A method and system for promoting achievement transformation based on precise docking of production, learning and research

Abstract

The application discloses a kind of based on industry-university-research precise docking and promotes achievement transformation method and system, it is related to industry-university-research technical field, including: the multi-modal data of industry-university-research main body is collected, and pretreatment is carried out, generates the dynamic knowledge graph of the association relationship of characterizing technical ability and resource;Potential cooperation scheme and corresponding resource demand vector are output based on atlas mining technology association path;In virtual resource pool, resource scheduling strategy is generated by multi-agent reinforcement learning model;Based on federal learning protocol, coordinate each participant resource data, execute cross-institution resource allocation;According to actual scheduling effect data, the associated weight of knowledge graph and reinforcement learning model parameter are synchronously corrected.The application realizes cross-institution resource optimization scheduling, dynamically adjusts knowledge graph, improves matching accuracy.Through feedback optimization strategy, adaptability and stability are enhanced, and combined with consistency index and version backtracking mechanism, improve system robustness and resource scheduling efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of industry-university-research cooperation, and specifically to a method and system for promoting achievement transformation based on precise connection between industry, academia and research. Background Art

[0002] In recent years, with the rapid development of information technology, especially the application of big data, artificial intelligence (AI), machine learning, and knowledge graphs, technology integration and resource optimization and scheduling between industry, academia, and research have gradually become a focus of attention. Data acquisition and processing technology has become an important means to promote the transformation of scientific research results. Academia and industry are constantly exploring how to promote the rapid application of scientific research results through technical means such as large-scale data analysis, machine learning, and knowledge graphs. In particular, significant progress has been made in the integration and application of multimodal data. Technologies such as natural language processing, image recognition, and sensor data fusion have been widely used in various fields. At the same time, reinforcement learning and federated learning, as the most cutting-edge technologies, have begun to be applied in areas such as resource scheduling and intelligent decision-making, providing strong technical support for the efficient allocation and collaborative cooperation of industry, academia, and research resources.

[0003] While existing technologies have made some progress in connecting industry, academia, and research resources, they still face numerous challenges, particularly regarding data diversity and complexity, optimizing resource scheduling strategies, and collaboration among diverse stakeholders. Significant technical bottlenecks remain. First, while existing knowledge graph construction methods can represent relationships between data, they are typically limited to generating static graphs and cannot dynamically adjust association weights and model parameters. This limits the adaptability and accuracy of knowledge graphs under changing resource and scheduling requirements. Second, existing resource scheduling methods often rely on traditional optimization algorithms, making it difficult to achieve flexible and real-time scheduling across institutions and multiple stakeholders. In particular, effective scheduling strategies and coordination mechanisms are lacking in multi-party collaboration scenarios. Finally, while multi-agent reinforcement learning models have been applied to resource scheduling in certain fields, most methods fail to effectively address multi-party data privacy protection and cross-institutional data collaboration. Existing reinforcement learning models often assume shared datasets, ignoring data privacy issues across institutions. Furthermore, in practical applications, they often lack real-time feedback mechanisms, making it difficult to optimize models and strategies in a timely manner. Consequently, existing technologies lack flexibility, real-time performance, and scalability when addressing multi-party collaboration, resource scheduling, and optimization, failing to meet practical needs. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problems solved by the present invention are: the existing resource scheduling and optimization methods are insufficient in real-time updating and correction of dynamic knowledge graphs, the resource scheduling strategies for collaboration among multiple subjects are not flexible and real-time, and there is a lack of effective privacy protection mechanisms and data sharing protocols during cross-institutional collaboration. In addition, the problems include how to improve resource utilization efficiency through intelligent algorithms in industry-university-research cooperation and promote the transformation of scientific research results quickly and effectively.

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

[0007] In a first aspect, an embodiment of the present invention provides a method for promoting the transformation of research results based on precise docking between industry, academia and research, including collecting multimodal data of industry, academia and research entities, and preprocessing it to generate a dynamic knowledge graph that characterizes the relationship between technical capabilities and resources;

[0008] Based on the associated paths of graph mining technology, potential cooperation plans and corresponding resource demand vectors are output;

[0009] In the virtual resource pool, a resource scheduling strategy is generated through a multi-agent reinforcement learning model;

[0010] Coordinate resource data of all participants based on the federated learning protocol and implement cross-institutional resource allocation;

[0011] Based on the actual scheduling effect data, the knowledge graph association weights and reinforcement learning model parameters are synchronously corrected.

[0012] As a preferred solution of the method for promoting achievement transformation based on precise docking between industry, academia and research institutes according to the present invention, the multimodal data collected from industry, academia and research institutes includes patent text data, experimental video data, equipment sensor data, scientific research paper data, technical report data and expert knowledge data;

[0013] The preprocessing includes performing spatiotemporal alignment on heterogeneous data sources to eliminate cross-device data delay deviation and establish a unified time reference coordinate system;

[0014] Use contrastive learning loss function to align vector spaces of different modalities;

[0015] When the preprocessed data is input into the dynamic knowledge graph, the domain knowledge graph is added for conflict detection.

[0016] As a preferred solution of the method for promoting the transformation of research results based on precise docking between industry, academia and research institutions described in the present invention, the generation of a dynamic knowledge graph representing the relationship between technical capabilities and resources includes multimodal entity alignment: cross-modal alignment of technical entities in patent texts, experimental videos and sensor data to generate entity embedding vectors in a unified semantic space;

[0017] Dynamically adjust the edge weights in the knowledge graph based on the evolution of association relationships using time series data analysis technology;

[0018] Load the inviolable axioms in the domain knowledge graph, perform logical verification on technical associations, and automatically cut off contradictory associations;

[0019] Map the resource pool state to virtual resource nodes and calculate the reachable path weight from technology nodes to resource nodes.

