A method and system for intelligent monitoring and optimization of nuclear power plant production business processes

Abstract

The present invention discloses a method for intelligent monitoring and optimization of nuclear power plant production business processes, which relates to the technical field of artificial intelligence, including building a hierarchical data fusion architecture; obtaining expert annotation feedback based on active learning and human-computer interaction technology, building an anomaly detection model for detecting production process anomalies; building a Bayesian optimization process decision model, and when an anomaly is detected, combining deep neural networks and reinforcement learning technology to give an optimization decision plan; developing a digital twin system based on virtual / augmented reality to present process monitoring status and optimization decision plans; according to the optimization decision plan, automatically issuing control instructions to the production site to form an automated closed loop of detection, decision-making, and execution. Through the efficient integration and closed-loop operation of anomaly detection, decision analysis, and execution control, the present invention can detect abnormal situations in production in real time, quickly make correct optimization decisions, and automatically execute them, greatly improving the safety and production efficiency of nuclear power plants.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and in particular to a method and system for intelligently monitoring and optimizing production processes of a nuclear power plant. Background Art

[0002] Nuclear power plant operations place extremely high demands on safety, reliability, and efficiency. Traditional production monitoring and optimization decisions rely primarily on manual experience, which presents several significant drawbacks. First, traditional methods rely primarily on limited historical data and manually extracted empirical rules, making it difficult to fully leverage the deep knowledge embedded in massive amounts of production and operation data. This results in insufficient accuracy and foresight in monitoring and decision-making. Second, the manual decision-making process is complex and inefficient, often preventing timely and appropriate responses when anomalies occur, leading to delayed decisions. Furthermore, manual decision-making relies too heavily on limited personal experience, resulting in significant subjective bias and uncertainty.

[0003] On the other hand, due to the complex and ever-changing production environment, manually constructed rules or models often have certain scenario limitations and lack sufficient versatility and robustness. Once the production environment or working conditions change, manual intervention is still required to build a new model, making effective cross-scenario migration impossible. Summary of the Invention

[0004] In view of the fact that traditional manual production monitoring and optimization decision-making methods can no longer fully meet the intelligent requirements of modern nuclear power production, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to use new technologies such as artificial intelligence and big data to realize intelligent abnormal monitoring, optimized decision-making and automated closed-loop control of the production process of nuclear power plants, thereby improving the safety, reliability and efficiency of production operations.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention provides an intelligent monitoring and optimization method for the production business process of a nuclear power plant, which includes constructing a layered data fusion architecture, deploying intelligent edge acquisition gateway devices in various production links of the nuclear power plant, and establishing a big data center in the cloud; obtaining expert annotation feedback based on active learning and human-computer interaction technology, and constructing an anomaly detection model for detecting production process anomalies; constructing a Bayesian optimization process decision model, and when an anomaly is detected, combining deep neural networks and reinforcement learning technology to provide an optimization decision plan; developing a digital twin system based on virtual / augmented reality to present process monitoring status and optimization decision plans; according to the optimization decision plan, automatically issuing control instructions to the production site to form an automated closed loop of detection, decision-making, and execution; when a major anomaly occurs, the system automatically explores the root cause of the anomaly and triggers local incremental updates to correct model drift; using domain adaptation technology, the optimization model of the mature site is migrated to the newly commissioned site for deployment.

[0008] As a preferred solution of the method for intelligent monitoring and optimization of nuclear power plant production business processes described in the present invention, the following steps are included: constructing an anomaly detection model: obtaining original operating data collected during the production process of the nuclear power plant and performing preprocessing; extracting a core feature subset closely related to anomaly detection from the preprocessed data based on domain knowledge and anomaly correlation; using all labeled historical data as the initial data set, and dividing it into a training set and a validation set according to a preset ratio; using a deep learning method and an ensemble learning method to train multiple initial anomaly detection models on the training set; evaluating the anomaly detection models in the model set on the validation set, and selecting the one with the best performance as the basic anomaly detection model; and selecting the selected The basic anomaly detection model is applied to unlabeled online data streams to automatically detect anomalies in the production process. Based on the confidence of the model's online predictions, an active learning strategy is used to intelligently extract suspected anomaly instances with a confidence level lower than a first preset threshold, and the extracted suspected anomaly instances are presented to experts for secondary confirmation. Using incremental learning technology, newly acquired labeled instances are incorporated into the training set, and the anomaly detection model is updated and optimized online. The loop process is repeated to continuously optimize and improve the model's anomaly detection capabilities, and the optimized complex anomaly detection model is compressed. Based on the hardware environment of the edge device, the computational graph structure of the compression model is optimized, and the optimized lightweight model is deployed to the edge gateway device.

