Multi-agent cooperative scheduling method and system for scientific and technological information analysis

Abstract

The application discloses a multi-agent collaborative scheduling method and system for scientific and technological information analysis, relates to the technical field of scientific and technological information analysis, and comprises the following steps: a task publishing layer outputs a standardized task description by performing task demand element analysis; an intelligent coordination and scheduling layer performs closed-loop collaborative decision after receiving the standardized task description and outputs a task allocation instruction; and a professional agent execution layer drives a data acquisition agent group to acquire scientific and technological information data through a data access adapter of a data resource layer and execute a scientific and technological information analysis task, and outputs a structured processing result, which is visually presented to a user through the task publishing layer. The application solves the technical problem that a scientific and technological information processing system in the prior art lacks intelligent task allocation, leading to low information processing efficiency, achieves four-layer architecture closed-loop collaboration and efficient scheduling of a professional agent group for scientific and technological information analysis tasks, and improves the technical effects of task allocation accuracy and information processing efficiency.

Description

Technical Field

[0001] This invention relates to the field of scientific and technological intelligence analysis technology, specifically to a multi-agent collaborative scheduling method and system for scientific and technological intelligence analysis. Background Technology

[0002] The field of scientific and technological intelligence analysis is increasingly demanding higher efficiency, quality, and professionalism in its processing. However, existing scientific and technological intelligence processing systems suffer from significant technical shortcomings: centralized systems are prone to processing bottlenecks due to the performance limitations of a single server; distributed systems rely on simple polling or random allocation strategies and lack global optimization capabilities; and workflow-based systems rely solely on time or event-triggered scheduling without intelligent resource allocation mechanisms. These systems generally suffer from problems such as a lack of intelligent task allocation, low resource utilization efficiency, insufficient consideration of domain specialization, weak collaboration capabilities, and poor adaptability, making it difficult to meet the actual needs of scientific and technological intelligence analysis for efficient, accurate, and professional processing.

[0003] Existing technology-based intelligence processing systems lack intelligent task allocation, resulting in low intelligence processing efficiency. Summary of the Invention

[0004] This application provides a multi-agent collaborative scheduling method and system for scientific and technological intelligence analysis, which addresses the technical problem of low intelligence processing efficiency caused by the lack of intelligent task allocation in existing scientific and technological intelligence processing systems.

[0005] In view of the above problems, this application provides a multi-agent cooperative scheduling method and system for scientific and technological intelligence analysis.

[0006] The first aspect of this application provides a multi-agent cooperative scheduling method for scientific and technological intelligence analysis, the method comprising:

[0007] After receiving the user's submitted scientific and technological intelligence analysis request, the task publishing layer parses the task requirement elements and outputs a standardized task description. The intelligent coordination and scheduling layer receives the standardized task description through a standardized interface and performs closed-loop collaborative decision-making via a capability assessment module, intelligent auction module, conflict resolution module, and collaborative optimization module, outputting a task allocation instruction. The professional intelligent agent execution layer receives the task allocation instruction from the intelligent coordination and scheduling layer through a standardized interface, drives a group of data acquisition intelligent agents to acquire scientific and technological intelligence data through the data access adapter of the data resource layer, and coordinates with data processing intelligent agents, knowledge mining intelligent agents, and decision support intelligent agents to execute the scientific and technological intelligence analysis task, outputting structured processing results. These structured processing results are then visualized and presented to the user through the task publishing layer.

[0008] A second aspect of this application provides a multi-agent cooperative scheduling system for scientific and technological intelligence analysis, the system comprising:

[0009] The system comprises the following layers: a task publishing layer, which receives and transforms user requests for scientific and technological intelligence analysis into standardized task descriptions; an intelligent coordination and scheduling layer, which includes a capability assessment module, an intelligent auction module, a conflict resolution module, and a collaborative optimization module interconnected via a distributed event bus to form a dynamic decision-making closed loop; a professional intelligent agent execution layer, which includes isolated and parallel groups of data acquisition intelligent agents, data processing intelligent agents, knowledge mining intelligent agents, and decision support intelligent agents; a data resource layer, which manages multi-source heterogeneous scientific and technological intelligence data sources and provides standardized data interfaces for data access services to the data acquisition intelligent agents, data processing intelligent agents, knowledge mining intelligent agents, and decision support intelligent agents through data access adapters; and a task publishing layer that visualizes the processing results generated by the professional intelligent agent execution layer to the user. These layers—data resource layer, professional intelligent agent execution layer, intelligent coordination and scheduling layer, and task publishing layer—are sequentially connected via standardized interfaces to form a loosely coupled distributed architecture.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] After receiving the scientific and technological intelligence analysis request submitted by the user, the task publishing layer parses the task requirement elements and outputs a standardized task description. The intelligent coordination and scheduling layer receives the standardized task description through a standardized interface and performs closed-loop collaborative decision-making via a capability assessment module, intelligent auction module, conflict resolution module, and collaborative optimization module, outputting a task allocation instruction. The specialized intelligent agent execution layer receives the task allocation instruction from the intelligent coordination and scheduling layer through a standardized interface, drives a group of data acquisition intelligent agents to acquire scientific and technological intelligence data through the data access adapter of the data resource layer, and coordinates with data processing intelligent agents, knowledge mining intelligent agents, and decision support intelligent agents to execute the scientific and technological intelligence analysis task, outputting structured processing results. These results are then visualized and presented to the user through the task publishing layer. This achieves a four-layer architecture for closed-loop collaboration in scientific and technological intelligence analysis tasks and efficient scheduling of specialized intelligent agent groups, improving the accuracy of task allocation and the efficiency of intelligence processing. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1A schematic diagram of the multi-agent cooperative scheduling method for scientific and technological intelligence analysis provided in this application embodiment;

[0014] Figure 2 A schematic diagram of the structure of a multi-agent collaborative scheduling system for science and technology intelligence analysis provided in this application embodiment.

[0015] Figure labeling: Task release layer 10, intelligent coordination and scheduling layer 20, professional intelligent agent execution layer 30, data resource layer 40. Detailed Implementation

[0016] This application provides a multi-agent collaborative scheduling method and system for science and technology intelligence analysis, which addresses the technical problem of low intelligence processing efficiency caused by the lack of intelligent task allocation in existing science and technology intelligence processing systems.

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Example 1, as Figure 1 As shown, this application provides a multi-agent cooperative scheduling method for scientific and technological intelligence analysis, the method comprising:

[0019] Step S100: After receiving the scientific and technological intelligence analysis request submitted by the user, the task publishing layer outputs a standardized task description by performing task requirement element parsing.

[0020] Specifically, the task publishing layer, serving as the interaction portal between users and external systems, first receives user-submitted scientific and technological intelligence analysis requests. These requests cover diverse scenarios such as literature reviews in specific technical fields, dynamic tracking of patent technologies, policy compliance analysis, and technology trend prediction. Initially, these requests are often unstructured descriptions, such as natural language text or simple request lists. Subsequently, the task publishing layer initiates a task requirement element analysis process, systematically extracting core elements from user requests. These elements include the technical field scope, such as specific fields like artificial intelligence, biomedicine, and new energy. They must match the first-level skill domains and second-level skill categories in a pre-built hierarchical skill tagging system, as well as data type requirements, such as academic literature, patent documents, policies and regulations, and industry reports. The layer clarifies the source priority and task urgency level of each data type, classifying them into levels 1-5. This directly relates to subsequent emergency reward calculations and quality standards, such as quantitative thresholds for accuracy ≥95%, completeness ≥90%, and consistency ≥85%. It also corresponds to quality assessment indicators and output format requirements for data processing and analysis, such as structured analysis reports, visualized knowledge graphs, decision suggestion lists, and data tables. After the element extraction is completed, the task release layer calls the preset requirement-standard mapping rule library to transform the extracted core elements into a standardized task description containing unified fields, a unique task ID, a set of technical field tags, a list of data requirements and priorities, an urgency level identifier, quality threshold parameters, output format specifications, and a task submission deadline. This ensures that the description can be accurately identified and parsed by the downstream intelligent coordination and scheduling layer through a standardized interface, providing a unified and standardized task benchmark for subsequent closed-loop collaborative decision-making.

[0021] Step S200: After receiving the standardized task description through the standardized interface, the intelligent coordination and scheduling layer performs closed-loop collaborative decision-making through the capability assessment module, intelligent auction module, conflict resolution module and collaborative optimization module, and outputs task allocation instructions.

