Science and technology project intelligent management method, device, equipment and medium

Abstract

The invention relates to an intelligent management method and device for science and technology projects, equipment and a medium. The method comprises the following steps: acquiring full-life-cycle heterogeneous data of science and technology projects, and preprocessing the full-life-cycle heterogeneous data of the science and technology projects to obtain project ontologies; embedding entity dynamic state time sequence data into the project ontology, and constructing to obtain a semantic-time sequence coupling knowledge graph; utilizing a semantic-time sequence coupling knowledge graph to constrain a causal exploration space, and performing ontology causal structure extraction to obtain causal priori; performing multi-strategy time sequence simulation based on a project digital twin model to obtain expected index distribution of each strategy; performing causal enhancement risk prediction according to the semantic-time sequence coupling knowledge graph, causal priori and expected index distribution of each strategy to obtain a risk prediction report and a causal interpretation report; and based on compliance requirements, performing multi-objective optimization to obtain a project association scheduling scheme. By adopting the method, causal traceability and simulation pre-judgment can be realized, and compliance and high-efficiency promotion of science and technology projects can be guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of information data management, and in particular relates to a method, device, equipment and medium for intelligent management of scientific and technological projects. Background Technology

[0002] With the development of information and intelligent technologies, project management has gradually evolved from traditional manual management and paper-based record-keeping to full-process control using information platforms and intelligent tools. This has led to the emergence of project management technologies based on information systems, which rely on project management software or collaboration platforms to achieve functions such as task decomposition, progress tracking, resource allocation, and results statistics, thereby assisting managers in making decisions.

[0003] Traditional project management typically involves setting project schedules, key milestones, and resource allocation lists. Monitoring project progress relies heavily on manual data entry and periodic checks, while risk identification and mitigation primarily depend on managers' experience and human judgment. When multiple sub-projects, cross-departmental collaboration, or external dependencies are involved, traditional methods often rely on meetings and manual coordination.

[0004] However, the aforementioned methods often result in data from different sources being scattered across various systems or tables, lacking a unified data model and hindering real-time aggregation and comprehensive analysis. Traditional management methods rely heavily on historical experience and manual judgment, lacking the ability to predict risks and extrapolate outcomes based on large-scale data and intelligent algorithms, making it difficult to provide timely warnings of project delays or resource bottlenecks. Existing project management systems primarily consist of static reports and progress tracking, failing to provide dynamic resource reallocation suggestions or optimized scheduling solutions, requiring managers to still rely on manual complex decision-making. Significant information silos exist between different projects or organizations, lacking unified data sharing and privacy protection mechanisms, making it difficult to transfer and reuse knowledge and experience on a larger scale. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, equipment, and medium for intelligent management of science and technology projects that can support data fusion, management support, and cross-organizational collaboration, addressing the aforementioned technical problems.

[0006] Firstly, this application provides an intelligent management method for science and technology projects, including:

[0007] Acquire heterogeneous data throughout the entire lifecycle of science and technology projects, and preprocess this data to obtain the project ontology. The heterogeneous data includes project metadata, resource data, text records, and compliance requirements. The project ontology includes entities, entity attributes, and static relationships between entities.

[0008] The dynamic state time-series data of entities is embedded into the project ontology, and a semantic-temporal coupled knowledge graph is constructed. The dynamic state time-series data of entities includes static relationships between entities that change over time. The semantic-temporal coupled knowledge graph includes project entity nodes and entity relationship edges.

[0009] Based on historical project data, we use semantic-temporal coupled knowledge graph to constrain the causal exploration space and obtain causal priors by extracting ontology causal structure from project entity nodes.

[0010] A digital twin model of the project is constructed based on a semantic-temporal coupled knowledge graph, causal priors, and resource data. Multi-strategy temporal simulation is then performed based on the digital twin model to obtain the expected index distribution of each strategy.

[0011] Based on the semantic-temporal coupled knowledge graph, causal priors, and the expected index distribution of each strategy, causal enhancement risk prediction is performed, resulting in a risk prediction report and a causal explanation report.

[0012] Based on compliance requirements, a project-related scheduling scheme is obtained through multi-objective optimization using expected indicator distribution, risk prediction reports, and causal explanation reports.

[0013] In one embodiment, heterogeneous data throughout the entire lifecycle of a science and technology project is preprocessed to obtain the project ontology, including:

[0014] Standardized cleansing is performed on heterogeneous data throughout the entire lifecycle of science and technology projects to obtain standardized science and technology project data.

[0015] Standardized science and technology project data are processed by a rule engine based on data type to extract entities and classify them into categories, resulting in a preliminary entity set. Data types include structured data, semi-structured data, and unstructured data.

[0016] Based on a pre-defined ontology mapping rule library for science and technology projects, the initial set of entities is mapped to the ontology of science and technology projects, resulting in ontology entity instances and a mapping rule table; the mapping rule table includes mapping confidence.

[0017] If the mapping confidence level is lower than the preset confidence threshold, a manual annotation instruction is generated; the manual annotation instruction is used to indicate the sending of a manual standard confirmation instruction and to obtain feedback on the manual entity annotation information.

[0018] The mapping rule table and ontology entity instances are updated based on the manually labeled entity information, and the updated mapping rule table and ontology entity instances are integrated to obtain the project ontology.

[0019] In one embodiment, dynamic state time-series data of entities are embedded into the project ontology, and a semantic-temporal coupled knowledge graph is constructed, including:

[0020] The entities are mapped to project entity nodes, and the static relationships between entities are mapped to entity relationship edges, resulting in the initial graph structure; project entity nodes include entity attributes;

[0021] By temporally binding entity attributes and entity relationship edges in the initial graph structure, a temporal graph structure is obtained.

[0022] Based on the entity behavior rules of science and technology projects, behavioral patterns are embedded into the entity nodes of each project in the time sequence graph structure to obtain a semantic-temporal coupled knowledge graph.

[0023] In one embodiment, based on historical project data, a semantic-temporal coupled knowledge graph is used to constrain the causal exploration space, and causal priors are obtained by extracting the ontology causal structure of project entity nodes, including:

[0024] Based on the semantic-temporal coupled knowledge graph, entity pairs with causal influence are selected, and hard and soft constraints are applied to the entity pairs to obtain causal candidate templates. Hard constraints correspond to the mandatory causal direction in the field of science and technology projects; soft constraints correspond to assigning weights to the causal relationship strength of entity pairs.

[0025] Extract time-synchronized feature sequences corresponding to causal candidate templates from the semantic-temporal coupled knowledge graph to obtain a feature pair dataset; the feature pair dataset includes causal features and outcome features.

[0026] By eliminating entity pairs in the causal candidate template whose cause and effect features do not conform to the causal direction through constraint-based causal orientation, and retaining entity pairs in the causal candidate template whose causal association strength weights meet the preset threshold conditions through scoring-based causal orientation, a local causal structure is obtained.

[0027] Using the fine-tuned small sample corresponding to the local causal structure as the prior, the causal effect of the local causal structure is calculated to obtain the causal prior.

[0028] In one embodiment, a digital twin model of the project is constructed based on a semantic-temporal coupled knowledge graph, causal priors, and resource data. Multi-strategy temporal simulation is then performed based on the project digital twin model to obtain the expected index distribution of each strategy, including:

[0029] The semantic-temporal coupled knowledge graph, causal priors, and resource data are decomposed into entity elements, their relational logic is sorted out, and behavioral rules are extracted to obtain a digital twin model of the project. The digital twin model of the project includes entity modules, relation modules, behavior modules, resource constraint modules, and causal intervention modules that are interconnected through data links.

[0030] Multiple project management strategies are derived through counterfactual reasoning based on project objectives and causal priors; these strategies include intervention targets, intervention methods, intervention timing, and intervention parameter values.

[0031] Based on the digital twin model of the project, the project management strategies are simulated to obtain the project timeline evolution trajectory corresponding to each project management strategy; the simulation corresponds to resource status updates, behavior pattern triggering, causal effect impact and data link synchronization;

[0032] The expected indicator distribution is obtained by extracting and statistically analyzing the time-series evolution trajectory of each project.

[0033] In one embodiment, causal enhancement risk prediction is performed based on a semantic-temporal coupled knowledge graph, causal priors, and the expected indicator distribution of each strategy, resulting in a risk prediction report and a causal explanation report, including:

[0034] Feature extraction is performed on the semantic-temporal coupled knowledge graph, causal priors, distributions of expected indicators, and text records, and the extracted features are standardized to obtain the prediction feature matrix; the semantic-temporal coupled knowledge graph corresponds to the extraction of dynamic trend features of the project and cross-entity impact features; the causal priors correspond to the extraction of causal impact strength features; the distributions of expected indicators correspond to the extraction of potential effect features of the strategy; and the text records correspond to the extraction of risk semantic features.

