Project cost progress management system based on artificial intelligence

Abstract

The invention discloses a project cost progress management system based on artificial intelligence, and relates to the technical field of artificial intelligence, and the system comprises a construction module which is used for constructing a dynamic causal network based on each work package in a project list and a multi-source influence factor extracted from multi-source data, and a prediction module which is used for predicting the project cost progress based on the dynamic causal network. And the optimization module is used for identifying parallel work packets in a project execution process, forming an experimental group and a control group, carrying out iterative optimization on the dynamic causal network according to a comparison result of the experimental group and the control group, and outputting a prediction result. According to the method, the dynamic causal network is constructed based on the influence factors extracted from the work packages and the multi-source data in the project list, the influence of risks on the total cost and the total construction period of the project is quantified, the micro-intervention strategy is generated, the experimental group and the control group are set, the strategy effect is quantified through differential analysis, the management precision and efficiency are continuously improved, and the risk is reduced. And full-period management and control of the project are realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based engineering cost and schedule management system. Background Technology

[0002] In recent years, with the continuous expansion of the construction industry and the significant increase in project complexity, project cost and schedule management have faced numerous challenges. The number of work packages involved in project implementation has increased, and the project is affected by a variety of factors. Traditional management methods are unable to fully integrate multi-source data, making it more difficult to control the overall progress of the project. At the same time, external factors such as market fluctuations, policy adjustments, and changes in natural conditions, as well as internal factors such as human resource allocation and material supply, all have a direct impact on project cost and schedule. More efficient and intelligent management methods are urgently needed to cope with these dynamic changes, ensure that the project proceeds as planned, and control costs within a reasonable range.

[0003] In existing technologies, traditional engineering cost and schedule management techniques mostly rely on static data and experience-based judgments, which cannot effectively integrate structured, semi-structured, and unstructured multi-source influencing factors. This results in incomplete coverage of project influencing factors, affecting the scientific nature of management decisions. The ability to predict project risks is limited, usually only able to deal with known risks, lacking the ability to quantitatively analyze and extrapolate hypothetical risks, unable to assess the impact of potential risks on the total project cost and duration in advance, and unable to provide forward-looking support for risk prevention and control. Furthermore, the lack of a dynamic iterative optimization mechanism means that during project execution, the relationships and influence weights between various elements cannot be adjusted in a timely manner based on real-time project data and intervention effects, making it difficult to continuously improve management accuracy and efficiency. Summary of the Invention

[0004] The technical problem addressed by this invention is that traditional engineering cost and schedule management techniques rely heavily on static data and experience-based judgments, failing to effectively integrate structured, semi-structured, and unstructured multi-source influencing factors. This results in incomplete coverage of project influencing factors, affecting the scientific nature of management decisions. Furthermore, the ability to predict project risks is limited, typically only addressing known risks. There is a lack of quantitative analysis and scenario simulation capabilities for hypothetical risks, making it impossible to assess the impact of potential risks on total project cost and duration in advance. This hinders the provision of forward-looking support for risk prevention and control, and the lack of a dynamic iterative optimization mechanism. During project execution, the relationships and influence weights between various elements cannot be adjusted promptly based on real-time project data and intervention effects, making it difficult to continuously improve management accuracy and efficiency.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: an artificial intelligence-based engineering cost and schedule management system, comprising a construction module, a prediction module, and an optimization module: The construction module is used to build a dynamic causal network based on the various work packages in the project list and the multi-source influencing factors extracted from multi-source data. The prediction module is used to quantify and predict hypothetical risks based on the dynamic causal network through scenario deduction, and output the prediction results. The optimization module is used to identify parallel work packages during project execution and form experimental and control groups. The dynamic causal network is iteratively optimized based on the comparison results between the experimental and control groups.

[0006] As a preferred embodiment of the artificial intelligence-based engineering cost and schedule management system of the present invention, the construction module includes a mapping unit, a forming unit, and an assignment unit; The mapping unit is used to map each work package in the project list and the multi-source influencing factors extracted from multi-source data into nodes in a dynamic causal network. A work package represents a project task in the project list, which includes the specific steps of the task, deliverables, and execution requirements. The multi-source influencing factors include structured data influencing factors, semi-structured data influencing factors, and unstructured data influencing factors; The structured data influencing factors include historical project cost databases and material price indices; The factors influencing the semi-structured data include project schedule documents and supplier contracts; The factors influencing unstructured data include weather data, social media information, policy and regulatory texts, and satellite remote sensing images; The forming unit is used to acquire historical project data, which includes multidimensional variables related to project cost and schedule. The historical project data is used as an observation sample set, and a causal discovery algorithm is used to perform statistical analysis on the observation sample set. Statistical analysis includes: testing the conditional independence between variables, eliminating spurious associations caused by common causes and accidental factors, determining the true causal direction and directed connection between variables at different nodes, solidifying the directed connection, and forming the topology of the initial causal network; The assignment unit is used to generate dynamic weights for each directed edge in the topology of the initial causal network to obtain the dynamic causal network. The dynamic weights are used to quantify the influence of cause nodes on result nodes and are continuously updated based on newly added project data. The logic for generating dynamic weights includes: Based on the node type and the characteristics of the processed project data, a matching calculation method is selected from the preset weight calculation method library. The new weight is calculated according to the calculation method to update the weight. The new weight is compared with the current dynamic weight. When the difference between the new weight and the current dynamic weight exceeds a preset difference value, the weight is updated.

[0007] As a preferred embodiment of the AI-based engineering cost and schedule management system of the present invention, wherein: the logic of respectively mapping to nodes in the dynamic causal network includes: Traverse each work package in the project list and map the work package to an execution node in the dynamic causal network. At the same time, map the multi-source influencing factors to different types of entity nodes in the dynamic causal network according to the data type and business semantics. The data types include structured data, semi-structured data, and unstructured data; The different types include resource nodes, environment nodes, risk nodes, and management nodes.

[0008] As a preferred embodiment of the artificial intelligence-based engineering cost and schedule management system of the present invention, the prediction module includes a scenario deduction unit; The scenario simulation unit is used to receive hypothetical risk conditions input by the user, and to perform counterfactual inference on the dynamic causal network through a counterfactual inference algorithm. It simulates the state changes of each node in the dynamic causal network when the hypothetical risk conditions are real, and quantitatively calculates the impact of each node's state change data to output the prediction results of the project's key indicators under hypothetical risk conditions. The hypothetical risk conditions include resource shock conditions, environmental mutation conditions, risk manifestation conditions, and management intervention conditions. The key performance indicators for the project include total project cost and total project duration.