[0020] As a preferred solution of the method for promoting achievement transformation based on precise docking of industry, academia and research according to the present invention, the output of potential cooperation plans and corresponding resource demand vectors includes selecting a target technology node as a starting point in the graph, performing a constrained random walk based on dynamic edge weights, and generating a set of candidate technology association paths P = {p1, p2, ..., p n}, where each path p i The weight score S i Calculated as:

[0021] Among them, w uv For node u The dynamic edge weight to v, d(u, v) is the number of path hops, and γ is the decay factor;

[0022] Use the K-means clustering algorithm to group the path set and remove redundant paths with similarity higher than the threshold θ;

[0023] For each technical path p i , extract its associated virtual resource nodes and calculate the resource demand vector R i =[E i , F i , T i ],in:

[0024] E i For equipment demand, the idle rate and scheduling cost of the required equipment types in the path are counted, and the Weighted summation; where idle rate k is the availability rate of the k-th type of equipment, cost k is the economic cost of scheduling the k-th type of equipment, and m is the number of equipment types;

[0025] F i For capital demand, based on the regression prediction of historical conversion data, let F i =ɑ·Technical complexity + β·Market size; where ɑ and β are regression coefficients, optimized through historical data training;

[0026] T i Build a skills matrix for talent needs Calculate T i =||M·v skill||2, v skill is the path skill requirement vector; where ||·||2 is the L2 norm, which calculates the comprehensive matching degree between talents and requirements; is the real number field, used to represent the dimension of the matrix, d is the skill dimension, and s is the number of talents;

[0027] The resource requirement vector is input into the pre-trained feasibility classifier, and the feasibility level of the solution is output as Y∈{high, medium, low}. The classification rule is:

[0028]

[0029] in, is the feasibility prediction value, σ is the Sigmoid function, W and b are the classifier parameter matrix and bias term respectively;

[0030] Only solutions with high feasibility levels and resource-accessible path weights less than or equal to the budget threshold are retained to form the final output set.

[0031] As a preferred solution of the method for promoting the transformation of research results based on the precise docking of industry, academia and research according to the present invention, in which: in the constrained random walk, the attenuation factor The path hop count limit is d(u, v)≤5, and the dynamic edge weight update rule is:

[0032]

[0033] in, Represents the weight between node u and node v at time step t; represents the weight between node u and node v at time step t-1; Δw is the weight adjustment step, σ(·) is the Sigmoid function, s r Verify the success rate for historical collaborations.

[0034] As a preferred solution of the method for promoting achievement transformation based on precise docking between industry, academia and research according to the present invention, the training loss function of the feasibility classifier is weighted cross entropy, and the formula is:

[0035]

[0036] Among them, ω i is the weight, r i is the case benefit, r max is the maximum benefit, y i is the true label, The success probability predicted by the model.

[0037] As a preferred solution of the method for promoting the transformation of research results based on the precise docking of industry, academia and research according to the present invention, wherein: the resource scheduling strategy generated by the multi-agent reinforcement learning model includes establishing independent agents for different industry, academia and research subjects, and the agent state space s i Including the urgency of technical requirements τ i , resource availability η i and historical cooperation success rate ρ i ;

[0038] The agent's action space includes resource request amounts and scheduling priorities, and the decision-making goal is to maximize resource matching and utilization;

[0039] Calculating the agent's local reward

[0040]

[0041] in, is the local reward of the ith agent, m i is the technical matching degree, d i is the scheduling delay, u i is the resource utilization rate, ɑ1 and β1 represent the reward weight coefficients;

[0042] Calculate the global reward r global :

[0043]

[0044] Among them, φ i is the effective utilization coefficient of the i-th agent, and n is the number of agents;

[0045] A multi-agent deep deterministic policy gradient algorithm is used to optimize agent strategies, update based on policy gradients, and iteratively solve Nash equilibrium.

[0046] If multiple agents request the same resource, the reallocation protocol is triggered to reallocate resources according to the agent priority weights, and compensation rewards are given to agents that fail to successfully obtain resources;

[0047] Generate resource scheduling reports, including resource allocation cause chains, Nash equilibrium convergence trajectories, and resource conflict resolution records.

[0048] As a preferred solution of the method for promoting the transformation of achievements based on the precise docking of industry, academia and research according to the present invention, the synchronous correction of the knowledge graph association weight and the reinforcement learning model parameters includes deploying a local resource model for each participant. Initialize global model parameters Θ0;

[0049] Use hybrid encryption protocol to process resource data D i,satisfy:

[0050]

[0051] in, is the encrypted local data, HE(·) is homomorphic encryption, is the differential privacy noise; ∈ is the privacy budget of differential privacy, which is used to measure the strength of privacy protection;

[0052] Each participant updates parameters based on local data:

[0053]

[0054] in, is the parameter of the i-th intelligent agent at time t+1, is the parameter of the i-th agent at time t, and η is the learning rate; is the gradient of the loss function to the agent parameters, that is, according to the local data D i and current parameters Calculated gradients, which help adjust parameters to optimize the strategy;

[0055] Aggregate global parameters by data volume ratio:

[0056]

[0057] Among them, Θ t+1 are the parameters of the global model at time step t+1;

[0058] Output the global model Input multi-agent reinforcement learning model, the optimization goal is:

[0059]

[0060] Among them, R g is the global reward, λ is the strategy constraint coefficient; π is the strategy of the intelligent agent; is the expectation, that is, the expected value of the reward obtained by the computing agent under the execution strategy π; T is the number of time steps; ρ t is the power of the discount factor ρ at time step t, which controls the degree of decay of future rewards; KL(π old ||π) is the KL divergence between the old and new strategies, which measures the degree of change in strategy;

[0061] According to the actual distribution data D real Update the local model:

[0062]

[0063] The distribution deviation δ=||R i-R real ||2 Feedback to the federation framework;

[0064] in, are the updated global model parameters, is the global model parameter before updating, α2 is the update step size, Θ global is the global model parameter, R real This is the actual resource scheduling result.

