Artificial intelligence-based scientific research project matching recommendation method and system

Abstract

The embodiment of the invention discloses a scientific research project matching recommendation method and system based on artificial intelligence. The method comprises the steps of obtaining a historical project demand data set of a scientific research project demand side and a historical resource supply data set of a scientific research resource side; performing demand feature extraction processing on the historical project demand data set to generate a demand project feature set; performing resource feature extraction processing on the historical resource supply data set to generate a resource supply feature set; calling a trained deep matching learning model, performing dynamic matching degree evaluation processing on the demand item feature set and the resource supply feature set, and generating a matching degree evaluation result between each demand item description text and the resource supply description text; and generating a scientific research project matching recommendation strategy according to a matching degree evaluation result, and synchronizing the scientific research project matching recommendation strategy to the corresponding scientific research project demand side and the scientific research resource side to trigger scientific research collaborative optimization operation.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of data mining and matching, and particularly to a method and system for matching and recommending scientific research projects based on artificial intelligence. Background Art

[0002] Current scientific research collaboration platforms generally adopt a resource matching mechanism based on keyword retrieval or tag classification, and associate project requirements with resource supply by combining a simple classification system manually annotated. Such technologies often rely on explicit conditions provided by the demand side (such as subject fields, budget ranges) and static attributes of the resource side (such as equipment types, personnel qualifications) for rule-based matching, lacking in-depth analysis of the internal complexity of requirements and the dynamic availability of resources. Moreover, most traditional matching models adopt shallow algorithms such as cosine similarity calculation, which are difficult to identify cross-dimensional correlation features hidden in text descriptions, resulting in matching results staying at the level of surface relevance. In addition, the related existing technologies do not consider the impact of the evolution of demand priorities and the fluctuation of resource status during the scientific research project cycle on the matching effect, often resulting in resource misallocation or outdated recommendations. Therefore, how to solve or partially solve the problems of resource misallocation and collaboration lag easily caused by traditional static matching technologies is a technical direction that needs to be overcome currently. Summary of the Invention

[0003] The embodiments of the present invention provide a method and system for matching and recommending scientific research projects based on artificial intelligence, which are used to solve or partially solve the problems of resource misallocation and collaboration lag easily caused by traditional static matching technologies.

[0004] In a first aspect, an embodiment of the present invention provides an artificial intelligence-based scientific research project matching and recommendation method, which is applied to a scientific research project matching and recommendation system. The method includes: obtaining a historical project requirement data set of scientific research project requesters and a historical resource supply data set of scientific research resource providers. The historical project requirement data set includes multiple requirement project description texts and corresponding requirement attribute tags, and the historical resource supply data set includes multiple resource supply description texts and corresponding resource attribute tags; performing requirement feature extraction processing on the historical project requirement data set to generate a requirement project feature set, where the requirement project feature set includes a requirement direction feature, a requirement complexity feature, and a requirement timeliness feature corresponding to each requirement project description text; performing resource feature extraction processing on the historical resource supply data set to generate a resource supply feature set, where the resource supply feature set includes a resource direction matching feature, a resource capability intensity feature, and a resource availability feature corresponding to each resource supply description text; calling a trained deep matching learning model to perform dynamic matching degree evaluation processing on the requirement project feature set and the resource supply feature set to generate a matching degree evaluation result between each requirement project description text and the resource supply description text; generating a scientific research project matching and recommendation strategy according to the matching degree evaluation result, and synchronizing the scientific research project matching and recommendation strategy to the corresponding scientific research project requesters and scientific research resource providers to trigger scientific research collaboration optimization operations.

[0005] In a second aspect, an embodiment of the present invention provides a scientific research project matching and recommendation system, including: a processor; a storage device on which a computer program is stored, when the computer program is executed by the processor, the processor implements any one of the artificial intelligence-based scientific research project matching and recommendation methods.

[0006] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the artificial intelligence-based scientific research project matching and recommendation method are implemented.

[0007] It can be seen that the embodiments of the present invention have the following beneficial effects: Through the multi-dimensional feature modeling and dynamic depth matching mechanism, the embodiments of the present invention achieve the precise adaptation of scientific research resources to project requirements. First, a three-dimensional requirement feature system covering direction characteristics, complexity levels, and timeliness indicators is constructed for historical project requirement data, breaking through the limitations of traditional single-label matching and being able to deeply deconstruct the implicit associations and priority relationships in requirement texts. At the same time, in the extraction of resource supply characteristics, a ternary evaluation dimension of direction matching degree quantification, ability strength grading, and availability dynamic tracking is innovatively integrated to realize the three-dimensional representation of resource attributes. Through the dynamic coupling analysis of the two-way feature set by a pre-trained depth matching model, the non-linear associations and dynamic evolution laws between requirements and resources can be captured, significantly improving the matching accuracy of cross-domain scientific research collaboration. Based on the real-time generated matching recommendation strategy, a two-way collaborative optimization mechanism can be triggered to promote the transformation of scientific research resource allocation from discrete response to intelligent scheduling, effectively solving the problems of resource misallocation and collaboration lag caused by traditional static matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a flowchart of a method for matching and recommending scientific research projects based on artificial intelligence provided by an embodiment of the present invention.

[0009] Figure 2 It is a schematic diagram of the basic structure of a scientific research project matching and recommending system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0010] To make the above objects, features, and advantages of the present invention more obvious and understandable, the embodiments of the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0011] See Figure 1 As shown, this figure is a flowchart of a method for matching and recommending scientific research projects based on artificial intelligence provided by an embodiment of the present invention, and this method can be applied to a scientific research project matching and recommending system. As Figure 1 shown, this method may include step 101-step 105.

[0012] Step 101: Obtain a historical project requirement data set of scientific research project requesters and a historical resource supply data set of scientific research resource providers. The historical project requirement data set includes multiple requirement project description texts and corresponding requirement attribute labels, and the historical resource supply data set includes multiple resource supply description texts and corresponding resource attribute labels.

[0013] In the embodiments of the present invention, the scientific research project requester refers to an entity or organization with the need to initiate a scientific research project, such as the scientific research management department of a university, the technology R & D center of an enterprise, or an industry research institution. The historical project requirement data set refers to a structured data set of the descriptive information of all scientific research projects submitted by the requester within a past time period and its associated attributes. The requirement project description text refers to the core content of the project requirements recorded in natural language, such as "Research on industrial defect detection algorithms based on deep learning" or "Application development of new composite materials in the aerospace field". The requirement attribute label is the metadata for structuring and annotating the requirement project description text, such as the technical direction labels annotated as "Intelligent manufacturing field", "Materials engineering application", "Algorithm optimization", and the project attribute labels such as "Key R & D topics", "School-enterprise cooperation projects", "Multi-year cycle".

[0014] In addition, the scientific research resource provider refers to an entity or team that provides scientific research capacity support, including university laboratories, independent research institutions, technical expert teams, or scientific research service platforms. The historical resource supply data set covers the scientific research resource information provided by the resource provider in the past and its attribute characteristics. The resource supply description text is such as "An interdisciplinary team with experience in computer vision algorithm optimization" or "Having an experimental platform for material synthesis and performance testing". The resource attribute labels include technical ability labels such as "Artificial intelligence direction", "Engineering materials field", "Senior research team", and availability labels such as "Projects that can be undertaken annually", "Technical service open".

[0015] Exemplarily, a certain industry scientific research foundation, as the requester, its historical project requirement data set contains information such as the technical solution summaries of the funded projects in the past five years and the annotated subject classification, research type, project cycle, etc.; the key laboratory of a comprehensive university, as the resource provider, its historical resource supply data set contains information such as the introduction of scientific research projects participated by the laboratory in recent years, the equipment list, and the annotated technical fields, core expertise, service time periods, etc. The data acquisition process is realized by docking the project management system of the requester and the resource management platform of the resource provider, and adopting a combination of data interface to retrieve structured data and text parsing to ensure that the data format is standardized and complies with data security specifications.

[0016] Step 102: Perform requirement feature extraction processing on the historical project requirement data set to generate a requirement project feature set, where the requirement project feature set contains the requirement direction feature, requirement complexity feature, and requirement timeliness feature corresponding to each requirement project description text.

[0017] Among them, the requirement feature extraction process refers to the process of semantic parsing and feature representation of unstructured text using natural language processing techniques. The requirement direction feature represents the technical field and application direction of the project, and is realized through domain term recognition and semantic network analysis. For example, for the project description of "Research on Quantum Encryption Communication Protocol", after semantic model parsing, three direction feature vectors of "information security", "quantum technology", and "communication protocol" are generated. The requirement complexity feature reflects the difficulty of project technical implementation, and an evaluation model is constructed based on indicators such as the hierarchical structure of the technical solution, the depth of interdisciplinary intersection, and the number of key technical nodes.

[0018] Specifically, the semantic role labeling technology is used to identify the technical implementation path in the project description, and the complexity weight is calculated in combination with the knowledge graph. For example, a project involving multiple-stage technical processes such as "development of new catalysts - reaction mechanism modeling - process parameter optimization" has a higher complexity feature value than a project with a single experimental verification. The requirement timeliness feature represents the time urgency of the project and is extracted through time element recognition and cycle analysis model. For example, for projects containing keywords such as "urgent technical research" and "completed within a quarter", high timeliness features are extracted; while for projects such as "long-term theoretical research", low timeliness features are generated.

[0019] In specific implementation, in the requirement project feature set of a scientific research management institution, after feature extraction of the "new energy power system R & D" project, the direction features are mapped to three dimensions of "energy technology", "power engineering", and "system control". The complexity feature is evaluated as a higher level based on the degree of interdisciplinary technology integration, and the timeliness feature is marked as a medium level due to the project cycle. The feature extraction process adopts a combination of pre-trained language models and domain knowledge bases, and optimizes the semantic representation space through feature fusion algorithms to ensure the accurate parsing of technical terms.

