Detection scheme generation method and system for special equipment, equipment and medium

Abstract

The invention provides a special equipment detection scheme generation method and system, equipment and a medium, and belongs to the technical field of special equipment and artificial intelligence application, and the method comprises the steps: collecting data of target equipment to construct a data file, employing LSTM, CNN and Transformer encoders to extract time sequence, image and text features respectively, and fusing the features to obtain a detection scheme; intelligent evaluation of the equipment is realized; based on the evaluation result, generating an initial test scheme in combination with rule constraint and historical case recommendation; performing simulation prediction on the scheme through a digital twinborn environment, and performing multi-objective optimization by adopting reinforcement learning to obtain a safe, economic and efficient-balanced optimal test scheme; and finally, outputting the scheme, and performing feedback updating on the evaluation model according to an actual test result to form a closed-loop learning mechanism. According to the method, the test scheme is converted from experience driving to data intelligent driving, and the capabilities of personalized formulation, look-ahead optimization and sustainable evolution are improved.

Description

Technical Field

[0001] This invention belongs to the field of special equipment and artificial intelligence application technology, specifically relating to a method, system, equipment and medium for generating detection schemes for special equipment. Background Technology

[0002] Special equipment (such as pressure vessels, boilers, pressure pipelines, elevators, lifting machinery, etc.) is an important infrastructure for the national economy and people's lives. Its safe operation is of paramount importance, and the inspection of special equipment is the standard for ensuring safe operation.

[0003] Traditional testing and inspection solutions rely heavily on the personal experience of testing personnel and their interpretation of mandatory national regulations. This results in a situation where massive amounts of historical equipment operation data, online monitoring data, and testing records are in "information silos," failing to be effectively integrated and used to predict equipment performance degradation trends. Furthermore, the experience of senior experts is difficult to quantify and solidify into the solution development process.

[0004] To break down information silos, an equipment management information system (such as EAM), an electronic inspection report database, and an online monitoring data platform were established, enabling centralized digital storage and retrieval of equipment files, historical records, and some real-time data. However, while the system primarily addresses data access, it lacks in-depth analysis, intelligent correlation, and decision support capabilities. The generation of solutions still requires inspection personnel to manually extract information from massive amounts of data and rely on experience-based judgment. It has not substantially transformed data into "fuel" to drive decision-making, and the value of the data has not been intelligently applied, thus limiting the foresight and accuracy of inspection solutions.

[0005] To overcome subjective biases and solidify some expert experience, a rule-based decision support system was developed. This system encodes important regulatory provisions and explicit expert judgment logic (such as "If the medium is extremely hazardous, the testing period is shortened to X years") into computer-executable rules. While this rule engine improves the standardization and consistency of solutions, it sacrifices necessary flexibility and adaptability. It lacks an understanding and ability to adapt to the complexity and dynamism of the real world, and it does not possess the advanced intelligence to continuously learn from data and proactively optimize under constraints. Therefore, it cannot generate truly forward-looking, accurate, and economical dynamically optimal solutions. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method, system, equipment, and medium for generating detection schemes for special equipment.

[0007] To achieve the above objectives, the present invention provides a method for generating a detection scheme for special equipment, comprising: The system collects the inherent settings attributes, dynamic operating data, historical diagnostic reports, and regulatory standard data of the target special equipment; the dynamic operating data includes the temperature and pressure curves of each component during equipment operation, and ultrasonic and X-ray images of each component.

[0008] Extract the temporal variation features of temperature and pressure curves of each component, extract the defect morphology features of ultrasonic and X-ray images of each component, and extract the textual semantic features of historical diagnostic reports of each component; dynamically weight and fuse the temporal variation features, defect morphology features, and textual semantic features to form a unified equipment status feature vector.

[0009] Using the device state feature vector, the safety assessment score of each component of the target device is obtained; based on the safety assessment score, inherent setting attributes and regulatory standard data constraints, an initial testing plan is generated that includes the inspection items, recommended inspection methods, key inspection areas and suggested inspection cycles for each component.

[0010] The initial detection scheme is simulated and executed in the pre-defined digital twin system. A reinforcement learning algorithm is used to construct a reward function with security, economy and efficiency as comprehensive optimization objectives. The parameters of the initial detection scheme are solved, and the optimal detection scheme that satisfies the minimum detection cost and the shortest detection time is output.

[0011] Preferably, the reinforcement learning algorithm is used to construct a reward function with safety, economy, and efficiency as comprehensive optimization objectives, and to solve for the parameters of the initial detection scheme. Specifically, the reinforcement learning algorithm designs the reward function with maximizing safety, minimizing detection cost, and minimizing detection time as its core. The detection scheme parameters are adjusted iteratively through interaction between the agent and the digital twin system. After iterating until the reward value converges, the Pareto optimal detection scheme is output. The detection scheme parameters include the combination of detection methods, detection order, detection range, and operation parameters.

