A nuclear power digital processing system

Abstract

The application discloses a nuclear power digital processing system, and relates to the technical field of nuclear power, which comprises an input layer, a cooperative control layer, a document understanding intelligent agent, a knowledge reasoning intelligent agent, an IO configuration intelligent agent, a drawing intelligent agent, a data interaction layer and an output layer; the document understanding intelligent agent is used for carrying out labeling analysis processing on a nuclear power digital design file to obtain a quality evaluation result; the knowledge reasoning intelligent agent is used for executing knowledge management and engineering rule reasoning verification processing to generate a reasoning result according to the quality evaluation result; the IO configuration intelligent agent is used for generating an IO list and carrying out cabinet machine distribution optimization processing according to the reasoning result to obtain an IO configuration result; and the drawing intelligent agent is used for executing power supply demand analysis, component selection and schematic design operation according to the IO configuration result to obtain a power supply cabinet drawing. The scheme can automatically output the power supply cabinet drawing, and realizes the collaborative improvement of efficiency, stability and design quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nuclear power control, and in particular to a nuclear power digital processing system. BACKGROUND

[0002] With the rapid development of nuclear power digital technology, the nuclear power digital processing system (DCS) as an important control platform for the development of nuclear power technology has been increasingly applied in nuclear power scenarios. The core of the system is to realize the automatic monitoring and control of various devices and process of nuclear power plants through computer technology, communication technology and software algorithm. In order to improve the safety of nuclear power operation and optimize the efficiency of nuclear power operation, it is particularly important to study how to intelligently design the DCS system.

[0003] At present, one way in the related art is to realize automatic signal distribution by using a nuclear power DCS cabinet and a card channel. However, this scheme is only designed for a single function module, and each link operates independently. When encountering rule conflicts, document format changes, configuration abnormalities and the like, manual intervention is required, resulting in low processing efficiency. Another way is to realize automatic drawing through a controller cabinet. However, this scheme uses a passive and fixed drawing template, which cannot handle complex layout optimization. SUMMARY

[0004] The purpose of the present application is to provide a nuclear power digital processing system.

[0005] To achieve the above purpose, the present application provides the following solutions.

[0006] In a first aspect, the present application provides a nuclear power digital processing system, comprising:

[0007] an input layer, a collaborative control layer, a plurality of intelligent agents, a data interaction layer and an output layer connected in sequence, wherein the plurality of intelligent agents comprise a document understanding intelligent agent, a knowledge reasoning intelligent agent, an IO configuration intelligent agent and a drawing intelligent agent connected in sequence; the drawing intelligent agent is connected with the output layer, and the document understanding intelligent agent is connected with the input layer;

[0008] The input layer is configured to: when receiving a document processing task, acquire a nuclear power digital design file;

[0009] The collaborative control layer is configured to: according to the nuclear power digital design file, perform collaborative control processing through the plurality of intelligent agents to obtain a power cabinet drawing; the collaborative control processing comprises at least one of the following: task scheduling, task negotiation, quality monitoring and exception handling;

[0010] The document understanding agent is configured to perform annotation analysis on the nuclear power digital design file to obtain a quality evaluation result; the knowledge reasoning agent is configured to perform knowledge management and engineering rule reasoning verification according to the quality evaluation result to generate a reasoning result; the IO configuration agent is configured to generate an IO list according to the reasoning result and perform cabinet allocation optimization processing to obtain an IO configuration result; and the drawing agent is configured to perform power supply demand analysis, component selection and schematic design according to the IO configuration result to obtain a power supply cabinet drawing.

[0011] The data interaction layer is configured to perform data interaction processing between the agents.

[0012] The output layer is configured to output the power supply cabinet drawing.

[0013] According to the embodiments provided in the present application, the following technical effects are disclosed:

[0014] The nuclear power digital processing system provided in the present application can obtain nuclear power digital design files through the input layer to provide accurate and complete initial data for the system and provide a reliable data basis for subsequent processing. The task scheduling, negotiation, quality monitoring and exception handling are realized through the collaborative control layer to break the limitation of independent operation of each link, actively cope with abnormal situations such as rule conflicts, reduce manual intervention, and improve the overall collaborative efficiency and stability of the system. The document understanding agent performs annotation analysis and quality evaluation to identify potential problems in the design file in advance, provides basis data with qualified quality for subsequent processing, and performs knowledge management and engineering rule reasoning verification through the knowledge reasoning agent to ensure that the terminology is used in a standard manner and in accordance with the engineering rules, avoid design errors caused by terminology ambiguity or rule inconsistency, and improve the professionalism and compliance of the design. The IO configuration agent generates an IO list and optimizes cabinet allocation to realize reasonable planning and efficient allocation of IO resources, improve the cabinet space utilization rate and the scientific nature of the configuration. The drawing agent completes power supply demand analysis, component selection and schematic design to accurately generate a power supply cabinet drawing based on the results of the previous processing, ensures that the drawing design meets the actual power supply demand, and improves the practicality and accuracy of the drawing. The data interaction layer realizes data interaction between the agents to realize real-time sharing and circulation of information, ensures that the processing of each link is based on the same basis, and improves the coherence and collaboration of the overall design process. The output layer outputs the power supply cabinet drawing in a standardized and clear form to present the design results. The intelligent optimization of each link is suitable for complex layout optimization, ensures the design quality and rationality of the power supply cabinet drawing, and realizes the collaborative improvement of efficiency, stability and design quality. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below only illustrate some of the embodiments of the present application, and for those skilled in the art, other drawings can be obtained based on these drawings without creative labor.

[0016] Figure 1 A structural schematic diagram of a nuclear power digital processing system in an embodiment of the present application is shown in FIG. 1.

[0017] Figure 2 A flowchart of a task negotiation decision mechanism provided in an embodiment of the present application is shown in FIG. 2.

[0018] Figure 3 A flowchart of a document analysis process of a document understanding agent provided in an embodiment of the present application is shown in FIG. 3.

[0019] Figure 4 A flowchart of a document standardization analysis process of a document understanding agent provided in another embodiment of the present application is shown in FIG. 4.

[0020] Figure 5 An architecture schematic diagram of an IO allocation optimization algorithm based on reinforcement learning provided in an embodiment of the present application is shown in FIG. 5.

[0021] Figure 6 A flowchart of a drawing drawing process performed by a drawing drawing agent provided in an embodiment of the present application is shown in FIG. 6.

[0022] Figure 7 A flowchart of a cooperation control process among various agents provided in an embodiment of the present application is shown in FIG. 7. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0025] In one way of the related technology, nuclear power DCS cabinet and card channel are adopted to realize automatic signal distribution, signal measurement points with distribution are acquired, relationship attributes of each signal measurement point are identified, and distribution rules are defined according to the relationship attributes; a distribution record table is constructed for each DCS cabinet, and the corresponding table position of the distribution record table is dynamically filled and deleted to realize distribution. However, this scheme is only designed for a single functional module and solves a single link of signal distribution, and when rule conflicts, document format changes, configuration abnormalities and the like occur, the system cannot autonomously coordinate; in another way, automatic drawing is realized through a control cabinet, including reading configuration file information of a target controller cabinet by a program file and parsing a mapping database, matching corresponding CAD blocks based on the configuration file information and the parsed mapping database; the program file opens a CAD template, and draws graphics in the CAD module according to the read configuration file information and the matched CAD blocks. However, this scheme focuses on the automation of the drawing link, and uses a fixed CAD drawing template, which cannot adaptively handle complex situations.

[0026] Based on the above defects, the present application provides a nuclear power digital processing system. Compared with the prior art, in the present system, the nuclear power digital design file is acquired through the input layer to provide accurate and complete initial data for the system and provide a reliable data basis for subsequent processing; task scheduling, negotiation, quality monitoring and abnormality handling are realized through the collaborative control layer to break the limitation of independent operation of each link, actively cope with abnormal situations such as rule conflicts, reduce manual intervention, and improve the overall collaborative efficiency and stability of the system; the document understanding agent performs labeling analysis and quality evaluation, identifies potential problems in the design file in advance, provides quality qualified basic data for subsequent processing, and performs knowledge management and engineering rule reasoning verification through the knowledge reasoning agent to ensure that the terminology is used in a standard manner and in accordance with the engineering rules, avoid design errors caused by terminology ambiguity or rule inconsistency, and improve the professionalism and compliance of the design; the IO configuration agent generates an IO list and optimizes cabinet allocation to realize reasonable planning and efficient allocation of IO resources, improve cabinet space utilization and the scientific nature of configuration; the drawing agent completes power supply demand analysis, component selection and schematic design, accurately generates power supply cabinet drawings in combination with the results of the previous processing, ensures that the drawing design fits the actual power supply demand, and improves the practicality and accuracy of the drawings; and data interaction between the agents is realized through the data interaction layer to realize real-time sharing and circulation of information, ensure consistency of processing in each link, and improve the coherence and collaboration of the overall design process; the output layer outputs the power supply cabinet drawings in a standardized and clear form to present the design results, the intelligent optimization of each link described above is suitable for complex layout optimization, ensures the design quality and rationality of the power supply cabinet drawings, and realizes collaborative improvement of efficiency, stability and design quality.

[0027] Please refer to Figure 1 as shown, Figure 1A structural schematic diagram of a nuclear power digital processing system provided by the embodiment of the application is shown in the figure, and the system comprises an input layer, a collaborative control layer, a plurality of agents, a data interaction layer and an output layer connected in sequence. The plurality of agents comprises a document understanding agent, a knowledge reasoning agent, an IO configuration agent and a drawing drawing agent connected in sequence; the drawing drawing agent is connected with the output layer, and the document understanding agent is connected with the input layer.