[0020] Step 103: Perform resource feature extraction processing on the historical resource supply data set to generate a resource supply feature set, where the resource supply feature set includes resource direction matching features, resource ability intensity features, and resource availability features corresponding to each resource supply description text.

[0021] Exemplarily, the resource direction matching feature represents the technical fit between resource supply and scientific research requirements, and is realized through semantic similarity calculation and domain classification matching. For example, for the resource description of a research team "biomedical data analysis and algorithm development", after topic model analysis, three direction matching features of "medical and health", "data science", and "algorithm engineering" are generated, forming a high match with the "intelligent medical image diagnosis" project of the demand side. The resource ability intensity feature quantitatively evaluates the professional technical level and service ability of the resource, and constructs a comprehensive evaluation system including team experience accumulation, equipment advancement, and result quality.

[0022] For example, the resource capacity intensity characteristics of a certain key laboratory are obtained by a multi-dimensional evaluation algorithm with a high intensity value based on data such as the academic qualifications of its core members, the advanced instrument and equipment equipped, and the large-scale scientific research projects completed. The resource availability characteristics reflect the time schedulability of resources and the service carrying capacity, and are extracted through service cycle analysis and load status modeling. For example, an analysis and testing center marked as "able to undertake new projects in the next stage" and with the current service schedule in an expandable state has a higher availability characteristic value than an organization with a service close to saturation.

[0023] In a specific case, after the resource supply description of an intelligent materials team in a university is feature-extracted, the direction matching features are mapped to the dimensions of "intelligent materials", "sensing technology", and "performance optimization". The capacity intensity characteristics are evaluated as a higher level based on the team's research accumulation and technical equipment conditions. The availability characteristics are marked as available because the current project acceptance volume is within a reasonable range. The feature extraction process integrates a context-aware neural network and a domain ontology model, effectively identifying implicit information in the resource description. Expressions such as "leading the formulation of industry standards" are automatically associated with the technical authority characteristics.

[0024] Step 104: Invoke the trained deep matching learning model to perform a dynamic matching degree evaluation process on the demand project feature set and the resource supply feature set, and generate a matching degree evaluation result between each demand project description text and the resource supply description text.

[0025] In the embodiment of the present invention, the deep matching learning model adopts a dual-encoder architecture, including a demand feature encoding module and a resource feature encoding module. The demand feature encoding module performs multi-feature fusion on the demand direction feature, complexity feature, and timeliness feature, and adjusts the importance of different feature dimensions through a dynamic weight allocation mechanism. For example, for a demand project with an urgent timeliness feature, the model automatically enhances the matching weight of the resource availability feature. The resource feature encoding module uses a graph attention network to process the correlation relationship between the resource direction matching feature, capacity intensity feature, and availability feature, and constructs a resource capacity relationship graph. The dynamic matching degree evaluation calculates the semantic space similarity of two encoded vectors and combines a time sensitivity function to handle the change of resource availability.

[0026] In a specific application, the matching degree evaluation result of a demand project of an enterprise "development of an autonomous driving perception system" (direction features: autonomous driving, sensor fusion, real-time computing; complexity: high; timeliness: urgent) and a resource of a scientific research team "research on multi-modal perception algorithms" (direction matching: highly compatible; capacity intensity: high; availability: idle in the next stage) is at a higher level, while the matching degree evaluation with the resource of "optimization of traditional control algorithms" is at a lower level.

[0027] Optionally, the model training adopts a triplet contrastive learning method, uses historical successful matching cases to construct positive and negative sample pairs, and optimizes the discrimination of the feature space. The dynamic nature is reflected in the model's real-time response to resource status updates. For example, when the maintenance plan of a laboratory device causes a change in service availability, the system automatically triggers the recalculation of the matching degree. The evaluation results are presented in a hierarchical quantification form, and a multi-dimensional matching analysis report is generated, detailing sub-dimensions such as the fit of technical directions, the degree of ability adaptation, and the status of time coordination.

[0028] Specifically, in the process of matching evaluation between scientific research projects and resources, an independent one-to-one matching degree calculation is performed between each requirement project description text and each resource supply description text to form a complete matching matrix. Specifically, it is assumed that the historical project requirement data set contains N requirement project description texts (such as N projects like "Research and Development of New Energy Power Systems" and "Research on Quantum Encryption Communication Protocols" of a certain industry scientific research foundation), and the historical resource supply data set contains M resource supply description texts (such as M resources like "Research Team on Multimodal Perception Algorithms" and "Material Synthesis Experimental Platform" of a certain university laboratory). Then, the deep matching learning model traverses and calculates all combinations, and finally generates N×M matching degree evaluation results. Each evaluation result accurately represents the dynamic matching degree of a single requirement project and a single resource supply in multi-dimensional features such as the fit of technical directions, ability adaptability, and time coordination.

[0029] For example, the project of "Development of Autonomous Driving Perception System" (one of the N requirements) of the demand side will be matched one by one with all M resource supply description texts of the resource side: a high matching degree is calculated with the "Research Team on Multimodal Perception Algorithms" (high overlap of direction features, ability strength meeting the requirements of complex algorithm development, resource availability coordinated with the project's urgent cycle), while a low matching degree is generated with the "Optimization of Traditional Control Algorithms" team (only partial association of direction features, technical ability not covering real-time computing requirements); similarly, the other N - 1 requirement projects are all independently matched with all M resource supplies, and finally the system outputs N×M quantitative evaluation values. The matching results are stored in a hierarchical matrix form, supporting the demand side to screen the best resource combination according to priority, or the resource side to query the matching projects in reverse, so as to achieve the global optimal supply-demand docking in the scientific research collaboration platform.

[0030] Step 105: Generate a scientific research project matching recommendation strategy according to the matching degree evaluation result, and synchronize the scientific research project matching recommendation strategy to the corresponding scientific research project demand side and scientific research resource side to trigger scientific research collaboration optimization operations.

[0031] In this step, the matching recommendation strategy generates a multi-dimensional sorting algorithm. On the basis of ensuring the core matching conditions, it balances the priorities of the demand side and the service characteristics of the resource side. For a research and development project in a key field, the system generates a gradient recommendation plan: first, recommend the research institutions with the highest matching degree and outstanding ability intensity; second, recommend the technical teams with good matching degree and fast service response; finally, recommend the collaborative units with qualified matching degree and better cost-effectiveness. The strategy synchronization is realized through the intelligent push system of the scientific research collaboration platform. The technical ability matrix and the successful case library of the recommended resources are displayed on the demand side interface, while the technical summary and the cooperation process description of the matching requirements are received on the resource side interface.

[0032] Specifically, the triggering of the scientific research collaboration optimization operation is specifically manifested as follows: the demand side initiates a technology docking invitation through the platform, and the resource side gives an online feedback on the intention to undertake. The two parties enter the process of demonstrating the technical solution and signing the cooperation agreement. For example, after a new material research and development demand project matches the recommended resource side, the demand side comprehensively compares the maturity of the technical solutions and the cooperation models of each resource, and selects the optimal partner to start joint research and development. The system also provides a collaboration process supervision module to track the whole process of elements such as project technical nodes, resource input progress, and output quality of achievements. The recommendation strategy also includes an adaptive adjustment mechanism. When the service status of the recommended resources changes, it automatically triggers the recommendation of alternative solutions and pushes update notifications.

[0033] In one implementation, the extracting of the demand characteristics from the historical project demand data set in step 102 to generate a demand project feature set includes: Step 1021: Perform semantic segmentation processing on the demand project description text in the historical project demand data set to obtain multiple demand semantic segments.

[0034] Specifically, the semantic segmentation processing uses a deep learning model based on natural language understanding to divide the demand project description text into semantic units. For example, after processing the demand project description text of a "Quantum Encryption Communication Protocol Research" project of a scientific research foundation, it is divided into three demand semantic segments: "Optimization of Quantum Key Distribution Mechanism", "Design of Anti-Quantum Computing Attack Algorithm", and "Verification of Communication Protocol Security", and each segment corresponds to a specific technical implementation module. During the semantic segmentation processing, the model identifies the semantic boundaries based on grammatical markers such as punctuation marks, conjunctions, and transition words, and at the same time combines the scientific and technical term dictionaries stored in the domain knowledge base to accurately locate the core elements in the technical solution description. For complex descriptions containing compound sentences, such as "Develop a material property prediction model based on a deep neural network and integrate an experimental data feedback mechanism", the model uses dependency syntactic analysis to identify the subject-predicate-object structure and divides it into two demand semantic segments: "Development of a Deep Neural Network Material Property Prediction Model" and "Integration of an Experimental Data Feedback Mechanism", ensuring that each segment has an independent technical connotation.

[0035] Step 1022: Invoke the pre-trained requirement feature encoder to perform direction recognition processing on the multiple requirement semantic segments, and generate requirement direction features corresponding to each requirement item description text.

[0036] Among them, the requirement feature encoder uses a pre-trained language model based on the Transformer architecture, and captures the correlation relationships between semantic segments through the multi-head attention mechanism. For example, for the three requirement semantic segments of the project "Research and Development of Intelligent Warehouse Robot Path Planning System" of an enterprise, namely "Optimization of Dynamic Environment Perception Algorithm", "Design of Multi-Robot Cooperative Scheduling Strategy", and "Research on Energy Consumption Efficiency Balancing Mechanism", after being processed by the encoder, the three direction feature vectors of "Machine Vision", "Operations Research Optimization", and "Energy Management" are respectively extracted, and the overall requirement direction feature is generated through feature fusion. During the direction recognition processing, the encoder maps each semantic segment to a preset scientific research field direction space, which contains 128 standard direction dimensions such as artificial intelligence, materials engineering, and biomedicine, and determines the belonging field by calculating semantic similarity. For interdisciplinary projects, such as "Construction of a Financial Risk Prediction Model Based on Bio-Inspired Algorithms", the encoder identifies the dual-direction features of "Computational Biology" and "Financial Engineering", and assigns feature intensity values based on segment weights.