[0035] The predicted feature matrix is ​​input into a pre-trained causal interpretable hybrid prediction model to obtain the risk prediction result; the causal interpretable hybrid prediction model is composed of a temporal feature encoder, a text feature encoder and a causal correction layer connected in sequence.

[0036] Based on the risk prediction results, risk types are classified and risk time trend analysis is performed to obtain a risk prediction report;

[0037] Based on the risk prediction report, causal priors, and semantic-temporal coupled knowledge graph relationship edges, a local causal contribution analysis and counterfactual intervention suggestions are conducted to obtain a causal explanation report.

[0038] In one embodiment, based on compliance requirements, a project-related scheduling scheme is obtained through multi-objective optimization based on the expected indicator distribution, risk prediction report, and causal explanation report, including:

[0039] Transform compliance requirements and resource data into project compliance rules and resource limits to obtain optimization goals and constraints.

[0040] Based on the optimization objective, reinforcement learning and mixed-integer programming are performed according to the expected indicator distribution, risk prediction report and causal explanation report to obtain the scheduling strategy set;

[0041] The scheduling strategy set is security verified based on optimization constraints to obtain a security verification report; the security verification report includes pass and fail.

[0042] The scheduling strategies that have passed the security verification report will be integrated into a project-related scheduling scheme.

[0043] Secondly, this application also provides an intelligent management device for science and technology projects, comprising:

[0044] The data module is used to acquire heterogeneous data throughout the entire lifecycle of science and technology projects, and to preprocess the heterogeneous data to obtain the project ontology; the heterogeneous data of science and technology projects includes project metadata, resource data, text records and compliance requirements;

[0045] The knowledge graph module is used to embed dynamic state time-series data of entities into the project ontology and construct a semantic-temporal coupled knowledge graph; the dynamic state time-series data of entities includes static relationships between entities that change over time; the semantic-temporal coupled knowledge graph includes project entity nodes and entity relationship edges;

[0046] The causal module is used to constrain the causal exploration space based on historical project data using a semantic-temporal coupled knowledge graph, and to obtain causal priors by extracting the ontology causal structure of project entity nodes.

[0047] The management simulation module is used to construct a digital twin model of the project based on semantic-temporal coupled knowledge graphs, causal priors, and resource data, and to perform multi-strategy temporal simulation based on the project digital twin model to obtain the expected index distribution of each strategy;

[0048] The prediction module is used to perform causal enhancement risk prediction based on the semantic-temporal coupled knowledge graph, causal priors, and the expected indicator distribution of each strategy, and to obtain risk prediction reports and causal explanation reports.

[0049] The management module is used to obtain a project-related scheduling scheme by performing multi-objective optimization based on compliance requirements, expected indicator distribution, risk prediction report, and causal explanation report.

[0050] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-mentioned intelligent management methods for scientific and technological projects.

[0051] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the above-mentioned intelligent management methods for scientific and technological projects.

[0052] The aforementioned intelligent management methods, devices, equipment, and media for science and technology projects, by covering heterogeneous data throughout the entire lifecycle of projects, including metadata, resource data, text records, and compliance requirements, and outputting a project ontology containing entities, entity attributes, and static relationships after preprocessing, break down the barriers of data being scattered across different systems and having incompatible formats in traditional management. This achieves semantic unification of heterogeneous data throughout the entire lifecycle, solving the problem of data fragmentation. By embedding dynamic state time-series data of entities on the project ontology, a semantic-temporal coupled knowledge graph is formed. This graph preserves the semantic relationships between entities and traces the temporal evolution of entity states and relationships, enabling dynamic traceability of projects. It allows project managers to quickly query the entity states and relationship changes at a specific historical point in time, providing a basis for project review and problem tracing. The semantic-temporal coupled knowledge graph constrains the causal exploration space, avoiding the blind search of traditional causal discovery. Furthermore, by extracting the ontology causal structure based on historical data, it generates transferable causal priors, solving the cold start and predictive interpretability problems of new projects. For newly launched science and technology projects, causal priors can be quickly reused without large amounts of local data, shortening the project management startup cycle and providing interpretable causal logic, rather than relying on black-box models. Through digital twin multi-strategy simulation, proactive management based on simulation before decision-making is achieved. This helps managers compare the advantages and disadvantages of different strategies before actual implementation, avoiding resource waste or schedule delays caused by trial and error, and improving the scientific nature of decision-making. Risk prediction reports can provide early warnings of potential project risks, while causal explanation reports provide clear directions for risk intervention, preventing managers from knowing the risks but not knowing how to respond. Using compliance requirements as a hard constraint, and combining expected indicator distribution, risk prediction, and causal explanation for multi-objective optimization, the generated project-related scheduling scheme not only meets regulatory and compliance rules but also takes into account objectives such as schedule and cost. This is suitable for the high compliance characteristics of technology projects, avoiding project suspension due to scheduling scheme violations. Furthermore, multi-objective optimization balances various management objectives, improving resource utilization and project success rate. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating the intelligent management method for scientific and technological projects according to the present invention;

[0055] Figure 2 This is a flowchart illustrating the steps of step S103.

[0056] Figure 3This is a flowchart illustrating the steps of step S104.

[0057] Figure 4 This is a structural diagram of the intelligent management device for scientific and technological projects according to the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] In one embodiment, such as Figure 1 As shown, a method for intelligent management of science and technology projects is provided. This embodiment illustrates the method by applying it to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0060] S101. Obtain heterogeneous data of the entire lifecycle of science and technology projects, and preprocess the heterogeneous data of the entire lifecycle of science and technology projects to obtain the project ontology; the heterogeneous data of science and technology projects includes project metadata, resource data, text records and compliance requirements; the project ontology includes entities, entity attributes and static relationships between entities.

[0061] Heterogeneous data throughout the entire lifecycle of a technology project refers to a data set covering all stages of a technology project, from initiation and execution to final acceptance, with significant differences in source, format, and type. Illustratively, project metadata is structured information describing the core framework of the project. For example, project identifiers uniquely identify the project, sub-projects reflect the internal hierarchy, milestones define key time nodes, and task tables contain details such as task identifiers, dependencies between tasks, estimated working hours, responsible persons, and milestone dependency types, forming the basic framework for project schedule management. Resource data focuses on various elements supporting project operation, covering human resources, equipment, and finance, directly determining the resource guarantee capability for project execution. Text records are carriers of unstructured project information, including email correspondence, meeting minutes, and requirement specifications, containing a wealth of implicit information influencing project decisions. Compliance requirements are the external regulations and internal norms that the project must adhere to, such as review deadlines set by regulatory agencies and the submission status of necessary documents, which are prerequisites for the legal and compliant progress of the project.

[0062] Furthermore, in science and technology project management, heterogeneous data is often stored in different systems, with incompatible formats and issues such as redundancy, missing data, and anomalies. Preprocessing is essential to transform this heterogeneous data into a project ontology with a unified semantic system, laying the data foundation for intelligent management throughout the entire process. Specifically, the preprocessing process proceeds step-by-step with data standardization, entity extraction, and semantic association as its core logic. For example, different types of data undergo differentiated processing for data cleaning and format unification. Further, entity-attribute mapping is performed based on a pre-defined science and technology project domain ontology framework. This framework includes common concepts and hierarchical relationships such as tasks, milestones, resources, experimental data, and compliance checkpoints. The cleaned data sources are matched with these ontology concepts. Simultaneously, static relationships between entities are established, such as tasks depending on resource allocation and compliance checkpoints constraining tasks, ultimately integrating to form the project ontology.

[0063] S102. Embed entity dynamic state time-series data into the project ontology and construct a semantic-temporal coupled knowledge graph; entity dynamic state time-series data includes entity static association relationships that change over time; semantic-temporal coupled knowledge graph includes project entity nodes and entity relationship edges.

[0064] Indicatively, a Semantic-Temporal Coupled Knowledge Graph (STKG) is a knowledge carrier that integrates semantic associations and temporal features. Its core structure consists of project entity nodes containing static and dynamic temporal attributes, and entity relationship edges containing static associations and temporal constraints. It retains the structured description of entity semantic relationships found in traditional knowledge graphs while dynamically tracking entity states and relationship evolution through temporal data embedding. Entity dynamic state temporal data refers to entity attributes and relationship data that change over time, such as daily task progress, hourly resource availability, experimental parameters changing over time, and entity relationships adjusting over time. Specifically, entities in the project ontology are mapped to STKG project entity nodes, each containing complete attributes; static relationships in the ontology are mapped to entity relationship edges, and relationship types and association strength weights are labeled, forming an initial graph structure containing only semantic associations. Furthermore, the process of embedding entity dynamic status time-series data follows a path of timestamp association, dynamic attribute binding, and time-series index construction. A timestamp tag is added to each entity's dynamic attribute to form a triple of attribute-timestamp-value, ensuring that the data is traceable to a specific point in time. The time-series triple is bound to the corresponding entity node in the project ontology, so that the entity node has both static attributes and dynamic time-series attributes. A time-series index is constructed, and the entity identifier is associated with a time window to support quick querying of the entity's dynamic status within a certain time period.