[0012] Preferably, a Long Short-Term Memory (LSTM) network is used to extract the temporal variation features of the temperature and pressure curves of each component, a Convolutional Neural Network (CNN) is used to extract the defect morphology features of the ultrasound and X-ray images of each component, and a Transformer encoder is used to extract the textual semantic features of the historical diagnostic reports of each component. An attention mechanism is used to dynamically weight and fuse the temporal variation features, defect morphology features, and textual semantic features to form a unified equipment status feature vector.

[0013] Preferably, after outputting the optimal detection scheme that satisfies the minimum detection cost and the shortest detection time, the method further includes: the digital twin system generating a prediction result of the optimal detection scheme; the optimal detection scheme being executed on-site by the target device, and the actual detection result data being collected; and the reinforcement learning being updated by using the difference between the actual detection result data and the prediction result.

[0014] Preferably, the inherent attribute parameters include design parameters, manufacturing information, material grade, structural drawings, and service life; dynamic operating data includes time-series parameters for pressure, temperature, flow rate, vibration, acoustic emission, and corrosion monitoring; historical diagnostic reports include previous test reports, defect records, maintenance history, and original data and images from non-destructive testing; and regulatory and standard data include structured TSG series procedures, GB / T national standards, and industry standards.

[0015] Preferably, the safety assessment score includes a comprehensive evaluation of component risk level, health score, and potential failure probability.

[0016] Preferably, an initial testing plan is generated by using a rule engine and a case recommendation system, which includes the testing items for each component, recommended testing methods, key testing areas, and suggested testing cycles. The rule engine ensures the compliance of the initial testing plan, and the case recommendation system retrieves the most similar successful cases from the historical case library, draws on their solutions, and generates a preliminary testing plan adapted to the current scenario.

[0017] The present invention also provides a special equipment testing scheme generation system, comprising: The data acquisition module is used to collect the inherent setting attributes, dynamic operating data, historical diagnostic reports, and regulatory standard data of the target special equipment; the dynamic operating data includes the temperature and pressure curves of each component during equipment operation, and ultrasonic and X-ray images of each component.

[0018] The feature extraction module is used to extract the temporal variation features of the temperature and pressure curves of each component, extract the defect morphology features of the ultrasonic and X-ray images of each component, and extract the textual semantic features of the historical diagnostic reports of each component. The temporal variation features, defect morphology features, and textual semantic features are dynamically weighted and fused to form a unified equipment status feature vector. Using the equipment status feature vector, the safety assessment score of each component of the target equipment is output. Based on the safety assessment score, inherent setting attributes, and constraints of regulatory and standard data, an initial testing plan is generated, which includes the inspection items, recommended inspection methods, key inspection areas, and suggested inspection cycles for each component.

[0019] The application module is used to simulate the execution of the initial detection scheme in a preset digital twin system. It uses a reinforcement learning algorithm to construct a reward function with security, economy and efficiency as comprehensive optimization objectives, solves the parameters of the initial detection scheme, and outputs the optimal detection scheme that satisfies the minimum detection cost and the shortest detection time.

[0020] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in the method for generating a detection scheme for the special equipment.

[0021] The present invention also provides a computer-readable storage medium storing a computer program, which, when loaded by a processor, is capable of executing any of the steps in the method for generating a detection scheme for special equipment.

[0022] The invention provides a method for generating detection schemes for special equipment, which has the following advantages: First, it utilizes LSTM, CNN, and Transformer neural networks to perform parallel deep feature extraction on multimodal data and achieves dynamic fusion through an attention mechanism, replacing the shallow logic of manual experience judgment and rule engines, and realizing data-driven cognition and quantitative evaluation of complex equipment states. Second, it simulates and predicts the effects of the generated personalized initial scheme in a digital twin environment, and uses reinforcement learning algorithms with safety, economy, and efficiency as comprehensive optimization goals. In the virtual environment, it autonomously iterates and dynamically optimizes the scheme parameters, thereby outputting an inspection scheme that achieves optimal cost and time while meeting safety constraints. This allows for accurate prediction and quantitative verification of the expected effects of the detection scheme, avoiding resource misallocation or unsatisfactory results caused by the inability to perform simulations in traditional methods. Ultimately, it achieves a fundamental leap from static experience-based decision-making to dynamic intelligent optimization decision-making. Attached Figure Description