[0022] Specifically, after receiving the standardized task description output by the task release layer through a standardized interface, the intelligent coordination and scheduling layer initiates a closed-loop collaborative decision-making process consisting of a capability assessment module, an intelligent auction module, a conflict resolution module, and a collaborative optimization module. First, the capability assessment module retrieves the agent status of all candidate agents from the professional agent execution layer, including current load, skill level, and historical task performance. Combined with the task requirements extracted from the standardized task description, such as technical field, data type, and quality requirements, it uses a six-dimensional capability assessment model covering technical expertise, data processing capability, load status, historical reputation, task matching degree, and collaborative capability to quantitatively evaluate the candidate agents. Technical expertise is calculated using a hierarchical skill tagging system, including first-level skill domains, second-level skill categories, third-level skill items, and agent skill level certification results. Data processing capability is calculated based on the processing speed and quality scores of different data types, namely documents, patents, and policy documents. Finally, a six-dimensional capability score is output for each candidate agent. Next, the intelligent auction module executes multiple rounds of sealed auctions: the pre-qualification round matches candidate agents in the technical field, checks basic capability thresholds, availability, and permissions. Agents that pass the pre-qualification enter the bidding round, where a basic bid is generated based on a bidding function that integrates basic cost, domain expertise coefficient, load penalty coefficient, reputation reward coefficient, data sensitivity coefficient, and urgency coefficient. The final bid is determined by combining the bidding multiple output by the deep Q-network. The comprehensive evaluation round integrates the capability score from the capability assessment module, the reliability score based on historical reputation and real-time load, the emergency reward based on the task urgency level and the agent's response capability, and the bid score after the bidding conversion. A dynamic weighted comprehensive evaluation function is used to calculate the comprehensive evaluation score for each agent. If the conflict resolution module detects conflicts in data source access, such as database concurrent connection limits, computing resources, or expert resources, it employs a priority preemption strategy based on a six-dimensional capability score and a comprehensive evaluation score. This involves calculating priorities according to task strategic importance and data sensitivity, and using a dynamic resource reallocation strategy. The NSGA-II algorithm is used to optimize resource allocation ratios or intelligent task decomposition strategies, maintaining semantic integrity while resolving conflicts by breaking down complex tasks, and outputting a preliminary task allocation plan. Finally, the collaborative optimization module, based on a collaborative model of the scientific and technological intelligence processing pipeline, covering data acquisition, preprocessing, information extraction, knowledge mining, value assessment, and decision support stages, optimizes the collaborative order and resource allocation of agents in the task allocation plan. This ensures efficient connection between agents at each stage of the pipeline, ultimately outputting accurate task allocation instructions.

[0023] Step S300: After receiving the task allocation instruction issued by the intelligent coordination and scheduling layer through the standardized interface, the professional intelligent agent execution layer drives the data acquisition intelligent agent group to acquire scientific and technological intelligence data through the data access adapter of the data resource layer, and coordinates the data processing intelligent agent group, knowledge mining intelligent agent group and decision support intelligent agent group to perform scientific and technological intelligence analysis tasks and output structured processing results.

[0024] The structured processing results are visualized and presented to the user through the task publishing layer.

[0025] Specifically, the professional intelligent agent execution layer, as the core execution unit for task implementation, receives task allocation instructions from the intelligent coordination and scheduling layer through standardized interfaces. Following the task division and collaboration requirements specified in the instructions, it initiates a collaborative workflow involving multiple intelligent agents. First, it drives the data acquisition intelligent agent group, including document collection agents, patent monitoring agents, and policy tracking agents, to access corresponding data sources through the data access adapter in the data resource layer. This adapter shields the interface differences between different data sources and supports simultaneous connection to academic literature databases such as IEEE Xplore and CNKI, patent databases such as CNIPA and USPTO, policy and regulatory databases such as government websites, industry association platforms, and knowledge graph databases. It accurately filters and acquires scientific and technological intelligence data according to task requirements, while simultaneously completing the initial standardization of data formats to ensure compatibility in subsequent processing. Subsequently, the data acquisition intelligence group transmits standardized raw data to the data processing intelligence group, which includes a text cleaning intelligence group, an entity recognition intelligence group, and a format conversion intelligence group. This group sequentially performs text deduplication, noise filtering, entity extraction (such as identification of organization names, technical concepts, and personnel information), semantic annotation, and format standardization, outputting structured data and processing logs, laying the foundation for the knowledge mining stage. Next, the knowledge mining intelligence group, including a problem discovery intelligence group, a technology association intelligence group, and a trend prediction intelligence group, constructs an entity relationship graph based on the structured data. Through association analysis, it identifies technical bottlenecks and cutting-edge directions; through time series models, it predicts technological development trends; and through anomaly detection, it locates potential risk points, outputting a knowledge graph, trend analysis reports, and risk warning information.

[0026] Finally, the decision support intelligence ensemble, comprising value assessment, solution recommendation, and decision list agents, combines task requirements with preliminary analysis results to conduct a multi-dimensional assessment of the technical value, economic impact, and policy compliance of scientific and technological intelligence. This generates priority ranking results, resource allocation suggestions, and implementation plans. The information output from all stages is integrated to form a structured result, which includes data source descriptions, processing flow records, core conclusions, visualizations, and decision recommendations, with field formats conforming to the preset standards of the task release layer. Ultimately, the structured result is returned to the task release layer via a standardized interface. The task release layer uses visualization tools, such as trend line charts, knowledge graph interactive interfaces, and report preview windows, to visually present the results to users, facilitating quick access to key information and decision-making basis.

[0027] In one possible implementation, step S200 further includes:

[0028] Step S210: The capability assessment module retrieves the states of multiple candidate agents from the professional agent execution layer.

[0029] Step S220: Extract task requirements from the standardized task description.

[0030] Step S230: The capability assessment module performs a six-dimensional capability assessment on the multiple candidate agents based on the task requirements and the states of multiple agents, and outputs multiple six-dimensional capability scores.

[0031] Step S240: Execute a multi-round sealed auction for the multiple candidate intelligent agents through the intelligent auction module, and output multiple comprehensive evaluation scores. The multi-round sealed auction includes a pre-qualification round, a bidding round, and a comprehensive evaluation round.

[0032] Step S250: After detecting resource conflicts, the conflict resolution module uses a multi-dimensional conflict resolution strategy to resolve the conflicts based on the multiple six-dimensional capability scores and multiple comprehensive evaluation scores, and outputs a task allocation scheme.

[0033] Step S260: The collaborative optimization module optimizes the task allocation scheme based on the pipeline collaborative model and outputs the task allocation instruction.

[0034] Specifically, the capability assessment module, as the basic support unit for decision-making in the intelligent coordination and scheduling layer, actively retrieves the agent status of all candidate agents related to the current scientific and technological intelligence analysis task through a pre-set standardized data interaction interface with the professional agent execution layer. The retrieved agent status encompasses multi-dimensional key information, specifically including: the agent's current load status, such as the number of tasks being executed, the complexity of each task, and the remaining processing time, which is used for subsequent calculations of relative load; skill level information that has passed system certification, corresponding to the level of the three-level skill items in the hierarchical skill tagging system, such as the intermediate or advanced certification results of the IEEE Xplore access skill under the document collection category, along with the certification time and validity period; historical task execution records, i.e., the success rate, quality score, timeliness performance, and innovation capability assessment of tasks completed in the past 3 months, providing data for historical reputation calculation; data processing capability parameters, processing speed for different types of scientific and technological intelligence data, such as document processing speed (articles / minute), patent processing speed (items / minute), and historical statistical data on processing quality; and agent collaboration records, i.e., the success rate of upstream and downstream collaboration with other agents, cross-agent collaboration efficiency, and communication quality score, supporting the assessment of collaborative capabilities. This comprehensive status data lays the core foundation for subsequent accurate six-dimensional capability assessment.

[0035] From the standardized task descriptions generated in the early stages, the core task requirements are precisely extracted, including the technical field scope corresponding to the task, such as the machine learning sub-direction under the field of artificial intelligence, which needs to match the second-level skill category in the hierarchical skill tag system, the data type to be processed, such as academic literature, patent documents or policies and regulations, the task urgency level, such as level 1-5, which affects the priority of subsequent bidding and resource allocation, quality requirements, such as quantitative thresholds of accuracy ≥95% and completeness ≥90%, and output format specifications, such as structured reports or knowledge graphs, to ensure that subsequent evaluation and scheduling can accurately match the core requirements of the task.

[0036] The capability assessment module is based on the task requirements extracted from the standardized task description, including core elements such as the technical field scope, target data type, task urgency level, quality requirement threshold, and output format specifications. It also combines the status of multiple candidate agents previously retrieved from the professional agent execution layer, covering information such as the current load, certified skill level, historical task performance, data processing parameters, and collaboration records of each agent. Through a six-dimensional capability assessment model designed specifically for scientific and technological intelligence analysis scenarios, it conducts a comprehensive quantitative assessment of each candidate agent and finally outputs a six-dimensional capability score for each agent.

[0037] In the specific evaluation process, the calculation of the six-dimensional capabilities is closely matched with the task requirements and the actual state of the agent: the technical expertise assessment is calculated using a weighted average method. The core of this method is to combine the importance of each technical field in the current task, the agent's skill level in the corresponding field, and skill certification status to accurately measure the agent's professional capabilities in a specific field of scientific and technological intelligence analysis. First, the domain weights of each technical field in the current task need to be determined. These weights reflect the importance of different technical fields in the task and range from 0 to 1. Next, the agent's skill level in each technical field is obtained, ranging from 1 to 10, reflecting the agent's basic ability level in the corresponding field. Then, the certification factor for each technical field is determined. The certification factor for verified skills is 1.2, and the certification factor for unverified skills is 1.0, used to distinguish the credibility of skills. Then, the numerator is calculated by multiplying the domain weight, corresponding skill level, and certification factor of each technical field, and then summing the results for all technical fields; simultaneously, the denominator is calculated by summing the domain weights of all technical fields. Finally, the sum of the numerators is divided by the sum of the denominators to obtain the agent's technical expertise.