[0065] S103. Based on historical project data, the causal exploration space is constrained by semantic-temporal coupled knowledge graph, and causal priors are obtained by extracting ontology causal structure from project entity nodes.

[0066] Historical project data refers to the full-process data of similar completed or ongoing scientific and technological projects, covering project metadata, resource consumption records, progress execution trajectory, risk event logs, causal analysis conclusions, etc. For newly launched scientific and technological projects, there is a lack of local data to support causal relationship mining in the early stages, while historical project data can provide causal patterns in similar scenarios. For example, based on the semantic associations of STKG (Site-Based Knowledge Group), causal candidate filtering can be performed. The categories and relationship edges of entity nodes in STKG reflect the domain logic, which can be used to filter entity pairs that may have causal influence, excluding combinations without logical connection, such as financial budgets and experimental reagent models, which usually have no direct causal relationship. Furthermore, based on the temporal characteristics of STKG, a preliminary judgment of causal direction can be made. Causal relationships must satisfy the temporal order of cause preceding effect, and the timestamps of the dynamic status time-series data of entities in STKG can provide key evidence. For example, STKG records that equipment failure occurs at time t1, and task progress stalls at time t2, and... This allows us to initially determine the causal relationship between equipment failure and task stagnation, eliminate the reverse correlation between task stagnation and equipment failure, and further narrow down the scope of our investigation.

[0067] Furthermore, ontology causal structure extraction is a process of gradually obtaining causal priors through data verification, structure optimization, and effect quantification within a constrained causal exploration space. Optionally, constrained causal verification is performed first, using historical project data to test the conditional independence of candidate entity pairs. For example, for candidate pairs of resource availability rate pointing to task delay days, after controlling for confounding variables such as personnel skill level and task complexity, it is tested whether there is still a significant correlation between the two. If there is a correlation, the candidate pair is retained; if there is no correlation, it is excluded. At the same time, the causal direction constraint is strictly followed to ensure that the verification results conform to the domain logic.

[0068] Furthermore, a scoring-based causal structure optimization is performed. For validated candidate entity pairs, the Bayesian Information Criterion (BIC) or the Akaike Information Criterion (AIC) is used to calculate the causal structure score. The score comprehensively considers the explanatory power of the causal relationship on the data and the number of parameters. If a candidate entity pair can better explain the changes in the outcome features and has fewer parameters, the score is higher. Combining the priority weights derived from the scores, the truly existing and significantly influential causal relationships are finally selected to form a local causal structure.

[0069] For each causal edge in the local causal structure, Bayesian structural learning is used for small-sample fine-tuning using historical project data and a small amount of local data from new projects. This involves using the local causal structure as a structural prior and updating the posterior confidence score using a Markov Chain Monte Carlo (MCMC) algorithm. For example, if the correlation between decreased resource availability and increased task delay days is stronger in the local data of a new project, the confidence score of that causal edge is increased from 0.85 to 0.92. Simultaneously, propensity score matching (PSM) or dual / re-biased machine learning is used to estimate the causal effect value and 95% confidence interval, quantifying the magnitude and uncertainty of the causal impact. Finally, the local causal structure, causal effect value, confidence score, and confidence interval are integrated to generate a standardized causal prior.

[0070] S104. Construct a digital twin model of the project based on the semantic-temporal coupled knowledge graph, causal priors and resource data, and perform multi-strategy temporal simulation based on the project digital twin model to obtain the expected index distribution of each strategy.

[0071] A project digital twin model is a digital mirror of all elements and processes of a technology project. Based on real-time data and causal logic, it can simulate the project evolution trajectory under different strategies, helping managers to simulate before making decisions and avoid trial-and-error costs. For example, strategies need to be combined with the project's core objectives, such as schedule assurance, cost control, and risk avoidance, and with counterfactual reasoning results based on causal priors, to generate multi-dimensional and differentiated intervention strategies. Each strategy needs to clearly define four elements: the intervention target, the intervention method, the intervention time, and the intervention parameter values. The simulation execution phase is based on time steps. Technology projects are typically set to one time step per day. Each strategy is executed sequentially in the digital twin model. Specifically, first, resource status is updated, adjusting the supply and demand data of the resource constraint module according to the strategy parameters; second, behavioral patterns are triggered, calling the behavior module to execute entity state transitions; then, causal effects are calculated, simulating the chain reactions triggered by the strategy through the causal intervention module; finally, the data link is synchronized, synchronizing the state changes of all modules in real time to ensure that the simulation results of each time step are based on the latest state. To cover uncertain scenarios, an additional 100-200 piecewise Monte Carlo simulations are performed for each strategy. Small random perturbations are added to the uncertainty parameters during each simulation to ensure that the simulation results reflect the risk boundary.

[0072] The results statistical phase extracts core project management indicators from the time-series evolution of each strategy, including schedule indicators, cost indicators, research output indicators, and compliance indicators. Statistical analysis is performed on each indicator; for example, the mean, quantiles, and probability distribution of the indicators are calculated, and the differences in indicators across strategies are displayed through charts, ultimately forming a multi-strategy expected indicator distribution table.

[0073] S105. Based on the semantic-temporal coupled knowledge graph, causal priors, and the expected index distribution of each strategy, perform causal enhancement risk prediction to obtain a risk prediction report and a causal explanation report.

[0074] This diagram illustrates the extraction of dynamic trend features and cross-entity impact features from STKG (Site-Driven Knowledge Group). Dynamic trend features are calculated using a sliding window to measure the statistics of entity time-series data, reflecting the current dynamics of the project. Cross-entity impact features are calculated based on the relationship edges in STKG, reflecting the potential impact of inter-entity relationships on risk. Causal impact strength features are extracted from causal priors, directly reading the confidence scores of causal edge effects in the causal priors to quantify the degree of influence of key factors on risk. Potential effect features of strategies are extracted from the distribution of expected indicators for each strategy, calculating the deviation of the current actual state from the optimal strategy, and the mean and extreme values ​​of risk probabilities, reflecting the risk boundaries under different strategies. Risk semantic features are extracted from text records, using domain-adjusted Transformer models such as BERT to perform entity recognition and sentiment analysis on meeting minutes and risk reports, outputting risk keyword frequencies and risk sentiment scores to capture implicit risk information in the text. All features are standardized and integrated into a predictive feature matrix containing time-granularity labels.

[0075] For example, causal-enhanced risk prediction is achieved by constructing a hybrid prediction model consisting of a temporal feature encoder, a textual feature encoder, and a causal correction layer. The temporal feature encoder uses a Temporal GNN, such as TGAT, to handle dynamic trends and cross-entity impact features, leveraging the graph structure modeling capabilities of GNNs to capture dynamic relationships between entities and output temporal embedding vectors. The textual feature encoder uses a fine-tuned Transformer model to handle risk semantic features, transforming unstructured text into structured semantic embedding vectors. The causal correction layer fuses the two types of embedding vectors with causal features from causal priors, adjusting feature weights through a causal attention mechanism. Specifically, higher attention weights are assigned to causal features with high confidence in the causal priors, while a causal constraint loss function is introduced. If the causal priors indicate that an increase in resource availability points to a decrease in risk, the model prediction must satisfy a negative correlation between the two; otherwise, a penalty loss is imposed to ensure that the model prediction results do not violate causal logic and avoid being misled by confounding variables.

[0076] Optionally, by combining STKG's real-time streaming data, such as newly added equipment failure events and reagent delivery delay notifications within the past hour, the prediction results can be continuously optimized. In other words, the real-time streaming data can be transformed into real-time feature increments and input into the trained hybrid prediction model for incremental updates. The model can quickly calculate the impact of real-time events on risk based on causal prior effect values, and the update time can be controlled within seconds, ensuring that the prediction results can reflect the latest risk status of the project in real time.