[0023] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a method for generating a detection scheme for special equipment according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the method for generating inspection and testing schemes according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the data structure of the multi-source data fusion module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the deep learning model in the equipment safety status assessment module of this invention. Figure 5 This is a schematic diagram of the workflow of the scheme simulation and optimization module in an embodiment of the present invention; Figure 6 A system architecture diagram for generating the inspection and testing scheme in this embodiment of the invention is provided. Detailed Implementation

[0025] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0026] like Figure 2 As shown, this invention first acquires equipment-related data through multi-source data acquisition. After data cleaning and fusion processing, if the data is normal, a deep learning model is used to analyze and complete the equipment safety status assessment, thereby generating an initial inspection plan. Subsequently, the effectiveness of the plan is predicted through digital twin simulation, and then optimized through reinforcement learning to determine whether the target is met. If not, the optimization is repeated; if it is met, the plan is advanced to the field for inspection. Finally, feedback data is collected to update and train the model, and the final detection plan is output.

[0027] Based on this, the present invention provides a method for generating inspection schemes for special equipment, and particularly relates to an intelligent method for generating inspection and testing schemes for special equipment that integrates digital twin technology and AI deep learning, specifically as follows: Figure 1 As shown, it includes: S1. Collect the inherent settings attributes, dynamic operation data, historical diagnostic reports, and regulatory standard data of the target special equipment; the dynamic operation data includes the temperature and pressure curves of each component during equipment operation, and ultrasonic and X-ray images of each component.

[0028] The multi-data source fusion module acquires data from various data sources for the regenerator. Static properties include: design pressure 0.35 MPa, material Q345R, and 12 years of service history. Dynamic data includes: temperature cycling records from the past year (200-720℃). Historical inspection data includes: ultrasonic (UT) testing three years ago revealed slight thinning of the inner wall (-0.5 mm). Real-time data includes: the acoustic emission monitoring system detected an increase in recent activity signals.

[0029] like Figure 3As shown, the multi-data source fusion module clearly divides equipment static attribute data, equipment dynamic operation data, historical inspection data, and regulatory and standard data into four core data categories. Among them, equipment static attribute data includes equipment model, design parameters, material information, and manufacturing information; equipment dynamic operation data covers pressure, temperature, vibration monitoring, acoustic emission monitoring, and corrosion monitoring data; historical inspection data involves annual inspection reports, equipment defect history, historical maintenance records, and annual inspection records; and regulatory and standard data encompasses TSG safety technical regulations, GB / T series national standards, NB / T series industry standards, and other relevant standards and specifications. This comprehensive collection of multi-dimensional data required for special equipment testing lays the data foundation for subsequent intelligent equipment assessment and testing solution generation.

[0030] S2. Extract the temporal variation characteristics of the temperature and pressure curves of each component, extract the defect morphology characteristics of the ultrasonic and X-ray images of each component, and extract the textual semantic features of the historical diagnostic reports of each component; dynamically weight and fuse the temporal variation characteristics, defect morphology characteristics, and textual semantic features to form a unified equipment status feature vector; using the equipment status feature vector, output the safety assessment score of each component of the target equipment; based on the safety assessment score, inherent setting attributes, and constraints of regulatory standard data, generate an initial testing plan that includes the inspection items, recommended inspection methods, key inspection areas, and suggested inspection cycles for each component.

[0031] The equipment safety status assessment module is activated. An LSTM network analyzes temperature cycling data and identifies thermal fatigue as the primary damage mode. A CNN network analyzes historical UT thickness measurement data to quantify the corrosion rate. After all features are fused using an attention mechanism, the multi-task model outputs: a medium risk level, a health score of 72, and potential failure modes including thermal fatigue cracking (65% probability) and uniform corrosion (30% probability).

[0032] like Figure 4As shown, the deep learning model structure in the special equipment safety status assessment module is divided into an input layer, a feature fusion layer, and an output layer from top to bottom. The input layer uses dedicated models to process four types of core data: an LSTM network to process time-series operational data such as pressure and temperature to extract time-series features; a CNN convolutional neural network to process historical inspection data such as UT and RT images to extract image features; a Transformer encoder to process text report data such as historical inspection reports to extract semantic features; and an FCL fully connected embedding layer to process structured attribute data such as equipment design parameters to complete feature encoding. The feature fusion layer dynamically weights and concatenates multi-source data features through an attention mechanism to form a unified equipment status feature vector. The output layer uses an MLP multilayer perceptron to achieve multi-task output, ultimately generating the equipment's risk level, health score, failure mode, and failure probability, providing accurate safety assessment basis for subsequent detection scheme generation.