[0009] As a preferred solution of the intelligent monitoring and optimization method for the production business process of a nuclear power plant described in the present invention, the method includes the following steps: using a deep learning method and an integrated learning method to train multiple initial anomaly detection models: designing and constructing a deep convolutional neural network architecture including a convolutional layer, a pooling layer and a fully connected layer, which is used to automatically learn and extract feature patterns from training data; designing and constructing a deep recurrent neural network architecture suitable for sequence data input to learn and capture the temporal dynamic laws of production data; dividing the preprocessed training data into samples, and inputting them into a deep convolutional neural network and a deep recurrent neural network respectively for end-to-end training; using a Boosting or Bagging integration method to integrate the trained multiple deep neural network models into a composite anomaly detector; using a random forest or ensemble tree algorithm to construct a decision tree integration model on the training data as another initial anomaly detection model; designing a variational autoencoder neural network architecture to train an anomaly detection classifier model in the latent feature space of the autoencoder; and combining the trained composite anomaly detector, decision tree integration model, and anomaly detection classifier model to form a model set.

[0010] As a preferred embodiment of the method for intelligent monitoring and optimization of nuclear power plant production business processes described in the present invention, the following steps are included: constructing a Bayesian optimization process decision model, including the following steps: constructing a Bayesian network or other probabilistic graphical model based on expert knowledge and historical decision data to represent a prior probability model reflecting the relationship between process status and optimization decisions; modeling the optimization decision process as a Markov decision process, defining a state space, a discrete action space, and a reward function; using the output of an anomaly detection model and production operation data as state data, and performing standardization, missing value processing, and feature engineering; using a deep neural network to encode high-dimensional original states into low-dimensional continuous state representations as input to a reinforcement learning model; using a deep reinforcement learning algorithm to train an intelligent agent in a simulated environment to obtain a reinforcement learning model; combining the prior probability model with the reinforcement learning model to perform Bayesian reasoning on optional decision plans when an anomaly is detected, and outputting the optimal decision plan with the highest expected reward; providing a human-computer interaction interface for experts to review and adjust the decision plan, using online learning technology to continuously incorporate new decision samples and update the prior model and reinforcement learning model; optimizing and deploying the decision model, and automatically distributing or providing the decision results to on-site operators for execution.

[0011] As a preferred solution of the intelligent monitoring and optimization method for the production business process of a nuclear power plant described in the present invention, the method includes the following steps: constructing a simulation environment model to describe the dynamic evolution process of the nuclear power plant production process and abnormal scenarios; selecting an applicable deep reinforcement learning algorithm framework for continuous state space and discrete action space, and designing the algorithm's network structure and training process; using a deep neural network to encode high-dimensional original state data to obtain a low-dimensional continuous state vector as the input of the reinforcement learning agent; selecting a fully connected output layer based on the discrete action space to parameterize the optional decision-making scheme as the output of the reinforcement learning agent; starting the reinforcement learning training process in the simulation environment, the intelligent agent interacts with the environment to obtain a state-action-reward sequence, and caching the training data using experience replay; regularly performing parameter updates according to the update strategy of the DQN algorithm, so that the reinforcement learning agent model gradually optimizes the decision-making behavior under the guidance of the reward function; monitoring training indicators, and evaluating the model performance in a verification environment, and saving the reinforcement learning agent model after reaching the expected level.

[0012] As a preferred solution of the intelligent monitoring and optimization method for the production business process of a nuclear power plant described in the present invention, the following steps are included: forming an automated closed loop of detection, decision-making and execution: ensuring the network interconnection between the decision execution system and the production control system and the anomaly detection system, and establishing a secure data exchange channel; defining a standardized control instruction set, and converting the optimization decision into an executable control instruction sequence; modeling the production process as a state machine, and automatically executing the corresponding decision plan and control instruction sequence based on the input and current state of the anomaly detection system; issuing the control instruction to the production control system for execution, and at the same time feeding back the execution feedback and real-time monitoring data to the anomaly detection and decision model; building a closed-loop mechanism between anomaly detection, decision analysis and control execution, and an automated closed loop of detection, decision-making and execution; recording the entire process log for auditing, and supporting manual intervention and adjustment.

