Science and technology project evaluation method and system

Abstract

The invention relates to a science and technology project evaluation method and system, and relates to the technical field of multi-index decision and intelligent modeling. Comprising the steps of collecting multi-dimensional data of finance, intellectual property, teams, historical projects and the like of science and technology projects; constructing an evaluation index system; constructing a cognitive map by taking the indexes as nodes and combining statistical correlation and expert rules; a graph attention network (GAT) is utilized to train the graph, and structural causal weights among indexes are extracted; fusing the structure weight and the original fuzzy relation matrix to form a structure enhanced fuzzy relation matrix; and in combination with a weight vector determined by a fuzzy analytic hierarchy process (FAHP), performing fuzzy weighted comprehensive evaluation, and outputting a project grade and an interpretable result. According to the method, the accuracy, self-adaptability and interpretability of science and technology project evaluation are effectively improved, and the method is suitable for complex project evaluation tasks in a multi-source heterogeneous data environment.

Description

Technical Field

[0001] The present invention relates to the field of scientific and technological projects, and in particular to a scientific and technological project evaluation method and system. Background Art

[0002] With the continued advancement of policies driven by scientific and technological innovation, governments and businesses are increasingly demanding scientific, quantitative, and intelligent decision-making methods for the establishment, management, and performance evaluation of science and technology projects. Science and technology projects are often characterized by high uncertainty, multiple influencing factors, and a mix of subjective and objective information. Traditional evaluation methods struggle to fully and accurately reflect a project's true value and development potential.

[0003] Currently, mainstream science and technology project evaluation methods often utilize multi-criteria decision-making models such as expert scoring, the Analytic Hierarchy Process (AHP), and fuzzy comprehensive evaluation. While these methods offer certain advantages in structured modeling, they still suffer from the following major issues: Weight setting relies on subjective judgment and lacks dynamic adaptability: Weights are typically determined based on one-time expert scoring or a fuzzy judgment matrix, failing to account for differences between projects or changing industry characteristics, which can easily lead to distorted evaluation results. Each criterion is considered independent, making it impossible to model a "causal relationship structure": In actual evaluations, innovation capabilities may influence market expectations, and team structure may constrain technical feasibility. Traditional models cannot explicitly capture this "structural dependency," limiting their interpretability and generalization capabilities. Summary of the Invention

[0004] (1) Technical problems solved To address the shortcomings of existing technologies, the present invention aims to provide a technology project evaluation method and system. This approach addresses the problems of existing technology project evaluation methods, such as the strong subjectivity of weights, unreasonable assumptions about indicator independence, and a lack of causal modeling and structural interpretation capabilities. A structurally enhanced evaluation method is proposed that integrates cognitive graphs with graph attention networks (GATs). This approach constructs evaluation indicators as cognitive graphs, introducing directed edges based on statistical correlations and expert rules to explicitly characterize the causal or dependency relationships between indicators. Using the graph attention network to train the graph, it automatically learns the nonlinear influence strengths between indicators and outputs a structural causal weight matrix, thereby enhancing the traditional fuzzy relationship matrix. Combined with subjective weights determined by the fuzzy analytic hierarchy process, the model achieves a weighted comprehensive evaluation that integrates structure and experience, improving evaluation accuracy and interpretability. Furthermore, the system supports online training and dynamic updating of causal weights as project data accumulates. This system exhibits excellent adaptability and scalability, enabling it to better adapt to multi-source heterogeneous data environments and evolving industry evaluation needs.

[0005] (2) Technical solution To achieve the above object, the present invention provides the following technical solution: a method for evaluating a scientific and technological project, comprising the following steps: Data collection: Collect basic project information, corporate financial data, intellectual property data, human resources data, and historical scientific and technological project data through data collection templates and interfaces; Data processing: missing value filling, outlier processing, data conversion, standardization, and correlation analysis are performed on the collected data in sequence to obtain a standardized feature matrix; Construction of evaluation index system: Based on the project evaluation requirements, a hierarchical factor set consisting of five primary indicators, including innovation, feasibility, market prospects, team capabilities and financial status, and their secondary indicators is established; Structural causal modeling: construct an indicator cognitive map with the factor set as nodes; generate weighted directed edges based on historical data correlation and expert rules to obtain an adjacency matrix; input the node feature matrix and the adjacency matrix into the graph attention neural network GAT for training, and output the structural causal weight matrix ; Combined with the original fuzzy membership matrix R, a structure-enhanced fuzzy relationship matrix is ​​formed. ; Determination of indicator weights: Use fuzzy analytic hierarchy process (FAHP) to construct a fuzzy judgment matrix and defuzzify it to obtain the weight vector W; Fuzzy comprehensive evaluation: using weight vector W and structure-enhanced fuzzy relationship matrix Calculate the comprehensive evaluation vector B and determine the project level according to the maximum membership principle; Result analysis and visual interpretation: Generate causal maps, risk warnings, and automated assessment reports based on indicator contribution, GAT attention coefficient, and similar case comparisons.