[0077] A risk forecasting report is a structured presentation of forecast results. It can be organized by risk type, time granularity, quantitative indicators, and scope of impact, ensuring managers have a clear understanding of the overall project risks. Risk types can be categorized as schedule risk, cost risk, compliance risk, and research output risk. The schedule risk section indicates the probability of default on core milestones and tasks, expected delays, and peak risk times. The cost risk section includes the probability of total budget overruns, expected overrun amounts, and key overrun items. The compliance risk section clarifies the probability of approval at each compliance review stage and potential violations. The research output risk section predicts experiment completion rates and sample delivery times. The report can also include risk trend charts and scope of impact analysis to provide early warnings, helping managers identify potential project risks in a timely manner and avoid being caught unprepared when risks suddenly arise.

[0078] The causal explanation report serves as an attribution and intervention guide for risk prediction results. It needs to be generated from three dimensions: causal contribution, counterfactual suggestions, and path visualization, based on the risk prediction report, causal priors, and the STKG relationship edges. This addresses the problem of knowing the risk but not understanding why it exists or how to intervene. Specifically, local causal contribution analysis calculates the causal contribution of each feature to the risk prediction results. For example, by combining the attention weights of the hybrid prediction model with the effect values ​​of the causal priors, the contribution is normalized, and the top 5 core influencing features and their contributions are output. Counterfactual intervention suggestions are crucial for the report's implementation. Based on the effect values ​​of the causal priors, they simulate the risk improvement effect after changing a certain feature value. For example, features with high and controllable causal contributions are selected, and the specific effects of intervention measures are calculated. Furthermore, causal path visualization is an intuitive supplement to the report, drawing the risk transmission path based on the STKG relationship edges and the causal graph of the causal priors.

[0079] S106. Based on compliance requirements, a project-related scheduling scheme is obtained through multi-objective optimization based on the expected indicator distribution, risk prediction report, and causal explanation report.

[0080] This is an illustrative example; the sources of multiple objectives need to cover four dimensions: compliance requirements, risk control, resource efficiency, and project output. Among these, compliance is the bottom-line objective, set based on compliance requirements and representing inviolable hard constraints. For example, all compliance review nodes must be completed by the deadline, research funding usage must not exceed 10% of the total budget, and experimental data recording completeness must reach 100%. Violations will result in project suspension or penalties, therefore, this objective must be prioritized in optimization. Risk control is the safeguard objective, set based on risk prediction reports, focusing on reducing high-probability, high-impact risks. The scheduling plan must control risks within an acceptable range. Resource efficiency is the optimization objective, set based on expected indicator distribution and resource data, focusing on improving resource utilization and reducing waste, maximizing resource value under the premise of compliance and controllable risk. Project output is the core objective, set based on the fundamental mission of scientific and technological projects, ensuring that the scheduling plan ultimately serves the delivery of project results.

[0081] Multi-objective optimization requires finding the optimal balance among conflicting objectives. Specifically, the generation of project-related scheduling schemes follows a core process of prioritizing compliance constraints, allocating multi-objective weights, intelligent optimization, and scheme verification. This ensures that the schemes not only comply with regulations but also achieve optimal multi-objective performance. For example, compliance requirements are transformed into mathematical constraints and embedded in the optimization model. These constraints are inviolable; if a scheme violates a constraint during optimization, it is immediately eliminated, ensuring the compliance foundation of all candidate schemes. Furthermore, the weights of each objective are determined using the Analytic Hierarchy Process (AHP) or expert scoring, combined with project priorities and management preferences. A hybrid algorithm of reinforcement learning (RL) and mixed-integer programming (MIP) is employed to model the scheduling problem as a Markov decision process (MDP). The states are the current resource status, task progress, and risk indicators; the actions are resource allocation adjustments, task priority adjustments, and time plan adjustments; and the reward is the comprehensive score of the multi-objectives—points are added for compliance and deducted for violation. Simulation data and historical scheduling data are used to train the multi-objective reinforcement learning model, generating multiple global scheduling candidate schemes. Furthermore, mixed integer programming (MIP) is used to locally optimize the candidate solutions. To address potential short-term resource conflicts in the RL solutions, MIP solves the time window allocation using integer variables to ensure resource conflicts-free operation. At the same time, MIP strictly embeds compliance constraints and risk thresholds to further screen locally feasible solutions and obtain the project-related scheduling solution.

[0082] The aforementioned intelligent management methods for science and technology projects include full lifecycle data coverage, ensuring no data omissions and supporting end-to-end management from project initiation to completion, avoiding management blind spots caused by missing data; dynamic temporal embedding enables the knowledge graph to reflect project status changes in real time, rather than being statically stored, meeting the management needs of frequent dynamic adjustments in science and technology projects; causal exploration space constraints reduce the calculation of invalid causal relationships, lowering the time cost and computing power consumption of causal discovery, and improving the efficiency of causal prior generation; ontology causal structure extraction ensures that causal priors conform to the logic of the science and technology project domain, avoiding causal relationships that violate common sense, and improving the credibility and reusability of causal priors. The multi-strategy expected indicator distribution output provides a quantitative basis for comparing the effects of subsequent scheduling, helping managers intuitively select the optimal strategy. Causal priors are used to correct prediction results, avoiding being misled by confounding variables in the data and improving the accuracy of risk prediction; risk prediction reports meet risk warning needs, and causal explanation reports meet risk intervention needs; the combination of the two forms a warning-response closed loop, improving risk management capabilities. Compliance constraints ensure that the scheduling plan complies with the regulatory and compliance rules of science and technology projects from the source, reducing the risk of project violations; multi-objective optimization takes into account both the optimal expected indicators and controllable risks, avoiding the loss of one aspect due to the optimization of a single objective, and generating a comprehensive optimal executable plan.

[0083] In one embodiment, heterogeneous data throughout the entire lifecycle of a science and technology project is preprocessed to obtain the project ontology, including:

[0084] S11. Standardize and clean heterogeneous data throughout the entire lifecycle of science and technology projects to obtain standardized science and technology project data.

[0085] For structured data throughout the entire lifecycle of science and technology projects, such as project metadata and resource ledgers; semi-structured data, such as meeting minutes and approval process sheets; and unstructured data, such as research reports and email correspondence, a differentiated cleaning strategy is employed to achieve data standardization. Specifically, for structured data, duplicate records are removed using field validation rules, missing values ​​such as equipment failure rates are corrected using mean-filling, and outliers are corrected based on business thresholds. Simultaneously, the time format is standardized to YYYY-MM-DDHH:MM:SS, and numerical units are standardized according to project specifications to ensure data format consistency. For semi-structured / unstructured data, text data such as meeting minutes and risk reports are processed using NLP tools for word segmentation and stop word removal, and key information such as time, personnel, and tasks is extracted using regular expressions. For semi-structured data such as approval process sheets, the correspondence between form fields and project ontology concepts is parsed, redundant format tags are removed, and the data is converted into structured fields. The final output is standardized science and technology project data with a uniform format, free of redundancy and outliers.

[0086] S12. Extract entities and classify categories from standardized science and technology project data according to the rule engine corresponding to the data type to obtain a preliminary entity set; the data types include structured data, semi-structured data, and unstructured data.

[0087] For example, for structured data, a preset field mapping rule engine is invoked to directly extract entities from project metadata and resource ledgers. For semi-structured data, a template matching rule engine is activated to extract compliance check nodes and personnel entities from approval process sheets, meeting minutes, and other metadata through field templates; experimental data entities are extracted from experimental records using keyword matching rules and categorized accordingly. For unstructured data entity extraction, a hybrid rule engine combining rules and a small amount of learning is enabled. For research reports and email texts, entities are initially located using domain keyword rules, then a fine-tuned NER model is invoked to extract compliance check node entities. Semantic similarity calculations are used to categorize entities, ultimately outputting a preliminary set of entities including tasks, resources, experimental data, milestones, and compliance check nodes, with each entity accompanied by basic attribute information.

[0088] S13. Based on the preset ontology mapping rule library for science and technology project domains, map the initial entity set to the ontology of science and technology project domains to obtain ontology entity instances and mapping rule tables; the mapping rule tables include mapping confidence.

[0089] The rule base stores the corresponding rules for structured fields, ontology attribute text entities, and ontology categories. For example, the responsible person field in the task table points to the name of the ontology resource, and the regulatory review time in the attribute text points to the deadline attribute of the ontology compliance check node. Illustratively, for each entity in the initial entity set, the corresponding ontology concept is matched according to the rule base, and the mapping process is recorded to generate a mapping rule table, which includes the original entity, ontology concept, mapping rule identifier, and mapping confidence. The confidence of rule-driven mapping is set to 0-9 by default, and the confidence of entity extracted by model assistance is assigned according to the recognition accuracy.

