Predictive maintenance method and device for overpressure protection system, equipment and storage medium

Abstract

The invention discloses a predictive maintenance method, device and equipment for an overpressure protection system and a storage medium, and relates to the technical field of nuclear power plants, and the method comprises the steps: comprehensively collecting the multi-source operation and maintenance historical data of the system, extracting health characterization features, pre-judging the state trend through a time sequence prediction model, and analyzing the fault root cause after early warning. The technical problems that traditional maintenance lacks data support, multi-source data are not effectively fused, prediction and decision are insufficient and the model is not continuously optimized are solved by combining maintenance constraints to generate decision suggestions and executing feedback to optimize the model, accurate prediction, early warning and intelligent maintenance decision are achieved, excessive or insufficient maintenance is avoided, and the maintenance efficiency is improved. The cost and the safety risk are reduced, a closed-loop intelligent maintenance system is formed, and the safe and stable operation of the system is guaranteed.

Description

Technical Field

[0001] This application relates to the field of nuclear power plant technology, and in particular to a predictive maintenance method, apparatus, equipment and storage medium for an overpressure protection system. Background Technology

[0002] The overpressure protection system of a nuclear power plant is a critical barrier to ensure the safe operation of nuclear power. Traditionally, its maintenance has relied on a combination of periodic testing, preventative replacement, and reactive repair. This approach suffers from problems such as over-maintenance increasing costs, under-maintenance failing to detect potential faults in a timely manner, and maintenance decisions relying on experience without data support. With the development of nuclear power digitalization, massive amounts of operational and maintenance data have been accumulated. However, existing technologies have failed to effectively integrate multi-source heterogeneous data to comprehensively characterize equipment status. Feature extraction lacks specificity, and time-series prediction models struggle to accurately capture complex degradation patterns and long-term dependencies of equipment. Furthermore, after a prediction triggers an early warning, there is a lack of effective analysis of the root causes of the fault, making it impossible to generate scientific maintenance decision recommendations based on maintenance constraints. Additionally, there is a lack of a mechanism for continuous optimization of the prediction model based on execution feedback, making it difficult to form a closed-loop intelligent maintenance system and failing to meet the needs of real-time monitoring, early warning, and precise maintenance for overpressure protection systems. Summary of the Invention

[0003] This application provides a predictive maintenance method, apparatus, equipment, and storage medium for an overpressure protection system. It addresses the problems in related technologies where traditional maintenance methods lack data support, fail to effectively integrate multi-source data, have insufficient targeted feature extraction, struggle to accurately capture equipment degradation patterns using time-series prediction models, lack effective root cause analysis, cannot generate scientific decision-making suggestions based on maintenance constraints, and lack a continuous model optimization mechanism.

[0004] According to a first aspect of this application, a predictive maintenance method for an overpressure protection system is provided, comprising: Collect multi-source operational data and maintenance history data generated during the operation of the overpressure protection system; Feature extraction is performed on the multi-source operating data and the maintenance history data to construct feature data characterizing the health status of the overpressure protection system; Based on the feature data, a prediction result is generated by a time-series prediction model to predict the future health status trend of the overpressure protection system. When the future health status trend meets the preset early warning conditions, root cause analysis is performed on the prediction results to generate characteristic factors that lead to changes in health status. Based on the future health status trend and the characteristic factors, combined with the preset maintenance constraints, maintenance decision suggestions are generated for the overpressure protection system. Based on the execution feedback of the maintenance decision recommendations, the machine learning model is updated and optimized.

[0005] According to a second aspect of this application, a predictive maintenance device for an overpressure protection system is provided, comprising: The acquisition module is configured to acquire multi-source operational data and maintenance history data generated by the overpressure protection system during operation. The extraction module is configured to extract features from the multi-source operating data and the maintenance history data to construct feature data characterizing the health status of the overpressure protection system. The prediction module is configured to generate prediction results based on the feature data using a time-series prediction model to predict the future health status trend of the overpressure protection system. The analysis module is configured to perform root cause analysis on the prediction results and generate characteristic factors that lead to changes in health status when the future health status trend meets preset warning conditions. The generation module is configured to generate maintenance decision suggestions for the overpressure protection system based on the future health status trend and the characteristic factors, combined with preset maintenance constraints. The optimization module is configured to update and optimize the machine learning model based on the execution feedback of the maintenance decision recommendations.

[0006] According to a third aspect of this application, an electronic device is provided, comprising: At least one processor; and memory that is communicatively connected to at least one processor; The memory stores instructions that can be executed by at least one processor, which enables the at least one processor to perform the predictive maintenance method for the overpressure protection system described in the first aspect above.

[0007] According to a fourth aspect of this application, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to perform the predictive maintenance method for an overpressure protection system described in the first aspect above.

[0008] According to a fifth aspect of this application, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the predictive maintenance method for an overpressure protection system as described in the first aspect above.

[0009] This application addresses several issues in related technologies. By collecting multi-source operational and maintenance historical data from overpressure protection systems and extracting features to construct characteristic data representing the system's health status, using time-series prediction models to predict future health trends, conducting root cause analysis when warning conditions are met, generating maintenance decision recommendations based on maintenance constraints, and updating and optimizing the model based on execution feedback, this approach solves problems such as lack of data support, ineffective integration of multi-source data, insufficient targeted feature extraction, difficulty in accurately capturing equipment degradation patterns using time-series prediction models, lack of effective root cause analysis, inability to generate scientific decision recommendations based on maintenance constraints, and lack of continuous model optimization mechanisms. This achieves the technical effects of accurate prediction, early warning, and intelligent maintenance decision-making for overpressure protection systems, avoiding over-maintenance and under-maintenance, reducing operating costs and safety risks, forming a closed-loop intelligent maintenance system, and ensuring the safe and stable operation of the system.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0011] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating a predictive maintenance method for an overpressure protection system provided in an embodiment of this application; Figure 2 A flowchart illustrating another predictive maintenance method for an overpressure protection system provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of a predictive maintenance device for an overpressure protection system provided in an embodiment of this application. Detailed Implementation

[0013] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0014] The following description, with reference to the accompanying drawings, describes a predictive maintenance method, apparatus, device, and storage medium for an overpressure protection system according to embodiments of this application.