[0065] As a preferred solution of the method for promoting the transformation of achievements based on the precise docking of industry, academia and research according to the present invention, wherein: the synchronous correction of the knowledge graph association weight and the reinforcement learning model parameters includes the acquisition technology verification success rate s v , commercial benefits c , resource utilization u r , construct the feedback matrix

[0066] Correct the knowledge graph weight and calculate the weight update amount:

[0067]

[0068] Where Δw uv is the weight update amount, η is the learning rate, is the loss function; σ(w uv ) is the Sigmoid function, which is used to convert w uv Converted to a probability value; Fw-y represents the difference between the predicted value Fw and the actual label y;

[0069] Constraint weight lower limit:

[0070]

[0071] in, is the updated weight at time step t+1, indicating the weight update process;

[0072] Policy gradient update:

[0073]

[0074] in, is the gradient of the objective function J with respect to the policy parameter θ, is the gradient of the secondary parameter θ, is the mathematical expectation operator, logπ θ (a||s) is the strategy π θ The logarithmic probability of choosing action a in state s; Q π (s,a) is the state-action value function, which represents the expected cumulative reward after executing action a in state s; KL(πold ||π)) is the KL divergence, which measures the old policy π old The difference from the new strategy π; α3 is the regularization coefficient;

[0075] Define consistency indicators:

[0076]

[0077] According to the calculation result of the consistency index C, if it is lower than the set standard, global retraining is triggered;

[0078] Among them, C is the consistency index; is an indicator function, the value is 1 when the condition is met, otherwise it is 0; N is the total number of paths, is the sign of the edge weight change of the i-th path in the knowledge graph, sign(Δθ i ) is the update direction symbol of the i-th parameter in the reinforcement learning model;

[0079] Generate a version hash snapshot:

[0080]

[0081] If 3 consecutive rounds v <0.6, roll back to the historical best version V t-k , k is the round offset of the historical best version;

[0082] Among them, V t is the snapshot hash value of version t, Hash(·) is the hash function, w (t) is the knowledge graph weight matrix after the tth round of iteration, θ (t) is the reinforcement learning model parameter after the tth round of iteration, is the concatenation operator.

[0083] In a second aspect, an embodiment of the present invention provides a system for promoting achievement transformation based on precise docking between industry, academia and research, including:

[0084] Data collection and preprocessing module: collects multimodal data from industry, academia, and research entities, performs preprocessing, and generates a dynamic knowledge graph that represents the relationship between technical capabilities and resources;

[0085] Graph mining module: Based on graph mining technology association paths, it outputs potential cooperation plans and corresponding resource demand vectors;

[0086] Multi-agent reinforcement learning module: Generates resource scheduling strategies in the virtual resource pool through a multi-agent reinforcement learning model;

[0087] Cross-institutional resource allocation module: Coordinates the resource data of each participant based on the federated learning protocol and performs cross-institutional resource allocation;

[0088] Model parameter correction module: Based on the actual scheduling effect data, the knowledge graph association weights and reinforcement learning model parameters are synchronously corrected.

[0089] The beneficial effects of this invention: By introducing a multi-agent reinforcement learning model and a real-time feedback mechanism, the present invention can dynamically adjust and correct the association weights in the knowledge graph based on actual scheduling effect data, thereby improving the adaptability and accuracy of the knowledge graph in a changing environment. This innovation significantly improves the accuracy and flexibility of the connection between industry, academia, and research resources.

[0090] This paper uses a multi-agent reinforcement learning (MARL) model for resource scheduling, generating intelligent resource allocation solutions based on the needs and resource availability of different agents. Compared with traditional static scheduling methods, this intelligent scheduling strategy can more efficiently and flexibly respond to dynamically changing resource demands, improving resource utilization.

[0091] By introducing a federated learning protocol, this paper effectively addresses the challenges of cross-institutional data sharing and privacy protection. Participants can collaboratively optimize the global resource allocation model without sharing local data, thereby enhancing the system's privacy protection capabilities and data security, ensuring that the interests of all parties are not infringed.

[0092] This invention promotes the rapid application and transformation of scientific research results through precise resource scheduling and optimized collaboration proposal recommendations. It can efficiently integrate resources from industry, academia, and research, promote the cross-border flow of technology and knowledge, and thus accelerate the practical application and commercialization of scientific research results. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which:

[0094] Figure 1 The first embodiment of the present invention provides an overall flow chart of a method for promoting the transformation of research results based on precise docking between industry, academia and research. DETAILED DESCRIPTION

[0095] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0096] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides a method for promoting the transformation of research results based on precise docking between industry, academia and research, including:

[0097] S1: Collect multimodal data from industry, academia, and research entities, and preprocess them to generate a dynamic knowledge graph that represents the relationship between technical capabilities and resources.

[0098] The multimodal data collected from industry, academia and research entities include patent text data, experimental video data, equipment sensor data, scientific research paper data, technical report data and expert knowledge data.

[0099] Patent text data: Contains information such as technical solutions, claims, background technology, inventors, and applicants, and is mainly used to characterize the distribution of technical capabilities, technological development trends, and innovations.

[0100] Experimental video data: records the experimental process, operation steps and key technical details, which can intuitively reflect the feasibility of scientific research results and experimental conditions.

[0101] Device sensor data: Data from IoT devices in laboratory or production environments, such as physical parameter data such as temperature, humidity, pressure, vibration, and power consumption, helps describe device status and resource availability.

[0102] Scientific research paper data: Contains academic research results, methodologies, experimental results, etc., which can be used to analyze scientific research trends, technology dependencies, and the evolution path of cutting-edge technologies.

[0103] Technical report data: usually provided by scientific research institutions, government projects, and corporate R&D departments, covering technical evaluation, feasibility analysis, market research, etc., and providing evaluation criteria for technology value and maturity.

[0104] Expert knowledge data: including expert review opinions, knowledge base of industry experts, technical classifications and technical reviews of patent examiners, etc., which can assist in technology correlation analysis and technology maturity assessment.

[0105] Preprocessing includes temporal and spatial alignment of heterogeneous data sources to eliminate cross-device data latency deviation and establish a unified time reference coordinate system;

[0106] Use contrastive learning loss function to align vector spaces of different modalities;

[0107] When inputting pre-processed data into the dynamic knowledge graph, the domain knowledge graph is added for conflict detection;

[0108] Generating a dynamic knowledge graph that represents the relationship between technical capabilities and resources includes: multimodal entity alignment: cross-modal alignment of technical entities in patent texts, experimental videos, and sensor data to generate entity embedding vectors in a unified semantic space;

[0109] Dynamically adjust the edge weights in the knowledge graph based on the evolution of association relationships using time series data analysis technology;

[0110] Load the inviolable axioms in the domain knowledge graph, perform logical verification on technical associations, and automatically cut off contradictory associations;

[0111] Map the resource pool state to virtual resource nodes and calculate the reachable path weight from technology nodes to resource nodes.