[0009] As a preferred embodiment of the artificial intelligence-based engineering cost and progress management system of the present invention, the optimization module includes an identification unit, a generation unit, a setting unit, a quantification unit, and an update unit. The identification unit is used to scan the project work breakdown structure in real time and identify parallel work packages during project execution. The generation unit is used to generate micro-intervention strategies for the identified parallel work packages; The setting unit is used to select one of the parallel work packages as the experimental group, implement a micro-intervention strategy on the experimental group, and use the remaining parallel work packages as the control group. The quantification unit is used to collect key data from the experimental group and the control group respectively, and to quantify the causal effect of the micro-intervention strategy on the performance of the work package through differential analysis to obtain the causal effect value. The key data includes schedule data, cost data, and quality data; The update unit is used to update the connection relationships of corresponding nodes in the dynamic causal network.

[0010] As a preferred embodiment of the AI-based engineering cost and schedule management system of the present invention, the logic for identifying parallel work packages includes: Collect relevant information about the project and use natural language processing technology to extract text clues describing parallel relationships from the relevant information, and use the text clues as a candidate set of parallel relationships; Obtain the text evidence strength of each pair of parallel relationships in the candidate set of parallel relationships. When the text evidence strength is greater than a preset effective threshold, confirm that the current parallel relationship is a valid parallel work package. The relevant information includes project planning documents, construction logs, and meeting minutes; The parallel relationships include simultaneous, synchronous, and overlapping operations.

[0011] As a preferred embodiment of the AI-based engineering cost and schedule management system of the present invention, the logic for generating the micro-intervention strategy includes: The basic attribute features of the current work package are extracted from the project list, and the correlation and influence features are extracted from multi-source data, including structured data, semi-structured data and unstructured data. The basic attribute features and correlation and influence features are quantized and encoded using natural language processing technology to obtain the feature vector of the current work package. The basic attribute characteristics include the task's professional category, planned duration, planned cost, and deliverable standards. The associated impact characteristics include the associated impact characteristics of structured data, the associated impact characteristics of semi-structured data, and the associated impact characteristics of unstructured data; Based on the feature vector of the current work package, retrieve and match the most similar successful intervention cases from the historical project database to construct an initial candidate strategy pool; Receive the target work package that needs intervention, extract the feature vector of the target work package, calculate the similarity between the feature vector and the feature vector of all work packages in the historical project database, select the top N historical work packages with the highest similarity scores from the historical project database, and extract the intervention measures corresponding to the top N historical work packages to form a candidate intervention strategy set. The system acquires the current resource status of the project in real time and converts the resource status into a numerical state vector. Based on the risk preference level set when the project starts, the system loads the corresponding risk preference parameters and encodes the numerical state vector and risk preference parameters into an environmental state vector, which is then used as the input to the reinforcement learning model. The risk preference levels are categorized as conservative, balanced, and aggressive. Load a pre-trained reinforcement learning model. The core of this model is a Q-value function, which evaluates the optimal long-term reward obtained by executing each policy from the candidate intervention policy set given the environment state vector. The expression for the Q-value function is as follows: ; in, Represents the environment state vector. Indicates candidate intervention strategies, Indicates immediate feedback. Indicates the discount factor; in, Indicates the implementation of candidate intervention strategies The state vector of the new environment after entering; express All candidate intervention strategies; Indicates the optimal long-term return in the future; The environmental state vector and the set of candidate intervention strategies are input into the reinforcement learning model. The Q-value of each candidate intervention strategy is calculated, and the candidate intervention strategy with the largest Q-value is selected as the output of the reinforcement learning model to generate a micro-intervention strategy.

[0012] As a preferred embodiment of the AI-based engineering cost and schedule management system of the present invention, the logic for converting the resource status into a numerical state vector includes: The resource status includes human resources status, equipment resources status, financial resources status, time resources status, and material resources status; Identify specific indicators for each resource status, determine the encoding strategy based on the corresponding data type, map the resource status to numbers using the encoding strategy, and normalize the numbers to obtain a numerical status vector. The data types include continuous, categorical, ordinal, and Boolean.

[0013] As a preferred embodiment of the AI-based engineering cost and schedule management system of the present invention, the logic for quantifying the causal effect of the micro-intervention strategy on the performance of the work package includes: At preset time points, key data of the experimental group and the control group are collected respectively, and the performance change of the experimental group before and after the implementation of the micro-intervention strategy is calculated and recorded as the performance change of the experimental group. At the same time, the performance change of the control group in the same time period is calculated and recorded as the performance change of the control group. The net effect value is obtained by calculating the difference between the performance change of the experimental group and the performance change of the control group, and the net effect value is used as the causal effect value.

[0014] As a preferred embodiment of the AI-based engineering cost and schedule management system of the present invention, the logic for updating the connection relationships of corresponding nodes in the dynamic causal network includes: Acquire causal effect records of the quantification unit, wherein the causal effect records include the characteristics of the micro-intervention strategy, the initial characteristics of the work package, and the causal effect values; The features of the micro-intervention strategy and the initial features of the work package are concatenated into input features, which are then input into the dynamic causal network in the construction module. Forward propagation calculation is performed on the dynamic causal network, and the predicted value of the causal effect is output. The difference between the predicted value and the causal effect value is calculated using a preset loss function. Based on the difference, backpropagation calculation is performed on the dynamic causal network. Starting from the output layer of the dynamic causal network, the partial derivatives of the loss function with respect to each weight parameter of the dynamic causal network are calculated layer by layer, and the partial derivatives are used as the gradient. Based on the gradient and a preset learning rate, the update amount of each weight parameter is calculated, and the optimizer iteratively updates the weight parameters in the dynamic causal network according to the update amount.

[0015] The beneficial effects of this invention are as follows: This invention constructs a dynamic causal network based on work packages in the project list and influencing factors extracted from multi-source data, achieving comprehensive integration of multi-source factors and providing support for management decisions to enhance their scientific rigor. Based on this network, through scenario simulation and counterfactual inference, it simulates the changes in node states when hypothetical risks such as resource shocks and environmental mutations occur, quantifying the impact of risks on the total project cost and duration, providing forward-looking support for risk prevention and control. Simultaneously, it can identify parallel work packages during project execution, generate micro-intervention strategies, and set up experimental and control groups. Through differential analysis, it quantifies the effect of the strategies and updates the parameters of the dynamic causal network to adapt to real-time data and intervention effects, continuously improving management accuracy and efficiency, and achieving full-cycle project control. Attached Figure Description

[0016] Figure 1 This is a basic flowchart of an artificial intelligence-based engineering cost and schedule management system provided as an embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example, refer to Figure 1 As one embodiment of the present invention, an artificial intelligence-based engineering cost and schedule management system is provided, including a construction module, a prediction module, and an optimization module: The building module is used to construct dynamic causal networks based on the various work packages in the project list and the multi-source influencing factors extracted from multi-source data.