[0015] Figure 1 This is a flowchart illustrating a predictive maintenance method for an overpressure protection system provided in an embodiment of this application.

[0016] like Figure 1 As shown, the method includes the following steps: Step 101: Collect multi-source operating data and maintenance history data generated by the overpressure protection system during operation.

[0017] In some embodiments, to comprehensively obtain complete information on the operating status and historical maintenance of the overpressure protection system and provide reliable data support for subsequent analysis, it is necessary to collect multi-source operating data and maintenance history data generated by the system during operation. Multi-source operating data originates from multiple related systems within the nuclear power plant, including digital control systems, safety instrumented systems, and environmental monitoring systems. This includes time-series data continuously collected by sensors, such as primary loop pressure, secondary loop pressure, pressure before and after the safety valve, valve opening, medium temperature, coil resistance, and vibration spectrum—core parameters directly reflecting the real-time operating status of the equipment. It also includes automatically generated status logs during equipment operation, detailing the number of valve actions, various functional test results, and alarm information triggered during operation. Environmental parameters affecting equipment operation, such as ambient temperature, humidity, and radiation levels recorded by the environmental monitoring system, are also included in the multi-source operating data. Maintenance history data comes from a dedicated maintenance database, comprehensively recording past maintenance work orders for each key component of the overpressure protection system, including the last calibration date, valve seat grinding, seal replacement, and other specific information on various maintenance operations, clearly presenting the equipment's historical maintenance trajectory and usage status. During the data collection process, various data types are correlated using unified device tags and timestamps to ensure accurate matching of data from different sources and of different types, achieving effective aggregation of multi-dimensional data. This comprehensive collection method can fully capture key information on the operating status and historical maintenance of the overpressure protection system, providing a comprehensive and accurate data foundation for subsequent feature extraction and fault prediction.

[0018] Step 102: Extract features from the multi-source operating data and the maintenance history data to construct feature data characterizing the health status of the overpressure protection system.

[0019] In some embodiments, after acquiring multi-source operational data and maintenance history data, systematic processing and targeted feature extraction are required to transform the raw data into feature data that accurately characterizes the health status of the overpressure protection system. First, both types of raw data are preprocessed to remove illegal values ​​exceeding reasonable ranges and abnormal data caused by sensor malfunctions or transmission interference. Missing values ​​encountered during data acquisition are filled using appropriate interpolation or statistical methods. Simultaneously, data from different acquisition frequencies and sources are time-series aligned to ensure consistency across all data, laying a high-quality data foundation for subsequent feature extraction. Based on this, deep feature construction is carried out by combining the operating mechanism of the overpressure protection system with domain expertise. For multi-source operational data, key features such as pressure fluctuation rate, average response time, leakage indication features, and vibration energy in specific frequency bands can be extracted. Pressure fluctuation rate reflects the stability of system pressure changes; average response time reflects the response efficiency of components such as valves; leakage indication features reflect the sealing status through the difference in relevant pressure parameters; and vibration energy features capture abnormal vibration signals during equipment operation. For historical maintenance data, derived features such as maintenance interval duration, cumulative number of similar maintenance operations, and runtime since the last maintenance can be constructed. These features can indirectly reflect the equipment's recovery status and long-term service stability after maintenance. Through the above preprocessing and feature construction process, scattered and messy raw data is transformed into feature data with reasonable dimensions and clear physical meaning. This feature data comprehensively depicts the health status of the overpressure protection system from multiple dimensions such as operating status and historical maintenance, providing accurate and effective data support for subsequent status prediction and fault diagnosis. Its beneficial effect lies in solving the problem that raw data cannot directly reflect the system's health status. Through targeted feature extraction, the characterization of health status becomes more targeted and accurate, laying a solid foundation for the efficient implementation of subsequent steps.

[0020] Step 103: Based on the feature data, generate prediction results through a time-series prediction model to predict the future health status trend of the overpressure protection system.

[0021] In some embodiments, the feature data characterizing the health status of the overpressure protection system constructed in step 102 is used as the core input of the time-series prediction model, fully leveraging the model's ability to deeply mine time-series data to accurately predict the future health status trend of the system. The time-series prediction model is specifically designed for data with temporal correlation, effectively capturing the changing patterns of feature data over time. Whether it's subtle fluctuations easily overlooked in the short term or gradually emerging degradation trends in the long term, these can all be presented through the model's learning and analysis. During the prediction process, the model fully utilizes historically accumulated feature data, combined with real-time updated feature information, to establish a mapping relationship between feature changes and the evolution of the system's health status, deeply learning the health status change patterns of the overpressure protection system under different operating conditions. Through such learning and computation, the final prediction results generated by the model not only include the overall trajectory of the system's health status change over a future period but also quantify the expected fluctuation range and degradation rate of health indicators, while estimating the remaining useful life of key system components, allowing maintenance personnel to clearly and intuitively predict the future health trend of the overpressure protection system. It breaks away from the reliance on experience in traditional forecasting methods and significantly improves the accuracy and reliability of predicting future health status trends through a data-driven time series forecasting model, providing solid technical support for timely detection of potential risks and the formulation of reasonable maintenance plans.

