Nuclear power circulating water pump multi-agent operation and maintenance system based on industrial physics AI engine

Abstract

The invention discloses a nuclear power circulating water pump multi-agent operation and maintenance system based on an industrial physics AI engine, and relates to the technical field of intelligent operation and maintenance of nuclear power station equipment, the nuclear power circulating water pump multi-agent operation and maintenance system comprises a data acquisition agent, the industrial physics AI engine, a coordination agent, an operation and maintenance knowledge base and an operation and maintenance assistant agent, according to the method, a data driving model and a physical mechanism model are fused, the method is specially designed for the nuclear power circulating water pump, predictive maintenance and intelligent decision making of the nuclear power circulating water pump are achieved through the fusion data driving and physical mechanism model of a multi-agent collaboration framework and an industrial physical AI engine, and integration of a dynamic knowledge base and natural language interaction; compared with a traditional method, the operation efficiency and safety are remarkably improved, an industrial physics AI engine serves as a core, seamless fusion of a data model and a mechanism model is ensured, the fault prediction accuracy, scientificity and interpretability are improved, and secondary disasters caused by sudden faults are avoided.

Description

A Multi-Agent Operation and Maintenance System for Nuclear Power Plant Circulating Water Pumps Based on an Industrial Physics AI Engine Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for nuclear power plant equipment, specifically to a multi-agent operation and maintenance system for nuclear power circulating water pumps based on an industrial physics AI engine. Background Technology

[0002] Nuclear power plant circulating water pumps are the core equipment of the nuclear power cooling system, responsible for transporting cooling water to maintain the safe operation of the reactor. Their reliability directly affects the safety and economy of the power plant. Traditional operation and maintenance of circulating water pumps mainly rely on manual inspection, regular maintenance and simple fault diagnosis technology. In recent years, multi-agent systems based on artificial intelligence have gradually emerged in the industrial field, for example, for optimizing the energy efficiency control of cooling water pumps or the safe operation and maintenance management of distributed equipment. At the same time, large language models (LLM) have also been introduced into industrial operation and maintenance to provide suggestions for natural language interaction. The application of these technologies in the operation and maintenance of nuclear power circulating water pumps is still in its early stages and has not yet formed a systematic multi-agent collaboration framework.

[0003] However, the current operation and maintenance of circulating water pumps has the following characteristics and shortcomings:

[0004] Operation and maintenance rely on human experience, and data analysis and decision-making are performed manually. The operating environment is complex and the amount of data is huge. The efficiency of manual identification of anomalies is low and highly subjective, making it difficult to achieve early warning.

[0005] Insufficient data utilization, inconsistent data formats and protocols across different systems (such as DCS and SCADA), and difficulties in data cleaning and fusion lead to inadequate real-time analysis and multi-dimensional feature extraction.

[0006] Mechanism analysis and intelligent diagnosis are disconnected. Commonly used mechanism calculations (such as NPSH net positive margin and bearing life L10 criterion) are usually independent of intelligent diagnostic models, which cannot be effectively combined with business scenarios. Furthermore, there is a lack of a unified platform to combine mechanism knowledge with data-driven models, making it difficult to balance interpretability and prediction accuracy.

[0007] There is a lack of knowledge accumulation and sharing. Operation and maintenance records, expert experience and industry standards are scattered in paper archives or independent databases. There is a lack of structured knowledge base. Existing operation and maintenance support systems are difficult to quickly call and reason through natural language interaction, and it is inconvenient for operation and maintenance personnel to obtain decision support.

[0008] To avoid the aforementioned technical problems, it is indeed necessary to provide a multi-agent operation and maintenance system for nuclear power circulating water pumps based on an industrial physics AI engine to overcome the deficiencies in the existing technology. Summary of the Invention

[0009] This invention provides a multi-agent operation and maintenance system for nuclear power plant circulating water pumps based on an industrial physics AI engine. It can effectively solve the problems mentioned in the background technology, such as operation and maintenance relying on human experience, insufficient data utilization, disconnect between mechanism analysis and intelligent diagnosis, and insufficient knowledge accumulation and sharing.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a multi-agent operation and maintenance system for nuclear power circulating water pumps based on an industrial physics AI engine, comprising a data acquisition agent, an industrial physics AI engine, a coordination agent, an operation and maintenance knowledge base, and an operation and maintenance assistant agent;

[0011] The industrial physics AI engine is the core component of the system, integrating data-driven models and physical mechanism models. It is specifically designed for nuclear power circulating water pumps and includes equipment baseline models, state analysis algorithms, and mechanism calculation algorithms.

[0012] After the system starts up, the specific workflow is as follows:

[0013] S1: Data Acquisition and Preprocessing. The data acquisition agent collects and processes the monitoring data of the nuclear power plant circulating water pump and transmits the data to the industrial physics AI engine.

[0014] S2: Establish a normal operation baseline model. Based on real-time data provided by the data acquisition agent, the industrial physics AI engine constructs a normal operation model of the circulating water pump. The baseline model is then transmitted to the state analysis algorithm for deviation comparison, and the predicted curve is sent to the mechanism calculation algorithm for fusion evaluation.