[0019] The prediction module is used to quantify hypothetical risks based on dynamic causal networks and through scenario extrapolation, and output the prediction results.

[0020] The optimization module is used to identify parallel work packages during project execution and form experimental and control groups. The dynamic causal network is iteratively optimized based on the comparison results between the experimental and control groups.

[0021] This system effectively enhances the intelligence and precision of project cost and schedule management. The construction module integrates work packages from the project list with influencing factors extracted from multi-source data to build a dynamic causal network, providing a comprehensive and project-realistic analytical foundation for subsequent management. This solves the problems of scattered data and unclear correlations in traditional management. The prediction module relies on the dynamic causal network to conduct scenario simulations, enabling quantitative predictions of hypothetical risks and outputting results. This allows project teams to understand the potential impact of risks on cost and schedule in advance, providing a forward-looking basis for risk prevention and response decisions. The optimization module identifies parallel work packages during project execution and sets up experimental and control groups. By comparing the results, iteratively optimizing the dynamic causal network, the network can continuously adapt to the real-time progress of the project, further improving the accuracy and practicality of the management model and helping to achieve efficient cost and schedule control throughout the entire project lifecycle.

[0022] The building blocks include mapping units, forming units, and assignment units.

[0023] The mapping unit is used to map each work package in the project list and the multi-source influencing factors extracted from multi-source data to nodes in a dynamic causal network.

[0024] A work package represents a project task in the project list. A project task includes the specific steps of the task, deliverables, and execution requirements.

[0025] Multi-source influencing factors include structured data influencing factors, semi-structured data influencing factors, and unstructured data influencing factors.

[0026] Factors influencing structured data include historical project cost databases and material price indices.

[0027] Factors affecting semi-structured data include project schedule documents and supplier contracts.

[0028] Factors influencing unstructured data include weather data, social media information, policy and regulatory texts, and satellite remote sensing images.

[0029] The forming unit is used to acquire historical project data, which includes multidimensional variables related to project cost and schedule. The historical project data is used as an observation sample set, and the causal discovery algorithm is used to perform statistical analysis on the observation sample set.

[0030] Statistical analysis includes: testing the conditional independence between variables, eliminating spurious associations caused by common causes and accidental factors, determining the true causal direction and directed connections between variables at different nodes, solidifying the directed connections, and forming the topology of the initial causal network.

[0031] To lay the foundation for the network structure, we extract genuine correlations between node variables from historical project data. Using historical project data covering multiple dimensions related to project cost and schedule as the observation sample set, we conduct a systematic analysis of the correlations between variables. First, we test the conditional independence between variables, systematically identifying and eliminating spurious correlations caused by common causes (such as an external factor simultaneously affecting two seemingly related variables) or accidental factors (such as data collection errors or short-term random fluctuations). This avoids such non-genuine correlations interfering with the accuracy of the network structure. Based on this, we further clarify and determine the true causal direction between different node variables, that is, clarify which variable is the cause and which is the effect. At the same time, we lock in the effective directed connections between variables to ensure that the correlations between nodes in the network conform to the causal logic of actual project operation. Finally, we solidify these verified directed connections to form the topology of the initial causal network, providing a clear and reliable structural framework for subsequent dynamic weight assignment and network iterative optimization, ensuring the accuracy of subsequent risk prediction and management optimization.

[0032] The assignment unit is used to generate dynamic weights for each directed edge in the topology of the initial causal network, thereby obtaining the dynamic causal network.

[0033] Dynamic weights are used to quantify the influence of cause nodes on result nodes and are continuously updated based on newly added project data.

[0034] The logic for generating dynamic weights includes: Based on the node type and the characteristics of the processed project data, a matching calculation method is selected from the preset weight calculation method library. The new weight is calculated according to the calculation method to update the weight. The new weight is compared with the current dynamic weight. When the difference between the new weight and the current dynamic weight exceeds the preset difference value, the weight is updated.

[0035] The assignment unit first analyzes the type of each node in the dynamic causal network (such as execution nodes, resource nodes, environment nodes, etc.), and clarifies the attribute characteristics of the processed project data (such as the degree of data structuring, volatility, sample size, etc.). Based on these analysis results, it selects suitable calculation methods from the preset weight calculation method library. For example, if the node is a structured type related to historical cost data and the data shows linear correlation characteristics, a linear regression method may be selected. If the node involves influencing factors derived from unstructured data and the data correlation is more complex, a machine learning weight calculation method may be selected.

[0036] After processing the newly added project data using the selected calculation method, new weights for updating are obtained. Then, the new weights are compared with the dynamic weights of the corresponding directed edges in the current dynamic causal network to determine the magnitude of the difference. When the difference exceeds the preset difference threshold, it means that the node influence relationship reflected by the new data has changed significantly. At this time, the system will trigger the weight update mechanism to replace the current dynamic weights with the new weights, ensuring that the weights of the dynamic causal network are always consistent with the latest actual situation of the project, and providing an accurate basis for the intensity of influence for subsequent risk prediction and intervention optimization.

[0037] Among the multi-source influencing factors, the scope of unstructured data influencing factors is set according to their direct correlation with project cost and progress. Social media information is limited to content related to the project supply chain, industry policies and surrounding environment, excluding irrelevant hot topics. Satellite remote sensing images only include the geographical environment and progress images of the construction area to avoid node redundancy and network efficiency caused by irrelevant regional data.

[0038] The causal discovery algorithm is selected based on the scale and dimensionality of historical project data. When the data sample size is large and the variable associations are complex, the Bayesian network algorithm is used first to mine causal associations. When the data dimension is small but the time series characteristics are obvious, the time series causal discovery algorithm is adapted to ensure the accuracy of variable association analysis. The adaptation logic is clearly defined according to the data characteristics.

[0039] A pre-defined weight calculation method library is constructed according to node type and data characteristics. Resource nodes are associated with continuous data and are assigned to linear regression, while environmental nodes are associated with categorical data and are assigned to logistic regression.

[0040] The preset difference value is adjusted according to the project type and management precision. For large-scale infrastructure projects that are sensitive to cost and schedule, the difference value is set to a small range (such as 5%). For small and sporadic projects, the difference value is relaxed to about 10% to balance precision and efficiency. There is no unified standard.