[0028] The input layer is configured to acquire a nuclear power digital design file when receiving a document processing task; the collaborative control layer is configured to perform collaborative control processing through the plurality of agents based on the nuclear power digital design file to obtain a power cabinet drawing; the collaborative control processing comprises at least one of task scheduling, task negotiation, quality monitoring and exception handling; the document understanding agent is configured to perform labeling analysis processing on the nuclear power digital design file to obtain a quality evaluation result; the knowledge reasoning agent is configured to perform knowledge management and engineering rule reasoning verification processing based on the quality evaluation result to generate a reasoning result; the IO configuration agent is configured to generate an IO list based on the reasoning result and perform cabinet machine allocation optimization processing to obtain an IO configuration result; the drawing drawing agent is configured to perform power supply demand analysis, component selection and schematic design operations based on the IO configuration result to obtain the power cabinet drawing; the data interaction layer is configured to perform data interaction processing between the agents; and the output layer is configured to output the power cabinet drawing.

[0029] It should be noted that the nuclear power digital processing system provided by the application refers to a nuclear power DCS system based on autonomous collaboration of a plurality of agents, which realizes autonomous processing of the whole process from technical document input to complete CAD drawing output, and comprises a multi-layer architecture, i.e., an input layer, a collaborative control layer, a data interaction layer and an output layer.

[0030] Specifically, the input layer is configured to acquire a nuclear power digital design file when receiving a document processing task. The document processing task can be a nuclear power DCS design file, and the nuclear power digital design file can be in different formats, such as PDF, Excel, Word and the like.

[0031] The collaborative control layer is an end-to-end autonomous collaborative control layer, which is the core control layer of the system and has four key functions, i.e., task scheduling, task negotiation, quality monitoring and exception handling. The task scheduling is configured to realize intelligent task distribution and flow control based on an event-driven model, the task negotiation is configured to realize real-time negotiation and decision optimization between the agents, the quality monitoring is configured to provide whole-process quality monitoring and preventive control, and the exception handling is configured to realize adaptive processing and automatic recovery of exceptions.

[0032] The system can further include an agent processing layer, which includes four agents, namely, a document understanding agent, a knowledge reasoning agent, an IO configuration agent, and a drawing agent. The document understanding agent is responsible for multi-format document parsing and nuclear power professional data extraction, adopts a multi-strategy adaptive parsing algorithm to realize intelligent identification of table structure and standardized processing of terms, has the ability of dynamic optimization of text recognition parameters and adaptive selection of parsing strategies, and performs term verification and quality optimization through cooperation with the knowledge agent. The knowledge reasoning agent is responsible for nuclear power DCS professional knowledge management and engineering rule reasoning verification, adopts a hierarchical rule management and forward chain reasoning algorithm to realize automatic detection and solution of rule conflicts, has the ability of reasoning confidence evaluation and dynamic updating of rule base, and provides real-time reasoning support and knowledge consistency coordination for other agents.

[0033] The IO configuration agent is responsible for intelligent generation and distribution optimization of IO lists, adopts a multi-rule fusion configuration algorithm and a multi-constraint IO distribution optimization algorithm to realize functions such as card selection, power supply calculation, and cabinet slot distribution, has the ability of automatic solution of constraint conflicts and dynamic balance adjustment of load, and realizes multi-objective trade-off decision through consultation with other agents. The drawing agent is responsible for generating unified power supply cabinet drawings, adopts intelligent template selection and standardized layout algorithm to realize CAD parameterized drawing and layout intelligent optimization, has the ability of automatic design of power supply scheme and automatic calculation of electrical parameters, can influence IO configuration decisions and coordinate design consistency in a reverse manner.

[0034] The data interaction layer includes a data interface layer and a data exchange layer. The data interface layer is a standardized structure for standard data exchange between agents, can realize unified data format, has API interface specifications, data verification mechanisms, and version compatibility management. For unified data format, a unified data format (including IO point data model, rule data model, configuration data model, drawing data model, etc.) can be defined based on JSON Schema to ensure consistent data structure. RESTful agent service interfaces are adopted to define unified request / response format, error code system, and version control mechanism, and to simplify the calling logic between agents. For data verification mechanism, format verification, business logic verification, and integrity verification are performed on input and output data to ensure the reliability of data exchange. For version compatibility management, backward compatible processing of data format upgrade is adopted to support independent upgrade and deployment of agents.

[0035] The data exchange layer is a communication infrastructure for collaboration between various agents, and is used for data interaction between various agents. The data interaction layer includes a message queue, and through an asynchronous message passing mechanism, the agents can complete data interaction and task assignment without real-time waiting for a response from the other party. For example, after the document understanding agent finishes parsing, it automatically pushes data to the knowledge reasoning agent through the queue, which not only avoids blocking the entire process due to a single agent being stuck, but also improves the flexibility of task distribution. State synchronization can also be achieved: the task state database records the running state and task progress of each agent in real time, ensuring that the collaboration control layer can clearly understand the entire process nodes and timely discover lagging links and allocate resources. The running state includes: "parsing", "verification passed", and "manual intervention" is required.

[0036] When there is a rule conflict (such as a conflict between the IO allocation scheme and the cabinet capacity), the standardized message format supports multiple agents (such as the IO configuration agent and the drawing agent) to quickly exchange opinions and propose alternative solutions, automatically reach a consensus through pre-set logic, and reduce human intervention. Quality feedback can also be performed, and quality information such as the rationality of IO configuration and the compliance of power supply design is transmitted bidirectionally through real-time channels. For example, if the drawing agent finds that the selection of a component does not match the IO parameters, it immediately feeds back to the IO configuration agent for adjustment, achieving dynamic optimization of the design process and avoiding rework of the final result.

[0037] The output layer is the final presentation of the system processing results, and its output content is closely related to the actual needs of nuclear power engineering, with both completeness and practicality. It can output an IO configuration list and a power supply cabinet drawing, including a power supply cabinet schematic diagram and a power supply cabinet wiring diagram. The IO configuration list includes a signal allocation table, a card configuration table, and a cabinet allocation table, providing accurate parameters for cabinet assembly and signal wiring. The signal allocation table can include the source, destination, and type of each signal, the card configuration table includes the card type, interface allocation, and parameter settings, and the cabinet allocation table includes the installation location and space occupation of each device in the cabinet. The schematic diagram clearly shows the electrical connection logic of the power supply circuit, and the wiring diagram labels the specific cable direction, terminal number, and connection relationship, which not only meets the standardization requirements of engineering design, but also provides an intuitive operation guide for construction personnel, ensuring that the on-site installation is consistent with the design scheme.

[0038] Compared with the prior art scheme that only realizes the single function of signal distribution, in the present application, four intelligent agents are constructed to sequentially perform end-to-end cooperation of document analysis, knowledge reasoning, IO configuration and drawing of drawings, so as to output the power cabinet drawings. Compared with the prior art scheme of a solidified CAD drawing template that cannot process complex layout problems, in the present scheme, the intelligent agents are cooperatively executed, the layout structure of the drawings is optimized, the corresponding power cabinet drawings are presented in a standard and clear manner, the scheme is suitable for complex layout optimization, the design quality and rationality of the power cabinet drawings are guaranteed, and the efficiency, stability and design quality are synergistically improved.

[0039] The above-mentioned cooperation control layer includes a plurality of core sub-mechanisms, respectively: task scheduling, task negotiation, quality monitoring and exception handling. For the task scheduling sub-mechanism, the system can adopt an event-driven model to realize automatic task scheduling between intelligent agents. After the upstream intelligent agent completes processing, the system performs multi-dimensional quality evaluation to obtain a comprehensive quality score, and the multi-dimensions include accuracy, integrity and confidence. When the comprehensive quality score exceeds the cooperation triggering threshold, the downstream intelligent agent is automatically triggered to process. The upstream intelligent agent or the downstream intelligent agent can include a document intelligent agent, a knowledge reasoning intelligent agent, an IO configuration intelligent agent and a drawing intelligent agent. In the process of task allocation, a load balancing strategy can be adopted to comprehensively consider the health state, current load, historical performance and other factors of the intelligent agent to select the most suitable intelligent agent to execute the task. The system maintains a global task state database to monitor the task progress and execution state in real time.

[0040] For the task negotiation sub-mechanism, when there is a rule conflict between intelligent agents, a four-stage negotiation mechanism can be started: the first stage is conflict detection: the system detects the rule applicability degree to score and determine the difference value, and when the difference value is less than the difference threshold, it is determined as a conflict; when the difference value is not less than the difference threshold, it is determined as no conflict. The second stage is negotiation invitation: the conflict initiator sends a negotiation request to the relevant intelligent agent and explains the conflict situation and the candidate scheme in the negotiation request. The third stage is professional scoring: each intelligent agent scores according to the dimensions of the professional field, for example, the knowledge reasoning intelligent agent focuses on rule compliance, the IO configuration intelligent agent focuses on engineering practicability, and the drawing intelligent agent focuses on layout rationality. The fourth stage is comprehensive decision: the system obtains the problem type and dynamically adjusts the weight according to the problem type, calculates the comprehensive score, and selects the optimal candidate scheme from all candidate schemes.

[0041] When the cooperation control processing is task negotiation, the system is also used for:

[0042] determine a negotiation request; the negotiation request is sent by the IO configuration agent to the knowledge reasoning agent and the drawing agent, and the negotiation request includes: conflict information and candidate processing strategies; obtain the rule complexity dimension score of the knowledge reasoning agent and the layout optimization dimension score of the drawing agent; determine the problem type according to the conflict information, and calculate the comprehensive score value based on the rule complexity dimension score, the layout optimization dimension score and the dynamic weight value; determine the target processing strategy from all candidate processing strategies based on the comprehensive score value and notify the IO configuration agent, the knowledge reasoning agent and the drawing agent.

[0043] Specifically, please refer to Figure 2 As shown, the conflict initiator can perform conflict detection, obtain a difference value through rule applicability score difference analysis, determine a conflict when the difference value is less than a pre-difference threshold, then perform conflict negotiation processing, send a negotiation invitation to related agents by the conflict initiator, and explain the conflict situation and candidate solutions, each agent scores according to its own professional field, the knowledge reasoning agent focuses on rule compliance, the IO configuration agent focuses on engineering practicality, and the drawing agent focuses on layout rationality, and determines the problem type according to the conflict problem, calculates the comprehensive score value based on the respective dimension scores, determines the optimal processing scheme according to the comprehensive score value and notifies all participants, and dynamically adjusts the weight value through weighted decision-making, and generates a decision record for result confirmation.