[0022] Step 104: When the future health status trend meets the preset early warning conditions, perform root cause analysis on the prediction results to generate characteristic factors that lead to changes in health status.

[0023] In some embodiments, the preset early warning conditions are formulated based on the safe operation standards of the overpressure protection system, historical fault data, and the failure threshold of core components. When the health indicators in the future health status trend generated in step 103 are lower than the predefined safety threshold, or the failure probability exceeds the set early warning line, the root cause analysis process is triggered. The root cause analysis process focuses on the correlation logic between the prediction results and feature data, using specialized analysis techniques to deeply dissect the intrinsic causes of changes in health status. By quantifying the contribution of each input feature to the current prediction result, key factors that play a dominant role in health status degradation are selected from numerous features. The analysis process fully integrates the operating mechanism of the overpressure protection system, focusing on features related to the function of core components, such as leakage indication features related to sealing status, vibration energy features related to component vibration, and pressure fluctuation features related to pressure stability. It clarifies which abnormal changes in features directly lead to a decline in health indicators or an increase in failure probability, accurately locating specific operating parameters, component status, or environmental influencing factors. Through such in-depth analysis, clear and specific feature factors are ultimately generated, allowing maintenance personnel to intuitively understand the core causes of changes in the health status of the overpressure protection system, rather than just knowing that there is a risk of failure. This solves the problem that traditional early warning systems can only indicate risks but cannot identify the root cause, providing a precise basis for developing targeted maintenance plans, avoiding blind maintenance, and improving the targeting and efficiency of fault handling.

[0024] Step 105: Based on the future health status trend and the characteristic factors, and combined with the preset maintenance constraints, generate maintenance decision suggestions for the overpressure protection system.

[0025] In some embodiments, after obtaining the future health status trend of the overpressure protection system and the characteristic factors leading to changes in health status, a comprehensive assessment needs to be conducted in conjunction with preset maintenance constraints to generate scientific and feasible maintenance decision recommendations. The preset maintenance constraints are comprehensive criteria based on nuclear power plant safety operation specifications, operation and maintenance management requirements, and actual operational needs. They cover core elements such as safety risk level classification standards, the schedulable range and time limits of maintenance resources (including personnel, spare parts, and tools), window period restrictions for unit operation (such as power reduction adjustment periods and planned overhaul cycles), cost thresholds corresponding to different maintenance operations, and assessment standards for potential downtime losses due to fault expansion. During the assessment process, the urgency of the system is first determined based on the future health status trend—if health indicators deteriorate rapidly, the probability of failure approaches the critical value, and the remaining useful life is short, the maintenance need is more urgent. Then, the maintenance objects and core priorities are identified by combining characteristic factors. Targeted maintenance directions are formulated for key components and abnormal parameters that lead to failure. For example, if internal leakage is caused by wear on the sealing surface, the focus is on sealing component repair; if abnormal vibration is the cause, the emphasis is on component tightening and wear detection. Based on this, a multi-dimensional balance is made against preset maintenance constraints: if the urgency is high but the unit is in a critical operating phase, the downtime losses far outweigh the maintenance risks, and there is a near-term operating window, then planned maintenance should be prioritized within that window; if the probability of failure is extremely high, and the safety risk exceeds the constraint threshold, then even with some downtime losses, immediate maintenance should be recommended to avoid major safety hazards; if the health status declines gradually, the probability of failure is low, and maintenance resources are temporarily strained, then continued monitoring and simultaneous resource allocation can be recommended. The final maintenance decision recommendations are clear and explicit, including not only precise maintenance timing (such as specific time nodes and operating windows) and appropriate maintenance types (such as online verification, component replacement, calibration and debugging), but also clear maintenance priorities, ensuring that maintenance work proceeds in an orderly and efficient manner. This achieves precise and scientific maintenance decisions, avoiding cost waste caused by over-maintenance and eliminating safety risks caused by under-maintenance, thus balancing the relationship between safety, cost, and operational efficiency.

[0026] Step 106: Update and optimize the machine learning model based on the execution feedback of the maintenance decision suggestions.

[0027] In some embodiments, the execution feedback of maintenance decision recommendations encompasses multiple key information aspects, including effectiveness verification data after maintenance personnel actually perform maintenance actions—such as whether the health indicators of the overpressure protection system have returned to a safe range after maintenance, whether potential faults have been completely eliminated, whether the actual root causes of faults are consistent with the characteristic factors analyzed by the model, and on-site information such as new operating conditions and new fault modes discovered during the maintenance process that were not captured by the model; it also includes specific parameter adjustments during maintenance execution, records of abnormal situations, and optimization suggestions given by maintenance personnel based on practical experience. This feedback information is comprehensively collected and standardized by the system. After removing invalid interference information, it is integrated with the model's original training dataset to form new effective samples encompassing the entire process of "prediction-decision-execution-verification." Based on this, incremental learning is used to update and optimize the machine learning model. Instead of retraining the entire model, targeted adjustments are made to the model's parameter weights and feature association logic based on the new samples, correcting any deviations that may have existed in previous predictions or root cause analyses. This allows the model to continuously learn new operating scenarios and fault patterns, gradually improving its ability to capture complex operating conditions and potential faults, and enhancing its prediction accuracy. It has constructed a continuously iterative closed-loop learning system, which enables the machine learning model to keep up with the actual operational changes of the overpressure protection system, avoids the model's accuracy decay due to long-term use, and ensures that subsequent prediction results and maintenance decision suggestions always maintain high reliability and pertinence, providing continuous optimization technical support for the long-term safe and stable operation of the system.