[0013] As a preferred solution of the intelligent monitoring and optimization method for nuclear power plant production business processes described in the present invention, triggering local incremental updates to correct model drift includes the following steps: setting anomaly severity criteria and thresholds, and triggering an anomaly root cause mining process when it is detected that the anomaly severity exceeds the threshold; recording key operating data, operation logs, and monitoring data before and after the anomaly occurs, and automatically extracting a core feature subset related to the anomaly; constructing a Bayesian network model of the anomaly cause based on domain knowledge and historical data, and performing probabilistic reasoning on the model using collected evidence data to obtain the most likely anomaly root cause; updating the knowledge base of the anomaly detection model and decision model based on the results of the anomaly root cause analysis, and incorporating newly discovered anomaly samples into the training data set; adopting an online incremental learning algorithm, updating model parameters using newly discovered training samples, and performing real-time local optimization of the anomaly detection model and decision model; evaluating the performance of the updated model on a validation set, and putting it into online operation if the performance meets the requirements, otherwise returning to the previous step for continued optimization; continuously monitoring the online operation of the model, and triggering incremental updates again if performance degradation is detected, forming a closed-loop learning cycle of anomaly detection, root cause mining, and model update.

[0014] In the second aspect, an embodiment of the present invention provides an intelligent monitoring and optimization system for the production business process of a nuclear power plant, which includes a layered architecture module for constructing a layered data fusion architecture, deploying intelligent edge acquisition gateway devices in various production links of the nuclear power plant, and establishing a big data center in the cloud; a model construction module for obtaining expert annotation feedback based on active learning and human-computer interaction technology, and constructing an anomaly detection model for detecting production process anomalies; a decision-making scheme module for constructing a Bayesian optimization process decision model, and when an anomaly is detected, combining deep neural networks and reinforcement learning technology to provide an optimization decision scheme; a twin system module for developing a digital twin system based on virtual / augmented reality to present process monitoring status and optimization decision schemes; an instruction issuing module for automatically issuing control instructions to the production site according to the optimization decision scheme, forming an automated closed loop of detection, decision-making, and execution; a local incremental module for automatically digging out the root cause of the anomaly when a major anomaly occurs, triggering a local incremental update to correct the model drift; and an adaptive migration module for using domain adaptive technology to migrate the optimization model of a mature site to a newly commissioned site for deployment.

[0015] In a third aspect, an embodiment of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the intelligent monitoring and optimization method for the production business process of a nuclear power plant as described in the first aspect of the present invention are implemented.

[0016] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the method for intelligent monitoring and optimization of nuclear power plant production business processes as described in the first aspect of the present invention are implemented.

[0017] The beneficial effects of the present invention are as follows: through the efficient integration and closed-loop operation of anomaly detection, decision analysis, and execution control, the present invention can detect abnormal situations in production in real time, quickly make correct optimization decisions and automatically execute them, thereby greatly improving the safety and production efficiency of nuclear power plants; using deep learning, integrated learning, active learning and other technologies, it can automatically learn abnormal patterns from complex production big data and continuously improve anomaly detection capabilities; combining Bayesian networks and reinforcement learning and other technologies, it can provide the optimal decision-making plan that meets multi-objective optimization; obtaining expert feedback through human-computer interaction, and using incremental learning technology to continuously optimize the model, thereby achieving continuous evolution of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 The framework flow chart of the intelligent monitoring and optimization method for the production business process of nuclear power plants.

[0020] Figure 2 The production process anomaly detection process is an intelligent monitoring and optimization method for the production business process of nuclear power plants.

[0021] Figure 3 A diagram of computer equipment used to develop an intelligent monitoring and optimization method for business processes in nuclear power plants. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0025] Example 1

[0026] Reference Figures 1 to 3 , which is the first embodiment of the present invention, provides a method for intelligent monitoring and optimization of production business processes of a nuclear power plant, including:

[0027] S1: Build a layered data fusion architecture, deploy intelligent edge acquisition gateway devices in various production links of nuclear power plants, and establish a big data center in the cloud.

[0028] S2: Based on active learning and human-computer interaction technology, we obtain expert annotation feedback and build an anomaly detection model to detect anomalies in the production process.

[0029] Specifically, the following steps are included:

[0030] S2.1: Obtain the original operating data collected during the production process of the nuclear power plant and perform preprocessing.

[0031] Specifically, preprocessing includes cleaning outliers and missing values ​​and standardizing heterogeneous data.

[0032] S2.2: Based on domain knowledge and anomaly correlation, a subset of core features closely related to anomaly detection is extracted from the preprocessed data.

[0033] Preferably, the features closely related to anomaly detection include key equipment operating parameters, process control indicators, radiation monitoring data, containment compartment detection data, and online detection and diagnostic data (coolant activity, core neutron flux distribution, and pipeline medium leakage detection).

[0034] S2.3: Use all labeled historical data as the initial dataset and divide it into training set and validation set according to the preset ratio.