[0044] Taking the conflict initiator as the IO configuration agent and the negotiation request receiver as the knowledge reasoning agent and the drawing agent as an example, the IO configuration agent generates a collaboration request and sends it to the knowledge reasoning agent and the drawing agent, the knowledge reasoning agent generates a rule complexity dimension score in response to the collaboration request, the drawing agent generates a layout optimization dimension score in response to the collaboration request, and determines the problem type, calculates the comprehensive score value based on the rule complexity dimension score, the layout optimization dimension score and the dynamic weight value; determines the target processing strategy from all candidate processing strategies based on the comprehensive score value and notifies the IO configuration agent, the knowledge reasoning agent and the drawing agent.

[0045] When the collaboration control processing is abnormal processing, the system is further used for:

[0046] detecting whether an exception occurs in each agent interaction, agent internal and system level, determining an exception type; the exception type includes: a first-level exception, a second-level exception, a third-level exception and a fourth-level exception; the first-level exception, the second-level exception, the third-level exception and the fourth-level exception are arranged in order from small to large according to the degree of exception; execute the corresponding exception processing strategy according to the exception type, and obtain an exception processing result.

[0047] For the abnormal processing sub-mechanism, the system can establish a four-level exception classification system: first-level exception, second-level exception, third-level exception and fourth-level exception. The first-level exception is a slight exception, and the slight exception adopts parameter automatic adjustment; the second-level exception is a medium exception, and the medium exception starts agent collaborative processing; the third-level exception is a serious exception, and the serious exception triggers system-level intervention; and the fourth-level exception is a fatal exception, and the fatal exception needs manual processing.

[0048] Specifically, the abnormal processing can adopt a hierarchical strategy, first performing autonomous adjustment within the agent, such as performing OCR parameter optimization, rule threshold adjustment, etc. When it cannot be solved autonomously within the agent, a collaborative recovery mechanism can be started to provide assistance from other agents. The above system can also establish an abnormal prevention mechanism to identify abnormal precursor indicators through historical data analysis. In this embodiment, through four-level exception classification and multi-level processing strategy, the system can autonomously cope with various abnormal situations such as document format changes, rule conflicts, device updates, etc., significantly improving the stability and continuity of the system.

[0049] In this embodiment, through the end-to-end autonomous collaborative control mechanism, an event-driven task scheduling model is adopted to establish an abnormal automatic detection and collaborative repair mechanism, to realize autonomous operation of the system. Through a multi-agent real-time negotiation decision mechanism, a standardized four-stage negotiation decision process is established to solve rule conflict problems. Through a dynamic weight adjustment mechanism and an asynchronous communication protocol, consistency of the solution is ensured when the system faces complex decision-making problems.

[0050] The application provides a nuclear power digital processing system. Compared with the prior art, in the scheme, a nuclear power digital design file is acquired through an input layer to provide accurate and complete initial data for the system and provide a reliable data basis for subsequent processing; task scheduling, negotiation, quality monitoring and abnormal processing are realized through a collaborative control layer to break the limitation of independent operation of each link, actively cope with abnormal situations such as rule conflicts, reduce manual intervention, and improve the overall collaborative efficiency and stability of the system; a document understanding agent performs labeling analysis and quality evaluation, identifies potential problems in the design file in advance, provides quality qualified basic data for subsequent processing, and performs knowledge management and engineering rule reasoning verification through a knowledge reasoning agent to ensure that the terminology is used in a standard and engineering rule-compliant manner, avoid design errors caused by terminology ambiguity or rule inconsistency, and improve the professionalism and compliance of the design; an IO configuration agent generates an IO list and optimizes cabinet machine allocation to realize reasonable planning and efficient allocation of IO resources, improve the cabinet machine space utilization rate and the scientific nature of the allocation; a drawing generation agent completes power supply demand analysis, component selection and schematic diagram design, accurately generates power supply cabinet drawings in combination with the previous processing results, ensures that the drawing design is in line with the actual power supply demand, and improves the practicability and accuracy of the drawing; and data interaction between the agents is realized through a data interaction layer to realize real-time sharing and circulation of information, ensure that the processing basis of each link is consistent, and improve the coherence and collaboration of the overall design process; the output layer outputs the power supply cabinet drawings in a standardized and clear form to present the design results, the intelligent optimization of the above-mentioned links is suitable for complex layout optimization, ensures the design quality and rationality of the power supply cabinet drawings, and realizes the collaborative improvement of efficiency, stability and design quality.

[0051] In one embodiment, referring to FIG. Figure 3 The document understanding agent can perform the following steps:

[0052] S201, identifying and processing the nuclear power digital design file to obtain a feature recognition result; the feature recognition result includes: a file name mode, a content feature, and a page layout.

[0053] S202, selecting a target analysis strategy using a decision tree model according to the feature recognition result; the target analysis strategy includes: a table analysis strategy, a robust analysis strategy and a mixed analysis strategy.

[0054] S203, when the target analysis strategy is the table analysis strategy or the robust analysis strategy, performing table boundary detection and cell segmentation processing to obtain a quality evaluation result.

[0055] S204, when the target analysis strategy is the mixed analysis strategy, performing terminology standardization processing to obtain a quality evaluation result.

[0056] It should be noted that the document understanding agent described above is used to process documents of various formats, can automatically identify table structures, and accurately extract IO data. The core innovation of this document understanding agent lies in its multi-strategy adaptive parsing and collaborative quality improvement mechanism.

[0057] Specifically, please see Figure 4 As shown, after obtaining the nuclear power digital design document through the input layer, a three-dimensional feature analysis is performed on the input nuclear power digital design document, including filename pattern recognition, content feature analysis, and page layout recognition. A pattern library and terminology density analysis of the nuclear power digital design document are also performed to obtain feature recognition results. Then, based on the feature recognition results, a decision tree model is used to select the optimal target parsing strategy from all parsing strategies. These parsing strategies can include: table parsing strategy, robust parsing strategy, and hybrid parsing strategy. When the feature recognition results indicate that the nuclear power digital design document is a high-quality standard document, the table parsing strategy is selected as the target parsing strategy; when the feature recognition results indicate that the nuclear power digital design document is a low-quality scanned document, the robust parsing strategy is selected as the target parsing strategy; when the feature recognition results include both low-quality scanned documents and high-quality standard documents, a hybrid parsing strategy is selected as the target parsing strategy. The hybrid parsing strategy can be a region classification parsing strategy.

[0058] When the target parsing strategy is a table parsing strategy and a robust parsing strategy, a multi-algorithm parallel strategy is adopted, including Hough transform, morphological operations, and connected component analysis, to perform table boundary detection and cell segmentation processing, obtaining quality assessment results. Optionally, a three-engine voting decision using Tesseract, PaddleOCR, and EasyOCR can be used to improve recognition accuracy. When the target parsing strategy is a hybrid parsing strategy, a hierarchical vocabulary of 5000+ nuclear power terminology can be used to perform terminology standardization processing using edit distance and context association, implementing abbreviation expansion and a dual verification mechanism to obtain quality assessment results.

[0059] When the term recognition credibility is less than the collaboration trigger threshold, a verification request is automatically sent, and term authority verification and revision suggestions are performed within a preset time period, so as to continue to generate a quality evaluation result through a four-dimensional quality evaluation system, and trigger task distribution of downstream intelligent agents according to the quality evaluation result. The four-dimensional quality evaluation system includes accuracy, integrity, consistency and credibility, and the weight values corresponding to each dimension are different. For example, the accuracy accounts for 35% of the corresponding weight value of the entire quality evaluation system, the integrity accounts for 25% of the corresponding weight value of the entire quality evaluation system, the consistency accounts for 25% of the corresponding weight value of the entire quality evaluation system, and the credibility accounts for 15% of the corresponding weight value of the entire quality evaluation system. The collaboration trigger threshold and the preset time period can be customized according to actual needs. For example, the collaboration trigger threshold can be 0.8, and the preset time period can be 30 seconds.

[0060] The document understanding intelligent agent in this embodiment can accurately extract key information (such as device parameters, signal types, connection relationships, etc.) in the file by performing labeling and analysis processing on the nuclear power digital design file and generating a quality evaluation result, and can ensure the consistency and integrity of information extraction through standardized analysis to avoid omissions or biases that may occur in manual interpretation. The quality evaluation result generated thereby can filter out unqualified or flawed design files in advance, providing a reliable data basis for subsequent intelligent agent processing and reducing repeated work or errors caused by input data problems in downstream links, thereby improving the processing efficiency and design quality of the entire system. Through more than 100 times of collaborative learning, the term recognition accuracy is improved from 82% to 94%, and the collaboration request frequency is reduced from 15% to 3%, significantly improving the autonomous processing capability.

[0061] In one of the embodiments, the knowledge reasoning intelligent agent is specifically used for:

[0062] When the quality evaluation result is greater than the collaboration trigger threshold, the term information and the fact condition are obtained based on the nuclear power digital design file; the term information includes a term list and term context information; the term information is evaluated to obtain a term verification result and is sent to the document understanding intelligent agent; and / or, the fact condition is reasoned to obtain a reasoning result and is sent to the IO configuration intelligent agent; and / or, when multiple intelligent agents conflict, a conflict type is determined and a preset rule knowledge base is used for arbitration based on the conflict type to obtain a coordination result; the preset rule knowledge base includes a domain division rule, a priority rule and a use range rule.

[0063] Specifically, a preset rule knowledge base is first constructed, which includes domain division rules, priority rules and usage range rules. The domain division rules include division by professional field, and the professional field includes signal classification, card selection, power supply calculation, terminal plate matching and safety level. The priority rules include: mandatory > standard > best practice > project preference. The usage range rules include general range, technology-specific range and project-specific range. The preset rule knowledge base is used for ordered storage and accurate positioning of various rules, ensuring that the most suitable rule can be quickly called in different scenarios.