[0041] The construction module, through the collaborative operation of mapping, forming, and assignment units, effectively solves the problems of difficult data integration and inaccurate correlation analysis in traditional engineering cost management. The mapping unit accurately maps each work package containing specific task steps, deliverables, and execution requirements from the project list, along with influencing factors extracted from multi-source data, into execution nodes and entity nodes of a dynamic causal network. This achieves systematic integration of core project tasks and multi-dimensional influencing factors. The forming unit uses multi-dimensional variables related to cost and schedule from historical projects as the observation sample set. Using causal discovery algorithms, it eliminates false associations caused by common causes and accidental factors, clarifies the true causal direction and directed connection relationships of node variables, and solidifies them into an initial causal network topology that fits the actual logic of the project. The assignment unit generates dynamic weights with quantifiable influence strength for the directed edges of the initial topology. It updates the weights based on newly accessed data and matching calculation methods, triggering updates only when the difference between the new and current weights exceeds a preset value, balancing timeliness and resource utilization. These three components collaboratively construct a comprehensive, realistic, and dynamically adaptable causal network that adapts to project progress, laying a solid foundation for subsequent risk quantification prediction in the prediction module and network iterative optimization in the optimization module.

[0042] The logic for mapping these nodes to nodes in a dynamic causal network includes: Iterate through each work package in the project list, map the work package to the execution node in the dynamic causal network, and map the multi-source influencing factors to different types of entity nodes in the dynamic causal network according to the data type and business semantics.

[0043] Data types include structured data, semi-structured data, and unstructured data.

[0044] Different types include resource nodes, environment nodes, risk nodes, and management nodes.

[0045] Work packages are designed based on project scale and management precision requirements, and work package splitting standards are established. Large projects are split into sub-projects (such as civil engineering and installation), while small projects are split into core task phases (such as preliminary preparation and main construction). After splitting, the workload of each work package is verified to ensure that there is no redundancy of "a single node containing multiple independent tasks" or overly detailed issues of "a single task being split into multiple nodes". A unified splitting specification document is generated and synchronized to the mapping unit.

[0046] This approach effectively solves the problem of ambiguous relationships between project tasks and influencing factors in traditional management. It maps each work package as an execution node, clearly presenting the independent attributes and execution logic of each task in the project list. This provides a clear node carrier for subsequent tracking of the cost and progress status of individual tasks. At the same time, combining data types and business semantics, it classifies and maps multi-source influencing factors into four types of entity nodes: resources, environment, risk, and management. This makes the relationship between influencing factors with different attributes and project tasks clearer. It not only achieves the orderly integration of multi-source data, but also lays a structured foundation for subsequent dynamic causal network analysis of the impact paths of various factors on tasks, improving the pertinence and efficiency of network analysis.

[0047] The prediction module includes a scenario simulation unit.

[0048] The scenario simulation unit receives hypothetical risk conditions input by the user and performs counterfactual inference on the dynamic causal network using a counterfactual inference algorithm. It simulates the state changes of each node in the dynamic causal network under the hypothetical risk conditions as real conditions, and quantifies the impact of each node's state change data on the project's key indicators based on the data, outputting the prediction results of the project's key indicators under the hypothetical risk conditions.

[0049] Hypothetical risk conditions include resource shock conditions, environmental mutation conditions, risk manifestation conditions, and management intervention conditions.

[0050] Key performance indicators for a project include total project cost and total project duration.

[0051] The impact on key project indicators is quantitatively calculated. First, the core nodes directly related to the total project cost (related to material prices, labor costs, equipment rental, etc.) and total project duration (related to the execution of each work package, weather impact, resource supply, etc.) are identified. Then, the status change data of these core nodes after the hypothetical risk occurs are obtained through counterfactual inference algorithms and compared with the node baseline values ​​under the "risk-free normal state" to calculate the node status differences. Finally, based on the dynamic weight of each node to the total cost and total project duration (quantifying the impact intensity), the node differences are converted into specific changes in key indicators (such as the increase in total cost and the delay in total project duration), thereby completing the quantitative calculation of the impact.

[0052] It can accurately receive four types of hypothetical risk conditions input by users: resource shock, environmental mutation, risk manifestation, and management intervention. Through counterfactual inference algorithms, it can deduce the dynamic causal network and realistically simulate the state changes of each node in the network when the risk actually occurs. This avoids the problem of the disconnect between the "static assessment" in traditional risk analysis and the actual dynamic operation of the project. Based on the node state change data, the unit can quantitatively calculate the impact of the risk on the two key indicators of the project's total cost and total schedule and output the prediction results. This allows the project team to understand the cost overrun and schedule delay under different risk scenarios in advance, providing data support for the formulation of targeted risk response strategies (such as reserving backup resources in advance and adjusting the construction plan), and significantly improving the initiative and accuracy of project risk management.

[0053] The optimization module includes an identification unit, a generation unit, a setting unit, a quantization unit, and an update unit.

[0054] The identification unit is used to scan the project work breakdown structure in real time during project execution and identify parallel work packages.

[0055] The generation unit is used to generate micro-intervention strategies for the identified parallel work packages.

[0056] The configuration unit is used to select one work package from the parallel work packages as the experimental group, implement micro-intervention strategies on the experimental group, and use the remaining parallel work packages as the control group.

[0057] The quantification unit is used to collect key data from the experimental and control groups respectively. Through differential analysis, the causal effect of the micro-intervention strategy on the performance of the work package is quantified, and the causal effect value is obtained.

[0058] Key data includes schedule data, cost data, and quality data.

[0059] The update unit is used to update the connection relationships of corresponding nodes in a dynamic causal network.

[0060] The identification unit scans the project's work breakdown structure in real time, extracting textual clues about parallel relationships from project plans, construction logs, and other information using natural language processing technology. This accurately identifies parallel work packages, addressing the problem of overlooking parallel tasks in traditional management. The generation unit combines the basic attribute features and multi-source correlation influence features of parallel work packages, using a reinforcement learning model to generate micro-intervention strategies adapted to the project's resource status and risk preferences, ensuring the intervention's relevance. The setting unit provides a control variable environment for verifying the intervention effect by dividing the experimental and control groups. The quantification unit collects key data on progress, cost, and quality, using differential analysis to quantify the causal effect of the micro-intervention strategy and outputs an objective causal effect value. The update unit iteratively optimizes the node connections of the dynamic causal network based on this effect value, continuously improving the system's management accuracy for the project.