[0035] S2.4: On the training set, use deep learning methods and ensemble learning methods to train multiple initial anomaly detection models.

[0036] Specifically, a deep convolutional neural network architecture is designed and constructed, including convolutional layers, pooling layers, and fully connected layers, which are used to automatically learn and extract feature patterns from training data; a deep recurrent neural network architecture is designed and constructed, which is suitable for sequence data input and learns to capture the temporal dynamics of production data; the preprocessed training data is divided into samples, which are input into a deep convolutional neural network and a deep recurrent neural network respectively for end-to-end training; the Boosting or Bagging integration method is used to integrate multiple trained deep neural network models into a composite anomaly detector; a random forest or ensemble tree algorithm is used to build a decision tree integration model on the training data as another initial anomaly detection model; a variational autoencoder neural network architecture is designed to train an anomaly detection classifier model in the latent feature space of the autoencoder; the trained composite anomaly detector, decision tree integration model, and anomaly detection classifier model are combined to form a model set.

[0037] S2.5: Evaluate the anomaly detection models in the model ensemble on the validation set and select the one with the best performance as the base anomaly detection model.

[0038] S2.6: Apply the selected basic anomaly detection model to unlabeled online data streams to automatically detect anomalies in the production process.

[0039] S2.7: Based on the confidence of the model's online prediction, use an active learning strategy to intelligently extract suspected anomaly instances with a confidence lower than a first preset threshold, and present the extracted suspected anomaly instances to experts for secondary confirmation.

[0040] S2.8: Use incremental learning technology to incorporate newly acquired labeled instances into the training set and update and optimize the anomaly detection model online.

[0041] S2.9: Repeat the loop process from S2.6 to S2.8 to continuously optimize and improve the model's anomaly detection capabilities, and perform model compression on the optimized complex anomaly detection model.

[0042] S2.10: Optimize the computational graph structure of the compression model based on the hardware environment of the edge device, and deploy the optimized lightweight model to the edge gateway device.

[0043] S3: Build a Bayesian optimization process decision model, and when an anomaly is detected, combine deep neural networks and reinforcement learning technology to provide an optimized decision solution.

[0044] Specifically, the following steps are included:

[0045] S3.1: Based on expert knowledge and historical decision data, construct a Bayesian network or other probabilistic graphical model to represent the prior probability model that reflects the relationship between process status and optimization decisions.

[0046] Specifically, historical operating data of nuclear power plants are collected and organized to construct an original data set; experts in the nuclear power field are invited to determine the key factors affecting process flow and decision-making and their causal relationships based on expert knowledge and experience; combining expert knowledge and data analysis results, potential causal structures are automatically mined from the data through structural learning algorithms to construct an initial Bayesian network topology; parameter learning is performed on the initial Bayesian network, and network parameters are estimated from historical data to obtain a complete parameterized probability model reflecting the state-decision relationship; a scoring function is designed to evaluate the predictive performance of the Bayesian network on the validation set, and a structural reduction strategy is used to optimize the network topology; an expert review is conducted on the optimized Bayesian network, and human-computer collaboration is used to improve the causal relationships and conditional probability distributions in the network to form the final prior probability model; a visual interface based on the prior probability model is developed to intuitively display the relationship between state and decision-making, providing support for subsequent decision-making reasoning.

[0047] S3.2: Model the optimization decision process as a Markov decision process, defining the state space (abnormalities and device states), discrete action space (optional decision plans), and reward function (optimization objective).

[0048] Preferably, the state space represents the various possible states of the system, the action space refers to a series of operational behaviors that the agent can perform (such as maintaining the current state, adjusting control parameters, switching to backup equipment, load shedding, orderly shutdown, emergency cooling, etc.), and the reward function is used to characterize the goal of the agent, that is, the long-term cumulative reward that the action sequence is expected to maximize. The reward function of the present invention can take into account system operation safety, production efficiency, economic cost, and other optimization goals.

[0049] S3.3: Use the output of the anomaly detection model and production operation data as status data, and perform normalization, missing value processing, and feature engineering.

[0050] S3.4: Use deep neural networks to encode high-dimensional raw states into low-dimensional continuous state representations as input to reinforcement learning models.

[0051] Preferably, the network can use a structure such as a convolutional network or a gated recurrent unit to extract features of anomalies and device states.

[0052] S3.5: Use deep reinforcement learning algorithms to train the agent in a simulated environment to obtain a reinforcement learning model.