[0006] Preferably, the data conversion includes: performing logarithmic transformation on skewed distribution data; and performing one-hot encoding on categorical data.

[0007] Preferably, in the structural causal modeling step, the indicator cognitive map construction includes: taking the evaluation indicators as nodes, constructing weighted directed edges based on historical data correlation and expert knowledge, and forming a directed graph structure that reflects the causal relationship between indicators.

[0008] Preferably, in the structural causal modeling step, the propagation formula of the graph attention neural network GAT is: Where, represents the representation vector of node i in layer l; represents the neighbor set of node i; : represents the weight matrix of the k-th attention head (trainable parameters); represents the attention weight of node j to node i (normalized causal strength); Represents the concatenation operation of each attention head; σ is the activation function LeakyReLU.

[0009] Preferably, in the structural causal modeling step, the structural causal weight matrix The expression is: , where It represents the structural weight influence strength of indicator j on indicator i.

[0010] Preferably, the indicator weight determination step includes: building a judgment hierarchy, constructing a fuzzy judgment matrix, calculating fuzzy weights, defuzzification processing and consistency testing.

[0011] Preferably, the judgment hierarchy includes a target layer, a criterion layer and a solution layer; the target layer is for evaluating the quality of scientific and technological projects, the criterion layer is a first-level indicator, and the solution layer is a second-level indicator.

[0012] Preferably, constructing the fuzzy judgment matrix includes: multiple experts make “two-by-two comparisons” on each pair of indicators and give importance judgments in the form of language quantification; converting language judgments into triangular fuzzy numbers , construct the fuzzy judgment matrix .

[0013] A science and technology project evaluation system, comprising: Data collection templates and interfaces: used to collect customer technology project-related data and patent information data; Central processing module: It is used to monitor, analyze and process the collected data; Data storage module: It is used to store the original collected data, processed data, intermediate data during model training, and final model parameters; User interaction module: It is used to provide a friendly user interface to facilitate the interaction between scientific and technological project staff and customers and the system; Report generation module: It is used to automatically generate detailed science and technology project evaluation reports based on the evaluation results; Model update module: It is used to start the model update process.

[0014] (3) Beneficial effects The purpose of the present invention is to provide a scientific and technological project evaluation method and system with significant technological advancement and practical application value. First, by constructing an indicator cognitive map that includes statistical correlation and expert knowledge, the limitation of the independence assumption between indicators in traditional evaluation methods is solved, so that the evaluation model can explicitly characterize the causal or dependency relationship between indicators. Secondly, the graph attention network (GAT) is introduced to train the map, automatically learn and quantify the nonlinear structural influence between indicators, and significantly improve the model's ability to model complex interactive information. After fusion with the original fuzzy relationship matrix, a structure-enhanced fuzzy relationship matrix can be formed, which greatly improves the accuracy and discrimination of the evaluation results while maintaining the interpretability and stability of the traditional fuzzy comprehensive evaluation.

[0015] In addition, the present invention combines the fuzzy analytic hierarchy process (FAHP) to determine the subjective weights of experts. Through mechanisms such as fuzzy number defuzzification and weight consistency verification, it improves the rationality and flexibility of weight setting and adapts to different review scenarios. The system also has online self-learning capabilities, supports dynamic updates of structural causal weights with project data, and enhances the adaptability and long-term effectiveness of the model. The final output results include not only comprehensive scores and grade judgments, but also visually display the impact paths and indicator contributions, providing decision makers with intuitive and reliable support. Overall, the present invention is superior to existing technologies in terms of accuracy, intelligence, interpretability, and adaptability, and is suitable for a variety of application scenarios such as government science and technology project approval, enterprise technology evaluation, and investment and financing review. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings: Figure 1 This is a flowchart of a technology project evaluation method and system in an embodiment of the present application.