[0061] The logic for identifying parallel work packages includes: Collect relevant information about the project and use natural language processing technology to extract textual clues describing parallel relationships from the relevant information, and use the textual clues as a candidate set of parallel relationships.

[0062] Obtain the textual evidence strength of each pair of parallel relationships in the candidate set of parallel relationships. When the textual evidence strength is greater than the preset effective threshold, confirm that the current parallel relationship is a valid parallel work package.

[0063] The relevant information includes project planning documents, construction logs, and meeting minutes.

[0064] Parallel relationships include simultaneous, synchronous, and overlapping operations.

[0065] The calculation of textual evidence strength and threshold setting first quantifies textual clues based on the dimensions of expression clarity and semantic relevance. The validity of clues is distinguished by comparing the expression precision (e.g., the strength of fully synchronous execution is higher than that of partial cross-progression). The preset effective threshold is adjusted according to the differences in project types. For large and complex projects (such as subway construction), the threshold is set at a higher level (e.g., ≥85 points) to avoid parallel conflicts due to the close correlation of processes. For small and simple projects (such as street-front shop renovation), the threshold is lowered (e.g., ≥60 points) to take into account management efficiency. At the same time, the reverse reasoning logic is clearly defined. If this logic is not followed (e.g., uniform threshold, calculation ignoring semantic details), there will be problems of mis-screening of effective relationships or misjudging of invalid relationships. Ultimately, this ensures the accurate identification of parallel work packages.

[0066] For each pair of relations in the candidate set of parallel relations, the text evidence strength is calculated. Only relations with strength exceeding the preset effective threshold are identified as valid parallel work packages. This can eliminate interference from clues such as "vague expression" and "semantic ambiguity". For example, in the construction log, expressions such as "preparing pipeline materials during the main construction period" that do not explicitly indicate "synchronous execution" will be filtered out due to insufficient evidence strength, while explicit expressions such as "the main structure construction and pipeline pre-embedding will be carried out simultaneously" will be accurately identified, which greatly improves the accuracy and reliability of the parallel work package identification results.

[0067] The logic for generating micro-intervention strategies includes: The basic attribute features of the current work package are extracted from the project list, and the correlation and influence features are extracted from multi-source data, including structured data, semi-structured data and unstructured data. The basic attribute features and correlation and influence features are quantized and encoded using natural language processing technology to obtain the feature vector of the current work package.

[0068] First, accurately extract the basic attribute characteristics of the current work package from the project list. This includes the professional category of the task (such as sub-fields like civil engineering and structural engineering, mechanical and electrical installation, etc.), planned duration (such as 30 days, 60 days, etc.), planned cost (such as budget amounts like 500,000 yuan, 1 million yuan, etc.), and deliverable standards (such as a 98% acceptance rate, compliance with industry standards, etc.). This ensures complete coverage of the core inherent attributes of the work package. Simultaneously, extract related impact characteristics from multi-source data. For structured data (such as historical project cost databases and material price indices), extract features such as cost fluctuations and material supply stability. For semi-structured data (such as project schedule documents and supplier contracts), extract features such as schedule milestone requirements and supplier performance. By extracting features such as contract deadlines and other characteristics from unstructured data (such as weather data and policy and regulatory texts), and features such as suitable construction days and policy compliance requirements, we can achieve a comprehensive capture of external influencing factors of the work package. We then use natural language processing technology to quantify and encode the extracted basic attribute features and related influence features. Textual features (such as professional categories and policy requirements) are converted into numerical vectors through word embedding algorithms, numerical features (such as planned construction period and planned cost) are normalized, and categorical features (such as the standard level of deliverables) are encoded using one-hot encoding. Finally, all quantified features are integrated into a unified and computable feature vector for the current work package, providing a standardized data foundation for subsequent matching of historical cases and generation of micro-intervention strategies.

[0069] The basic attribute characteristics include the professional category of the task, the planned duration, the planned cost, and the standard characteristics of the deliverables.

[0070] The characteristics of association effects include the association effects of structured data, the association effects of semi-structured data, and the association effects of unstructured data.

[0071] Based on the feature vector of the current work package, the most similar successful intervention cases are retrieved and matched from the historical project database to construct an initial candidate strategy pool.

[0072] Receive the target work package that needs intervention, extract the feature vector of the target work package, calculate the similarity between the feature vector and the feature vector of all work packages in the historical project database, select the top N historical work packages with the highest similarity scores from the historical project database, and extract the intervention measures corresponding to the top N historical work packages to form a set of candidate intervention strategies.

[0073] Upon receiving the target work package requiring intervention, first extract the basic attribute features (professional category, planned duration, etc.) and related impact features (multi-source data association features) of the target work package, encode them into feature vectors, and then calculate the similarity between this vector and the feature vectors of historical normal work packages. If the similarity is lower than a preset threshold (e.g., <60%), or if its key performance indicators (schedule, cost, etc.) deviate from the planned threshold (e.g., schedule delay exceeds 10%, cost overrun exceeds 8%), then intervention is deemed necessary.

[0074] When selecting the top N work packages with the highest similarity from the historical project database, the value of N should be set reasonably. If the value of N is too small (e.g., N=1), the candidate strategies may be insufficient due to the limitations of a single case. If the value of N is too large (e.g., N=20), ineffective strategies for low-similarity cases may be introduced. Based on the size of the historical database and the distribution of case similarity, a reasonable range of 5-10 should be set to balance the diversity and effectiveness of strategies.

[0075] The system acquires the current resource status of the project in real time and converts it into a numerical state vector. Based on the risk preference level set at the start of the project, it loads the corresponding risk preference parameters and encodes the numerical state vector and risk preference parameters into an environmental state vector, which is then used as the input to the reinforcement learning model.

[0076] Risk preference levels include conservative, balanced, and aggressive.

[0077] Load the pre-trained reinforcement learning model. The core of the reinforcement learning model is a Q-value function, used to evaluate the optimal long-term reward obtained by implementing each policy from the set of candidate intervention policies given the environment state vector. The expression for the Q-value function is: .

[0078] in, Represents the environment state vector. Indicates candidate intervention strategies, Indicates immediate feedback. Indicates the discount factor; in, Indicates the implementation of candidate intervention strategies The state vector of the new environment after entering; express All candidate intervention strategies; Indicates the optimal long-term return in the future; The environmental state vector and the set of candidate intervention strategies are input into the reinforcement learning model. The Q-value of each candidate intervention strategy is calculated, and the candidate intervention strategy with the largest Q-value is selected as the output of the reinforcement learning model to generate a micro-intervention strategy.

