A method and system for intelligent evaluation and risk early warning of a science and technology project based on big data and AI

Abstract

This invention discloses a method and system for intelligent assessment and risk warning of science and technology projects based on big data and AI. The method includes: collecting multi-source heterogeneous science and technology project data to generate a normalized multimodal project feature set; based on the normalized multimodal project feature set, using a self-supervised learning algorithm to automatically extract technical entities, R&D teams, resource elements, and their dynamic relationships, constructing a domain knowledge graph that evolves over time; inputting the domain knowledge graph into a pre-trained dynamic graph neural network risk propagation model to output multi-level risk warning signals for a preset future period; and generating an adaptive project intervention strategy set based on the multi-level risk warning signals using a multi-objective optimization algorithm to achieve intelligent assessment and early warning optimization of science and technology projects. Using this invention, deep semantic association and risk propagation simulation of the entire lifecycle of science and technology project data can be achieved, improving the accuracy and foresight of risk warnings.

Description

Technical Field

[0001] This invention belongs to the field of AI technology, specifically a method and system for intelligent assessment and risk warning of scientific and technological projects based on big data and AI. Background Technology

[0002] Science and technology project evaluation and risk warning are core components of scientific research management and industrial innovation, directly impacting resource allocation efficiency and R&D success rates. Currently, project evaluation primarily relies on expert review and structured indicator reporting, resulting in a single data source and outdated updates, making it difficult to comprehensively capture dynamic changes and potential risks during the R&D process. With the development of big data and artificial intelligence technologies, some systems have begun to attempt to integrate multi-source information; however, when processing unstructured data (such as R&D logs and experimental images), semantic fragmentation and feature heterogeneity remain issues, leading to low data fusion and insufficient correlation mining capabilities.

[0003] Existing risk warning methods are mostly based on static indicator thresholds or single-dimensional trend analysis, neglecting the complex dynamic relationships between project elements. In actual R&D networks, risk factors such as technical bottlenecks, team changes, and resource constraints often transmit to each other through implicit paths, creating a cumulative amplifying effect. Traditional methods struggle to simulate such complex transmission mechanisms, limiting the accuracy and timeliness of warnings. Furthermore, existing warning systems rely heavily on human experience when generating intervention strategies, lacking the ability to comprehensively optimize for multiple objective constraints (such as cost, timeline, and technical feasibility), making adaptive control difficult. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent assessment and risk warning of science and technology projects based on big data and AI, so as to overcome the shortcomings of the existing technology, realize deep semantic association and risk propagation simulation of data throughout the entire life cycle of science and technology projects, and improve the accuracy and foresight of risk warning.

[0005] One embodiment of this application provides a method for intelligent assessment and risk warning of technology projects based on big data and AI, the method comprising: Collect multi-source heterogeneous science and technology project data, including structured application data and unstructured R&D process data, and transform unstructured text, image and time-series log data into unified semantic vector representations through cross-modal alignment technology to generate a normalized multimodal project feature set; Based on the normalized multimodal project feature set, a self-supervised learning algorithm is used to automatically extract technical entities, R&D teams, resource elements and their dynamic relationships to construct a domain knowledge graph that evolves over time. The domain knowledge graph is input into a pre-trained dynamic graph neural network risk propagation model to simulate the transmission path and superposition effect of multiple risk factors in the project network, and output multi-level risk warning signals within a preset period in the future. Based on the multi-level risk warning signals, an adaptive project intervention strategy set is generated using a multi-objective optimization algorithm. The strategy set covers resource allocation adjustment, technical route correction, and collaborative key point reorganization, so as to realize intelligent evaluation and early warning optimization of science and technology projects.

[0006] Optionally, the collection of multi-source heterogeneous science and technology project data includes structured application data and unstructured R&D process data. Cross-modal alignment technology is used to transform unstructured text, images, and time-series log data into a unified semantic vector representation, generating a normalized multimodal project feature set, including: By extracting structured application form data and unstructured R&D documents, design drawings and time sequence operation logs through API interfaces and data acquisition engines, the original multi-source heterogeneous dataset is generated. Data cleaning and standardization are performed on the original multi-source heterogeneous dataset. Data integrity verification and outlier correction algorithms are used for structured data. Named entity recognition technology is used to extract key entities for unstructured text data. Image enhancement algorithms are used to improve the clarity of image data. Timestamp alignment algorithms are used to unify the time base for time-series log data, generating cleaned and standardized multimodal data. The cleaned and standardized multimodal data is input into the cross-modal alignment coding network. The multimodal Transformer model with attention mechanism maps text semantics, image features and temporal patterns to a unified high-dimensional semantic space to generate an initial alignment semantic vector. The initial aligned semantic vector is processed by feature normalization. Layer normalization technique is used to eliminate the dimensional differences of features of different modalities. Key information is retained by feature dimensionality reduction algorithm, and finally a normalized multimodal item feature set is generated.

[0007] Optionally, based on the normalized multimodal project feature set, a self-supervised learning algorithm is used to automatically extract technical entities, R&D teams, resource elements, and their dynamic relationships to construct a domain knowledge graph that evolves over time, including: Technical term vectors, R&D personnel feature vectors, and resource identifier vectors are extracted from the normalized multimodal project feature set. A self-supervised contrastive learning algorithm is used to train the entity recognition model to generate an initial entity set of technical entities, R&D teams, and resource elements. Based on the initial entity set, a self-supervised relationship prediction task is designed. The potential association patterns between entities are learned through the masked entity prediction algorithm to generate an entity relationship prediction model. The entity relationship prediction model is used to analyze the temporal evolution patterns of multimodal project feature sets, identify the creation, strengthening and disappearance processes of entity relationships under different time slices, and generate dynamic relationship evolution sequences. By integrating the initial entity set and the dynamic relationship evolution sequence, a temporal knowledge graph construction algorithm is used to organize entities and relationships according to the time dimension, and finally construct a domain knowledge graph that evolves over time.

[0008] Optionally, the step of inputting the domain knowledge graph into a pre-trained dynamic graph neural network risk propagation model to simulate the transmission paths and superposition effects of multiple risk factors in the project network, and outputting multi-level risk warning signals within a preset future period, includes: The domain knowledge graph that evolves over time is represented as a dynamic graph sequence data, with each time step containing entity node features and relation edge weights, resulting in a dynamic graph sequence for input. Load a pre-trained dynamic graph neural network risk propagation model. This model adopts a spatiotemporal graph convolutional network architecture, which can capture the spatiotemporal dependencies between nodes in the graph for risk propagation calculation. By inputting dynamic graph sequences into a dynamic graph neural network risk propagation model, the transmission process of technical risks, resource risks, and team collaboration risks in the project network is simulated through a multi-hop message passing mechanism, generating risk propagation simulation results. Based on the results of risk transmission simulation, the risk superposition effect analysis algorithm is used to predict the degree of risk accumulation of each entity node in the future within a preset period. The risk levels are divided into high, medium and low risk levels according to the risk level threshold, and finally multi-level risk warning signals are output.

[0009] Optionally, based on the multi-level risk warning signals, an adaptive project intervention strategy set is generated using a multi-objective optimization algorithm. This strategy set encompasses resource allocation adjustments, technical route corrections, and collaborative key-point reorganization to achieve intelligent evaluation and early warning optimization of science and technology projects, including: Analyze multi-level risk warning signals, extract high-risk entity nodes and their relationships, construct a multi-objective optimization problem for risk mitigation, with objective functions including maximizing risk reduction, minimizing resource consumption, and minimizing the impact on project schedule, and generate a multi-objective optimization model. A non-dominated sorting genetic algorithm is used to solve the multi-objective optimization model. The Pareto optimal solution set is searched through population evolution iteration to generate a set of candidate intervention strategies. Based on the set of candidate intervention strategies, a strategy feasibility assessment index system is constructed, including assessments of technical feasibility, resource accessibility, and team execution capability, and strategy feasibility assessment results are generated. Based on the feasibility assessment results of the strategies, the optimal resource allocation adjustment plan, technical route correction suggestions, and collaborative key point reorganization plan are selected and integrated into an adaptive project intervention strategy set. The effect is verified through strategy execution simulation, and finally, intelligent evaluation and early warning optimization of science and technology projects are achieved.

[0010] Another embodiment of this application provides a smart assessment and risk warning system for technology projects based on big data and AI, the system comprising: The data acquisition module is used to collect multi-source heterogeneous science and technology project data, including structured application data and unstructured R&D process data. It also uses cross-modal alignment technology to transform unstructured text, images, and time-series log data into a unified semantic vector representation, generating a normalized multimodal project feature set. The module is used to automatically extract technical entities, R&D teams, resource elements and their dynamic relationships based on the normalized multimodal project feature set and using a self-supervised learning algorithm to construct a domain knowledge graph that evolves over time. The simulation module is used to input the domain knowledge graph into a pre-trained dynamic graph neural network risk propagation model to simulate the transmission path and superposition effect of multiple risk factors in the project network, and output multi-level risk warning signals within a preset period in the future. The generation module is used to generate an adaptive project intervention strategy set based on the multi-level risk warning signals using a multi-objective optimization algorithm. The strategy set covers resource allocation adjustment, technical route correction, and collaborative key point reorganization, so as to realize intelligent evaluation and early warning optimization of science and technology projects.