[0017] Figure 2 This is a schematic diagram of an evaluation index system in a technology project evaluation method and system in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The following is a summary of the examples of the present invention. Figure 1-Figure 2 A clear and complete description of the technical solutions in the embodiments of the present invention is provided. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0019] A science and technology project evaluation system: including 1. Data collection templates and interfaces: These are used to collect various types of customer data, including financial data. These templates can directly connect to enterprise financial software databases or support the import of common financial statements (such as Excel balance sheets, income statements, and cash flow statements). These templates can analyze and extract key financial indicators, such as operating income, net profit, total assets, and R&D expenses.

[0020] 2. Central processing module: It is used to monitor, analyze and process the collected data; calculate the relevance and importance indicators of the data, analyze the relationship between the various data features and their influence on the evaluation of scientific and technological projects.

[0021] 3. Data Storage Module: This module is used to store raw collected data, processed data, intermediate data during model training, and final model parameters. It utilizes an efficient database management system to ensure data security, integrity, and scalability. It enables categorized storage and rapid retrieval of data, facilitating subsequent data query, analysis, and model updates.

[0022] 4. User Interaction Module: Provide a user-friendly interface to facilitate interaction between the system and the technology project department staff and clients. Staff can use this interface to enter basic client information, initiate data collection and evaluation processes, and view evaluation results and model recommendations. Clients can view relevant evaluation information and improvement suggestions. The interface design should be simple and intuitive, with operational guidance and prompts, such as the use of visual charts to display data and evaluation results, to enhance user experience and understanding of system outputs.

[0023] 5. Report Generation Module: Automatically generates detailed technology project evaluation reports based on the evaluation results. These reports include a client data overview, project evaluation scores and ratings, strengths and weaknesses analysis, model-based improvement recommendations (such as patent application strategies, R&D investment adjustments, and financial indicator optimization paths), and future project development forecasts. The reports are formatted in a standardized manner and can be exported to common document formats (such as PDF and Word) for easy archiving and sharing.

[0024] 6. Model Update Module: As time passes and data accumulates, this module initiates the model update process when model performance degrades or when significant changes occur in the business environment (such as regulatory adjustments or industry technological innovations). It recollects the latest data and updates and redeploys the existing model according to the data evaluation and model training optimization process, ensuring that the system can always adapt to new situations and provide accurate and effective evaluation results.

[0025] A method for evaluating a science and technology project comprises the following steps: 1. Data Collection By collecting the required data from the enterprise's internal systems and external databases. Including: 1. Financial data, used to reflect the financial health of the enterprise's ability to support scientific and technological projects, including but not limited to: Operating income: the total annual revenue from the company's main business; Net profit margin: net profit divided by operating income, reflecting profitability; Total assets: all assets owned by the enterprise; Current ratio: Current assets ÷ current liabilities, measures short-term debt repayment ability; Debt-to-asset ratio: total liabilities divided by total assets, reflecting financial risk; R&D expense ratio: annual R&D investment divided by operating income, reflecting the importance a company places on R&D; Data source: Directly connect to the enterprise ERP / financial system; support uploading of standardized reports (such as balance sheets, income statements, and cash flow statements in Excel format); the system automatically extracts and verifies core financial indicators.

[0026] 2. Intellectual property data, focusing on measuring the company's technological innovation capabilities and the degree of intellectual property protection: Total number of patents: the total number of patents currently applied for by the enterprise; Percentage of valid patents: Number of valid patents divided by total number of patents, reflecting patent maintenance capabilities and effectiveness; Number of core patents: the number of patents that are highly relevant to the core technology of this project; Patent citation count: The frequency with which a company's patents are cited by others, indirectly reflecting its technological influence; Data source: Connected to a third-party patent analysis platform; the system supports keyword search, patent number query, and automatic identification of patent types and fields of application; it also has a built-in patent quality scoring module that can mark high-value patents (such as those that have been cited multiple times).