[0028] Compared with related technologies, in this embodiment, multi-source operational data and maintenance history data generated during the operation of the overpressure protection system are collected; features are extracted from the multi-source operational data and the maintenance history data to construct feature data characterizing the health status of the overpressure protection system; based on the feature data, a time-series prediction model is used to generate prediction results to predict the future health status trend of the overpressure protection system; when the future health status trend meets preset warning conditions, root cause analysis is performed on the prediction results to generate feature factors leading to changes in health status; based on the future health status trend and the feature factors, combined with preset maintenance constraints, maintenance decision suggestions for the overpressure protection system are generated; based on the execution feedback of the maintenance decision suggestions, the machine learning model is updated and optimized. It can solve the problems in related technologies, such as the lack of data support in traditional maintenance methods, the ineffective integration of multi-source data, insufficient targeted feature extraction, the inability of time series prediction models to accurately capture equipment degradation patterns, the lack of effective root cause analysis of failures, the inability to generate scientific decision-making suggestions in combination with maintenance constraints, and the lack of a continuous model optimization mechanism. It can achieve the technical effects of accurate prediction, early warning and intelligent maintenance decision-making of overpressure protection systems, avoid over-maintenance and under-maintenance, reduce operating costs and safety risks, form a closed-loop intelligent maintenance system, and ensure the safe and stable operation of the system.

[0029] Figure 2 A flowchart illustrating another predictive maintenance method for an overpressure protection system provided in this application embodiment includes the following steps: Step 201: Collect sensor timing data, system event logs, and environmental parameter data related to the overpressure protection system from the process control system, safety protection system, maintenance history database, and environmental monitoring system of the nuclear power plant.

[0030] In some embodiments, the process control system, as the core monitoring and control carrier of the nuclear power plant's production process, collects sensor time-series data that directly reflects the real-time operating status of key components of the overpressure protection system. This includes high-frequency dynamic data such as primary loop pressure, secondary loop pressure, pressure before and after the safety valve, valve opening degree, medium temperature, coil resistance, and vibration spectrum. These data are recorded in continuous time series, capturing subtle fluctuations and trends in equipment operation. The safety protection system focuses on real-time monitoring of equipment safety operation. Its collected system event logs record in detail the number of valve actions, various functional test results, alarm information under abnormal operating conditions, and fault triggering events, clearly presenting the equipment's safety status and response at different operating stages. The maintenance history database stores all maintenance-related records since the overpressure protection system was put into operation, covering the time of each maintenance, specific operation content, replaced component models, and verification results, providing a basis for tracing the equipment's historical status and analyzing degradation patterns. The environmental monitoring system is responsible for collecting external environmental parameters that affect equipment operation, mainly including data on ambient temperature, relative humidity, and radiation levels within the nuclear power plant. While these parameters do not directly determine the core functions of the equipment, they affect the aging rate of components and operational stability over long periods, making them an indispensable part of comprehensively assessing the system's health status. During the data collection process, a unified equipment tag is used to establish a correspondence between various data types and specific components of the overpressure protection system. Simultaneously, precise timestamps are used to align the time sequence of data from different sources, ensuring consistency across all collected data in both spatial and temporal dimensions, forming a structured and correlated comprehensive dataset. This implementation method achieves comprehensive coverage and effective correlation of multi-dimensional data, avoiding the limitations of data collected from a single system, and providing rich and reliable basic data support for subsequent feature extraction and status analysis.

[0031] Step 202: Perform data cleaning, data alignment, and data filling processing on the multi-source operating data and the maintenance history data.

[0032] In some embodiments, data cleaning, data alignment, and data imputation are crucial steps in ensuring the quality of raw data and providing a reliable foundation for subsequent feature construction. During data cleaning, systematic processing is required to address various anomalies that may exist in multi-source operational data and maintenance historical data: illegal data exceeding equipment ranges or clearly violating physical laws (such as pressure values ​​far exceeding normal ranges or negative vibration energy data) are directly removed; noise data caused by sensor interference or transmission link fluctuations is smoothed to reduce its impact on the overall data trend; and duplicate records generated during multi-source data fusion due to system interactions are removed to avoid redundant information interfering with the analysis results. Data alignment uses timestamps as the core benchmark. Due to differences in data acquisition frequencies from different sources (such as millisecond-level data from high-frequency sensors versus second-level records in system logs), interpolation is used to unify all data to the same time granularity, ensuring that operating parameters such as pressure, temperature, and vibration are accurately matched with equipment status logs and maintenance records in the time dimension, achieving spatiotemporal consistency of the data. Data imputation addresses missing values ​​caused by temporary sensor malfunctions or signal interruptions during data acquisition. It combines the time-series characteristics of the data with the operating patterns of the overpressure protection system, employing methods such as the mean, median, or trend extrapolation based on historical similar operating conditions to reasonably fill in the missing values. This ensures data integrity without disrupting the inherent correlations and trends of the original data. This step effectively eliminates invalid interference information, resolves the spatiotemporal differences and integrity issues of multi-source data, significantly improves data quality, and lays a solid foundation for the accuracy of subsequent feature construction.

[0033] Step 203: Based on the working principle of the overpressure protection system, construct characteristic data reflecting the performance degradation or potential faults of the overpressure protection system from the processed data.