[0079] The pre-trained reinforcement learning model is trained with sufficient historical project data as training samples. The samples need to cover intervention cases of different types of projects (such as large-scale infrastructure and small-scale civil projects), including the environmental status (resource status, risk preference), intervention strategies and corresponding returns (progress improvement, cost savings, etc.) of each case. The model parameters are optimized through repeated iterative training to ensure that the model has the ability to evaluate the effectiveness of the strategy.

[0080] By extracting the basic attribute features of the work package (task professional category, planned duration, etc.) and the multi-source correlation influence features (structured, semi-structured, and unstructured data features), and quantifying and encoding them into feature vectors using natural language processing technology, the unique attributes and affected dimensions of the work package can be accurately captured. Then, by combining historical project databases to match similar successful cases, a candidate strategy pool is constructed to ensure that the initial strategy fits the actual needs of the work package and avoids the problem of insufficient strategy adaptability. The project resource status is transformed into a numerical state vector in real time. Combined with risk preference parameters such as conservative, balanced, and aggressive, an environmental state vector is constructed and input into a pre-trained reinforcement learning model. The model evaluates the immediate return and the optimal long-term return of each candidate strategy through a Q-value function. Finally, the strategy with the largest Q-value is selected as the output, which not only adapts to the current project resources and risk preferences, but also takes into account long-term management benefits, realizing the upgrade of strategy from experience matching to intelligent optimization.

[0081] The logic for transforming resource status into a numerical state vector includes: Resource status includes human resources status, equipment resources status, financial resources status, time resources status, and material resources status.

[0082] Identify specific indicators for each resource status, determine the encoding strategy based on the corresponding data type, map the resource status to numbers through the encoding strategy, and normalize the numbers to obtain a numerical status vector.

[0083] Data types include continuous, categorical, ordinal, and boolean.

[0084] The coding strategy is determined based on the corresponding data type. First, the data type of each resource indicator (such as the amount of available funds and equipment type) is determined, clarifying whether it belongs to continuous, categorical, or ordinal, or Boolean type. Then, the coding method is matched according to the type: continuous data uses standardized normalization or Min-Max normalization (both map the data to a unified interval), categorical data uses one-hot coding (tower crane, excavator, corresponding to [1,0], [0,1]), ordinal data uses sequential coding (sufficient-moderate-shortage is set to 3, 2, 1), and Boolean data uses 0-1 coding ("available" is 1, "unavailable" is 0). Finally, the coding is ensured to match the type, avoiding resource information distortion and ensuring the effectiveness of numerical state vectors.

[0085] By normalizing all resource indicators to the same numerical range, for continuous data such as financial amounts and remaining construction period, the original values ​​are mapped to the [0,1] range through Min-Max normalization. For example, if the available funds for a project are between 500,000 and 1,500,000 yuan, 1,000,000 yuan can be converted to 0.5 after calculation. For coded Boolean and ordinal data, they are synchronously verified and adjusted to the corresponding range (e.g., Boolean values ​​of 0-1 are directly retained in the [0,1] range, and the 3-2-1 ordinal codes corresponding to sufficient-moderate-shortage are converted to 0.67-0.33-0 through calculation). In the end, all resource indicators are ensured to be on the same scale, so that the numerical state vector can objectively reflect the relative importance of various resources, and provide balanced data support for reinforcement learning models to accurately assess the environmental state and generate optimal micro-intervention strategies.

[0086] By transforming the status of five core resources—human resources, equipment resources, financial resources, time resources, and material resources—into numerical state vectors through an indicator identification-coding mapping-normalization process, this approach solves the problem of inconsistent analysis of different resource types (such as monetary data for financial resources, personnel quantity data for human resources, and equipment condition data for equipment resources) due to their different forms. It converts previously scattered and unstructured resource information into standardized data that can be input into the model for calculation, providing a unified data foundation for subsequent reinforcement learning model evaluation of micro-intervention strategies. This approach clearly covers the five core resources that affect the progress of work packages during project execution, including physical resources such as human resources, equipment, and materials, as well as supporting resources such as finance and time. This ensures that the quantitative results of resource status comprehensively reflect the actual resource supply capacity of the project, avoiding biases in the model's judgment of environmental conditions due to missing resource dimensions, and improving the accuracy of micro-intervention strategy generation.

[0087] The logic behind quantifying the causal effect of micro-intervention strategies on job package performance includes: Key data were collected from the experimental group and the control group at preset time points. The performance change of the experimental group before and after the implementation of the micro-intervention strategy was calculated and recorded as the performance change of the experimental group. At the same time, the performance change of the control group during the same time period was calculated and recorded as the performance change of the control group. The net effect value was obtained by calculating the difference between the performance change of the experimental group and the performance change of the control group. The net effect value was used as the causal effect value.

[0088] Preset time nodes: Based on the work package duration and the intervention effectiveness pattern, the time nodes are set in layers—short-term tasks are set to 1 day before intervention and daily after intervention, while medium- and long-term tasks are set to 1 week before intervention, every 3 days during intervention, and 1 week after intervention, to ensure that the experimental group and the control group are completely aligned (e.g., both are collected at T0, T3, and T7) and cover the entire chain, providing complete data support for the calculation of net effect value.

[0089] By simultaneously collecting key data from the experimental and control groups, calculating the performance changes of both groups within the same time period, and then subtracting the performance changes of the control group from the experimental group's performance change to obtain the net effect value, the interference of external environmental factors (such as material price fluctuations and weather changes) and the overall project rhythm can be effectively eliminated. This allows for the precise extraction of the specific impact of micro-intervention strategies on work package performance, avoiding the misattribution of performance changes caused by non-intervention factors to intervention effects, ensuring the authenticity of causal effect values, and transforming causal effects into directly comparable net effect values ​​(such as the difference in the number of days the schedule is ahead of schedule or the difference in the amount of cost savings). This breaks the ambiguity of traditional qualitative descriptions of intervention effects, allowing managers to clearly judge the actual magnitude of the micro-intervention strategy's effect (such as a strategy that shortens the construction period by 5 days in the experimental group compared to the control group), providing clear data basis for whether to promote the strategy and adjust strategy parameters in the future.

[0090] The logic for updating the connectivity of corresponding nodes in a dynamic causal network includes: Acquire causal effect records for the quantification unit, which include the characteristics of the micro-intervention strategy, the initial characteristics of the work package, and the causal effect values.

[0091] The features of the micro-intervention strategy and the initial features of the working package are concatenated as input features, which are then input into the dynamic causal network in the building module. Forward propagation calculation is performed on the dynamic causal network, and the predicted value of the causal effect is output. The difference between the predicted value and the causal effect value is calculated using a preset loss function.