[0015] S3: Performs operational status assessment and early warning. It receives data from the industrial physical AI engine to obtain pre-processed data and baseline models of intelligent agents, and performs operational status analysis algorithms to conduct real-time comparisons and fault warnings.

[0016] S4: Conduct mechanism analysis and life prediction. The industrial physics AI engine receives data to obtain the classification data of the intelligent agent, the baseline model prediction curve, the early warning signal obtained by the state assessment algorithm, and the knowledge query of the operation and maintenance knowledge base.

[0017] S5: The system coordinates intelligent agents to optimize multi-agent collaboration. The coordinating intelligent agents receive status reports, health and lifespan predictions output by the industrial physical AI engine.

[0018] S6: Build and update the operation and maintenance knowledge base, receive data from the operation and maintenance knowledge base to obtain multimodal data of intelligent agents, and support knowledge storage and dynamic mining;

[0019] S7: Provides intelligent operation and maintenance interaction, receiving evaluation results and analysis reports from the industrial physical AI engine, as well as specifications and experience from the operation and maintenance knowledge base through the operation and maintenance assistant intelligent agent.

[0020] According to the above technical solution, in S1, the data acquisition agent serves as the system's data input layer, connecting to the sensor network of the nuclear power plant's circulating water pump, including vibration sensors, temperature sensors, pressure sensors, and flow sensors. The collected real-time data includes time-series signals, and the acquisition frequency is set according to the pump's operating characteristics, not lower than 1kHz to capture high-frequency fault signals. After acquisition, data preprocessing is performed: Kalman filtering algorithm is used for noise reduction, with the following formula:

[0021] (1)

[0022] (2)

[0023] in, For posterior state estimation, For Kalman gain, Let H be the observed values, H be the observation matrix, P be the covariance matrix, and R be the observation noise covariance.

[0024] According to the above technical solution, in S2, the industrial physics AI engine has the function of establishing a baseline model of the equipment. It uses real-time data provided by the data acquisition agent to establish a normal operation model of the circulating water pump. A Transformer neural network is used to process time-series data to capture multivariate dependencies and graph structure features. The Transformer model structure includes a multi-head self-attention mechanism and a feedforward network. The calculation formula for the attention mechanism is:

[0025] (3)

[0026] Where Q, K, and V are the query, key, and value matrices, The key dimension is [batch_size, sequence_length, features]. The input data is a multi-channel time series with dimensions [batch_size, sequence_length, features]. Training uses historical normal operation data, and the goal is to generate a baseline model to simulate the behavior parameters of the pump under normal conditions.

[0027] The model training uses the Adam optimizer, and the loss function is the mean squared error combined with Kullback-Leibler divergence to improve distribution matching.

[0028] (4)

[0029] (5)

[0030] in, To ensure interpretability, attention weights are visualized and key features are highlighted. After training, the baseline model outputs the normal operation threshold and prediction curve for subsequent evaluation.

[0031] According to the above technical solution, in S3, the industrial physical AI engine receives data to acquire the preprocessed data and baseline model of the intelligent agent, and runs the status evaluation algorithm.

[0032] The comparison methods include bias analysis: calculating the residuals between real-time data and baseline predictions. If the residual exceeds the threshold, an early warning is triggered. A multi-core support vector machine is used to enhance evaluation accuracy. The MK-SVM classification hyperplane formula is:

[0033] (6)

[0034] in, Let m be the m-th kernel function, and let Gaussian kernel function be used. For the core weights, optimize the solution using quadratic programming;

[0035] The early warning mechanism includes multi-level alerts: yellow - minor deviation, orange - potential fault, and red - immediate shutdown. To improve interpretability, SHAP values ​​are used to interpret model decisions, the marginal contribution of features to model output is calculated, and the model is interpreted from both global and local levels. The output includes status reports.

[0036] According to the above technical solution, in S4, the industrial physics AI engine receives data from the intelligent agent's classification data, prediction model, early warning signals, and knowledge queries from the operation and maintenance knowledge base to perform health and lifespan assessments. The industrial physics AI engine integrates mechanism calculation algorithms and possesses nuclear power industry-specific calculation methods, including NPSH (Net Positive Suction Head) calculation.

[0037] (7)

[0038] in, Atmospheric pressure Vaporization pressure, which is temperature-dependent. For density, It is the acceleration due to gravity. For flow rate, , The suction height and friction loss are calculated based on the pipe length;

[0039] Including bearing life calculation, the L10 criterion is adopted:

[0040] (8)

[0041] Where C is the basic dynamic load rating, P is the equivalent dynamic load, and the uncertainty in the lifetime prediction is quantified using the Weibull distribution:

[0042] (9)

[0043] in, For shape parameters, The scale parameter is used to output a health score (0-100) and an estimate of remaining lifespan.