[0037] Step 1023: Perform complexity quantification processing on the requirement attribute labels in the historical project requirement data set, and generate requirement complexity features corresponding to each requirement item description text.

[0038] Exemplarily, the complexity quantification processing uses the analytic hierarchy process to construct an evaluation index system, which includes core dimensions such as the number of technical levels, the degree of interdisciplinary intersection, and the number of key technical nodes. For example, for a project "Research and Development of Thermal Protection System for Aerospace Vehicles" marked with three requirement attribute labels of "Multi-Physical Field Coupling Analysis", "Extreme Environment Experiment Verification", and "Material-Structure Collaborative Optimization", the complexity scores of 8.7, 9.2, and 8.5 (full score of 10) are obtained respectively after quantification processing, and the requirement complexity feature value of 8.8 is generated by weighted average. During the processing, the system invokes the domain knowledge graph analysis technology to implement the technical path, analyzes the three interdisciplinary features of fluid mechanics, thermodynamics, and materials science for the label of "Multi-Physical Field Coupling Analysis", and improves the complexity score based on the preset discipline span weight coefficient. For the attribute label marked as "Key Technology Research", the system automatically associates a high complexity level, while the label of "Mature Technology Integration" corresponds to a lower score.

[0039] Step 1024: Perform timeliness analysis processing on the requirement attribute labels in the historical project requirement data set, and generate requirement timeliness features corresponding to each requirement item description text.

[0040] In this step, the timeliness parsing process identifies the time constraint conditions in the project cycle annotation through a time element extraction model. For example, for the requirement attribute label of "two-year cycle, acceptance in three phases" marked in the "Smart City Traffic Control Platform Construction" project of a local government, after parsing, a timeliness feature vector containing stage time nodes [2024Q2, 2024Q4, 2025Q2] is generated. During the processing, the system establishes a mapping rule from the time expression to the standard time axis, converting "urgent research" into a 3-month cycle and "medium- and long-term research" into a 36-month cycle. For compound time constraints, such as "complete prototype development in 6 months in the first stage and iterate and optimize in the subsequent 12 months", the parsing model generates multi-dimensional timeliness features through time interval segmentation. The system also considers the influence of the project priority label. For projects marked as "key projects", the time sensitivity weight is automatically increased by 20%.

[0041] Step 1025: Perform feature fusion processing on the demand direction feature, the demand complexity feature, and the demand timeliness feature to generate the demand project feature set.

[0042] In the actual application process, the feature fusion processing adopts a combination of cascaded splicing and attention weighting to construct a 512-dimensional unified feature vector. For example, the demand direction feature (new energy materials, electrochemical engineering), the demand complexity feature (8.5 / 10), and the demand timeliness feature (24-month cycle) of a project on "development of new fuel cell catalysts" in a certain university are fused to form a comprehensive feature representation including material synthesis methods, electrochemical test indicators, and time schedule nodes. During the fusion process, the direction feature dynamically adjusts the dimension weights through the multi-head attention mechanism, and the complexity feature and the timeliness feature are normalized and then spliced with the direction feature respectively. The system sets a feature verification mechanism. When a logical conflict is detected between the direction feature and the complexity feature (such as a high complexity score matching the "basic theory research" direction), the manual review process is triggered to ensure data consistency.

[0043] In one implementation, the resource feature extraction process for the historical resource supply data set in step 103 to generate the resource supply feature set includes: Step 1031: Perform resource type classification processing on the resource supply description text in the historical resource supply data set to obtain multiple resource type identifiers.

[0044] In this step, the classification process adopts a multi-level classification system that combines a rule engine and machine learning. The first level is divided into major categories such as experimental equipment, technical teams, and computing resources, and the second level is refined to specific professional fields. For example, after the classification process of the resource description of a "high-energy physics experimental data computing cluster" in a certain laboratory, a three-level type identifier of "computing resources - high-performance computing - physical simulation" is obtained. During the processing, the system extracts the core noun "data computing cluster" through named entity recognition and determines the professional field attribution in combination with the modifier "high-energy physics experiment". For complex resource descriptions, such as an "interdisciplinary team with material characterization and biocompatibility testing capabilities", the system generates double type identifiers of "experimental equipment - material characterization" and "technical team - biomedical engineering" at the same time and establishes an association relationship graph.

[0045] Step 1032: Perform a direction matching degree analysis process on the resource supply description text based on the multiple resource type identifiers, and generate a resource direction matching feature corresponding to each resource supply description text.

[0046] For another example, the analysis process adopts a semantic similarity calculation method based on ontology to map the resource type to the standard scientific research direction space. For example, after the analysis of the resource type identifier "technical service - automatic control" of the "industrial robot motion control algorithm optimization service" in a certain enterprise research institute, a matching degree feature vector [0.92, 0.88, 0.75] with the three directions of "intelligent manufacturing", "robot technology", and "control theory" is generated. During the processing, the system constructs an association matrix between the resource type and the scientific research direction and determines the matching degree by calculating the cosine similarity. For cross-type resources, such as "providing both gene sequencing experiments and bioinformatics analysis services" at the same time, the system calculates the direction matching degree corresponding to each type respectively, and then takes the maximum value as the final feature value to ensure that the resource ability boundary is fully reflected.

[0047] Step 1033: Perform an ability strength evaluation process on the resource attribute labels in the historical resource supply data set, and generate a resource ability strength feature corresponding to each resource supply description text.

[0048] In this embodiment, the evaluation process establishes an evaluation model including dimensions such as equipment advancement, team qualifications, and achievement quality. For example, the resource attribute labels of a certain key laboratory are marked as "equipped with the third-generation synchrotron radiation light source", "the team includes 3 academicians", and "5 papers have been published in Nature in the past five years". After evaluation, an ability strength feature value of 9.3 / 10 is obtained. During the processing, the system calls the equipment parameter database to verify the technical indicators of the "third-generation synchrotron radiation light source", combines the academic influence model of academicians and the journal impact factor data, and obtains a comprehensive score through weighted calculation. For service-type resources, such as "the annual sample processing volume exceeds 10,000 pieces", the system converts it into an equipment throughput index and incorporates it into the evaluation system.

[0049] Step 1034: Perform availability cycle analysis on the resource attribute tags in the historical resource supply data set to generate resource availability features corresponding to each resource supply description text.

[0050] Exemplarily, the parsing process extracts the available time window through time expression recognition and scheduling analysis algorithm. For example, the resource attribute labels of a certain analysis and testing center are marked with "equipment maintenance period in Q2, 2024" and "the current undertaking volume reaches 70% of the capacity". After parsing, an availability feature vector [2024Q1 availability rate 85%, 2024Q2 availability rate 30%] is generated. During the processing, the system establishes a dynamic monitoring model for resource status, converts the "equipment maintenance period" into an availability degradation coefficient for a specific time period, and calculates the remaining service capacity in combination with the current load rate. For open service resources, such as "accepting appointments throughout the year, with a response cycle of 3-5 working days", the system parses it into a continuous availability feature and marks the efficiency reduction coefficient during the peak period.

[0051] Step 1035: Perform feature fusion processing on the resource direction matching feature, the resource capability strength feature and the resource availability feature to generate the resource supply feature set.

[0052] In this embodiment, the fusion process adopts a deep neural network architecture that combines feature crossover with a gating mechanism. For example, the resource direction matching features (smart materials 0.95, sensor technology 0.88), resource capability strength features (9.1 / 10), and resource availability features (2024 annual availability 92%) of a certain university's "intelligent material research and development platform" are fused to form a 384-dimensional feature vector containing technical adaptability, service reliability, and time accessibility. During the fusion process, the gating mechanism dynamically adjusts the interaction weights of the direction matching features and the capability strength features. When high direction matching but low capability strength is detected, the impact of this dimension on the final feature is automatically reduced. The system sets a feature standardization layer to eliminate the scale differences between features of different dimensions.

[0053] In one implementation, the calling of the trained deep matching learning model in step 104 performs dynamic matching evaluation processing on the demand item feature set and the resource supply feature set to generate a matching evaluation result between each demand item description text and the resource supply description text, including: Step 1041: Calculate the direction adaptability of the demand direction feature in the demand item feature set and the resource direction matching feature in the resource supply feature set to generate a direction matching index.

[0054] Among them, the calculation and processing adopt a multi-dimensional space projection algorithm based on improved cosine similarity. For example, the requirement direction features of a project of "Autonomous Driving Multi-Sensor Fusion Algorithm Development" (Autonomous Driving 0.93, Sensor Technology 0.87, Real-Time Computing 0.79) are matched with the resource direction matching features of a research institute's "Environmental Perception Algorithm Optimization" resources (Autonomous Driving 0.88, Computer Vision 0.82, Signal Processing 0.75) to calculate the fitness, and a direction matching degree index of 0.89 is obtained. During the calculation process, the system sets domain weight coefficients for each direction dimension, and the contribution value of the matching degree in key fields (such as autonomous driving) is increased by 30%. For projects with multiple direction dimensions, the maximum matching principle is adopted to select the highest matching value in each dimension for weighted averaging.

[0055] Step 1042: Perform an ability fitness calculation and processing on the requirement complexity feature in the requirement project feature set and the resource ability intensity feature in the resource supply feature set to generate an ability matching degree index.