[0092] Based on the difference, backpropagation calculation is performed on the dynamic causal network. Starting from the output layer of the dynamic causal network, the partial derivatives of the loss function with respect to each weight parameter of the dynamic causal network are calculated layer by layer, and the partial derivatives are used as the gradient.

[0093] The partial derivatives of the loss function with respect to each weight parameter in the dynamic causal network are calculated layer by layer. The starting point for the calculation is clearly defined as the output layer of the dynamic causal network. Based on the difference between the predicted causal effect value calculated by the preset loss function and the actual causal effect value, according to the chain rule, the partial derivatives of the loss function with respect to the output value of the output layer neurons are first solved. Then, combined with the derivatives of the activation functions of the output layer neurons, the partial derivatives of the loss function with respect to the weight parameters of the output layer are obtained. Then, based on the partial derivatives of the output layer, the process is reversed to the hidden layer. The partial derivatives of the loss function with respect to the output value of each neuron in the hidden layer are calculated in turn. Combining the relationship between the derivatives of the activation functions of the hidden layer and the weight parameters of that layer, the partial derivatives of the loss function with respect to each weight parameter of the hidden layer are gradually solved. Finally, the process continues to backpropagate to the input layer and repeat the above derivation process until the partial derivatives of the loss function with respect to the weight parameters of all layers of the network are calculated. These partial derivatives are mapped one by one to each weight parameter to form gradients that can be used for subsequent parameter updates.

[0094] Based on the gradient and a preset learning rate, the update amount of each weight parameter is calculated, and the optimizer iteratively updates the weight parameters in the dynamic causal network according to the update amount.

[0095] The preset learning rate is set with the core objective of balancing update efficiency and parameter stability. A higher value (0.01-0.1) is set in the initial training phase to accelerate parameter optimization. In the mid-to-late stages, it is reduced to 0.001-0.01 to avoid oscillations. When data fluctuations are large, it is further reduced to 0.0001-0.001 to reduce interference from abnormal data.

[0096] The default loss function is set so that the causal effect value is continuous. The mean squared error loss function (MSE) is used first to accurately quantify the prediction bias and ensure smooth gradient calculation. When there are a few outliers, the mean absolute error loss function (MAE) is used to enhance robustness and avoid the network from over-adjusting parameters.

[0097] By acquiring the causal effect records (including micro-intervention strategy features, initial features of the work package, and causal effect values) output by the quantization unit, the two types of features are concatenated into input features and input into the dynamic causal network. After forward propagation to calculate the predicted value and backpropagation to calculate the gradient, the network weight parameters are iteratively updated by combining the learning rate and the optimizer. This allows the network to absorb new intervention effect data in real time, continuously optimize the correlation logic between intervention strategy, work package features, and causal effect, improve the accuracy of subsequent causal effect prediction, and avoid judgment bias caused by data lag. Each update is based on real causal effect records, rather than relying on fixed initial data. This allows the dynamic causal network to gradually adapt to changes in the project execution process (such as adjustments to resource supply patterns and the impact of external environmental fluctuations on intervention effects). For example, if a certain type of micro-intervention strategy has a causal effect of 5% on cost savings in the early stage of the project, and as material prices rise, subsequent records show that the effect drops to 3%, the network can accurately capture this change through weight updates, making subsequent decisions based on the network more in line with the actual needs of the scenario.

[0098] This invention constructs a dynamic causal network based on work packages in the project list and influencing factors extracted from multi-source data. This achieves comprehensive integration of multi-source factors, providing support for management decisions and enhancing their scientific rigor. Based on this network, scenario simulation and counterfactual inference are used to simulate node state changes when hypothetical risks such as resource shocks and environmental mutations occur, quantifying the impact of risks on total project cost and duration. This provides forward-looking support for risk prevention and control. Simultaneously, parallel work packages can be identified during project execution, generating micro-intervention strategies and setting up experimental and control groups. The effect of the strategies is quantified through differential analysis, and the dynamic causal network parameters are updated to adapt to real-time data and intervention effects, continuously improving management accuracy and efficiency, and achieving full-cycle project control.

[0099] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based engineering cost progress management system, characterized by, The system comprises a construction module, a prediction module and an optimization module. The construction module is configured to construct a dynamic causal network based on each work package in a project list and multi-source influencing factors extracted from multi-source data. The prediction module is configured to quantitatively predict a hypothetical risk through scenario deduction based on the dynamic causal network, and output a prediction result. The optimization module is configured to identify parallel work packages during project execution, form an experimental group and a control group, and iteratively optimize the dynamic causal network according to a comparison result of the experimental group and the control group.

2. The artificial intelligence-based engineering cost progress management system of claim 1, wherein: The construction module comprises a mapping unit, a forming unit and an assignment unit. The mapping unit is configured to map each work package in the project list and the multi-source influencing factors extracted from the multi-source data into nodes in the dynamic causal network, respectively. One work package represents a project task in the project list, and the project task comprises task specific steps, delivery results and execution requirements. The multi-source influencing factors comprise structured data influencing factors, semi-structured data influencing factors and unstructured data influencing factors. The structured data influencing factors comprise a historical project cost database and a material price index. The semi-structured data influencing factors comprise a project schedule plan file and a supplier contract. The unstructured data influencing factors comprise weather data, social media information, policy and regulation texts and satellite remote sensing images. The forming unit is configured to obtain historical project data comprising multi-dimensional variables related to project cost and schedule, take the historical project data as an observation sample set, and perform statistical analysis on the observation sample set by using a causal discovery algorithm. The statistical analysis comprises: determining real causal directions and directed connection relationships between variables of different nodes by excluding false correlations caused by common causes and accidental factors through testing conditional independence between the variables, solidifying the directed connection relationships, and forming a topology structure of an initial causal network. The assignment unit is configured to generate dynamic weights for each directed edge in the topology structure of the initial causal network, and obtain a dynamic causal network. The dynamic weights are used to quantitatively determine the influence intensity of a cause node on a result node, and are continuously updated according to newly accessed project data. The logic of generating dynamic weights comprises: According to the node type and the characteristics of the processed project data, a matching calculation method is selected from a preset weight calculation method library, a new weight is calculated according to the calculation method for weight updating, and the new weight is compared with the current dynamic weight, and when the difference between the new weight and the current dynamic weight exceeds a preset difference value, the weight is updated.