[0038] Data processing capability is calculated by comprehensively considering the speed, quality, and scale of the intelligent agent's ability to process scientific and technological intelligence data. The core formula is: Data Processing Capability = (Processing Speed ​​× Quality Score × Scale Factor) / Baseline Performance. Processing speed is categorized by data type: document processing speed is measured in the number of documents processed per unit time; patent processing speed is measured in the number of patents processed per unit time; and policy document processing speed is measured in the number of policy documents processed per unit time. The quality score is a comprehensive evaluation of the accuracy, completeness, and consistency of the processing results, obtained by multiplying accuracy by 0.4, completeness by 0.3, and consistency by 0.3, and then summing the results. The scale factor uses logarithmic... 10 (Maximum concurrent tasks + 1) / log 10 (11) Acquire and comprehensively measure the efficiency and quality of the agent in processing target data.

[0039] Load state calculation considers not only the number of tasks currently being executed by the agent, but also, more importantly, the actual resource consumption of the agent by task complexity, quantifying it through the concept of relative load. First, the complexity of each current task is calculated. This complexity is determined by the data volume factor, algorithm complexity, domain difficulty coefficient, and time constraint coefficient; multiplying these four factors yields the complexity value for a single task. Next, the total load of all current tasks is calculated, which is the sum of the product of the complexity of each task and its corresponding remaining processing time. Then, this total load is divided by the agent's maximum processing capacity to obtain the load occupancy ratio. Finally, subtracting this load occupancy ratio from 1 gives the load state, ranging from 0 to 1. A value closer to 1 indicates a lighter load, while a value closer to 0 indicates a heavier load.

[0040] Historical reputation is calculated using a domain- and time-based weighted approach, combining the agent's historical performance in science and technology intelligence tasks with a time decay mechanism to comprehensively assess its reliability. First, by technical field, the agent's recent performance is calculated for each field. This performance is derived by weighting four dimensions: success rate, quality score, timeliness, and innovation. Success rate accounts for 0.4, quality score for 0.3, timeliness for 0.2, and innovation for 0.1. These four factors are multiplied and summed to obtain the recent performance value for a single field. Next, a time decay mechanism is introduced, calculating the time decay coefficient using an exponential decay function. This coefficient is related to the interval between the current time and the task completion time; the longer the interval, the smaller the coefficient, thus reducing the impact of long-term performance on the current reputation assessment. The time decay is calculated as: [d] = e^(-0.1 × number of months). Then, the recent performance for each field is multiplied by the corresponding time decay coefficient and the field weight, and the results for all fields are summed. The final sum is the agent's historical reputation, which dynamically reflects the agent's reliability in different technical fields and at different times, providing a reputation reference for task allocation.

[0041] Task matching degree is calculated using a geometric mean to ensure that the agent's capabilities are well-matched with task requirements across four dimensions: technology, data type, urgency, and output format. Specifically, technology matching degree is derived by comparing the similarity between the agent's skill vector and the task requirement vector. The calculation method is to sum the minimum and maximum values, where the minimum value is the smaller of the agent's skills and task requirements in each domain multiplied by the domain weight, and the maximum value is the larger of the agent's skills and task requirements in each domain multiplied by the domain weight. The sum of the former is then divided by the sum of the latter. Data type matching degree is calculated using cosine similarity, i.e., summing the product of the agent's data capabilities and the task's data type, then dividing by the square root of the product of the squares of the agent's data capabilities and the squares of the task's data types. Urgency matching degree assesses the match between the agent's response capability and the task's urgency level, taking the smaller value between the agent's response capability divided by the task's urgency level and 1.0. Output format matching degree calculates the overlap between the output formats supported by the agent and the required format. Finally, the matching degree results across these four dimensions are geometrically averaged to obtain the final task matching degree.

[0042] Collaborative capability is calculated using a geometric mean to comprehensively evaluate an agent's ability to collaborate with other agents to complete complex scientific and technological intelligence tasks, specifically covering four core dimensions. The upstream success rate measures the success rate of an agent in receiving and processing upstream data, calculated as the number of tasks that successfully received and processed upstream data divided by the total number of collaborative tasks. The downstream success rate assesses the success rate of an agent in transmitting data to downstream agents, calculated as the number of tasks in which a downstream agent successfully processed its own output data divided by the total number of transmission tasks. Cross-agent collaboration efficiency compares the time efficiency of collaborative tasks with that of completing tasks individually, calculated as the average completion time of collaborative tasks divided by the time to complete similar tasks individually. Communication quality comprehensively evaluates message delivery, data compatibility, and response timeliness, calculated by multiplying message delivery success rate by 0.4, data format compatibility by 0.3, and response timeliness by 0.3, then summing the results. Finally, the evaluation results of these four dimensions are geometrically averaged to obtain the final collaborative capability value.

[0043] The intelligent auction module executes multiple rounds of sealed auctions for candidate agents in the order of pre-qualification rounds, bidding rounds, and comprehensive evaluation rounds, ultimately outputting the comprehensive evaluation score for each candidate agent. In the pre-qualification round, candidate agents undergo pre-screening checks, including technical field matching checks, basic capability threshold checks, availability checks, and permission checks. Only candidate agents that pass all checks can proceed to subsequent rounds. In the bidding round, basic costs, domain expertise coefficients, load penalty coefficients, reputation reward coefficients, data sensitivity coefficients, and urgency coefficients are pre-calculated. A basic bid is then calculated and output via a bidding function. A state space vector integrating task feature vectors, agent state vectors, environmental context vectors, and historical performance vectors is pre-constructed and input into a deep Q-network to output the bidding multiplier selection result. The final bid is calculated and output based on the basic bid and bid multiplier selection results. In the comprehensive evaluation round, the capability scores of candidate agents are first obtained from the capability assessment module. Then, the reliability score is calculated by combining the historical reputation provided by the capability assessment module and the real-time load reported by the professional agent execution layer. At the same time, the emergency reward is calculated by integrating the task urgency level parsed by the task release layer and the agent response capability output by the capability assessment module. After converting the final bid into a bid score, the capability score, reliability score, emergency reward and bid score are input into the dynamic weight comprehensive evaluation function. The comprehensive evaluation score of each candidate agent is generated through the function calculation.

[0044] The conflict resolution module first monitors the resource usage in the system in real time. When it detects conflicts in data source resources, such as the number of connections to the literature database and API call quotas; computing resources, such as CPU cores and GPU resources; storage resources, such as disk space and database connections; or expert resources, such as the number of domain expert consultation slots, i.e., the total resource demand exceeds the available capacity, it will use a multi-dimensional conflict resolution strategy to handle the conflict based on the six-dimensional capability scores of multiple candidate agents output by the capability assessment module and the comprehensive bidding scores output by the intelligent auction module. The multi-faceted conflict resolution strategies include a priority preemption strategy, a dynamic resource reallocation strategy, and a task intelligent decomposition strategy. The priority preemption strategy first calculates task priorities, combining basic priority, urgency factors, strategic importance, data sensitivity, international competition factors, and policy relevance. If high-priority tasks lack sufficient resources, and low-priority tasks are less than 80% complete, and the benefits of preemption outweigh the costs, then low-priority task resources are preempted. The dynamic resource reallocation strategy constructs a multi-objective optimization model using the NSGA-II algorithm. Under constraints such as resource capacity and minimum requirements, it minimizes priority-weighted delays, resource waste, and quality degradation, maximizing expected value and adjusting resource allocation ratios. The task intelligent decomposition strategy, when conflict severity exceeds a threshold, task complexity is too high, or estimated execution time is too long, splits tasks according to semantic integrity, dependency minimization, load balancing, and professional matching principles, breaking down complex tasks into multiple sub-tasks to distribute resource pressure. After handling resource conflicts through these strategies, the conflict resolution module outputs a task allocation scheme after conflict resolution, clearly defining the tasks and resource ratios corresponding to each candidate agent.