[0056] Among them, the calculation and processing establish a complexity-ability intensity mapping function, and a perfect matching degree is obtained when the resource ability intensity value is greater than or equal to the requirement complexity value. For example, the requirement complexity feature value of 8.5 of a new material R & D project is matched with the resource ability intensity feature value of 9.2 of a laboratory, and the ability matching degree index 0.92 is calculated through the formula min(1, resource ability intensity / requirement complexity). During the processing, the system sets a non-linear attenuation function. When the resource ability intensity is lower than the requirement complexity, the matching degree drops exponentially. For example, when the ability intensity is 7.0 corresponding to the complexity of 8.5, the matching degree drops to 0.68. This mechanism effectively avoids the occurrence of resource overload.

[0057] Step 1043: Perform a time window fitness calculation and processing on the requirement timeliness feature in the requirement project feature set and the resource availability feature in the resource supply feature set to generate a timeliness matching degree index.

[0058] Among them, the calculation and processing adopt a dynamic time warping algorithm to align the project time requirements and the resource availability cycle. For example, the requirement timeliness feature of an urgent technology research and development project requires completion in Q2 2024, which is matched with the available rate of 65% in Q2 2024 marked by the resource availability feature of a laboratory, and the timeliness matching degree index 0.78 is obtained through the time overlap calculation. During the processing, the system establishes a time sensitivity attenuation model and sets a higher weight coefficient for the change of resource availability near the deadline. For multi-stage projects, such as the completion requirements of the requirement timeliness feature including three stages of [Q1 2024, Q2 2024, Q3 2024], the system calculates the matching degree of each stage and takes the geometric mean as the final index.

[0059] Step 1044: Construct a comprehensive matching degree evaluation matrix based on the direction matching degree index, the ability matching degree index, and the timeliness matching degree index.

[0060] Exemplarily, the matrix construction adopts a method combining weighted summation and non-linear activation, and sets the direction matching degree weight to 0.5, the ability matching degree weight to 0.3, and the timeliness matching degree weight to 0.2. For example, the three matching degree indexes of a medical image analysis project and a certain AI computing resource are 0.92, 0.85, and 0.78 respectively. After weighted calculation, the comprehensive matching degree score is 0.89, forming the evaluation value at the corresponding position in the matrix. During the matrix construction process, the system introduces fuzzy logic to handle boundary cases. When a certain index is lower than the threshold but other indexes are particularly prominent, the expert rule base is activated for score correction. The matrix dimension expansion mechanism supports dynamically adding new demand items or resource supplies to maintain the scalability of the evaluation system.

[0061] Step 1045: Sort the matching association relationships between each demand item description text and the resource supply description text according to the comprehensive matching degree evaluation matrix to generate the matching degree evaluation result.

[0062] Optionally, the sorting process adopts a multi-objective optimization algorithm based on Pareto optimality to find the optimal solution set on the premise of ensuring the core matching conditions. For example, a provincial key R & D project obtains comprehensive scores of 0.93, 0.88, and 0.82 respectively with three resource supplies in the matching degree matrix. The system generates a recommended priority sequence: first recommend the laboratory resource with a matching degree of 0.93, then select the key university team with a matching degree of 0.88, and finally consider the enterprise technology center with a matching degree of 0.82. During the sorting process, the system sets industry domain preference parameters to increase the sorting weight of resource supplies that meet the regional industrial development plan by 10%. The evaluation result output module also generates a matching analysis report, which details the sub-indexes of the matching degree in each dimension and improvement suggestions.

[0063] As an optional technical solution, the training process of the deep matching learning model includes: Step 201: Obtain a set of historical project matching data for training, and the set of historical project matching data for training contains multiple paired samples of demand item description texts and resource supply description texts with marked matching results.

[0064] In specific implementation, the historical project matching data set for training is sourced from the real project docking records accumulated on the scientific research collaboration platform. Each paired sample includes the complete text description of the demand project, the text description of the resource supply, and the manually annotated matching result label. For example, the data set includes a paired sample of the "Quantum Encryption Communication Protocol Research" demand project described in step 101 and the "Multi-modal Perception Algorithm Research Team" resource mentioned in step 103, with the annotation result being "Highly Matched". During the data acquisition process, the system removes invalid or contradictory samples through a data cleaning module, such as deleting samples with significantly conflicting technical directions or incomplete annotation information. Each paired sample is appended with metadata such as the project cycle, field classification, and actual cooperation effectiveness to ensure the spatio-temporal consistency and logical completeness of the training data.

[0065] Step 202: Perform training demand feature extraction processing on the text description of the demand project in the paired sample to generate a training demand project feature set.

[0066] In the actual application process, the training demand feature extraction processing can adopt a technical process similar to that of steps 1021 to 1025 to ensure the consistency of the feature extraction logic. For example, for the text description of the "Intelligent Warehouse Robot Path Planning System R & D" demand project in step 101, it is segmented into three demand semantic segments: "Optimization of Dynamic Environment Perception Algorithm", "Design of Multi-robot Cooperative Scheduling Strategy", and "Research on Energy Consumption Efficiency Balancing Mechanism" through semantic segmentation processing. Direction features in the directions of "Machine Vision", "Operations Research Optimization", and "Energy Management" are generated through direction recognition processing. Combining the complexity quantification and timeliness analysis results, a multi-dimensional demand project feature vector is finally formed. During the processing, the system establishes a feature alignment mechanism to ensure the consistent distribution of the feature space in the training set and the validation set, avoiding the decline in the model's generalization ability due to data deviation.

[0067] Step 203: Perform training resource feature extraction processing on the text description of the resource supply in the paired sample to generate a training resource supply feature set.

[0068] Exemplarily, the training resource feature extraction processing can refer to steps 1031 to 1035 to ensure the symmetry between the resource features and the demand features. For example, for the text description of the "Material Synthesis Experimental Platform" resource supply in step 103, an identifier of "Experimental Equipment - Materials Engineering" is obtained through resource type classification processing. Direction matching features in the directions of "Intelligent Materials" and "Chemical Synthesis" are generated through direction matching degree analysis. Combining the ability strength evaluation and availability cycle analysis results, a multi-dimensional resource supply feature vector is finally formed. During the processing, the system conducts feature importance analysis to identify the feature dimensions that have a significant impact on the matching result. For example, it is found that the weight coefficients of the resource direction matching features are generally higher in interdisciplinary projects than in single-field projects.

[0069] Step 204: Perform a matching degree prediction process on the training requirement item feature set and the training resource supply feature set based on the initial deep matching learning model to generate a predicted matching degree result.

[0070] Among them, the initial deep matching learning model adopts a dual-encoder architecture, including a requirement feature encoding module and a resource feature encoding module. For example, input the requirement item feature vector of "automatic driving perception system development" described in step 102 into the requirement encoding module, and generate a requirement embedding representation through multi-layer non-linear transformation; at the same time, input the resource feature vector of "environmental perception algorithm optimization" described in step 103 into the resource encoding module to generate a resource embedding representation; finally, output a predicted matching degree value through the similarity calculation module. In the model initialization stage, an orthogonal weight initialization strategy is adopted to avoid the problems of gradient disappearance or explosion, and the hidden layer dimension setting of the encoding module is matched with the information density of the feature vector.

[0071] Step 205: Construct a model loss function according to the difference between the predicted matching degree result and the labeled matching result.

[0072] In this step, the loss function adopts a composite loss mechanism, combining the matching degree regression loss and the contrastive learning loss. For example, for the positive sample pair labeled as "highly matched", calculate the mean square error between the predicted matching degree and the labeled value; at the same time, construct a negative sample pair, and increase the matching degree difference between the positive and negative sample pairs through the contrastive loss function. A hard sample mining strategy is introduced in the process of constructing the loss function to automatically identify the samples with large differences between the predicted results and the labeled results, and dynamically adjust their loss weights to improve the model's learning ability for complex cases. For the sample of the key R & D project mentioned in step 101, the system sets a higher loss weight coefficient to enhance the matching accuracy of the model in key areas.

[0073] Step 206: Perform iterative optimization processing on the model parameters of the initial deep matching learning model through the backpropagation algorithm until the model loss function converges to a preset threshold, and generate the trained deep matching learning model.

[0074] It can be understood that the iterative optimization processing adopts an adaptive learning rate adjustment strategy, setting a higher learning rate in the initial stage to accelerate the convergence speed, and gradually reducing the learning rate in the later stage to improve the parameter optimization accuracy. For example, during the optimization process, the predicted matching degree of the deep matching learning model for the sample pair of "quantum encryption communication protocol research" and "multi-modal perception algorithm research team" described in step 201 gradually approaches the labeled value from the initial value, and the loss function curve shows a stable downward trend until it reaches the preset convergence standard. Regularization constraints and gradient clipping strategies are implemented during the optimization to effectively control the model complexity and training stability, and prevent overfitting.

[0075] As an alternative technical solution, the iterative optimization process of the model parameters of the initial deep matching learning model by the backpropagation algorithm in step 206 until the model loss function converges to a preset threshold to generate the trained deep matching learning model includes: Step 2061: Perform gradient calculation on the model loss function to generate the gradient update direction of each model parameter in the initial deep matching learning model.

[0076] Among them, the gradient calculation is implemented through automatic differentiation technology, and the error signal is propagated backward along the computational graph. For example, for the weight parameters of the fully connected layer of the demand feature encoding module, calculate its partial derivative with respect to the total loss function to determine the parameter update direction and amplitude. During the processing, the gradient checkpointing technology is used to optimize the memory usage efficiency, selectively store and recompute the intermediate calculation results in the deep network, and balance the computational resource consumption and accuracy requirements.

[0077] Step 2062: Perform parameter adjustment on the weight matrix of the initial deep matching learning model according to the gradient update direction to generate an intermediate optimized deep matching learning model.

