Informatization project management system based on big data analysis

Abstract

The invention discloses an informatization project management system based on big data analysis, which belongs to the field of big data and comprises a data acquisition module, a time sequence modeling module, a task coupling analysis module, a risk clustering identification module, a resource allocation prediction module and the like. The data acquisition module asynchronously and parallelly acquires structured and unstructured data and uniformly encodes the structured and unstructured data; the time sequence modeling module constructs a multi-dimensional time sequence based on an autoregressive residual network; the task coupling analysis module fuses the task trajectory and the dependency relationship to generate a task influence directed graph; the risk clustering identification module identifies risks through variational graph auto-encoder mapping; the resource allocation prediction module constructs a dynamic resource priority based on a graph attention mechanism; a progress deviation traceability module identifies a deviation causal chain; the knowledge graph decision-making module corrects resource priorities and path strategies in a cross-graph manner; and the project global control module dynamically adjusts key paths and resource configuration and performs closed-loop self-correction. The beneficial effect is that the intelligent level and the risk response capability of project management are improved.

Description

Technical Field

[0001] The present invention relates to the field of big data, and more specifically, to an information-based project management system based on big data analysis. Background Art

[0002] With the continuous expansion of the scale of information-based projects, project management has gradually developed from traditional manual planning and empirical scheduling to digital management based on information systems. Early project management systems mainly focused on static task arrangement and manual resource allocation. Although the visualization level has been improved, they are still insufficient in dealing with dynamic task changes, risk chain identification, and resource conflict prediction. In recent years, with the application of big data and artificial intelligence technologies, project management systems have begun to introduce data-driven auxiliary decision-making mechanisms to achieve statistical analysis of historical data and preliminary trend prediction, improving the accuracy and real-time performance of some decisions.

[0003] However, there are still several deficiencies in the existing technologies: First, the time series modeling ability is limited, making it difficult to accurately predict the task evolution trend; second, resource scheduling relies on static plans and lacks a mechanism for dealing with cross-cycle resource coupling and dynamic conflicts; third, the risk identification path is shallow, and it is impossible to form a propagation chain analysis only relying on experience or single deviation; fourth, there is a lack of a dynamic adjustment mechanism for the logical relationship between tasks, making it difficult to build a flexible response strategy. Therefore, there is an urgent need for an intelligent project management system that integrates deep time series modeling, risk path identification, resource prediction optimization, and graph structure control to improve the global coordination and dynamic response capabilities. Summary of the Invention

[0004] The purpose of the present invention is to provide an information-based project management system based on big data analysis to solve the problems raised in the above background art: First, the time series modeling ability is limited, making it difficult to accurately predict the task evolution trend; second, resource scheduling relies on static plans and lacks a mechanism for dealing with cross-cycle resource coupling and dynamic conflicts; third, the risk identification path is shallow, and it is impossible to form a propagation chain analysis only relying on experience or single deviation; fourth, there is a lack of a dynamic adjustment mechanism for the logical relationship between tasks, making it difficult to build a flexible response strategy.

[0005] Technical Solution: The information-based project management system based on big data analysis includes a data collection module, a time series data modeling module, a task coupling analysis module, a risk clustering and identification module, a resource allocation prediction module, a progress deviation tracing module, a knowledge graph decision-making module, and a project global control module; The data acquisition module parallelly acquires structured data and unstructured data based on asynchronous pipeline - type trigger logic, and encodes them through a unified nested vector standard; the time - series data modeling module constructs a multi - dimensional project time - series model based on an autoregressive residual network and a multi - scale sliding window mechanism; the task coupling and parsing module fuses the task execution trajectory output by the multi - dimensional project time - series model with the static dependency matrix in the project task library to form a high - dimensional interaction tensor structure, and performs tensor graph decomposition on the high - dimensional interaction tensor to generate a task influence directed graph; the risk clustering and identification module takes the task influence directed graph as the graph input, generates multi - dimensional clustering embeddings through a variational graph auto - encoder, and maps them to the clustering distribution space to identify potential risk points; the resource allocation prediction module nests a resource sensitivity network based on a graph attention mechanism in the clustering distribution space to construct a dynamic priority queue for resource allocation; the progress deviation tracing module maps the identified risk points and resource scheduling offsets based on a deviation propagation chain graph modeling strategy, and identifies the deviation causal chain through multi - hop path accumulation; the knowledge graph decision module takes the risk clustering output and the progress deviation chain as inputs, performs cross - graph edge - weight tensor reconstruction in a unified graph semantic layer, and corrects the resource allocation priority and path selection strategy through a bilateral attention mechanism; the project global control module receives the correction result in real - time, dynamically rewrites the critical path, non - critical path, and resource arrangement table using a hierarchical boundary value strategy, and constructs an information - based project management system with closed - loop self - correction based on big data analysis.

