Multi-source data driven project progress intelligent deduction method and system

Abstract

The present application relates to the technical field of project progress management, and more particularly to a multi-source data driven project progress intelligent deduction method and system. The method obtains associated progress, resource and dependency relationship data in multiple heterogeneous data sources, performs multi-scale coupling mapping based on time sequence alignment rules to generate standardized project state representation; a dynamic causal diagram model is constructed according to multi-dimensional features, causal transmission paths and constraint conditions are mined through reverse causal inference, and driving factors and their weight distribution are determined by causal decoupling and tracing; according to the driving factors and weight distribution, combined with project constraint conditions and resource availability, multiple candidate deduction paths are generated by multi-scenario simulation and risk propagation evaluation is performed, and the target deduction path is selected.

Description

Technical Field

[0001] This invention relates to the field of project schedule management technology, and in particular to a method and system for intelligent project schedule extrapolation driven by multi-source data. Background Technology

[0002] In the current field of project schedule management, conventional practices mainly rely on a single data source (such as planning data in project management software or task records in enterprise resource planning systems) for schedule projection. These methods are typically based on fixed critical path analysis or Gantt chart logic, generating linear or tree-like project execution plans through pre-defined dependencies and resource allocations. During the projection process, schedule assessment often uses static threshold judgments or statistical regression based on historical data, such as predicting subsequent nodes through percentage completion or milestone achievement rates. Some systems introduce Monte Carlo simulations or Bayesian networks, but the core remains dominated by single time series data, lacking the ability to integrate multi-source heterogeneous data.

[0003] Conventional approaches suffer from two significant drawbacks. First, the limited availability of data sources prevents schedule projections from reflecting the complex coupling relationships in actual project operation. For example, resource data (such as manpower allocation and equipment availability) and dependency data (such as logical constraints between tasks) often come from different systems with inconsistent time granularities. Conventional methods struggle to achieve synchronized mapping across the time and logical dimensions, making projection results prone to deviating from reality, especially in cross-departmental collaborations or dynamic resource adjustments, where prediction errors increase significantly. Second, existing projection mechanisms lack in-depth analysis of causal relationships. Most methods focus only on statistical correlations or temporal relationships between data, failing to identify the fundamental factors driving schedule evolution, such as hidden bottlenecks in the critical path or the transmission paths of resource conflicts. When the project environment changes, correlation-based models struggle to adapt, leading to rigid projection paths and an inability to effectively address unforeseen risks, such as resource shortages or cascading delays caused by changes in dependencies. Summary of the Invention

[0004] This invention provides a method and system for intelligent project progress prediction driven by multi-source data, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a multi-source data-driven intelligent project progress prediction method, comprising:

[0006] Acquire progress data, resource data, and dependency data associated with the target project from multiple heterogeneous data sources. Based on time-series alignment rules, perform multi-scale coupling mapping of the progress data, resource data, and dependency data in the time and logical dimensions to generate a standardized project status representation.

[0007] A dynamic causal graph model is constructed based on the multi-dimensional features in the standardized project status representation. The causal transmission paths and constraints between the multi-dimensional features are explored through reverse causal inference. The causal decoupling and tracing method is used to determine the driving factors affecting the progress evolution and their weight distribution.

[0008] Based on the driving factors and their weight distribution, combined with the preset project constraints and resource availability, a multi-scenario simulation mechanism is used to generate multiple candidate progress projection paths, and a risk propagation assessment is performed on each candidate progress projection path to select the target projection path that meets the confidence requirements.

[0009] The target derivation path is deconstructed into a phased execution strategy. At each execution stage, the actual progress status and the state deviation vector of the target derivation path are collected. Based on the state deviation vector, the causal transmission path in the dynamic causal graph model is locally reconstructed. The weight distribution of the driving factors is adjusted through gradient backpropagation to drive the iterative convergence of the derivation model.

[0010] Based on time-series alignment rules, the progress data, resource data, and dependency data are subjected to multi-scale coupling mapping of time and logical dimensions to generate a standardized project status representation, including:

[0011] The timestamp sequences of progress data, resource data, and dependency data are decomposed into multiple granularities. Time anchors are identified at each granularity level, and a cross-data source time alignment benchmark is constructed. Based on the time alignment benchmark, heterogeneous time scales are decoupled and reconstructed at multiple levels, forming a unified time series expression of each data source in a unified time coordinate system.

[0012] Under the unified time coordinate system, a logical hierarchy topology graph is constructed based on the logical constraints in the dependency data. The progress data and the resource data are then projected in layers according to the node hierarchy of the logical hierarchy topology graph to generate state subspaces at each level.

[0013] Multi-scale tensor reconstruction is performed on the state subspaces of each level. By calculating the coupling strength between different granularity levels and logical levels, the coupling strength is used as a modulation factor to perform hierarchical fusion and dimensional compression of each state subspace, forming a standardized project state representation.

[0014] A logical hierarchy topology graph is constructed based on the logical constraints in the dependency data. The progress data and resource data are then projected hierarchically according to the node hierarchy of the logical hierarchy topology graph to generate state subspaces for each level, including:

[0015] Transitive closure expansion is performed on the logical constraint relationships in the dependency data to identify the direct and indirect dependency paths between data entities. A directed dependency graph is constructed based on the direct and indirect dependency paths. The directed dependency graph is then hierarchically analyzed and the topological distance of each node is calculated. Based on the topological distance, the nodes are divided into different logical levels to construct a logical hierarchical topology graph.

[0016] Based on the hierarchical affiliation and semantic identifier of each node in the logical hierarchical topology diagram, multi-dimensional feature matching is performed on the progress data and resource data. The matched progress data and resource data are semantically projected and spatially decomposed along the node hierarchy of the logical hierarchical topology diagram. Within each level, the projected data is reconstructed into state vectors to generate a state subspace with hierarchical isolation.

[0017] Based on the multi-dimensional features in the standardized project status representation, a dynamic causal graph model is constructed. Through reverse causal inference, the causal transmission paths and constraints between these multi-dimensional features are explored. A causal decoupling and source tracing method is used to determine the driving factors influencing schedule evolution and their weight distribution, including:

[0018] The standardized project state representation is deconstructed into multiple semantic layers. Multi-dimensional features are expanded into a set of nodes and an initial directed graph structure is constructed. Reverse intervention signals are implanted and activated in each node of the initial directed graph structure. By tracking the propagation path and attenuation mode of the reverse intervention signals between nodes, the causal transmission path and transmission intensity threshold between multi-dimensional features are restored according to the propagation dynamics mode and condensed into constraints.

[0019] Based on the causal transmission path and the constraints, causal decoupling analysis is performed on the feature nodes in the initial directed graph structure. By constructing a counterfactual scenario, the marginal contribution of each feature node to the progress evolution target under the assumed removal state is calculated. Feature nodes whose marginal contribution exceeds the significance discrimination criterion are locked into the driving factor set through significance separation.

[0020] An interaction coupling degree analysis is performed on each driving factor in the set of driving factors. Based on the interaction coupling degree, the marginal contribution of each driving factor is redundantly corrected. The corrected marginal contribution is then mapped to the weight distribution through global weight redistribution. The weight distribution is then injected into the corresponding node of the initial directed graph structure and drives the evolution to generate a dynamic causal graph model.

[0021] Based on the aforementioned driving factors and their weight distributions, and combined with preset project constraints and resource availability, a multi-scenario simulation mechanism is used to generate multiple candidate schedule projection paths. A risk propagation assessment is then performed on each candidate schedule projection path, and the target projection paths that meet the confidence requirements are selected, including:

[0022] The driving factors and their weight distributions are transformed into state evolution driving vectors through multi-scale feature decomposition and vectorization encoding. The state evolution driving vectors are coupled with preset project constraints to construct a constraint evolution space. In the constraint evolution space, the resource availability is decomposed into multiple perturbations through a multi-scenario simulation mechanism to generate multiple sets of scenario parameter configurations. Parallel simulation is performed on each scenario parameter configuration using the multi-scenario simulation mechanism to generate multiple candidate progress simulation paths and their state transition sequences.

[0023] The state transition sequence of each candidate progress projection path is deconstructed into state node topology and each node is assigned a probability value based on historical transition frequency. A risk propagation topology network is constructed based on the deconstructed state transition nodes and their probability assignments. External risk sources are constructed and activated at each state transition node. The risk sources are driven to cascade and spread along the state transition sequence through the probability assignments. The risk accumulation and path robustness of each candidate progress projection path are calculated.

[0024] The accumulated risk and the robustness of the path are fused together to form a dual evaluation index for confidence quantification. The confidence quantification result is compared with a preset confidence threshold to lock in the target deduction path that meets the confidence threshold requirement.