[0011] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0012] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0013] Compared with existing technologies, the present invention provides a method for intelligent assessment and risk warning of science and technology projects based on big data and AI, which can realize deep semantic association and risk propagation simulation of data throughout the entire life cycle of science and technology projects, thereby improving the accuracy and foresight of risk warning. Attached Figure Description

[0014] Figure 1 A hardware structure block diagram of a computer terminal for a method of intelligent assessment and risk warning of science and technology projects based on big data and AI, provided for an embodiment of the present invention; Figure 2 A flowchart illustrating a method for intelligent assessment and risk warning of technology projects based on big data and AI, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a technology project intelligent assessment and risk warning system based on big data and AI, provided as an embodiment of the present invention. Detailed Implementation

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] The present invention first provides a method for intelligent assessment and risk warning of technology projects based on big data and AI. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0017] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a method of intelligent assessment and risk warning of science and technology projects based on big data and AI, provided as an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0018] See Figure 2 The embodiments of the present invention provide a method for intelligent assessment and risk warning of technology projects based on big data and AI, which may include the following steps: S201 collects multi-source heterogeneous science and technology project data, including structured application data and unstructured R&D process data, and transforms unstructured text, image and time-series log data into a unified semantic vector representation through cross-modal alignment technology to generate a normalized multimodal project feature set. Specifically, structured application form data and unstructured R&D documents, design drawings and time sequence operation logs can be extracted through API interfaces and data acquisition engines to generate original multi-source heterogeneous datasets; The core of this step is to achieve comprehensive collection of data from multiple sources and types of science and technology projects. Through the collaborative work of API interfaces and data collection engines, it covers all data from project application to the R&D process, ultimately forming a raw, multi-source, heterogeneous dataset. The specific implementation method is as follows: The API interface is primarily used to connect to structured data sources such as official project application systems and R&D management platforms. It employs the RESTful communication protocol and has a fixed call frequency of once per hour to avoid congestion caused by frequent calls. A timeout and retry mechanism is also implemented, with a 3-second timeout and 3 retries to ensure stable extraction of structured data. The extracted structured application form data covers core basic information about science and technology projects, specifically including fields such as project name, applicant organization, R&D field, R&D budget, core technical indicators, R&D cycle, number of participants, and collaborating organizations. The data is uniformly formatted as CSV for easy batch processing.

[0019] The data acquisition engine is mainly used to extract unstructured data. It adopts a combination of distributed crawling and local collection. On the one hand, it crawls unstructured R&D documents uploaded to the R&D management platform (such as technical feasibility reports, mid-term progress reports, test reports, patent application documents, etc.), supporting multiple text formats such as PDF and DOCX. On the other hand, it connects to R&D equipment and collaboration platforms to collect unstructured design drawings (such as product structural design drawings, circuit schematics, component assembly drawings, etc., in formats including PNG, JPG, and CAD) and time-series operation logs (such as R&D equipment operation logs, team collaboration platform operation logs, server access logs, code submission logs, etc.).

[0020] The time-series operation logs are collected in real-time monitoring mode at a frequency of 100 milliseconds per collection, ensuring the capture of detailed information for each operation. Each log entry includes core information such as operation time, operator, operation content, device identifier, and operation status. In the example, structured data from a technology project is extracted via an API interface: Project Name = "Research and Development of Intelligent Risk Control System Based on Big Data", Applicant = "XX Technology Co., Ltd.", R&D Budget = 8 million yuan, R&D Period = 24 months, Participants = 15 people. Unstructured data for the project is extracted via a data acquisition engine: the R&D document is the "Technical Feasibility Report of Intelligent Risk Control System" (PDF format), the design drawing is the "Risk Control System Architecture Design Diagram" (PNG format), and the time-series operation logs include entries such as "2024-05-01 10:00:00, Zhang San, submitted risk control algorithm code, Server A, successful". All extracted structured and unstructured data are aggregated, categorized and stored according to data type, generating an original multi-source heterogeneous dataset. This dataset covers multi-dimensional data throughout the project's entire lifecycle, laying the foundation for subsequent processing.

[0021] Data cleaning and standardization are performed on the original multi-source heterogeneous dataset. Data integrity verification and outlier correction algorithms are used for structured data. Named entity recognition technology is used to extract key entities for unstructured text data. Image enhancement algorithms are used to improve the clarity of image data. Timestamp alignment algorithms are used to unify the time base for time-series log data, generating cleaned and standardized multimodal data. The core of this step is to eliminate noise, bias, and format differences in the original data. Through targeted cleaning and standardization, data of different types and formats are brought to a unified standard, facilitating subsequent cross-modal alignment. The specific implementation method is as follows: The processing of structured data primarily employs data integrity verification and outlier correction algorithms. Data integrity verification checks for missing core fields, with a preset threshold of 5%. If the missing rate of a core field (project name, R&D budget, R&D cycle) exceeds 5%, the data entry is directly removed. If the missing rate is below 5%, appropriate methods are used for filling in missing values: numeric fields (R&D budget, number of participants) are filled with the mean, and character fields (applicant unit, R&D field) are filled with the mode. The outlier correction algorithm uses the 3σ principle. It first calculates the mean μ and standard deviation σ of the field, and identifies data exceeding the range [μ-3σ, μ+3σ] as outliers. Outliers are then corrected using the median to prevent them from affecting subsequent analysis. In the example, the R&D budget for a certain project in the structured data is filled in as 80 million yuan. The mean of this field is μ=7.5 million yuan, the standard deviation is σ=2 million yuan, and the range of [μ-3σ,μ+3σ] is [1.5 million,13.5 million yuan]. 80 million yuan is an outlier, so the median of 7.2 million yuan is used to correct it to ensure that the data is reasonable.

[0022] Named entity recognition (NAME) technology is used to process unstructured text data. This technology is used to automatically extract key entities related to the technology project from the text, mainly including technical entities (such as "big data algorithm," "intelligent risk control model," and "blockchain technology"), R&D team entities (such as "Zhang San," "Li Si," and "algorithm R&D group"), resource entities (such as "server," "laboratory equipment," and "R&D funding"), and achievement entities (such as "patent," "software copyright," and "test report"). During the recognition process, a context-based entity recognition model is used, combined with a technology-specific dictionary to improve recognition accuracy. After recognition, the extracted key entities are deduplicated to generate a set of key text entities. In the example, the technical entities "LSTM neural network" and "feature engineering algorithm," the R&D team entity "Wang Wu (algorithm engineer)," the resource entity "2 high-performance servers," and the achievement entity "1 software copyright" are extracted from the "Technical Feasibility Report of Intelligent Risk Control System," completing the standardization processing of the text data.

[0023] Image enhancement algorithms are employed for image data processing, primarily addressing issues such as blurriness, uneven grayscale, and unclear details in scanned design drawings. This improves image quality and facilitates subsequent feature extraction. Histogram equalization is used to adjust the image's grayscale distribution, enhancing contrast and making lines, text, and other details in the drawings clearer. Simultaneously, Gaussian denoising removes noise points while preserving core features. The processed data is uniformly converted to PNG format with a resolution of 1920×1080 pixels. In the example, the original image of a circuit schematic was blurry with unclear lines. After histogram equalization, image contrast increased by 30%, and line edges became sharper. Gaussian denoising reduced noise points by 80%, clearly revealing core information such as circuit nodes and wiring connections, thus standardizing the image data.

[0024] The processing of time-series log data employs a timestamp alignment algorithm. Due to clock deviations across different R&D equipment and collaboration platforms, the timestamps in the time-series logs are inconsistent, with deviations typically ranging from 10 to 50 milliseconds. Therefore, they need to be unified to a common time reference. Using the system's unified reference clock (synchronized with the international standard clock via the NTP protocol, achieving a synchronization accuracy within 1 millisecond) as the standard, the deviation Δt between the original timestamp of each log entry and the reference clock is calculated as: Δt = Reference clock time - Original log timestamp. The correction formula is t_sync = t_original + Δt, where t_sync is the aligned timestamp and t_original is the original log timestamp. In the example, the system base clock time is 1714567890123 milliseconds, and the original timestamp of a certain log entry is 1714567890103 milliseconds. Δt = 20 milliseconds. After correction, t_sync = 1714567890103 + 20 = 1714567890123 milliseconds, achieving synchronization with the base clock. After all logs are corrected, they are sorted in chronological order to generate standardized time-series log data. The cleaning and standardization results of all data types are then summarized to generate cleaned and standardized multimodal data.

[0025] The cleaned and standardized multimodal data is input into the cross-modal alignment coding network. The multimodal Transformer model with attention mechanism maps text semantics, image features and temporal patterns to a unified high-dimensional semantic space to generate an initial alignment semantic vector. The core of this step is to address the semantic gap between different modalities of data (text, image, time series). Through a multimodal coding model, various data types are mapped to a unified high-dimensional semantic space, achieving cross-modal data alignment and generating an initial aligned semantic vector. The specific implementation method is as follows: Cross-modal alignment coding network is a network architecture specifically designed for multimodal data fusion and alignment. It consists of a text coding branch, an image coding branch, a temporal coding branch, and an attention fusion layer. The three coding branches are responsible for processing the standardized data of their respective types, while the attention fusion layer is responsible for fusing the coding results of the three branches and mapping them to a unified high-dimensional semantic space. The network input consists of cleaned and standardized text, image, and temporal data, and the output is an initial alignment semantic vector of a unified dimension. The vector dimension is set to 512 dimensions to balance feature representation capability and computational efficiency.

[0026] A multimodal Transformer model employing an attention mechanism is used as the core encoding model. This model captures the dependencies within single-modal data through self-attention and the relationships between different modalities through cross-attention, achieving deep fusion and alignment of multimodal data. The text encoding branch converts the set of key entities in standardized text into a text sequence, inputting it into the Transformer text encoder. The self-attention mechanism captures the relationships between entities (such as the dependency between "LSTM neural network" and "intelligent risk control model"), generating a 512-dimensional text semantic vector. The image encoding branch inputs enhanced image data into the Transformer image encoder, extracting texture, contour, and other features through convolutional layers. Then, a self-attention mechanism focuses on the core regions of the image (such as key components in design drawings), generating a 512-dimensional image feature vector. The temporal encoding branch converts aligned temporal log data into a temporal sequence, inputting it into the Transformer temporal encoder to capture temporal patterns of operational behavior (such as the timing of code submissions and changes in equipment operating status), generating a 512-dimensional temporal feature vector.