[0045] The collaborative optimization module, based on the task allocation scheme output by the conflict resolution module, combines a pipeline collaborative model specifically designed for scientific and technological intelligence analysis to perform optimization work, ultimately outputting task allocation instructions that can directly guide the execution layer operations of professional intelligent agents. This pipeline collaborative model divides scientific and technological intelligence processing into six key stages: data acquisition, data preprocessing, information extraction, knowledge mining, value assessment, and decision support. Each stage has clear input and output standards and corresponding intelligent agent group divisions of labor. The collaborative optimization module first matches the six-dimensional capabilities of candidate intelligent agents in the task allocation scheme according to the characteristics and requirements of each stage's tasks. For example, in the data acquisition stage, it matches intelligent agents with outstanding technical expertise and data processing capabilities within the data acquisition intelligent agent group, ensuring a high degree of compatibility between the agent's capabilities and the stage's tasks. Subsequently, it optimizes the execution process through various collaborative strategies, including: 1. Intelligent preprocessing caching strategy: predicting requirements based on the historical processing patterns, task types, and data characteristics of downstream intelligent agents, and caching preprocessed data in advance to reduce waiting time; 2. Incremental transmission optimization mechanism: transmitting large datasets in chunks according to the optimal batch size for each intelligent agent, supporting... 1. Parallel processing by intelligent agents; 2. Quality feedback mechanism: real-time monitoring of accuracy and completeness indicators at each stage of processing; if indicators are not met, feedback is sent to upstream intelligent agents to adjust parameters; 3. Load prediction and dynamic adjustment strategy: predicting the load of intelligent agents through time series analysis and external factor assessment, and adjusting resource allocation in advance to avoid overload; at the same time, relying on a unified standardized interface, including data format and control information standards, to ensure seamless connection of data transmission and instruction interaction between intelligent agents at each stage, and finally integrating the optimized task division, collaboration sequence, resource allocation and quality requirements into task allocation instructions to ensure efficient and accurate advancement of the entire process of scientific and technological intelligence processing.

[0046] In one possible implementation, step S250 further includes:

[0047] Step S251: The multi-faceted conflict resolution strategy includes a priority preemption strategy, a dynamic resource reallocation strategy, and a task intelligent decomposition strategy.

[0048] Specifically, the multi-faceted conflict resolution strategy is the core means for the conflict resolution module to address resource conflicts, encompassing three progressive strategies: priority preemption strategy, dynamic resource reallocation strategy, and intelligent task decomposition strategy. Among these, the priority preemption strategy designs its priority calculation logic around the specific characteristics of science and technology intelligence tasks. It comprehensively considers basic priority, urgency factor, strategic importance, data sensitivity, international competition factor, and policy relevance to derive task priority. Basic priority is determined by task type weight and user level weight. Task type weights are categorized as follows: Urgent Intelligence 1.0, Strategic Task 0.9, Routine Monitoring 0.6, and Exploration Task 0.4. User level weights are categorized as follows: National Level 1.0, Ministry Level 0.8, Institutional Level 0.6, and Individual Level 0.4. The urgency factor is calculated based on the task deadline, ranging from 1.0 to 2.0. Strategic importance is composed of three sub-factors: technological criticality, national security relevance, and economic impact. Data sensitivity considers information security requirements. The international competition factor reflects the intensity of technological competition, while the policy relevance assessment evaluates the degree of alignment with national and industry policies, with a value ranging from 0.5 to 1.0. When high-priority tasks face resource shortages, the progress of low-priority tasks is checked first. If low-priority tasks are more than 80% complete, preemption is not recommended. A cost-benefit analysis is then conducted to ensure that the benefits of preemption outweigh the costs. Preemption costs include schedule loss, context switching costs, relationship impacts, and restart costs. Schedule loss equals the current task's completed progress multiplied by the task value. Context switching costs include resource switching overhead and data transmission costs. Relationship impacts consider the negative impact on agent collaboration relationships. Restart costs include resetting time and rework penalties. Once the conditions are met, resource preemption is executed to ensure the progress of important tasks. The dynamic resource reallocation strategy is initiated when preemption is not feasible. It constructs a multi-objective optimization model that simultaneously considers minimizing priority-weighted delay, minimizing resource waste, minimizing quality degradation, and maximizing expected value. Priority-weighted delay is the sum of the products of task delay and priority; resource waste is the sum of all types of resource waste; quality degradation is the sum of the quality degradation of all tasks; and expected value is the sum of the products of task value and completion probability. The model's constraints include: resource capacity constraint (total allocation of each resource does not exceed capacity); minimum requirement constraint (minimum resource requirement of each task must be met); task integrity constraint (total resource allocation of each task reaches the minimum requirement); and agent capability constraint (resources allocated to agents do not exceed their maximum processing capacity). The NSGA-II algorithm is used to solve the model to find the Pareto optimal solution set. The algorithm process includes initializing the population to generate random resource allocation schemes, evaluating fitness and calculating the objective function value of each scheme, non-dominated sorting based on the Pareto optimality principle, crowding distance calculation to maintain solution diversity, selection of crossover and mutation evolution operations, iterating repeatedly until convergence, and finally adjusting the resource allocation ratio of each agent based on the solution results.The intelligent task decomposition strategy targets complex scientific and technological intelligence tasks with a conflict severity exceeding 0.8, task complexity exceeding 8.0, estimated execution time exceeding 4 hours, or data volume exceeding the processing limit of a single agent. It follows the principles of semantic integrity, dependency minimization, load balancing, and professional matching. Semantic integrity ensures that subtasks have complete semantic meaning; dependency minimization reduces data dependencies between subtasks; load balancing ensures that the complexity of subtasks is as uniform as possible; and professional matching ensures that the subtask type matches the professional expertise of available agents. Different decomposition strategies are adopted according to the task type; for literature comprehensive analysis tasks, decomposition is performed by technical field and time period. The process involves decomposition: for patent technology tracking tasks, the tasks are decomposed by patent classification number, applicant, and geographical region; for technology trend prediction tasks, the tasks are decomposed by technology direction and time span. Integration or comprehensive analysis tasks are added to integrate the results of all sub-tasks. After decomposition, sub-task quality checks are performed, including completeness checks (the union of the scope of all sub-tasks equals the scope of the original task), no-overlap checks (the overlap of sub-task scopes is less than 20%), dependency checks (the dependency relationship between sub-tasks should be a DAG structure), and resource requirement checks (the total resource requirement of the sub-tasks does not exceed 120% of the original task). These steps disperse resource pressure to resolve conflicts.

[0049] In one possible implementation, step S230 further includes:

[0050] Step S231: The capability assessment module performs a six-dimensional capability assessment of the multiple candidate agents based on the six-dimensional capability assessment model.

[0051] Step S232: The six-dimensional capability assessment model evaluates six dimensions of capability, including technical expertise, data processing capability, load status, historical reputation, task matching degree, and collaboration capability.

[0052] Specifically, when evaluating the capabilities of multiple candidate agents, the capability assessment module does not use a simple measurement based on a single dimension. Instead, it uses a six-dimensional capability assessment model designed specifically for science and technology intelligence analysis scenarios for quantitative evaluation. This model fully considers the comprehensive requirements of science and technology intelligence tasks on agents' professional capabilities, processing efficiency, resource status, and other aspects. It can comprehensively and accurately reflect the actual capabilities of candidate agents to adapt to tasks, providing a scientific basis for subsequent task allocation.

[0053] The six-dimensional capability assessment model covers the following specific capability dimensions: technical expertise, data processing capability, workload status, historical reputation, task matching degree, and collaborative capability. Technical expertise measures the agent's professional knowledge and processing ability in a specific technical field. Each agent maintains a multi-dimensional skill vector, including the professional level of each technical field, and the calculation requires consideration of domain weights, skill levels, and certification factors. Data processing capability comprehensively considers the agent's data processing speed, quality, and scalability. Speed ​​is statistically differentiated according to data type, quality is calculated using a weighted average of accuracy, completeness, and consistency, and scalability is reflected by a scale factor. Workload status considers not only the number of tasks but also the resource consumption of the agent due to task complexity, calculating the relative load by comparing task complexity with remaining time. Historical reputation is based on… The historical performance of intelligent agents in scientific and technological intelligence tasks is evaluated using a domain- and time-based weighted calculation method, while also incorporating a time decay mechanism to highlight the reference value of recent performance. Task matching degree measures the degree to which the capabilities and characteristics of intelligent agents match the current task requirements. It is comprehensively evaluated from four dimensions: technical field, data type, urgency, and output format, to ensure that the capabilities of intelligent agents are highly compatible with task requirements. Collaboration capability assesses the ability of intelligent agents to collaborate with other intelligent agents to complete complex scientific and technological intelligence tasks. It is measured by four indicators: upstream collaboration success rate, downstream collaboration success rate, cross-agent collaboration efficiency, and communication quality, to ensure the smooth operation of multi-agent collaborative work.

[0054] In one possible implementation, step S232 further includes:

[0055] A hierarchical skill tagging system is pre-constructed, which includes first-level skill domains, second-level skill categories, and third-level skill items.

[0056] Based on the multiple Level 3 skills declared by the first candidate agent, standard test tasks are distributed to the first candidate agent.

[0057] Based on the first quality output of the first candidate agent in completing the standard test task, the skill level of the first candidate agent is evaluated based on the hierarchical skill tagging system to obtain the first skill level certification result.

[0058] Based on the results of the first skill level certification, a technical expertise assessment will be conducted.