[0090] S14. If the mapping confidence is lower than the preset confidence threshold, generate a manual annotation instruction; the manual annotation instruction is used to indicate the sending of a manual standard confirmation instruction and to obtain feedback on the manual entity annotation information.

[0091] The mapping rule table is traversed. If the confidence score of an entity mapping is lower than the threshold, it is automatically marked as pending confirmation. For mappings pending confirmation, manual annotation instructions containing the original entity content, candidate ontology concepts, confidence scores, and data source fragments are generated and pushed to the project management terminal through the system interface, instructing the annotators to confirm the entity category and attribute mapping relationship.

[0092] S15. Update the mapping rule table and ontology entity instances based on the manually labeled entity information, and integrate the updated mapping rule table and ontology entity instances to obtain the project ontology.

[0093] The manually annotated mapping relationships are updated in the mapping rule table, and the corresponding confidence levels are adjusted. Simultaneously, the categories and attributes of ontology entity instances are corrected; for example, entities originally classified as milestone technical review entities are reclassified as compliance check node entities, and attributes such as review standards and required documents are added. Optionally, the updated ontology entity instances and mapping rule tables are summarized and integrated in RDF / OWL or a custom JSON-LD format.

[0094] In one embodiment, dynamic state time-series data of entities are embedded into the project ontology, and a semantic-temporal coupled knowledge graph is constructed, including:

[0095] S21. Map entities to project entity nodes and map static entity relationships to entity relationship edges to obtain the initial graph structure; project entity nodes include entity attributes.

[0096] In a schematic representation, based on entity categories such as tasks, resources, experimental data, milestones, and compliance check nodes in the project ontology, each entity is mapped to a project entity node in the knowledge graph, and each node must fully carry the attribute information of the corresponding entity. Furthermore, the static relationships between entities in the project ontology are mapped to entity relationship edges in the initial graph structure. Optionally, to clearly distinguish relationship types and strengths, each entity relationship edge needs to be labeled with a semantic category and weight value. The weight value can be set according to the tightness of the relationship between entities, such as the frequency of resource allocation to task execution in historical projects. Through the mapping and construction of entity nodes and entity relationship edges, an initial graph structure containing only semantic relationships and without embedded temporal information is finally formed.

[0097] S22. Bind the entity attributes and entity relationship edges in the initial graph structure in a time sequence to obtain a time sequence graph structure.

[0098] This example illustrates the extraction of time-series data corresponding to the dynamic attributes of project entity nodes. Optionally, for high-frequency time-series data such as equipment monitoring, to avoid node data redundancy, the complete time-series sequence is not stored in the project entity nodes. Instead, a connection between the node and the time-series database is established through time-series identifiers. During queries, the corresponding time-series data for the specified period can be retrieved in real time using the node identifier, TS-identifier, and time window. For low-frequency time-series data such as task status changes, time-status key-value pairs are directly stored as dynamic node attributes to ensure the time traceability of entity status. For entity relationship edges, a time window attribute is added to reflect the dynamic nature of the relationship. For relationships with time-limited constraints, such as temporary resource allocation to tasks or task-stage dependencies on milestones, the effective and ineffective times are marked on the entity relationship edges. For dependencies between tasks, time-series constraint attributes are also added to convert them into computable constraints. Through the time-series binding of entity attributes and entity relationship edges, the initial graph structure is upgraded to a time-series graph structure containing a time dimension, enabling time-related operations such as querying the entity status at a specific point in time and determining the validity of relationships for a specific time period.

[0099] S23. Based on the entity behavior rules of science and technology projects, the behavior patterns of each project entity node in the time sequence graph structure are embedded to obtain a semantic-temporal coupled knowledge graph.

[0100] This section illustrates typical behavioral rules for various entity categories in a technology project, encompassing both entity state transition logic and triggering conditions and time constraints for state changes. For example, the behavioral rules for a task entity are defined as state transition paths: pending, in progress, paused, completed, and closed. The transition from pending to in progress requires confirmation from the responsible person and completion of prerequisite tasks. Furthermore, these behavioral rules are transformed into lightweight, state machine-style behavioral model templates, matched to corresponding project entity nodes in the time sequence diagram structure according to entity category, thus embedding the behavioral patterns. During embedding, a connection must be established between the behavioral model and the entity's time sequence data. For instance, the task entity's behavioral model can automatically read progress change records from the time sequence database for the past 14 days to verify whether the current state transition conforms to preset rules. Through the embedding of behavioral patterns, the time sequence diagram structure further integrates the three core elements of semantic association, temporal features, and behavioral logic, ultimately forming a semantic-temporal coupled knowledge graph.

[0101] In one embodiment, such as Figure 2 As shown, based on historical project data, a semantic-temporal coupled knowledge graph is used to constrain the causal exploration space, and causal priors are obtained by extracting the ontology causal structure of project entity nodes, including:

[0102] S201. Filter out entity pairs with causal influence based on the semantic-temporal coupled knowledge graph, and apply hard and soft constraints to the entity pairs to obtain causal candidate templates; hard constraints correspond to the mandatory causal direction specified in the field of science and technology projects; soft constraints correspond to assigning weights to the causal relationship strength of entity pairs.

[0103] In a schematic manner, by traversing the entity categories and relation edges of STKG and combining the logic of the technology project domain, combinations of entities without direct relationships are excluded, and potentially causal entity pairs such as resource-task-external dependency-internal execution are prioritized. For the selected entity pairs, further hard and soft constraints are applied. Specifically, hard constraints enforce the causal direction based on the objective logic of the technology project domain to avoid misjudgments of reverse causality. For example, if equipment failure occurs before task delay, the causal direction of resource equipment failure is forced to point to task delay. Soft constraints assign causal relationship strength weights based on the association frequency and time-series co-occurrence of the entity pair in STKG. For example, if the proportion of records where equipment failure and task delay occur simultaneously reaches 85%, the association strength weight of the entity pair is set to 0.85. Finally, a causal candidate template containing entity pairs, causal direction constraints, and association strength weights is formed.

[0104] S202. Extract the time-synchronized feature sequences corresponding to the causal candidate templates from the semantic-temporal coupled knowledge graph to obtain the feature pair dataset; the feature pair dataset includes causal features and result features.

[0105] From the semantic-temporal coupled knowledge graph, for each entity pair in the causal candidate template, the corresponding time-synchronized feature sequence is extracted to construct a feature pair dataset. During feature extraction, it is crucial to ensure accurate timestamp matching between causal and result features to avoid spurious associations caused by time misalignment. For example, for the entity pair "resource device availability rate points to task delay days," daily device availability rate is extracted from the STKG time-series database as the causal feature, and the daily task delay days for the same period are extracted as the result feature. Optionally, the extracted feature sequences are preprocessed, including outlier removal and time step adjustment based on the lag of causal influence, ultimately forming a feature pair dataset containing both causal and result features.

[0106] S203. Eliminate entity pairs in the causal candidate template whose cause and effect features do not conform to the causal direction by using the constraint-based causal direction, and retain entity pairs in the causal candidate template whose causal association strength weights meet the preset threshold conditions by using the scoring-based causal direction, thus obtaining the local causal structure.

[0107] A dual screening mechanism of constraint-based and scoring-based methods is employed to extract local causal structures from causal candidate templates. Specifically, constraint-based causal direction screening verifies the causal direction of cause and effect features in the feature pair dataset based on hard constraints, eliminating entity pairs that violate the hard constraints. For example, if the actual causal direction of an entity pair is opposite to the resource-directed task specified by the hard constraints, the entity pair is directly excluded.

[0108] Subsequently, a scoring-based causal direction screening method was used, combined with the association strength weights assigned by soft constraints, and a causal structure scoring algorithm was employed to calculate the causal association significance score of entity pairs. A preset threshold was set, and entity pairs with scores higher than the threshold and corresponding association strength weights were retained. Through two rounds of screening, a local causal structure containing clear causal directions and significant associations was obtained.

[0109] S204. Using the fine-tuned small sample corresponding to the local causal structure as the prior, calculate the causal effect of the local causal structure to obtain the causal prior.

[0110] In a schematic manner, a Bayesian structural learning approach is employed, using the local causal structure as the structural prior. The posterior confidence of the causal edges is updated using fine-tuned small sample data. For example, if a 10% decrease in equipment availability in the current project data leads to an average increase of 12 days in task delays, and this correlation is stronger than in historical data, the confidence score of the corresponding causal edge is increased from 0.85 to 0.92. Simultaneously, propensity score matching (PSM) or dual machine learning methods are used to quantify the causal effect value and confidence interval of each causal edge in the local causal structure. For instance, a 1% increase in resource equipment availability leads to an average decrease of 0.06 days in task delays, with a 95% confidence interval of [-0.07, -0.05]. Integrating the updated local causal structure, causal effect value, and confidence interval ultimately forms a standardized causal prior, providing causal logic support for subsequent digital twin simulations and risk prediction.