[0025] A multi-scenario simulation mechanism is used to decompose resource availability into multiple sets of scenario parameter configurations within the constrained evolution space, generating a multi-dimensional perturbation decomposition. Parallel simulation is then performed on each scenario parameter configuration using this mechanism to generate multiple candidate progress paths and their state transition sequences, including:

[0026] Based on the resource availability boundary in the constrained evolution space, a multi-dimensional perturbation matrix is ​​constructed through a multi-scenario simulation mechanism. Different perturbation intensities and directions are set in each dimension of the multi-dimensional perturbation matrix to perform hierarchical perturbation decomposition of resource availability and generate multiple sets of scenario parameter configurations.

[0027] Each scenario parameter configuration is injected into an independent simulation thread of the multi-scenario simulation mechanism. Within each independent simulation thread, a time-series deduction calculation is performed based on the state evolution driving vector and the corresponding scenario parameter configuration. The dynamic evolution process of the project state in each simulation thread is tracked, the state nodes and the transition relationships between nodes in each evolution process are captured, and candidate progress deduction paths and their state transition sequences corresponding to each scenario parameter configuration are generated.

[0028] At each execution stage, a state deviation vector is collected between the actual progress status and the target projection path. Based on the state deviation vector, the causal transmission path in the dynamic causal graph model is locally reconstructed, including:

[0029] In each execution phase of the project, a status acquisition node is set up. At each status acquisition node, the actual progress status at the current moment and the expected progress status of the target projection path at the corresponding moment are acquired synchronously. The difference between the actual progress status and the expected progress status is quantitatively calculated in multiple status dimensions to generate a status deviation vector that represents the degree of deviation in each status dimension. The status deviation vector is then classified by deviation magnitude and labeled with deviation direction.

[0030] The state deviation vector is projected onto the causal node space of the dynamic causal graph model. Sensitive causal nodes affected by deviations in the dynamic causal graph model are identified hierarchically based on the deviation magnitude of the state deviation vector. The causal transmission intensity of the sensitive causal nodes is adjusted. The causal transmission path between the sensitive causal nodes is topologically reconstructed according to the deviation direction of the state deviation vector, and a locally reconstructed dynamic causal graph model is generated.

[0031] A second aspect of this invention provides a multi-source data-driven intelligent project progress prediction system, comprising:

[0032] The data fusion unit is used to acquire progress data, resource data and dependency data associated with the target project from multiple heterogeneous data sources, and perform multi-scale coupling mapping of the progress data, resource data and dependency data in time dimension and logical dimension based on time alignment rules to generate a standardized project status representation.

[0033] The causal analysis unit is used to construct a dynamic causal graph model based on the multi-dimensional features in the standardized project status representation, to mine the causal transmission paths and constraints between the multi-dimensional features through reverse causal inference, and to determine the driving factors and their weight distribution that affect the progress evolution by using the causal decoupling and tracing method.

[0034] The simulation and evaluation unit is used to generate multiple candidate progress simulation paths based on the driving factors and their weight distribution, combined with preset project constraints and resource availability, using a multi-scenario simulation mechanism, and to conduct risk propagation assessment on each candidate progress simulation path, and select the target simulation path that meets the confidence requirements.

[0035] The iterative optimization unit is used to deconstruct the target derivation path into a phased execution strategy, collect the state deviation vector between the actual progress state and the target derivation path at each execution stage, locally reconstruct the causal transmission path in the dynamic causal graph model based on the state deviation vector, and adjust the weight distribution of the driving factors through gradient backpropagation to drive the iterative convergence of the derivation model.

[0036] A third aspect of the present invention provides an electronic device, comprising:

[0037] processor;

[0038] Memory used to store processor-executable instructions;

[0039] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0040] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0041] Compared with existing technologies, this method significantly improves the accuracy and dynamic adaptability of project schedule projection. Through multi-scale coupling mapping of time and logic across multi-source heterogeneous data, it effectively eliminates temporal deviations and logical conflicts between different data sources, forming a more complete and consistent representation of the project status and laying a solid data foundation for subsequent projections. The introduction of a dynamic causal graph model overcomes the limitations of traditional correlation analysis, accurately identifying the deep causal drivers of schedule evolution and their weights, avoiding misjudgments caused by spurious associations, and resulting in more interpretable and robust projection results.

[0042] The multi-scenario simulation mechanism, combined with causal decoupling and source tracing results, automatically generates multiple candidate projection paths. Through risk propagation assessment, it selects the target path with high confidence, enabling the projection process to comprehensively cover uncertain scenarios and significantly reduce the deviation risk of a single predicted path. The target projection path is deconstructed into phased execution strategies, achieving refined control over project progress and facilitating segmented verification and adjustments during actual execution.

[0043] In the actual execution phase, the state deviation vector drives the local reconstruction of the causal graph model and the backpropagation of weight gradients, enabling the inference model to absorb new data in real time and continuously optimize the causal transmission path and the weights of driving factors. This closed-loop iterative mechanism endows the method with self-learning and adaptive capabilities, enabling it to cope with changes in the project environment and sudden disturbances. As the project progresses, the inference model continuously converges to a better state, significantly improving the stability and reliability of long-term inference. Overall, this method achieves a leapfrog progress from static prediction to intelligent inference through the synergistic effect of data fusion, causal reasoning, scenario simulation, and dynamic optimization. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the multi-source data-driven intelligent project progress prediction method according to an embodiment of the present invention.

[0045] Figure 2 This is a flowchart of the local reconstruction process of the dynamic causal graph model based on the state deviation vector in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0048] Figure 1 This is a flowchart illustrating the multi-source data-driven intelligent project progress prediction method according to an embodiment of the present invention.

[0049] Multi-source data-driven intelligent project schedule extrapolation methods include:

[0050] Acquire progress data, resource data, and dependency data associated with the target project from multiple heterogeneous data sources. Based on time-series alignment rules, perform multi-scale coupling mapping of the progress data, resource data, and dependency data in the time and logical dimensions to generate a standardized project status representation.

[0051] A dynamic causal graph model is constructed based on the multi-dimensional features in the standardized project status representation. The causal transmission paths and constraints between the multi-dimensional features are explored through reverse causal inference. The causal decoupling and tracing method is used to determine the driving factors affecting the progress evolution and their weight distribution.

[0052] Based on the driving factors and their weight distribution, combined with the preset project constraints and resource availability, a multi-scenario simulation mechanism is used to generate multiple candidate progress projection paths, and a risk propagation assessment is performed on each candidate progress projection path to select the target projection path that meets the confidence requirements.

[0053] The target derivation path is deconstructed into a phased execution strategy. At each execution stage, the actual progress status and the state deviation vector of the target derivation path are collected. Based on the state deviation vector, the causal transmission path in the dynamic causal graph model is locally reconstructed. The weight distribution of the driving factors is adjusted through gradient backpropagation to drive the iterative convergence of the derivation model.

[0054] In one optional implementation, a multi-scale coupling mapping of the progress data, resource data, and dependency data based on time-series alignment rules is performed to generate a standardized project status representation, including:

[0055] The timestamp sequences of progress data, resource data, and dependency data are decomposed into multiple granularities. Time anchors are identified at each granularity level, and a cross-data source time alignment benchmark is constructed. Based on the time alignment benchmark, heterogeneous time scales are decoupled and reconstructed at multiple levels, forming a unified time series expression of each data source in a unified time coordinate system.

[0056] Under the unified time coordinate system, a logical hierarchy topology graph is constructed based on the logical constraints in the dependency data. The progress data and the resource data are then projected in layers according to the node hierarchy of the logical hierarchy topology graph to generate state subspaces at each level.

[0057] Multi-scale tensor reconstruction is performed on the state subspaces of each level. By calculating the coupling strength between different granularity levels and logical levels, the coupling strength is used as a modulation factor to perform hierarchical fusion and dimensional compression of each state subspace, forming a standardized project state representation.

[0058] After acquiring progress, resource, and dependency data, the timestamps from different data sources often exhibit inconsistent sampling frequencies and significant differences in recording granularity. For example, progress data records milestone completion on a daily basis, while resource data tracks manpower input on an hourly basis, and dependency data uses logical trigger events of task nodes as time markers. To address this heterogeneous time characteristic, the timestamp sequences from each data source are first decomposed into multiple granularities. The original timeline is hierarchically segmented according to multiple granularity levels such as day, week, and month, forming a set of granularity levels. Within each granularity level, representative time anchors are identified by detecting data density abrupt changes, key event trigger points, and common record moments across data sources. The criteria for identifying time anchors include: multiple data sources have valid records at that moment; the moment corresponds to a project milestone or resource scheduling node; and the moment has statistically significant state transition characteristics.