[0064] After the preset rule knowledge base is constructed, an intelligent rule matching operation is performed. Two-stage measurement can be used. First, the rule matching range is quickly narrowed through index table set intersection operation, and then the applicable rule is locked by accurately calculating the matching rate score (matching rate ≥ 80%). The matching rate score of each rule can be obtained by matching rate score = 0.5 × condition matching degree + 0.3 × priority weight + 0.2 × historical success rate, and the rules are sorted to determine the most suitable rule, avoiding rule abuse and ensuring that the recommended rule meets both technical requirements and project actual effects.

[0065] Optionally, the knowledge reasoning agent can also use a breadth-first reasoning strategy from cause to effect. The reasoning depth of this reasoning strategy can be within five layers. The dynamic fact base records the complete reasoning chain, which ensures reasoning efficiency and avoids bias caused by excessive reasoning. At the same time, it supports "reverse verification". For a specific target, a probabilistic conclusion can be obtained, for example, when the specific target is "whether a certain card is applicable", the probabilistic result can be "confidence 90%", thereby providing flexible support for decision-making.

[0066] As an implementation manner, when the quality evaluation result is greater than the collaboration triggering threshold, the term information and fact conditions are obtained based on the nuclear power digital design file, and the term information is evaluated, including evaluating the term standardization, such as unifying temperature signals into standard expressions, and evaluating the matching degree of the term with the context, obtaining the term verification result and transferring it to the document understanding agent, to ensure the consistency of the basic data.

[0067] As another implementation manner, the obtained fact conditions include device parameters and signal types. Based on the obtained fact conditions, a reasoning result is obtained, thereby providing rule reasoning for the IO configuration agent. The reasoning result may, for example, include reasoning information, confidence and alternative solutions. The reasoning information may be which card should be configured for a certain signal, thereby providing configuration optimization reference.

[0068] As a further possible implementation, when multiple intelligent agents conflict, a conflict type is determined, which can include direct conflict, conditional conflict, and priority contradiction, etc., and based on the conflict type, a decision is made based on preset domain division rules, priority rules and usage range rules, thereby obtaining a coordination result. Wherein, the direct conflict is, for example, opposite rule conclusions, and the condition overlap is, for example, the intersection of rule application ranges.

[0069] In this embodiment, the knowledge reasoning intelligent agent can realize nuclear power digital professional knowledge management and engineering rule reasoning verification, and provide decision support for other intelligent agents.

[0070] In one of the embodiments, the IO list includes: card type, power supply mode and terminal board capacity; and the IO configuration intelligent agent is specifically used for:

[0071] Based on the reasoning result and the feature recognition result, multi-dimensional features, power type coding information and sensor power supply demand information are determined; the multi-dimensional features include: description semantics, electrical characteristics and engineering units; the description semantics is analyzed to identify signal functions, the electrical characteristics are analyzed to determine signal types, and a mapping relationship between engineering units and signal types is established; the card type is determined according to the signal function, the signal type, the electrical characteristics, the mapping relationship and the communication protocol; the power supply mode is determined based on the power type coding information and the sensor power supply demand information; and the card type, the power supply mode and the electrical characteristics are combined and matched, a matching mapping table is established, and the terminal board capacity is determined according to the matching mapping table.

[0072] Specifically, the IO configuration intelligent agent can obtain the feature recognition result and the reasoning result, and adopt a multi-feature fusion strategy to comprehensively analyze three dimensions of description semantics, electrical characteristics and engineering gear. Among them, the description semantics analysis is to identify signal functions through a keyword library, the electrical characteristics are to determine signal types according to key parameters, and the engineering unit analysis is to establish a mapping relationship between units and signal types, and the key parameters can include, for example: "4-20mA", "on-off signal" and the like. The weights of the features corresponding to each dimension can be different, and a weighted voting mechanism can be adopted: the weight of the electrical characteristics is 40%, the weight of the description semantics is 35%, and the weight of the engineering unit is 25%.

[0073] After multi-dimensional feature analysis, card type intelligent selection processing can be performed, which can be based on conditions such as signal function, signal type, electrical characteristics, mapping relationship and communication protocol to perform layer-by-layer screening to determine the card type, wherein different signals or scenes correspond to different card types, AI signals can select AI216, AI232, AI221 and other card types according to electrical characteristics, DI signals select DI211 card, and communication signals select special card according to protocol. Special scene processing includes factors such as redundancy configuration, isolation requirement, explosion-proof grade, etc.

[0074] After determining the card type, the power supply mode is calculated, the power type code information and sensor power supply requirement information are obtained, intelligent matching is performed based on the power type code information and sensor power supply requirement information, the sensor type can be determined first, and the power supply mode is determined according to the sensor type. For two-wire sensors, 24VDC loop power is provided by DCS, and four-wire sensors need additional power supply. Among them, the power supply mode of different signal types is determined according to the characteristics and power requirements of the equipment. After determining the power supply mode, based on the combination matching of card type, power supply mode and electrical characteristics, a complete matching mapping table is established, and the terminal board capacity is determined according to the matching mapping table, and the terminal board capacity is automatically verified by the system to ensure that the electrical requirements are met.

[0075] In one of the embodiments, the IO intelligent agent is also used for:

[0076] Based on the IO list, multiple state space information is extracted and feature fusion is performed to obtain comprehensive feature information; the state space information includes: cabinet machine state information, IO point state information and constraint state information; the cabinet machine state information includes: load rate, available slot, safety level, the IO point state information includes: type code, power level, priority; the constraint state information is used to reflect the load balancing degree and constraint situation; the comprehensive feature information is passed through an initial decision network to obtain a decision result; according to the weight value corresponding to the preset index of the decision result, a reward value is calculated, and the initial decision network is updated based on the reward value to construct a decision network; the preset index includes: load balancing, maintenance convenience, cost optimization, safety compliance; the decision network is used for cabinet allocation processing to obtain an IO configuration result.

[0077] Specifically, after generating the IO list, the IO allocation problem can be modeled as a multi-constraint optimization problem, and multiple state space information is extracted based on the IO list, including cabinet state information, IO point state information, and constraint state information. The state space information is the environment information that the agent can perceive when making decisions, such as including the remaining space of the cabinet, the type and demand limit of the IO point, etc., which can provide a basis for decision-making. And design and establish functions, comprehensively consider four goals of load balancing, maintenance convenience, cost optimization, safety compliance, and the corresponding weight values are 30%, 20%, 20%, and 30% respectively. Load balancing is calculated by standard deviation reciprocal, maintenance convenience considers signal aggregation degree, cost optimization includes cabinet use cost and standardized cost, and safety compliance focuses on 1E level signal isolation and redundant signal separation requirements.

[0078] An initial decision network is acquired, which adopts a hierarchical feature extraction and fusion design, processes cabinet state information, IO point state information and constraint state information respectively, and fuses the three kinds of state information through a feature fusion layer to obtain comprehensive feature information. Then the comprehensive feature information is processed through the initial decision network, which can include three hidden layers and use a ReLU activation function, adopts Dropout to prevent overfitting, and outputs actions and corresponding reward values. For example, in the state of "cabinet has empty + IO is 1E level signal", the model can learn that the action of "preferentially allocating to an independent cabinet" is better.

[0079] Among them, the historical experience (state→action→reward) of the agent can be stored, randomly sampled for training, and the experience data that can bring large rewards or correct large errors is given a higher weight, so that the model learns "key experience" preferentially, and the convergence process is accelerated. For example, first let the model learn the basic allocation logic in a simple scenario (such as many cabinets and few constraints), then gradually increase the complexity of the scenario (many constraints and various IO types), and let the model learn more smoothly. The Adam optimizer is used for parameter updating to make the training more stable and ensure that the model converges to a usable allocation strategy, thereby constructing a decision network, and performing cabinet machine allocation processing through the trained decision network to obtain an IO configuration result.

[0080] Please refer to Figure 5 As shown in the figure, a reinforcement learning environment modeling is constructed, and after the IO list is acquired, the IO list is input into the reinforcement learning environment, which includes a state space, an action space and a reward function. The state space includes cabinet state, IO point state, and constraint state information; the action space defines the operations executable by the agent, which is represented by a four-tuple (cabinet, cage, slot, channel), i.e., the allocation and adjustment of cabinets, cages, slots and channels, which is a specific behavior to change the state of the environment. The reward function sets a multi-objective optimization logic, dynamically adjusts the reward value by weighted summation, and encourages the agent to make better allocation decisions, such as higher resource utilization and more constraint-satisfying allocation behaviors to obtain higher rewards.

[0081] The DeepQ-Network network architecture is a decision-making network of the agent, which is realized based on deep reinforcement learning: the input layer receives state space information, and after network operation, the output layer outputs action space selection and corresponding reward prediction. Through continuous iteration training, the agent learns to choose which actions (such as how to allocate cabinet resources) can obtain higher rewards in different states, optimizes the allocation strategy, and realizes "learning optimal decisions from experience". The allocation result is post-processed and optimized to ensure that the scheme is more in line with actual engineering needs. First, load rebalancing is performed, and through threshold rules such as standard deviation, the load of IO on equipment such as cabinets is checked and adjusted to avoid local overload or underload and improve resource utilization balance. Related signal aggregation is performed: analyze signals according to functional association, optimize allocation, and make closely related signals more collaborative in physical layout or logical configuration to facilitate management and operation. Maintenance path optimization is performed by calculating device maintenance paths and layouts to make subsequent operation more convenient and reduce operation cost and difficulty. The state space is comprehensively checked and processed: the allocation result is comprehensively evaluated for quality, potential problems (such as whether constraints are violated) are identified, and an evaluation report is output to provide a reference for final scheme determination, ensuring that the allocation scheme is reliable and practical, thereby outputting the optimal allocation scheme, i.e., the IO configuration result.