[0027] 3. Human resources data, used to measure the overall strength of the team and the stability of project implementation: Total number of R&D personnel: the total number of personnel engaged in R&D positions in the enterprise; Ratio of personnel with senior professional titles / master's and doctoral degrees: the proportion of personnel with high academic qualifications or high professional titles in the R&D team; Team stability indicators: such as the average turnover rate of R&D personnel in the past three years, reflecting personnel mobility; Core members' technical background, project experience, etc.: optional collection, used for building the team capability model; Data sources: enterprise personnel system (HR system); project application form, employee files, manually uploaded resume documents or expert lists.

[0028] 4. Historical science and technology project execution and award-winning data to measure the company's past execution capabilities and the impact of its technological achievements: Project success rate: the number of successfully completed projects divided by the total number of projects approved over the years, reflecting the ability to implement the project; Number of award-winning projects: the number of projects that have won national, provincial, and municipal science and technology awards; Relevance score of the winning project: whether the winning project is highly consistent with the field of the current project; Data sources: enterprise science and technology management system; matching government public science and technology project platforms (such as the achievement disclosure systems of provincial and municipal science and technology departments); The data collection is achieved through the system's built-in data collection template and collection interface; the template unifies field formats, unit standards and logical verification rules; the interface supports multiple data source types: API docking (such as interfaces with financial systems and patent libraries); file upload (such as Excel, CSV, PDF structured analysis) and manual assisted entry; after the data enters the system, preliminary format checks and field integrity checks are automatically performed.

[0029] 2. Data Processing Step 1: Data cleaning: (1) Missing value processing: Field identification: traverse each field in the data table and calculate the missing rate; Numeric fields: If the missing rate is ≤ 20%, the industry mean or median is used for filling; if the missing rate is high, its impact on the assessment accuracy is evaluated and then removed as appropriate; Categorical fields: Use the mode to fill in; or use the conditional probability distribution of related variables to make inferences and fill in; (2) Outlier processing: Identification method: Use the standard deviation method (Z-score): |Z| ≥ 3 is considered abnormal; Processing method: directly eliminate those with input errors or abnormal formats; perform Winsor correction (truncation method) on those with extreme values ​​but reasonable values; Step 2: Data Transformation Numerical variable transformation: For variables with skewed distribution (such as operating income, number of patents, etc.), logarithmic transformation (log(x+1)) is performed to alleviate long-tail distribution and improve model stability; processing method: .

[0030] (2) Categorical variable encoding: All non-numeric variables (such as "industry" and "patent type") are uniformly one-hot encoded. For example, for the industry field: "manufacturing", "software", "pharmaceutical" → converted into three binary fields; Step 3: Data normalization The purpose is to make data of different dimensions and magnitudes comparable on a unified scale; the processing methods include: Min-Max normalization (suitable for non-normal distribution data): Where X is the original data, Xmin and Xmax are the minimum and maximum values ​​of the data field respectively, and Xnew is the normalized result.

[0031] Z-score standardization (suitable for approximately normally distributed data): .

[0032] Step 4: Data Correlation Analysis (1) Pearson correlation coefficient analysis: Calculate the Pearson correlation coefficient matrix between all pairs of numeric fields. The correlation coefficient is defined as: .

[0033] The correlation coefficient ranges from -1 to 1. The closer the absolute value is to 1, the stronger the correlation. Positive correlation means that the two variables change in the same direction, and negative correlation means that they change in the opposite direction.

[0034] Draw a correlation matrix: Display the correlation coefficients between all data fields in matrix form, and quickly discover the correlation patterns between data in a visual way.

[0035] 3. Construction of the evaluation indicator system Step 1: Determine the primary evaluation indicators: Based on the characteristics of the science and technology project and the evaluation objectives, determine several main evaluation dimensions, including: a. Innovation: used to evaluate the novelty and breakthrough of the project in terms of technical path, concept, product form, etc. b. Feasibility: used to evaluate the project's practical implementation capabilities and risk control capabilities from the perspectives of technology, funding, and operations; c. Market prospects: used to examine the scale, growth, competitive situation and sustainable development potential of the market to which the project is targeting d. Team Capabilities: used to analyze the team member structure, technical background, stability, and project execution capabilities of the project implementation team; e. Financial status: used to measure the financial health of the project unit or undertaker and its ability to support the project; These five first-level indicators constitute the factor set .