[0053] Specifically, a simulation environment model is constructed to describe the dynamic evolution process of the nuclear power plant production process and abnormal scenarios; for the continuous state space and discrete action space, an applicable deep reinforcement learning algorithm framework is selected, and the algorithm's network structure and training process are designed; the deep neural network in S3.4 is used to encode the high-dimensional original state data to obtain a low-dimensional continuous state vector as the input of the reinforcement learning agent; according to the discrete action space designed in S3.2, a fully connected output layer is selected to parameterize the optional decision-making schemes as the output of the reinforcement learning agent; the reinforcement learning training process is started in the simulation environment, and the agent interacts with the environment to obtain the state-action-reward sequence, and the training data is cached using experience replay; parameter updates are performed regularly according to the update strategy of the DQN algorithm, so that the reinforcement learning agent model gradually optimizes the decision-making behavior under the guidance of the reward function; the training indicators are monitored, and the model performance is evaluated in a verification environment. After reaching the expected level, the reinforcement learning agent model is saved.

[0054] S3.6: Combine the prior probability model with the reinforcement learning model to perform Bayesian reasoning on the available decision options when an anomaly is detected, and output the optimal decision option with the highest expected reward.

[0055] S3.7: Provide a human-computer interaction interface for experts to review and adjust decision plans, use online learning technology to continuously incorporate new decision samples, and update prior models and reinforcement learning models.

[0056] S3.8: Optimize and deploy the decision model, and automatically distribute or provide the decision results to on-site operators for execution.

[0057] S4: Develop a digital twin system based on virtual / augmented reality to present process monitoring status and optimization decision-making solutions.

[0058] S5: Based on the optimized decision-making plan, control instructions are automatically issued to the production site, forming an automated closed loop of detection, decision-making, and execution.

[0059] Specifically, ensure the network interconnection between the decision-making execution system and the production control system and the anomaly detection system, and establish a secure data exchange channel; define a standardized control instruction set to convert optimization decisions into executable control instruction sequences; model the production process as a state machine, and automatically execute the corresponding decision-making plan and control instruction sequence based on the input and current status of the anomaly detection system; send the control instructions to the production control system for execution, and at the same time feed back the execution feedback and real-time monitoring data to the anomaly detection and decision model; build a closed-loop mechanism between anomaly detection, decision analysis, and control execution, and automate the closed loop of detection, decision-making, and execution; record the entire process log for auditing, and support manual intervention and adjustment when necessary.

[0060] S6: When a major anomaly occurs, the system automatically explores the root cause of the anomaly and triggers local incremental updates to correct model drift.

[0061] Specifically, set the severity criteria and thresholds of the anomaly. When the severity of the anomaly is detected to exceed the threshold, the anomaly root cause mining process is triggered; record the key operating data, operation logs, and monitoring data before and after the anomaly occurs, and automatically extract the core feature subset related to the anomaly; based on domain knowledge and historical data, build a Bayesian network model of the cause of the anomaly, and use the collected evidence data to perform probabilistic reasoning on the model to obtain the most likely root cause of the anomaly; based on the results of the anomaly root cause analysis, update the knowledge base of the anomaly detection model and decision model, and incorporate the newly discovered anomaly samples into the training data set; adopt an online incremental learning algorithm, use the newly added training samples to update the model parameters, and perform local optimization of the anomaly detection model and decision model in real time; evaluate the performance of the updated model on the validation set, and put it into online operation if the performance meets the requirements, otherwise return to the previous step to continue optimization; continuously monitor the online operation of the model, and trigger incremental updates again if performance degradation is found, forming a closed-loop anomaly detection, root cause mining, and model update cycle learning.

[0062] S7: Use domain adaptation technology to migrate the optimization model of mature sites to newly commissioned sites for deployment.

[0063] Furthermore, this embodiment also provides an intelligent monitoring and optimization system for the production business process of a nuclear power plant, including a layered architecture module for building a layered data fusion architecture, deploying intelligent edge acquisition gateway devices in various production links of the nuclear power plant, and establishing a big data center in the cloud; a model construction module for obtaining expert annotation feedback based on active learning and human-computer interaction technology, building an anomaly detection model for detecting production process anomalies; a decision-making solution module for building a Bayesian optimization process decision model, and when an anomaly is detected, combining deep neural networks and reinforcement learning technology to provide an optimization decision solution; a twin system module for developing a digital twin system based on virtual / augmented reality to present process monitoring status and optimization decision solutions; an instruction issuance module for automatically issuing control instructions to the production site according to the optimization decision solution, forming an automated closed loop of detection, decision-making, and execution; a local incremental module for automatically discovering the root cause of the anomaly when a major anomaly occurs, triggering a local incremental update to correct model drift; and an adaptive migration module for using domain adaptive technology to migrate the optimization model of a mature site to a newly commissioned site for deployment.