[0033] The personalized inspection plan generation module was triggered. The rule engine confirmed that periodic inspections must be carried out in accordance with TSG21-2016 "Safety Supervision Regulations for Stationary Pressure Vessels" and GB150-2024 "Pressure Vessels". Based on the characteristics of "medium risk" and "thermal fatigue", the recommendation system matched similar cases from the knowledge base and generated a preliminary plan. The key inspection areas are the regenerator lower head and nozzle welds. The inspection methods are 100% phased array ultrasonic testing (PAUT) and metallographic inspection. The recommended cycle is immediate inspection.

[0034] Component risk level quantification: A 1-5 point system is used (1 point for low risk, 5 points for high risk), calculated based on weighted scores from secondary factors. Taking the main beam as an example, the component's functional importance is scored 4 points (core load-bearing component of the crane), the severity of failure consequences is scored 4.5 points (failure may lead to a major safety accident), and the severity of the operating environment is scored 3 points (outdoor operation, exposed to wind and rain). Therefore, the risk level score of the main beam component = 4 × 20% + 4.5 × 12% + 3 × 8% = 0.8 + 0.54 + 0.24 = 1.58 points, corresponding to a medium-high risk level.

[0035] Health score quantification: A 0-100 point scale is used, calculated based on data mapping of each dimension of the equipment status feature vector. Taking the drum as an example, based on data such as vibration amplitude feature value 0.68, defect area feature value 0.43, and wear rate feature value 0.39, the scores of each secondary factor are calculated through a preset mapping function: wear degree 72 points, defect condition 81 points, operational stability 78 points, and maintenance quality 85 points. The drum health score = 72×12%+81×10%+78×8%+85×5%=8.64+8.1+6.24+4.25=27.23 points (this is a weighted score; the final health score needs to be weighted and included in the total score).

[0036] Potential failure probability quantification: A 0-100% range quantification is used, and the failure probability prediction model is calculated by fitting historical data. Taking steel wire rope as an example, combined with data such as material fatigue characteristic value of 0.75, running time characteristic value of 0.88, and historical failure frequency characteristic value of 0.62, the potential failure probability is calculated to be 18.5% by substituting them into the model.

[0037] The overall safety assessment score is calculated as follows: Component Risk Level Score × 40% + Health Score × 35% + Potential Failure Probability × 25% (Note: Potential failure probability needs to be converted to a 0-5 point scale for calculation; the conversion formula is: 5 × Potential Failure Probability). Taking the hook assembly as an example, with a component risk level score of 1.8, a health score of 75, and a potential failure probability of 12% (converted to 0.6), the overall safety assessment score for the hook assembly is: 1.8 × 40% + 75 × 35% + 0.6 × 25% = 0.72 + 26.25 + 0.15 = 27.12 points (out of 50; a higher score indicates a worse safety condition).

[0038] Safety assessment score constraints: High-risk components (score ≥ 30 points) require additional inspection items and a shorter inspection cycle; medium-to-high-risk components (25 points ≤ score < 30 points) require enhanced inspection of key items; medium-risk components (20 points ≤ score < 25 points) should be inspected according to routine requirements; and medium-to-low-risk components (score < 20 points) may have their inspection procedures appropriately simplified.

[0039] Inherent setting attribute constraints: Combine the inherent attributes of the components (e.g., if the wire rope is made of high-strength alloy steel, fatigue damage needs to be tested in detail), structural design (e.g., if the main beam is a box structure, internal weld defects need to be paid attention to), service life (if it has been in service for 5 years, the inspection of worn parts needs to be strengthened) to determine the inspection focus.

[0040] S3. Simulate the execution of the initial detection scheme in the preset digital twin system. Use reinforcement learning algorithm to construct a reward function with security, economy and efficiency as comprehensive optimization objectives. Solve for the parameters of the initial detection scheme and output the optimal detection scheme that satisfies the minimum detection cost and the shortest detection time.

[0041] The simulation and optimization module loads the proposed solution into the digital twin model of the regenerator. Simulation results show that PAUT achieves a Point of Detection (POD) of 85% for this type of fine thermal fatigue crack. Using a reinforcement learning algorithm, a reward function is constructed with safety, economy, and efficiency as comprehensive optimization objectives. The parameters of the initial detection scheme are solved using this reward function, which serves as the reinforcement learning execution function. After multiple iterations, the reinforcement learning agent suggests reducing the PAUT's scanning speed by 15% and increasing the probe frequency to 10MHz. This optimized solution improves the POD to 94% while keeping the inspection time within an acceptable range. This optimized solution was adopted as the final solution.

