-time data. The damage discrimination unit is used to determine whether the main damage mode of real-time data is time-dependent. If the main damage pattern in the real-time data is time-dependent, then the first inspection unit is triggered; If the main damage pattern in the real-time data is independent of time, then the second test unit is triggered; The first inspection unit is connected to the damage discrimination unit. It is used to pre-set several special inspection schemes, select the special inspection scheme corresponding to the main damage mode of the real-time data, inspect the current pressure vessel according to the selected special inspection scheme, output the inspection results, and perform the steps of safety status level assessment. The second inspection unit is connected in communication with the damage judgment unit. It is used to pre-set a general inspection plan, inspect the current pressure vessel according to the general inspection plan, output the inspection results, and perform the steps of safety status level assessment. The safety status assessment unit is connected in communication with the first inspection unit and the second inspection unit. It is used to assess the current safety status level of the pressure vessel based on the inspection results and evaluate the results of the current pressure vessel's extended service life.

Images