[0111] In one embodiment, such as Figure 3 As shown, a digital twin model of the project is constructed based on a semantic-temporal coupled knowledge graph, causal priors, and resource data. Multi-strategy temporal simulations are then performed based on this digital twin model to obtain the expected index distributions for each strategy, including:

[0112] S301. The semantic-temporal coupled knowledge graph, causal priors, and resource data are decomposed into entity elements, sorted out in relational logic, and extracted in behavioral rules to obtain the project digital twin model. The project digital twin model includes entity modules, relation modules, behavioral modules, resource constraint modules, and causal intervention modules that are interconnected through data links.

[0113] As an illustration, the entity module directly maps to the project entity nodes in STKG. That is, it extracts core entities such as tasks, resources, experimental data, milestones, and compliance check nodes from STKG, uses the static attributes of the entities as the basic parameters of the module, and uses the dynamic temporal attributes of the entities as the real-time state variables of the module, ensuring that the attributes and states of the entities in the module are updated synchronously with STKG.

[0114] The association logic module integrates the entity relationship edges of STKG with the causal chains of causal priors. By transforming static associations such as dependencies, assignments, and constraints in STKG into topological relationship rules within the module, and transforming causal edges and effect values ​​in causal priors into causal association rules, the module ensures that it can simulate semantic associations and causal chain reactions between entities.

[0115] The behavior module is constructed based on the entity behavior model of STKG and the effect rules of causal priors. That is, it extracts the lightweight behavior model of the entity from STKG, combines the causal conditions of behavior triggering in the causal priors, and transforms it into a computable behavior function.

[0116] The resource constraint module is the boundary condition of the digital twin. It integrates resource data with the time constraints of STKG, transforming the upper limit of man-hours, available equipment time, budget amount, etc. in the resource data into the module's resource supply and demand rules, and transforming the time constraints of entity relationship edges in STKG into time window rules, ensuring that the resource allocation and task progress simulated by the module do not exceed the resource and time boundaries of the real project.

[0117] The causal intervention module is specifically designed to implement intervention logic based on causal priors, that is, to transform the causal edge and effect values ​​in the causal priors into executable operations. The module can receive intervention strategies from external input and calculate the chain causal effects caused by the intervention measures.

[0118] S302. Based on project objectives and causal priors, counterfactual reasoning is used to derive multiple project management strategies; project management strategies include intervention targets, intervention methods, intervention time, and intervention parameter values.

[0119] In a demonstrative sense, counterfactual reasoning, based on the core causal chain revealed by causal priors, simulates the impact of changing a certain causal variable on project outcomes, and then transforms it into an actionable management strategy. Simultaneously, other strategies are generated around dimensions such as process optimization and risk management. For example, considering the causal relationship between reagent delivery delays and experimental data initiation delays in the causal priors, a risk management strategy for switching to a backup reagent supplier is deduced. The intervention parameter is set to a 7-day delivery cycle for the backup supplier, 3 days shorter than the original supplier. This forms a project management strategy set including baseline strategies, resource adjustments, risk management, and process optimization. Each strategy clearly defines four elements in a parameterized form to ensure it can be parsed and executed by the digital twin model.

[0120] S303. Based on the digital twin model of the project, simulate the project management strategies to obtain the project timeline evolution trajectory corresponding to each project management strategy; simulate the resource status update, behavior pattern triggering, causal effect impact and data link synchronization.

[0121] Based on the established digital twin model of the project, time-series simulations are executed sequentially for each project management strategy, step by step. A closed-loop process is generated, from resource status updates to behavioral pattern triggers, causal effects, and data link synchronization, generating the project's time-series evolution trajectory for each strategy. Specifically, during the simulation initialization phase, the entity states of all modules in the digital twin model are frozen to the current time point to ensure consistency in the simulation starting point for each strategy. For example, taking the strategy of increasing experimental personnel as an example, the supply and demand data of the resource constraint module are updated, synchronizing the work hours of the two newly added experimental personnel to the system to correct the work hour gap in the experimental data; the state transition of the experimental data entity in the behavior module is triggered, and due to sufficient work hours, the experimental data changes from pending to in progress, with dynamic status updated at 20% progress per day; the causal intervention module calculates the delay probability of the target milestone from the initial 30% to 24% based on the effect value of the experimental data progress pointing to the milestone delay probability in the causal prior; and the state changes of all modules are synchronized in real time through the data link, providing the latest foundation for the simulation of the next time step.

[0122] S304. Extract and statistically analyze the core indicators of the time-series evolution trajectory of each project to obtain the expected indicator distribution.

[0123] By extracting core indicators of concern for science and technology project management from the temporal evolution trajectory of various project management strategies, and quantifying the characteristics of these indicators through statistical analysis, a multi-strategy expected indicator distribution is ultimately formed. For example, in the core indicator extraction stage, differentiated statistical methods are used for different types of indicators: for schedule indicators, the expected value of the peak delay probability and the number of delayed days at the end of the simulation in the time-series trajectory are calculated; for cost indicators, the percentage of the difference between total expenditure and budget and the percentage of overspending days are calculated; for research output indicators and compliance indicators, the percentage of the final completed quantity / time passed is directly extracted from the trajectory.

[0124] In one embodiment, causal enhancement risk prediction is performed based on a semantic-temporal coupled knowledge graph, causal priors, and the expected indicator distribution of each strategy, resulting in a risk prediction report and a causal explanation report, including:

[0125] S31. Extract features from the semantic-temporal coupled knowledge graph, causal priors, distributions of expected indicators, and text records respectively, and standardize the extracted features to obtain a prediction feature matrix; extract dynamic trend features and cross-entity impact features corresponding to the semantic-temporal coupled knowledge graph; extract causal impact strength features corresponding to the causal priors; extract potential effect features corresponding to the expected indicator distributions; and extract risk semantic features corresponding to the text records.

[0126] Indicatively, project dynamic trend features are calculated using a sliding window to measure the statistics of entity time-series data, quantifying real-time dynamic changes in the project. Cross-entity impact features are calculated based on STKG relation edges to determine the strength of associations, reflecting the potential impact of inter-entity associations on risk. Causal impact strength features are extracted based on causal priors, directly reading the causal side effect values ​​and confidence scores from the causal priors to clarify the degree of influence of key factors on risk. Potential effect features of strategies are extracted from the distribution of expected indicators for each strategy, calculating the deviation between the current actual state and the optimal strategy, and the mean and extreme values ​​of risk probabilities under all strategies, defining the risk boundaries under different strategies. For text records, a domain-fine-tuned Transformer model is used to extract risk semantic features, outputting the frequency of risk keywords and risk sentiment scores, transforming unstructured text into computable numerical features. All extracted features are standardized and organized according to the dimension of prediction time point - feature category, forming a prediction feature matrix where each row corresponds to one time granularity and each column corresponds to one feature, providing structured input for subsequent model predictions.

[0127] S32. Input the predicted feature matrix into the pre-trained causal interpretable hybrid prediction model to obtain the risk prediction result; the causal interpretable hybrid prediction model is composed of a temporal feature encoder, a text feature encoder and a causal correction layer connected in sequence.

[0128] The constructed prediction feature matrix is ​​input into a pre-trained causal interpretable hybrid prediction model. The temporal feature encoder uses TemporalGNN, which receives project dynamic trend features and cross-entity impact features from the prediction feature matrix. Leveraging the modeling advantages of GNN for graph-structured data, it captures the dynamic relationships between task-resource-experiment data in STKG, aggregates neighborhood features through a multi-layer graph attention mechanism, and outputs a temporal embedding vector with dimensions of 64 / 128, accurately depicting the project's temporal dynamic patterns. The text feature encoder uses a BERT model pre-trained in the project domain, focusing on processing risk semantic features. Through word segmentation, semantic encoding, and pooling operations, it transforms text features into text semantic embedding vectors with the same dimensions as the temporal embedding vectors, effectively mining the risk information hidden in the text. The causal correction layer integrates temporal embedding vectors, textual semantic embedding vectors, and causal influence strength features. It assigns higher weights to high-confidence causal features through a causal attention mechanism and introduces causal constraint loss to ensure that the prediction results are consistent with the causal logic. It outputs risk prediction results at different time granularities, including prediction targets, prediction probabilities, prediction values, and prediction confidence intervals. It also records the result update timestamp to ensure the timeliness and traceability of the prediction.

[0129] S33. Based on the risk prediction results, classify the risk types and analyze the risk time trend to obtain a risk prediction report.