[0042] Using a 10t steam boiler (special equipment category: boiler, equipment number: GL-2017-123) in use at a power plant as the target equipment, a 1:1 high-fidelity digital twin was constructed based on its 3D structural drawings, material physical properties (boiler drum material Q245R, water-cooled wall material 20G), and operating parameters (rated evaporation capacity 10t / h, rated pressure 3.82MPa, rated temperature 450℃). This twin integrates a multiphysics model, including a temperature field model (simulating the temperature distribution of the furnace and heating surfaces), a pressure field model (reproducing the pressure transmission law within the boiler), and a defect evolution model (simulating crack propagation and corrosion deepening processes based on fracture mechanics theory). It also connects to the equipment's real-time operating data interface to ensure real-time synchronization between the virtual model and the physical equipment's status.

[0043] The generated initial inspection plan is loaded into the digital twin system. The core content of the plan includes the inspection parameters of key components such as: boiler drum (medium-high risk, inspection items are weld defect detection and wall thickness measurement, inspection methods are ultrasonic testing and radiographic testing, inspection cycle is 6 months), water-cooled wall (medium risk, inspection items are corrosion detection and deformation measurement, inspection methods are magnetic particle testing and laser ranging, inspection cycle is 12 months), and safety valve (low risk, inspection items are start-up pressure verification, inspection method is calibration bench calibration, inspection cycle is 18 months).

[0044] The digital twin system simulated the complete inspection process according to the initial plan and output the simulation results: the overall defect detection probability (POD) was 83%, of which the detection rate of micro-cracks (≤1mm) in the boiler drum weld was only 75%; the total inspection cost was 28,000 yuan (including equipment downtime loss of 15,000 yuan and inspection consumables and labor costs of 13,000 yuan); the total inspection time was 52 hours (of which boiler drum inspection accounted for 24 hours, requiring two separate shutdowns); the safety assurance coefficient was 0.86 (safety assurance coefficient = 1 - probability of failure due to undetected defects).

[0045] Simulation analysis identified the following problems with the initial design: First, the simultaneous use of ultrasonic and X-ray inspections on the boiler drum resulted in overlapping procedures, leading to excessively long inspection times and high costs. Second, the water-cooled wall inspection did not cover areas severely affected by flue gas erosion, creating blind spots. Third, the dispersed inspection cycles for different components caused frequent equipment downtime, impacting production efficiency.

[0046] Reward Function Construction: With safety, economy, and efficiency as comprehensive optimization objectives, a linear weighted reward function is designed, as shown in the following formula: ; Wherein: S comprehensively reflects the safety assurance capability of the testing scheme, S=α×POD+β×S_f, α and β are weighting coefficients (α=0.6, β=0.4), POD is the defect detection probability (value range 0-1), and S_f is the safety assurance coefficient (value range 0-1); E reflects the efficiency of the testing process, E=1-(T_actual / T_standard), T_actual is the total simulated testing time, T_standard is the total industry standard testing time (preset to 60 hours), score range 0-1; C reflects the rationality of the testing cost, C=1-(C_actual / C_budget), C_actual is the total simulated testing cost, C_budget is the preset testing budget (30,000 yuan), score range 0-1; ω1, ω2, ω3 are target weights (set according to the safety priority of special equipment: ω1=0.45, ω2=0.3, ω3=0.25), satisfying ω1+ω2+ω3=1. The reward function has a value range of 0-1, and the higher the score, the better the solution.

[0047] Based on the initial simulation state (R=0.68), the intelligent agent first adjusts the combination of detection methods: cancels the X-ray inspection of the boiler drum and replaces it with phased array ultrasonic inspection (PAUT), taking advantage of PAUT's high detection rate of defects in thick-walled welds to make up for the inadequacy of ultrasonic inspection in detecting micro-cracks; at the same time, it optimizes the inspection sequence by synchronizing the boiler drum inspection and water-cooled wall inspection within the same downtime window to reduce the number of downtimes.

[0048] The optimized parameters were input into the digital twin system for resimulation, and the output results were as follows: POD increased to 89% (the detection rate of micro-cracks in the boiler drum reached 90%), the safety assurance coefficient increased to 0.91, the total detection time was shortened to 42 hours, the total detection cost was reduced to 23,000 yuan, and the reward value R=0.79, which is 16.2% higher than the first round.

[0049] In subsequent iterations, the intelligent agent focused on adjusting the detection range and operating parameters: For the water-cooled wall, based on the flue gas scouring simulation results of the digital twin, the detection range was adjusted from full-area coverage to focused coverage of the two sides of the furnace and the rear heating surface (accounting for 60% of the total area), while reducing the detection density in the low-scouring area at the top of the water-cooled wall; the PAUT detection operating parameters were adjusted, increasing the probe frequency from 5MHz to 8MHz and the scanning speed from 8mm / s to 6mm / s, balancing the detection rate and detection efficiency; the safety valve inspection cycle was extended from 18 months to 24 months, while annual spot checks on sealing were added (the cost of spot checks is only 15% of that of full calibration).