[0006] Preferably, the data acquisition module includes a structured data acquisition unit, an unstructured text transcoding unit, an unstructured speech transcribing unit, and a data normalization sub - module; The structured data acquisition unit uses a time - tag partitioning mechanism based on Kafka streams to capture log data in ERP, PMIS, and WBS systems in the form of event blocks, and generates a multi - level timestamp index; the unstructured text transcoding unit extracts entity - pair relationships in project emails, meeting minutes, and contract documents through a sentence - word aggregation network based on the Transformer architecture, and uniformly maps them to a nested triple vector group; The unstructured speech transcribing unit uses a deep attention audio recognition network to transcribe engineering meeting recordings into a segment structure and segment them into task - related semantic segments; the data normalization sub - module performs unified time - domain filling and vector scale normalization on all acquired content, and compresses it to a fixed - dimension nested vector standard.

[0007] Preferably, the data normalization sub-module adopts a heterogeneous vector rescaling mechanism based on an improved BatchNorm strategy and combines a reversible linear normalization function based on entropy weights, so that the vector dimension distributions generated by each data source still retain the main component covariance structure unchanged after multi-source mixed input. In the compression process, a multi-head attention screening mechanism is used to only retain the vector components with the top 10% of the largest clustering entropy value contributions, and the compression accuracy is not less than 93%.

[0008] Preferably, the task coupling parsing module represents the coupling relationship between task entities based on a sixth-order coupling tensor structure. The six dimensions are task ID, estimated duration, resource type, historical deviation rate, external interference index, and responsible unit number. A tensor graph decomposition algorithm is used to decompose the potential dependency relationships between tasks into multiple observable directed graphs, and the main sub-graphs set dynamic edge weights according to the deviation propagation frequency.

[0009] Preferably, the tensor kernel in the tensor graph decomposition algorithm adopts a low-rank kernel tensor with a truncated rank r = 6. The embedding space constructs an edge weight perturbation adjustment mechanism based on the graph Laplace eigen-distribution, imposes a regularization constraint that maximizes the edge connection stability on the coupling tensor offset that appears in multi-period task scheduling, and uses residual gradient guidance to converge to a steady-state influence sub-graph in iterative optimization.

[0010] Preferably, the steady-state influence sub-graph extracts a set of core risk propagation paths through a centrality screening mechanism, maps the set of core risk propagation paths to the priority scheduling graph in the resource allocation prediction module, and constructs a weighted impedance matrix based on the mutual impedance score between tasks. The weighted impedance matrix guides the resource scheduling order to re-arrange priorities among similar tasks.

[0011] Preferably, the risk clustering and identification module adopts a graph-structured variational auto-encoder mechanism VGAE-GNN. The encoder consists of two layers of graph convolutional units, and a gated fusion unit based on the high-order reconstruction error distribution is appended after the node representation. The node vectors from the task influence sub-graph are mapped to a high-dimensional clustering embedding space, and the clustering output is guided by the K-L divergence to fall into the potential risk hot spot distribution area. The hot spot risk distribution forms a connected sub-graph through the DBSCAN density threshold clustering strategy, and each sub-graph matches the isomorphic risk paths in the historical deviation graph.