[0034] In some embodiments, based on the working principle of the overpressure protection system, the focus is on its core functional links such as pressure regulation, sealing protection, and component action response, and feature data that can accurately reflect system performance degradation or potential faults are deeply constructed from the processed high-quality data. Based on the operational logic of the system's core components, characteristics such as pressure fluctuation rate (the magnitude of pressure change per unit time), deviation of pressure peak value from long-term average, and pressure stabilization time are constructed for pressure control-related data. These characteristics can intuitively reflect the system's ability to regulate pressure fluctuations. Abnormal fluctuations or prolonged stabilization time often indicate performance degradation of pressure control components. For critical actuators such as valves, characteristics such as average action time (the average time of multiple valve opening and closing actions), action response delay (the time difference between issuing a command and the valve executing the action), and cumulative rate of valve action are constructed. Component wear or jamming will directly lead to longer action times and increased response delays. For sealing performance, leakage indication characteristics (the pressure difference between the inlet and outlet when the valve is closed) are constructed. A continuous increase in this characteristic usually indicates wear on the sealing surface or aging of the sealing components. For vibration data, vibration energy integral characteristics of specific frequency bands are extracted, focusing on frequency band energy changes related to valve core impact and component loosening. Abnormal vibration energy is an important signal of component degradation. At the same time, combined with maintenance history data, characteristics such as maintenance interval duration, number of recurrences of similar faults, and continuous system operation time after the last maintenance are constructed. These characteristics can indirectly reflect the maintenance effect and long-term service stability of the equipment. The feature data constructed in this step is closely aligned with the system's working principle, enabling it to accurately capture early signals of performance degradation and potential faults, providing targeted and physically meaningful core data support for subsequent health status prediction and fault diagnosis.

[0035] Step 204: Using an unsupervised learning method, construct system health indicators based on historical feature data under normal operating conditions.

[0036] In some embodiments, unsupervised learning methods are used to construct system health indicators. The core of this approach is to fully utilize the massive historical feature data accumulated under normal operating conditions of the overpressure protection system, achieving effective characterization of the health status without relying on fault samples. First, data from periods of long-term stable operation of the overpressure protection system, with no fault records and parameters meeting safety standards, needs to be selected as the training basis. This data covers pre-processed features such as pressure, vibration, action response, and environmental influences, comprehensively reflecting the normal operating state of the system. Unsupervised learning methods construct health assessment standards by autonomously mining the inherent patterns in the data. Common implementation methods include reconstruction error analysis based on autoencoders and decision distance calculation using a type of support vector machine. Autoencoders learn the distribution patterns and correlation patterns of feature data under normal operating conditions, accurately reconstructing input data under normal conditions. When the system experiences performance degradation or potential faults, the data deviates from the normal pattern, and the reconstruction error increases significantly. This error value can serve as a core indicator directly reflecting the health status. Support vector machines learn the decision boundaries of normal data, mapping data in a high-dimensional feature space to a specific decision domain. The distance between a data point and the decision boundary can be converted into a health scalar; the greater the distance, the more severely the system deviates from its normal state. This type of unsupervised learning method transforms complex, multi-dimensional feature data into a one-dimensional health indicator. This indicator remains at a low level and fluctuates smoothly during normal system operation, but shows a continuous upward trend as components degrade or potential faults appear, providing an intuitive and clear quantification of the system's health status. This solves the problem of scarce fault samples in overpressure protection systems, making it difficult to construct health assessment standards, and achieves accurate quantification of health status, providing an intuitive and easily monitored core basis for subsequent predictions.

[0037] Step 205: Input the historical sequence of the health indicators and the feature data into the trained time series prediction model to generate health indicator values ​​and corresponding fault occurrence probabilities for future periods in order to predict the future health status trend of the overpressure protection system.

[0038] In some embodiments, the time-series prediction model needs to be fully trained beforehand. The training process uses the historical health indicator sequence and corresponding feature data of the overpressure protection system as core samples, allowing the model to deeply learn the correlation between health indicators and various features, the changing patterns over time, and the evolution patterns under different operating conditions. This ensures that the model has the ability to capture short-term feature fluctuations and long-term health degradation trends. During the prediction phase, the historical sequence of health indicators and the real-time updated feature data are used as inputs to the model. The historical sequence of health indicators reflects the trajectory of the system's past health status changes, providing a trend basis for prediction. The feature data contains dynamic information such as real-time pressure, vibration, and action response, providing support for the model to capture immediate state changes. During model execution, local dynamic patterns in the feature data are first extracted through a specific network structure. Then, a network with an attention mechanism focuses on key historical information highly correlated with the current prediction, capturing the long-range dependencies of the health status. Finally, feature fusion and the output layer generate predicted health indicator values ​​for multiple consecutive time steps in the future, fully presenting the trajectory of the system's health status changes over future periods. Simultaneously, probability estimation techniques are employed to analyze the uncertainty of the prediction results. By generating a probability distribution of predicted health indicator values ​​through multiple samplings and combining this with preset health indicator failure thresholds, the specific probability of system failures occurring in future time periods is calculated, clarifying the risk level at different time points. This step achieves accurate output of both health status trends and failure risks, not only allowing maintenance personnel to understand future health changes but also quantifying the probability of failures. This provides trend-based and quantitative data support for subsequent early warning triggering and maintenance decisions, enhancing the practical value of the prediction results.

[0039] Step 206: Use interpretable artificial intelligence technology to invert the prediction results generated by the time series prediction model.