[0027] The core function of the attention fusion layer is to weightedly fuse the encoded vectors from the three branches. Different attention weights are assigned based on the importance of different modalities in evaluating the science and technology project: the text semantic vector has a weight of 0.4, the image feature vector has a weight of 0.3, and the temporal feature vector has a weight of 0.3. These weights can be dynamically adjusted according to the project type. During the fusion process, a cross-attention mechanism is used to calculate the similarity between vectors from different modalities and adjust the weight allocation to ensure that the fused vector comprehensively reflects the core information of the multimodal data. In the example, the text semantic vector, image feature vector, and temporal feature vector of a science and technology project are fused through attention to generate a 512-dimensional initial aligned semantic vector. Each dimension of the vector corresponds to a core feature of the multimodal data; for example, dimensions 1-100 correspond to technical entity features in the text, dimensions 101-256 correspond to design features in the image, and dimensions 257-512 correspond to operational mode features in the temporal data, thus completing the generation of the initial aligned semantic vector.

[0028] The initial aligned semantic vector is processed by feature normalization. Layer normalization technique is used to eliminate the dimensional differences of features of different modalities. Key information is retained by feature dimensionality reduction algorithm, and finally a normalized multimodal item feature set is generated.

[0029] The core of this step is to eliminate dimensional differences and redundant information in the initial aligned semantic vectors. Through normalization and dimensionality reduction, a normalized multimodal item feature set with reasonable dimensions and prominent features is generated, which facilitates subsequent knowledge graph construction and risk assessment. The specific implementation method is as follows: Layer normalization is employed to normalize the features of the initial aligned semantic vector. This technique eliminates dimensional differences between different modal features. Due to the different feature scales of text, images, and time-series data, dimensional differences still exist even after encoding alignment (e.g., the range of text semantic vector values ​​is [-1, 1], while the range of image feature vector values ​​is [0, 255]). Failure to eliminate these differences can lead to subsequent algorithms overemphasizing features with larger values, affecting the accuracy of the analysis results. The formula for layer normalization is x_norm = (x - μ) / σ, where x is a feature value in the initial aligned semantic vector, μ is the mean of all feature values ​​in the vector, and σ is the standard deviation of all feature values. This formula maps all feature values ​​to a uniform range of [-1, 1], eliminating dimensional differences. In the example, the mean μ=0.3 and the standard deviation σ=0.2 of an initial aligned semantic vector, and the feature value x=0.7 in the vector, after normalization, x_norm=(0.7-0.3) / 0.2=2.0. Values ​​outside the range [-1,1] are treated as boundary values ​​to ensure that all feature values ​​are on a uniform scale and complete the normalization process.

[0030] The core purpose of dimensionality reduction algorithms is to preserve key features, remove redundant information, reduce subsequent computational complexity, and avoid the curse of dimensionality. Principal Component Analysis (PCA) is used as the feature dimensionality reduction algorithm. This algorithm maps high-dimensional vectors to a low-dimensional space through linear transformation, retaining principal components with a cumulative variance contribution rate of ≥95%, ensuring that the dimensionality-reduced data retains the core information of the original data. During dimensionality reduction, the covariance matrix of the normalized vector is first calculated, then the eigenvalues ​​and eigenvectors of the covariance matrix are solved, and principal components with larger eigenvalues ​​(cumulative variance contribution rate ≥95%) are selected. The normalized vector is then projected onto these principal components to obtain the dimensionality-reduced feature vectors. In the example, after PCA dimensionality reduction of a 512-dimensional normalized vector, the first 128 principal components are selected, with a cumulative variance contribution rate of 96.2%, preserving the core features of the original vector. The dimensionality of the dimensionality-reduced vector is 128-dimensional, reducing computational complexity while ensuring the integrity of the features.

[0031] The dimensionality-reduced feature vectors of all technology projects are aggregated and organized according to project identifiers to generate a normalized multimodal project feature set. This feature set contains a 128-dimensional normalized feature vector for each technology project. Each dimension of the vector corresponds to a key feature of the multimodal data, and auxiliary information such as project identifier and data collection time are also labeled to ensure the traceability and usability of the feature set. In the example, the normalized multimodal project feature set contains feature vectors of 100 technology projects, each with 128 dimensions. The feature vector of project 1 contains multi-dimensional key information such as technical entities, design features, and operation modes, which can be directly used in subsequent steps such as technical entity extraction and knowledge graph construction to complete the entire process of this step.

[0032] S202, Based on the normalized multimodal project feature set, a self-supervised learning algorithm is used to automatically extract technical entities, R&D teams, resource elements and their dynamic relationships to construct a domain knowledge graph that evolves over time; Specifically, technical term vectors, R&D personnel feature vectors, and resource identifier vectors can be extracted from the normalized multimodal project feature set. A self-supervised contrastive learning algorithm is used to train the entity recognition model to generate an initial entity set of technical entities, R&D teams, and resource elements. The core of this step is to separate three types of core vectors from the normalized feature set, train an entity recognition model through self-supervised learning, and automatically extract entities related to science and technology projects to form an initial entity set. The specific implementation method is as follows: The normalized multimodal item feature set is a 128-dimensional vector. Each vector integrates the core information of text, image, and time series multimodal features. When extracting the three types of target vectors, it is necessary to combine the distribution pattern of each modality feature and achieve accurate separation by filtering the feature dimensions. The technical terminology vectors primarily correspond to the technically relevant features of the text modality (R&D documents, application materials) and image modality (design drawings) in the feature set. Vector fragments from dimensions 1-40 (corresponding to text technical entities and image technical features) in the feature set are selected and linearly mapped to transform them into 64-dimensional technical terminology vectors. These vectors can represent the semantic features and associated attributes of technical terms. The R&D personnel feature vectors correspond to the personnel-related features of the time-series modality (operation logs) and text modality (participant information) in the feature set. Vector fragments from dimensions 41-80 are selected and transformed into 64-dimensional R&D personnel feature vectors, representing information such as the R&D personnel's professional fields, operating permissions, and participation frequency. The resource identifier vectors correspond to the resource-related features of the text modality (budget, equipment list) and image modality (equipment drawings). Vector fragments from dimensions 81-128 are selected and transformed into 64-dimensional resource identifier vectors, representing information such as the type, quantity, specifications, and ownership of resources.

[0033] Self-supervised contrastive learning algorithms are the core of training entity recognition models. Their core principle is to achieve accurate entity recognition by constructing positive and negative sample pairs without requiring manual data labeling. This allows the model to automatically learn the characteristics of similar entity vectors and dissimilar entity vectors. During training, the rules for constructing sample pairs are as follows: vectors belonging to the same category within the same project (e.g., two different technical term vectors) are used as positive sample pairs, while vectors from different categories (e.g., a technical term vector and a R&D personnel feature vector) are used as negative sample pairs. A temperature parameter τ=0.07 is introduced (to adjust the weight of sample pair similarity and prevent gradient vanishing). The model parameters are optimized using a contrastive loss function.

[0034] The entity recognition model employs an encoder-classifier architecture. The encoder receives a 64-dimensional target vector and extracts deep semantic features. The classifier outputs entity categories (technical entities, R&D personnel, resource elements) and entity identifiers. In the example, technical term vectors V1 (corresponding to "LSTM neural network") and V2 (corresponding to "feature engineering algorithm"), R&D personnel feature vectors U1 (corresponding to "Zhang San, algorithm engineer") and U2 (corresponding to "Li Si, test engineer"), and resource identifier vectors W1 (corresponding to "high-performance servers, 2 units") and W2 (corresponding to "R&D funding, 500,000 yuan") are extracted from the normalized feature set of a certain technology project. Positive sample pairs (V1, V2), (U1, U2), and (W1, W2) and negative sample pairs (V1, U1), (W1, V2) are constructed and input into the model for training. After 100 iterations, the model converges, achieving a recognition accuracy of over 92%. After training, the model identifies and classifies the three types of vectors for all projects, removes duplicates, and generates an initial entity set. The technical entities include "LSTM neural network", "feature engineering algorithm", "intelligent risk control model", etc. The R&D team entities include "algorithm R&D group", "testing group", "Zhang San", "Li Si", etc. The resource element entities include "high-performance server", "R&D funds", "laboratory equipment", etc., thus completing the generation of the initial entity set.

[0035] Based on the initial entity set, a self-supervised relationship prediction task is designed. The potential association patterns between entities are learned through the masked entity prediction algorithm to generate an entity relationship prediction model. The core of this step is to use a self-supervised relation prediction task to allow the model to automatically learn the association patterns between initial entities, optimize the model through a masked entity prediction algorithm, and finally generate a model that can accurately predict entity associations. The specific implementation method is as follows: Based on an initial entity set, a self-supervised relation prediction task is designed. The core of this task is to construct a sequence of entity pairs and mine the potential associations between entities. These associations are mainly categorized into three types: associations between technical entities (e.g., the "support" relationship between "feature engineering algorithm" and "LSTM neural network"), associations between R&D teams and technical entities (e.g., the "responsibility" relationship between "algorithm R&D group" and "LSTM neural network"), and associations between resource elements and technical entities (e.g., the "support" relationship between "high-performance server" and "intelligent risk control model"). During task design, entities in the initial entity set are randomly combined into entity pairs. Each entity pair is labeled to indicate whether an association exists (positive / negative). Simultaneously, contextual information about the entities (e.g., their location and temporal sequence in the multimodal feature set) is introduced to enrich the training data and enhance the model's ability to learn association patterns.

[0036] Masked entity prediction is a core implementation method for self-supervised relation prediction tasks. Its principle involves randomly masking one entity in an entity pair (the masking ratio is set to 15%, ensuring sufficient prediction space for the model while avoiding excessive masking that could lead to missing contextual information). The model then predicts the masked entity based on the unmasked entity and contextual information, thereby learning the relationship patterns between entities. For example, constructing an entity pair ("Algorithm R&D Group", "LSTM Neural Network"), masking the "LSTM Neural Network", and using contextual information about the "Algorithm R&D Group" (such as the group's operation logs mainly involving algorithm development, and text mentioning that the group is responsible for core algorithm development) to predict the masked entity. Through backpropagation of prediction errors, the model parameters are optimized, allowing the model to gradually grasp the "responsible" relationship between the "Algorithm R&D Group" and the "LSTM Neural Network".