[0059] Specifically, the first step is to develop a hierarchical skill labeling system, which is divided into three levels: first-level skill domains, second-level skill categories, and third-level skill items, based on the scope and granularity of skill coverage. The primary skill domains are set around the core business scenarios of science and technology intelligence analysis, specifically covering the domains of literature analysis, patent processing, policy interpretation, technology trend prediction, and risk warning. Each primary skill domain is further subdivided into multiple secondary skill categories. For example, the literature analysis domain includes literature retrieval, literature abstract generation, literature correlation analysis, and literature deduplication; the patent processing domain includes patent retrieval, patent claim analysis, patent family analysis, and patent legal status tracking; and the policy interpretation domain includes policy clause extraction, policy scope definition, and policy impact analysis. Each secondary skill category is further broken down into directly measurable and verifiable tertiary skill items. For example, the literature retrieval category includes tertiary skill items such as IEEE Xplore literature retrieval, CNKI literature retrieval, Web of Science literature retrieval, and multi-source literature integration retrieval; and the patent claim analysis category includes tertiary skill items such as invention patent claim decomposition, utility model patent claim decomposition, and design patent protection scope definition, forming a comprehensive and clearly hierarchical skill classification framework.

[0060] When assessing the skills and technical expertise of the first candidate agent, we first collect multiple Level 3 skills declared by the first candidate agent. These declared skills must be clearly mapped to specific Level 3 skills in the hierarchical skill tagging system. For example, the first candidate agent declares that it possesses three Level 3 skills: IEEE Xplore literature retrieval, invention patent claim breakdown, and policy clause extraction. Subsequently, based on the standard requirements of each Level 3 skill item in the hierarchical skill tagging system, corresponding standard test tasks were distributed to the first candidate agent. Each Level 3 skill item was matched with at least one set of standard test tasks: For the IEEEXplore literature retrieval skill item, retrieval test tasks were distributed with constraints including specific technical keywords such as "machine learning, image recognition", time range such as "last 5 years", and document type such as "journal articles, conference papers", explicitly requiring the output of a list of search results, a ranking of document relevance, and an explanation of the search strategy; For the invention patent claim decomposition skill item, invention patent documents from different technical fields such as electronic information and biomedicine were distributed, requiring the output of the division between independent claims and dependent claims, the results of technical feature extraction, and a preliminary definition of the scope of protection; For the policy clause extraction skill item, science and technology policy documents issued by the state or industry were distributed, requiring the output of key information such as core policy clauses, responsible entities, implementation periods, and support measures, and all standard test tasks were pre-set with a unified input data format, execution process specifications, and expected output standards to ensure the comparability of test results.

[0061] After the first candidate agent completes all standard test tasks, its first quality output is collected, and the quality output is quantitatively evaluated from multiple dimensions: accuracy dimension, the matching rate between the search results and the expected target documents, the completeness and correctness of the patent claim breakdown results, and the accuracy of the information extracted from policy clauses are calculated; timeliness dimension, the actual time taken for the first candidate agent to complete each set of test tasks is recorded and compared with the preset standard time to calculate the time compliance rate; consistency dimension, the first candidate agent is allowed to repeat the same set of test tasks 3 times, and the overlap of the output results of the 3 times is calculated. Then, combining the grading standards of each of the three levels of skill items in the hierarchical skill tagging system, the standard divides skill levels into four levels: basic, intermediate, professional, and expert. Each level corresponds to a specific quality threshold. For example, the basic level requires accuracy ≥70%, timeliness ≥80%, and consistency ≥85%; the intermediate level requires accuracy ≥80%, timeliness ≥85%, and consistency ≥90%; the professional level requires accuracy ≥90%, timeliness ≥90%, and consistency ≥95%; and the expert level requires accuracy ≥95%, timeliness ≥95%, and consistency ≥98%. Based on the quantitative evaluation results of the first quality output, and in accordance with the grading standards, the skill level of the first candidate agent in each declared level three skill item is determined. Finally, the grading results of all level three skill items are integrated to form a first skill level certification result that includes the skill item name, corresponding level, grading basis, and quantitative scores of each dimension. For example, IEEE Xplore literature retrieval skill is at the professional level, invention patent claim decomposition skill is at the intermediate level, and policy clause extraction skill is at the professional level.

[0062] When conducting technical expertise assessments, the results of the first skill level certification are used as the core basic data. In conjunction with the needs of current science and technology intelligence analysis tasks, the first level skill domain, second level skill category, and third level skill item in the hierarchical skill tagging system involved in the task are first identified. Then, based on the task's dependence on each skill item, a domain weight is assigned to each involved third level skill item. The weight of core skill items is set at 0.6-0.8, the weight of important skill items is set at 0.3-0.5, and the weight of auxiliary skill items is set at 0.1-0.2. The skill levels of the first candidate agent in each of the three skill levels involved in the task are then converted into corresponding scores: 1-3 points for basic level, 4-6 points for intermediate level, 7-9 points for professional level, and 10 points for expert level. Technical expertise is then calculated using the formula: Technical Expertise = (Score of each skill level × Corresponding Domain Weight) / (Score of each skill level domain weight). A certification factor is introduced to correct the results. For skill levels that have passed multiple standard tests and have stable results, the certification factor is set to 1.2; for skill levels that pass certification for the first time, the certification factor is set to 1.0. This results in the first candidate agent's technical expertise score for the current task, ensuring that the score reflects both the agent's true skill level and accurately matches the task requirements.

[0063] In one possible implementation, step S240 further includes:

[0064] Step S241: In the prequalification round, a pre-screening condition check is performed on the second candidate agent. The pre-screening condition check includes a technical field matching check, a basic capability threshold check, an availability check, and a permission check.

[0065] Step S242: If the pre-approval is successful, proceed to the bidding round.

[0066] Step S243: In the bidding round, generate the second bid for the second candidate agent based on the bidding function.

[0067] Step S244: In the comprehensive evaluation round, perform the following steps:

[0068] Step a: Obtain the capability score of the second candidate agent from the capability assessment module;

[0069] Step b: Calculate the reliability score based on the historical reputation provided by the capability assessment module and the real-time load reported by the professional intelligent agent execution layer;

[0070] Step c: Combine the task urgency level parsed by the task issuing layer with the agent's response capability output by the capability assessment module to calculate the emergency reward;

[0071] Step d: After converting the second bid into a bid score, input the capability score, reliability score, emergency reward and bid score into the dynamic weighted comprehensive evaluation function to calculate and generate the second comprehensive evaluation score.

[0072] Specifically, the pre-qualification rounds of the multi-round sealed auction are primarily based on four pre-screening checks on the second candidate agents: technical field matching check, basic capability threshold check, availability check, and permission check, in order to select agents that are qualified to participate in subsequent bidding. The process involves several key steps: First, a technology matching check compares the technology set of the second candidate agent with the technology set required by the current science and technology intelligence task. If there is no overlap, the agent is eliminated, ensuring a connection between the agent's domain and the task requirements. Second, a basic capability threshold check verifies the second candidate agent's skill level in the corresponding skill item, referencing the three-level skill items in a hierarchical skill tagging system. Agents failing to meet the minimum requirements cannot proceed to the next stage. Third, a usability check monitors the current load of the second candidate agent in real time to determine if it exceeds the system's set load threshold and operating status, confirming its "available" status and eliminating agents with excessive load or abnormal states, such as malfunctions or offline states, to prevent subsequent task execution from being hindered. Fourth, a permissions check compares the security level of the task data (e.g., the sensitivity level classification of science and technology intelligence data) with the security permission level of the second candidate agent, ensuring the agent has the necessary access permissions to process data at the corresponding security level. Agents without sufficient permissions are eliminated, guaranteeing data processing security and compliance.

[0073] In the multi-round sealed auction process, the connection between the pre-qualification round and the bidding round determines whether the second candidate agent can proceed to the next stage based on the pre-qualification results. If the second candidate agent successfully passes the following four checks during the pre-screening process: a technical field matching check (the agent's technical field set intersects with the task requirement field set); a basic capability threshold check (the agent's skill level meets the minimum requirements for the task); an availability check (the current load does not exceed the threshold and the status is "available"); and a permission check (the security permission level meets the task data security requirements), a pre-qualification pass notification is generated and sent to the second candidate agent. Simultaneously, the agent is added to the bidding candidate list, allowing it to enter subsequent bidding rounds. If the second candidate agent fails any of the above pre-screening checks, the specific reason for failure is provided, such as technical field mismatch or insufficient skill level, and it is removed from the candidate queue, no longer allowed to participate in subsequent stages of the auction. This ensures that all agents entering the bidding round possess the basic qualifications to participate in the scientific and technological intelligence analysis task, laying the foundation for the effectiveness of subsequent bidding and the reliability of task execution.