[0012] Preferably, the gated fusion unit introduces an edge weight masking mechanism based on the inversion of edge perturbation probability, performs reverse weight suppression on the instability of edge connections in the task influence graph, and adopts a maximum entropy compression strategy to reduce the distortion propagation probability from high-noise areas to the clustering space, so that the risk hot spot matching accuracy in the final clustering graph is more than 10% higher than that of existing risk identification algorithms.

[0013] Preferably, the knowledge graph decision module collaboratively executes by using a graph edge reconstruction mechanism and a policy edge weight adjustment mechanism. The graph edge reconstruction mechanism adopts an edge embedding update algorithm based on third-order path frequency regularization to reconstruct the connection weights of key nodes based on the progress deviation chain path. The policy edge weight adjustment mechanism uses a bilateral cross-attention mechanism to guide the fine-tuning of the edge weights of each policy edge in the project control graph, automatically reducing the transfer probability value of the edges related to the failure path and simultaneously increasing the alternative path in the risk surge state, so as to update the output path of the final project global control module from a static path graph to a dynamic path graph calculated based on the current situation.

[0014] Compared with the prior art, the advantages of the present invention are as follows: (1) Using neural networks such as LSTM to perform temporal modeling on task execution data, dynamically capturing the evolution trend of task states, and enhancing the prediction ability for future deviations.

[0015] (2) Integrating the attention mechanism and the resource coupling matrix to predict the resource usage amount and mediate conflicts across tasks and cycles, replacing the traditional static resource plan.

[0016] (3) Based on the task dependency graph and the deviation propagation influence index, automatically constructing a deviation propagation chain to trace the key tasks and the cascading risks they trigger, replacing manual tracking.

[0017] (4) Dynamically reconstructing the edge weights of the project critical path, reordering the priorities according to the task deviation history and risk frequency, and enhancing the scheduling flexibility and risk avoidance ability.

[0018] (5) Through dual-graph fusion and the A* path optimization algorithm, constructing an optimal scheduling plan that integrates the current task state and the historical execution success path to achieve global optimization control.

[0019] (6) Each functional module of the system can be independently deployed and docked with the existing PMIS through the API to achieve on-demand access and expansion, improving the system's implementability and compatibility.

[0020] (7) For the first time, the task deviation prediction result feedback is used for resource scheduling decision-making, forming a closed-loop control process of prediction-evaluation-optimization.

[0021] (8) The system supports complex structure modeling such as multi-level task dependencies, parallel / serial tasks, and loop tasks, enhancing the adaptability to large-scale projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the overall system schematic diagram of an information-based project management system based on big data analysis of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] For the embodiments, please refer toFigure 1 An information-based project management system based on big data analysis includes a data collection module, a time series data modeling module, a task coupling analysis module, a risk clustering and identification module, a resource allocation prediction module, a progress deviation tracing module, a knowledge graph decision-making module, and a project global control module; The data collection module parallelly collects structured data and unstructured data based on asynchronous pipeline trigger logic and encodes them through a unified nested vector standard; the time series data modeling module constructs a multi-dimensional project time series model based on an autoregressive residual network and a multi-scale sliding window mechanism; the task coupling analysis module fuses the task execution trajectories output by the multi-dimensional project time series model with the static dependency matrix in the project task library to form a high-dimensional interaction tensor structure, and performs tensor graph decomposition on the high-dimensional interaction tensor to generate a task influence directed graph; the risk clustering and identification module takes the task influence directed graph as the graph input, generates multi-dimensional clustering embeddings through a variational graph autoencoder, and maps them to the clustering distribution space to identify potential risk points; the resource allocation prediction module nests a resource sensitivity network based on a graph attention mechanism in the clustering distribution space to construct a dynamic priority queue for resource allocation; the progress deviation tracing module maps the identified risk points and resource scheduling offsets based on a deviation propagation chain graph modeling strategy, and identifies the deviation causal chain through multi-hop path accumulation; the knowledge graph decision-making module takes the risk clustering output and the progress deviation chain as inputs, performs cross-graph edge weight tensor reconstruction in the unified graph semantic layer, and corrects the resource allocation priority and path selection strategy through a bilateral attention mechanism; the project global control module receives the correction result in real time, dynamically rewrites the critical path, non-critical path, and resource arrangement table using a hierarchical boundary value strategy, and constructs a closed-loop self-correcting information-based project management system based on big data analysis.