[0112] It should be noted that because multimodal data (such as text, images, and sensor data) are expressed differently in feature space, directly calculating the similarity between them can lead to large errors. This paper uses comparative learning to aggregate different modal data of the same technical entity in the embedding space, thereby reducing the inconsistency of cross-modal information.

[0113] In the context of industry-university-research collaboration, the relationship between technological capabilities and resources is not static. As research progresses and market demand changes, the impact of technology and the availability of resources will continue to evolve. Therefore, this paper introduces a time series data analysis method to dynamically adjust the edge weights between technological entities.

[0114] Since data from different sources may have logical conflicts, such as a technology being classified into multiple contradictory fields, or a resource being claimed by multiple entities at the same time, it is necessary to introduce a domain knowledge graph as a constraint to automatically cut off unreasonable technical associations.

[0115] In order to effectively manage and schedule industry-university-research resources in the knowledge graph, the present invention maps the actual resource status into virtual resource nodes and calculates the reachable path weight from technology nodes to resource nodes.

[0116] S2: Based on the graph mining technology association path, output potential cooperation plans and corresponding resource demand vectors.

[0117] Select the target technology node as the starting point in the graph, perform constrained random walk based on dynamic edge weights, and generate a set of candidate technology association paths P = {p1, p2, ..., p n}, where each path p i The weight score S i Calculated as:

[0118]

[0119] Among them, w uv is the dynamic edge weight from node u to v, d(u, v) is the number of path hops, and γ is the attenuation factor.

[0120] Furthermore, through grid search and cross-validation experiments, it is verified that when γ = 0.2, the path weight score S i The system is highly sensitive to path length (moderate attenuation of long paths prevents overfitting of local correlations). In industry-university-research scenarios, technology association paths that are too long (e.g., more than five hops) can easily introduce noise (e.g., cross-domain technology misassociation). Experimental data shows that limiting the number of hops improves path quality by 41%. For example, starting from "graphene preparation technology," within five hops, it can be linked to "lithium-ion battery negative electrode applications." However, if it exceeds five hops, it may be misassociate with unrelated fields (e.g., "food additives").

[0121] Use the K-means clustering algorithm to group the path set and remove redundant paths with similarity higher than the threshold θ;

[0122] For each technical path p i , extract its associated virtual resource nodes and calculate the resource demand vector R i =[E i ,F i ,T i ],in:

[0123] E i For equipment demand, the idle rate and scheduling cost of the required equipment types in the path are counted, and the Weighted summation; where idle rate k is the availability rate of the k-th type of equipment, cost k is the economic cost of scheduling the k-th type of equipment, and m is the number of equipment types.

[0124] F i For capital demand, based on the regression prediction of historical conversion data, let F i =α·Technical complexity+β·Market regulation; where α and β are regression coefficients, which are optimized through historical data training.

[0125] T i Build a skills matrix for talent needs Calculate T i =‖M·v skill ‖2, v skill is the path skill requirement vector; where ||·||2 is the L2 norm, which calculates the comprehensive matching degree between talents and requirements; is the real number field, used to represent the dimension of the matrix, d is the skill dimension, and s is the number of talents.

[0126] It should be noted that the skill matrix M is constructed using a hierarchical clustering method, merging skill items with a similarity higher than 0.9 (such as “Python programming” and “machine learning framework usage” into “AI development skills”) to reduce redundant dimensions.

[0127] The resource requirement vector is input into the pre-trained feasibility classifier, and the feasibility level of the solution is output as Y∈{high, medium, low}. The classification rule is:

[0128]

[0129] in, is the feasibility prediction value, σ is the Sigmoid function, W and b are the classifier parameter matrix and bias term respectively.

[0130] Only solutions with high feasibility levels and resource-accessible path weights less than or equal to the budget threshold are retained to form the final output set.

[0131] Furthermore, in the constrained random walk, the decay factor γ∈[0.1, 0.3], the path hop count is limited to d(u, v)≤5, and the dynamic edge weight update rule is:

[0132]

[0133] in, Represents the weight between node u and node v at time step t; represents the weight between node u and node v at time step t-1; Δw is the weight adjustment step, σ(·) is the Sigmoid function, s r Verify the success rate for historical collaborations.

[0134] Furthermore, the training loss function of the feasibility classifier is weighted cross entropy, and the formula is:

[0135]

[0136] Among them, ω i is the weight, r i is the case benefit, r max is the maximum benefit, y i is the true label, The success probability predicted by the model.

[0137] It should be noted that unlike traditional technology matching methods based on static data, this paper constructs a dynamic knowledge graph and utilizes multimodal data fusion technology (patent text, experimental videos, sensor data, etc.) to continuously update the relationship between technology capabilities and resources. Through multimodal entity alignment, data from different modalities is mapped into a unified semantic space, improving the accuracy of data fusion and enhancing the reliability of technology association analysis.

[0138] This paper uses multi-agent reinforcement learning, combined with a federated learning protocol, to train resource scheduling strategies within a virtual resource pool and dynamically optimize cross-institutional resource allocation. Multi-agent collaborative training enables resource scheduling strategies to optimize resource matching across different task scenarios, improving resource utilization and reducing conflicting allocations. The introduction of a federated learning protocol ensures that all participants can collaboratively train a global model without sharing original data, improving the system's applicability and data security.

[0139] Federated learning protocols are used to coordinate cross-institutional data and perform secure data collaborative computing, enabling resources from different organizations to be shared and optimally allocated without leaking data. Through secure multi-party computing and federated learning methods, participants can conduct joint modeling without exposing their data, improving the security and effectiveness of cross-institutional data collaboration.

[0140] S3: Generate resource scheduling strategies in the virtual resource pool through a multi-agent reinforcement learning model.