[0130] As an illustration, the risk type classification stage categorizes the prediction results according to four dimensions: schedule risk, cost risk, compliance risk, and research output risk. The risk time trend analysis stage plots risk probability-time variation curves to visually demonstrate the evolution of risks over time, while also analyzing the scope of risk impact. Integrating the classification results with trend analysis generates a risk prediction report that includes risk level, quantitative indicators, trend charts, and scope of impact, providing project managers with clear risk warning information.

[0131] S34. Based on the risk prediction report, causal priors, and semantic-temporal coupled knowledge graph relationship edges, conduct local causal contribution analysis and counterfactual intervention suggestions to obtain a causal explanation report.

[0132] Indicatively, the local causal contribution analysis revolves around the relationship between features and risks. Combining the model's causal attention weights and causal prior effect values, it calculates and normalizes the contribution of each feature to the risk prediction results, outputting the Top 5 core influencing features and their contribution directions. Counterfactual intervention recommendations are generated based on causal prior effect values, selecting features with high and controllable causal contributions to simulate the risk improvement effect of intervention measures. Simultaneously, based on STKG relationship edges and causal prior causal graphs, the core causal path of risk transmission is drawn, using directed icons to annotate effect values ​​and current node states, supporting managers in tracing the source of risk. Integrating the above analysis results, a causal explanation report is generated, including a contribution analysis table, intervention recommendations, and path visualization diagrams. This report, in conjunction with the risk prediction report, provides comprehensive support for project decision-making, encompassing early warning, attribution, and response.

[0133] In one embodiment, based on compliance requirements, a project-related scheduling scheme is obtained through multi-objective optimization based on the expected indicator distribution, risk prediction report, and causal explanation report, including:

[0134] S41. Transform compliance requirements and resource data into project compliance rules and resource limits to obtain optimization goals and constraints.

[0135] This illustrative approach breaks down compliance requirements into quantifiable project compliance rules, serving as hard constraints. Violations of the scheduling plan will directly lead to project suspension, and these rules must be prioritized during the optimization process. Simultaneously, resource data is transformed into resource ceiling constraints, covering key dimensions such as manpower, equipment, and finance.

[0136] In setting optimization objectives, based on the needs of science and technology project management, objective functions are defined around four dimensions: risk control, cost optimization, schedule assurance, and resource efficiency. For example, the risk control objective is a milestone default probability ≤ 5% and task delay days ≤ 2 days; the cost optimization objective is a total budget overrun probability ≤ 10% and additional procurement expenditure ≤ 5%; the schedule assurance objective is to minimize the deviation between the total project duration and the planned duration; and the resource efficiency objective is a staff hour utilization rate ≥ 85% and equipment availability rate ≥ 90%. Weights are assigned to each objective using the analytic hierarchy process (AHP), forming a clear system of optimization objectives and constraints.

[0137] S42. Based on the optimization objective, reinforcement learning and mixed integer programming are performed according to the expected indicator distribution, risk prediction report and causal explanation report to obtain the scheduling strategy set.

[0138] A hierarchical optimization method combining reinforcement learning and mixed-integer programming is employed, integrating expected indicator distributions, risk prediction reports, and causal explanation reports to generate multiple feasible scheduling strategies, forming a scheduling strategy set. Specifically, in the global strategy generation phase of reinforcement learning, the project scheduling problem is modeled as a Markov Decision Process (MDP). The state space includes the current resource state, task progress, and risk indicators; the action space covers resource allocation adjustments, task priority adjustments, and time plan adjustments; the reward function is designed as a multi-objective comprehensive score, awarding points if the strategy meets the optimization objective and deducting points if it violates the optimization constraints. A deep reinforcement learning model is trained using historical scheduling data and simulation data from the expected indicator distribution. Through iterative learning, the model grasps the dynamic correlation between resources, tasks, and risks, quickly generating 2-3 sets of global scheduling strategies covering long-term resource allocation and task ranking.

[0139] Subsequently, mixed-integer programming was used to locally optimize the global strategy, focusing on resolving short-term resource conflicts and time window details. Simultaneously, counterfactual intervention suggestions from the causal explanation report were incorporated to fine-tune the resource allocation ratio. Finally, the global strategy and local optimization results were integrated, and after deduplication and filtering, 3-5 sets of differentiated scheduling strategies were retained to form a scheduling strategy set. Each strategy set includes a task time plan, a resource allocation table, and expected target values.

[0140] S43. Perform security verification on the scheduling strategy set according to the optimization constraints to obtain a security verification report; the security verification report includes pass and fail.

[0141] Based on optimization constraints, a formal verification method is used to perform security verification on each set of scheduling strategies, determining whether all constraints are met, and finally generating a security verification report containing pass / fail results. Illustratively, the scheduling strategy is transformed into a computable state transition system (STS), defining system states and transition rules using mathematical symbols. State variables include task progress, resource status, time, and risk indicators; transition rules correspond to resource allocation and task advancement logic in the strategy. Optimization constraints are transformed into temporal logic formulas, using linear temporal logic (LTL) or computational tree logic (CTL) to describe the constraints. A model verification tool automatically traverses all possible paths of the state transition system to verify whether the strategy satisfies the temporal logic formula on all paths. If the tool returns that all paths satisfy the formula, the strategy passes security verification, generating a verification pass record containing verification time, constraint satisfaction status, and path coverage; if a counterexample path is returned, it is marked as verification fail, and counterexample details are output for subsequent strategy modification reference. Finally, the verification results of all strategies are integrated to form a security verification report, clearly distinguishing between passing and failing scheduling strategies.

[0142] S44. Integrate the scheduling strategies that have passed the security verification report into a project-related scheduling scheme.

[0143] This approach illustratively demonstrates how validated scheduling strategies undergo multi-objective comprehensive scoring. Based on defined objective weights, scores for each strategy across risk control, cost optimization, schedule assurance, and resource efficiency are calculated and summed. The selected optimal strategy is then structurally integrated, breaking it down into executable plans with clear time granularity and assigned responsibilities, specifying corresponding adjustment measures. The integrated plans are then formatted and packaged into both human-readable and system-executable versions, outputting a complete project-related scheduling scheme that provides clear guidance for the efficient and compliant execution of technology projects.

[0144] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0145] Based on the same inventive concept, this application also provides a technology project intelligent management device for implementing the above-mentioned intelligent management method for technology projects. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the technology project intelligent management device provided below can be found in the limitations of the technology project intelligent management method described above, and will not be repeated here.

[0146] In one exemplary embodiment, such as Figure 4 As shown, a smart management device for science and technology projects is provided, comprising:

[0147] Data module 401 is used to acquire heterogeneous data throughout the entire lifecycle of science and technology projects, and to preprocess the heterogeneous data to obtain the project ontology; the heterogeneous data of science and technology projects includes project metadata, resource data, text records and compliance requirements;

[0148] The knowledge graph module 402 is used to embed dynamic state time-series data of entities into the project ontology and construct a semantic-temporal coupled knowledge graph; the dynamic state time-series data of entities includes static association relationships of entities that change over time; the semantic-temporal coupled knowledge graph includes project entity nodes and entity relationship edges;

[0149] Causal module 403 is used to constrain the causal exploration space based on historical project data using semantic-temporal coupled knowledge graphs, and to obtain causal priors by extracting ontology causal structure from project entity nodes.

[0150] The management simulation module 404 is used to construct a digital twin model of the project based on the semantic-temporal coupled knowledge graph, causal priors and resource data, and to perform multi-strategy temporal simulation based on the project digital twin model to obtain the expected index distribution of each strategy;

[0151] The prediction module 405 is used to perform causal enhanced risk prediction based on the semantic-temporal coupled knowledge graph, causal priors and the expected index distribution of each strategy, and to obtain a risk prediction report and a causal explanation report.

[0152] Management module 406 is used to obtain a project-related scheduling scheme by performing multi-objective optimization based on compliance requirements, expected indicator distribution, risk prediction report and causal explanation report.

[0153] In one embodiment, an entity extraction module is also included, for:

[0154] Standardized cleansing is performed on heterogeneous data throughout the entire lifecycle of science and technology projects to obtain standardized science and technology project data.

[0155] Standardized science and technology project data are processed by a rule engine based on data type to extract entities and classify them into categories, resulting in a preliminary entity set. Data types include structured data, semi-structured data, and unstructured data.

[0156] Based on a pre-defined ontology mapping rule library for science and technology projects, the initial set of entities is mapped to the ontology of science and technology projects, resulting in ontology entity instances and a mapping rule table; the mapping rule table includes mapping confidence.