[0078] Among them, the parameter adjustment adopts the stochastic gradient descent algorithm with momentum, and the momentum coefficient is set in a reasonable range to balance the convergence speed and stability. For example, for the convolution kernel parameters of the resource feature encoding module, combine the current gradient and the historical update direction for weighted averaging to effectively suppress the oscillation phenomenon during the parameter update process. The weight decay strategy is implemented during the adjustment process, and the model complexity is controlled through L2 regularization constraints to improve the generalization ability.

[0079] Step 2063: Perform training data verification on the intermediate optimized deep matching learning model, and calculate the matching accuracy of the intermediate optimized deep matching learning model on the validation set data.

[0080] Exemplarily, the validation set contains sample data independent of the training set, such as the "Application Development of New Composite Materials in the Aerospace Field" project in step 101 and the resource supply samples of the "Material Performance Testing Platform" mentioned in step 103. The accuracy evaluation adopts a multi-index comprehensive evaluation system, including core indicators such as Top-K hit rate and area under the precision-recall curve. During the verification process, it is found that the matching accuracy of the model for interdisciplinary projects has been significantly improved, and the validation loss value continues to decrease, indicating that the model optimization direction is correct.

[0081] Step 2064: When the matching accuracy does not reach the preset accuracy threshold, use the intermediate optimized deep matching learning model as the new initial deep matching learning model, and repeat the gradient calculation process, the parameter adjustment process, and the training data verification process.

[0082] In this step, for example, after a certain iteration, the matching prediction accuracy of the model for the pair of the "Research and Development of Thermal Protection System for Aerospace Vehicles" project in step 102 and the "Multi-Physical-Field Coupling Analysis Platform" resource in step 103 does not meet the standard. The system automatically triggers the learning rate decay mechanism and increases the sampling frequency of this type of sample. After multiple rounds of optimization, it finally reaches the preset threshold. The early stopping strategy is implemented in the cyclic optimization process. When the validation loss has not improved for multiple consecutive rounds, the training is automatically terminated to avoid ineffective calculations.

[0083] Step 2065: When the matching accuracy reaches the preset accuracy threshold, use the intermediate optimized deep matching learning model as the trained deep matching learning model.

[0084] It can be understood that the trained model is verified through an independent test set, and the test set contains all project type and resource type samples described in steps 101 to 105. For example, the evaluation result of the matching degree of the "Autonomous Driving Perception System Development" project in step 104 and the "Environmental Perception Algorithm Optimization" resource in step 103 by the model, and the error from the expert review conclusion remains within the allowable range, indicating that the model has the ability for actual deployment. When the final model is exported, quantization compression and format conversion are implemented to ensure its efficient operation in the distributed computing environment of the scientific research collaboration platform and to respond to the matching recommendation request described in step 105 in real time.

[0085] In a preferred embodiment, generating a scientific research project matching recommendation strategy according to the matching degree evaluation result in step 105 includes: Step 1051: Perform threshold screening processing on the matching degree evaluation result to obtain target matching association pairs that meet the preset matching degree threshold.

[0086] Specifically, the threshold screening processing is based on the comprehensive matching degree evaluation matrix generated in step 104, and filters out eligible project-resource combinations according to the preset matching degree threshold range. For example, for the "Research on Quantum Encryption Communication Protocol" demand project described in step 101, its matching degree evaluation result with the "Multi-Modal Perception Algorithm Research Team" resource in step 103 is 0.89. When the preset threshold is 0.85, this association pair is retained as a target matching association pair. During the screening process, the system establishes a multi-level threshold system, appropriately reduces the threshold standard for the demand projects with high complexity features described in step 1023 to ensure resource coverage for key scientific research tasks. For the cross-domain project matching matrix constructed in step 1044, a dynamic threshold adjustment mechanism is adopted to automatically adjust the screening criteria according to the domain priority coefficient.

[0087] Step 1052: Perform a priority sorting process on the requirement item description text and resource supply description text in the target matching association pair to generate a recommended priority sequence.

[0088] Exemplarily, the sorting process combines the matching degree metric and business rule weights. For example, for a project marked as a "key research topic" in Step 101, its sorting priority is increased based on the screening results of Step 1051. In a specific case, the association pair of the "Research and Development of Intelligent Warehouse Robot Path Planning System" requirement item and the "Multi-Robot Cooperative Scheduling Algorithm Platform" resource ranks among the top three in the sorting because the direction matching degree feature reaches 0.93 and meets the timeliness constraint of Step 1024. The sorting algorithm integrates the function of domain knowledge graph analysis and gives additional weights to the interdisciplinary ability identifiers included in the resource supply feature set in Step 1035.

[0089] Step 1053: Construct a dynamic recommendation list according to the recommended priority sequence. The dynamic recommendation list includes the detailed matching degree information of each target matching association pair and a recommended operation interface.

[0090] Among them, the list construction adopts a multi-dimensional information fusion technology. For example, it displays the direction matching degree metric in Step 1041, the sub-item values of the ability adaptation degree in Step 1042, and the time window matching status in Step 1043. For the "Development of Autonomous Driving Perception System" project described in Step 105, the dynamic recommendation list synchronously displays the equipment list, successful cases, and service response timeliness of the matching resource party, the "Environmental Perception Algorithm Optimization Team". The list supports an interactive filtering function and can be quickly retrieved according to the direction feature dimension in the requirement item feature set set in Step 1025.

[0091] Step 1054: Perform an association mapping process on the dynamic recommendation list and the identity identifiers of the scientific research project requester and the scientific research resource provider to generate an identity-customized recommendation strategy.

[0092] In this embodiment, the association mapping process realizes the personalized adaptation of the recommended content by docking the institutional attribute library of the scientific research project requester described in Step 101 and the service record database of the resource provider described in Step 103. For example, for the scientific research management department of a university, the recommendation strategy highlights resources for academic achievement transformation; for an enterprise R & D center, it preferentially recommends technical teams with strong engineering implementation capabilities. During the identity mapping process, the system automatically associates the priority sequence generated in Step 1052 with the institutional cooperation preference model to ensure that the recommendation strategy conforms to the historical cooperation mode of the requester.

[0093] Step 1055: Package the identity-customized recommendation strategy into a to-be-processed recommendation instruction to generate the scientific research project matching recommendation strategy.

[0094] It can be understood that the encapsulation process includes data format standardization and security encryption to ensure that sensitive information (such as the undisclosed service capabilities of the resource party) in the dynamic recommendation list in step 1053 complies with data security specifications during transmission. For example, the device parameter details of the "material synthesis experimental platform" resource in step 103 are encapsulated in the recommendation instruction using a hierarchical authority control mechanism and are fully displayed only to the demand party that has passed the identity authentication.

[0095] In a preferred embodiment, the step 105 of synchronizing the scientific research project matching recommendation strategy to the corresponding scientific research project demander and scientific research resource provider to trigger the scientific research collaboration optimization operation includes: Step 1056: Send the pending recommendation instructions in the scientific research project matching recommendation strategy to the demand-side device corresponding to the scientific research project demander.

[0096] The sending process integrates the multi-terminal synchronization mechanism of the scientific research project demander in step 101 to ensure the real-time consistency of the recommended instructions on PC workstations, mobile terminals and management system backgrounds. Taking the industry scientific research foundation in step 101 as an example, after receiving the recommended instructions, its demand-side device automatically triggers the update of the visual dashboard, highlighting the resource matching information of the "Quantum Technology Research Laboratory" ranked first in step 1052.

[0097] Step 1057: Send the pending recommendation instruction to the resource-end device corresponding to the scientific research resource party involved in the scientific research project matching recommendation strategy.

[0098] In this step, the instruction transmission adopts a two-way confirmation mechanism, and the receiving device of the resource party needs to return a digital signature receipt to confirm the integrity of the instruction. For example, after receiving the recommended instruction encapsulated in step 1055, the resource-side device of the "Multimodal Perception Algorithm Research Team" in step 103 automatically parses the technical points of the matching demand project (such as the real-time computing feature requirements described in step 102) and generates a task reminder to be responded to on the resource management interface.

[0099] Step 1058: Receive recommendation feedback data returned by the demand-side device and the resource-side device, wherein the recommendation feedback data includes an operation identifier for accepting a recommendation instruction or rejecting a recommendation instruction; and update the recommendation status identifier of the target matching association pair in the dynamic recommendation list according to the recommendation feedback data.

[0100] In this embodiment, the feedback processing mechanism synchronizes the operation states of both parties in real time. For example, when the enterprise R & D center in step 101 accepts the matching recommendation of the "New Composite Material Development" project and the "Material Performance Testing Platform" in step 103, the system immediately updates the status of this associated pair to "accepted" and automatically hides other competitive recommendation entries. For rejection feedback cases, the system records the specific reasons (such as resource availability period conflicts) and feeds them back to the matching degree evaluation model in step 104 for subsequent optimization.

[0101] Step 1059: When the recommendation status flag is an accept recommendation instruction, trigger the generation operation of the scientific research collaboration protocol and synchronize the generated scientific research collaboration protocol to the corresponding demand-side device and resource-side device.

[0102] In this step, the protocol generation integrates the demand project characteristics in step 1025 and the resource supply characteristics in step 1035, and automatically fills in the core contents such as technical indicators, delivery cycle, and intellectual property terms. Taking the successfully matched "Intelligent Warehouse Robot" project in step 105 as an example, the protocol template automatically references the feature parameters of the "Machine Vision" direction identified in step 1022 and the equipment capability strength indicators evaluated in step 1033 to form a legally binding cooperation document. The protocol synchronization process uses blockchain deposit and proof technology to ensure the immutability of the protocol versions obtained by both parties' terminals.