[0059] Based on the identified time anchors, a cross-data source time alignment benchmark is constructed. The construction process involves calculating the offset between the timestamps of each data source and the anchor point at each granularity level, using the anchor point time as a reference. Linear interpolation or piecewise constant interpolation is then used to map the data from the misaligned times to the anchor point time. For intervals with missing data, a weighted padding strategy based on adjacent valid values ​​is adopted, with the weight decreasing as the time distance increases, ensuring the padding values ​​are reasonable in terms of temporal continuity. After anchor alignment is completed, multi-level decoupling and reconstruction operations are performed for the heterogeneous time scales between different data sources: high-frequency data is aggregated and compressed at the coarse-grained level, and low-frequency data is interpolated and expanded at the fine-grained level, ensuring that each data source has a uniform time resolution at each granularity level. Finally, in a unified time coordinate system, the time series data from each data source are aligned with the same time index, forming a unified time series representation for each data source and eliminating time series misalignment caused by differences in sampling frequency.

[0060] After establishing a unified time coordinate system, the logical dimension mapping stage begins. Based on the logical constraints recorded in the dependency data, such as preconditions, parallel constraints, and resource mutual exclusion constraints, a logical hierarchical topology graph is constructed. The logical hierarchical topology graph uses task nodes as vertices and dependencies as directed edges. A topological sorting algorithm determines the hierarchical affiliation of each node, grouping nodes with the same dependency depth into the same logical layer. For parallel tasks with resource sharing constraints, resource contention edges are introduced into the topology graph, making resource constraints explicitly expressed in the graph structure. After the logical hierarchical topology graph is constructed, the progress data and resource data are projected hierarchically according to the node hierarchy of the topology graph. Specifically, for the first... The first logical level extracts the progress status values ​​and resource usage values ​​corresponding to all nodes within that level, and arranges them on a time axis of a unified time coordinate system to form the second logical level. Layer state subspace State subspace It is a two-dimensional matrix, where the row dimension corresponds to the set of nodes within the corresponding level, and the column dimension corresponds to the time step sequence in a unified time coordinate system. The matrix elements are the state feature vectors of the corresponding nodes at their respective times. By performing the above hierarchical projection operation on all logical levels, a set of state subspaces covering all logical levels is obtained. ,in This represents the total number of logical levels.

[0061] The generation of the state subspace set provides a structured input for subsequent multi-scale tensor reconstruction. Tensor expansion is performed on the state subspaces at each level, introducing the time granularity dimension as a third tensor dimension, so that each state subspace is expanded from a two-dimensional matrix to a third-order tensor. The three dimensions correspond to the node dimension, time step dimension, and granularity level dimension, respectively. Based on this, the coupling strength between different granularity levels and the logical level is calculated. The coupling strength reflects the degree to which temporal changes at a certain granularity level affect the state of a certain logical level, and is quantified by calculating the mutual information or partial correlation coefficient between the two. Let the granularity level index be... Logical level index is The coupling strength between the two is denoted as Its calculation is based on the statistical dependency between the temporal characteristics at this granularity level and the state characteristics at this logical level. Coupling strength matrix It describes the global association structure between the granular space and the logical space.

[0062] Coupling strength As a modulation factor, hierarchical fusion is performed on each state subspace. The specific process of hierarchical fusion is as follows: for each logical level... , from different granular levels State information according to the corresponding coupling strength Weighted superposition and weight normalization are performed to ensure the consistency of the fusion results across numerical scales. The fusion yields a comprehensive state representation for each logical level, which includes both fine-grained short-term fluctuations and retains long-term trend characteristics at the coarse-grained level. After hierarchical fusion, dimensionality compression is performed on the fusion results. Principal component analysis or an autoencoder-based nonlinear dimensionality reduction method is used to map the high-dimensional fusion features to a low-dimensional compact representation space, removing redundant dimensions and retaining core feature components that discriminate the evolution of project states. The compressed low-dimensional representations of each logical level are concatenated in logical order to form a standardized project state representation covering the entire project structure. This standardized project state representation has a unified time coordinate in the time dimension, retains the task dependency structure in the logical dimension, and undergoes multi-scale coupling modulation in the feature dimension, providing high-quality structured input for the subsequent construction of dynamic causal graph models.

[0063] In practical engineering projects, timestamps from different data sources may exhibit time zone differences or system clock drift, necessitating time zone unification and clock calibration preprocessing before multi-granularity decomposition. For data sources with periodic sampling patterns, their periodicity can be leveraged to assist in the identification of time anchors, improving the robustness of anchor detection. When cyclic dependencies exist in dependency data, loop detection and breaking are required before constructing the logical hierarchy topology graph. This can be achieved by introducing virtual nodes or relaxing constraints to transform the cyclic graph into a directed acyclic graph, ensuring the feasibility of topology sorting. Calculating coupling strength requires a certain amount of data. When the number of cross-samples between a certain granularity level and the logical level is insufficient, regularization estimation methods can be used to improve the stability of coupling strength estimation and avoid distortion of coupling strength due to sample sparsity.

[0064] In one optional implementation, a logical hierarchy topology graph is constructed based on the logical constraints in the dependency data. The progress data and the resource data are then projected hierarchically according to the node hierarchy of the logical hierarchy topology graph to generate state subspaces for each level, including:

[0065] Transitive closure expansion is performed on the logical constraint relationships in the dependency data to identify the direct and indirect dependency paths between data entities. A directed dependency graph is constructed based on the direct and indirect dependency paths. The directed dependency graph is then hierarchically analyzed and the topological distance of each node is calculated. Based on the topological distance, the nodes are divided into different logical levels to construct a logical hierarchical topology graph.

[0066] Based on the hierarchical affiliation and semantic identifier of each node in the logical hierarchical topology diagram, multi-dimensional feature matching is performed on the progress data and resource data. The matched progress data and resource data are semantically projected and spatially decomposed along the node hierarchy of the logical hierarchical topology diagram. Within each level, the projected data is reconstructed into state vectors to generate a state subspace with hierarchical isolation.

[0067] After obtaining the dependency data, the logical constraints are first processed using transitive closure expansion. Transitive closure expansion refers to iteratively deriving and making explicit all indirect dependencies based on the original dependencies. Specifically, if data entities... Directly dependent on entities ,entity Directly dependent on entities Then, the transitive closure operation can be used to identify... right There is an indirect dependency path. Assume the dependency data contains... Each data entity, in an adjacency matrix This represents the initial direct dependency, where Representing entities Directly dependent on entities .right Performing the transitive closure operation yields the complete transitive closure matrix. ,in Representing entities To the entity There exists at least one dependency path (direct or indirect). This is achieved through... and The difference analysis can clearly distinguish between direct and indirect dependency paths, providing complete path information for the subsequent construction of a directed graph of dependency relationships.

[0068] After identifying direct and indirect dependency paths, a directed dependency graph is constructed, using data entities as nodes and dependency relationships as directed edges. ,in For a set of nodes, This is a set of directed edges. Direct dependencies correspond to solid edges in the directed graph, while indirect dependencies are labeled as dashed edges to distinguish between strong and weak dependencies in subsequent analysis. After construction, [the following steps are performed]. A hierarchical analysis is performed, employing a topological sorting algorithm based on breadth-first traversal. Starting from the root node of the graph (i.e., the node without any incoming edges), the algorithm traverses downwards level by level, calculating the topological distance of each node relative to the root node. Defined as from the root node to the node The longest directed path length. In cases where multiple paths lead to the same node, the maximum path length is taken as the topological distance to that node to ensure the strictness and completeness of the hierarchical division.

[0069] Based on the topological distance of each node All nodes are grouped according to their topological distance values, with nodes having the same topological distance grouped into the same logical level, thus constructing a logical hierarchical topology graph. Let the total number of logical levels be . (Represents the total number of logical levels), the first The set of nodes contained in each logical level is denoted as ,satisfy Furthermore, the sets of nodes at each level are mutually exclusive. The logical hierarchy topology graph vertically reflects the hierarchical dependencies between data entities; the lower the level (i.e., the more dependent the data entity), the higher the hierarchy. Nodes with smaller values ​​are closer to the source of the project's fundamental constraints, while higher-level nodes reflect more complex dependency relationships. In project schedule management scenarios, lower-level nodes typically correspond to basic resource constraints or milestone events, while higher-level nodes correspond to comprehensive schedule goals or cross-departmental collaborative tasks.