[0074] In the multi-round sealed auction bidding process, the core is that the second candidate agent generates a second bid based on the bidding function. This process is designed to fully integrate the professional and dynamic characteristics of scientific and technological intelligence tasks, specifically divided into two key stages: basic bid calculation and bid multiplier optimization. First, the basic bid is calculated, requiring the pre-calculation of six core coefficients and their substitution into the bidding function: the basic cost is calculated as task complexity × estimated execution time × unit cost; task complexity is determined with reference to data volume factors, algorithm complexity, and domain difficulty coefficients; the domain expertise coefficient is calculated as 0.5 + 0.5 × (agent's domain skill / 10), with a lower coefficient indicating a higher skill level in the task-related domain; the load penalty coefficient is calculated as 1 + (current load / maximum load). 2 The calculations are as follows: the higher the agent's current load, the larger the coefficient to reflect resource consumption costs; the reputation reward coefficient is calculated as 1 - 0.3 × (historical reputation - 0.5), with a lower coefficient indicating better historical reputation for the agent, thus providing bidding discounts; the data sensitivity coefficient is calculated as 1 + 0.5 × task security level / 10, with a larger coefficient indicating higher task data security levels, such as those involving core technological intelligence; the urgency coefficient is calculated as 1 + (task urgency level - 1) × 0.3, with a larger coefficient indicating higher task urgency levels to incentivize priority response. Multiplying these six coefficients together yields the base bid. Subsequently, the bidding multiplier is optimized. First, a state space vector is pre-constructed, which integrates task feature vectors (including one-hot encoding of the technical field, data volume, urgency level, etc.), agent state vectors (including current load, recent success rate, etc.), environmental context vectors (including total system load, number of competitors, etc.), and historical performance vectors (including recent task success rate, quality score trend, etc.). This state space vector is then input into a deep Q-network, which outputs the bidding multiplier selection result from 11 discrete options: [0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4, 1.5]. Finally, the second candidate agent calculates and generates a second bid by multiplying the base bid by the bidding multiplier selection result, ensuring that the bid aligns with its own capabilities and state, while also meeting the actual needs of the scientific intelligence task.

[0075] In the comprehensive evaluation round of a multi-round sealed auction, the second comprehensive evaluation score of the second candidate agent needs to be calculated according to a four-step process. Step a: The capability score of the second candidate agent is retrieved from the capability assessment module. This score is calculated as the average of three core capabilities: technical expertise, task matching degree, and data processing capability. Technical expertise is obtained by weighting the certification results of a hierarchical skill tagging system with domain weights. Task matching degree is calculated by the geometric mean of technical field, data type, urgency, and output format. Data processing capability is calculated by averaging processing speed, quality score, scale factor, and baseline performance. Step b: The reliability score is calculated by combining the historical reputation provided by the capability assessment module with the real-time load reported by the professional agent's execution layer. Historical reputation is obtained by weighted summation of recent performance in different domains and time decay coefficients. Real-time load is calculated by calculating the relative load based on the current task complexity and remaining time. Finally, the reliability score is obtained by multiplying the historical reputation by (1 minus the real-time load). In step c, the emergency reward is calculated by combining the task urgency level parsed by the task release layer with the agent's response capability output by the capability assessment module. The task urgency level is divided into 1-5 levels. The agent's response capability is assessed based on historical task response time and emergency task processing efficiency. The emergency reward is generated by multiplying the task urgency level by the agent's response capability and then dividing by 25. In step d, the second bid is first converted into a bid score, calculated using the formula (highest bid minus current bid) divided by (highest bid minus lowest bid), ensuring that the lower the bid, the higher the score. Then, the capability score, reliability score, emergency reward, and bid score are input into a dynamic weighted comprehensive evaluation function. The weights of each indicator in the function are adjusted according to the task characteristics: for non-urgent tasks, the bid score weight is increased to 0.3-0.5; for complex tasks, the capability score weight is increased to 0.4-0.5; for important tasks, the reliability score weight is increased to 0.2-0.3; and for urgent tasks, the emergency reward weight is increased to 0.1-0.2. The second comprehensive evaluation score is generated by weighted summation, providing a core decision-making basis for subsequent task allocation.

[0076] In one possible implementation, step S243 further includes:

[0077] Step S2431: After pre-calculating the basic cost, domain expertise coefficient, load penalty coefficient, reputation reward coefficient, data sensitivity coefficient, and urgency coefficient, the basic bid is calculated and output through the bidding function.

[0078] Step S2432: Preconstruct a state space vector, wherein the state space vector integrates the task feature vector, the agent state vector, the environmental context vector, and the historical performance vector.

[0079] Step S2433: Input the state space vector into the deep Q network and output the bidding multiple selection result.

[0080] Step S2434: Based on the selection results of the base bid and bid multiple, calculate and output the second bid.

[0081] Specifically, six core parameters need to be pre-calculated: basic cost, domain expertise coefficient, load penalty coefficient, reputation reward coefficient, data sensitivity coefficient, and urgency coefficient. Then, the basic bid is calculated and output via a bidding function. The basic cost is calculated by multiplying task complexity by estimated execution time and then by unit cost. Task complexity is determined by data volume factor, algorithm complexity, domain difficulty coefficient, and time constraint coefficient. Estimated execution time is estimated based on task scale and the agent's historical processing speed. Unit cost is the system's preset unit time cost for processing scientific and technological intelligence tasks. The domain expertise coefficient is calculated by adding 0.5 to the agent's domain skill and dividing by 10. The agent's domain skill is its level in the task-related technical field, ranging from 1 to 10. The load penalty coefficient is calculated by adding 1 to the current load and dividing by the maximum load. The load is calculated by multiplying the current load by the agent's real-time resource occupancy ratio, and the maximum load by the agent's resource capacity limit. The reputation reward coefficient is calculated by subtracting 0.3 from 1 and multiplying by the historical reputation coefficient minus 0.5. The historical reputation coefficient is weighted based on the agent's recent performance in its domain and a time decay coefficient. The data sensitivity coefficient is calculated by adding 0.5 to 1 and multiplying by the task security level by 10. The task security level is divided into 1 to 10 based on the sensitivity of the technological intelligence data. The urgency coefficient is calculated by adding 1 to the task urgency level coefficient minus 1 and multiplying by 0.3. The task urgency level is divided into 1 to 5. Multiplying these six coefficients together yields the base bid.

[0082] A pre-constructed state space vector is generated, integrating four main categories of information: task feature vector, agent state vector, environmental context vector, and historical performance vector. The task feature vector includes a technology area vector, standardized data volume, urgency, quality requirements, and estimated complexity. The technology area vector uses one-hot encoding to represent the technical field involved in the task. The standardized data volume is the normalized value of the task data size. The urgency ranges from 1 to 5, the quality requirements range from 1 to 5, and the estimated complexity ranges from 1 to 10. The agent state vector includes current load ratio, recent success rate, resource availability, skill matching degree, and historical bidding success rate. The current load ratio ranges from 0 to 1. The recent success rate is the percentage of recent task completions by the agent. Resource availability is the percentage of remaining resources available to the agent. Skill matching degree is the degree of matching between the agent's skills and task requirements. Historical bidding success rate is the percentage of successful bids won by the agent in past auctions. The environmental context vector includes system... The system's total load, number of competitors, time factor, task queue length, and average task complexity are defined as follows: total system load is the proportion of resources used by the entire system; number of competitors is the total number of agents participating in this auction; time factor is the ratio of the current time to the task deadline; task queue length is the number of tasks waiting to be processed in the system; and average task complexity is the average complexity of the tasks waiting to be processed in the system. The historical performance vector includes the success rate of the last 10 tasks, the average completion time ratio, the quality score trend, and collaboration efficiency. The success rate of the last 10 tasks is the percentage of an agent's successful tasks in the last 10 attempts; the average completion time ratio is the average ratio of the agent's actual task completion time to the estimated time; the quality score trend is the recent trend in the quality score of the agent's tasks; and collaboration efficiency is the efficiency value of the agent in collaborating with other agents to complete tasks.

[0083] The constructed state space vector is input into a deep Q-network, which outputs the bid multiplier selection result. The deep Q-network employs a four-layer fully connected architecture. The input layer receives a 25-dimensional state vector. The first hidden layer contains 128 neurons using ReLU activation, the second hidden layer contains 64 neurons using ReLU activation, the third hidden layer contains 32 neurons using ReLU activation, and the output layer contains 11 neurons corresponding to 11 bid multiplier options, using a linear activation function. The network training parameters are set as follows: learning rate 0.001, batch size 32, experience replay buffer size 10000, target network update frequency 100 steps, exploration rate decaying from 1.0 to 0.01 over 10000 steps, and discount factor 0.95. Through this network operation, the bidding multiple selection result corresponding to the second candidate agent is output from 11 discrete bidding multiple options: 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4, and 1.5.

[0084] Based on the obtained base bid and the obtained bid multiplier selection result, the second bid of the second candidate agent is output by multiplying the base bid by the bid multiplier selection result. This bid comprehensively reflects the agent's cost, capabilities, state and task characteristics, providing a key basis for subsequent comprehensive evaluation.

[0085] Example 2, as Figure 2 As shown, this application provides a multi-agent cooperative scheduling system for scientific and technological intelligence analysis, wherein the system includes:

[0086] Task publishing layer 10 is used to receive and convert users' scientific and technological intelligence analysis requirements into standardized task descriptions.