[0040] In some embodiments, interpretable artificial intelligence technology can break through the "black box" limitations of time-series prediction models, providing clear and traceable logical support for prediction results. When inverting prediction results, this technology, based on the operating mechanism and data correlation patterns of the overpressure protection system, traces back the core basis for the predicted future health status trend to arrive at the model's result. By reconstructing the model's processing path for input features, weight allocation logic, and decision-making process, the abstract model output is transformed into a concrete analysis process closely integrated with the actual operating scenario. Whether it's judging the downward trend of health indicators or issuing a risk warning of exceeding the fault probability limit, this technology can break down the process of action of specific input features, making the originally difficult-to-understand model prediction results transparent and interpretable. This avoids relying solely on model output without logical corroboration, ensuring that root cause analysis has solid technical support. Its beneficial effect lies in improving the credibility of prediction results, laying a logical foundation for subsequent accurate fault location, and meeting the requirements of high-confidence analysis in the nuclear safety field.

[0041] Step 207: Quantify the degree of influence of each input feature on the prediction result of the future health status trend.

[0042] In some embodiments, the process of quantifying the influence of each input feature requires combining the health status evaluation standards of the overpressure protection system and using specialized analysis algorithms to accurately calculate the correlation between each input feature and the prediction result. The algorithm comprehensively considers the magnitude of feature data changes, frequency of occurrence, and physical correlation with the system's health status, assigning a corresponding quantitative value to each feature. This intuitively reflects the feature's driving or inhibiting effect on future health status trends. Features with higher values ​​indicate that their abnormal changes contribute more significantly to the decline in health indicators or the increase in failure probability, and are key factors dominating changes in health status; while features with lower values ​​have a weak impact on the prediction result and can be temporarily disregarded as the focus of root cause analysis. This quantification method allows for the rapid selection of core influencing factors from numerous input features, avoiding aimless root cause analysis. It achieves precise quantification of feature influence, abandons the vague judgment of feature importance in traditional analysis, and provides data support for subsequent fault root cause localization.

[0043] Step 208: Based on the features with the highest degree of influence, locate the associated subsystems or physical processes as feature factors for potential root cause analysis of failures.

[0044] In some embodiments, after obtaining the quantification results of the influence of each input feature, the core feature with the highest value and most significant impact is focused on. Combined with the structural composition and working principle of the overpressure protection system, the specific subsystem or physical process associated with this feature is located. For example, if the leakage indication feature has the highest influence, it corresponds to the sealing subsystem and the medium sealing physical process in the overpressure protection system; if the vibration energy feature ranks highly, it is associated with the valve core movement subsystem and the physical processes related to component impact and loosening. Through this precise association, abstract feature data is transformed into specific equipment operation links or component states, clarifying that these associated subsystems and physical processes are the core causes leading to changes in the system's health status, ultimately identifying them as feature factors. Its beneficial effect lies in realizing the transformation from "feature data" to "actual fault causes," closely integrating root cause analysis results with actual equipment operation, and providing clear targets for subsequent development of targeted maintenance plans.

[0045] Step 209: Construct a state space based on the system health status, operating conditions, and resource conditions, and use different maintenance actions as the action space.

[0046] In some embodiments, constructing a state space and action space is the foundation for reinforcement learning-driven maintenance decisions, aiming to provide clear decision boundaries and optional behavioral frameworks for subsequent benefit assessment. The construction of the state space comprehensively integrates three types of core information: system health status directly adopts previously generated quantitative data such as health indicators, failure probabilities, and remaining useful life, accurately reflecting the current and future health level of the equipment; operating condition information covers real-time operating scenario parameters such as nuclear power plant unit load and operating phase (e.g., normal full-load operation, power reduction adjustment, standby state), reflecting the operating environment of the equipment; resource conditions focus on the feasibility of maintenance implementation, including the scheduling of maintenance personnel, the inventory quantity and allocation cycle of required spare parts, and the availability of special tools, ensuring that decisions are aligned with the current status of operation and maintenance resources. The action space is designed based on the maintenance needs and actual operating scenarios of the overpressure protection system, including four core maintenance actions: immediate maintenance, suitable for scenarios with extremely high fault risk and requiring emergency handling; planned maintenance, for situations where the risk is controllable and can be implemented during subsequent operating windows; continued monitoring, for situations where the health status degradation is gradual, requiring no substantial maintenance but continuous tracking; and testing, for scenarios where the root cause of the fault is not yet clear and further testing and verification are needed. These four actions comprehensively cover maintenance options under different risk levels and actual needs. Its beneficial effect lies in clarifying the environmental constraints and behavioral options for decision-making, laying a structured foundation for subsequent accurate evaluation of the benefits of each maintenance action, and avoiding decisions that are divorced from actual operation and resource conditions.

[0047] Step 210: Use a reinforcement learning algorithm to evaluate the long-term benefits of different actions in the state space. The benefits are calculated by comprehensively considering maintenance costs, downtime losses, and security risks.

[0048] In some embodiments, reinforcement learning algorithms simulate the execution process of different maintenance actions in the current state space to achieve a comprehensive evaluation of the long-term benefits of each action. The core of this evaluation lies in the comprehensiveness and long-term perspective of the benefit calculation. The benefit calculation uses maintenance costs, downtime losses, and safety risks as three core dimensions to form a quantified reward function: maintenance costs cover direct inputs such as spare parts procurement or replacement costs, maintenance personnel labor costs, and the cost of using specialized equipment; downtime losses focus on indirect losses such as power generation losses and production plan delays caused by unit shutdowns or load reductions due to maintenance actions, requiring precise calculation based on unit operating efficiency and downtime; safety risks are incorporated into the benefit evaluation system by quantifying the potential risks of equipment damage, safety barrier failure, and even radiation leakage that may result from fault expansion. The algorithm does not only consider the immediate benefits of a single action but also simulates the evolution trajectory of the system state after the action is executed, as well as the subsequent maintenance needs and risk changes that may be triggered, calculating the long-term accumulated comprehensive benefits to ensure that the evaluation results align with the core objective of long-term safe operation of the nuclear power plant. Its beneficial effect lies in breaking through the limitations of traditional decision-making that only focuses on short-term costs or single risks, and realizing a comprehensive assessment of maintenance actions over multiple dimensions and long cycles, providing a scientific quantitative basis for selecting the optimal decision.