[0050] After 120 iterations, the reward function value tends to stabilize (the rate of change is ≤0.8% for 10 consecutive rounds), reaching the convergence condition. At this point, the Pareto optimal detection scheme is output.

[0051] The probability of defect detection (POD) increased to 94%, and the safety assurance coefficient increased to 0.95, representing improvements of 13.3% and 10.5% respectively compared to the initial plan, completely resolving the problem of insufficient detection rate for microcracks. The total inspection cost decreased from 28,000 yuan to 21,000 yuan, a reduction of 25%; the unit inspection cost decreased from 2,800 yuan / part to 2,100 yuan / part, with the safety valve inspection cost decreasing by 40% (due to extended cycle and sampling inspection). The total inspection time was shortened from 52 hours to 38 hours, a reduction of 26.9%; the number of equipment downtimes decreased from 3 to 1, and the downtime duration was reduced from 52 hours to 38 hours, reducing production interruption losses by more than 30%, fully meeting the optimization goals of "minimum inspection cost and shortest inspection time."

[0052] like Figure 5 As shown, the workflow of the simulation and optimization module in the special equipment inspection scheme is based on the initial inspection scheme (including the list of inspection items, selection of inspection methods, resource allocation plan, and inspection cycle setting) and the digital twin environment (including equipment geometric model, equipment physical model, and equipment operation model). First, a virtual inspection process simulation is carried out in the digital twin environment to calculate indicators such as defect detection probability, estimated time consumption, inspection cost, and risk level. Then, reinforcement learning optimization is performed with the DDPG algorithm as the core. By constructing a reward function, iterating the scheme, and judging convergence, the optimal inspection scheme that satisfies the balance between safety, economy, and efficiency is output. Finally, feedback learning is carried out by combining on-site inspection data (including defect type, defect image, defect size, and defect location) to calculate the result error and realize incremental model updates.

[0053] Output and Feedback: The solution output and feedback learning module generates a detailed inspection work order. The inspection team executes the plan and ultimately finds a tiny crack in the designated location. This result (including crack size, location image) is entered into the system. The system compares the on-site inspection results with the model's predictions and uses this new data to fine-tune the deep learning model, improving its future prediction accuracy for similar cracks.

[0054] The primary objective of this invention is to overcome the shortcomings of existing technologies and provide an intelligent system capable of automatically generating personalized, high-precision inspection and testing solutions for special equipment. Another objective of this invention is to improve inspection efficiency and economy by introducing "digital twin" and deep learning technologies to simulate, verify, and optimize the generated inspection and testing solutions for special equipment, thereby ensuring safety.

[0055] like Figure 6As shown, the three-layer architecture of the special equipment inspection and testing scheme generation system based on AI deep learning is illustrated. The data layer includes a static attribute library for equipment and a regulatory standard database, providing multi-source basic data for the system. The AI ​​layer has four core modules: multi-data source fusion, equipment safety assessment, inspection scheme generation, and scheme simulation and optimization. It uses models such as LSTM and CNN and reinforcement learning algorithms to complete data processing, assessment, scheme generation, and optimization. The application layer, through the scheme output and feedback learning module, realizes standardized scheme generation, visualization, data collection feedback, and model update training, thus constructing a closed-loop intelligent inspection scheme generation system.

[0056] Multi-data source fusion module: Used to extract and clean data from Enterprise Resource Planning (ERP), Supervisory Control and Data Acquisition (SCADA), Inspection Report Management System (IRMS), and Internet of Things (IoT) sensors. The holographic equipment profile built by this module includes:

[0057] Equipment static attribute library: design parameters, manufacturing information, material grades, structural drawings.

[0058] Dynamic operating database: time-series operating parameters such as pressure, temperature, flow rate, and vibration.

[0059] Historical Inspection Records Database: All inspection reports, defect records, maintenance history, and raw data and images from non-destructive testing (such as UT, RT, MT).

[0060] Regulatory and Standards Knowledge Base: The latest versions of the TSG series of safety technical regulations, GB / T national standards and industry standards are stored in a structured format.

[0061] Equipment safety status assessment module: This module is one of the core components of this invention, employing a multi-task deep learning model to analyze the fused data. The model architecture includes:

[0062] Input encoding layer: Long Short-Term Memory (LSTM) network is used to process time-series running data, Convolutional Neural Network (CNN) is used to process historical test image data, Transformer encoder is used to process text report data, and Fully Connected Layer (FCL) is used to process structured attribute data.

[0063] Feature fusion layer: Employs an attention-based fusion method to dynamically weight the feature contributions from different data sources, forming a unified device state feature vector.