[0141] Establish independent intelligent agents for different industry-university-research entities, and the state space of the intelligent agent s i Including the urgency of technical requirements τ i , resource availability η i and historical cooperation success rate ρ i ;

[0142] The agent's action space includes resource request amounts and scheduling priorities, and the decision-making goal is to maximize resource matching and utilization;

[0143] Calculating the agent's local reward

[0144]

[0145] in, is the local reward of the ith agent, m i is the technical matching degree, d i is the scheduling delay, u i is the resource utilization rate; α1 and β1 represent the reward weight coefficients, which are used to adjust the importance of different factors in the reward calculation to ensure the fairness and rationality of the system;

[0146] Calculate the global reward rglobal :

[0147]

[0148] Among them, φ i is the effective utilization coefficient of the i-th agent, which is used to adjust each agent's contribution to the global reward to ensure a reasonable allocation of resources; n is the number of agents. This global reward mechanism can promote overall system optimization while also taking into account the local optimality of individual agents.

[0149] A multi-agent deep deterministic policy gradient algorithm is used to optimize agent strategies, update based on policy gradients, and iteratively solve Nash equilibrium.

[0150] If multiple agents request the same resource, the reallocation protocol is triggered to reallocate resources according to the agent priority weights, and compensation rewards are given to agents that fail to successfully obtain resources;

[0151] Generate resource scheduling reports, including resource allocation cause chains, Nash equilibrium convergence trajectories, and resource conflict resolution records.

[0152] It should be noted that the present invention adopts Multi-Agent Reinforcement Learning (MARL) to model the resource scheduling process, enabling the agent to autonomously learn the optimal scheduling strategy based on interactive experience rather than relying on manually set heuristic rules.

[0153] A competition and cooperation mechanism among agents is introduced, and conflict resolution protocols are used to ensure rational resource allocation, avoiding the negative impact of resource competition. Resources are allocated based on agent priority and weight, meaning that high-priority agents have a higher probability of obtaining resources, ensuring the timely execution of urgent tasks. A compensation mechanism is implemented for agents that fail to obtain resources. If an agent fails to obtain resources for multiple consecutive rounds, its scheduling weight is appropriately increased in subsequent scheduling, avoiding long-term uneven resource distribution.

[0154] The policy gradient method is used to optimize the agent's strategy, allowing the agent to continuously learn and improve its decision-making ability in complex environments.

[0155] After resource scheduling is completed, the scheduling status of each agent is recorded and key parameters (such as scheduling priority, scheduling time, resource utilization, etc.) are stored for subsequent analysis and optimization.

[0156] S4: Coordinate the resource data of each participant based on the federated learning protocol and implement cross-institutional resource allocation.

[0157] Deploy local resource models for each participant Initialize global model parameters Θ0;

[0158] Use hybrid encryption protocol to process resource data D i ,satisfy:

[0159]

[0160] in, is the encrypted local data, HE(·) is homomorphic encryption, is the differential privacy noise; ∈ is the privacy budget of differential privacy, which is used to measure the strength of privacy protection;

[0161] Each participant updates parameters based on local data:

[0162]

[0163] in, is the parameter of the i-th intelligent agent at time t+1, is the parameter of the i-th agent at time t, and η is the learning rate; is the gradient of the loss function to the agent parameters, that is, according to the local data D i and current parameters Calculated gradients, which help adjust parameters to optimize the strategy;

[0164] Aggregate global parameters by data volume ratio:

[0165]

[0166] Among them, Θ t+1 are the parameters of the global model at time step t+1;

[0167] Output the global model Input multi-agent reinforcement learning model, the optimization goal is:

[0168]

[0169] Among them, R g is the global reward, λ is the strategy constraint coefficient; π is the strategy of the intelligent agent; is the expectation, that is, the expected value of the reward obtained by the computing agent under the execution strategy π; T is the number of time steps; ρ t is the power of the discount factor ρ at time step t, which controls the degree of decay of future rewards; KL(π old ||π) is the KL divergence between the old and new strategies, which measures the degree of change in strategy;

[0170] According to the actual distribution data D real Update the local model:

[0171]

[0172] The distribution deviation δ = ‖R i -R real ‖2 is fed back to the federated framework; when δ>1.0, the federated model retraining is triggered to prevent cumulative errors.

[0173] in, are the updated global model parameters, is the global model parameter before updating, α2 is the update step size, Θ global is the global model parameter, R real This is the actual resource scheduling result.

[0174] It should be noted that combining homomorphic encryption (HE) and differential privacy (DP) for data protection enables participants to perform computations in a secure environment without directly sharing raw data. This addresses the security risks inherent in traditional federated learning data sharing, making cross-institutional data sharing and computation feasible. This approach is particularly suitable for sensitive data scenarios, such as those in healthcare and finance.

[0175] By using a weighted approach based on the proportion of data volume for parameter aggregation, rather than simple averaging, we prevent institutions with small data volumes from having a significant impact on the global model. This solves the model bias problem caused by data imbalance in traditional federated learning and improves the generalization ability of the global model.

[0176] By optimizing resource allocation through reinforcement learning (RL), intelligent agents can autonomously learn optimal strategies in a dynamic environment, rather than relying on fixed rules. This overcomes the limitations of traditional static rule-based scheduling methods and can adapt to the ever-changing resource supply and demand environment. Through reinforcement learning, agents can autonomously discover optimal scheduling strategies, improving resource utilization and allocation fairness.

[0177] KL divergence is used to constrain policy changes, avoiding training instability caused by overly rapid policy updates. This solves the policy oscillation problem that can exist in traditional reinforcement learning, making the system more stable during updates and reducing the occurrence of abnormal scheduling.

[0178] S5: Based on the actual scheduling effect data, the knowledge graph association weights and reinforcement learning model parameters are synchronously corrected.

[0179] Collection technology verification success rate v , commercial benefits c , resource utilization u r , construct the feedback matrix

[0180] Correct the knowledge graph weight and calculate the weight update amount:

[0181]

[0182] Where Δw uv is the weight update amount, η is the learning rate, is the loss function; σ(w uv ) is the Sigmoid function, which is used to convert w uv Converted into a probability value; Fw-y represents the difference between the predicted value Fw and the actual label y; this loss function ensures that the weight of the knowledge graph does not fluctuate too much, and at the same time optimizes the weight of the edge in combination with feedback data to improve the adaptability of the knowledge graph to actual resource scheduling conditions.