[0049] Step 211: Output the maintenance action with the highest long-term benefit as a decision recommendation including maintenance timing and type.

[0050] In some embodiments, after the reinforcement learning algorithm completes the long-term benefit evaluation of all maintenance actions, it automatically selects the maintenance action with the highest long-term benefit and transforms it into specific and feasible maintenance decision suggestions, realizing the implementation from algorithm evaluation results to actual operation and maintenance guidance. In the decision suggestions, the maintenance type directly corresponds to the selected optimal action, clearly informing operation and maintenance personnel that immediate maintenance, planned maintenance, continued monitoring, or testing should be performed; the maintenance timing is precisely determined by combining the operating condition information and resource conditions in the state space. For example, planned maintenance will clearly point to suitable operating windows such as the unit's power reduction adjustment period and the planned overhaul cycle, ensuring that the maintenance action does not conflict with the unit's operating plan, while matching the resource availability time. In addition, the decision suggestions will also indirectly associate the characteristic factors obtained from root cause analysis, so that the maintenance type corresponds precisely to the root cause of the fault. For example, the fault risk caused by the wear of the sealing surface will clearly point to the replacement of the seal or the repair of the sealing surface in the planned maintenance suggestion. Its beneficial effects lie in the fact that the maintenance decision recommendations it outputs are both scientific and operable. It ensures a balance between safety, cost and efficiency through long-term benefit assessment, and provides specific timing and type guidance so that maintenance personnel can implement them without additional analysis, effectively improving the accuracy of maintenance decisions and the efficiency of maintenance work.

[0051] Step 212: Update and optimize the machine learning model based on the execution feedback of the maintenance decision suggestions.

[0052] For a description of step 212, please refer to the description of step 106 in the above embodiment. This embodiment will not repeat the details further.

[0053] Figure 3 This is a schematic diagram of the structure of a predictive maintenance device for an overpressure protection system provided in an embodiment of this application, as shown below. Figure 3 As shown, it includes: acquisition module 301, extraction module 302, prediction module 303, analysis module 304, generation module 305, and optimization module 306.

[0054] The acquisition module 301 is configured to acquire multi-source operating data and maintenance history data generated by the overpressure protection system during operation; The extraction module 302 is configured to extract features from the multi-source operating data and the maintenance history data to construct feature data characterizing the health status of the overpressure protection system. Prediction module 303 is configured to generate prediction results based on the feature data using a time-series prediction model to predict the future health status trend of the overpressure protection system. The analysis module 304 is configured to perform root cause analysis on the prediction results and generate characteristic factors that lead to changes in health status when the future health status trend meets preset early warning conditions. The generation module 305 is configured to generate maintenance decision suggestions for the overpressure protection system based on the future health status trend and the characteristic factors, combined with preset maintenance constraints. The optimization module 306 is configured to update and optimize the machine learning model based on the execution feedback of the maintenance decision suggestions.

[0055] In some examples of this embodiment, the acquisition module 301 is specifically configured to acquire sensor timing data, system event logs, and environmental parameter data related to the overpressure protection system from the process control system, safety protection system, maintenance history database, and environmental monitoring system of the nuclear power plant.

[0056] In some examples of this embodiment, the extraction module 302 is specifically configured to perform data cleaning, data alignment and data filling processing on the multi-source operating data and the maintenance history data; based on the working principle of the overpressure protection system, feature data reflecting the performance degradation or potential faults of the overpressure protection system is constructed from the processed data.

[0057] In some examples of this embodiment, the prediction module 303 is specifically configured to use an unsupervised learning method to construct system health indicators based on historical feature data under normal operating conditions. The historical sequence of the health indicators and the feature data are input into a trained time series prediction model to generate health indicator values ​​and corresponding fault occurrence probabilities for future periods in order to predict the future health status trend of the overpressure protection system.

[0058] In some examples of this embodiment, the analysis module 304 is specifically configured to use interpretable artificial intelligence technology to invert the prediction results generated by the time series prediction model: quantify the degree of influence of each input feature on the future health status trend prediction results; and locate the associated subsystems or physical processes based on the features with the highest degree of influence as feature factors for potential root cause analysis of failures.

[0059] In some examples of this embodiment, the generation module 305 is specifically configured to construct a state space based on system health status, operating conditions, and resource conditions, and to use different maintenance actions as the action space; to use a reinforcement learning algorithm to evaluate the long-term benefits of different actions in the state space, wherein the benefits are calculated by comprehensively considering maintenance costs, downtime losses, and safety risks; and to output the maintenance action with the highest long-term benefits as a decision recommendation that includes maintenance timing and type.

[0060] It should be noted that other corresponding descriptions of the functional units involved in the predictive maintenance device for overpressure protection system provided in this embodiment can be found in [reference needed]. Figure 1 , Figure 2The corresponding descriptions in [the document] will not be repeated here.