[0064] This embodiment also provides a computer device suitable for the intelligent monitoring and optimization method of nuclear power plant production business processes, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the intelligent monitoring and optimization method of nuclear power plant production business processes proposed in the above embodiment.

[0065] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0066] This embodiment also provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the method for intelligently monitoring and optimizing the production process of a nuclear power plant as proposed in the above embodiment is implemented.

[0067] In summary, the present invention can detect abnormal situations in production in real time through the efficient integration and closed-loop operation of anomaly detection, decision analysis, and execution control, quickly make correct optimization decisions and automatically execute them, thereby greatly improving the safety and production efficiency of nuclear power plants; using deep learning, integrated learning, active learning and other technologies, it can automatically learn abnormal patterns from complex production big data and continuously improve anomaly detection capabilities; combining Bayesian networks and reinforcement learning and other technologies, it can provide optimal decision-making solutions that meet multi-objective optimization; obtaining expert feedback through human-computer interaction, and using incremental learning technology to continuously optimize the model, thereby achieving continuous evolution of the model.

[0068] Example 2

[0069] Reference Figures 1 to 3 , which is the second embodiment of the present invention, provides an intelligent monitoring and optimization method for the production business process of a nuclear power plant. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0070] Specifically, a virtualized nuclear power plant conducted a six-month full-process simulation test on this solution. First, a layered big data architecture was built, 32 acquisition gateway devices were deployed in the production process, and a big data center was built in the cloud. Secondly, the anomaly detection model was trained based on three years of historical operating data (including 50 million labeled samples). The feature dimension used in training was 256, covering multiple fields such as key equipment parameters and process control indicators. In the first phase, six deep convolutional networks, four recurrent networks, and three variational autoencoders were constructed. The integrated composite anomaly detection model had an F1 score of 0.928 on the validation set. The detector model was compressed from 350M to 45M, and the inference time was reduced from 32ms to 9ms, meeting the requirements of edge devices. The compressed lightweight model was deployed to the on-site gateway.

[0071] Furthermore, an optimization decision model was constructed in parallel. The prior probability model was constructed based on domain knowledge and 10 years of historical data (1 million samples). The modeling variables included 73 state variables and 27 control decision variables. The final network contained 128 nodes and achieved a log-likelihood score of 0.86. The simulation environment covered 18 typical abnormal scenarios, with a 512-dimensional state space and 15 action spaces. The reinforcement learning model used a 4-layer convolutional network to encode the state and a 3-layer fully connected network to model the actions. The model was trained in the environment for 12 million iterations using the Proportional Proportional Optimization (PPO) algorithm, achieving an average score of 0.91 in the test scenarios.

[0072] Furthermore, a compressed anomaly detector and optimized decision-making model were integrated and deployed on 32 edge gateways. An augmented reality digital twin system was developed to visually display monitoring data and decision-making processes. A standard control instruction set was developed, and the production process was modeled as a state machine, ensuring high-bandwidth connectivity between gateway devices and the control system.

[0073] Furthermore, over the next nine months of continuous operation, 3,762 anomalies were simulated and injected, with an average detection delay of 9.6 seconds and a detection rate of 94.2%. For 2,019 low-confidence judgment results, the system intelligently sampled 529 of them, submitted them for expert review and annotation, and continuously optimized the detection model. For 356 major anomalies, the system conducted root cause analysis and incremental updates to model parameters, correcting model drift. The automated decision response took an average of 5.3 seconds, and 95.7% of anomalies were resolved within 30 minutes. The system continued to operate for 162 hours without human intervention.

[0074] After long-term operation, the model performance has gradually improved, the detector F1 score has increased from 0.928 to 0.962, and the average decision score has increased from 0.91 to 0.97.

[0075] Preferably, the comparison indicators of the present invention and the traditional method are shown in Table 1.

[0076] Table 1 Comparative indexes of the present invention and the traditional method

[0077] index The present invention Traditional methods Anomaly detection accuracy 95.6％ 75％～88％ Detection response time 9.6s >5 minutes Decision-making time 5.3s >30 minutes Abnormal recovery time 95.7% of abnormalities recovered within 30 minutes 1 to 8 hours Model update cycle Continuous online growth June to December Deployment complexity Low (automation) high

[0078] Specifically, as shown in Table 1, the present invention integrates multiple advanced artificial intelligence technologies and has a high detection accuracy rate (95.6%), which is significantly better than traditional methods (75% to 88%). Rapid response is a key advantage of this solution, with the detection response time shortened from minutes to 9.6 seconds, and the decision-making time reduced from several hours to 5.3 seconds. Through optimized decision-making and automated closed-loop execution, 95.7% of anomalies can be recovered within 30 minutes, significantly shortening the duration of the abnormal state. The use of an online incremental learning mechanism can continuously optimize and improve the model, avoiding the problem of outdated and stable models in traditional methods. Based on standardized workflows, the deployment and migration complexity of this solution is low, reducing the cost of engineering implementation.