[0037] The entity relationship prediction model adds a relationship prediction head to the entity recognition model in step one. The prediction head uses two fully connected layers. The input is a 128-dimensional vector resulting from the concatenation of the feature vectors of two entities. The output is the association type and association probability between the entities (values ​​range from 0 to 1; higher probability indicates a stronger association). During training, a cross-entropy loss function is used, combined with the loss from masked entity prediction, to jointly optimize the model. After 80 iterations, the model converges, achieving an association prediction accuracy of over 90%. In the example, the model inputs an entity pair ("high-performance server", "intelligent risk control model"), and outputs an association type of "support" with an association probability of 0.93, indicating a strong association. Inputting an entity pair ("test group", "R&D funds"), the model outputs an association probability of 0.12, indicating no significant association. After training, the entity relationship prediction model is generated. This model can accurately predict the association type and strength of two input entities, laying the foundation for subsequent dynamic relationship analysis.

[0038] The entity relationship prediction model is used to analyze the temporal evolution patterns of multimodal project feature sets, identify the creation, strengthening and disappearance processes of entity relationships under different time slices, and generate dynamic relationship evolution sequences. The core of this step is to combine time-series information and analyze the changing patterns of entity relationships over time using a relationship prediction model, identifying the three states of relationship creation, strengthening, and disappearance, and forming a dynamic relationship evolution sequence. The specific implementation method is as follows: First, the multimodal project feature set is divided into time slices according to the time dimension. The division of time slices is based on the R&D cycle of science and technology projects, and is set to 3 months per slice (adapting to the stage division of most science and technology projects, which can capture the dynamic changes of relationships and avoid computational redundancy caused by overly fine slices). Each time slice corresponds to a stage of project R&D (such as early stage, mid stage, and late stage). Each slice contains all normalized multimodal feature vectors within that stage and is labeled with a slice timestamp (e.g., from 2024-01-01 to 2024-03-31, the timestamp is recorded as T1).

[0039] An entity relationship prediction model is used to analyze entity relationships within each time slice. For each pair of entities in the initial entity set, the model is input to predict its association type and association strength within that slice. Association strength is quantified into a score of 0-100 (derived from association probability, probability × 100), with higher scores indicating stronger associations. By comparing the association status of the same entity pair in different time slices, three evolutionary processes of the relationship are identified: the creation process, where an entity pair is unrelated in previous slices (association strength < 20) but becomes related in the current slice (association strength ≥ 20); the strengthening process, where the entity pair is related in consecutive slices, and the association strength increases by ≥ 10; and the disappearance process, where the entity pair is related in previous slices (association strength ≥ 20) but is unrelated in the current slice (association strength < 20).

[0040] In the example, a technology project is divided into four time slices: T1 (early stage), T2 (mid-stage 1), T3 (mid-stage 2), and T4 (late stage). The entity pair ("Algorithm R&D Team", "LSTM Neural Network") has a correlation strength of 30 in slice T1 (created, indicating that the team began to be responsible for the algorithm R&D in the early stage), 55 in slice T2 (strengthened, R&D depth increased), 80 in slice T3 (strengthened, the correlation became closer), and 75 in slice T4 (no significant change, not reaching the threshold for strengthening or disappearing). The entity pair ("LSTM Neural Network", "Traditional Regression Algorithm") has a correlation strength of 25 in slices T1-T2 (correlated, used for algorithm comparison), 15 in slice T3 (disappeared, indicating that the project abandoned the comparison with the traditional algorithm), and 12 in slice T4 (remaining in a state of disappearance). The entity pair ("R&D Funds", "Intelligent Risk Control Model") has a correlation strength of 18 in slice T1 (no correlation), 40 in slice T2 (created, funds began to be invested in model R&D), and 60 and 70 in slices T3-T4, respectively (continuously strengthened). The evolution of all entity pairs is recorded in chronological order by time slices, and the evolution type, change in association strength, and slice timestamp are labeled to generate a dynamic relationship evolution sequence. This sequence fully presents the dynamic changes of entity relationships as the project development progresses.

[0041] By integrating the initial entity set and the dynamic relationship evolution sequence, a temporal knowledge graph construction algorithm is used to organize entities and relationships according to the time dimension, and finally construct a domain knowledge graph that evolves over time.

[0042] The core of this step is to fuse the initial entities and the dynamic relationship evolution sequence. Using a temporal knowledge graph construction algorithm, entities and relationships are organized along the time dimension to form a domain knowledge graph that reflects the dynamic evolution of entity relationships. The specific implementation method is as follows: The process integrates the initial entity set and the dynamic relationship evolution sequence. First, the initial entity set is supplemented and improved, and then updated with new entities appearing in the dynamic relationship evolution sequence (such as the "Transformer algorithm" entity added during the evolution process) to ensure entity integrity. Simultaneously, the dynamic relationship evolution sequence is deduplicated and validated, removing abnormal evolution records (such as records with abrupt changes in association strength without reasonable context) to ensure the rationality of relationship evolution. The integrated data includes an entity set (containing entity identifiers, entity types, and core attributes) and a dynamic relationship set (containing entity pairs, association types, association strengths, evolution types, and time slices).

[0043] The core of the temporal knowledge graph construction algorithm is to organize entities and relationships according to the time dimension, constructing a spatiotemporally integrated graph structure. The algorithm treats each time slice as a time layer of the graph, with each time layer containing all entity nodes and associated edges within that slice. Entity node attributes include entity identifier, type, and core features (extracted from normalized feature vectors). Associated edge attributes include association type, association strength, and evolution type. A time attribute (slice timestamp) is also added to enable temporal tracing of entity relationships. The algorithm employs a graph storage structure, with entities as nodes and relationships as directed edges (e.g., "Algorithm Development Group" → "LSTM Neural Network," with the edge label "responsible"). Each node and edge is bound to time slice information, supporting queries of graph states at different stages according to the time dimension.

[0044] In the example, the integrated entity set includes entities such as "Algorithm R&D Group", "LSTM Neural Network", "R&D Funds", and "Transformer Algorithm". The dynamic relationship set contains the evolution records of each entity pair. After constructing the algorithm using a time-series knowledge graph, the graph in time layer T1 (initial stage) includes nodes "Algorithm R&D Group", "LSTM Neural Network", and "R&D Funds", with the edge "Algorithm R&D Group" → "LSTM Neural Network" (association strength 30, evolution type creation). In time layer T2 (mid-term 1), the edge "R&D Funds" → "Intelligent Risk Control Model" (association strength 40, evolution type creation) is added, and the association strength of the original edge "Algorithm R&D Group" → "LSTM Neural Network" is increased to 55 (evolution type strengthening). In time layer T3 (mid-term 2), the entity "Transformer Algorithm" is added, and the edge "Algorithm R&D Group" → "Transformer Algorithm" (association strength 35, evolution type creation) is added. The association strength of the edge "R&D Funds" → "Intelligent Risk Control Model" is increased to 60 (strengthening), and the edge "LSTM Neural Network" → "Traditional Regression Algorithm" is deleted. In time layer T4 (late stage), the association strength of each edge is fine-tuned to maintain the core association relationships. The final domain knowledge graph can clearly present the entity distribution and relationship evolution at different R&D stages, support time-series queries (such as querying the changes in entity relationships between slices T2 and T3), provide structured knowledge support for subsequent risk propagation simulation, and complete the entire process of this step.

[0045] S203, input the domain knowledge graph into the pre-trained dynamic graph neural network risk propagation model to simulate the transmission path and superposition effect of multiple risk factors in the project network, and output multi-level risk warning signals within a preset period in the future; Specifically, the domain knowledge graph that evolves over time can be represented as a dynamic graph sequence data, with each time step containing entity node features and relation edge weights, to obtain a dynamic graph sequence for input; The core of this step is to transform the time-series domain knowledge graph into a dynamic graph sequence format that the model can recognize. By quantifying the features of entity nodes and the weights of relation edges, the spatiotemporal evolution information of entity relationships is preserved, laying the data foundation for subsequent risk propagation simulation. The specific implementation method is as follows: The domain knowledge graph, evolving over time, exists in the form of multiple time layers. Each time layer corresponds to the entity association state at a specific development stage. The dynamic graph sequence arranges these time layers chronologically, forming continuous time-series graph data. Each time step corresponds to one time layer, and the interval between time steps is consistent with the time slices of the knowledge graph, set to 3 months to ensure the continuity of time-series information. Each time step contains two core data points: entity node features and relation edge weights, which together constitute the basic unit of the dynamic graph sequence.

[0046] Entity node features are quantitative representations of each entity node. Based on a normalized multimodal project feature set extraction, each entity node corresponds to a 64-dimensional feature vector. The vector dimension is determined by the entity type. Technical entity node features focus on indicators such as technology maturity and complexity; R&D team node features focus on indicators such as team size and professional matching degree; and resource element node features focus on indicators such as resource sufficiency and utilization rate. During feature extraction, the corresponding dimension vectors in the normalized feature set are converted into node features through linear mapping. Node type identifiers are added (Technology Entity = 1, R&D Team = 2, Resource Element = 3) to facilitate the model's differentiation of the risk transmission characteristics of different types of nodes. In the example, the node feature vector of "LSTM Neural Network" (Technology Entity) includes quantitative indicators such as technology maturity (0.8) and complexity (0.7); the node feature vector of "Algorithm R&D Group" (R&D Team) includes indicators such as team size (0.6) and professional matching degree (0.9); and the node feature vector of "R&D Funds" (Resource Element) includes indicators such as sufficiency (0.5) and utilization rate (0.8).