[0044] Including efficiency calculation models:

[0045] (10)

[0046] Where Q is the flow rate and H is the head. Shaft power;

[0047] The industrial physics AI engine transmits the analysis and prediction results to the coordinating agent for decision arbitration, sends the analysis report to the operation and maintenance assistant agent for interactive output, and receives feedback decisions from the coordinating agent to optimize the analysis parameters.

[0048] According to the above technical solution, in S5, the system coordinating agent is located in the coordination and knowledge layer, receives the state report and analysis and prediction results output by the industrial physics AI engine, is responsible for communication and decision optimization among multiple agents, and uses a multi-agent reinforcement learning algorithm to achieve cooperation. The state space includes the output of each agent, the action space is resource allocation, and the value function is decomposed as follows:

[0049] (11)

[0050] in, For a single agent, the Q-value. It is a hybrid function, where s is the global state, and coordination ensures system autonomy;

[0051] The system coordinating agent provides knowledge queries to the industrial physics AI engine for physical model supplementation, sends specifications and experience to the operation and maintenance assistant agent for suggestion generation, and receives decisions from the coordinating agent to trigger knowledge updates.

[0052] According to the above technical solution, in S6, the operation and maintenance knowledge base is constructed as a dynamic knowledge graph, which stores operation records, national industry standards, accident reports and expert experience, supports multimodal data integration, including data in the form of images and videos, and uses knowledge mining algorithms for the update mechanism to automatically extract new knowledge from external databases. The query adopts the SPARQL graph traversal algorithm.

[0053] According to the above technical solution, in S7, the operation and maintenance assistant intelligent agent is located in the interaction and integration layer. It receives the evaluation results and analysis reports of the industrial physical AI engine, as well as the specifications and experience of the operation and maintenance knowledge base. Based on the large language model LLM, it provides natural language interaction, inputs user queries, and uses prompt engineering to generate output: Prompt = "Based on the knowledge base and evaluation results, provide operation and maintenance suggestions for nuclear power circulating water pumps: {query}". It integrates a gatekeeper mechanism: rule-based filtering to ensure that the output complies with nuclear safety specifications, and uses XAI tools to interpret LLM decisions. It supports scenario simulation: the user inputs assumptions, and LLM generates reports.

[0054] The operations and maintenance assistant intelligent agent outputs user suggestions as the final operations and maintenance decision, and receives decisions from the coordinating intelligent agent to adjust the interaction logic.

[0055] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention achieves predictive maintenance and intelligent decision-making for nuclear power plant circulating water pumps through the integration of multi-agent collaborative architecture, industrial physics AI engine-driven data and physical mechanism models, and dynamic knowledge base and natural language interaction. Compared with traditional methods, it significantly improves operational efficiency and safety. The industrial physics AI engine, as the core, ensures seamless integration of data models and mechanism models, improves the accuracy, scientific nature, and interpretability of fault prediction, and avoids secondary disasters caused by sudden faults. The multi-agent collaboration supports on-demand maintenance, reduces unplanned downtime and lowers operation and maintenance costs. The integration of interpretable AI and gatekeeper mechanisms ensures transparent decision-making and compliance with nuclear safety regulations, reducing the risk of human error. Natural language suggestions enable rapid reuse of expert experience, improve decision-making efficiency, promote the intelligent transformation of the nuclear energy industry, and ensure sustainable development and energy security. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0057] In the attached diagram:

[0058] Figure 1 is a system operation flowchart of the present invention;

[0059] Figure 2 is a schematic diagram of the data acquisition intelligent agent workflow of the present invention;

[0060] Figure 3 is an example diagram of the knowledge graph of the operation and maintenance knowledge base of the present invention. Detailed Implementation

[0061] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0062] Example 1:

[0063] This invention provides a technical solution: a multi-agent operation and maintenance system for nuclear power circulating water pumps based on an industrial physics AI engine, comprising a data acquisition agent, an industrial physics AI engine, a coordination agent, an operation and maintenance knowledge base, and an operation and maintenance assistant agent.

[0064] The Industrial Physics AI Engine is the core component of the system, ensuring seamless integration of data models and mechanistic models. This improves the accuracy, scientific validity, and interpretability of fault prediction, preventing secondary disasters caused by sudden failures. It integrates data-driven models and physical mechanism models and is specifically designed for nuclear power plant circulating water pumps. The Industrial Physics AI Engine includes equipment baseline models, state analysis algorithms, and mechanism calculation algorithms. Based on this engine, the system uses the collaboration of multiple intelligent agents to call engine functions, enabling real-time monitoring, state assessment, life prediction, and intelligent operation and maintenance decisions for nuclear power plant circulating water pumps. It supports on-demand maintenance, reduces unplanned downtime, and lowers operation and maintenance costs.

[0065] As shown in Figure 1, the system operation flowchart illustrates the following workflow after system startup:

[0066] S1: Data Acquisition and Preprocessing. The data acquisition agent collects and processes monitoring data of the nuclear power plant circulating water pump and transmits the data to the industrial physics AI engine. In the industrial physics AI engine, real-time data is used to build the equipment baseline model, preprocessed data is used for the state analysis algorithm, and classified data is provided to the mechanism calculation algorithm. In addition, the data acquisition agent also inputs multimodal data into the operation and maintenance knowledge base to realize data distribution and preliminary interaction.