[0103] In an exemplary embodiment, the method further includes: Step 301: Monitor the recommendation execution effect data of the scientific research project matching recommendation strategy in real time. The recommendation execution effect data includes the number of successful matches, resource utilization rate indicators, and project completion timeliness indicators.

[0104] Specifically, the monitoring is carried out by docking with the scientific research collaboration protocol execution system generated in step 105 to collect the operation logs and result data in the project docking stage in real time. For example, for the matching case of the "Quantum Encryption Communication Protocol Research" project and the "Multi-modal Perception Algorithm Research Team", the system continuously tracks the signing status of the cooperation agreement between the two parties, the resource call records, and the project milestone achievement time. During the monitoring process, the system establishes a data collection interface to integrate the recommendation feedback data and the protocol execution data to form a complete execution effect tracking chain. For projects with high-complexity features, the system implements an enhanced monitoring strategy to increase the collection frequency of special indicators such as the acceptance status of technical nodes.

[0105] Step 302: Perform a quantitative evaluation process on the recommendation execution effect data to generate a recommendation effect evaluation report.

[0106] Among them, the evaluation process adopts a multi-dimensional index fusion method. For example, for the selected target matching correlation pairs, analyze the deviation law between their actual execution effects and predicted matching degrees. Taking the matching case of the "material synthesis experimental platform" resource and the "new composite material development" project as a benchmark, systematically quantify and evaluate the actual utilization rate of this resource during the project cycle and its impact on the project timeliness. The evaluation model integrates the processing logics from step 3021 to step 3027, ensures the objectivity of the evaluation results through a cross-validation mechanism, and maps the analysis conclusions to the aforementioned matching degree evaluation matrix optimization requirements.

[0107] Step 303: Perform an online incremental optimization process on the model parameters of the deep matching learning model according to the recommended effect evaluation report, and generate an optimized deep matching learning model.

[0108] Among them, the optimization process adopts a continuous learning mechanism. Without interrupting the matching recommendation service in step 105, fine-tune the model parameters based on the evaluation report data generated above. For example, for the resource availability feature extraction granularity optimized in step 3026, synchronously adjust the dimension of the parsing model embedding layer in step 1034, so that the resource encoding module can more accurately represent the fine-grained availability information. During the optimization process, implement model version control, retain the basic model architecture completed in training in step 206, and only update the parameters of specific feature processing modules to ensure system stability.

[0109] Step 304: Dynamically update the optimized deep matching learning model to the deep matching learning model that has completed training.

[0110] It can be understood that the update process adopts a hot-swap technology to complete the model replacement during the continuous operation of the recommendation service in step 105. For example, when the optimization of the direction adaptability calculation module is completed, the system gradually routes some requests to the new model through an AB test mechanism, and fully enables the updated model after verifying that the effect is stable. During the update process, the system keeps the model document synchronized and updated as set in step 2065, records the parameter change range and performance improvement indicators involved in each optimization, and forms a complete model iteration knowledge base.

[0111] In an exemplary embodiment, the effect quantification evaluation process for the recommended execution effect data to generate a recommended effect evaluation report includes: Step 3021: Perform a statistical process on the number of successful matches to generate a matching success rate indicator.

[0112] Specifically, based on the above-recorded recommended feedback data, the statistical processing calculates the proportion of matching associated pairs that accept the recommendation instructions in the total number of recommendations. For example, for the historical recommendation records of the industry scientific research foundation, statistics show that the average matching success rate of the top three resources in the generated recommendation priority sequence reaches 78%. During the processing, the system differentiates the success rate differences of projects with different timeliness levels set, sets up an independent statistical unit for urgent research and development projects, and identifies the matching pattern characteristics in special scenarios.

[0113] Step 3022: Conduct trend analysis processing on the resource utilization rate indicator to generate a resource utilization efficiency curve.

[0114] Among them, the analysis processing integrates the parsed resource availability characteristics and the actual call records during protocol execution to construct a trend map of the resource load rate changing over time. For example, after the "Multi-physical Field Coupling Analysis Platform" resource is matched with the "Research and Development of Thermal Protection System for Aerospace Vehicles" project, its monthly utilization rate increases from 45% before matching to 82%. Based on this, the system generates an efficiency curve showing an upward trend. During the analysis process, the system correlates the generated set of resource supply characteristics to identify the correlation rules between resource types and utilization rate fluctuations. For example, the fluctuation range of the utilization rate of experimental equipment resources is generally lower than that of technical service resources.

[0115] Step 3023: Conduct time window comparison processing on the project completion timeliness indicator to generate a timeliness optimization coefficient.

[0116] It can be understood that the time window comparison processing calculates the time deviation rate and converts it into an optimization coefficient by comparing the set demand timeliness characteristics with the recorded actual completion time. For example, the expected cycle of the "Research and Development of Intelligent Warehouse Robot Path Planning System" project is 18 months, and it is actually completed in 16 months after matching the resources in step 103. The system generates a positive timeliness optimization coefficient of 1.125. During the processing, the system establishes a time deviation type classification model to distinguish delays caused by insufficient resource supply from natural delays due to the difficulty of technical research and development exceeding expectations, ensuring that the optimization coefficient accurately reflects the timeliness improvement effect of the matching strategy.

[0117] Step 3024: Conduct report integration processing on the matching success rate indicator, the resource utilization efficiency curve, and the timeliness optimization coefficient to generate the recommended effect evaluation report.

[0118] It can be understood that the report integration processing adopts a method combining a visual dashboard and structured data. For example, the statistically calculated matching success rate is presented by classifying the demand side types, and at the same time, the superimposed generated resource utilization efficiency curve reflects the cross-institutional collaboration effect. The report content is automatically associated with the constructed comprehensive matching degree evaluation matrix, and the execution effect distribution characteristics of high matching degree areas are displayed through a heat map, providing data support for subsequent optimization.

[0119] Step 3025: Adjust the numerical range of the preset matching degree threshold according to the matching success rate index in the recommended effect evaluation report.

[0120] Among them, the adjustment process implements a dynamic threshold management mechanism. For example, when statistics show that the matching success rate in a certain field is continuously lower than expected, the system dynamically lowers the matching degree threshold benchmark for that field according to the direction adaptability calculation rule. During the adjustment process, the system retains the original multi-level threshold system architecture, and sets different adjustment amplitudes for projects with different complexity levels to ensure that the resource matching coverage of high-complexity projects is not affected by threshold changes.

[0121] Step 3026: Optimize the extraction granularity of the resource availability feature in the resource supply feature set according to the resource utilization efficiency curve.

[0122] In the specific implementation process, the optimization process of this step refines the parsing accuracy of the availability cycle. For example, the original quarterly availability parsing is improved to monthly parsing. Taking the resources of the "Environmental Perception Algorithm Optimization Team" as an example, its resource availability feature is optimized from "Available rate in Q2 2024: 65%" to "Available rate in April 2024: 70%, Available rate in May 2024: 60%, Available rate in June 2024: 65%", thereby improving the accuracy of the time window adaptability calculation. During the optimization process, the system establishes a resource status prediction model, predicts the future availability trend based on the efficiency curve in Step 3022, and incorporates it into the feature extraction process.

[0123] Step 3027: Adjust the calculation weight of the time window adaptability calculation process between the demand timeliness feature and the resource availability feature according to the timeliness optimization coefficient.

[0124] It can be understood that the adjustment process implements a dynamic weight allocation strategy. For example, when the calculated timeliness optimization coefficient shows that the actual cycle of a certain type of project is generally shorter than expected, the system increases the weight ratio of the time window adaptability index in the comprehensive matching degree evaluation. During the adjustment process, the system associates the classification result of the demand timeliness feature, and sets an upper limit value for the time window weight of urgent projects to avoid weakening core indicators such as direction matching degree or ability adaptability due to excessive optimization.

[0125] In an alternative non-limiting embodiment, the method further includes: Receiving the recommended feedback data returned by the scientific research project requester and the scientific research resource provider for the scientific research project matching recommendation strategy, where the recommended feedback data includes an acceptance status identifier for the matching recommendation result and a text of modification suggestions; Performing intent parsing processing on the recommended feedback data to extract the demand adjustment feature and the resource adjustment feature in the text of the modification suggestions; Update the corresponding requirement direction feature, requirement complexity feature, and requirement timeliness feature in the requirement item feature set according to the described requirements adjustment features; Update the corresponding resource direction matching feature, resource capacity intensity feature, and resource availability feature in the resource supply feature set according to the described resource adjustment features; Based on the updated requirement item feature set and resource supply feature set, re-execute the dynamic matching degree evaluation process, generate an optimized matching degree evaluation result, and update the scientific research project matching recommendation strategy.

[0126] Specifically, during the operation of the scientific research collaboration platform, the system continuously collects feedback data from the demand side and the resource side on the recommendation results to achieve dynamic optimization of the matching strategy. For example, when a scientific research management department of a university (the demand side) receives a recommendation list for the project of "Research and Development of Intelligent Warehouse Robot Path Planning System", if it is found that the recommended resource of "Multi-Robot Cooperative Scheduling Algorithm Platform" has a deviation in the technical direction, it can submit a modification suggestion through the feedback interface, pointing out that "the adaptability of the dynamic environment perception algorithm needs to be strengthened". The system analyzes this suggestion through natural language processing technology, identifies the requirement adjustment feature of "dynamic environment perception", and automatically updates the weight of the corresponding direction feature in the requirement item feature set. At the same time, if the resource side (such as a robot technology laboratory) feedbacks that the device load rate is approaching saturation, the system will extract "insufficient experimental equipment throughput" as the resource adjustment feature and correspondingly reduce the value of the resource capacity intensity feature of this resource. After completing the feature update, the system re-executes the matching degree evaluation and generates an optimized recommendation list. For example, the "Machine Vision Algorithm Optimization Team" ranked third in the original recommendation sequence may rise to the first place due to the improvement of the direction feature adaptability. This mechanism effectively solves the problem of inaccurate recommendations caused by feature extraction deviation in the initial matching model. For example, in the research project of quantum encryption communication protocol, through multiple iterative feedback optimizations, a scientific research team with experience in anti-quantum computing attack algorithms is finally accurately matched.