[0070] After the logical hierarchical topology diagram is constructed, multi-dimensional feature matching is performed on the progress data and resource data based on the hierarchical affiliation and semantic identifiers of each node. Semantic identifiers are semantic description labels assigned to each data entity during the data preprocessing stage, covering dimensions such as task type, resource category, and time granularity. The core of multi-dimensional feature matching is to semantically associate each record in the progress and resource data with nodes in the logical hierarchical topology diagram. The matching rules comprehensively consider three dimensions: task code correspondence, time window overlap, and resource affiliation identifiers. Specifically, for task records in the progress data... Calculate its relationship with the node Overall matching score The score is obtained by weighted summation of three factors: task coding similarity, time window overlap rate, and resource ownership consistency. Exceeding the preset matching threshold When the record hits a node, it is determined that the record has been hit. The logical level at which it is located.

[0071] For matched progress and resource data, semantic projection and spatial decomposition are performed along the node hierarchy of the logical topology graph. Semantic projection refers to extracting the feature dimensions that are semantically relevant to the current level from the original data record to form a local feature representation for that level. Spatial decomposition, on the other hand, divides the mixed features across levels according to the level boundaries, ensuring that the feature representation of each level only contains information directly related to the logical constraints of that level. Taking resource data as an example, a resource record may involve tasks at multiple levels. Spatial decomposition extracts the resource consumption of different levels in the record and assigns them to their respective levels, avoiding information overlap between levels.

[0072] Within each level, the projected data undergoes state vector reconstruction to generate a state subspace with hierarchical isolation. The state vector reconstruction process includes three steps: feature normalization, missing value imputation, and dimension alignment. Feature normalization employs an adaptive normalization method based on the statistical distribution within the level, uniformly scaling features of different dimensions to ensure comparability across numerical ranges. Missing value imputation uses interpolation estimation based on the state values ​​of adjacent nodes within the level; for structurally missing values ​​(i.e., nodes at a certain level have no data records within a specific time period), zero vectors are used for imputation with a missing value marker. Dimension alignment ensures that the state vectors of all nodes within the same level have a uniform feature dimension, facilitating subsequent vector operations and distance calculations within the state subspace.

[0073] After the reconstruction is completed, the first Each logical level's state subspace is composed of the set of state vectors of all nodes within that level, denoted as . . It exhibits hierarchical isolation, meaning that the state subspaces at different levels are semantically independent, and the relationships between levels are only reflected through the edge structure of the logical hierarchical topology graph, rather than being mixed and encoded within the state vectors. This hierarchical isolation design allows subsequent multi-scale coupling mapping to accurately characterize the cross-level dependency transmission mechanism while maintaining the independence of states within each level, thus providing a structurally clear and semantically explicit input representation for the construction of dynamic causal graph models.

[0074] In actual engineering projects, the number of layers in the logical hierarchy topology diagram Typically, the depth corresponds to the project's Work Breakdown Structure (WBS), generally between 3 and 7 levels. For large-scale projects, the number of nodes in the directed graph of dependencies... When the computational complexity reaches thousands of units, the transitive closure operation is relatively high. Optimization algorithms based on sparse matrices or block-based parallel computing strategies can be used to improve processing efficiency. Furthermore, when cyclic dependencies exist in the dependency data (i.e., cycles exist in the directed graph), it is necessary to perform strongly connected component detection before expanding the transitive closure. Nodes within strongly connected components should be merged into a single supernode to ensure the directed acyclicity of the logical hierarchy topology graph, thereby ensuring the effectiveness of topological distance calculation and hierarchy partitioning.

[0075] In one optional implementation, a dynamic causal graph model is constructed based on the multi-dimensional features in the standardized project state representation. The causal transmission paths and constraints between these multi-dimensional features are then mined through reverse causal inference. A causal decoupling and tracing method is used to determine the driving factors influencing the progress evolution and their weight distribution, including:

[0076] The standardized project state representation is deconstructed into multiple semantic layers. Multi-dimensional features are expanded into a set of nodes and an initial directed graph structure is constructed. Reverse intervention signals are implanted and activated in each node of the initial directed graph structure. By tracking the propagation path and attenuation mode of the reverse intervention signals between nodes, the causal transmission path and transmission intensity threshold between multi-dimensional features are restored according to the propagation dynamics mode and condensed into constraints.

[0077] Based on the causal transmission path and the constraints, a causal decoupling analysis is performed on the feature nodes in the initial directed graph structure. By constructing a counterfactual scenario, the marginal contribution of each feature node to the progress evolution target under the assumed removal state is calculated. Feature nodes whose marginal contribution exceeds the significance discrimination criterion are locked into the driving factor set through significance separation.

[0078] An interaction coupling degree analysis is performed on each driving factor in the set of driving factors. Based on the interaction coupling degree, the marginal contribution of each driving factor is redundantly corrected. The corrected marginal contribution is then mapped to the weight distribution through global weight redistribution. The weight distribution is then injected into the corresponding node of the initial directed graph structure and drives the evolution to generate a dynamic causal graph model.

[0079] After obtaining a standardized project state representation, it is necessary to extract multi-dimensional features and construct a dynamic model that reflects the causal relationships between features. Multi-dimensional features typically cover several categories such as task completion rate, resource utilization rate, critical path float time, and external dependency satisfaction. These features are not independent of each other during project evolution, but rather influence each other through complex causal chains.

[0080] When performing multi-layer semantic deconstruction on the standardized project state representation, each dimension of the feature is mapped to an independent node in a graph structure. Each node carries the numerical vector of that feature within the current time window and its historical trend. Let the set of feature nodes be... ,in This represents the total number of features involved in the modeling. Initial directed graph structure. The set of directed edges in Initialization is performed based on prior knowledge of the domain and historical correlation analysis. At this point, the direction of the edges only represents the direction of influence of the preliminary hypothesis and has not yet been verified causally.

[0081] Implanting reverse intervention signals is the core operation for uncovering the true causal transmission path. Each node Apply reverse intervention signals sequentially This signal simulates the effect on the node. The forced perturbation is set with the amplitude standardized based on the historical fluctuation range of this feature. After activating the reverse intervention signal, the propagation process of the signal along each directed edge in the graph structure is tracked, and the arrival of the signal at each adjacent node is recorded. Response strength at time and transmission delay If the signal comes from After multiple hops of propagation, the origin is in The intensity of the response triggered at that location exceeds the preset propagation sensing threshold. Then it is believed right There is an effective causal transmission path.

[0082] By summarizing the propagation dynamics patterns through reverse intervention experiments at all nodes, the attenuation patterns of the signal along different paths can be identified. The differences in attenuation patterns reveal the directional and intensity differences in causal propagation: if the signal travels from... arrive The attenuation rate is significantly lower than that from arrive The attenuation rate then confirms To determine the dominant causal direction, all verified causal transmission paths are summarized, while those with transmission strength below a threshold are excluded. pseudo-correlated edges from After removing outdated elements, a refined set of causal transmission paths is formed, and the upper and lower bounds of the transmission strength corresponding to each path are condensed into a set of constraints. It is used to constrain the rationality of causal relationships in subsequent deduction processes.

[0083] After obtaining the causal transmission path and constraints, the node set... Causal decoupling analysis is performed on the feature nodes in the model. The core idea of ​​constructing a counterfactual scenario is: assuming that a certain feature node... Remove from the graph structure and calculate the progress evolution objective function under this assumption. The change in quantity, thus quantified. Marginal contribution to the overall progress evolution . Specifically, ,in This represents the objective function value for progress evolution under a complete set of nodes. Indicates the removal of a node. The corresponding value after that. Constructing a counterfactual scenario requires maintaining the structural integrity of the causal transmission path, that is, removing... At the same time, the incoming and outgoing edges in the causal graph are merged according to the transmission strength weighting to avoid estimation bias caused by graph structure breakage due to the removal of a single node.

[0084] For all After each node completes the counterfactual calculation, the marginal contribution vector is obtained. Let the significance criterion be... ,in for The mean, Standard deviation This is the significance factor (typically between 1.5 and 2.0, which can be adaptively adjusted based on project size and complexity). Marginal contribution. Nodes exceeding this discrimination criterion are identified as candidate driving factors with significant causal influence, and are further narrowed down into the set of driving factors through a significance separation operation. Significance separation not only identifies high-contribution nodes, but also simultaneously verifies whether the contributions of these nodes remain stable across different time windows, eliminating occasional high-contribution nodes caused by data noise.