[0087] Specifically, the task publishing layer 10 is the core layer for interaction between the multi-agent collaborative scheduling system for science and technology intelligence analysis and users. It primarily undertakes the key functions of receiving, transforming, and providing feedback on user requests. On one hand, this layer receives various science and technology intelligence analysis requests submitted by users. These requests cover a variety of science and technology intelligence-related scenarios, including literature analysis, patent tracking, technology trend prediction, policy interpretation, and risk warning. Users can submit request details through the system's interactive interface, including but not limited to information such as target technology areas, data type preferences, task completion deadlines, and result quality requirements. On the other hand, the task publishing layer 10 performs a task requirement element parsing process on the received user requests, extracting key elements such as core objectives, data scope, time constraints, quality standards, and security levels. Following the system's preset unified format and semantic specifications, it transforms unstructured or semi-structured user requests into standardized task descriptions that are uniformly formatted, complete in information, and semantically clear. This ensures that the subsequent intelligent coordination and scheduling layer 20 can accurately understand the task requirements, providing a clear task basis for subsequent capability assessment, intelligent auction, and task allocation. At the same time, the task release layer 10 will also present the structured processing results generated by the professional intelligent agent execution layer 30 after completing the scientific and technological intelligence analysis to the user in a visual form, such as analysis reports, data charts, knowledge graphs, etc., so that the user can intuitively obtain and utilize the results of the scientific and technological intelligence analysis.

[0088] The intelligent coordination and scheduling layer 20 includes a capability assessment module, an intelligent auction module, a conflict resolution module, and a collaborative optimization module that are interconnected through a distributed event bus to form a dynamic decision-making closed loop.

[0089] Specifically, the intelligent coordination and scheduling layer 20 is the core decision-making layer for system task allocation and resource scheduling. Its core components are a capability assessment module, an intelligent auction module, a conflict resolution module, and a collaborative optimization module, interconnected via a distributed event bus. These four modules work together to form a dynamic decision-making closed loop, ensuring efficient and accurate allocation of scientific and technological intelligence analysis tasks. The capability assessment module retrieves the agent states of multiple candidate agents from the professional agent execution layer. Combined with task requirements extracted from standardized task descriptions, it evaluates the capabilities of candidate agents based on a six-dimensional capability assessment model covering technical expertise, data processing capabilities, load status, historical reputation, task matching degree, and collaborative capabilities, outputting a six-dimensional capability score to provide capability data support for subsequent decisions. The intelligent auction module receives the six-dimensional capability score output by the capability assessment module and executes a multi-round sealed auction process. First, a pre-qualification round checks candidate agents on their technical field matching, basic capability thresholds, availability, and permissions. Agents that pass the pre-qualification enter the bidding round, generating bids based on a bidding function. Finally, a comprehensive evaluation round calculates the capability score. The system uses a combination of six-dimensional capability scores, reliability scores, emergency reward scores, and bidding scores to ultimately output a comprehensive evaluation score. The conflict resolution module monitors resource conflicts in the system in real time, including data sources, computing, storage, and expert resources. If a conflict is detected, it combines the six-dimensional capability score with the comprehensive evaluation score, employing multiple strategies such as priority preemption, dynamic resource reallocation, and intelligent task decomposition to resolve the conflict and output a task allocation scheme. The collaborative optimization module, based on the task allocation scheme output by the conflict resolution module, leverages a pipeline collaborative model that divides scientific and technological intelligence processing into six stages: data acquisition, preprocessing, information extraction, knowledge mining, value assessment, and decision support. It optimizes the scheme through strategies such as intelligent preprocessing caching, incremental transmission optimization, quality feedback, load prediction, and dynamic adjustment, ultimately outputting task allocation instructions. The four modules achieve real-time information interaction and data synchronization through a distributed event bus. The scoring results of the capability assessment module are transmitted to the intelligent auction module. The bidding score of the intelligent auction module and the capability score together support the conflict resolution module in handling conflicts. The task allocation scheme of the conflict resolution module is input into the collaborative optimization module for optimization. The optimized task allocation instructions can then be fed back to each module for subsequent decision-making adjustments, forming a dynamic and efficient decision-making closed loop that adapts to the professional and complex needs of scientific and technological intelligence analysis tasks.

[0090] The professional intelligent agent execution layer 30 includes isolated and parallel data acquisition intelligent agent groups, data processing intelligent agent groups, knowledge mining intelligent agent groups, and decision support intelligent agent groups.

[0091] Specifically, the professional intelligent agent execution layer 30 is the core execution layer for scientific and technological intelligence analysis tasks. It adopts an isolated parallel architecture design, comprising four main categories of intelligent agents: data acquisition agents, data processing agents, knowledge mining agents, and decision support agents. Each agent group is both independent and isolated to avoid the impact of a single group failure on the overall process, and can collaborate in parallel to improve task processing efficiency, jointly completing the entire process of scientific and technological intelligence processing from data acquisition to decision support. Among them, the data acquisition agent group is responsible for, according to the task allocation instructions issued by the intelligent coordination and scheduling layer 20, using the standardized data interface of the data resource layer 40 to acquire data from multi-source heterogeneous scientific and technological intelligence data sources, such as IEEE Xplore, CNIPA patent database, government policy and regulation databases, and Web of... Data is collected from sources such as Science, and data format standardization and initial quality checks are performed. Standardized raw datasets and data quality reports are output, providing foundational data support for subsequent processing. The data processing intelligence group receives the standardized raw datasets from the data acquisition intelligence group and performs preprocessing operations such as text cleaning, deduplication, format conversion, and encoding standardization. This generates cleaned structured data and processing logs, eliminating data format differences and quality issues, laying the data foundation for in-depth mining. The knowledge mining intelligence group, based on structured data, uses entity recognition, relation extraction, event detection, and semantic annotation to extract structured information and construct entity relation graphs. In-depth mining is then conducted through pattern discovery, trend analysis, anomaly detection, and predictive modeling, outputting knowledge graphs, analysis reports, and prediction results to uncover the potential value and patterns in scientific and technological intelligence. The decision support intelligence group combines the output of the knowledge mining intelligence group with user needs analyzed at the task release layer 10 to perform multi-dimensional value assessment, importance ranking, and impact analysis, generating value assessment reports and priority rankings. This provides decision support services such as solution generation, team matching, resource planning, and risk assessment, outputting decision lists, implementation plans, and risk reports. Each intelligent agent within each intelligent agent group possesses independent perception, reasoning, learning, and execution capabilities, enabling it to autonomously complete corresponding sub-tasks. Furthermore, it achieves data transmission and collaboration with other intelligent agent groups through standardized interfaces, ensuring efficient and accurate advancement of the entire process of scientific and technological intelligence analysis. Ultimately, the structured processing results are fed back to the task release layer 10, where they are visualized and presented to the user.

[0092] The data resource layer 40 is used to manage multi-source heterogeneous scientific and technological intelligence data sources. The data resource layer provides standardized data interfaces for data access services to the data acquisition intelligent agent group, data processing intelligent agent group, knowledge mining intelligent agent group and decision support intelligent agent group through data access adapters.

[0093] Specifically, the data resource layer 40 is the underlying data support layer of the entire multi-agent collaborative scheduling system, primarily responsible for the unified management and standardized data service provision of multi-source heterogeneous scientific and technological intelligence data sources. The scientific and technological intelligence data sources managed by this layer cover various types, including academic literature databases, patent databases, policy and regulation databases, and knowledge graph databases. The academic literature databases encompass resources from platforms such as IEEE Xplore, ACM Digital, Scopus, and CNKI; the patent databases contain patent information from organizations such as CNIPA, USPTO, EPO, and WIPO; the policy and regulation databases include policy documents published by government websites, industry associations, and standards organizations; and the knowledge graph database stores constructed knowledge-related data in the scientific and technological field. Through centralized management of these multi-source heterogeneous data sources, the system ensures that it can acquire comprehensive scientific and technological intelligence data. Simultaneously, the data resource layer 40 deploys data access adapters to shield and uniformly convert the access protocols and data format differences of different data sources, constructing a standardized data interface. This standardized data interface provides data access services to the data acquisition intelligent agent group, data processing intelligent agent group, knowledge mining intelligent agent group, and decision support intelligent agent group in the professional intelligent agent execution layer 30. Regardless of the data source or original format of the data required by each intelligent agent group, it can efficiently and compliantly obtain data that meets its own processing needs through this standardized interface without having to adapt to the access rules of different data sources. This reduces the data acquisition complexity of each intelligent agent group and ensures the consistency and stability of data transmission, laying a data foundation for the smooth execution of subsequent scientific and technological intelligence analysis tasks.

[0094] The task publishing layer 10 visualizes the processing results generated by the professional intelligent agent execution layer 30 and presents them to the user.