[0047] Relationship edge weights are a quantitative representation of the strength of the relationship between entities, ranging from 0 to 1. A higher weight indicates a stronger relationship between the two entities, leading to a higher probability and faster risk transmission. The weight value is derived from the relationship strength of entities in the domain knowledge graph using the formula: weight = relationship strength / 100. Here, the relationship strength is a score of 0-100 quantified in the previous dynamic relationship evolution sequence, which is then standardized to a weight of 0-1 using this formula. In the example, the relationship strength between "Algorithm R&D Group" and "LSTM Neural Network" is 80 points, resulting in a weight of 0.8 after conversion, indicating a strong relationship and high risk transmission efficiency. Conversely, the relationship strength between "Testing Group" and "R&D Funds" is 12 points, resulting in a weight of 0.12 after conversion, indicating a weak relationship and difficulty in risk transmission.

[0048] The entity node features and relation edge weights at each time step are organized according to a graph structure, labeled with corresponding time step identifiers (e.g., T1, T2, T3, T4), and concatenated in chronological order to form a dynamic graph sequence. In the example, the dynamic graph sequence of a technology project contains four time steps. Time step T1 includes nodes such as "Algorithm R&D Group" and "LSTM Neural Network" and their features, as well as corresponding relation edge weights. Time step T2 adds the "R&D Funds" node and related edge weights. Subsequent time steps update node features and edge weights sequentially, ultimately forming a complete dynamic graph sequence that can be directly input into a dynamic graph neural network risk propagation model.

[0049] Load a pre-trained dynamic graph neural network risk propagation model. This model adopts a spatiotemporal graph convolutional network architecture, which can capture the spatiotemporal dependencies between nodes in the graph for risk propagation calculation. The core of this step is to load the pre-trained risk propagation model, clarify the model architecture and core functions, and ensure that the model can accurately capture the spatiotemporal dependencies between entity nodes, providing reliable model support for risk transmission simulation. The specific implementation method is as follows: The pre-trained dynamic graph neural network risk propagation model is a mature model obtained after multiple rounds of iterative training and optimization based on a large amount of dynamic graph sequence data and risk event data of historical science and technology projects. The core value of the model lies in its ability to simultaneously capture the spatial dependency relationship (the association between different nodes in the same time step) and the temporal dependency relationship (the state change of the same node in different time steps) between entity nodes, thereby accurately simulating the transmission law of risk in the spatiotemporal dimension.

[0050] The model employs a spatiotemporal graph convolutional network architecture, consisting of spatial convolutional layers, temporal convolutional layers, and fully connected layers. These three layers work together to capture spatiotemporal dependencies. The spatial convolutional layer captures the spatial dependencies between entity nodes within the same time step. By weighted aggregation of the features of each node's neighboring nodes, it mines the impact of node associations on risk propagation. The weights are determined by the edge weights; the closer the association between neighboring nodes, the greater the risk impact on the current node. The temporal convolutional layer captures the temporal dependencies of the same node across different time steps. By performing convolution operations on the feature changes of the same node across consecutive time steps, it captures the temporal evolution of node risk states. A 1D convolutional kernel with a kernel size of 3 (considering node features from the current time step and the two preceding and following time steps simultaneously) is used to ensure the continuity of temporal features. The fully connected layer fuses the features extracted by the spatiotemporal convolutional layer, outputting the risk prediction value for each node and completing the risk propagation calculation.

[0051] During model pre-training, dynamic graph sequences of 1000 historical science and technology projects were used as training data. Each project was labeled with real risk events and their impact range. A risk prediction loss function (calculated based on the deviation between the actual and predicted risk values ​​of nodes) was used to optimize model parameters. After 120 training iterations, the model converged, achieving a risk prediction accuracy of over 91%, meeting practical application requirements. When loading the model, pre-trained model parameters (such as convolutional kernel weights and fully connected layer parameters) were loaded simultaneously. The model input dimension was set to the node feature dimension (64 dimensions) and the number of time steps (default maximum support of 20 time steps, suitable for long-term R&D projects). The output dimension was the risk value (0-1) of each node. A model inference speed threshold was also set to ensure that the latency of each round of risk propagation simulation was controlled within 50 milliseconds, meeting real-time early warning requirements. In the example, after loading the model, it automatically identified the number of time steps and nodes in the input dynamic graph sequence, adapting to the current R&D cycle and entity scale of the science and technology project, and prepared to conduct risk propagation simulation.

[0052] By inputting dynamic graph sequences into a dynamic graph neural network risk propagation model, the transmission process of technical risks, resource risks, and team collaboration risks in the project network is simulated through a multi-hop message passing mechanism, generating risk propagation simulation results. The core of this step is to utilize the model's multi-hop message passing mechanism to simulate the transmission paths and processes of three types of core risks, passing the risks from the initial risk node to related nodes, and finally generating simulation results containing the risk status of each node. The specific implementation method is as follows: The multi-hop message passing mechanism is the core logic of the model simulating risk transmission. Its principle is to treat the risk state of each entity node as a "message," which is transmitted to its neighboring nodes through relational edges. The transmission process can span multiple nodes (i.e., multiple hops). Each hop message is attenuated according to the relational edge weight and the transmission distance to ensure the rationality of risk transmission. The number of hops is set to 3 (which can cover the main related nodes while avoiding excessive hops that could lead to distorted risk diffusion). During transmission, the strength of the message transmission is determined by the relational edge weight and the risk attenuation coefficient. The risk attenuation coefficient is set to 0.9, meaning that the risk intensity decreases by 10% with each hop. The attenuation coefficient can be dynamically adjusted according to the project type; for technology-intensive projects, the attenuation coefficient can be appropriately reduced (e.g., 0.85), while for resource-intensive projects, it can be appropriately increased (e.g., 0.95).

[0053] The simulated risks are categorized into three types: technical risk, resource risk, and team collaboration risk. These three risks are transmitted independently yet influence each other, collectively constituting the overall risk transmission process of the project. Technical risks mainly originate from technical entity nodes, such as technical defects in the "LSTM neural network" or insufficient adaptability of the "feature engineering algorithm." The initial risk value is set to 0.7 (ranging from 0 to 1, with 0.7 representing a higher initial risk). The risk transmission path is primarily technical entity → R&D team → other technical entities. In the example, the technical risk (0.7) of the "LSTM neural network" is transmitted to the "algorithm R&D team" through a relation edge with a weight of 0.8. After transmission, the technical risk value of the "algorithm R&D team" is 0.7 × 0.8 × 0.9 = 0.504. Then, the "algorithm R&D team" transmits the risk to the "intelligent risk control model" (relationship edge weight 0.7), resulting in a risk value of 0.504 × 0.7 × 0.9 ≈ 0.317.

[0054] Resource risks mainly originate from resource element nodes, such as insufficient "R&D funds" and failure of "high-performance servers". The initial risk value is set to 0.8. The transmission path is mainly resource element → technical entity → R&D team. In the example, the resource risk (0.8) of "R&D funds" is passed to the "intelligent risk control model" through the relation edge with a weight of 0.6. After the transmission, the risk value is 0.8×0.6×0.9=0.432. Then it is passed to the "test group" (relation edge weight 0.5). After the transmission, the risk value is 0.432×0.5×0.9≈0.194. Team collaboration risks mainly originate from nodes within the R&D team, such as insufficient personnel in the "testing group" and low collaboration efficiency in the "algorithm R&D group." The initial risk value is set at 0.6, and the transmission path is mainly R&D team → technical entity → resource element. In the example, the collaboration risk (0.6) of the "testing group" is passed to the "intelligent risk control model" through a relation edge with a weight of 0.4. After the transmission, the risk value is 0.6 × 0.4 × 0.9 = 0.216. Then it is passed to the "high-performance server" (relation edge weight 0.3), and after the transmission, the risk value is 0.216 × 0.3 × 0.9 ≈ 0.059.

[0055] During the risk transmission simulation, the model calculates the transmission values ​​of the three types of risks for each node at each time step, sums them to obtain the total risk value of the node, and records the transmission path, number of hops, and risk decay process for each risk. After the simulation is completed, the risk transmission simulation results are generated. These results include the total risk value of all entity nodes within each time step, the individual risk values ​​of each type of risk, and a list of risk transmission paths, clearly showing the transmission process and scope of impact of risks in the project network. In the example, the simulation results show that the total risk value of the "intelligent risk control model" at time step T3 is 0.317 (technical risk) + 0.432 (resource risk) + 0.216 (collaboration risk) = 0.965, with the risks mainly originating from the "LSTM neural network" and "R&D funding" nodes.

[0056] Based on the results of risk transmission simulation, the risk superposition effect analysis algorithm is used to predict the degree of risk accumulation of each entity node in the future within a preset period. The risk levels are divided into high, medium and low risk levels according to the risk level threshold, and finally multi-level risk warning signals are output.

[0057] The core of this step is to predict future risk accumulation through risk superposition effect analysis, classify risk levels, and generate multi-level early warning signals, providing a clear basis for the formulation of subsequent intervention strategies. The specific implementation method is as follows: The core of the risk superposition effect analysis algorithm is to consider the time accumulation characteristics and mutual superposition effects of risks, and predict the risk accumulation degree of each entity node within a preset period. The preset period is set to 6 months based on the R&D stage of the technology project and the risk warning requirements, that is, to predict the risk accumulation situation of the next two time steps (3 months each). The algorithm adopts a nonlinear superposition model, considering the time decay characteristics of risks, and sets the time decay coefficient to 0.95, which means that the risk weight of each future time step is reduced by 5% compared with the current time step. The later the time step, the smaller the impact of the risk on the current warning. The formula for calculating the risk accumulation degree is risk_accum=Σ(risk_t×0.95^(t-t0)), where risk_accum is the risk accumulation degree, risk_t is the node risk value at the t-th time step, t0 is the current time step, t is the future time step (t>t0), and the summation range is all time steps within the preset period.