[0064] Multi-task output layer: Simultaneously outputs risk level (high risk, medium risk, low risk), health score (continuous value from 0 to 100), and possible failure modes of key components (such as uniform corrosion, pitting, fatigue cracks, creep, etc.) as well as the probability of equipment failure.

[0065] Personalized Inspection Solution Generation Module: This module integrates a rule engine and a recommendation system. The rule engine ensures that the generated solutions strictly comply with the mandatory requirements in the regulatory and standard knowledge base. The recommendation system, based on the output of the equipment safety status assessment module, uses collaborative filtering and knowledge graph technology to retrieve similar cases from a historical high-quality solution database, generating a preliminary solution that includes a list of inspection items, recommended inspection methods (such as visual inspection, ultrasonic inspection, and radiographic inspection), key inspection areas, and suggested inspection cycles.

[0066] The simulation and optimization module introduces digital twin technology to create a high-fidelity virtual model of the physical equipment. The initial solution is simulated within the digital twin model to predict its defect detection rate (POD), estimated time, cost, and risk level. Based on this, a reinforcement learning algorithm based on Deep Deterministic Policy Gradient (DDPG) is employed, using a comprehensive reward function that maximizes safety, minimizes cost, and optimizes efficiency to iteratively optimize the solution parameters. Convergence is determined by a reward function change rate <1%, ultimately outputting the Pareto optimal solution.

[0067] The solution output and feedback learning module generates structured, executable inspection task sheets from the optimized solutions, supplemented with visual charts. After the inspection personnel complete the task, they feed back data such as the actual defect findings and inspection time to the system through this module. The system calculates the difference between the predicted and actual values ​​and uses this difference data to incrementally learn or fine-tune the deep learning model in the equipment safety status assessment module, forming a technical chain of decision-making, execution, feedback, and optimization.

[0068] From experience-driven to data-driven: This invention deeply mines the value of data through deep learning models, enabling the formulation of solutions to be based on objective and quantitative analysis, reducing interference from human factors. (2) Achieving personalized customization for each machine: This invention comprehensively considers the unique "genes" (static attributes) and "history" (dynamic data) of each device to generate the most targeted inspection strategy. (3) Possessing foresight and dynamic adaptability: This invention, through digital twin simulation and reinforcement learning, can predict and optimize the effect of the solution before its implementation, and can dynamically adjust the inspection plan according to changes in the equipment status. (4) Forming closed-loop intelligence and continuous evolution: The feedback learning mechanism enables the system to learn from each inspection practice, continuously correct its prediction model, and achieve continuous improvement in intelligent performance.

[0069] Based on the same inventive concept, the present invention also provides a special equipment detection scheme generation system, comprising: The data acquisition module is used to collect the inherent setting attributes, dynamic operating data, historical diagnostic reports, and regulatory standard data of the target special equipment; the dynamic operating data includes the temperature and pressure curves of each component during equipment operation, and ultrasonic and X-ray images of each component.

[0070] The feature extraction module is used to extract the temporal variation features of the temperature and pressure curves of each component, extract the defect morphology features of the ultrasonic and X-ray images of each component, and extract the textual semantic features of the historical diagnostic reports of each component. The temporal variation features, defect morphology features, and textual semantic features are dynamically weighted and fused to form a unified equipment status feature vector. Using the equipment status feature vector, the safety assessment score of each component of the target equipment is output. Based on the safety assessment score, inherent setting attributes, and constraints of regulatory and standard data, an initial testing plan is generated, which includes the inspection items, recommended inspection methods, key inspection areas, and suggested inspection cycles for each component.

[0071] The application module is used to simulate the execution of the initial detection scheme in a preset digital twin system. It uses a reinforcement learning algorithm to construct a reward function with security, economy and efficiency as comprehensive optimization objectives, solves the parameters of the initial detection scheme, and outputs the optimal detection scheme that satisfies the minimum detection cost and the shortest detection time.

[0072] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the aforementioned method for generating a detection scheme for special equipment.

[0073] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described method for generating a detection scheme for special equipment.