[0082] In this embodiment, the IO configuration agent can efficiently generate a reasonable IO list and optimize cabinet resource allocation by means of standardized data interaction, accurate rule reasoning, and dynamic optimization capabilities, ensuring that the IO configuration meets engineering rules and actual needs, providing a reliable foundation for power supply cabinet design, improving the scientificity, compliance, and efficiency of IO resource configuration in the overall design process, reducing human configuration errors and resource waste, and helping to output high-quality power supply cabinet design results. Through deep reinforcement learning global optimization, the IO allocation problem is modeled as a multi-constraint optimization problem, global optimization is realized through the Deep Q-Network algorithm, a multi-objective reward function is designed, and optimal allocation decisions are made under multiple constraint conditions. The constraint satisfaction rate is more than 90%, the load balancing standard deviation is reduced by 60%, the rule conflict self-solution rate is improved to more than 70%, and the power supply cabinet design automation rate is more than 85%, which significantly improves the design quality.

[0083] In one of the embodiments, the drawing drawing agent is specifically used for:

[0084] According to the IO configuration result, the power supply demand information is counted; the power supply demand information includes: the working current and power consumption information of various cards.

[0085] Based on the power supply demand information, the power supply capacity is calculated according to the safety margin rule.

[0086] Determine the element spacing and partition layout of each element in the power supply cabinet, and the cross-sectional area and bus connection path of the bus; the partition layout includes: functional partition layout, vertical layout and horizontal layout; the bus includes: main bus and distribution bus.

[0087] Obtain the power supply configuration type, and select the power supply cabinet drawing template library according to the power supply configuration type; the power supply configuration type includes: single power supply, double power supply and USB power supply.

[0088] Based on the power supply cabinet drawing template library, retrieve each component symbol and draw it through the connection line, generate a drawing and perform quality detection processing; the component includes at least one of the following: circuit breaker, contactor, relay, indicator light.

[0089] When the quality detection passes, generate the power supply cabinet drawing according to the drawing output type; the power supply cabinet drawing includes: power supply cabinet principle diagram and power supply cabinet wiring diagram.

[0090] Specifically, please refer to Figure 6 As shown in the drawing, after obtaining the IO configuration result, the drawing intelligent agent is used to perform load demand statistics, automatically analyze the IO configuration result, and count various power supply demand information, such as DCS internal power supply demand including the working current and power consumption of various card components, 24VDC external power supply demand including two-wire transmitter, solenoid valve, indicator light and other device loads, and 220VAC power supply demand including actuator, heater, high-power device and other loads. Then, based on the power supply demand information, the load classification algorithm is used to classify and summarize according to the voltage level and power supply mode. Among them, the DCS internal +24V power supply power can be determined by the following formula:

[0091] P_DCS =Σ(card component number_i×single card power consumption_i);

[0092] Where i represents different card component types, and P_DCS is the power supply power.

[0093] The DCS external 24VDC power supply power calculation includes transmitter loop power consumption and auxiliary device power consumption. The 220VAC power supply power needs to consider the difference between the starting current and the working current of the device.

[0094] After obtaining the power supply demand information, the power supply capacity is calculated according to the 125% safety margin rule, that is, the main power supply capacity is selected according to 125% of the total load, and the standby power supply is configured according to 150% of the key load. The selection of the circuit breaker capacity considers factors such as load current, starting current, and environmental temperature, and selects the minimum standard specification greater than the calculated value. The redundant power supply configuration is determined according to the importance level of the load. The 1E level key load adopts a triple redundant configuration of dual power supply + UPS backup, the non-1E level important load adopts a dual power supply configuration, and the general load adopts a single power supply configuration.

[0095] The above drawing intelligent agent is also configured with a template matching algorithm, which follows the selective principle that the action time of the upper circuit breaker is greater than that of the lower circuit breaker. The main circuit breaker adopts a time delay action characteristic, and the branch circuit breaker adopts a fast action characteristic. The short-circuit current calculation considers factors such as transformer impedance, line impedance, and contact resistance.

[0096] Optionally, the drawing intelligent agent is also configured with a power cabinet layout algorithm for realizing the reasonable arrangement and connection design of internal components of the power cabinet by adopting a rule-based layout strategy. The component spacing and partition layout mode of each component in the power cabinet are determined, and the power cabinet can adopt a functional partition layout, which is arranged from top to bottom as a main circuit breaker area, a distribution circuit breaker area, a control component area, and a wiring terminal area. The main circuit breaker area is located at the upper part of the cabinet body, which is convenient for operation and maintenance. The distribution circuit breaker area is arranged according to the grouping of loop functions, and related loops are placed adjacently. The vertical layout considers the heat dissipation requirement, and high-power components are arranged in a position with good ventilation. The horizontal layout reserves a maintenance channel to ensure that all components have sufficient operation space. The component spacing is determined according to the electrical safety distance and heat dissipation requirement. The busbar design adopts a layered structure, the main busbar carries the total current, and the distribution busbar connects each branch circuit. The busbar cross-sectional area is selected according to the current-carrying capacity and voltage drop requirement, and the connection mode adopts bolt connection or plug-in connection. The control circuit is arranged separately from the main loop to avoid electromagnetic interference. The control circuit adopts a plurality of soft wires, and the main loop adopts a single hard wire or copper bar. The line identification is color-coded and text-labeled according to the national standard.

[0097] The drawing intelligent agent can pre-establish a power cabinet drawing template library, which includes standard templates of different configurations such as single power supply, dual power supply, and UPS power supply. Each template defines basic elements such as drawing frame, title bar, component layout, and connection relationship. The template selection is based on the complexity of the power supply configuration and the redundancy requirement. For simple configuration selection, a single power supply template can be selected, for complex configuration selection, a multi-power supply template can be selected, and for high reliability requirement, a redundant power supply template can be selected.

[0098] After selecting the corresponding template from the power supply cabinet drawing template library, the component symbol can be called from the standard symbol library of the corresponding template, which includes common electrical elements such as circuit breakers, contactors, relays, and indicator lights. Among them, the standard symbol library is established according to the GB / T symbol standard to ensure the standardization of the drawing. The connection line drawing adopts a right-angle connection method to avoid diagonal line connection. The connection line color can be distinguished according to the voltage level, with red for 380V, brown for 220V, blue for 24V, and green for control lines. Thus, the connection line is drawn and processed to generate the drawing paper. The wire number marking adopts an automatic numbering mechanism to ensure the uniqueness and traceability of the numbering.

[0099] After the drawing is completed, the drawing paper is automatically subjected to quality inspection to obtain a quality inspection result. The quality inspection can include connection integrity inspection, marking correctness inspection, and specification compliance inspection. The connection integrity inspection is used to ensure that all elements have correct power and control connections. The marking correctness inspection is used to verify the accuracy of element numbering, wire number, and technical parameter marking information. When the quality inspection passes, the power supply cabinet drawing is generated according to the drawing output type; when the quality inspection fails, the drawing processing is continued until the quality inspection passes, and the power supply cabinet drawing is output. The drawing output type can include power supply cabinet schematic diagram and power supply cabinet wiring diagram. The power supply cabinet schematic diagram is used to display the electrical principle and control logic, and the power supply cabinet wiring diagram is used to display the actual connection and construction information. The generated power supply cabinet drawing format can support multiple standard formats such as DWG and PDF.

[0100] In this embodiment, the drawing intelligent agent can realize CAD parameterized drawing and layout intelligent optimization using intelligent template selection and standardized layout algorithm, ensuring the design quality and rationality of the power supply cabinet drawing, and achieving the coordinated improvement of efficiency, stability, and design quality.

[0101] In one embodiment, the data interface layer uses a unified data model for standardized data description to define the standard format for communication between intelligent agents. The data model can use a hierarchical structure, and the message includes a message header and a message payload. The message header includes metadata fields such as message identification, message type, sender, receiver, timestamp, priority, timeout, and association identification.

[0102] It should be noted that the message header can include message identification, message type, sender, receiver, timestamp, priority, timeout, and association identification. The message identification can use a unique coding format of "MSG_year-month-day_hour-minute-second_sequence number". The message type distinguishes between request, response, notification, and collaboration. The priority is set to low, medium, high, and urgent. The timeout range is set to 30 seconds to 3600 seconds, which is dynamically adjusted according to the task complexity.

[0103] The message payload part is defined with different data structures according to different service types. The service types include document processing, term verification, configuration generation, and drawing generation. The document processing payload contains fields such as document path, processing options, and extraction results. The term verification payload contains fields such as term list, context information, and verification results. The configuration generation payload contains fields such as IO point data, configuration scheme, and optimization indicators. The drawing generation payload contains fields such as design requirements, drawing file, and quality assessment.

[0104] In this embodiment, a service interface architecture based on the Hypertext Transfer Protocol can also be used to establish a standardized service calling mechanism between agents, including basic service interfaces and special service interfaces. Each agent can serve as an independent service node, providing standardized service interfaces for other agents to call.

[0105] The basic service interface described above can include different functional interfaces, including three types of core functions: health status check, capability query, and business processing. The health status check interface is used to return information such as the running state, resource usage, and service availability of the agent. The capability query interface is used to return capability descriptions such as the service types supported by the agent, processing capacity, and interface version. The business processing interface is used to receive specific business requests, execute corresponding processing logic, and return processing results. The special service interface described above is designed according to the professional functions of the agent. Different agents have different special service interfaces. The document understanding agent provides special interfaces for document processing, data verification, and retry processing. The knowledge reasoning agent provides special interfaces for term verification, rule reasoning, and conflict resolution. The IO configuration agent provides special interfaces for configuration generation, scheme optimization, and load analysis. The drawing drawing agent provides special interfaces for scheme design, drawing generation, and quality check.

[0106] In this embodiment, a message queue implementation mechanism is used to achieve reliable message transmission between agents using an asynchronous message queue architecture based on a lightweight message transmission protocol. The message queue system includes four core components: message publisher, message subscriber, message broker, and message storage. The message broker is responsible for message routing and forwarding and queue management, supporting two message transmission modes: topic subscription mode and point-to-point mode. The topic subscription mode is suitable for one-to-many broadcast communication, such as state notification and event distribution. The point-to-point mode is suitable for one-to-one direct communication, such as service calling and data transmission.