[0058] In the example, the current time step is T4, and the preset future period is 6 months (corresponding to two time steps, T5 and T6). The risk value of the "intelligent risk control model" at T4 is 0.965, the model predicts that the risk value at time step T5 is 0.92 (based on the previous transmission pattern), and the risk value at time step T6 is 0.88. Substituting into the formula, the risk accumulation level is calculated as 0.92×0.95^(5-4)+0.88×0.95^(6-4)=0.92×0.95+0.88×0.9025≈0.874+0.794=1.668. The higher the risk accumulation level, the greater the risk that the node will face in the future.

[0059] Risk level thresholds are set based on historical risk event statistics and project risk tolerance, using a three-part division principle, and dynamically adjusted according to project type. The general thresholds are: high risk level (risk accumulation ≥ 1.2), medium risk level (0.6 ≤ risk accumulation < 1.2), and low risk level (risk accumulation < 0.6). The thresholds are set based on historical data: when the risk accumulation level is ≥ 1.2, over 85% of nodes will experience actual risk events; when 0.6 ≤ risk accumulation < 1.2, 30%-85% of nodes will experience risk events; and when < 0.6, only less than 5% of nodes will experience risk events, ensuring the rationality of the threshold division. In the example, the risk accumulation level of the "intelligent risk control model" is 1.668 ≥ 1.2, classifying it as high risk; the risk accumulation level of the "algorithm development team" is 0.85, classifying it as medium risk; and the risk accumulation level of the "high-performance server" is 0.32, classifying it as low risk.

[0060] Multi-level risk warning signals are generated based on risk level classification. The "multi-level" aspect is reflected in three dimensions: node level, relationship level, and overall project level. Node-level warning signals target individual high, medium, and low-risk nodes, clearly defining the node identifier, risk level, risk accumulation degree, and main risk sources. Relationship-level warning signals target key related edges in the risk transmission path, clarifying high-risk transmission paths (e.g., "R&D funds → intelligent risk control model"). Overall project-level warning signals are based on the risk level distribution of all nodes, judging the overall risk level of the project (e.g., if the number of high-risk nodes is ≥3, the project is judged as high-risk). Warning signals include core information such as warning type, warning level, affected nodes, risk transmission path, prediction period, and emergency alerts. In the example, the high-risk warning signal for the "intelligent risk control model" is: "Warning type: Technology + Resource Risk Warning; Warning level: High risk; Affected node: Intelligent risk control model; Main risk sources: LSTM neural network (technology risk), R&D funds (resource risk); Prediction period: Next 6 months; Emergency alert: Timely investigation of technical defects and replenishment of R&D funds are required." All warning signals are sorted by risk level and then aggregated to form a complete multi-level risk warning signal, thus completing the entire process of this step.

[0061] S204. Based on the multi-level risk warning signals, an adaptive project intervention strategy set is generated using a multi-objective optimization algorithm. The strategy set covers resource allocation adjustment, technical route correction, and collaborative key point reorganization, so as to realize intelligent evaluation and early warning optimization of science and technology projects.

[0062] Specifically, it can analyze multi-level risk warning signals, extract high-risk entity nodes and their relationships, construct a multi-objective optimization problem for risk mitigation, with objective functions including maximizing risk reduction, minimizing resource consumption, and minimizing the impact on project schedule, and generate a multi-objective optimization model; The core of this step is to extract key decision-making information from multi-level risk warning signals, clarify the core objectives and constraints of risk mitigation, and construct a multi-objective optimization model by quantifying the objective function, providing a standardized framework for subsequent strategy solving. The specific implementation method is as follows: Multi-level risk warning signals include warning information at the node level, relationship level, and overall project level. The analysis process needs to be broken down hierarchically, prioritizing the extraction of high-risk entity nodes and their key relationships, while eliminating low-risk and irrelevant redundant information. During analysis, firstly, entity nodes with a high risk level (risk accumulation degree ≥ 1.2) are selected from the node-level warning signals, and the core risk type (technical risk, resource risk, team collaboration risk), risk accumulation degree, and main risk source nodes are labeled for each high-risk node. Next, high-risk transmission paths are extracted from the relationship-level warning signals, clarifying the association type and strength between high-risk nodes and source nodes, forming a high-risk association network. Finally, combined with the overall project-level warning signals, the overall project risk constraints are determined (e.g., the overall risk level needs to be reduced from high risk to medium risk or below).

[0063] In the example, after analyzing the multi-level risk warning signals, the high-risk entity node is extracted as "intelligent risk control model" (risk accumulation level 1.668, core risks are technology risk + resource risk). Its correlation is as follows: it has a strong correlation with "LSTM neural network" (source of technology risk, correlation strength 0.8) and "R&D funds" (source of resource risk, correlation strength 0.6). The high-risk transmission path is "LSTM neural network → intelligent risk control model" and "R&D funds → intelligent risk control model". The overall risk level of the project is high risk, and intervention is needed to reduce the overall risk to medium risk.

[0064] The core of constructing a multi-objective optimization problem for risk mitigation is to define three mutually constraining objective functions and set reasonable constraints to ensure that the optimization problem aligns with the actual needs of the project. All three objective functions are defined quantitatively, as follows: The objective function for maximizing risk reduction is quantified by the risk reduction rate at high-risk nodes, calculated as risk_reduce = (risk_before - risk_after) / risk_before, where risk_before represents the risk accumulation at high-risk nodes before intervention, and risk_after represents the risk accumulation after intervention. The goal is to maximize risk_reduce (ideally ≥ 0.4, i.e., a risk reduction of over 40%). The objective function for minimizing resource consumption is quantified by the total resource consumption cost of the intervention strategy, encompassing resources such as funds, equipment, and manpower, calculated as resource_c. ost = Σ(resource_i × cost_i), where resource_i is the input amount of the i-th type of resource and cost_i is the unit cost of the i-th type of resource. The goal is to minimize resource_cost (not exceeding 30% of the total remaining available resources of the project). The objective function for minimizing the impact on project schedule is quantified by the delay time of the intervention strategy on the project schedule. The calculation formula is schedule_delay = Σ(strategy_j × delay_j), where strategy_j is the j-th intervention strategy and delay_j is the number of days of schedule delay caused by the strategy. The goal is to minimize schedule_delay (not exceeding 7 days).

[0065] The constraints mainly include resource constraints (intervention resource input ≤ total remaining available resources of the project), schedule constraints (total delay ≤ 7 days), and risk constraints (after intervention, the cumulative risk level of high-risk nodes < 1.2, and the overall risk level of the project is reduced to medium risk or below). Combining the high-risk correlation network, the objective function and constraints are integrated to generate a multi-objective optimization model. In the example, the remaining available R&D funds for the project are 2 million yuan, the remaining manpower is 10 people, and the remaining project schedule is 6 months. The objective of the multi-objective optimization model is to achieve a risk reduction rate of ≥ 0.4 for the "intelligent risk control model," resource consumption ≤ 600,000 yuan (2 million × 30%), and schedule delay ≤ 7 days. The constraints are: capital input ≤ 2 million yuan, manpower input ≤ 10 people, and delay ≤ 7 days, ultimately generating a complete multi-objective optimization model.

[0066] A non-dominated sorting genetic algorithm is used to solve the multi-objective optimization model. The Pareto optimal solution set is searched through population evolution iteration to generate a set of candidate intervention strategies. The core of this step is to utilize the global search capability of the non-dominated sorting genetic algorithm to solve the multi-objective optimization model, find the Pareto optimal solution set that takes into account the three objective functions, and transform the solution set into a set of candidate intervention strategies. The specific implementation method is as follows: The non-dominated sorting genetic algorithm is an intelligent algorithm specifically designed to solve multi-objective optimization problems. Its core principle is to simulate the genetic evolution of biological populations, screening, crossovering, and mutating potential intervention strategies (individuals) through iterative optimization until a Pareto optimal solution set is found—meaning no single strategy is superior to another across all objective functions, and each strategy is one of the optimal choices under the current conditions. The core parameters of the algorithm include population size, number of iterations, crossover probability, and mutation probability. Parameter settings must balance search efficiency and optimization effect. Specific parameters are: population size 50 (i.e., retaining 50 potential intervention strategies in each iteration), number of iterations 30 (ensuring algorithm convergence and avoiding computational redundancy due to excessive iteration), crossover probability 0.8 (controlling the feature fusion probability between two strategies; 0.8 means 80% of individuals will undergo crossover, increasing population diversity), and mutation probability 0.1 (controlling the feature mutation probability of a single strategy; 0.1 means 10% of individuals will mutate, preventing the algorithm from getting trapped in local optima).

[0067] The algorithm execution process consists of four stages: Population initialization, which randomly generates 50 potential intervention strategies based on high-risk nodes and their relationships. Each strategy includes one or more operations such as resource allocation adjustment, technical route correction, and collaborative attack point reorganization. The resource consumption, schedule delay, and risk reduction rate of each strategy are labeled to ensure that the initial population covers all possible intervention directions; Non-dominated sorting, which sorts the population in each iteration round and divides the strategies into different non-dominated levels according to the performance of the three objective functions. The strategy in the optimal level (first level) is the current Pareto optimal solution and is retained for the next iteration; Genetic operation, which performs crossover and mutation operations on the retained strategies. The crossover operation is to merge the core operations of two different strategies (such as merging "supplementing R&D funds" and "optimizing the LSTM algorithm" into a new strategy), and the mutation operation is to randomly adjust the parameters of a single strategy (such as adjusting the investment from 500,000 yuan to 550,000 yuan); Convergence judgment, which stops the algorithm when the Pareto optimal solution set does not change significantly after 3 consecutive iterations or when the preset number of iterations (30 rounds) is reached, and outputs the final Pareto optimal solution set.