[0085] set of driving factors There is an interactive coupling effect among the driving factors, meaning that the combined effect of two driving factors existing simultaneously is not equal to the sum of their independent effects. Any two driving factors and Calculate its interaction coupling degree , defined as the difference between the joint marginal contribution and the sum of the independent marginal contributions: .when When, it indicates that there is positive cooperative coupling between the two; when This indicates that there is negative redundant coupling between the two. The interaction coupling degree is calculated for all driving factor pairs, and a coupling degree matrix is ​​constructed. .

[0086] Based on the coupling degree matrix Redundancy correction is applied to the marginal contributions of each driving factor. For driving factor pairs with strong negative redundant coupling, the one with the relatively lower contribution needs to be adjusted according to the redundancy ratio. The value is reduced, and the reduction factor is from and The ratio determines the corrected contribution vector. It can accurately reflect the independent explanatory power of each driving factor after removing redundancy. For driving factor pairs with positive synergistic coupling, their contribution is not reduced, but they need to be treated as coupled units in the subsequent weight allocation.

[0087] The corrected contribution vector Mapped to a weight distribution through global weight redistribution. Normalization ensures ,in For positively coupled driving factor pairs, their joint weights are distributed between them according to the coupling strength ratio, ensuring the interpretability and consistency of the weight distribution in terms of numerical values.

[0088] Weight distribution Inject initial directed graph structure The corresponding nodes are assigned to each driving factor node, allowing each node to carry its influence weight information in the progress evolution. The weights of non-driving factor nodes are set to a near-zero minimum value, but their topological position in the graph structure is retained to maintain the integrity of the causal transmission path. After weight injection and causal path refinement, the initial directed graph structure evolves into a dynamic causal graph model. ,in This is a refined set of directed edges that has been validated through reverse intervention. The dynamic causal graph model not only encodes the causal transmission relationships between feature nodes, but also quantifies the relative importance of each driving factor to the progress evolution through weight distribution, providing a structured causal reasoning basis for subsequent multi-scenario simulation and risk propagation assessment.

[0089] In one optional implementation, based on the driving factors and their weight distribution, combined with preset project constraints and resource availability, a multi-scenario simulation mechanism is used to generate multiple candidate schedule projection paths. A risk propagation assessment is then performed on each candidate schedule projection path, and the target projection path that meets the confidence requirements is selected, including:

[0090] The driving factors and their weight distributions are transformed into state evolution driving vectors through multi-scale feature decomposition and vectorization encoding. The state evolution driving vectors are coupled with preset project constraints to construct a constraint evolution space. In the constraint evolution space, the resource availability is decomposed into multiple perturbations through a multi-scenario simulation mechanism to generate multiple sets of scenario parameter configurations. Parallel simulation is performed on each scenario parameter configuration using the multi-scenario simulation mechanism to generate multiple candidate progress simulation paths and their state transition sequences.

[0091] The state transition sequence of each candidate progress projection path is deconstructed into state node topology and each node is assigned a probability value based on historical transition frequency. A risk propagation topology network is constructed based on the deconstructed state transition nodes and their probability assignments. External risk sources are constructed and activated at each state transition node. The risk sources are driven to cascade and spread along the state transition sequence through the probability assignments. The risk accumulation and path robustness of each candidate progress projection path are calculated.

[0092] The accumulated risk and the robustness of the path are fused together to form a dual evaluation index for confidence quantification. The confidence quantification result is compared with a preset confidence threshold to lock in the target deduction path that meets the confidence threshold requirement.

[0093] After constructing the dynamic causal graph model and determining the set of driving factors and their weight distribution, the causal analysis results need to be transformed into quantitative inputs for simulation. For each driving factor in the set, its feature representation at multiple time scales is extracted, including short-cycle fluctuation features, medium-cycle trend features, and long-cycle structural features. Each scale feature component is extracted separately using wavelet decomposition or multi-resolution filtering, and then weighted and fused with corresponding normalized weights to obtain the multi-scale feature vector of each driving factor. Finally, the multi-scale feature vectors of all driving factors are weighted and concatenated according to the weight coefficients in the weight distribution vector to form the state evolution driving vector. This vector comprehensively reflects the combined driving force of various driving factors on the evolution of project schedule at different time scales.

[0094] State evolution driving vector Coupled with preset project constraints, a constraint evolution space is constructed. The construction process of the constraint evolution space is as follows: Hard constraints in the project constraints (such as project deadline, total resource limit, incompressible nodes on the critical path, etc.) are mapped to the boundary hyperplane of the constraint evolution space; soft constraints (such as resource utilization preference intervals, limits on the number of parallel tasks, etc.) are mapped to a penalty potential field within the constraint evolution space. This ensures that the trajectory of the state evolution driving vector within the constraint evolution space is rigidly constrained by hard constraints and elastically guided by soft constraints. Resource availability is represented by a multi-dimensional time series matrix. This indicates that each row corresponds to a type of resource, each column corresponds to a time slot, and the matrix elements represent the proportion of that type of resource available within the corresponding time slot.

[0095] The core of the multi-scenario simulation mechanism lies in the multi-dimensional perturbation decomposition of resource availability, using the resource availability matrix. Based on this, independent perturbations are applied along the resource dimension and the time dimension, respectively. The perturbation amplitude is determined by the perturbation intensity parameter. The control measures categorize disturbances into three types: uniform disturbances, normal disturbances, and extreme scenario disturbances. Uniform disturbances are used to simulate scenarios with slight resource fluctuations, normal disturbances to simulate scenarios with random changes in resource supply and demand, and extreme scenario disturbances to simulate scenarios with sudden resource interruptions or significant reductions. For each disturbance type, several sets of disturbance samples are generated. These samples are then superimposed on the original resource availability matrix and pruned according to boundary conditions in the constraint evolution space to obtain multiple sets of scenario parameter configurations. ,in This represents the total number of scenarios.

[0096] Configure parameters for each scenario ( ), in constrained evolution space Internal state evolution driving vector Parallel simulation is performed for the simulation engine. During the simulation, the current state of the project is used as the initial state, and the simulation is performed step by step according to the time step. At each step, the state transition direction is determined based on the available resources and driving vector under the current scenario parameter configuration, and candidate progress simulation paths are generated for that scenario. and its corresponding state transition sequence ,in Representing a path In the The state nodes at each time step For path The total number of steps. After multiple candidate progress simulation paths are generated in parallel, the risk propagation assessment phase begins.

[0097] State transition sequence for each candidate progress derivation path Perform state node topology deconstruction. Deconstruct each state node in the sequence. Treating a node as a node in a risk propagation topology network, and constructing paths using state transition relationships as directed edges. Corresponding risk propagation topology network Based on the frequency of transitions between nodes of the same state in historical project data, a transition probability is assigned to each directed edge. This probability reflects the historical state from which the state was determined. Transfer to The actual frequency of occurrence. If there are insufficient corresponding transition samples in the historical data, a Bayesian smoothing method is used to supplement the estimation with a prior uniform distribution to ensure that the sum of the outgoing edge probabilities of each node is normalized.

[0098] External risk sources are constructed at each state transition node, and the initial risk intensity of the external risk sources is determined. The risk level is determined based on the frequency of abnormal deviations observed at the node in historical data and the risk sensitivity of the node's current stage. After activating an external risk source, the risk propagates in a cascading manner along the state transition sequence: at the node... Location, risk level With transition probability Propagate to subsequent nodes, while considering the risk decay coefficient. (The range of values ​​is) This causes the risk to gradually decrease as the path length increases during propagation. The cumulative risk value at each node is the sum of the risk intensities propagated from all predecessor nodes to that node and the sum of the node's own initial risk intensity. Path Accumulated risk It is defined as the weighted sum of the cumulative risk values ​​of all nodes on the path. The weight is the temporal position weight of each node in the path. The closer the node is to the end of the path, the higher the weight, so as to highlight the impact of later risks on the overall progress.

[0099] Path robustness This reflects the ability of candidate progress projection paths to maintain state transition consistency under various scenario perturbations, through path calculation. The overlap rate between the path and the state of all other candidate paths at key milestone nodes is used as the metric. Specifically, the state vectors of all candidate paths at preset key milestone time points are extracted, and the path is calculated. The mean cosine similarity with other paths at these time points indicates that the path has stronger consistency under different scenarios and higher robustness.

[0100] Accumulated risk and path robustness A dual evaluation index is formed through fusion calculation, and confidence is quantified. Confidence score. The calculation method is as follows: for path robustness Perform positive contribution calculations on the cumulative risk amount. After normalization, the negative value is used as a negative penalty term, which is then applied through the fusion weights. (Used to control the balance between robustness and risk, the value range is...) The two are linearly fused to obtain a comprehensive confidence score. ,in and These are the path robustness and risk accumulation values ​​after max-min normalization, respectively.