[0095] Specifically, in addition to receiving user requests for science and technology intelligence analysis and converting them into standardized task descriptions, the task publishing layer 10 is also responsible for presenting the processing results generated by the professional intelligent agent execution layer 30 to users in a visual format. The data acquisition, data processing, knowledge mining, and decision support intelligent agent groups within the professional intelligent agent execution layer 30 will progressively complete data collection, preprocessing, deep mining, and decision support work according to task allocation instructions, ultimately generating processing results including structured data, knowledge graphs, analysis reports, prediction results, decision lists, and risk assessment reports. After receiving these processing results, the task publishing layer 10 will present them in an intuitive and easy-to-understand visualization method based on user needs and result types. For example, it may use data charts to display changes in technological trends, knowledge graphs to present technological relationships, structured reports to summarize core analysis conclusions, and checklists to list decision recommendations. This allows users to quickly and clearly understand the results of the science and technology intelligence analysis, intuitively obtain key information, and effectively support users in making subsequent decisions or carrying out work based on the analysis results.

[0096] The data resource layer 40, the professional intelligent agent execution layer 30, the intelligent coordination and scheduling layer 20, and the task publishing layer 10 are connected sequentially through standardized interfaces to form a loosely coupled distributed architecture.

[0097] Specifically, the data resource layer 40, the professional intelligent agent execution layer 30, the intelligent coordination and scheduling layer 20, and the task release layer 10 constitute the four-layer core architecture of the system. Each layer establishes communication connections with the others through standardized interfaces, forming a loosely coupled distributed architecture. The data resource layer 40 serves as the underlying data support, connecting upwards to the professional intelligent agent execution layer 30 via a standardized interface, providing unified data access services for various intelligent agent groups within this layer. The professional intelligent agent execution layer 30 interacts with the upper-level intelligent coordination and scheduling layer 20 through a standardized interface, receiving task allocation instructions from the scheduling layer and providing feedback on the intelligent agent status and processing results during task execution. The intelligent coordination and scheduling layer 20 communicates with the top-level task publishing layer 10 through a standardized interface, obtaining standardized task descriptions generated by the task publishing layer, and simultaneously synchronizing the relevant information of the final optimized task allocation instructions to the task publishing layer. Each layer relies on standardized interfaces to achieve efficient data and instruction transmission, with clear boundaries of responsibility and low dependency between layers. This ensures the independent operation and maintenance of a single layer while allowing for flexible expansion or adjustment of functional modules at a particular layer based on business needs, without requiring large-scale modifications to other layers. This fully adapts to the flexibility, scalability, and stability requirements of scientific and technological intelligence analysis tasks.

[0098] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0099] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0100] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A multi-agent cooperative scheduling method for scientific and technological intelligence analysis, characterized in that, The method includes: After receiving the technology intelligence analysis request submitted by the user, the task publishing layer outputs a standardized task description by parsing the task requirement elements. After receiving the standardized task description through a standardized interface, the intelligent coordination and scheduling layer performs closed-loop collaborative decision-making through the capability assessment module, intelligent auction module, conflict resolution module, and collaborative optimization module, and outputs task allocation instructions. After receiving the task allocation instructions issued by the intelligent coordination and scheduling layer through the standardized interface, the professional intelligent agent execution layer drives the data acquisition intelligent agent group to acquire scientific and technological intelligence data through the data access adapter of the data resource layer, and coordinates the data processing intelligent agent group, knowledge mining intelligent agent group and decision support intelligent agent group to perform scientific and technological intelligence analysis tasks and output structured processing results. The structured processing results are visualized and presented to the user through the task publishing layer. The intelligent coordination and scheduling layer receives the standardized task description through a standardized interface, and then performs closed-loop collaborative decision-making through the capability assessment module, intelligent auction module, conflict resolution module, and collaborative optimization module, outputting task allocation instructions, including: The capability assessment module performs six-dimensional capability assessment of multiple candidate agents based on a six-dimensional capability assessment model. The six-dimensional capability assessment model evaluates six capabilities, including technical expertise, data processing capability, load status, historical reputation, task matching degree, and collaboration capability. A hierarchical skill tagging system is pre-constructed, which includes first-level skill domains, second-level skill categories, and third-level skill items; Based on the multiple Level 3 skills declared by the first candidate agent, standard test tasks are distributed to the first candidate agent. Based on the first quality output of the first candidate agent in completing the standard test task, the skill level of the first candidate agent is evaluated based on the hierarchical skill tagging system to obtain the first skill level certification result. Based on the results of the first skill level certification, a technical expertise assessment will be conducted.

2. The multi-agent cooperative scheduling method for scientific and technological intelligence analysis as described in claim 1, characterized in that, After receiving the standardized task description through a standardized interface, the intelligent coordination and scheduling layer performs closed-loop collaborative decision-making via a capability assessment module, an intelligent auction module, a conflict resolution module, and a collaborative optimization module, and outputs a task allocation instruction. The method includes: The capability assessment module retrieves the states of multiple candidate agents from the professional agent execution layer. Extract task requirements from the standardized task description; The capability assessment module performs a six-dimensional capability assessment on the multiple candidate agents based on the task requirements and the states of multiple agents, and outputs multiple six-dimensional capability scores. The intelligent auction module executes multiple rounds of sealed auctions for the multiple candidate intelligent agents, and outputs multiple comprehensive evaluation scores. The multiple rounds of sealed auctions include a pre-qualification round, a bidding round, and a comprehensive evaluation round. After detecting resource conflicts, the conflict resolution module uses a multi-dimensional conflict resolution strategy to resolve the conflicts based on the multiple six-dimensional capability scores and multiple comprehensive evaluation scores, and outputs a task allocation scheme. The collaborative optimization module optimizes the task allocation scheme based on the pipeline collaborative model and outputs the task allocation instruction.

3. The multi-agent cooperative scheduling method for scientific and technological intelligence analysis as described in claim 2, characterized in that, The multi-faceted conflict resolution strategy includes a priority preemption strategy, a dynamic resource reallocation strategy, and a task intelligent decomposition strategy.

4. The multi-agent cooperative scheduling method for scientific and technological intelligence analysis as described in claim 2, characterized in that, The method involves executing multiple rounds of sealed-bid auctions for the multiple candidate agents through the intelligent auction module, and outputting multiple comprehensive evaluation scores. In the prequalification round, a pre-screening condition check is performed on the second candidate agent. The pre-screening condition check includes a technical field matching check, a basic capability threshold check, an availability check, and a permission check. If the preliminary review is approved, the bidding round will proceed. In the bidding round, a second bid is generated based on the bidding function for the second candidate agent; In the comprehensive evaluation round, the following steps are performed: Step a: Obtain the capability score of the second candidate agent from the capability assessment module; Step b: Calculate the reliability score based on the historical reputation provided by the capability assessment module and the real-time load reported by the professional intelligent agent execution layer; Step c: Combine the task urgency level parsed by the task issuing layer with the agent's response capability output by the capability assessment module to calculate the emergency reward; Step d: After converting the second bid into a bid score, input the capability score, reliability score, emergency reward and bid score into the dynamic weighted comprehensive evaluation function to calculate and generate the second comprehensive evaluation score.

5. The multi-agent cooperative scheduling method for scientific and technological intelligence analysis as described in claim 4, characterized in that, In the bidding round, a second bid is generated based on the bidding function to create a second bid for the second candidate agent, the method comprising: After pre-calculating the basic cost, domain expertise coefficient, load penalty coefficient, reputation reward coefficient, data sensitivity coefficient, and urgency coefficient, the basic bid is calculated and output through the bidding function. A pre-constructed state space vector is provided, wherein the state space vector integrates the task feature vector, the agent state vector, the environmental context vector, and the historical performance vector; The state space vector is input into a deep Q-network, which outputs the bidding multiplier selection result. Based on the selection results of the base bid and bid multiplier, the second bid is calculated and output.

6. A multi-agent cooperative scheduling system for scientific and technological intelligence analysis, characterized in that, The system is used to execute the multi-agent cooperative scheduling method for scientific and technological intelligence analysis as described in any one of claims 1 to 5, and the system includes: The task publishing layer is used to receive and convert users' scientific and technological intelligence analysis requirements into standardized task descriptions. The intelligent coordination and scheduling layer includes a capability assessment module, an intelligent auction module, a conflict resolution module, and a collaborative optimization module, which are interconnected through a distributed event bus to form a dynamic decision-making closed loop. The professional intelligent agent execution layer includes isolated and parallel data acquisition intelligent agent groups, data processing intelligent agent groups, knowledge mining intelligent agent groups, and decision support intelligent agent groups; The data resource layer is used to manage multi-source heterogeneous scientific and technological intelligence data sources. The data resource layer provides standardized data interfaces for data access services to the data acquisition intelligent agent group, data processing intelligent agent group, knowledge mining intelligent agent group and decision support intelligent agent group through data access adapters. The task publishing layer visualizes the processing results generated by the professional intelligent agent execution layer and presents them to the user. The data resource layer, professional intelligent agent execution layer, intelligent coordination and scheduling layer, and task release layer are connected sequentially through standardized interfaces to form a loosely coupled distributed architecture.

Images