[0079] It should be noted that the above embodiments are only used to illustrate the technical solutions 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 preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for intelligent monitoring and optimization of production process of a nuclear power plant, characterized by: include, Build a layered data fusion architecture, deploy intelligent edge data collection gateway devices in various production links of nuclear power plants, and establish a big data center in the cloud; Based on active learning and human-computer interaction technology, we obtain expert annotation feedback and build an anomaly detection model to detect production process anomalies. Build a Bayesian optimization process decision model, and when anomalies are detected, combine deep neural networks and reinforcement learning technology to provide optimized decision solutions; Develop digital twin systems based on virtual / augmented reality to present process monitoring status and optimize decision-making solutions; According to the optimized decision-making plan, control instructions are automatically issued to the production site, forming an automated closed loop of detection, decision-making, and execution; When a major anomaly occurs, the system automatically explores the root cause of the anomaly and triggers a local incremental update to correct model drift. Using domain adaptation technology, the optimization model of mature sites is migrated to newly commissioned sites for deployment; The construction of the anomaly detection model includes the following steps: Acquire the original operating data collected during the production process of the nuclear power plant and perform preprocessing; Based on domain knowledge and anomaly correlation, a subset of core features closely related to anomaly detection is extracted from preprocessed data; All labeled historical data are used as the initial data set, and the training set and validation set are divided according to the preset ratio; On the training set, multiple initial anomaly detection models are trained using deep learning methods and ensemble learning methods; Evaluate the anomaly detection models in the model set on the validation set and select the one with the best performance as the base anomaly detection model; Apply the selected basic anomaly detection model to unlabeled online data streams to automatically detect anomalies in production processes; Based on the confidence of the model's online prediction, an active learning strategy is used to intelligently extract suspected anomaly instances with a confidence lower than a first preset threshold, and the extracted suspected anomaly instances are presented to experts for secondary confirmation; Use incremental learning technology to incorporate newly acquired labeled instances into the training set and update and optimize the anomaly detection model online; Repeat the loop process to continuously optimize and improve the model's anomaly detection capabilities, and compress the optimized complex anomaly detection model; Based on the hardware environment of the edge device, the computational graph structure of the compression model is optimized, and the optimized lightweight model is deployed to the edge gateway device.

2. The method for intelligent monitoring and optimization of nuclear power plant production process according to claim 1, characterized in that: The method of training multiple initial anomaly detection models using a deep learning method and an ensemble learning method includes the following steps: Design and build a deep convolutional neural network architecture consisting of convolutional layers, pooling layers, and fully connected layers to automatically learn and extract feature patterns from training data; Design and build deep recurrent neural network architectures suitable for sequential data inputs and learn to capture the temporal dynamics of production data; The preprocessed training data is divided into samples and input into deep convolutional neural networks and deep recurrent neural networks for end-to-end training; Using Boosting or Bagging ensemble methods, multiple trained deep neural network models are integrated into a composite anomaly detector; Use the random forest or ensemble tree algorithm to build a decision tree ensemble model on the training data as another initial anomaly detection model; Design a variational autoencoder neural network architecture and train an anomaly detection classifier model in the autoencoder's latent feature space; The trained composite anomaly detector, decision tree ensemble model, and anomaly detection classifier model are combined to form a model set.

3. The method for intelligent monitoring and optimization of nuclear power plant production process according to claim 2, characterized in that: The Bayesian optimization process decision model construction includes the following steps: Based on expert knowledge and historical decision data, a Bayesian network or other probabilistic graphical model is constructed to represent the prior probability model reflecting the relationship between process status and optimization decision; Model the optimization decision process as a Markov decision process, defining the state space, discrete action space and reward function; Use the output of the anomaly detection model and production operation data as status data, and perform normalization, missing value processing, and feature engineering; Use deep neural networks to encode high-dimensional raw states into low-dimensional continuous state representations as input to the reinforcement learning model; Use deep reinforcement learning algorithms to train intelligent agents in a simulated environment to obtain a reinforcement learning model; Combining the prior probability model with the reinforcement learning model, Bayesian reasoning is performed on the available decision options when an anomaly is detected, and the optimal decision option with the highest expected reward is output; Provide a human-computer interaction interface for experts to review and adjust decision plans, use online learning technology to continuously incorporate new decision samples, and update prior models and reinforcement learning models; Optimize and deploy decision models, and automatically distribute or provide decision results to on-site operators for execution.