[0101] Calculate the confidence score for each candidate path. Compared with the preset confidence threshold Compare and filter out all that meet the requirements. Candidate paths are selected. If the number of paths meeting the criteria is greater than 1, they are further sorted in descending order of confidence score, and the path with the highest score is selected as the target derivation path output; if no candidate path meets the confidence threshold requirement, the perturbation intensity parameter is adaptively reduced. The scenario parameter configuration is then regenerated and simulation is performed until a target simulation path that meets the requirements is generated. The final target simulation path will serve as the input for subsequent phased execution strategy deconstruction, supporting the iterative optimization and dynamic adjustment of the simulation model.

[0102] In one optional implementation, a multi-scenario simulation mechanism is used to perform multi-dimensional perturbation decomposition of resource availability within the constrained evolution space to generate multiple sets of scenario parameter configurations. The multi-scenario simulation mechanism is then used to perform parallel simulation for each scenario parameter configuration, generating multiple candidate progress derivation paths and their state transition sequences, including:

[0103] Based on the resource availability boundary in the constrained evolution space, a multi-dimensional perturbation matrix is ​​constructed through a multi-scenario simulation mechanism. Different perturbation intensities and directions are set in each dimension of the multi-dimensional perturbation matrix to perform hierarchical perturbation decomposition of resource availability and generate multiple sets of scenario parameter configurations.

[0104] Each scenario parameter configuration is injected into an independent simulation thread of the multi-scenario simulation mechanism. Within each independent simulation thread, a time-series deduction calculation is performed based on the state evolution driving vector and the corresponding scenario parameter configuration. The dynamic evolution process of the project state in each simulation thread is tracked, the state nodes and the transition relationships between nodes in each evolution process are captured, and candidate progress deduction paths and their state transition sequences corresponding to each scenario parameter configuration are generated.

[0105] like Figure 2 As shown, the method includes:

[0106] After calculating the weight distribution of driving factors and determining the constraint evolution space, a systematic multi-dimensional perturbation decomposition of resource availability is needed to cover various uncertainties that may arise during actual project execution. Resource availability in project simulation is not a single scalar quantity, but a vectorized description composed of multiple dimensions, including human resources, equipment resources, material resources, and external collaboration resources. To address this characteristic, a multi-dimensional perturbation matrix is ​​constructed. The matrix rows correspond to resource dimensions, the columns correspond to disturbance scenario numbers, and the matrix elements... Indicates the first The resource dimension in the first The disturbance values ​​under each scenario.

[0107] The perturbation matrix is ​​constructed following the principle of hierarchical perturbation decomposition. For the constrained evolution space... Given known resource availability boundaries, the availability intervals of each resource dimension are divided into several disturbance levels, each corresponding to a different disturbance intensity level. The disturbance intensity levels are set based on the statistical distribution of resource fluctuations in historical project data, typically using three basic levels: low, medium, and high disturbance. Positive and negative disturbances are then superimposed on these levels, forming six basic disturbance combinations. In the multi-dimensional resource space, disturbances between different dimensions can be configured independently, or coupled disturbances can be introduced to simulate scenarios of simultaneous tension or relaxation of multiple resource types. Let the total number of resource dimensions be... Then the multidimensional perturbation matrix The scale is ,in The total number of scenarios is represented by the column vectors, which together form a complete scenario parameter configuration.

[0108] Regarding the setting of disturbance direction, a positive disturbance indicates resource availability exceeding baseline expectations, such as early arrival of manpower or increased equipment utilization; a negative disturbance indicates resource availability falling below baseline expectations, such as increased staff absenteeism or increased equipment failure rate. Disturbance strength parameters. Controlling the amplitude range of disturbances in each dimension, specifically, the first... Each resource dimension in the context The disturbance quantity below satisfy ,in For the first The upper bound of the perturbation strength specific to each dimension is determined by the resource availability matrix. The standard deviation of the historical fluctuation in the corresponding dimension is multiplied by a preset perturbation amplification factor. This dimension-adaptive perturbation intensity setting method can avoid applying excessive perturbation to resource dimensions with relatively small fluctuations, thereby ensuring the physical rationality of scenario configuration.

[0109] After generating multiple sets of scenario parameter configurations, each configuration is injected into an independent simulation thread within the multi-scenario simulation mechanism. Each independent simulation thread corresponds to a set of scenario parameter configurations, and the threads are completely isolated and do not interfere with each other. Within each independent simulation thread, a state evolution driving vector is used. Based on the driving force, the resource disturbances in the current scenario parameter configuration are superimposed onto the resource availability matrix. At the corresponding time step, a corrected resource availability sequence for this scenario is formed. subscript This indicates the scenario number. The revised resource availability sequence will directly affect the feasible domain of state transitions at each time step. That is, in resource-constrained time steps, some state transition paths will be marked as unreachable due to insufficient resources, thereby guiding the simulation thread to advance along different evolution trajectories.

[0110] Time series extrapolation calculations are performed according to time steps. The incremental order unfolds step by step, and within each time step, based on the dynamic causal graph model... The causal transmission path in the data, combined with the corrected resource availability at the current time step. The corresponding column vectors are used to calculate the transition probabilities of each reachable successor state starting from the current state node. The calculation of transition probabilities comprehensively considers the weight distribution of driving factors. Resource availability constraints and a set of preset project constraints A positive probability is assigned to the successor state that satisfies the constraint, and a zero probability is assigned to the successor state that violates the constraint. All positive probabilities are then normalized.

[0111] When tracking the dynamic evolution of the project state in each simulation thread, the state node at each time step is... The system records complete information such as resource consumption, constraint satisfaction, and risk triggering events to form an evolution log for this scenario. The sequence of state nodes in the evolution log constitutes the backbone structure of the candidate progress projection path, and the transition relationships between adjacent state nodes are recorded as state transition sequences. Directed edges in a graph, each edge carrying a transition probability. Attribute information such as resource consumption vector and constraint test results.

[0112] To address the state bifurcation phenomenon encountered by simulation threads during the simulation process—that is, when multiple reachable successor states exist at a given time step—a Monte Carlo sampling mechanism is used to randomly select a successor state from the normalized transition probability distribution as the actual transition target for the thread, ensuring the uniqueness and random diversity of each candidate path. When simulation threads under different scenario parameter configurations select different successor states at the same time step, the corresponding candidate paths will bifurcate at that time step, thus enabling… The candidate paths cover a relatively broad evolutionary space as a whole.

[0113] There are two types of termination conditions for simulation threads: one is when the number of simulation time steps reaches the total number of path steps. The simulation thread has two parameters: a preset upper limit and a final state. When the simulation thread terminates normally, the corresponding candidate path is marked as a complete path. If the simulation thread has not reached any final state by the time step limit, the corresponding candidate path is marked as an overdue path, and the deviation between its final state and the expected final state is recorded for use in the subsequent risk propagation assessment phase.

[0114] all After each independent simulation thread completes its simulation, the candidate progress paths and their state transition sequences output by each thread are aggregated to form a candidate path set. The candidate path set undergoes preliminary deduplication, merging duplicate paths with identical state nodes at all time steps into a single path, and accumulating its frequency as an auxiliary reference indicator for path prior confidence. This deduplicated candidate path set will then serve as input for subsequent risk propagation assessment, entering the confidence scoring and target simulation path selection stages.

[0115] In actual engineering projects, the number of resource dimensions Typically between 5 and 20, total number of scenarios The settings need to strike a balance between computational resource consumption and scenario coverage. When When the disturbance matrix is ​​large, orthogonal experimental design methods can be used to test it. The column vectors are sparsely sampled to ensure that the perturbation effects in each dimension are fully explored. The scale is kept within an acceptable range. The scenario parameter configuration after orthogonal sampling has a uniform distribution characteristic in the multidimensional perturbation space, which can achieve efficient coverage of the resource uncertainty space with a small number of scenarios, thereby improving the representativeness of the candidate path set to the real project evolution trajectory.

[0116] In one optional implementation, collecting the state deviation vector between the actual progress status and the target deduced path at each execution stage, and performing local reconstruction of the causal transmission path in the dynamic causal graph model based on the state deviation vector includes:

[0117] In each execution phase of the project, a status acquisition node is set up. At each status acquisition node, the actual progress status at the current moment and the expected progress status of the target projection path at the corresponding moment are acquired synchronously. The difference between the actual progress status and the expected progress status is quantitatively calculated in multiple status dimensions to generate a status deviation vector that represents the degree of deviation in each status dimension. The status deviation vector is then classified by deviation magnitude and labeled with deviation direction.