[0127] In an alternative non-limiting embodiment, the method further includes: Real-time monitor the resource availability feature in the resource supply feature set corresponding to the scientific research resource side, and identify the dynamic change data of the resource callable time range in the resource availability feature; Perform incremental update processing on the resource availability feature according to the dynamic change data to generate a real-time updated resource supply feature set; Input the real-time updated resource supply feature set and the current requirement item feature set into the deep matching learning model, re-execute the dynamic matching degree evaluation process, and generate a real-time matching degree evaluation result; Based on the real-time matching degree evaluation result, perform priority rearrangement processing on the dynamic recommendation list in the scientific research project matching recommendation strategy to generate a recommended priority sequence after real-time adjustment; Push the updated recommended operation interface to the corresponding scientific research project demand side and scientific research resource side according to the recommended priority sequence after real-time adjustment.

[0128] For example, the system tracks the dynamic changes in the resource supply status in real time to ensure that the recommendation strategy is updated synchronously with the actual situation. Taking a certain material synthesis experimental platform as an example, when its resource availability feature changes from "quarterly availability rate of 85%" to "remaining available machine hours this month is 120 hours", the system immediately triggers the feature update process. In the updated resource supply feature set, the availability sub-feature of this platform is refined to the hour-level accuracy and rematched with the features of all current demand projects. For example, for the "development of new composite materials" project, the platform was ranked at the top in the original recommendation list because the quarterly matching degree reached the standard, but after real-time update, it was found that the remaining machine hours this month could not meet the urgent needs of the project, so the system will automatically lower the matching degree score and increase the recommendation priority of the "rapid prototyping laboratory" with immediate availability. At the same time, for sudden resource releases (such as a certain analytical test equipment having a gap due to project delays), the system can complete the update of resource availability features within minutes and re-evaluate the matching relationship. This dynamic adjustment mechanism has been typically applied in the "development of autonomous driving perception system" project: when a certain sensor calibration platform suddenly opens an emergency reservation period next week, the system immediately dynamically matches this resource with the perception algorithm project with high timeliness requirements, increasing the resource utilization rate by 37%.

[0129] In an alternative non-limiting embodiment, the method further includes: Obtain the project execution progress data corresponding to the matching association pairs that have triggered scientific research collaboration optimization operations, where the project execution progress data includes the actual resource invocation time node and the project stage completion status; Perform deviation analysis processing based on the actual resource invocation time node and the resource callable time range in the resource availability feature to generate a resource scheduling deviation index; Perform timeliness comparison processing based on the project stage completion status and the expected completion time constraint condition in the demand timeliness feature to generate a project progress delay coefficient; Perform parameter calibration processing on the resource capacity intensity feature and the demand complexity feature based on the resource scheduling deviation index and the project progress delay coefficient to generate a calibrated resource supply feature set and a demand project feature set; Input the calibrated feature set into the deep matching learning model for re-evaluation of the matching degree and update the target matching association pairs in the scientific research project matching recommendation strategy.

[0130] In this embodiment, by continuously monitoring the execution process of the matched items, the system realizes the closed-loop calibration of the characteristic parameters. During the execution of a project named "Research and Development of Thermal Protection System for Aerospace Vehicles", the system found that the actual frequency of invoking the "Multi-Physical-Field Coupling Analysis Platform" exceeded the predicted value of the resource capacity intensity characteristic by 20%, and immediately triggered the parameter calibration process. The calibrated resource capacity intensity characteristic will increase the "attenuation coefficient of high-concurrency task processing ability", which can more accurately characterize the performance change law of the platform during continuous operation. At the same time, the project progress delay data (such as the third stage was postponed for two weeks due to the unexpected extension of the material test cycle) will be converted into a correction parameter for the demand complexity characteristic, enhancing the prediction accuracy of the timeliness of subsequent similar projects. This calibration mechanism is particularly important in interdisciplinary projects. For example, in a project named "Biomedical Data Analysis", the project progress lagged behind due to the underestimated algorithm complexity. The system calibrated the demand complexity characteristic value from 7.8 to 8.5 through calibration, making subsequent matching more inclined to select a computing resource team with the ability to tackle problems across fields.

[0131] In an alternative non-limiting embodiment, the method further includes: Identifying the rejection recommendation instructions corresponding to the unaccepted recommendation operation interfaces in the scientific research project matching recommendation strategy, and extracting the rejection reason text in the rejection recommendation instructions; Performing keyword extraction processing on the rejection reason text to generate a demand mismatch label and a resource shortage label; Adjusting the weight parameter in the direction adaptability calculation process of the demand direction characteristic and the resource direction matching characteristic according to the demand mismatch label; Adjusting the quantization threshold in the ability adaptability calculation process of the resource capacity intensity characteristic according to the resource shortage label; Based on the adjusted weight parameter and quantization threshold, re-executing the dynamic matching degree evaluation process, generating a corrected matching degree evaluation result and updating the recommendation priority sequence.

[0132] It can be understood that for the case where the recommendation is rejected, the system deeply mines the reasons for rejection to optimize the matching logic. When a certain enterprise's R & D center repeatedly rejects the recommendation of "traditional control algorithm optimization" resources, the system extracts the key requirement mismatch label of "strengthening real-time computing capabilities" from the rejection reason text. Accordingly, the model automatically increases the weight coefficient of the "real-time computing" direction in the demand direction features, making subsequent recommendations more focused on resources with edge computing acceleration capabilities. At the same time, if a certain laboratory rejects a matching request due to "the annual inspection of the spectroscopic analysis equipment is out of service", the system will generate a resource shortage label and dynamically adjust the evaluation threshold of the equipment capacity strength feature. This mechanism performs significantly in complex scenarios. For example, after a "quantum communication prototype test" project was rejected three times in a row due to "conflicts in booking the low-temperature experimental chamber", the system adjusted the feature threshold and redirected the matching recommendation to a cooperative institution with redundant low-temperature experimental resources, increasing the success rate from 42% to 79%.

[0133] In an alternative non-limiting embodiment, the method further includes Technical Solution 5: Detect high-frequency matching requests from multiple scientific research project requesters for the resource supply description text of the same scientific research resource provider, and generate a resource competition conflict identifier; Analyze and process the conflict reasons for the target matching association pairs involved according to the resource competition conflict identifier, and extract resource demand overlap features and time window conflict features; Based on the resource demand overlap features, perform domain subdivision processing on the demand direction features to generate refined demand direction sub-features; Based on the time window conflict features, perform segmented scheduling processing on the resource availability features to generate resource availability sub-features for multiple time periods; Re-input the refined demand direction sub-features and the resource availability sub-features for multiple time periods into the deep matching learning model to generate a matching degree evaluation result after conflict resolution and update the dynamic recommendation list.

[0134] In the specific application process, the system optimizes the resource competition problem through a conflict resolution mechanism. When multiple demand parties (such as an autonomous driving enterprise and a quantum computing laboratory) simultaneously request "high-performance computing cluster" resources at a high frequency, the system identifies the overlapping characteristics of resource requirements and conducts nanoscale subdivision of the demand direction characteristics. For example, "high-performance computing" is decomposed into sub-directions of "parallel algorithm optimization" and "large-scale numerical simulation", and different computing resources with CUDA acceleration experience and expertise in fluid dynamics simulation are respectively matched. At the same time, for the conflict characteristics of time windows (such as a material simulation project requiring exclusive use of computing resources for two weeks), the system decomposes the resource availability characteristics into multiple spliceable time periods to achieve time-sharing scheduling recommendations across projects. This mechanism plays a key role in the sharing of major scientific research infrastructures. For example, through the management of time-period availability characteristics in the National Supercomputing Center, projects such as "climate prediction models" and "gene sequence alignment" can use the same computing cluster at off-peak times. While the peak resource utilization rate drops from 91% to 68%, the average project completion time is shortened by 15%.

[0135] In the embodiment of the present invention, through the multi-dimensional feature modeling and dynamic depth matching mechanism, the precise adaptation of scientific research resources and project requirements is realized. First, a three-dimensional requirement feature system covering direction characteristics, complexity levels, and timeliness indicators is constructed for historical project requirement data, breaking through the limitations of traditional single-label matching and being able to deeply deconstruct the implicit associations and priority relationships in the requirement text. At the same time, in the extraction of resource supply characteristics, a ternary evaluation dimension of direction matching degree quantification, ability strength grading, and availability dynamic tracking is innovatively integrated to realize the three-dimensional characterization of resource attributes. Through the dynamic coupling analysis of the two-way feature set by a pre-trained deep matching model, the non-linear associations and dynamic evolution laws between requirements and resources can be captured, significantly improving the matching accuracy of cross-domain scientific research collaboration. Based on the real-time generated matching recommendation strategy, a two-way collaborative optimization mechanism can be triggered to promote the transformation of scientific research resource allocation from discrete response to intelligent scheduling, effectively solving the problems of resource mismatch and collaboration lag caused by traditional static matching.

[0136] See Figure 2 As shown, this figure is a schematic diagram of the basic structure of a scientific research project matching recommendation system 200 provided by an embodiment of the present invention. The scientific research project matching recommendation system 200 includes: A processor 201; A storage device 202, on which a computer program 2020 is stored; When the computer program 2020 is executed by the processor 201, the processor 201 implements any one of the artificial intelligence-based scientific research project matching recommendation methods.