[0183] Constraint weight lower limit:

[0184]

[0185] in, is the updated weight at time step t+1, indicating the weight update process;

[0186] Policy gradient update:

[0187]

[0188] in, is the gradient of the objective function J with respect to the policy parameter θ, is the gradient of the secondary parameter θ, is the mathematical expectation operator, logπ θ (a‖s) is the strategy π θ The logarithmic probability of choosing action a in state s; Q π (s, a) is the state-action value function, which represents the expected cumulative reward after executing action a in state s; KL(π old ||π)) is the KL divergence, which measures the old policy π old The difference from the new strategy π; α3 is the regularization coefficient;

[0189] Define consistency indicators:

[0190]

[0191] According to the calculation result of the consistency index C, if it is lower than the set standard, global retraining is triggered;

[0192] Among them, C is the consistency index; is an indicator function, the value is 1 when the condition is met, otherwise it is 0; N is the total number of paths, is the sign of the edge weight change of the i-th path in the knowledge graph, sign(Δθ i) is the update direction sign of the i-th parameter in the reinforcement learning model. In a multi-agent system, the knowledge graph and the reinforcement learning model must maintain consistency, otherwise the optimization goal may deviate. Therefore, this metric is used to determine whether retraining is necessary to improve system stability.

[0193] Generate a version hash snapshot:

[0194]

[0195] If 3 consecutive rounds v <0.6, roll back to the historical best version V t-k , k is the round offset of the historical best version.

[0196] Among them, V t is the snapshot hash value of version t, Hash(·) is the hash function, w (t) is the knowledge graph weight matrix after the tth round of iteration, θ (t) is the reinforcement learning model parameter after the tth round of iteration, is a splicing operator; since parameters are constantly changing during training, this hash snapshot helps detect unstable oscillations and roll back when necessary to avoid parameter updates causing system performance degradation.

[0197] It should be noted that most existing knowledge graphs are static and cannot dynamically adjust weights during resource scheduling, resulting in path search results that may not match actual conditions. This invention updates weights based on feedback data, allowing the knowledge graph to adapt to scheduling conditions and improve decision-making accuracy.

[0198] During the training process of reinforcement learning models, the stability of policy updates is key. This paper uses KL divergence to constrain the range of policy changes, ensuring a smooth transition between the old and new policies and reducing performance fluctuations caused by drastic policy changes.

[0199] In a multi-agent environment, knowledge graph updates must be synchronized with reinforcement learning model updates; otherwise, resource scheduling strategies may fail. This paper monitors this synchronization using a consistency metric, C, and triggers global retraining when necessary. Furthermore, version hashing is used to store snapshots, allowing rollback to the optimal historical version when necessary to prevent error propagation.

[0200] Example 2 is the second embodiment of the present invention, which is different from the previous embodiment in that:

[0201] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the current technical solution, can be embodied in the form of a software product. The current computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0202] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0203] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0204] Example 3 is an embodiment of the present invention, which provides a system for promoting the transformation of research results based on the precise docking of industry, academia and research, including a data acquisition and preprocessing module, a graph mining module, a multi-agent reinforcement learning module, a cross-institutional resource allocation module and a model parameter correction module.

[0205] Data collection and preprocessing module: collects multimodal data from industry, academia, and research entities, performs preprocessing, and generates a dynamic knowledge graph that represents the relationship between technical capabilities and resources;

[0206] Graph mining module: Based on graph mining technology association paths, it outputs potential cooperation plans and corresponding resource demand vectors;

[0207] Multi-agent reinforcement learning module: Generates resource scheduling strategies in the virtual resource pool through a multi-agent reinforcement learning model;

[0208] Cross-institutional resource allocation module: Coordinates the resource data of each participant based on the federated learning protocol and performs cross-institutional resource allocation;

[0209] Model parameter correction module: Based on the actual scheduling effect data, the knowledge graph association weights and reinforcement learning model parameters are synchronously corrected.

[0210] Example 4 is an embodiment of the present invention, which provides a method for promoting the transformation of research results based on the precise docking of industry, academia and research. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation / comparative experiments.

[0211] This experiment aims to verify the effectiveness of a resource optimization and scheduling method based on precise integration between industry, academia, and research. The experiment uses multimodal data from research institutions, universities, and enterprises, including heterogeneous information sources such as patent texts, research papers, equipment sensor data, experimental videos, and technical reports, to construct a dynamic knowledge graph and optimize resource allocation strategies.

[0212] The experiment selected five universities, three research institutions, and five enterprises in Province A as data sources, collecting over 10,000 patents, over 5,000 papers, over 1,200 hours of experimental video data, over 2TB of sensor data, and over 3,000 technical reports from the past three years. To ensure data quality, the heterogeneous data was first spatiotemporally aligned, using a bidirectional LSTM neural network to eliminate cross-device latency deviations and establish a unified time base. Secondly, a contrastive learning loss function was used to align features of data from different modalities, generating embedding vectors for subsequent knowledge graph construction.

[0213] During the knowledge graph construction phase, multimodal entity alignment is first performed. For example, experimental data from a university is linked to corresponding papers and patent information to ensure data integrity. Then, time series data analysis methods are used to model the evolution of technology associations and dynamically adjust the edge weights of the knowledge graph. To enhance the logical consistency of the knowledge graph, constraint rules (such as technical standards and axioms) from the domain knowledge graph are loaded, and unreasonable technology association paths are automatically cut off. Available resources in the resource pool are mapped to virtual resource nodes to calculate the weight of the reachable path from technology nodes to resource nodes.

[0214] In the resource scheduling phase, the experiment used a multi-agent reinforcement learning (MARL) model for strategy optimization. Different industry-university-research entities were set up as agents, and each agent independently learned and made resource scheduling decisions based on its own resource needs, technical capabilities, and historical collaboration data. The experiment designed three different scheduling strategies for comparison:

[0215] Strategy 1 (traditional static rules): adopts fixed scheduling rules based on expert experience.

[0216] Strategy 2 (based on single-agent learning): only considers individual optimization and does not perform federated learning.

[0217] Strategy 3 (method of the present invention): Combine federated learning to coordinate resources and dynamically adjust the weight of the knowledge graph.