[0118] The state deviation vector is projected onto the causal node space of the dynamic causal graph model. Sensitive causal nodes affected by deviations in the dynamic causal graph model are identified hierarchically based on the deviation magnitude of the state deviation vector. The causal transmission intensity of the sensitive causal nodes is adjusted. The causal transmission path between the sensitive causal nodes is topologically reconstructed according to the deviation direction of the state deviation vector, and a locally reconstructed dynamic causal graph model is generated.

[0119] During project execution, the start and end times of each execution phase are pre-marked as status acquisition nodes. These status acquisition nodes are evenly distributed along the timeline or according to key milestones to ensure continuous monitoring and coverage of project progress. At each status acquisition node, two types of data are extracted simultaneously: one is the actual progress status aggregated from on-site feedback channels, engineering logs, and sensor data streams; the other is the expected progress status at the corresponding moment read from the target projection path. The actual progress status includes multiple status dimensions such as task completion rate, resource consumption, and process connection status, while the expected progress status also has corresponding baseline values ​​in the same dimensions.

[0120] The actual progress status and the expected progress status are quantified dimension by dimension, assuming the total number of status dimensions is . , No. The actual observed values ​​for each state dimension are The corresponding expected benchmark value is Then the first The deviation components in each dimension are , by all The deviation components in each dimension constitute the state deviation vector. This vector fully represents the degree and direction of deviation from the expected state in each dimension during the current execution phase.

[0121] For the state deviation vector Deviation magnitude is classified and deviation direction is identified. Deviation magnitude classification is based on the absolute value of the deviation components in each dimension. Compare with a pre-defined hierarchical threshold sequence. Let the hierarchical threshold sequence be... ,when When the deviation in this dimension is slight, it is labeled as "slight deviation"; when When, it is marked as "moderate deviation"; when When, it is marked as "significant deviation"; when When this occurs, it is marked as "serious deviation". The direction of deviation is determined by the sign of the deviation component: This indicates that the actual situation is ahead of expectations (positive deviation). This indicates that the actual state lags behind the expectation (negative deviation). This indicates no deviation. After completing the classification and direction labeling, the state deviation vector... It carries complete amplitude level labels and direction symbols, providing a directional basis for subsequent local reconstruction of the causal graph.

[0122] The state deviation vector Projecting onto the causal node space of the dynamic causal graph model is the core operation for achieving local reconstruction. Each causal node in the dynamic causal graph model corresponds to a linear or nonlinear combination of state dimensions, which are mapped using a pre-constructed projection mapping matrix. (in (total number of causal nodes), Mapping to the causal node space yields the bias response vector at each causal node. , its first Each component Indicates the first The intensity of the biased excitation received by each causal node.

[0123] Based on deviation response vector Based on the deviation magnitude classification results, sensitive causal nodes in the dynamic causal graph model are identified. The absolute value of the deviation response exceeding the sensitivity identification threshold is considered. The nodes are marked as sensitive causal nodes, forming a set of sensitive nodes. For nodes classified as "slightly biased," even if the absolute value of their bias response reaches the threshold, they are only included as candidate sensitive nodes without immediately triggering reconstruction, in order to avoid unnecessary model modifications caused by noise disturbances; for nodes classified as "moderately biased" or higher, they are directly included. This will trigger subsequent adjustment procedures.

[0124] Adjust the causal transmission strength of sensitive causal nodes, assuming As a sensitive causal node, To and There exist adjacent nodes with directly directed edges, and the original causal propagation strength is denoted as . The deviation direction indicator determines the adjustment direction: when the deviation direction is positive (actually ahead), from... point to The conductivity is increased proportionally, with an adjustment amount of . ,in The step size coefficient is adjusted to adjust the conduction strength; when the deviation direction is negative (actual lag), the conduction strength is weakened by the same proportion, i.e. The adjusted conductivity is And truncate the results to ensure Keep within the preset reasonable range Internally, to prevent physically unreasonable extreme values ​​of conduction intensity.

[0125] After adjusting the conduction strength, topological reconstruction is performed on the causal conduction paths between sensitive causal nodes based on the deviation direction identifier. Topological reconstruction includes three types of operations: edge strengthening, edge weakening, and edge addition / removal. For node pairs with deviation amplitudes reaching "significant deviation" or "severe deviation" and consistent deviation directions... If the original directed graph does not contain a source... arrive The directed edge exists, but bias response analysis indicates a statistically significant causal relationship between the two. Therefore, this directed edge is added to the locally reconstructed dynamic causal graph model, with an initial transmission strength set to... For node pairs in the original directed graph that have directed edges but whose deviation analysis after multiple consecutive execution stages shows that the edges do not significantly contribute to the deviation evolution, the directed edges are removed from the locally reconstructed graph structure to simplify the causal structure and reduce model redundancy.

[0126] The topology reconstruction process also introduces a local consistency verification mechanism to ensure that the reconstructed dynamic causal graph model does not generate new directed loops. After each edge addition or deletion operation is completed, a directed acyclic graph (DAG) validity check is performed on the involved node subgraphs: the connected subgraphs containing the start and end points of the newly added edge are traversed in a depth-first search manner. If a loop is detected, the edge addition operation is canceled, and the node pair is marked as a "topology conflict node pair" and recorded in the conflict log for subsequent manual review.

[0127] After adjusting the conduction strength and reconstructing the topology, a locally reconstructed dynamic causal graph model is output. This model retains the global structure of the original dynamic causal graph model that is unaffected by biases, and only modifies the set of sensitive nodes. The edge set within the model and its neighborhood was locally modified, thereby integrating the actual progress deviation information of the current execution stage into the structural representation of the causal graph while maintaining the overall stability of the model. The locally reconstructed dynamic causal graph model will serve as the input for the next round of gradient backpropagation weight adjustment, driving the inference model to further iterate and converge in subsequent execution stages, so that the prediction accuracy of the target inference path continues to improve as execution progresses.

[0128] In practical engineering scenarios, the density of status acquisition nodes can be dynamically adjusted based on project complexity and the number of tasks on the critical path. For task nodes on high-risk critical paths, the status acquisition frequency can be increased to once per working day; for auxiliary tasks on non-critical paths, the acquisition frequency can be appropriately reduced to balance monitoring accuracy and data acquisition costs. Deviation amplitude grading threshold sequence. and sensitivity identification threshold All of them can be adaptively calibrated based on project type and historical data statistics to ensure that the grading results are consistent with actual engineering experience, thereby making the local reconstruction operation sufficiently robust and applicable to engineering.

[0129] A second aspect of this invention provides a multi-source data-driven intelligent project progress prediction system, comprising:

[0130] The data fusion unit is used to acquire progress data, resource data and dependency data associated with the target project from multiple heterogeneous data sources, and perform multi-scale coupling mapping of the progress data, resource data and dependency data in time dimension and logical dimension based on time alignment rules to generate a standardized project status representation.

[0131] The causal analysis unit is used to construct a dynamic causal graph model based on the multi-dimensional features in the standardized project status representation, to mine the causal transmission paths and constraints between the multi-dimensional features through reverse causal inference, and to determine the driving factors and their weight distribution that affect the progress evolution by using the causal decoupling and tracing method.

[0132] The simulation and evaluation unit is used to generate multiple candidate progress simulation paths based on the driving factors and their weight distribution, combined with preset project constraints and resource availability, using a multi-scenario simulation mechanism, and to conduct risk propagation assessment on each candidate progress simulation path, and select the target simulation path that meets the confidence requirements.

[0133] The iterative optimization unit is used to deconstruct the target derivation path into a phased execution strategy, collect the state deviation vector between the actual progress state and the target derivation path at each execution stage, locally reconstruct the causal transmission path in the dynamic causal graph model based on the state deviation vector, and adjust the weight distribution of the driving factors through gradient backpropagation to drive the iterative convergence of the derivation model.