4. The method for intelligent monitoring and optimization of nuclear power plant production process according to claim 3, characterized in that: The method of using a deep reinforcement learning algorithm to train an intelligent agent in a simulated environment includes the following steps: Construct a simulation environment model to describe the dynamic evolution of nuclear power plant production processes and abnormal scenarios; For continuous state space and discrete action space, select the appropriate deep reinforcement learning algorithm framework and design the algorithm's network structure and training process; Use deep neural networks to encode high-dimensional raw state data to obtain low-dimensional continuous state vectors as input for reinforcement learning agents; According to the discrete action space, a fully connected output layer is selected to parameterize the optional decision options as the output of the reinforcement learning agent; Initiate reinforcement learning training in a simulated environment. The agent interacts with the environment to obtain state-action-reward sequences and caches training data using experience replay. Regularly update parameters according to the DQN algorithm's update strategy, so that the reinforcement learning agent model gradually optimizes its decision-making behavior under the guidance of the reward function; Monitor training metrics and evaluate model performance on a validation environment, saving the reinforcement learning agent model once it reaches the expected level.

5. The method for intelligent monitoring and optimization of production process of a nuclear power plant according to claim 4, characterized in that: The automated closed loop of detection, decision-making, and execution includes the following steps: Ensure the network interconnection between the decision-making execution system, production control system, and anomaly detection system, and establish a secure data exchange channel; Define standardized control instruction sets to convert optimization decisions into executable control instruction sequences; Model the production process as a state machine, and automatically execute the corresponding decision-making plan and control instruction sequence based on the input and current state of the anomaly detection system; Send control instructions to the production control system for execution, and feed execution feedback and real-time monitoring data into anomaly detection and decision-making models; Build a closed-loop mechanism between anomaly detection, decision analysis, and control execution, and automate the closed loop of detection, decision-making, and execution; Record the entire process log for auditing and support manual intervention and adjustment.

6. The method for intelligent monitoring and optimization of production process of a nuclear power plant according to claim 5, characterized in that: The triggering of local incremental update to correct model drift includes the following steps: Set the abnormality severity criteria and thresholds. When the abnormality severity exceeds the threshold, the abnormality root cause mining process is triggered. Record key operational data, operation logs, and monitoring data before and after an anomaly occurs, and automatically extract a subset of core features related to the anomaly; Based on domain knowledge and historical data, a Bayesian network model of the cause of the anomaly is constructed. The model is probabilistically reasoned using the collected evidence data to obtain the most likely root cause of the anomaly. Based on the results of anomaly root cause analysis, the knowledge base of the anomaly detection model and decision model is updated, and newly discovered anomaly samples are incorporated into the training dataset; Adopting online incremental learning algorithm, using newly added training samples to update model parameters, and performing local optimization of anomaly detection model and decision model in real time; Evaluate the performance of the updated model on the validation set. If the performance meets the requirements, run it online. Otherwise, return to the previous step to continue optimization. Continuously monitor the online operation of the model and trigger incremental updates again once performance degradation is detected, forming a closed-loop learning cycle of anomaly detection, root cause mining, and model update.

7. An intelligent monitoring and optimization system for nuclear power plant production business processes, based on the intelligent monitoring and optimization method for nuclear power plant production business processes according to any one of claims 1 to 6, characterized in that: Also includes, A layered architecture module is used to build a layered data fusion architecture, deploy intelligent edge acquisition gateway devices in various production links of nuclear power plants, and establish a big data center in the cloud; The model building module is used to obtain expert annotation feedback based on active learning and human-computer interaction technology to build anomaly detection models for detecting production process anomalies; The decision-making solution module is used to build a Bayesian optimization process decision model. When an anomaly is detected, it combines deep neural networks and reinforcement learning technology to provide an optimized decision solution; Twin system module, used to develop digital twin systems based on virtual / augmented reality to present process monitoring status and optimize decision-making solutions; The instruction issuing module is used to automatically issue control instructions to the production site based on the optimized decision-making plan, forming an automated closed loop of detection, decision-making, and execution; The local incremental module is used to automatically discover the root cause of a major anomaly and trigger a local incremental update to correct model drift when a major anomaly occurs. The adaptive migration module is used to utilize domain adaptation technology to migrate the optimization model of mature sites to newly commissioned sites for deployment.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for intelligent monitoring and optimization of nuclear power plant production business processes according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for intelligent monitoring and optimization of nuclear power plant production business processes according to any one of claims 1 to 6 are implemented.

Images