[0067] S2: Establish a normal operation baseline model. Based on real-time data provided by the data acquisition agent, the industrial physics AI engine constructs a normal operation model of the circulating water pump. The baseline model is then transmitted to the state analysis algorithm for deviation comparison, and the predicted curve is sent to the mechanism calculation algorithm for fusion evaluation.

[0068] S3: Performs operational status assessment and early warning. It receives data from the industrial physical AI engine to obtain the preprocessed data and baseline model of the intelligent agent, and performs real-time comparison and fault warning through the operational status analysis algorithm. It transmits the warning signal to the mechanism calculation algorithm for diagnosis, sends the status report to the coordinating intelligent agent for optimization, and provides the evaluation results to the operation and maintenance assistant intelligent agent for report generation. At the same time, it receives feedback from the coordinating intelligent agent to adjust the threshold.

[0069] S4: Conduct mechanism analysis and life prediction. The industrial physics AI engine receives data to obtain the classification data of the intelligent agent, the baseline model prediction curve, the early warning signal obtained by the state assessment algorithm, and the knowledge query of the operation and maintenance knowledge base. It runs the mechanism calculation algorithm to evaluate the health and remaining life, and transmits the health and life prediction to the coordinating intelligent agent for arbitration. The analysis report is sent to the operation and maintenance assistant intelligent agent for interaction. At the same time, it receives feedback from the coordinating intelligent agent to optimize parameters.

[0070] S5: The system coordinates the intelligent agent to optimize multi-agent collaboration. By receiving status reports and health and lifespan predictions from the industrial physical AI engine, the coordinating intelligent agent enables autonomous system decision-making and feeds back the coordination decisions to the industrial physical AI engine, the operation and maintenance knowledge base, and the operation and maintenance assistant intelligent agent to achieve dynamic adjustment and global optimization.

[0071] S6: Build and update the operation and maintenance knowledge base, receive data from the operation and maintenance knowledge base to obtain multimodal data of the intelligent agent, support knowledge storage and dynamic mining, provide knowledge queries to the industrial physical AI engine for model supplementation, and send specifications and experience to the operation and maintenance assistant intelligent agent for suggestion generation. At the same time, receive the decision of the coordinating intelligent agent to trigger updates.

[0072] S7: Provides intelligent operation and maintenance interaction. It receives the evaluation results and analysis reports of the industrial physical AI engine, as well as the specifications and experience of the operation and maintenance knowledge base, through the operation and maintenance assistant intelligent agent. It generates suggestions and reports based on natural language models and outputs user suggestions as the final operation and maintenance decision. At the same time, it receives the decisions of the coordination intelligent agent to adjust the interaction logic.

[0073] In S1, the data acquisition agent serves as the system's data input layer, connecting to the sensor network of the nuclear power plant's circulating water pump. This network includes vibration sensors, temperature sensors, pressure sensors, and flow sensors. The workflow of the data acquisition agent is shown in Figure 2. The collected real-time data includes time-series signals such as vibration amplitude, temperature changes, pressure fluctuations, and rotational speed. The acquisition frequency is set according to the pump's operating characteristics, typically not lower than 1kHz to capture high-frequency fault signals. After acquisition, data preprocessing is performed: denoising is achieved using the Kalman filtering algorithm, with the following formula:

[0074] (1)

[0075] (2)

[0076] in, For posterior state estimation, For Kalman gain, Here, H is the observation matrix, P is the covariance matrix, and R is the observation noise covariance.

[0077] Subsequently, data processing and classification were performed: the data was categorized into normal operation data, abnormal data, and historical data. K-means clustering was used for classification, with cluster centers initialized using the k-means++ method. Data quality was ensured to meet nuclear safety standards (IEC 61513) and verified via ZLIB. conduct

[0078] The data acquisition agent transmits data to the industrial physics AI engine. Real-time data is used for baseline model building, preprocessed data is sent to the state analysis algorithm for real-time comparison, and classified data is provided to the mechanism calculation algorithm for physical calculation. In addition, multimodal data, including images and videos, are input into the operation and maintenance knowledge base for knowledge updates.

[0079] In S2, the industrial physics AI engine has the function of establishing equipment baseline models. It uses real-time data provided by the data acquisition agent to build a normal operation model of the circulating water pump. It uses a Transformer neural network to process time series data to capture multivariate dependencies and graph structure features. The Transformer model structure includes a multi-head attention mechanism and a feedforward network. The attention mechanism calculation formula is as follows:

[0080] (3)

[0081] Where Q, K, and V are the query, key, and value matrices, The key dimension is [batch_size, sequence_length, features]. The input data is a multi-channel time series (such as vibration and temperature). The training uses historical normal operation data. The goal is to generate a baseline model to simulate the behavior parameters of the pump under normal conditions (such as vibration spectrum and efficiency curve).