[0218] To ensure the security of cross-institutional resource data, the experiment employed federated learning (FL) to coordinate data from all participating parties and implement cross-institutional resource allocation. Data was processed using a hybrid encryption protocol (homomorphic encryption + differential privacy), and the parameters of each agent were aggregated using a federated averaging algorithm. The experiment ran 20 training iterations, with each round synchronously updating the knowledge graph edge weights and reinforcement learning parameters based on actual scheduling performance data to improve resource matching accuracy.

[0219] The experimental reference data is shown in Table 1.

[0220] Table 1 Experimental data records

[0221]

[0222] From the experimental data, it can be seen that the method of the present invention has obvious advantages in resource scheduling optimization, which is reflected in the following aspects:

[0223] (1) Resource matching and task completion rates have been significantly improved

[0224] In experiments, the resource demand matching rate increased from an initial 62.4% to 79.5%, and the task completion rate increased from 78.9% to 94.2%, demonstrating that the proposed method can significantly improve the accuracy of resource allocation. This is due to the dynamic adjustment of the knowledge graph, which makes the connection between technical capabilities and resources more precise.

[0225] (2) Optimize resource utilization and avoid waste

[0226] After implementing the proposed method, resource utilization increased from 65.3% to 81.2%. This demonstrates that optimizing resource scheduling strategies through reinforcement learning resulted in more efficient resource allocation and reduced idle resources. Furthermore, the collaborative optimization mechanism of federated learning ensures efficient cross-institutional resource sharing, avoiding redundant computation and duplicate resource allocation.

[0227] (3) Reduced computing costs and reduced expenses

[0228] Experimental data shows that computational cost was reduced from 145ms to 108ms, a savings rate of 16.4%. This demonstrates that the proposed method not only improves scheduling efficiency but also effectively reduces computational overhead. This is primarily attributed to the optimization of the reinforcement learning model, which enables the agent to converge to the optimal scheduling strategy more quickly.

[0229] (4) Improved scheduling success rate and enhanced system stability

[0230] Over six rounds of experiments, the scheduling success rate increased from 85.4% to 95.1%, demonstrating that this method effectively reduces the probability of scheduling failures. Traditional methods, lacking dynamic adjustments, are prone to matching failures. However, this method, through real-time correction of the knowledge graph and continuous optimization through reinforcement learning, enables the scheduling system to maintain efficient operation even in complex environments.

[0231] (5) Innovation compared with existing technologies

[0232] Traditional rule-based approaches (Strategy 1) suffer from static decision-making issues and struggle to adapt to dynamically changing resource demands. Single-agent reinforcement learning (Strategy 2) can optimize local resource allocation but lacks cross-organizational collaboration. This invention, by combining federated learning with knowledge graphs, enables all participants to share optimization experience while protecting data privacy, thereby improving overall scheduling efficiency and matching accuracy.

[0233] (6) Dynamic modification of knowledge graph enhances system adaptability

[0234] By continuously optimizing the edge weights in the knowledge graph, the system can dynamically respond to changes in technical relationships. This feature is particularly important in long-term resource scheduling, effectively improving the accuracy of resource matching and enhancing the system's adaptability and scalability.

[0235] This example demonstrates the advantages of a method for promoting the transformation of research results through precise industry-university-research integration in resource scheduling optimization. Experimental results show that compared to traditional static scheduling methods, this method significantly improves resource matching, task completion rate, and scheduling success rate, while reducing computational overhead and resource waste. This innovative method achieves more accurate and efficient cross-institutional resource scheduling through dynamic adjustment of knowledge graphs, reinforcement learning strategy optimization, and a federated learning collaborative mechanism, providing new technical support for collaborative innovation between industry, academia, and research.

[0236] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for promoting the transformation of research achievements based on precise docking between industry, academia and research, characterized in that: include: Collect multimodal data from industry, academia, and research entities, perform preprocessing, and generate a dynamic knowledge graph that represents the relationship between technical capabilities and resources; Based on the associated paths of graph mining technology, potential cooperation plans and corresponding resource demand vectors are output; In the virtual resource pool, a resource scheduling strategy is generated through a multi-agent reinforcement learning model; Coordinate resource data of all participants based on the federated learning protocol and implement cross-institutional resource allocation; Based on the actual scheduling effect data, the knowledge graph association weights and reinforcement learning model parameters are simultaneously modified; Select the target technology node in the graph as the starting point, perform constrained random walks based on dynamic edge weights, and generate a set of candidate technology association paths , where each path Weight score Calculated as: in, is the dynamic edge weight from node u to v, is the number of path hops, is the attenuation factor; Use K-means clustering algorithm to group the path set and remove the paths with similarity higher than the threshold. Redundant paths; For each technical path , extract its associated virtual resource nodes and calculate the resource demand vector ,in: For equipment demand, the idle rate and scheduling cost of the required equipment types in the path are counted, and the Weighted sum; where For the Availability of equipment of this type, For scheduling The economic cost of such equipment, is the number of device types; For capital demand, based on the regression prediction of historical conversion data, ;in, These are all regression coefficients, which are optimized through historical data training; Build a skills matrix for talent needs ;calculate , is the path skill requirement vector; where, is the L2 norm, which calculates the comprehensive matching degree between talents and needs; is the real number field, used to represent the dimension of the matrix, d is the skill dimension, and s is the number of talents; Input the resource requirement vector into the pre-trained feasibility classifier and output the feasibility level of the solution , the classification rules are: in, , is the feasibility prediction value, is the Sigmoid function, , are the classifier parameter matrix and bias term respectively; Only solutions with high feasibility levels and resource-accessible path weights less than or equal to the budget threshold are retained to form the final output set.

2. The method for promoting achievement transformation based on precise docking between industry, academia and research as claimed in claim 1, characterized in that: The multimodal data collected from industry, academia and research entities include patent text data, experimental video data, equipment sensor data, scientific research paper data, technical report data and expert knowledge data; The preprocessing includes performing spatiotemporal alignment on heterogeneous data sources to eliminate cross-device data delay deviation and establish a unified time reference coordinate system; Use contrastive learning loss function to align vector spaces of different modalities; When the preprocessed data is input into the dynamic knowledge graph, the domain knowledge graph is added for conflict detection.