[0137] On the basis described above, a readable storage medium is provided, and a program or instructions are stored on the readable storage medium. When the program or instructions are executed by a processor, the steps of the above method are implemented.

[0138] It should be noted that the various embodiments in this specification are described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the systems or devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple. For the relevant parts, reference can be made to the descriptions in the method part.

Claims

1. A scientific research project matching and recommendation method based on artificial intelligence, characterized in that, Including: Obtain the historical project requirement data set of scientific research project requesters and the historical resource supply data set of scientific research resource providers. The historical project requirement data set contains multiple requirement project description texts and corresponding requirement attribute tags, and the historical resource supply data set contains multiple resource supply description texts and corresponding resource attribute tags; Perform requirement feature extraction processing on the historical project requirement data set to generate a requirement project feature set. The requirement project feature set contains the requirement direction feature, requirement complexity feature, and requirement timeliness feature corresponding to each requirement project description text; Perform resource feature extraction processing on the historical resource supply data set to generate a resource supply feature set. The resource supply feature set contains the resource direction matching feature, resource ability intensity feature, and resource availability feature corresponding to each resource supply description text; Call the trained deep matching learning model to perform dynamic matching degree evaluation processing on the requirement project feature set and the resource supply feature set, and generate the matching degree evaluation result between each requirement project description text and the resource supply description text; Generate a scientific research project matching recommendation strategy according to the matching degree evaluation result, and synchronize the scientific research project matching recommendation strategy to the corresponding scientific research project requesters and scientific research resource providers to trigger scientific research collaboration optimization operations.

2. The method according to claim 1, wherein The performing requirement feature extraction processing on the historical project requirement data set to generate a requirement project feature set includes: Perform semantic segmentation processing on the requirement project description texts in the historical project requirement data set to obtain multiple requirement semantic segments; Call the pre-trained requirement feature encoder to perform direction recognition processing on the multiple requirement semantic segments to generate the requirement direction feature corresponding to each requirement project description text. The requirement direction feature is used to characterize the scientific research field direction to which the requirement project description text belongs; Perform complexity quantification processing on the requirement attribute tags in the historical project requirement data set to generate the requirement complexity feature corresponding to each requirement project description text. The requirement complexity feature is used to characterize the scientific research implementation difficulty level of the requirement project description text; Perform timeliness analysis processing on the requirement attribute tags in the historical project requirement data set to generate the requirement timeliness feature corresponding to each requirement project description text. The requirement timeliness feature is used to characterize the expected completion time constraint condition of the requirement project description text; Perform feature fusion processing on the requirement direction feature, the requirement complexity feature, and the requirement timeliness feature to generate the requirement project feature set.

3. The method according to claim 2, characterized in that, The performing resource feature extraction processing on the historical resource supply data set to generate a resource supply feature set includes: Perform resource type classification processing on the resource supply description texts in the historical resource supply data set to obtain multiple resource type identifiers; Perform a direction matching degree analysis process on the resource supply description text based on the multiple resource type identifiers to generate a resource direction matching feature corresponding to each resource supply description text, where the resource direction matching feature is used to characterize the adaptability of the resource supply description text to different scientific research field directions; Perform a capability intensity evaluation process on the resource attribute tags in the historical resource supply data set to generate a resource capability intensity feature corresponding to each resource supply description text, where the resource capability intensity feature is used to characterize the quantitative index of the scientific research resource supply capability corresponding to the resource supply description text; Perform an availability period parsing process on the resource attribute tags in the historical resource supply data set to generate a resource availability feature corresponding to each resource supply description text, where the resource availability feature is used to characterize the time range during which the resource corresponding to the resource supply description text can be invoked; Perform a feature fusion process on the resource direction matching feature, the resource capability intensity feature, and the resource availability feature to generate the resource supply feature set.

4. The method according to claim 3, wherein Invoke the trained deep matching learning model to perform a dynamic matching degree evaluation process on the demand project feature set and the resource supply feature set, and generate a matching degree evaluation result between each demand project description text and the resource supply description text, including: Perform a direction adaptability calculation process on the demand direction feature in the demand project feature set and the resource direction matching feature in the resource supply feature set to generate a direction matching degree index; Perform a capability adaptability calculation process on the demand complexity feature in the demand project feature set and the resource capability intensity feature in the resource supply feature set to generate a capability matching degree index; Perform a time window adaptability calculation process on the demand timeliness feature in the demand project feature set and the resource availability feature in the resource supply feature set to generate a timeliness matching degree index; Construct a comprehensive matching degree evaluation matrix based on the direction matching degree index, the capability matching degree index, and the timeliness matching degree index; Sort the matching association relationships between each demand project description text and the resource supply description text according to the comprehensive matching degree evaluation matrix to generate the matching degree evaluation result.

5. The method according to claim 4, wherein The training process of the deep matching learning model includes: Obtain a training historical project matching data set, where the training historical project matching data set contains multiple paired samples of demand project description texts and resource supply description texts with labeled matching results; Perform a training demand feature extraction process on the demand project description texts in the paired samples to generate a training demand project feature set; Perform a training resource feature extraction process on the resource supply description texts in the paired samples to generate a training resource supply feature set; Perform a matching degree prediction process on the training demand project feature set and the training resource supply feature set based on the initial deep matching learning model to generate a predicted matching degree result; Construct a model loss function according to the difference between the predicted matching degree result and the labeled matching result; Iteratively optimize the model parameters of the initial deep matching learning model through the backpropagation algorithm until the model loss function converges to a preset threshold, and generate the trained deep matching learning model.

6. The method according to claim 5, wherein The iterative optimization process of the model parameters of the initial deep matching learning model through the backpropagation algorithm includes: Perform gradient calculation on the model loss function to generate the gradient update direction of each model parameter in the initial deep matching learning model; Adjust the parameters of the weight matrix of the initial deep matching learning model according to the gradient update direction to generate an intermediate optimized deep matching learning model; Perform training data verification on the intermediate optimized deep matching learning model, and calculate the matching accuracy rate of the intermediate optimized deep matching learning model on the validation set data; When the matching accuracy rate does not reach the preset accuracy threshold, use the intermediate optimized deep matching learning model as the new initial deep matching learning model, and repeat the gradient calculation process, the parameter adjustment process, and the training data verification process; When the matching accuracy rate reaches the preset accuracy threshold, use the intermediate optimized deep matching learning model as the trained deep matching learning model.

7. The method according to claim 1, wherein The generation of the scientific research project matching recommendation strategy according to the matching degree evaluation result includes: Perform threshold screening on the matching degree evaluation result to obtain target matching association pairs that meet the preset matching degree threshold; Perform priority ranking on the requirement project description text and the resource supply description text in the target matching association pairs to generate a recommended priority sequence; Construct a dynamic recommendation list according to the recommended priority sequence, where the dynamic recommendation list includes the detailed matching degree information of each target matching association pair and a recommended operation interface; Perform associated mapping on the dynamic recommendation list with the identity identifiers of the scientific research project demand side and the scientific research resource side to generate an identity-customized recommendation strategy; Package the identity-customized recommendation strategy into a recommendation instruction to be processed to generate the scientific research project matching recommendation strategy.

8. The method according to claim 7, wherein The synchronization of the scientific research project matching recommendation strategy to the corresponding scientific research project demand side and scientific research resource side to trigger scientific research collaboration optimization operations includes: Send the recommendation instruction to be processed in the scientific research project matching recommendation strategy to the demand-side device corresponding to the scientific research project demand side; Send the recommendation instruction to be processed to the resource-side device corresponding to the scientific research resource side involved in the scientific research project matching recommendation strategy; Receive the recommended feedback data returned by the demand-side device and the resource-side device, where the recommended feedback data includes an operation identifier for accepting or rejecting the recommendation instruction; Update the recommendation status identifier of the target matching association pair in the dynamic recommendation list according to the recommended feedback data; When the recommendation status identifier is an accepted recommendation instruction, trigger the generation operation of the scientific research collaboration protocol, and synchronize the generated scientific research collaboration protocol to the corresponding demand-side device and resource-side device.

9. The method according to claim 1, characterized in that The method further includes: Monitor the recommendation execution effect data of the scientific research project matching recommendation strategy in real time. The recommendation execution effect data includes the number of successful matches, the resource utilization rate index, and the project completion timeliness index; Perform effect quantification evaluation processing on the recommendation execution effect data to generate a recommendation effect evaluation report; Perform online incremental optimization processing on the model parameters of the deep matching learning model according to the recommendation effect evaluation report to generate an optimized deep matching learning model; Dynamically update the optimized deep matching learning model to the trained deep matching learning model; The performing effect quantification evaluation processing on the recommendation execution effect data to generate a recommendation effect evaluation report includes: Perform statistical processing on the number of successful matches to generate a matching success rate index; Perform trend analysis processing on the resource utilization rate index to generate a resource utilization efficiency curve; Perform time window comparison processing on the project completion timeliness index to generate a timeliness optimization coefficient; Perform report integration processing on the matching success rate index, the resource utilization efficiency curve, and the timeliness optimization coefficient to generate the recommendation effect evaluation report; Adjust the numerical range of the preset matching degree threshold according to the matching success rate index in the recommendation effect evaluation report; Optimize the extraction granularity of the resource availability feature in the resource supply feature set according to the resource utilization efficiency curve; Adjust the calculation weight of the time window adaptability calculation processing between the demand timeliness feature and the resource availability feature according to the timeliness optimization coefficient.

10. A scientific research project matching and recommendation system, characterized in that Including: A processor; A storage device on which a computer program is stored. When the computer program is executed by the processor, the processor implements the artificial intelligence-based scientific research project matching recommendation method according to any one of claims 1-9.

Images