[0157] If the mapping confidence level is lower than the preset confidence threshold, a manual annotation instruction is generated; the manual annotation instruction is used to indicate the sending of a manual standard confirmation instruction and to obtain feedback on the manual entity annotation information.

[0158] The mapping rule table and ontology entity instances are updated based on the manually labeled entity information, and the updated mapping rule table and ontology entity instances are integrated to obtain the project ontology.

[0159] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0160] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0161] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0162] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent management of science and technology projects, characterized in that, The method comprises: acquiring heterogeneous data of a science and technology project full life cycle, and preprocessing the heterogeneous data of the science and technology project full life cycle to obtain a project ontology; the science and technology project heterogeneous data comprises project metadata, resource data, text records and compliance requirements; the project ontology comprises entities, entity attributes and entity static association relationships; embedding entity dynamic state time series data into the project ontology, and constructing a semantic-time coupling knowledge graph; the entity dynamic state time series data comprises the entity static association relationships changing over time; the semantic-time coupling knowledge graph comprises project entity nodes and entity relationship edges; based on historical project data, constraining a causal exploration space by using the semantic-time coupling knowledge graph, and obtaining causal priors by extracting an ontology causal structure of the project entity nodes; constructing a project digital twin model according to the semantic-time coupling knowledge graph, the causal priors and the resource data, and performing multi-strategy time series simulation based on the project digital twin model to obtain expected index distributions of each strategy; performing causal enhanced risk prediction according to the semantic-time coupling knowledge graph, the causal priors and the expected index distributions of each strategy to obtain a risk prediction report and a causal explanation report; based on the compliance requirements, performing multi-objective optimization according to the expected index distributions, the risk prediction report and the causal explanation report to obtain a project correlation scheduling scheme.

2. The method of claim 1, wherein, The preprocessing of the science and technology project full life cycle heterogeneous data to obtain a project ontology comprises: standardizing and cleaning the science and technology project full life cycle heterogeneous data to obtain standardized science and technology project data; performing entity extraction and category division on the standardized science and technology project data according to a rule engine corresponding to a data type to obtain a preliminary entity set; the data type comprises structured data, semi-structured data and unstructured data; mapping the preliminary entity set to a science and technology project domain ontology based on a preset science and technology project domain ontology mapping rule library to obtain ontology entity instances and a mapping rule table; the mapping rule table comprises a mapping confidence; if the mapping confidence is lower than a preset confidence threshold, generating an artificial labeling instruction; the artificial labeling instruction is used to instruct sending an artificial standard confirmation instruction and obtaining artificial entity labeling information feedback; updating the mapping rule table and the ontology entity instances according to the artificial entity labeling information, and integrating the updated mapping rule table and the ontology entity instances to obtain the project ontology.

3. The method of claim 2, wherein, The embedding of entity dynamic state time series data into the project ontology and the construction of a semantic-time coupling knowledge graph comprise: mapping the entities to the project entity nodes and mapping the entity static association relationships to the entity relationship edges to obtain an initial graph structure; the project entity nodes comprise entity attributes; performing time series binding on the entity attributes in the initial graph structure and the entity relationship edges to obtain a time series graph structure; based on entity behavior rules of a science and technology project, performing behavior mode embedding on each project entity node in the time series graph structure to obtain the semantic-time coupling knowledge graph.

4. The method of claim 3, wherein, The historical project data is used to constrain a causal exploration space by using the semantic-temporal coupling knowledge graph, and a causal prior is obtained by performing ontology causal structure extraction on the project entity nodes, including: According to the semantic-temporal coupling knowledge graph, entity pairs with causal effects are screened out, and hard constraints and soft constraints are processed on the entity pairs to obtain causal candidate templates; the hard constraints correspond to mandatory causal directions in the scientific and technological project field; and the soft constraints correspond to weights assigned to the causal correlation strength of the entity pairs; Time synchronization feature sequences corresponding to the causal candidate templates are extracted from the semantic-temporal coupling knowledge graph to obtain a feature pair dataset; the feature pair dataset includes cause features and result features; Through constraint causal direction, entity pairs in which the cause features and the result features in the causal candidate templates do not conform to the causal direction are removed, and through scoring causal direction, entity pairs in which the causal correlation strength weight in the causal candidate templates conforms to a preset threshold condition are retained to obtain a local causal structure. The causal effect of the local causal structure is calculated by taking the fine-tuning small sample corresponding to the local causal structure as a prior to obtain the causal prior.

5. The method of claim 1, wherein, The project digital twin model is constructed according to the semantic-temporal coupling knowledge graph, the causal prior and the resource data, and the expected index distribution of each strategy is obtained through multi-strategy temporal simulation based on the project digital twin model, including: The semantic-temporal coupling knowledge graph, the causal prior and the resource data are subjected to entity element disassembly, relationship logic carding and behavior rule extraction to obtain a project digital twin model; the project digital twin model includes entity modules, relationship modules, behavior modules, resource constraint modules and causal intervention modules that are connected to each other through data links; A plurality of project management strategies are obtained through counterfactual reasoning according to a project target and the causal prior; the project management strategies include intervention objects, intervention methods, intervention times and intervention parameter values; The project management strategies are simulated based on the project digital twin model to obtain project temporal evolution trajectories corresponding to the project management strategies; the simulation corresponds to resource state updating, behavior mode triggering, causal effect influence and data link synchronization; Core indexes are extracted and counted from the project temporal evolution trajectories to obtain the expected index distribution.

6. The method of claim 1, wherein, The risk prediction report and the causal explanation report are obtained through causal enhancement risk prediction according to the semantic-temporal coupling knowledge graph, the causal prior and the expected index distribution of each strategy, including: Features are extracted from the semantic-temporal coupling knowledge graph, the causal prior, each expected index distribution and the text record respectively, and the extracted features are standardized to obtain a prediction feature matrix; the semantic-temporal coupling knowledge graph corresponds to project dynamic trend feature extraction and cross-entity influence feature extraction; the causal prior corresponds to causal influence strength feature extraction; the expected index distribution corresponds to strategy potential effect feature extraction; and the text record corresponds to risk semantic feature extraction. inputting the predicted feature matrix into a pre-trained causally interpretable hybrid prediction model to obtain a risk prediction result; the causally interpretable hybrid prediction model is composed of a time sequence feature encoder, a text feature encoder, and a causal correction layer connected in sequence; performing risk type classification and risk time trend analysis according to the risk prediction result to obtain the risk prediction report; performing local causal contribution analysis and counterfactual intervention suggestion according to the risk prediction report, the causal prior, and the semantic-time coupling knowledge graph relationship edge to obtain a causal explanation report.

7. The method of claim 6, wherein, The multi-objective optimization is performed according to the expected index distribution, the risk prediction report, and the causal explanation report based on the compliance requirements to obtain a project correlation scheduling scheme, including: The compliance requirements and the resource data are converted into project compliance rules and resource upper limits to obtain optimization objectives and optimization constraints; Based on the optimization objectives, reinforcement learning and mixed integer programming are performed according to the expected index distribution, the risk prediction report, and the causal explanation report to obtain a scheduling strategy set; According to the optimization constraints, the scheduling strategy set is subjected to safety verification to obtain a safety verification report; the safety verification report includes pass and fail; The scheduling strategy of the safety verification report is integrated into the project correlation scheduling scheme.

8. A scientific project intelligent management device, characterized in that, The device includes: A data module configured to acquire heterogeneous data of a whole life cycle of a science and technology project, and to pre-process the heterogeneous data of the whole life cycle of the science and technology project to obtain a project ontology; the heterogeneous data of the science and technology project includes project metadata, resource data, text records, and compliance requirements. A knowledge graph module configured to embed entity dynamic state time sequence data into the project ontology, and to construct a semantic-time coupling knowledge graph; the entity dynamic state time sequence data includes time-varying entity static correlation relationships; the semantic-time coupling knowledge graph includes project entity nodes and entity relationship edges. A causal module configured to constrain a causal exploration space by using the semantic-time coupling knowledge graph based on historical project data, and to obtain a causal prior by extracting a causal structure of the project entity nodes. A management simulation module configured to construct a project digital twin model based on the semantic-time coupling knowledge graph, the causal prior, and the resource data, and to perform multi-strategy time sequence simulation based on the project digital twin model to obtain an expected index distribution of each strategy. A prediction module configured to perform causally enhanced risk prediction based on the semantic-time coupling knowledge graph, the causal prior, and the expected index distribution of each strategy to obtain a risk prediction report and a causal explanation report. A management module configured to perform multi-objective optimization based on the compliance requirements to obtain a project correlation scheduling scheme according to the expected index distribution, the risk prediction report, and the causal explanation report. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1 to 7.

Images