[0036] Step 2: Decompose the second-level and lower evaluation indicators. For each first-level evaluation indicator, further subdivide it into specific second-level evaluation indicators to more accurately measure the performance of the project in various aspects. Specifically include: (1) Innovation index (u1): Technological novelty: used to indicate whether the technology used is proposed or applied for the first time, and whether it is disruptive; Technological advancement: used to indicate the degree of advancement compared with international / industry advanced levels; Technological differentiation: used to indicate the differences from existing products or solutions in terms of technical principles, functions, performance, etc. Application scalability: used to indicate whether the project results can be replicated and promoted, and whether they have industry platform capabilities; (2) Feasibility index (u2): Technology maturity: used to indicate the stage of the technology life cycle: concept, prototype, pilot, mass production, etc. Cost controllability: This indicates whether R&D and production costs are controllable and whether mass production is economical. Resource availability: used to indicate whether the equipment, technology, personnel, etc. required for implementation are easy to obtain; Risk prediction and management capabilities: This indicates whether the project has identified major risks and developed effective response plans; (3) Market outlook indicator (u3): Market demand intensity: used to indicate whether there are significant "pain point" needs for target users; Market size: used to indicate the current and next five years' industry market capacity; Market competition level: used to indicate the existing competition situation and whether there are obvious entry barriers; Business expansion capability: used to indicate whether the results can be extended to other fields / markets; (4) Team capability indicator (u4): Ratio of senior R&D talents: used to indicate the proportion of personnel with master's and doctoral degrees and senior professional titles in the R&D team; Team stability: used to indicate the average turnover rate of R&D team members in the past three years; Team collaboration and project experience: This indicates whether core members have worked on similar projects together in the past and whether there is a basis for collaboration. Academic and Patent Achievements: used to indicate the number and quality of scientific research papers and patent authorizations in recent years; (5) Financial status indicator (u5): Debt-to-asset ratio: an important indicator used to measure a company's financial risk Net profit margin: used to reflect profitability R&D expense ratio: a direct reflection of R&D investment intensity Cash flow health: used to indicate whether the company has a stable source of operating cash flow Step 3: Indicator quantifiability definition and data mapping Each secondary indicator is mapped through observable enterprise data fields; Example: “R&D expense ratio” → data field: R&D expense ÷ operating income "Proportion of senior talents" → Number of personnel with a master's degree or above / senior professional title ÷ total number of R&D personnel; "Market demand intensity" → Analyze and score the keyword frequency in the market report using the NLP model. Step 4: Define and number factor sets The standard factor set U that takes the first-level and second-level indicator numbers as model input is: .

[0037] in It represents the mth second-level sub-item under the nth first-level indicator and is an important component of the subsequent fuzzy relationship matrix R and GAT node input.

[0038] 4. Structural Causal Modeling This step introduces the cognitive graph structure and graph neural network embedding mechanism based on the traditional fuzzy comprehensive evaluation model to model the nonlinear causal relationship structure between evaluation indicators, thereby improving the accuracy, dynamic adaptability and interpretability of project evaluation results. This step includes the following sub-processes: Step 1: Build a cognitive map of science and technology project indicators (1) Graph node construction: All the first-level and second-level evaluation indicators constructed in step 3 are As a node set V. Each node represents an evaluation indicator item, and the node number corresponds to the factor set definition.

[0039] (2) Graph edge construction (causal relationship edge construction) In the process of constructing a cognitive map of scientific and technological project indicators, in addition to defining each indicator as a node, the key lies in reasonably defining the directed edge relationship between indicators, that is, constructing a "causal path" between indicators to indicate "whether the change of a certain indicator will have a direct or indirect impact on another indicator."

[0040] The construction of this edge mainly consists of two parts: data-driven statistical correlation edges and expert knowledge-based causal rule edges. The two complement each other to form a graph structure with explanatory and generalization capabilities.

[0041] a. Build edges based on statistical data ( ) Data source: Take the original indicator data of all evaluated projects in the historical science and technology project database; each indicator is regarded as a variable dimension to form a data matrix , where s is the number of samples and n is the number of indicators.

[0042] Correlation calculation: Use the Pearson correlation coefficient to calculate the degree of linear correlation between each pair of indicators: Edge weight generation rule: If When it is greater than the preset threshold, the node An edge is generated between , indicating that the change of variable j has an impact on i.

[0043] Sparse processing: To control the number of edges and prevent the graph from being too dense, each node can be set to accept at most the top k most strongly related edges (top-k restriction).