[0082] The model training uses the Adam optimizer, with the loss function being the mean squared error (MSE) combined with Kullback-Leibler divergence (KL) to improve distribution matching.

[0083] (4)

[0084] (5)

[0085] in, To ensure interpretability, attention weights are visualized and key features are highlighted. After training, the baseline model outputs the normal operation threshold and prediction curve for subsequent evaluation.

[0086] In the industrial physics AI engine, the baseline model is transmitted to the state analysis algorithm for deviation comparison, and the prediction model is sent to the mechanism calculation algorithm for fusion physical assessment.

[0087] In S3, the industrial physics AI engine receives data to acquire preprocessed data and baseline models of intelligent agents, and runs state evaluation algorithms;

[0088] The comparison methods include bias analysis: calculating the residuals between real-time data and baseline predictions. If the residual exceeds the threshold, an early warning is triggered. Multi-Kernel SVM (MK-SVM) is used to enhance evaluation accuracy. The MK-SVM classification hyperplane formula is:

[0089] (6)

[0090] in, Let m be the m-th kernel function, and let Gaussian kernel function be used. For the core weights, optimize the solution using quadratic programming;

[0091] The early warning mechanism includes multi-level alerts: yellow - minor deviation, orange - potential fault, and red - immediate shutdown. To improve interpretability, SHAP (SHapley Additive exPlanations) values ​​are used to interpret model decisions, calculate the marginal contribution of features to model output, and interpret the model from both global and local levels. For example, the contribution of vibration features to fault prediction is quantified. Outputs include status reports, such as "abnormal pump vibration, deviation rate 15%, possible cause: bearing wear".

[0092] The industrial physics AI engine transmits early warning signals to the mechanism calculation algorithm for further diagnosis, sends status reports to the coordinating agent for global optimization, provides evaluation results to the operation and maintenance assistant agent for generating user reports, and receives feedback decisions from the coordinating agent to adjust evaluation thresholds.

[0093] In S4, the Industrial Physics AI Engine receives data from the intelligent agent's classification data, prediction models, early warning signals, and knowledge queries from the operation and maintenance knowledge base to perform health and lifespan assessments. The Industrial Physics AI Engine integrates mechanism calculation algorithms and possesses nuclear power industry-specific calculation methods, including NPSH (Net Positive Suction Head) calculation.

[0094] (7)

[0095] in, Atmospheric pressure Vaporization pressure, which is temperature-dependent. For density, It is the acceleration due to gravity. For flow rate, , The suction height and friction loss are calculated based on the pipe length;

[0096] Including bearing life calculation, the L10 criterion is adopted:

[0097] (8)

[0098] Where C is the basic dynamic load rating, P is the equivalent dynamic load, and the uncertainty in the lifetime prediction is quantified using the Weibull distribution:

[0099] (9)

[0100] in, For shape parameters, The scale parameter is used to output a health score (0-100) and an estimate of remaining lifespan.

[0101] Including efficiency calculation models:

[0102] (10)

[0103] Where Q is the flow rate and H is the head. Shaft power;

[0104] The industrial physics AI engine transmits the analysis and prediction results to the coordinating agent for decision arbitration, sends the analysis report to the operation and maintenance assistant agent for interactive output, and receives feedback decisions from the coordinating agent to optimize the analysis parameters.

[0105] In S5, the system coordinating agent resides in the coordination and knowledge layer. It receives state reports and analysis / prediction results from the industrial physics AI engine, and is responsible for communication and decision optimization among multiple agents. It employs a multi-agent reinforcement learning (MARL) algorithm to achieve collaboration. The state space includes the outputs of each agent, the action space is for resource allocation (such as priority adjustment), and the value function is decomposed as follows:

[0106] (11)

[0107] in, For a single agent, the Q-value. For the hybrid function, s is the global state, which coordinates to ensure system autonomy, such as prioritizing mechanism analysis and handling conflicts under high load;

[0108] The system coordinating agent provides knowledge queries to the industrial physics AI engine for physical model supplementation, sends specifications and experience to the operation and maintenance assistant agent for suggestion generation, and receives decisions from the coordinating agent to trigger knowledge updates.

[0109] In S6, the operations and maintenance knowledge base is built as a dynamic knowledge graph, as shown in Figure 3. It stores operation records, national and industry standards, accident reports and expert experience, and supports multimodal data integration, including data in the form of images and videos. The update mechanism uses knowledge mining algorithms to automatically extract new knowledge from external databases, and the query uses the SPARQL graph traversal algorithm.

[0110] In S7, the operations and maintenance assistant intelligent agent resides in the interaction and integration layer. It receives evaluation results and analysis reports from the industrial physics AI engine, as well as specifications and experience from the operations and maintenance knowledge base. Based on the large language model LLM, it provides natural language interaction. When a user query is input (such as "pump life prediction?"), the prompting engineering generates the output: Prompt = "Based on the knowledge base and evaluation results, provide operation and maintenance suggestions for nuclear power circulating water pumps: {query}". It integrates a gatekeeper mechanism: rule-based filtering to ensure that the output complies with nuclear safety specifications. It also uses the XAI tool to interpret LLM decisions and supports scenario simulation: the user inputs assumptions, and the LLM generates a report.