3. The artificial intelligence-based engineering cost progress management system of claim 2, wherein: The logic of mapping into nodes in the dynamic causal network comprises: Each work package in the project list is traversed, and the work package is mapped into an execution node in the dynamic causal network, and multi-source influencing factors are mapped into entity nodes of different types in the dynamic causal network according to data types and business semantics. The data types comprise structured data, semi-structured data and unstructured data. The different types comprise resource nodes, environment nodes, risk nodes and management nodes.

4. The artificial intelligence-based engineering cost progress management system of claim 3, wherein: The prediction module comprises a scenario deduction unit. The scenario deduction unit is configured to receive a user-inputted hypothetical risk condition, and perform counterfactual inference on the dynamic causal network through a counterfactual inference algorithm, simulate state changes of each node in the dynamic causal network when the hypothetical risk condition is true, and quantitatively calculate an affected degree of a project key indicator based on state change data of each node, and output a prediction result of the project key indicator under the hypothetical risk condition. The hypothetical risk condition includes a resource impact type condition, an environmental mutation type condition, a risk manifestation type condition, and a management intervention type condition. The project key indicator includes a total project cost and a total project duration.

5. The artificial intelligence-based engineering cost progress management system of claim 4, wherein: The optimization module includes an identification unit, a generation unit, a setting unit, a quantification unit, and an updating unit. The identification unit is configured to scan a project work breakdown structure in real time during project execution, and identify parallel work packages. The generation unit is configured to generate a micro-intervention strategy for the identified parallel work packages. The setting unit is configured to take one of the parallel work packages as an experimental group, implement the micro-intervention strategy on the experimental group, and take the remaining parallel work packages as a control group. The quantification unit is configured to collect key data of the experimental group and the control group respectively, quantitatively calculate a causal effect of the micro-intervention strategy on work package performance through a difference analysis method, and obtain a causal effect value. The key data includes progress data, cost data, and quality data. The updating unit is configured to update a connection relationship of a corresponding node in the dynamic causal network.

6. The artificial intelligence-based engineering cost progress management system of claim 5, wherein: The logic of identifying parallel work packages includes: Collecting relevant information of a project, and extracting text clues describing parallel relationships in the relevant information by using natural language processing technology, and taking the text clues as a parallel relationship candidate set; Obtaining a text evidence strength of each pair of parallel relationships in the parallel relationship candidate set, and confirming a current parallel relationship as a valid parallel work package when the text evidence strength is greater than a preset effective threshold; The relevant information includes project plan documents, construction logs, and meeting minutes; The parallel relationship includes simultaneous, synchronous, and cross operations.

7. The artificial intelligence-based engineering cost progress management system of claim 6, wherein: The logic of generating a micro-intervention strategy includes: Extracting basic attribute features of a current work package from a project list, and extracting associated impact features from multi-source data including structured data, semi-structured data, and unstructured data, and quantitatively encoding the basic attribute features and the associated impact features through natural language processing technology to obtain a feature vector of the current work package; The basic attribute features include task belonging to a professional category feature, a planned duration feature, a planned cost feature, and a delivery result standard feature; The associated impact features include structured data associated impact features, semi-structured data associated impact features, and unstructured data associated impact features; Based on the feature vector of the current work package, retrieving and matching the most similar successful intervention cases from a historical project database to construct an initial candidate strategy pool; Receiving a target work package requiring intervention, extracting a feature vector of the target work package, performing similarity calculation on the feature vector with feature vectors of all work packages in a historical project database, screening top N historical work packages with highest similarity scores from the historical project database, and extracting intervention measures corresponding to the top N historical work packages to form a candidate intervention strategy set; Real-time acquisition of resource status of a current project, and conversion of the resource status into a numerical state vector, loading of a corresponding risk preference parameter according to a risk preference level set at project start, and encoding of the numerical state vector and the risk preference parameter into an environment state vector as an input of a reinforcement learning model; The risk preference level includes conservative, balanced and aggressive types; loading a pre-trained reinforcement learning model, the core of which is a Q-value function for evaluating the optimal long-term return of executing each policy in a set of candidate intervention policies under an environmental state vector, the Q-value function expressed as ; wherein, represents an environmental state vector, represents a candidate intervention policy, represents an immediate reward, represents a discount factor; wherein, representing executing the candidate intervention strategy a new environment state vector entered after representing all candidate intervention strategies under E[RT] represents the expected future optimal long-term return; Inputting the environment state vector and the candidate intervention strategy set into the reinforcement learning model, calculating Q values of each candidate intervention strategy, selecting a candidate intervention strategy with the maximum Q value as an output of the reinforcement learning model, and generating a micro intervention strategy.

8. The artificial intelligence-based engineering cost progress management system of claim 7, wherein: The logic of converting the resource status into a numerical state vector includes: The resource status includes human resource status, equipment resource status, financial resource status, time resource status and material resource status; Identifying specific indicators of each resource status, determining an encoding strategy according to a corresponding data type, mapping the resource status into a number through the encoding strategy, and obtaining a numerical state vector through normalization processing of the number; The data type includes continuous type, category type, ordinal type and Boolean type.

9. The artificial intelligence-based engineering cost progress management system of claim 8, wherein: The logic of quantifying the causal effect of the micro intervention strategy on work package performance includes: At a preset time node, key data of the experimental group and the control group are collected respectively, and the performance change amount of the experimental group before and after the implementation of the micro intervention strategy is calculated, which is recorded as the experimental group performance change amount, and the performance change amount of the control group in the same time period is calculated, which is recorded as the control group performance change amount, and the difference between the experimental group performance change amount and the control group performance change amount is calculated to obtain a net effect value, which is used as a causal effect value. 10.The artificial intelligence-based engineering cost progress management system of claim 9, wherein: The logic of updating the connection relationship of the corresponding node in the dynamic causal network includes: Obtaining a causal effect record of the quantification unit, the causal effect record including features of the micro intervention strategy, initial features of the work package and a causal effect value; Splicing the features of the micro intervention strategy and the initial features of the work package into input features, inputting the input features into the dynamic causal network in the construction module, performing forward propagation calculation on the dynamic causal network, and outputting a predicted value of the causal effect, and calculating a difference between the predicted value and the causal effect value by using a preset loss function; Based on the difference, the dynamic causal network is calculated by back propagation, starting from the output layer of the dynamic causal network, and the partial derivative of the loss function to each weight parameter in the dynamic causal network is calculated layer by layer, and the partial derivative is used as a gradient; Based on the gradient, combining a preset learning rate, the update amount of each weight parameter is calculated, and the optimizer is used to iteratively update the weight parameters in the dynamic causal network according to the update amount.

Images