[0044] b. Building edges based on expert knowledge ( ) Expert input method: Provide a visual configuration interface or structured template, and invite field experts or policymakers to mark the "indicator impact path"; For example: "R&D investment ratio" → "Innovation"; "Proportion of highly educated team members" → "Feasibility"; "Technological differentiation" → "Market prospects".

[0045] Edge weight assignment method: Causal strength is quantified based on expert perception as follows (using triangular fuzzy numbers or fixed real values): (3) Adjacency matrix synthesis: Combine statistical edges and expert edges to form a complete adjacency matrix: in, is the fusion ratio (configurable or adaptively adjusted according to the sample size); if the same pair of indicators (i, j) are associated in both types of graphs, a weighted average or expert weight priority retention strategy is adopted.

[0046] Step 2: Construct node feature matrix Each indicator node Construct a d-dimensional feature vector , forming the input matrix: , Feature dimensions include but are not limited to: the actual value of the indicator in the current sample (normalized); the mean and standard deviation of the indicator in historical samples; the expert scoring weight corresponding to the indicator; and the one-hot encoding of the first-level indicator category to which it belongs.

[0047] Step 3: Graph Attention Network (GAT) training model Cognitive Map Input the multi-head graph attention network model (GAT) to learn causal weights. The GAT propagation formula is as follows: in, represents the representation vector of node i in layer l; represents the neighbor set of node i; : represents the weight matrix of the k-th attention head (trainable parameters); represents the attention weight of node j to node i (normalized causal strength); Represents the concatenation operation of each attention head; σ is the activation function LeakyReLU.

[0048] The calculation formula of attention coefficient is: in, and Represents the input feature vectors of nodes i and j respectively; W represents the trainable weight matrix; It represents feature concatenation, which connects the transformed features of node i and its neighbor j into an overall vector to represent the "interaction relationship between them"; Learnable scoring vectors, is the activation function, is the final attention weight.

[0049] This mechanism means that when updating the representation of node i, GAT will automatically "pay attention" to its more closely related neighbor node j, thereby achieving adaptive learning of indicator causal weights.

[0050] Step 4: Output structural weight matrix After training is completed, the attention weights of the final layer are extracted from the GAT model and normalized to form a structural causal matrix: , in, The matrix represents the structural weighted influence of indicator j on indicator i. This matrix reflects the dynamic causal network structure between indicators. The weights are learned from project data and do not require subjective expert assignment. It is highly interpretable and automatically updated with data.

[0051] Step 5: Fusion of structural causal weights and fuzzy relationship matrix GAT output weight matrix Combined with the original fuzzy membership matrix R, a structure-enhanced fuzzy relationship matrix is ​​generated : .in Indicates row-aligned matrix multiplication or weighted fusion; The result is a new fuzzy relationship matrix , which is used for the next fuzzy comprehensive evaluation. Compared with the traditional method, this structure enables the model to not only consider the membership degree of each indicator itself during evaluation, but also introduce the causal influence of other indicators on it, forming a structural linkage evaluation mechanism.

[0052] 5. Determine indicator weights: Calculate the weight coefficients of each evaluation indicator by combining expert judgment with fuzzy mathematics. The fuzzy analytic hierarchy process (FAHP, FuzzyAHP) is used to address the issues of fuzzy scoring and large differences in expert preferences in traditional AHP, making weight calculation more scientific, flexible, and tolerant of inconsistencies. This includes the following sub-steps: Step 1: Build a judgment hierarchy The overall evaluation goal ("the quality of scientific and technological project evaluation") is set as the top-level goal; the second level is the first-level indicators (innovation, feasibility, market prospects, team capabilities, and financial status); the third level is the second-level indicators under each first-level indicator; forming a standard three-layer AHP hierarchical structure (which can be further expanded to a four-layer structure).

[0053] Step 2: Construct the fuzzy judgment matrix (1) Collection of expert ratings: Several industry experts gave relative importance evaluations based on the "pairwise comparison method" and used fuzzy scales instead of integer scores, selecting triangular fuzzy numbers to represent "fuzzy judgment preferences": (2) Construct fuzzy judgment matrix: For each pair of indicators , construct a triangular fuzzy number matrix: Among them: If the experts believe Significantly stronger than ,but =(3,5,7); then =(1 / 7, 1 / 5, 1 / 3).

[0054] Step 3: Fuzzy weight calculation (1) Calculate the fuzzy weight sum of each row: (2) Normalized fuzzy weight: Get the fuzzy weight vector , which is still in the form of triangular fuzzy numbers.