[0074] Specific limitations regarding the computational system for generating inspection plans for special equipment can be found in the limitations described above for the method of generating inspection plans for special equipment, and will not be repeated here. Each module in the aforementioned system for generating inspection plans for special equipment can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device, or stored in the memory of the computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0075] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for generating a testing scheme for special equipment, characterized in that, include: Collect the inherent settings attributes, dynamic operating data, historical diagnostic reports, and regulatory and standard data of the target special equipment; The dynamic operating data includes temperature and pressure curves of each component during equipment operation, and ultrasonic and X-ray images of each component. Extract the temporal variation features of temperature and pressure curves of each component, extract the defect morphology features of ultrasound and X-ray images of each component, and extract the textual semantic features of historical diagnostic reports of each component. The temporal variation features, defect morphology features, and text semantic features are dynamically weighted and fused to form a unified device state feature vector; Using the device state feature vector, the safety assessment score of each component of the target device is obtained; based on the safety assessment score, inherent setting attributes and regulatory standard data constraints, an initial testing plan is generated that includes the inspection items, recommended inspection methods, key inspection areas and suggested inspection cycles for each component. The initial detection scheme is simulated and executed in the pre-defined digital twin system. A reinforcement learning algorithm is used to construct a reward function with security, economy and efficiency as comprehensive optimization objectives. The parameters of the initial detection scheme are solved, and the optimal detection scheme that satisfies the minimum detection cost and the shortest detection time is output.

2. The method for generating a detection scheme for special equipment according to claim 1, characterized in that, The method employs a reinforcement learning algorithm to construct a reward function with safety, economy, and efficiency as comprehensive optimization objectives, and solves for the parameters of the initial detection scheme. Specifically, the reinforcement learning algorithm designs the reward function with maximizing safety, minimizing detection cost, and minimizing detection time as its core objectives. The detection scheme parameters are adjusted iteratively through interaction between the agent and the digital twin system. After iterating until the reward value converges, the Pareto optimal detection scheme is output. The detection scheme parameters include the combination of detection methods, detection order, detection range, and operation parameters.

3. The method for generating a detection scheme for special equipment according to claim 1, characterized in that, The temporal variation features of temperature and pressure curves of each component are extracted using a Long Short-Term Memory (LSTM) network. The defect morphology features of ultrasound and X-ray images of each component are extracted using a Convolutional Neural Network (CNN). The textual semantic features of historical diagnostic reports of each component are extracted using a Transformer encoder. An attention mechanism is used to dynamically weight and fuse the temporal variation features, defect morphology features, and textual semantic features to form a unified equipment status feature vector.

4. The method for generating a detection scheme for special equipment according to claim 1, characterized in that, After outputting the optimal detection scheme that satisfies the minimum detection cost and the shortest detection time, the method further includes: the digital twin system generating a prediction result of the optimal detection scheme; the optimal detection scheme being executed on-site by the target device, and the actual detection result data being collected; and the reinforcement learning being updated by using the difference between the actual detection result data and the prediction result.

5. The method for generating a detection scheme for special equipment according to claim 1, characterized in that, The inherent attribute parameters include design parameters, manufacturing information, material grades, structural drawings, and service life; dynamic operating data includes time-series parameters for pressure, temperature, flow rate, vibration, acoustic emission, and corrosion monitoring; historical diagnostic reports include all previous test reports, defect records, maintenance history, and raw data and images from non-destructive testing; regulatory and standard data include structured TSG series procedures, GB / T national standards, and industry standards.

6. The method for generating a detection scheme for special equipment according to claim 1, characterized in that, The safety assessment score includes a comprehensive evaluation of component risk level, health score, and potential failure probability.

7. The method for generating a detection scheme for special equipment according to claim 1, characterized in that, Through a rule engine and a case recommendation system, an initial testing plan is generated, which includes inspection items for each component, recommended inspection methods, key inspection areas, and suggested inspection cycles. The rule engine ensures the compliance of the initial testing plan, and the case recommendation system retrieves successful cases most similar to the current problem from the historical case library, draws on their solutions, and generates a preliminary testing plan adapted to the current scenario.

8. A system for generating detection schemes for special equipment, characterized in that, include: The data acquisition module is used to collect the inherent settings attributes, dynamic operating data, historical diagnostic reports, and regulatory and standard data of the target special equipment. The dynamic operating data includes temperature and pressure curves of each component during equipment operation, and ultrasonic and X-ray images of each component. The feature extraction module is used to extract the temporal variation features of the temperature and pressure curves of each component, extract the defect morphology features of the ultrasound and X-ray images of each component, and extract the textual semantic features of the historical diagnostic reports of each component. The temporal change features, defect morphology features, and text semantic features are dynamically weighted and fused to form a unified equipment state feature vector; using the equipment state feature vector, the safety assessment score of each component of the target equipment is output; based on the safety assessment score, inherent setting attributes, and constraints of regulatory and standard data, an initial testing plan is generated that includes inspection items, recommended inspection methods, key inspection areas, and suggested inspection cycles for each component; The application module is used to simulate the execution of the initial detection scheme in a preset digital twin system. It uses a reinforcement learning algorithm to construct a reward function with security, economy and efficiency as comprehensive optimization objectives, solves the parameters of the initial detection scheme, and outputs the optimal detection scheme that satisfies the minimum detection cost and the shortest detection time.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 7.

Images