[0107] The message queue supports a quality of service guarantee mechanism, including at most once, at least once, and exactly once delivery guarantee levels. At most once is suitable for scenarios that are insensitive to message loss, such as state monitoring. At least once is suitable for scenarios that are sensitive to message loss, such as task distribution. Exactly once is suitable for scenarios that are sensitive to message repetition, such as configuration updates. The message persistence mechanism stores important messages in persistent storage, supporting reliable recovery and replay of messages. The message storage supports multi-dimensional queries and management by time, by topic, and by priority. The dead letter queue mechanism handles messages that cannot be normally delivered, supporting manual intervention and automatic retry.

[0108] The above data exchange layer can also perform data verification and error handling, and through the establishment of a three-layer data verification system, the reliability and consistency of data exchange between intelligent agents are ensured. The first layer of data verification is format verification, which is used to verify the integrity of the data structure and the correctness of the field type. The second layer of data verification is business verification, which is used to verify the logical reasonableness of the data content and the compliance of the business rules. The third layer of data verification is integrity verification, which is used to verify the integrity and consistency constraints of the data. Among them, the format verification adopts a structured data mode verification mechanism to define the type, format, value range and other constraint conditions of each data field. The string field supports length limit, format pattern, enumeration value and other constraints. The numerical value field supports minimum value, maximum value, precision, unit and other constraints. The array field supports element type, length limit, uniqueness and other constraints.

[0109] Business verification adopts a rule engine mechanism to define various business rules and verification logic, and performs data verification through business rules. Among them, the business rules include data correlation check, logical consistency verification, professional knowledge compliance verification and other verification items. The rule engine supports dynamic loading and updating of rules to adapt to changes in business requirements. The error handling mechanism needs to establish a unified error classification and processing strategy. Error types include timeout error, verification error, processing error, and communication error. Among them, the timeout error adopts a retry mechanism, sets the number of retries and retry intervals according to the severity of the error; the verification error returns detailed error information and correction suggestions, and does not recommend retry; the processing error adopts parameter adjustment, degradation processing, manual intervention and other strategies according to the error reason; the communication error adopts recovery mechanisms such as backup channel, message retransmission, and service switching.

[0110] In one embodiment, the system further includes a global state machine and a task scheduling module, and the global state machine includes a task state database for recording the current execution state, execution agent, completion progress and quality score of each task.

[0111] The task scheduling module is configured to create a global task record and assign a task identification and configure task information for the document processing task when receiving the document processing task; the task information includes a task priority and an expected completion time.

[0112] In this embodiment, an event-driven intelligent agent interaction timing control mechanism is established, and a global state machine is used to manage the cooperation process among the intelligent agents. A task state database is maintained by the system, which is used to record the current execution state, execution intelligent agent, completion progress and quality score of each task, and other key information.

[0113] Specifically, in the task starting phase, after the system receives the document processing task, the task scheduling module first creates a global task record, assigns a unique task identification to the document processing task, sets the task priority and the expected completion time. Then, a task processing request is sent to the document understanding intelligent agent, the current execution state in the task state database is updated to “document parsing”, and a 30-second period state monitoring mechanism is started. After the document understanding intelligent agent completes the processing, a quality evaluation result is obtained, and the system determines the subsequent process according to the output quality evaluation result. When the quality evaluation result is greater than the cooperation trigger threshold, the term verification process of the knowledge reasoning intelligent agent is automatically triggered; when the quality evaluation result is between the abnormal trigger threshold and the cooperation trigger threshold, the retry mechanism is started, and the OCR parameters are adjusted for reprocessing; when the quality evaluation result is less than the abnormal trigger threshold, the task is marked as abnormal and manual intervention is requested. The cooperation trigger threshold and the abnormal trigger threshold can be customized according to actual needs, the cooperation trigger threshold is greater than the abnormal trigger threshold, the cooperation trigger threshold can be 0.85, and the abnormal trigger threshold can be 0.7.

[0114] During the term verification process of the knowledge reasoning intelligent agent, multiple reasoning engines can be called in parallel to process complex reasoning problems. The system coordinates multiple parallel tasks through the cooperation request manager to ensure the consistency and timeliness of the results. After all the parallel tasks are completed, the reasoning results are summarized and transmitted to the IO configuration intelligent agent. The IO configuration intelligent agent generates an IO configuration result according to the reasoning result and sends it to the drawing preparation intelligent agent. The drawing preparation intelligent agent may find that the configuration is unreasonable during the layout design process. At this time, the system can start a feedback loop mechanism, the drawing preparation intelligent agent sends optimization suggestions to the IO configuration intelligent agent, the IO configuration intelligent agent adjusts the configuration scheme according to the optimization suggestions, updates the IO configuration result, and sends it to the drawing preparation intelligent agent, forming an iterative optimization loop, until the generated drawing quality detection passes (reaches the preset quality threshold) or reaches the maximum number of iterations, so as to generate a power cabinet drawing according to the drawing output type.

[0115] Optionally, after each agent completes the corresponding processing flow, the system can automatically perform quality evaluation, which includes four dimensions: accuracy, completeness, consistency, and reliability. When the comprehensive quality score exceeds the quality evaluation threshold, the processing flow of the downstream agent is automatically triggered based on the preset trigger mechanism. The trigger mechanism can be in the form of asynchronous message passing, avoiding direct coupling between agents.

[0116] The system can maintain a dependency graph between the agents to clearly define the preconditions for data transmission. The output of the document understanding agent is the input premise of the knowledge reasoning agent and the IO configuration agent. The verification result of the knowledge reasoning agent is an important basis for the IO configuration agent to optimize its decision. The power supply demand of the IO configuration agent is a necessary input for the design of the drawing agent. The system automatically determines the triggering time based on the dependency relationship.

[0117] To improve data transmission efficiency, the system establishes a data caching mechanism between agents. The output data of the upstream agent is temporarily stored in a shared cache, and the downstream agent can directly obtain the required data from the shared cache, avoiding repeated transmission. A time-based and version-based invalidation strategy is used in the shared cache to ensure the timeliness and consistency of the data. When the output of the upstream agent changes locally, the system uses an incremental update mechanism to transmit only the changed data. The downstream agent receives the incremental update and determines whether the change affects the generated results, deciding whether to reprocess or adjust locally. By using the incremental update mechanism, the response efficiency of the system can be greatly improved.

[0118] In one embodiment, the document understanding agent is further configured to: when the quality evaluation result is less than a preset collaboration trigger threshold, generate a term verification collaboration request and send it to the knowledge reasoning agent through a message queue, so that the knowledge reasoning agent performs term verification processing; the term verification collaboration request includes: request type, problem description information, expected result, urgency, maximum waiting time.

[0119] The IO configuration agent is further configured to: when there is a rule conflict or the constraint condition is not met, generate a rule reasoning collaboration request and send it to the knowledge reasoning agent and the drawing agent, so that the knowledge reasoning agent and the drawing agent participate in negotiation.

[0120] The drawing agent is further configured to: when the layout structure of the power cabinet drawing does not meet the preset layout requirements, generate a configuration adjustment collaboration request and send it to the IO configuration agent, so that the IO configuration agent adjusts and updates the IO configuration result in response to the configuration adjustment collaboration request after passing the capability matching check.

[0121] Specifically, each agent may encounter problems beyond its own capabilities during processing, and when encountering problems beyond its own capabilities, an automatic collaboration request is identified. For example, when the preset collaboration trigger threshold is 0.8, the document understanding agent initiates a term verification collaboration request to the knowledge reasoning agent when the term recognition confidence is less than 0.8, so that the knowledge reasoning agent performs term verification processing. When the IO configuration agent encounters rule conflicts or constraint dissatisfaction, it initiates a rule reasoning collaboration request and sends it to the knowledge reasoning agent and the drawing agent, so that the knowledge reasoning agent and the drawing agent participate in negotiation. When the drawing agent encounters a layout structure that does not meet the preset layout requirements, it initiates a configuration adjustment collaboration request and sends it to the IO configuration agent, so that the IO configuration agent, after passing the capability matching check, responds to the configuration adjustment collaboration request and adjusts and updates the IO configuration result.

[0122] The system automatically selects the target agent most suitable for collaboration according to the collaboration demand type and the agent capability matching, and sends a collaboration request. The selection algorithm when determining the target agent for collaboration needs to consider three factors: professional matching degree, current load state, and historical collaboration success rate. The professional matching degree is calculated based on the matching degree of the agent's capability description and the problem domain. The load state is used to reflect the current processing capacity and response time of the agent. The historical success rate is based on the success and quality evaluation of previous collaboration.

[0123] The above-mentioned collaboration request can be in a standardized message format, which contains information such as request type, problem description, expected result, urgency, maximum waiting time, etc. The current agent in the system sends a collaboration request to the target agent through a message queue and starts a timeout monitoring mechanism. After receiving the collaboration request, the target agent first performs a capability matching check to determine whether it can collaborate and process based on the matching capability, and sends an acceptance confirmation message when it determines that it can collaborate and process. After completing the collaboration processing, the target agent obtains the collaboration result and returns it to the current agent of the request party through a response message. The response message contains information such as processing result, confidence evaluation, suggested solution, processing time, etc. After receiving the response message, the request party integrates the collaboration result into its own processing flow and sends a confirmation message to the target agent indicating that the collaboration is complete.

[0124] The system can also maintain a distributed global state database that records the running state, task progress, resource usage, performance indicators, and other information of each agent in real time. The global state database can use a key-value storage structure to support high-concurrency read and write operations and millisecond-level query responses. Each agent actively pushes state update information to the global state database at key nodes. The state update information includes task start, progress change, quality evaluation, task completion, and abnormal occurrence events. The system uses an asynchronous update mechanism to avoid state synchronization affecting the normal processing performance of the agent. At the same time, a state change notification mechanism is established, and related agents can subscribe to specific state change notifications.

[0125] To ensure state consistency in a distributed environment, the system can use a timestamp-based version control mechanism. Each state update is assigned an incremental version number and an accurate timestamp. When a state conflict occurs, the system resolves the conflict based on the timestamp order to ensure the eventual consistency of the state. For critical state changes, a two-phase commit protocol is used to ensure strong consistency. When an agent fails or the network is interrupted, the system uses state snapshots and operation logs to achieve fast recovery. State snapshots save the complete state information of the agent, and operation logs record all changes after the state snapshot. During fault recovery, the system starts from the latest state snapshot and replays the operation log to recover to the state before the failure.