[0055] Step 4: Defuzzification: In order to facilitate the subsequent fuzzy relationship matrix operation, the fuzzy weight vector is converted into a clear numerical weight: Then use the maximum membership method to get the final normalized clear weight vector ,satisfy .

[0056] Step 5: Consistency Check Due to the subjectivity of experts, the judgment matrix may have inconsistencies, so the consistency indicators CI and CR are calculated first: , .

[0057] in, is the maximum eigenvalue of the judgment matrix; The average random consistency index corresponding to the n-order matrix; if CR < 0.1, the judgment matrix is ​​considered to have acceptable consistency.

[0058] 6. Fuzzy comprehensive evaluation calculation This step integrates the constructed weight vector with the structured fuzzy relationship matrix, and outputs the membership score vector of each scientific and technological project at each evaluation level through the weighted operation method in fuzzy mathematics, and determines the final level evaluation result based on this.

[0059] Step 1: Input parameter preparation, including (1) Weight vector W: From step 5, the indicator weight is determined in the form of: .

[0060] (2) Structural Enhanced Fuzzy Relationship Matrix :From step 4 structural causal modeling, the form is: Among them, the number of rows n represents the number of indicators; the number of columns m represents the number of preset grade divisions (such as excellent, good, medium, qualified, and unqualified).

[0061] Step 2: Fuzzy weighted comprehensive calculation The fuzzy weighted average (FWA) model is used to calculate the fuzzy comprehensive evaluation vector B: Among them, each Indicates that the item belongs to the level The membership degree of the project; B means the fuzzy evaluation vector of the project at each level.

[0062] Step 3: Level Judgment Strategy According to the fuzzy evaluation result B, one of the following strategies is used to determine the final project level: (1) MaxMembership Principle: The grade obtained is Maximum corresponding level; (2) Weighted average method (for continuous scoring): The numerical scores corresponding to the preset levels are V=[100,80,60,40,20]; Calculate the weighted score: Step 4: Fuzzy evaluation result output The final output is as follows: project grade: such as "excellent"; fuzzy evaluation vector B: such as [0.12, 0.40, 0.25, 0.15, 0.08]; grade explanation: such as "the project has outstanding performance in innovation and market prospects, and has good development potential"; recommendation: such as "priority support", "it is recommended to apply after improvement", etc.

[0063] Step 5: Structured output interface: The evaluation results can be output in the following formats: JSON format result structure; Word / PDF automatically generates project evaluation reports; system API response results are used for subsequent links such as intelligent approval and archiving.

[0064] 7. Result Analysis and Visual Interpretation After completing the fuzzy comprehensive evaluation (step six) and obtaining the comprehensive score and grade, this step further conducts an explanatory analysis of the evaluation results from multiple dimensions, including ranking the contribution of key indicators, graphical display of causal paths, scoring trend deduction, risk and recommendation generation, etc., and supports the automatic generation of evaluation reports.

[0065] Step 1: Indicator contribution analysis (1) Contribution calculation: Analyze the impact of each indicator on the final evaluation level based on the following two information fusion calculations: Indicator weight ; Contribution rate of fuzzy membership to the final grade (from the structure-enhanced fuzzy relationship matrix); Contribution formula (based on level For example): .

[0066] Get the contribution vector of each indicator to the final level: .

[0067] (2) Contribution ranking and explanation: Arrange C in descending order to obtain the ranking of the main influencing factors; output example: The comprehensive contribution of "Innovation-Technological Novelty" to the grade of "Excellent" is 22.6%; "Financial Status-Debt-to-Asset Ratio" has a smaller impact, accounting for only 6.4%.

[0068] Step 2: Visualize the Structural Causal Graph Based on the attention matrix output during GAT model training , draw a causal influence map between indicators; Node: evaluation indicators (primary / secondary); Directed edge: the intensity of attention between indicators, the thicker the edge, the stronger the influence; Step 3: Evaluate grade trends and compare with similar cases (1) Trend deduction: Introducing historical data and scoring models to simulate the changes in project scores under different parameter assumptions; Example: If the proportion of R&D expenditure increases by 20%, the overall project score will increase by 5 points, and the level will move from "good" to "excellent".