[0111] The operations and maintenance assistant intelligent agent outputs user suggestions as the final operations and maintenance decision, and receives decisions from the coordinating intelligent agent to adjust the interaction logic.

[0112] Example 2:

[0113] In the cooling system of a nuclear power plant, the Benduo intelligent agent operation and maintenance system was applied to a circulating water pump. The circulating water pump model is KSBRDL200-400, with a rated power of 500kW and a speed of 1480rpm. The system is deployed on an industrial-grade server with an Intel Xeon CPU and 64GB of RAM. The sensor network is connected via the Modbus protocol. The implementation process is as follows:

[0114] Step 1: Data Acquisition and Preprocessing: The intelligent agent connects to the sensor network on the pump to collect real-time data;

[0115] During normal operation of the circulating water pump at 80% load, vibration amplitude was 4.2 mm / s (normal threshold <5 mm / s), temperature was 45°C (normal range 40-50°C), pressure was 2.5 MPa (normal range 2-3 MPa), and flow rate was 1500 m³ / h (normal range 1400-1600 m³ / h). The sampling frequency was 2 kHz, lasting for 1 hour, and a total of 10,000 data points were collected. Kalman filtering was used for noise reduction, and the standard deviation of the vibration data decreased from 0.5 mm / s to 0.3 mm / s after filtering. Subsequently, K-means classification (k=3) was performed: normal data accounted for 85%, abnormal data accounted for 10%, and historical data accounted for 5%. The data was distributed to the industrial physics AI engine in JSON format: real-time data was transmitted to the baseline model, preprocessed data was transmitted to the state analysis algorithm, and classification data was transmitted to the mechanism calculation algorithm. In addition, multimodal data was transmitted to the operation and maintenance knowledge base.

[0116] Step 2: Establish a normal operating baseline model: Receive the real-time data from Step 1 through the industrial physics AI engine and train the model using the Transformer network;

[0117] The input sequence is 1024 bytes long with a feature dimension of 4. Features include vibration, temperature, pressure, and flow rate. The training dataset consists of 10,000 hours of historical normal operation data. The Adam optimizer has a learning rate of 1e-4, and the loss function used is MSE+KL. After training, the model outputs a normal vibration prediction curve (peak value < 4.5 mm / s) and an efficiency curve (> 95%). The baseline model is transmitted to the state analysis algorithm for comparison, and the prediction model is sent to the mechanism calculation algorithm for fusion.

[0118] Step 3: Perform operational status assessment and early warning: Receive the preprocessed data from Step 1 through the industrial physics AI engine, and perform deviation analysis in conjunction with the baseline model created in Step 2;

[0119] Calculate the residuals: The vibration residual is 0.3 mm / s (<3σ=0.9 mm / s, no warning); simulating an abnormal scenario, the vibration increases to 6.0 mm / s, the residual... This triggers an orange alert (potential fault). MK-SVM classification is used, employing an RBF kernel + linear kernel, achieving 98% accuracy. The output status report states: "Pump operation is normal, deviation rate 5%, under abnormal simulation, bearing wear may occur." The alert signal is transmitted to the mechanism calculation algorithm, the status report to the coordinating agent, and the evaluation results to the maintenance assistant agent. Simultaneously, feedback from the coordinating agent is received to adjust the threshold. To adapt to load changes;

[0120] Step 4: Conduct mechanism analysis and life prediction: The industrial physics AI engine receives the classification data from Step 1, the prediction model from Step 2, the early warning signals from Step 3, and the knowledge query from the operation and maintenance knowledge base, and runs the mechanism calculation algorithm.

[0121] The physical parameters obtained through the monitoring system and database are as follows:

[0122]

[0123] Calculate NPSH net positive suction head:

[0124]

[0125] If the depth exceeds the requirement by 8m, there is no risk of cavitation.

[0126] Calculate the life of bearing L10:

[0127]

[0128] The L10 lifetime is approximately 3 years. Using a Weibull distribution, [the following is taken as...]. , In 2018, the remaining lifespan was calculated and predicted to be 2.5 years with a confidence level of 90%, and a health score of 92 / 100 was output. The analysis and prediction results were transmitted to the coordinating agent, and the analysis report was sent to the operations and maintenance assistant agent. Simultaneously, feedback from the coordinating agent was received to optimize the Weibull parameters: adjustments were made. Up to 1.8;

[0129] Step 5: System Coordinating Agent Optimizes Multi-Agent Collaboration: The coordinating agent receives the state report from Step 3 and the analysis and prediction results from Step 4, and uses the QMIXMARL algorithm with a state space dimension of 10 and an action space of 5, including priority adjustment;

[0130] Under abnormal simulation, the coordination decision is as follows: the priority of mechanism analysis is increased by 20%, more computing resources are allocated, the CPU utilization is increased from 50% to 70%, the coordination decision is fed back to the industrial physics AI engine, the threshold of the state analysis algorithm is adjusted and updated, the parameters of the mechanism calculation algorithm are optimized, the operation and maintenance knowledge base is triggered to be updated, and the operation and maintenance assistant intelligent agent is dynamically interacted.