[0134] A third aspect of the present invention provides an electronic device, comprising:

[0135] processor;

[0136] Memory used to store processor-executable instructions;

[0137] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0138] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0139] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-source data-driven intelligent project progress prediction method, characterized in that, include: Acquire progress data, resource data, and dependency data associated with the target project from multiple heterogeneous data sources. Based on time-series alignment rules, perform multi-scale coupling mapping of the progress data, resource data, and dependency data in the time and logical dimensions to generate a standardized project status representation. A dynamic causal graph model is constructed based on the multi-dimensional features in the standardized project status representation. The causal transmission paths and constraints between the multi-dimensional features are explored through reverse causal inference. The causal decoupling and tracing method is used to determine the driving factors affecting the progress evolution and their weight distribution. Based on the driving factors and their weight distribution, combined with the preset project constraints and resource availability, a multi-scenario simulation mechanism is used to generate multiple candidate progress projection paths, and a risk propagation assessment is performed on each candidate progress projection path to select the target projection path that meets the confidence requirements. The target derivation path is deconstructed into a phased execution strategy. At each execution phase, a state deviation vector between the actual progress status and the target derivation path is collected. Based on this state deviation vector, the causal transmission path in the dynamic causal graph model is locally reconstructed, including: In each execution phase of the project, a status acquisition node is set up. At each status acquisition node, the actual progress status at the current moment and the expected progress status of the target projection path at the corresponding moment are acquired synchronously. The difference between the actual progress status and the expected progress status is quantitatively calculated in multiple status dimensions to generate a status deviation vector that represents the degree of deviation in each status dimension. The status deviation vector is then classified by deviation magnitude and labeled with deviation direction. The state deviation vector is projected onto the causal node space of the dynamic causal graph model. Sensitive causal nodes affected by deviations in the dynamic causal graph model are identified hierarchically based on the deviation magnitude of the state deviation vector. The causal transmission strength of the sensitive causal nodes is adjusted. The causal transmission path between the sensitive causal nodes is topologically reconstructed according to the deviation direction of the state deviation vector. A locally reconstructed dynamic causal graph model is generated. The weight distribution of the driving factors is adjusted through gradient backpropagation to drive the iterative convergence of the inference model.

2. The method according to claim 1, characterized in that, Based on time-series alignment rules, the progress data, resource data, and dependency data are subjected to multi-scale coupling mapping of time and logical dimensions to generate a standardized project status representation, including: The timestamp sequences of progress data, resource data, and dependency data are decomposed into multiple granularities. Time anchors are identified at each granularity level, and a cross-data source time alignment benchmark is constructed. Based on the time alignment benchmark, heterogeneous time scales are decoupled and reconstructed at multiple levels, forming a unified time series expression of each data source in a unified time coordinate system. Under the unified time coordinate system, a logical hierarchy topology graph is constructed based on the logical constraints in the dependency data. The progress data and the resource data are then projected in layers according to the node hierarchy of the logical hierarchy topology graph to generate state subspaces at each level. Multi-scale tensor reconstruction is performed on the state subspaces of each level. By calculating the coupling strength between different granularity levels and logical levels, the coupling strength is used as a modulation factor to perform hierarchical fusion and dimensional compression of each state subspace, forming a standardized project state representation.

3. The method according to claim 2, characterized in that, A logical hierarchy topology graph is constructed based on the logical constraints in the dependency data. The progress data and resource data are then projected hierarchically according to the node hierarchy of the logical hierarchy topology graph to generate state subspaces for each level, including: Transitive closure expansion is performed on the logical constraint relationships in the dependency data to identify the direct and indirect dependency paths between data entities. A directed dependency graph is constructed based on the direct and indirect dependency paths. The directed dependency graph is then hierarchically analyzed and the topological distance of each node is calculated. Based on the topological distance, the nodes are divided into different logical levels to construct a logical hierarchical topology graph. Based on the hierarchical affiliation and semantic identifier of each node in the logical hierarchical topology diagram, multi-dimensional feature matching is performed on the progress data and resource data. The matched progress data and resource data are semantically projected and spatially decomposed along the node hierarchy of the logical hierarchical topology diagram. Within each level, the projected data is reconstructed into state vectors to generate a state subspace with hierarchical isolation.

4. The method according to claim 1, characterized in that, Based on the multi-dimensional features in the standardized project status representation, a dynamic causal graph model is constructed. Through reverse causal inference, the causal transmission paths and constraints between these multi-dimensional features are explored. A causal decoupling and source tracing method is used to determine the driving factors influencing schedule evolution and their weight distribution, including: The standardized project state representation is deconstructed into multiple semantic layers. Multi-dimensional features are expanded into a set of nodes and an initial directed graph structure is constructed. Reverse intervention signals are implanted and activated in each node of the initial directed graph structure. By tracking the propagation path and attenuation mode of the reverse intervention signals between nodes, the causal transmission path and transmission intensity threshold between multi-dimensional features are restored according to the propagation dynamics mode and condensed into constraints. Based on the causal transmission path and the constraints, a causal decoupling analysis is performed on the feature nodes in the initial directed graph structure. By constructing a counterfactual scenario, the marginal contribution of each feature node to the progress evolution target under the assumed removal state is calculated. Feature nodes whose marginal contribution exceeds the significance discrimination criterion are locked into the driving factor set through significance separation. An interaction coupling degree analysis is performed on each driving factor in the set of driving factors. Based on the interaction coupling degree, the marginal contribution of each driving factor is redundantly corrected. The corrected marginal contribution is then mapped to the weight distribution through global weight redistribution. The weight distribution is then injected into the corresponding node of the initial directed graph structure and drives the evolution to generate a dynamic causal graph model.

5. The method according to claim 1, characterized in that, Based on the aforementioned driving factors and their weight distributions, and combined with preset project constraints and resource availability, a multi-scenario simulation mechanism is used to generate multiple candidate schedule projection paths. A risk propagation assessment is then performed on each candidate schedule projection path, and the target projection paths that meet the confidence requirements are selected, including: The driving factors and their weight distributions are transformed into state evolution driving vectors through multi-scale feature decomposition and vectorization encoding. The state evolution driving vectors are coupled with preset project constraints to construct a constraint evolution space. In the constraint evolution space, the resource availability is decomposed into multiple perturbations through a multi-scenario simulation mechanism to generate multiple sets of scenario parameter configurations. Parallel simulation is performed on each scenario parameter configuration using the multi-scenario simulation mechanism to generate multiple candidate progress simulation paths and their state transition sequences. The state transition sequence of each candidate progress projection path is deconstructed into state node topology and each node is assigned a probability value based on historical transition frequency. A risk propagation topology network is constructed based on the deconstructed state transition nodes and their probability assignments. External risk sources are constructed and activated at each state transition node. The risk sources are driven to cascade and spread along the state transition sequence through the probability assignments. The risk accumulation and path robustness of each candidate progress projection path are calculated. The accumulated risk and the robustness of the path are fused together to form a dual evaluation index for confidence quantification. The confidence quantification result is compared with a preset confidence threshold to lock in the target deduction path that meets the confidence threshold requirement.

6. The method according to claim 5, characterized in that, A multi-scenario simulation mechanism is used to decompose resource availability into multiple sets of scenario parameter configurations within the constrained evolution space, generating a multi-dimensional perturbation decomposition. Parallel simulation is then performed on each scenario parameter configuration using this mechanism to generate multiple candidate progress paths and their state transition sequences, including: Based on the resource availability boundary in the constrained evolution space, a multi-dimensional perturbation matrix is ​​constructed through a multi-scenario simulation mechanism. Different perturbation intensities and directions are set in each dimension of the multi-dimensional perturbation matrix to perform hierarchical perturbation decomposition of resource availability and generate multiple sets of scenario parameter configurations. Each scenario parameter configuration is injected into an independent simulation thread of the multi-scenario simulation mechanism. Within each independent simulation thread, a time-series deduction calculation is performed based on the state evolution driving vector and the corresponding scenario parameter configuration. The dynamic evolution process of the project state in each simulation thread is tracked, the state nodes and the transition relationships between nodes in each evolution process are captured, and candidate progress deduction paths and their state transition sequences corresponding to each scenario parameter configuration are generated.

7. A multi-source data-driven intelligent project progress simulation system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The data fusion unit is used to acquire progress data, resource data and dependency data associated with the target project from multiple heterogeneous data sources, and perform multi-scale coupling mapping of the progress data, resource data and dependency data in time dimension and logical dimension based on time alignment rules to generate a standardized project status representation. The causal analysis unit is used to construct a dynamic causal graph model based on the multi-dimensional features in the standardized project status representation, to mine the causal transmission paths and constraints between the multi-dimensional features through reverse causal inference, and to determine the driving factors and their weight distribution that affect the progress evolution by using the causal decoupling and tracing method. The simulation and evaluation unit is used to generate multiple candidate progress simulation paths based on the driving factors and their weight distribution, combined with preset project constraints and resource availability, using a multi-scenario simulation mechanism, and to conduct risk propagation assessment on each candidate progress simulation path, and select the target simulation path that meets the confidence requirements. The iterative optimization unit is used to deconstruct the target derivation path into a phased execution strategy, collect the state deviation vector between the actual progress state and the target derivation path at each execution stage, locally reconstruct the causal transmission path in the dynamic causal graph model based on the state deviation vector, and adjust the weight distribution of the driving factors through gradient backpropagation to drive the iterative convergence of the derivation model.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images