[0069] (2) Similar case comparison: Based on vector similarity calculation (such as cosine similarity), match the most similar cases in historical projects; Step 4: Risk warning and suggestion output Automatically identify potential issues and generate intelligent recommendations based on indicator performance and industry experience. If an indicator is significantly lower than the industry average, generate an early warning. If a critical path influencing factor is weak, recommend enhancements. While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating scientific and technological projects, characterized in that: The following steps are involved: Data collection: Collect basic project information, corporate financial data, intellectual property data, human resources data, and historical scientific and technological project data through data collection templates and interfaces; Data processing: missing value filling, outlier processing, data conversion, standardization, and correlation analysis are performed on the collected data in sequence to obtain a standardized feature matrix; Construction of evaluation index system: Based on the project evaluation requirements, a hierarchical factor set consisting of five primary indicators, including innovation, feasibility, market prospects, team capabilities and financial status, and their secondary indicators is established; Structural causal modeling: construct an indicator cognitive map with the factor set as nodes; generate weighted directed edges based on historical data correlation and expert rules to obtain an adjacency matrix; The node feature matrix and the adjacency matrix are input into the graph attention neural network GAT for training, and the structural causal weight matrix is ​​output. ; Combined with the original fuzzy membership matrix R, a structure-enhanced fuzzy relationship matrix is ​​formed. ; Determination of indicator weights: Use fuzzy analytic hierarchy process (FAHP) to construct a fuzzy judgment matrix and defuzzify it to obtain the weight vector W; Fuzzy comprehensive evaluation: using weight vector W and structure-enhanced fuzzy relationship matrix Calculate the comprehensive evaluation vector B and determine the project level according to the maximum membership principle; Result analysis and visual interpretation: Generate causal maps, risk warnings, and automated assessment reports based on indicator contribution, GAT attention coefficient, and similar case comparisons.

2. A scientific and technological project evaluation method according to claim 1, characterized in that: In the data processing step, data conversion includes: performing logarithmic transformation on skewed distribution data; and performing one-hot encoding on categorical data.

3. A scientific and technological project evaluation method according to claim 1, characterized in that: In the structural causal modeling step, the indicator cognitive map construction includes: taking the evaluation indicators as nodes, constructing weighted directed edges based on the historical data correlation and expert knowledge, and forming a directed graph structure that reflects the causal relationship between indicators.

4. A scientific and technological project evaluation method according to claim 1, characterized in that: In the structural causal modeling step, the propagation formula of the graph attention neural network GAT is: Where, represents the representation vector of node i in layer l; represents the neighbor set of node i; : represents the weight matrix of the k-th attention head (trainable parameters); represents the attention weight of node j to node i (normalized causal strength); Represents the concatenation operation of each attention head; σ is the activation function LeakyReLU.

5. A scientific and technological project evaluation method according to claim 1, characterized in that: In the structural causal modeling step, the structural causal weight matrix The expression is: , where It represents the structural weight influence strength of indicator j on indicator i.

6. A scientific and technological project evaluation method according to claim 1, characterized in that: The indicator weight determination step includes: building a judgment hierarchy, constructing a fuzzy judgment matrix, calculating fuzzy weights, defuzzification processing and consistency testing.

7. A scientific and technological project evaluation method according to claim 6, characterized in that: The judgment hierarchy includes a target layer, a criterion layer and a solution layer; the target layer is for evaluating the quality of scientific and technological projects, the criterion layer is a first-level indicator, and the solution layer is a second-level indicator.

8. A scientific and technological project evaluation method according to claim 6, characterized in that: The construction of the fuzzy judgment matrix includes: multiple experts make "two-by-two comparisons" on each pair of indicators and give importance judgments in the form of language quantification; convert the language judgments into triangular fuzzy numbers , construct the fuzzy judgment matrix .

9. A system for executing the scientific and technological project evaluation method according to any one of claims 1 to 8, characterized in that: include: Data collection templates and interfaces: used to collect customer technology project-related data and patent information data; Central processing module: It is used to monitor, analyze and process the collected data; Data storage module: It is used to store the original collected data, processed data, intermediate data during model training, and final model parameters; User interaction module: It is used to provide a friendly user interface to facilitate the interaction between scientific and technological project staff and customers and the system; Report generation module: It is used to automatically generate detailed science and technology project evaluation reports based on the evaluation results; Model update module: It is used to start the model update process.

Images