3. The method for promoting achievement transformation based on precise docking between industry, academia and research as claimed in claim 2, characterized in that: The generation of a dynamic knowledge graph representing the relationship between technical capabilities and resources includes multimodal entity alignment: cross-modal alignment of technical entities in patent texts, experimental videos, and sensor data to generate entity embedding vectors in a unified semantic space; Dynamically adjust the edge weights in the knowledge graph based on the evolution of association relationships using time series data analysis technology; Load the inviolable axioms in the domain knowledge graph, perform logical verification on technical associations, and automatically cut off contradictory associations; Map the resource pool state to virtual resource nodes and calculate the reachable path weight from technology nodes to resource nodes.

4. The method for promoting achievement transformation based on precise docking between industry, academia and research as claimed in claim 3, characterized in that: In the constrained random walk, the attenuation factor , the path hop limit is , the dynamic edge weight update rule is: in, Represents the weight between node u and node v at time step t; Represents the weight between node u and node v at time step t-1; is the weight adjustment step size, is the Sigmoid function, Verify the success rate for historical collaborations.

5. The method for promoting achievement transformation based on precise docking between industry, academia and research as claimed in claim 4, characterized in that: The training loss function of the feasibility classifier is weighted cross entropy, and the formula is: in, is the weight, ; For case benefits, For maximum benefit, is the true label, The success probability predicted by the model.

6. The method for promoting achievement transformation based on precise docking between industry, academia and research as claimed in claim 5, characterized in that: The resource scheduling strategy generated by the multi-agent reinforcement learning model includes establishing independent agents for different industry-university-research entities, and the agent state space Including the urgency of technical requirements , resource availability and historical cooperation success rate ; The agent's action space includes resource request amounts and scheduling priorities, and the decision-making goal is to maximize resource matching and utilization; Calculating the agent's local reward : in, is the local reward of the ith agent, For technical matching, For scheduling delay, is the resource utilization rate, and Represents the reward weight coefficient; Calculating global rewards : in, is the effective utilization coefficient of the i-th agent, is the number of agents; A multi-agent deep deterministic policy gradient algorithm is used to optimize agent strategies, update based on policy gradients, and iteratively solve Nash equilibrium. If multiple agents request the same resource, the reallocation protocol is triggered to reallocate resources according to the agent priority weights, and compensation rewards are given to agents that fail to successfully obtain resources; Generate resource scheduling reports, including resource allocation cause chains, Nash equilibrium convergence trajectories, and resource conflict resolution records.

7. The method for promoting achievement transformation based on precise docking between industry, academia and research as claimed in claim 6, characterized in that: The synchronous modification of the knowledge graph association weights and reinforcement learning model parameters includes deploying a local resource model for each participant. , initialize the global model parameters ; Use hybrid encryption protocol to process resource data ,satisfy: in, For encrypted local data, For homomorphic encryption, is the differential privacy noise; The privacy budget of differential privacy is used to measure the strength of privacy protection; Each participant updates parameters based on local data: in, is the parameter of the i-th intelligent agent at time t+1, is the parameter of the ith intelligent agent at time t, is the learning rate; is the gradient of the loss function to the agent parameters, that is, according to the local data and current parameters Calculated gradients, which help adjust parameters to optimize the strategy; Aggregate global parameters by data volume ratio: in, are the parameters of the global model at time step t+1; Output the global model Input multi-agent reinforcement learning model, the optimization goal is: in, For global rewards, is the strategy constraint coefficient; strategies for intelligent entities; is the expectation, that is, the computing agent executes the strategy The expected value of the reward obtained; is the number of time steps; Discount factor The power at time step t controls the degree of decay of future rewards; is the KL divergence between the old and new strategies, which measures the degree of change in strategy; According to the actual distribution data Update the local model: Allocation bias Feedback into the federal framework; in, are the updated global model parameters, is the global model parameter before updating, is the update step size, is the global model parameter, This is the actual resource scheduling result.

8. The method for promoting achievement transformation based on precise docking between industry, academia and research as claimed in claim 7, characterized in that: The synchronous correction of the knowledge graph association weight and reinforcement learning model parameters includes the acquisition technology verification success rate , commercial benefits , resource utilization , construct the feedback matrix ; Correct the knowledge graph weight and calculate the weight update amount: in, is the weight update amount, is the learning rate, is the loss function; is the Sigmoid function, which is used to Convert to probability value; Represents the predicted value With the actual label the differences between; Constraint weight lower limit: in, is the updated weight at time step t+1, indicating the weight update process; Policy gradient update: in, is the objective function Strategy parameters The gradient, Secondary parameter The gradient, is the mathematical expectation operator, For strategy In state Select Action The logarithmic probability of is the state-action value function, which means that in the state Execute an action Expected cumulative rewards after is the KL divergence, which measures the old strategy With new strategies differences; is the regularization coefficient; Define consistency indicators: According to the consistency index If the calculation result is lower than the set standard, global retraining is triggered; in, To be consistent indicator; It is an indicator function, the value is 1 when the condition is met, otherwise it is 0; is the total number of paths, is the edge weight change symbol of the i-th path in the knowledge graph, is the update direction symbol of the i-th parameter in the reinforcement learning model; Generate a version hash snapshot: If 3 consecutive rounds , roll back to the best historical version , is the round offset of the historical best version; in, is the snapshot hash value of version t, is a hash function, is the knowledge graph weight matrix after the tth round of iteration, is the reinforcement learning model parameter after the tth round of iteration, is the concatenation operator.

9. A system for promoting the transformation of research achievements based on the precise docking of industry, academia and research, used to implement the method for promoting the transformation of research achievements based on the precise docking of industry, academia and research as described in any one of claims 1 to 8, characterized in that: include: Data collection and preprocessing module: collects multimodal data from industry, academia, and research entities, performs preprocessing, and generates a dynamic knowledge graph that represents the relationship between technical capabilities and resources; Graph mining module: Based on graph mining technology association paths, it outputs potential cooperation plans and corresponding resource demand vectors; Multi-agent reinforcement learning module: Generates resource scheduling strategies in the virtual resource pool through a multi-agent reinforcement learning model; Cross-institutional resource allocation module: Coordinates the resource data of each participant based on the federated learning protocol and performs cross-institutional resource allocation; Model parameter correction module: Based on the actual scheduling effect data, the knowledge graph association weights and reinforcement learning model parameters are synchronously corrected.

Images