[0131] Step Six: Build and update the operation and maintenance knowledge base: Receive the multimodal data from Step One through the operation and maintenance knowledge base, embed entities using BERT, and build a knowledge graph with 5000 nodes, including the relationship between "bearing wear" and "abnormal vibration".

[0132] The system automatically mines new standards from IAEA reports, updates L10 guidelines, provides knowledge queries to the industrial physics AI engine, supplements NPSH data, L10 calculation results, and sends standards / experiences to the operations assistant agent. At the same time, it receives decision-triggered updates from the coordination agent and adds an anomaly simulation report.

[0133] Step 7: Provide intelligent operation and maintenance interaction: Receive the evaluation results from Step 3, the analysis report from Step 4, and the specifications / experiences from Step 6 through the operation and maintenance assistant intelligent agent. Use LLM to process user queries: "Pump current status?" Output suggestions: "Pump health 92%, bearing inspection recommended; remaining life 2.5 years." Integrate a gatekeeper mechanism to ensure output complies with IEC61513. Output user suggestions as the final decision: "Continue operation, planned maintenance in 6 months." Simultaneously, receive decision adjustment prompt templates from the coordinating intelligent agent.

[0134] In this embodiment, the system response time is less than 5 minutes, proving the system's practicality. The following is an example output table:

[0135]

[0136] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-agent operation and maintenance system for nuclear power plant circulating water pumps based on an industrial physics AI engine, characterized in that: The system comprises a data acquisition agent, an industrial physics AI engine, a coordination agent, an operation and maintenance knowledge base, and an operation and maintenance assistant agent. The industrial physics AI engine is the core component of the system, integrating a data-driven model and a physical mechanism model, specifically designed for nuclear power plant circulating water pumps. The industrial physics AI engine includes an equipment baseline model, a state analysis algorithm, and a mechanism calculation algorithm. After system startup, the specific workflow is as follows: S1: Data acquisition and preprocessing: The data acquisition agent collects and processes monitoring data from the nuclear power plant circulating water pumps and transmits the data to the industrial physics AI engine; S2: Establishing a normal operation baseline model: The industrial physics AI engine constructs a normal operation model of the circulating water pump based on real-time data provided by the data acquisition agent, and transmits the baseline model to the state analysis algorithm for deviation comparison and sends the predicted curve to the mechanism calculation algorithm for fusion evaluation. S3: Performs operational status assessment and early warning. It receives data from the industrial physical AI engine to obtain pre-processed data and baseline models of intelligent agents, and performs operational status analysis algorithms to conduct real-time comparisons and fault warnings. S4: Conduct mechanism analysis and life prediction. The industrial physics AI engine receives data to obtain the classification data of the intelligent agent, the baseline model prediction curve, the early warning signal obtained by the state assessment algorithm, and the knowledge query of the operation and maintenance knowledge base. S5: The system coordinates intelligent agents to optimize multi-agent collaboration, receiving status reports, health status, and lifespan predictions from the industrial physical AI engine through the coordinating intelligent agent; S6: The system builds and updates the operation and maintenance knowledge base, receiving multimodal data from intelligent agents through the operation and maintenance knowledge base, supporting knowledge storage and dynamic mining; S7: The system provides intelligent operation and maintenance interaction, receiving evaluation results and analysis reports from the industrial physical AI engine, as well as specifications and experience from the operation and maintenance knowledge base through the operation and maintenance assistant intelligent agent.

2. The nuclear power plant circulating water pump multi-agent operation and maintenance system based on an industrial physics AI engine according to claim 1, characterized in that: In step S1, the data acquisition agent serves as the system's data input layer, connecting to the sensor network of the nuclear power plant's circulating water pump. This network includes vibration sensors, temperature sensors, pressure sensors, and flow sensors. The collected real-time data includes time-series signals. The acquisition frequency is set according to the pump's operating characteristics, not lower than 1kHz to capture high-frequency fault signals. After acquisition, data preprocessing is performed: Kalman filtering is used for noise reduction, with the following formula: (1) (2) Among them, For posterior state estimation, For Kalman gain, Let H be the observed values, H be the observation matrix, P be the covariance matrix, and R be the observation noise covariance.

3. The nuclear power plant circulating water pump multi-agent operation and maintenance system based on an industrial physics AI engine according to claim 1, characterized in that: In S2, the industrial physics AI engine has the function of establishing a baseline model of the equipment. It uses real-time data provided by the data acquisition agent to establish a normal operation model of the circulating water pump. It uses a Transformer neural network to process time series data to capture multivariate dependencies and graph structure features. The Transformer model structure includes a multi-head self-attention mechanism and a feedforward network. The calculation formula for the attention mechanism is: (3) Where Q, K, and V are query, key, and value matrices, The key dimension is [batch_size, sequence_length, features]. The input data is a multi-channel time series with dimensions [batch_size, sequence_length, features]. Training uses historical normal operation data, and the goal is to generate a baseline model to simulate the behavior parameters of the pump under normal conditions. The model training uses the Adam optimizer, and the loss function is mean squared error combined with Kullback-Leibler divergence to improve distribution matching. (4) (5) Among them, To ensure interpretability, attention weights are visualized and key features are highlighted. After training, the baseline model outputs the normal operation threshold and prediction curve for subsequent evaluation.