[0068] In the example, after 30 iterations, the algorithm converges to 12 Pareto optimal solutions, each corresponding to a potential intervention strategy. These strategies are then compiled to generate a candidate intervention strategy set. The candidate strategies in the set cover three types of operations. For example, Strategy 1: Supplement R&D funding by 600,000 yuan (invested in the "R&D Funds" node), optimize the LSTM neural network algorithm (adjust the technical parameters of the "LSTM Neural Network" node), consuming 600,000 yuan in resources, delaying progress by 3 days, and reducing risk by 0.42; Strategy 2: Assign 3 senior algorithm engineers to the "Algorithm R&D Team," reorganize collaborative attack points (merge the Algorithm R&D Team and the Testing Team), consuming 450,000 yuan in resources, delaying progress by 5 days, and reducing risk by 0.41; Strategy 3: Change the core technology route (replacing part of the LSTM algorithm with the Transformer algorithm), supplement with 2 high-performance servers, consuming 580,000 yuan in resources, delaying progress by 6 days, and reducing risk by 0.45. All candidate strategies meet the constraints and take into account the optimization requirements of the three objective functions.

[0069] Based on the set of candidate intervention strategies, a strategy feasibility assessment index system is constructed, including assessments of technical feasibility, resource accessibility, and team execution capability, and strategy feasibility assessment results are generated. The core of this step is to establish a scientific feasibility assessment index system, comprehensively evaluate candidate intervention strategies, quantify the feasibility score of each strategy, select strategies that meet the actual implementation conditions of the project, and generate feasibility assessment results. The specific implementation method is as follows: The strategy feasibility assessment index system includes three primary indicators: technical feasibility, resource accessibility, and team execution capability. Each primary indicator has several secondary quantitative indicators. All indicators use a quantitative scoring standard of 0-100 points. The higher the score, the stronger the strategy feasibility. The weight of the indicators is set according to the project requirements. Among them, technical feasibility has a weight of 0.4 (a core indicator that directly affects the intervention effect), resource accessibility has a weight of 0.3, and team execution capability has a weight of 0.3. The comprehensive feasibility score is calculated as feasibility_score = 0.4 × tech_score + 0.3 × resource_score + 0.3 × team_score, where tech_score is the technical feasibility score, resource_score is the resource accessibility score, and team_score is the team execution capability score.

[0070] The technical feasibility assessment mainly determines whether the technical operations in the intervention strategy are feasible. The secondary indicators include technical maturity (0-100 points, such as 90 points for the Transformer algorithm maturity and 40 points for the maturity of a new unverified algorithm), technical adaptability (0-100 points, assessing the degree of adaptability of the technology to the existing architecture of the project), and technical risk controllability (0-100 points, assessing whether the potential risks in the implementation process of the technology are controllable). The scoring adopts a combination of expert scoring and historical data comparison. In the example, the technical feasibility assessment of strategy 3 is as follows: Transformer algorithm maturity 90 points, adaptability to intelligent risk control model 85 points, technical risk controllability 80 points, tech_score=(90+85+80) / 3=85 points.

[0071] Resource accessibility assessment primarily determines whether a project possesses the various resources required to implement its strategy. Secondary indicators include funding availability (0-100 points, assessing whether the required funding is within the project's remaining budget), equipment availability (0-100 points, assessing whether the required equipment can be allocated in a timely manner), and manpower availability (0-100 points, assessing whether the required manpower can be provided). The score is calculated based on the project's existing resource reserves. In the example, the resource accessibility assessment for Strategy 1 is as follows: the required funding of 600,000 yuan is within the remaining budget of 2 million yuan (100 points), no new equipment is needed (100 points), and no additional manpower is needed (100 points). Therefore, resource_score = (100 + 100 + 100) / 3 = 100 points.

[0072] The team execution capability assessment primarily determines whether the R&D team possesses the ability to implement the strategy. Secondary indicators include team professional matching degree (0-100 points, assessing the degree of matching between the team members' professional expertise and the strategy requirements), team collaboration efficiency (0-100 points, assessing the team's collaborative ability to execute the strategy), and strategy execution experience (0-100 points, assessing whether the team has similar strategy execution experience). The score is based on the team's current personnel composition and historical execution data. In the example, the team execution capability assessment for Strategy 2 is as follows: the professional matching degree of 3 senior algorithm engineers is 90 points, the team's historical collaboration efficiency is 80 points, and the experience in deploying similar personnel is 75 points. The team_score is approximately 81.7 points (rounded down to 82 points).

[0073] Each of the 12 candidate intervention strategies was scored, and a comprehensive feasibility score was calculated. A feasibility threshold of 70 points was set (a score ≥70 points indicates a feasible strategy, and <70 points indicates an infeasible strategy), generating a strategy feasibility assessment result. In the example, the comprehensive score of strategy 1 = 0.4×80 + 0.3×100 + 0.3×85 = 32 + 30 + 25.5 = 87.5 points (feasible), the comprehensive score of strategy 2 = 0.4×82 + 0.3×90 + 0.3×82 = 32.8 + 27 + 24.6 = 84.4 points (feasible), and the comprehensive score of strategy 3 = 0.4×85 + 0.3×95 + 0.3×78 = 34 + 28.5 + 23.4 = 85.9 points (feasible). Finally, 8 feasible strategies were selected, forming a feasibility assessment result, which clarifies the comprehensive score, scores of each indicator, and advantages and disadvantages of each feasible strategy.

[0074] Based on the feasibility assessment results of the strategies, the optimal resource allocation adjustment plan, technical route correction suggestions, and collaborative key point reorganization plan are selected and integrated into an adaptive project intervention strategy set. The effect is verified through strategy execution simulation, and finally, intelligent evaluation and early warning optimization of science and technology projects are achieved.

[0075] The core of this step is to select the optimal solution from feasible strategies, integrate them into a complete strategy set according to three intervention directions, verify the effect through simulation execution, ensure that the strategy can effectively mitigate risks, and achieve intelligent project assessment, early warning and optimization. The specific implementation method is as follows: The optimal strategy selection logic is based on the feasibility assessment results, combined with the optimization effects of three objective functions. A weighted comprehensive scoring method is used to conduct a secondary screening of eight feasible strategies, selecting the optimal solutions for three intervention directions (resource allocation adjustment, technical route correction, and collaborative key point reorganization), while ensuring that the three types of solutions are compatible and conflict-free. The screening weights are set as follows: feasibility score 0.5, risk reduction rate 0.3, and resource consumption control rate 0.2 (resource consumption control rate = 1 - resource consumption / maximum allowable consumption). The optimal strategy comprehensive score is calculated as: optimal_score = 0.5 × feasibility_score + 0.3 × risk_reduce + 0.2 × resource_control, where resource_control is the resource consumption control rate; the higher the score, the better the strategy.

[0076] In the example, after secondary screening, the optimal solutions for the three intervention directions were determined: the resource allocation adjustment plan is "supplementing R&D funds of 550,000 yuan, prioritizing investment in intelligent risk control model R&D, and allocating 2 high-performance servers to assist in algorithm optimization," with a feasibility score of 92 points, a risk reduction rate of 0.43, and a resource consumption control rate of 0.08 (550,000 yuan < 600,000 yuan maximum allowable consumption), and an optimal_score of 0.5×92+0.3×0.43+0.2×0.08≈46+0.129+0.016=46.145 points; the technical route correction suggestion is "optimizing the LSTM neural network algorithm parameters, introducing the Transformer algorithm module to assist in feature extraction, and eliminating traditional regression algorithms with poor adaptability." The "Legal Branch" has a feasibility score of 88, a risk reduction rate of 0.46, a resource consumption control rate of 0.03, and an optimal score of approximately 0.5×88+0.3×0.46+0.2×0.03≈44+0.138+0.006=44.144. The collaborative tackling key reorganization plan is to "reorganize the algorithm R&D group and the testing group, establish a special task force for intelligent risk control models, clarify the responsibilities of each group, and establish a weekly collaborative tackling meeting mechanism." Its feasibility score is 86, a risk reduction rate of 0.42, a resource consumption control rate of 0.1, and an optimal score of approximately 0.5×86+0.3×0.42+0.2×0.1≈43+0.126+0.02=43.146.

[0077] By integrating the three optimal solutions, and supplementing the strategy execution sequence (resource allocation adjustments are implemented first, technical route corrections are promoted simultaneously, and collaborative tackling of key reorganization runs through the entire process), the division of responsibilities (resource allocation is the responsibility of the finance team, technical corrections are the responsibility of the algorithm R&D team, and the reorganization plan is the responsibility of the project management team), and the execution nodes (resource allocation is in place within 1 week, technical corrections are completed within 3 weeks, and the reorganization plan is executed immediately), an adaptive project intervention strategy set is formed. This set can be dynamically adjusted according to changes in project risk to ensure the continuity of intervention effects.

[0078] The strategy execution simulation verification employs a historical data-based simulation method to build a project risk evolution simulation model. An adaptive project intervention strategy set is input, simulating project risk changes, resource consumption, and progress over the next 6 months (preset warning period) to verify whether the strategies meet the preset objectives. During the simulation, the risk accumulation level, total resource consumption, and schedule delay duration at high-risk nodes are monitored in real time, and various indicators are compared before and after intervention. In the example, the simulation results show that after intervention, the risk accumulation level of the "intelligent risk control model" decreased to 0.98 (<1.2, down to medium risk), with a risk reduction rate of 0.41, meeting the objective; the total resource consumption was 520,000 yuan (<600,000 yuan), meeting the resource consumption control standard; the schedule delay was 2 days (<7 days), not affecting the overall project progress, and the overall project risk level decreased from high risk to medium risk, achieving the expected intervention effect.

[0079] After successful simulation and verification, the set of adaptive project intervention strategies will be officially implemented. At the same time, combined with the full-link monitoring module, the implementation effect of the strategies will be tracked in real time, and the strategy parameters will be dynamically adjusted according to changes in risk. Ultimately, intelligent evaluation and early warning optimization of science and technology projects will be achieved to ensure the smooth progress of projects and reduce risk losses.