[0126] The system can establish a multi-level exception detection system that includes three levels of agent internal exception detection, inter-agent interaction exception detection, and system-level exception detection. The agent internal exception detection is used to monitor internal states such as processing performance, resource usage, and quality indicators. The inter-agent interaction exception detection is used to monitor interaction processes such as message transmission, collaborative response, and data consistency. The system-level exception detection is used to monitor global states such as overall performance, load distribution, and service availability.

[0127] The system can classify exceptions into four levels: Level 1 exception, Level 2 exception, Level 3 exception, and Level 4 exception. Level 1 exception can be understood as a minor exception, Level 2 exception can be understood as a medium exception, Level 3 exception can be understood as a serious exception, and Level 4 exception can be understood as a fatal exception. Minor exceptions are handled by the agent itself and automatically recovered through parameter adjustment and retry mechanisms. Medium exceptions trigger inter-agent collaboration processing to solve problems with the support of collaboration partners. Serious exceptions trigger system-level intervention, which may require task redistribution or service degradation. Fatal exceptions will cause the task to fail and require manual intervention.

[0128] When a single agent cannot solve the exception autonomously, the system starts the collaborative recovery mechanism. First, identify the target agent of the collaboration that can provide help, and then send a collaboration request explaining the abnormal situation and the expected type of assistance. The target agent decides whether to accept the collaboration request according to its own capabilities and current load. Among them, a timeout protection mechanism can be used in the collaboration process to avoid the collaboration process from being stuck in infinite waiting. When the system load is too high or the key agent is unavailable, the system can start the service degradation mechanism. The degradation strategy includes three types of function degradation, performance degradation, and quality degradation. Function degradation turns off non-critical functions to concentrate resources on core tasks. Performance degradation can extend the processing time and reduce concurrency to reduce system pressure. Quality degradation reduces the quality threshold to accept relatively low-quality processing results to ensure task completion.

[0129] Exemplarily, taking the processing of the IO list PDF document of the reactor coolant system (RCS) of a certain nuclear power project as an example, the complete interaction process between agents is described in detail. First, the task is started and the state is initialized. After the system receives the document processing task of "AA-RCS system-IO list-RevB.pdf", the global task record is created through the task scheduling module, and the task identification is set as TASK_20241206_140000_001. The task priority is set as high, and the expected completion time is 180 seconds.

[0130] Please refer to Figure 7 As shown in the figure, the system creates a task state record in the global state database, sets the current execution state as "to be processed", and sends a task processing request including the document processing task to the document understanding agent, so as to enter the document understanding agent processing stage. After receiving the document processing task, the document understanding agent first updates the current task state to "document parsing" in the task state database, and starts a 30-second period progress reporting mechanism. The document understanding agent can use a multi-strategy adaptive parsing algorithm to process the technical document, identify the standard format engineering table document, select the target parsing strategy as the table parsing strategy and process it. Among them, the above technical document can include 25 pages, and the progress state is updated once every 5 pages during processing. The system monitors the CPU usage and memory occupation of the document understanding agent.

[0131] After the document understanding agent completes processing, a four-dimensional quality assessment process of accuracy, integrity, consistency, and credibility is performed to obtain a quality assessment result (i.e., a comprehensive quality score), for example, an accuracy score of 0.92, an integrity score of 0.94, a consistency score of 0.89, a credibility score of 0.91, and a comprehensive quality score of 0.915. When the comprehensive quality score exceeds the preset collaboration trigger threshold of 0.85, the system automatically starts the quality-driven trigger mechanism, stores the feature recognition result to the shared cache, and sends a processing notification to the knowledge reasoning agent. When the document understanding agent finds that the recognition confidence of 12 terms is lower than 0.8 in the quality assessment process, including the key signal bit numbers “TT-101-AI”, “PT-102-AI”, etc. The system automatically identifies the collaboration requirement as term verification, determines the knowledge reasoning agent as the best collaboration object according to the collaboration partner selection algorithm and takes it as the target agent. The document understanding agent constructs a collaboration request message, which contains: a list of terms to be verified, context information, expected confidence, etc., and sends the collaboration request to the target agent through the message queue.

[0132] After the document understanding agent sends the collaboration request to the knowledge reasoning agent, the system updates the current execution state in the task state database to “term verification in progress”, and starts the collaboration timeout monitoring for 120 seconds. The knowledge reasoning agent obtains the feature recognition result from the shared cache and starts the term verification process. During the processing, the knowledge reasoning agent regularly pushes progress updates to the task state database, and the document understanding agent can monitor the collaboration progress in real time. After receiving the collaboration request, the knowledge reasoning agent first performs capability matching check to confirm whether it has the capability of nuclear power term verification, and sends an acceptance confirmation message to the document understanding agent when it confirms that it has the capability of nuclear power term verification. The knowledge reasoning agent finds the related terms from the nuclear power term knowledge base, finds that “TT-101-AI” matches the temperature transmitter standard term with a confidence of 0.98, and “PT-102-AI” matches the pressure transmitter standard term with a confidence of 0.96. After the processing is completed, the collaboration response message is constructed according to the term verification result to return the term verification result.

[0133] After receiving the collaborative response message from the knowledge reasoning agent, the document understanding agent integrates the term verification result into the original extraction data, and updates the confidence score of the related terms. At this time, the integrated data quality score is improved to 0.945, which exceeds the collaborative trigger threshold that triggers downstream processing. The system automatically stores the integrated data into the shared cache, and determines the IO configuration agent as the next processing node through the dependency judgment mechanism, and sends a processing trigger notification. After receiving the processing notification, the IO configuration agent obtains the feature recognition result and the reasoning result from the shared cache. The IO configuration agent first performs data integrity check to confirm that it has obtained complete information of 150 IO points, and then starts the multi-rule fusion configuration algorithm to perform signal type judgment, card selection, power supply calculation, terminal selection and other processing for each IO point. During the processing of the IO configuration, progress updates are regularly pushed to the task state database to update the current task execution state.

[0134] Among them, the IO configuration agent encounters a rule conflict during processing. For the signal "TT-103-AI", two different rules can produce different card recommendations: rule R001 recommends AI216 card (applicability 0.75), and rule R005 recommends AI232 card (applicability 0.72). The system automatically identifies the conflict, and the IO configuration agent initiates a conflict resolution negotiation request, inviting the knowledge reasoning agent and the drawing agent to participate in the negotiation. This negotiation process is executed according to a four-stage process. The first stage is that the IO configuration agent sends a negotiation request to the knowledge reasoning agent and the drawing agent, describing the conflict situation and the candidate solution. The second stage is that the knowledge reasoning agent and the drawing agent confirm participation in the negotiation, and the system sets a negotiation timeout of 180 seconds. When the negotiation timeout is exceeded, it is defaulted as unconfirmed participation in the negotiation. The third stage is that each agent scores according to the professional dimension, and the knowledge reasoning agent scores the AI232 solution 89 points from the rule compliance perspective, and the drawing agent scores the AI232 solution 82 points from the layout optimization perspective. The fourth stage is that the system calculates the comprehensive score according to the dynamic weight of the knowledge reasoning agent and the drawing agent, and then the AI232 solution wins with 85.6 points. The negotiation result is notified to all participants, including the IO configuration agent, the knowledge reasoning agent and the drawing agent.

[0135] During the negotiation of each agent, the system continuously monitors the response state of each agent. When the response time of the drawing agent exceeds 60 seconds, the system sends a state query message to confirm the state of the agent and generates a drawing agent report. Among them, the current load of the drawing agent report is high but can still be processed normally, the system will adjust its weight and continue to negotiate, and finally it can be negotiated to complete within 156 seconds, without triggering timeout exception. The IO configuration agent adjusts the IO configuration scheme of TT-103-AI according to the negotiation result, and only updates the changed part by using the incremental update mechanism to avoid recalculation of the entire configuration scheme. The updated configuration data is automatically synchronized to the shared cache, and the cache manager ensures that the latest configuration information is obtained by subsequent agents according to the data version number. The system also cleans up expired cache data to release storage space.

[0136] After receiving the power supply demand data from the IO configuration agent, the drawing agent starts to execute the drawing generation task. Among them, the drawing request contains the Hualong No. 1 demonstration project RCS system project information, drawing number E-DCS-RCS-001, version number A, power supply demand data, 1E01 cabinet configuration information, physical location of 1E cabinet room A area, standard template type, compact layout style, detailed annotation level of drawing preference. The drawing agent uses intelligent template selection and standardized layout algorithm for processing, first performs intelligent analysis of power supply demand, determines the redundant power supply scheme according to the 1E level key load and 150 watt power demand. Then perform power supply cabinet layout design, adopt function partition layout strategy, arrange main circuit breaker area, distribution circuit breaker area, control element area, wiring terminal area from top to bottom. Then perform CAD parameterized drawing, call circuit breaker, power module and other electrical element symbols from standard element library, draw drawings according to national standard specifications.

[0137] After the drawing is completed, quality testing is performed, and when the quality testing is passed, the power supply cabinet drawing is output according to the drawing output type, which includes two types of standard drawings of power supply cabinet principle diagram and power supply cabinet wiring diagram. The power supply cabinet principle diagram file name can be "AA-RCS system-power supply cabinet principle diagram-RevA point dwg", including QF1 main circuit breaker 100A / 400V specification protection 1E level DCS system, PS1 power module DC24V_5A input 220VAC output 24VDC N+1 redundant configuration and other key components. The power supply cabinet wiring diagram file name can be "AA-RCS system-power supply cabinet wiring diagram-RevA point dwg", including QF1-2 to PS1-L W001 brown BV-2.5 square millimeter cable 850 millimeter length and other detailed connection information. The design summary information shows that the total power level is 1200 watts, the N+1 redundancy level, the selective protection cooperation, the cost estimate of 45000 yuan, and it meets the standards of GB50055-2011, IEEE-603, RCC-E. The quality inspection result shows that the connection integrity is 100%, the labeling accuracy is 99%, the standard compliance is 98%, and the layout optimization degree is 91%. The feedback suggestion points out that the optimization space of 15% heat dissipation efficiency can be improved by load redistribution.