[0061] Based on the above, Figure 1 , Figure 2 The present embodiment provides a predictive maintenance method for an overpressure protection system. Correspondingly, this embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method. Figure 1 , Figure 2 This paper presents a predictive maintenance method for an overpressure protection system.

[0062] Based on the above, Figure 1 , Figure 2 The present embodiment provides a predictive maintenance method for an overpressure protection system. Correspondingly, this embodiment also provides a computer program product on which a computer program is stored. When executed by a processor, this computer program implements the above-described predictive maintenance method. Figure 1 , Figure 2 This paper presents a predictive maintenance method for an overpressure protection system.

[0063] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0064] Based on the above, Figure 1 , Figure 2 A predictive maintenance method for an overpressure protection system is shown, and Figure 3 To achieve the above objectives, the present application also provides an electronic device, such as a personal computer or a server, in the illustrated virtual device embodiment. This device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to implement the above-described virtual device. Figure 1 , Figure 2 This paper presents a predictive maintenance method for an overpressure protection system.

[0065] In some embodiments, the aforementioned physical device may further include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, an input unit such as a keyboard, etc., and optionally, a USB interface, a card reader interface, etc. In some embodiments, the network interface may include a standard wired interface, a wireless interface (such as a Wi-Fi interface), etc.

[0066] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0067] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0068] The above are merely specific embodiments of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to these embodiments, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A predictive maintenance method for an overpressure protection system, characterized in that, include: Collect multi-source operational data and maintenance history data generated during the operation of the overpressure protection system; Feature extraction is performed on the multi-source operating data and the maintenance history data to construct feature data characterizing the health status of the overpressure protection system; Based on the feature data, a prediction result is generated by a time-series prediction model to predict the future health status trend of the overpressure protection system. When the future health status trend meets the preset early warning conditions, root cause analysis is performed on the prediction results to generate characteristic factors that lead to changes in health status. Based on the future health status trend and the characteristic factors, combined with the preset maintenance constraints, maintenance decision suggestions are generated for the overpressure protection system. Based on the execution feedback of the maintenance decision recommendations, the machine learning model is updated and optimized.

2. The predictive maintenance method for the overpressure protection system according to claim 1, characterized in that, The data collected during the operation of the overpressure protection system includes multi-source operational data and maintenance history data, such as: Collect sensor timing data, system event logs, and environmental parameter data related to the overpressure protection system from the nuclear power plant's process control system, safety protection system, maintenance history database, and environmental monitoring system.

3. The predictive maintenance method for the overpressure protection system according to claim 1, characterized in that, The step of extracting features from the multi-source operational data and the maintenance history data to construct feature data characterizing the health status of the overpressure protection system includes: The multi-source operational data and the maintenance historical data are subjected to data cleaning, data alignment, and data completion processing. Based on the working principle of the overpressure protection system, characteristic data reflecting the performance degradation or potential faults of the overpressure protection system are constructed from the processed data.

4. The predictive maintenance method for the overpressure protection system according to claim 1, characterized in that, The step of generating prediction results based on the feature data using a time-series prediction model to predict the future health status trend of the overpressure protection system includes: An unsupervised learning method is used to construct system health indicators based on historical feature data under normal operating conditions. The historical sequence of the health indicators and the feature data are input into a trained time series prediction model to generate health indicator values ​​and corresponding fault occurrence probabilities for future periods in order to predict the future health status trend of the overpressure protection system.

5. The predictive maintenance method for the overpressure protection system according to claim 1, characterized in that, When the future health status trend meets the preset early warning conditions, root cause analysis is performed on the prediction results to generate characteristic factors leading to changes in health status, including: The prediction results generated by the time series prediction model are inverted using interpretable artificial intelligence technology: Quantify the degree of influence of each input feature on the future health status trend prediction result; Based on the characteristics with the highest degree of influence, locate the related subsystems or physical processes, and use these characteristics as the basis for potential root cause analysis of failures.

6. The predictive maintenance method for the overpressure protection system according to claim 1, characterized in that, The step of generating maintenance decision recommendations for the overpressure protection system based on the future health status trend and the characteristic factors, combined with preset maintenance constraints, includes: The state space is constructed based on the system's health status, operating conditions, and resource conditions, while the action space is constructed based on different maintenance actions. The long-term benefits of different actions in the state space are evaluated using reinforcement learning algorithms. These benefits are calculated by comprehensively considering maintenance costs, downtime losses, and security risks. Output the maintenance actions with the highest long-term benefits, as a decision recommendation including the timing and type of maintenance.

7. A predictive maintenance device for an overpressure protection system, characterized in that, include: The acquisition module is configured to acquire multi-source operational data and maintenance history data generated by the overpressure protection system during operation. The extraction module is configured to extract features from the multi-source operating data and the maintenance history data to construct feature data characterizing the health status of the overpressure protection system. The prediction module is configured to generate prediction results based on the feature data using a time-series prediction model to predict the future health status trend of the overpressure protection system. The analysis module is configured to perform root cause analysis on the prediction results and generate characteristic factors that lead to changes in health status when the future health status trend meets preset warning conditions. The generation module is configured to generate maintenance decision suggestions for the overpressure protection system based on the future health status trend and the characteristic factors, combined with preset maintenance constraints. The optimization module is configured to update and optimize the machine learning model based on the execution feedback of the maintenance decision recommendations.

8. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the predictive maintenance method for the overpressure protection system according to any one of claims 1-6.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the predictive maintenance method for the overpressure protection system according to any one of claims 1-6.

10. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the predictive maintenance method for an overpressure protection system according to any one of claims 1-6.

Images