[0080] Another embodiment of the present invention provides a smart assessment and risk warning system for technology projects based on big data and AI, see [link to relevant documentation]. Figure 3 The system may include: The acquisition module 301 is used to collect multi-source heterogeneous science and technology project data, including structured application data and unstructured R&D process data, and to convert unstructured text, image and time-series log data into a unified semantic vector representation through cross-modal alignment technology, generating a normalized multimodal project feature set. Module 302 is used to automatically extract technical entities, R&D teams, resource elements and their dynamic relationships based on the normalized multimodal project feature set and using a self-supervised learning algorithm to construct a domain knowledge graph that evolves over time. The simulation module 303 is used to input the domain knowledge graph into a pre-trained dynamic graph neural network risk propagation model to simulate the transmission path and superposition effect of multiple risk factors in the project network, and output multi-level risk warning signals within a preset period in the future. The generation module 304 is used to generate an adaptive project intervention strategy set based on the multi-level risk warning signal using a multi-objective optimization algorithm. The strategy set covers resource allocation adjustment, technical route correction and collaborative key point reorganization, so as to realize intelligent evaluation and early warning optimization of science and technology projects.

[0081] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0082] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0083] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0084] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for intelligent assessment and risk warning of technology projects based on big data and AI, characterized in that, The method includes: Collect multi-source heterogeneous science and technology project data, including structured application data and unstructured R&D process data, and transform unstructured text, image and time-series log data into unified semantic vector representations through cross-modal alignment technology to generate a normalized multimodal project feature set; Based on the normalized multimodal project feature set, a self-supervised learning algorithm is used to automatically extract technical entities, R&D teams, resource elements and their dynamic relationships to construct a domain knowledge graph that evolves over time. The domain knowledge graph is input into a pre-trained dynamic graph neural network risk propagation model to simulate the transmission path and superposition effect of multiple risk factors in the project network, and output multi-level risk warning signals within a preset period in the future. Based on the multi-level risk warning signals, an adaptive project intervention strategy set is generated using a multi-objective optimization algorithm. The strategy set covers resource allocation adjustment, technical route correction, and collaborative key point reorganization, so as to realize intelligent evaluation and early warning optimization of science and technology projects.

2. The method according to claim 1, characterized in that, The collection of multi-source, heterogeneous science and technology project data includes structured application data and unstructured R&D process data. Cross-modal alignment technology is used to transform unstructured text, images, and time-series log data into a unified semantic vector representation, generating a normalized multimodal project feature set, including: By extracting structured application form data and unstructured R&D documents, design drawings and time sequence operation logs through API interfaces and data acquisition engines, the original multi-source heterogeneous dataset is generated. Data cleaning and standardization are performed on the original multi-source heterogeneous dataset. Data integrity verification and outlier correction algorithms are used for structured data. Named entity recognition technology is used to extract key entities for unstructured text data. Image enhancement algorithms are used to improve the clarity of image data. Timestamp alignment algorithms are used to unify the time base for time-series log data, generating cleaned and standardized multimodal data. The cleaned and standardized multimodal data is input into the cross-modal alignment coding network. The multimodal Transformer model with attention mechanism maps text semantics, image features and temporal patterns to a unified high-dimensional semantic space to generate an initial alignment semantic vector. The initial aligned semantic vector is processed by feature normalization. Layer normalization technique is used to eliminate the dimensional differences of features of different modalities. Key information is retained by feature dimensionality reduction algorithm, and finally a normalized multimodal item feature set is generated.

3. The method according to claim 2, characterized in that, The normalized multimodal project feature set is used to automatically extract technical entities, R&D teams, resource elements, and their dynamic relationships using a self-supervised learning algorithm, constructing a domain knowledge graph that evolves over time, including: Technical term vectors, R&D personnel feature vectors, and resource identifier vectors are extracted from the normalized multimodal project feature set. A self-supervised contrastive learning algorithm is used to train the entity recognition model to generate an initial entity set of technical entities, R&D teams, and resource elements. Based on the initial entity set, a self-supervised relationship prediction task is designed. The potential association patterns between entities are learned through the masked entity prediction algorithm to generate an entity relationship prediction model. The entity relationship prediction model is used to analyze the temporal evolution patterns of multimodal project feature sets, identify the creation, strengthening and disappearance processes of entity relationships under different time slices, and generate dynamic relationship evolution sequences. By integrating the initial entity set and the dynamic relationship evolution sequence, a temporal knowledge graph construction algorithm is used to organize entities and relationships according to the time dimension, and finally construct a domain knowledge graph that evolves over time.

4. The method according to claim 3, characterized in that, The process of inputting the domain knowledge graph into a pre-trained dynamic graph neural network risk propagation model simulates the transmission paths and superposition effects of multiple risk factors in the project network, and outputs multi-level risk warning signals within a preset future period, including: The domain knowledge graph that evolves over time is represented as a dynamic graph sequence data, with each time step containing entity node features and relation edge weights, resulting in a dynamic graph sequence for input. Load a pre-trained dynamic graph neural network risk propagation model. This model adopts a spatiotemporal graph convolutional network architecture, which can capture the spatiotemporal dependencies between nodes in the graph for risk propagation calculation. By inputting dynamic graph sequences into a dynamic graph neural network risk propagation model, the transmission process of technical risks, resource risks, and team collaboration risks in the project network is simulated through a multi-hop message passing mechanism, generating risk propagation simulation results. Based on the results of risk transmission simulation, the risk superposition effect analysis algorithm is used to predict the degree of risk accumulation of each entity node in the future within a preset period. The risk levels are divided into high, medium and low risk levels according to the risk level threshold, and finally multi-level risk warning signals are output.

5. The method according to claim 4, characterized in that, Based on the multi-level risk warning signals, an adaptive project intervention strategy set is generated using a multi-objective optimization algorithm. This strategy set encompasses resource allocation adjustments, technical route corrections, and collaborative tackling of key challenges, aiming to achieve intelligent evaluation and early warning optimization of science and technology projects. This includes: Analyze multi-level risk warning signals, extract high-risk entity nodes and their relationships, construct a multi-objective optimization problem for risk mitigation, with objective functions including maximizing risk reduction, minimizing resource consumption, and minimizing the impact on project schedule, and generate a multi-objective optimization model. A non-dominated sorting genetic algorithm is used to solve the multi-objective optimization model. The Pareto optimal solution set is searched through population evolution iteration to generate a set of candidate intervention strategies. Based on the set of candidate intervention strategies, a strategy feasibility assessment index system is constructed, including assessments of technical feasibility, resource accessibility, and team execution capability, and strategy feasibility assessment results are generated. Based on the feasibility assessment results of the strategies, the optimal resource allocation adjustment plan, technical route correction suggestions, and collaborative key point reorganization plan are selected and integrated into an adaptive project intervention strategy set. The effect is verified through strategy execution simulation, and finally, intelligent evaluation and early warning optimization of science and technology projects are achieved.

6. A smart assessment and risk warning system for technology projects based on big data and AI, characterized in that, The system includes: The data acquisition module is used to collect multi-source heterogeneous science and technology project data, including structured application data and unstructured R&D process data. It also uses cross-modal alignment technology to transform unstructured text, images, and time-series log data into a unified semantic vector representation, generating a normalized multimodal project feature set. The module is used to automatically extract technical entities, R&D teams, resource elements and their dynamic relationships based on the normalized multimodal project feature set and using a self-supervised learning algorithm to construct a domain knowledge graph that evolves over time. The simulation module is used to input the domain knowledge graph into a pre-trained dynamic graph neural network risk propagation model to simulate the transmission path and superposition effect of multiple risk factors in the project network, and output multi-level risk warning signals within a preset period in the future. The generation module is used to generate an adaptive project intervention strategy set based on the multi-level risk warning signals using a multi-objective optimization algorithm. The strategy set covers resource allocation adjustment, technical route correction, and collaborative key point reorganization, so as to realize intelligent evaluation and early warning optimization of science and technology projects.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: By extracting structured application form data and unstructured R&D documents, design drawings and time sequence operation logs through API interfaces and data acquisition engines, the original multi-source heterogeneous dataset is generated. Data cleaning and standardization are performed on the original multi-source heterogeneous dataset. Data integrity verification and outlier correction algorithms are used for structured data. Named entity recognition technology is used to extract key entities for unstructured text data. Image enhancement algorithms are used to improve the clarity of image data. Timestamp alignment algorithms are used to unify the time base for time-series log data, generating cleaned and standardized multimodal data. The cleaned and standardized multimodal data is input into the cross-modal alignment coding network. The multimodal Transformer model with attention mechanism maps text semantics, image features and temporal patterns to a unified high-dimensional semantic space to generate an initial alignment semantic vector. The initial aligned semantic vector is processed by feature normalization. Layer normalization technique is used to eliminate the dimensional differences of features of different modalities. Key information is retained by feature dimensionality reduction algorithm, and finally a normalized multimodal item feature set is generated.

8. The system according to claim 7, characterized in that, The building module is specifically used for: Technical term vectors, R&D personnel feature vectors, and resource identifier vectors are extracted from the normalized multimodal project feature set. A self-supervised contrastive learning algorithm is used to train the entity recognition model to generate an initial entity set of technical entities, R&D teams, and resource elements. Based on the initial entity set, a self-supervised relationship prediction task is designed. The potential association patterns between entities are learned through the masked entity prediction algorithm to generate an entity relationship prediction model. The entity relationship prediction model is used to analyze the temporal evolution patterns of multimodal project feature sets, identify the creation, strengthening and disappearance processes of entity relationships under different time slices, and generate dynamic relationship evolution sequences. By integrating the initial entity set and the dynamic relationship evolution sequence, a temporal knowledge graph construction algorithm is used to organize entities and relationships according to the time dimension, and finally construct a domain knowledge graph that evolves over time.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images