[0138] The total time of the whole end-to-end processing flow in this embodiment is 138.6 seconds, which is greatly shortened compared with 5 to 8 days of traditional manual design. The total number of data exchange messages is 12, the number of agent collaboration is 8 times, and the final quality score is 94.2%, realizing zero manual intervention and full automation processing. The specific decomposition is that the document understanding agent processes for 67.5 seconds, the knowledge reasoning agent verifies for 2.1 seconds, the IO configuration agent configures for 45.3 seconds, and the drawing agent draws for 23.7 seconds. Data transmission includes standardized data flow such as document extraction result, term verification request, verification result response, configuration generation request, configuration result output, drawing generation request, and drawing output result, realizing efficient, accurate, and standardized nuclear power DCS design full-process automation.

[0139] In the embodiment, the artificial intervention rate is reduced to less than 30% from the traditional artificial dominant through the end-to-end autonomous cooperation control mechanism, the design cycle is shortened from 20-28 weeks to 8-12 weeks, and the design efficiency is improved by 2-3 times. It is verified through experiments that for a large project containing 200 IO points and 20 cabinets, the traditional artificial design needs 20-30 days, and the processing time of the application can be 2-3 days; a whole-process preventive quality control network is established, the quality problem detection rate is more than 70%, and the design error rate is reduced by more than 50%. And support multiple document formats, multiple nuclear power technology routes, and multiple equipment types of design documents, which can handle design requirements of different technology routes such as AP1000 and Hualong No. 1. By setting the document understanding agent, knowledge reasoning agent, IO configuration agent and drawing agent, human intervention can be reduced, each link can run independently, and can automatically process when encountering abnormal conditions, through the construction of end-to-end autonomous cooperation control layer and data interaction layer, the whole-process autonomous processing from technical document input to complete CAD drawing set output is realized, the design quality and rationality of power supply cabinet drawing are guaranteed, and the efficiency, stability and design quality are improved.

[0140] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0141] The principles and implementation modes of the present application are described by specific examples in this paper, and the above embodiment description is only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A nuclear power digital processing system characterized by, The nuclear power digital processing system comprises an input layer, a collaborative control layer, a plurality of agents, a data interaction layer and an output layer connected in sequence, the plurality of agents comprise a document understanding agent, a knowledge reasoning agent, an IO configuration agent and a drawing agent connected in sequence, the drawing agent is connected with the output layer, and the document understanding agent is connected with the input layer; The input layer is configured to acquire a nuclear power digital design file when receiving a document processing task; The collaborative control layer is configured to perform collaborative control processing on the nuclear power digital design file through the plurality of agents to obtain a power cabinet drawing, and the collaborative control processing comprises at least one of task scheduling, task negotiation, quality monitoring and exception handling; The document understanding agent is configured to perform labeling and analysis processing on the nuclear power digital design file to obtain a quality evaluation result, the knowledge reasoning agent is configured to perform knowledge management and engineering rule reasoning verification processing according to the quality evaluation result to generate a reasoning result, the IO configuration agent is configured to generate an IO list according to the reasoning result and perform cabinet and machine allocation optimization processing to obtain an IO configuration result, and the drawing agent is configured to perform power supply demand analysis, component selection and schematic design operations according to the IO configuration result to obtain the power cabinet drawing; The data interaction layer is configured to perform data interaction processing between the agents; The output layer is configured to output the power cabinet drawing.

2. The nuclear power digital processing system of claim 1, wherein, The document understanding agent is specifically configured to: perform identification processing on the nuclear power digital design file to obtain a feature recognition result; the feature recognition result comprises a file name mode, a content feature or a page layout; adopt a decision tree model to select a target analysis strategy according to the feature recognition result, and the target analysis strategy comprises a table analysis strategy, a robust analysis strategy or a hybrid analysis strategy; when the target analysis strategy is the table analysis strategy or the robust analysis strategy, perform table boundary detection and cell segmentation processing to obtain a quality evaluation result; when the target analysis strategy is the hybrid analysis strategy, perform term standardization processing to obtain a quality evaluation result.

3. The nuclear power digital processing system of claim 1, wherein, The knowledge reasoning agent is specifically configured to: when the quality evaluation result is greater than a collaboration triggering threshold, acquire term information and factual conditions based on the nuclear power digital design file, the term information comprises a term list and term context information; evaluate the term information to obtain a term verification result and send the term verification result to the document understanding agent; and / or perform reasoning according to the factual conditions to obtain a reasoning result and send the reasoning result to the IO configuration agent; and / or when a plurality of agents conflict, determine a conflict type and adopt a preset rule knowledge base to make a ruling based on the conflict type to obtain a coordination result, and the preset rule knowledge base comprises a domain division rule, a priority rule and a use range rule.

4. The nuclear power digital processing system of claim 2, wherein, The IO list comprises a card type, a power supply mode and a terminal plate capacity, and the IO configuration agent is specifically configured to: determine multi-dimensional features, power type coding information, and sensor power supply requirement information based on the inference result and the feature recognition result; the multi-dimensional features include description semantics, electrical characteristics, and engineering units; analyze the description semantics to identify signal functions, analyze the electrical characteristics to determine signal types, and establish a mapping relationship between the engineering units and the signal types; determine a card type according to the signal functions, signal types, electrical characteristics, mapping relationship, and communication protocol; determine a power supply mode based on the power type coding information and the sensor power supply requirement information; combine and match the card type, the power supply mode, and the electrical characteristics, establish a matching mapping table, and determine a terminal board capacity according to the matching mapping table.

5. The nuclear power digital processing system of claim 1, wherein, The IO configuration intelligent agent is also used to: extract multiple state space information based on the IO list and perform feature fusion to obtain comprehensive feature information; the state space information includes cabinet machine state information, IO point state information, and constraint state information; the cabinet machine state information includes load rate, available slot, and safety level, the IO point state information includes type code, power level, and priority; the constraint state information reflects load balancing degree and constraint condition; obtain a decision result through an initial decision network; calculate a reward value according to a weight value corresponding to a preset index, and update parameters of the initial decision network based on the reward value to construct a decision network; the preset index includes load balancing, maintenance convenience, cost optimization, and safety compliance; perform cabinet machine allocation processing through the decision network to obtain an IO configuration result.

6. The nuclear power digital processing system of claim 1, wherein, The drawing drawing intelligent agent is specifically used to: statistically obtain power supply requirement information according to the IO configuration result; the power supply requirement information includes working current and power consumption information of various card types; calculate power supply capacity according to a safety margin rule based on the power supply requirement information; determine element spacing and partition layout mode of each element in a power supply cabinet, and busbar cross-sectional area and busbar connection path; the partition layout mode includes functional partition layout mode, vertical layout mode, and horizontal layout mode; the busbar includes main busbar and distribution busbar; obtain a power supply configuration type, select a power supply cabinet drawing template library according to the power supply configuration type; the power supply configuration type includes single power supply, double power supply, and USB power supply; based on the power supply cabinet drawing template library, call each component symbol and draw through a connection line to generate a drawing and perform quality detection processing; the components include at least one of the following: circuit breaker, contactor, relay, indicator light; when the quality detection passes, generate a power supply cabinet drawing according to a drawing output type; the power supply cabinet drawing includes a power supply cabinet schematic diagram and a power supply cabinet wiring diagram.

7. The nuclear power digital processing system of claim 1, wherein, The system further includes a global state machine and a task scheduling module; the global state machine includes a task state database for recording current execution state, execution intelligent agent, completion progress, and quality score of each task. The task scheduling module is configured to create a global task record and assign a task identification and task information to the document processing task when the document processing task is received.

8. The nuclear power digital processing system of claim 1, wherein, The document understanding agent is further configured to generate a term verification collaboration request and send the term verification collaboration request to the knowledge reasoning agent through a message queue when the quality evaluation result is less than a preset collaboration triggering threshold, so that the knowledge reasoning agent performs term verification processing. The term verification collaboration request includes a request type, problem description information, an expected result, an urgency level, and a maximum waiting time. The IO configuration agent is further configured to generate a rule reasoning collaboration request and send the rule reasoning collaboration request to the knowledge reasoning agent and the drawing generation agent when a rule conflict or a constraint condition is not satisfied, so that the knowledge reasoning agent and the drawing generation agent participate in negotiation. The drawing generation agent is further configured to generate a configuration adjustment collaboration request and send the configuration adjustment collaboration request to the IO configuration agent when a layout structure of the power supply cabinet drawing does not meet a preset layout requirement, so that the IO configuration agent adjusts and updates the IO configuration result in response to the configuration adjustment collaboration request after passing an ability matching check.

9. The nuclear power digital processing system of claim 1, wherein, When the collaboration control processing is task negotiation, the system is further configured to: determine a negotiation request; the negotiation request is sent by the IO configuration agent to the knowledge reasoning agent and the drawing generation agent, and the negotiation request includes conflict information and candidate processing strategies; obtain a rule complexity dimension score of the knowledge reasoning agent and a layout optimization dimension score of the drawing generation agent; determine a problem type according to the conflict information, and calculate a comprehensive score value based on the rule complexity dimension score, the layout optimization dimension score, and a dynamic weight value; determine a target processing strategy from all candidate processing strategies based on the comprehensive score value, and notify the IO configuration agent, the knowledge reasoning agent, and the drawing generation agent.

10. The nuclear power digital processing system of claim 1, wherein, When the collaboration control processing is abnormal processing, the system is further configured to: detect whether an exception occurs in each agent interaction, agent internal, and system level, and determine an exception type; the exception type includes a first-level exception, a second-level exception, a third-level exception, and a fourth-level exception; the first-level exception, the second-level exception, the third-level exception, and the fourth-level exception are arranged in order from small to large according to an exception degree; perform a corresponding exception processing strategy according to the exception type to obtain an exception processing result.

Images