4. The nuclear power plant circulating water pump multi-agent operation and maintenance system based on an industrial physics AI engine according to claim 1, characterized in that: In step S3, the industrial physics AI engine receives preprocessed data and a baseline model from the data acquisition agent, and runs a status evaluation algorithm; the comparison method includes deviation analysis: calculating the residual between real-time data and baseline prediction. If the residual exceeds the threshold, an early warning is triggered. A multi-core support vector machine is used to enhance evaluation accuracy. The MK-SVM classification hyperplane formula is: (6) Among them, Let m be the m-th kernel function, and let Gaussian kernel function be used. The core weights are optimized using quadratic programming for solving; the early warning mechanism includes multi-level alerts: yellow - slight deviation, orange - potential fault, and red - immediate shutdown. To improve interpretability, SHAP values ​​are used to interpret model decisions, the marginal contribution of features to model output is calculated, and the model is interpreted from both global and local levels. The output includes a status report.

5. The nuclear power plant circulating water pump multi-agent operation and maintenance system based on an industrial physics AI engine according to claim 1, characterized in that: In step S4, the industrial physics AI engine receives classification data, prediction models, early warning signals, and knowledge queries from the operation and maintenance knowledge base of the data acquisition intelligent agent to perform health and lifespan assessments. The industrial physics AI engine integrates mechanism calculation algorithms and possesses specific calculation methods for the nuclear power industry, including NPSH (Net Positive Suction Head) calculation. (7) Among them, Atmospheric pressure Vaporization pressure, which is temperature-dependent. For density, It is the acceleration due to gravity. For flow rate, 、 Suction height and friction loss are calculated based on pipe length; bearing life calculation is performed using the L10 criterion. (8) Where C is the basic dynamic load rating, P is the equivalent dynamic load, and the uncertainty of the lifetime prediction is quantified using the Weibull distribution: (9) Among them, For shape parameters, The scale parameter outputs a health score (0-100) and a remaining lifespan estimate; it includes an efficiency calculation model. (10) Where Q is the flow rate and H is the head. For shaft power; the industrial physics AI engine transmits the analysis and prediction results to the coordinating agent for decision arbitration, sends the analysis report to the operation and maintenance assistant agent for interactive output, and receives feedback decisions from the coordinating agent to optimize the analysis parameters.

6. The nuclear power plant circulating water pump multi-agent operation and maintenance system based on an industrial physics AI engine according to claim 1, characterized in that: In S5, the system coordinating agent resides in the coordination and knowledge layer. It receives state reports and analysis / prediction results from the industrial physics AI engine, and is responsible for communication and decision optimization among multiple agents. It employs a multi-agent reinforcement learning algorithm to achieve collaboration. The state space includes the outputs of each agent, the action space is resource allocation, and the value function is decomposed as follows: (11) Among them, For a single agent, the Q-value. For the hybrid function, s is the global state, and coordination ensures system autonomy; the system coordinating agent provides knowledge queries to the industrial physics AI engine for physical model supplementation, sends specifications and experience to the operation and maintenance assistant agent for generating suggestions, and receives decisions from the coordinating agent to trigger knowledge updates.

7. The nuclear power plant circulating water pump multi-agent operation and maintenance system based on an industrial physics AI engine according to claim 1, characterized in that: In S6, the operation and maintenance knowledge base is constructed as a dynamic knowledge graph, storing operation records, national industry standards, accident reports and expert experience. It supports multimodal data integration, including data in the form of images and videos. The update mechanism uses knowledge mining algorithms to automatically extract new knowledge from external databases, and the query uses the SPARQL graph traversal algorithm.

8. The nuclear power plant circulating water pump multi-agent operation and maintenance system based on an industrial physics AI engine according to claim 1, characterized in that: In S7, the operation and maintenance assistant agent resides in the interaction and integration layer. It receives the evaluation results and analysis reports from the industrial physics AI engine, as well as the specifications and experience from the operation and maintenance knowledge base. Based on the large language model LLM, it provides natural language interaction, accepts user queries, and generates output using prompt engineering: Prompt = "Based on the knowledge base and evaluation results, provide operation and maintenance suggestions for nuclear power plant circulating water pumps: {query}". It integrates a gatekeeper mechanism: rule-based filtering, to ensure that the output complies with nuclear safety specifications. It also uses the XAI tool to interpret LLM decisions and supports scenario simulation: the user inputs assumptions, and LLM generates reports. The operation and maintenance assistant agent outputs user suggestions as the final operation and maintenance decision and receives decisions from the coordinating agent to adjust the interaction logic.

Images