[0024] Specifically, deploy a Kafka event listener in a microservice architecture, divide the operation logs from ERP, PMIS, and WBS into granularities such as "task creation", "task status change", and "resource adjustment" according to event types, bind the task ID and time tag, and push them into the stream processing queue. The system uses an asynchronous controller to concurrently manage data extraction tasks through a Redis distributed lock, and finally constructs a unified nested vector structure (length 512), where the structured field vector and the text encoding vector are spliced and synthesized. All modules are deployed using container orchestration (such as a Kubernetes cluster) for service-level elastic management.

[0025] Specifically, the time series data modeling module is used to model the dynamic characteristics during project execution, identify the time evolution law of task status, and construct the basis for task deviation prediction.

[0026] Specifically, the time series data modeling module includes the following steps: S101. Data collection and preprocessing: Obtain task execution records, resource usage trajectories, project plans, and actual execution data from the project management system to form an initial time series matrix , where N represents the number of tasks, T represents the time step, and F represents the feature dimension at each step (such as task progress, status, resource load, etc.); S102. Perform standardization processing on the data, such as using Z-score or Min-Max normalization methods, and at the same time remove outliers through the IQR rule or the sliding window method; S103. Construct a time series modeling unit based on the Long Short-Term Memory Network (LSTM). The LSTM network adopts a two-layer stacked structure, with a hidden layer dimension of 128, an activation function of Tanh, and a Dropout set to 0.3 to capture the deep dependencies of task states changing over time; S104. Assume that there are N tasks in the system, and the state of each task at time step t is composed of features in dimensions, and the features include but are not limited to the current progress ratio, remaining construction period, risk level, resource pressure factor, task priority index, etc. Record the task state vector at each time point output by the model for subsequent deviation identification and dynamic resource scheduling.

[0027] Specifically, the resource allocation prediction module is used to predict the resource requirements and configuration priorities in the future period based on task deviation identification, improving the timeliness and efficiency of resource scheduling.

[0028] Specifically, the resource allocation prediction module includes the following steps: S201. Receive the task state prediction vector output from the time series data modeling module and construct a resource requirement tensor in combination with the task structure diagram, where M represents the types of resources, including human resources, equipment, funds, etc.; S202. Use a multi-head neural network based on the attention mechanism to predict the future resource requirements of each task. The model structure includes 3 attention heads, and the dimensions of Key / Query / Value are all 64, and the weight adjustment is carried out by introducing the resource scarcity factor and the task urgency factor; S203. Establish a resource coupling matrix between tasks to form a resource conflict graph, and complete the preliminary optimization allocation through the greedy conflict solution method and the alternative resource mapping matrix; S204. Output resource scheduling suggestions, including task numbers, required resource numbers, estimated allocation time periods, and priority scores, and support docking with the scheduling management system in JSON format.

[0029] Specifically, in another preferred embodiment, the progress deviation traceability module is used to construct a deviation propagation path diagram from the identified task progress deviations and identify the key node tasks that cause global scheduling risks.

[0030] The specific steps are as follows: S301. Construct a task dependency graph and a time path graph, align the task dependencies with the time axis, and generate a "task-time" two-way coupling graph; S302. Use a sliding window mechanism to scan the sub-paths in the deviation set. If 3 or more consecutive nodes are delayed within the window period, mark them as suspicious propagation paths; S303. Calculate the deviation transfer influence of each node, defined as follows: ; where is a Boolean value indicating whether the successor task is delayed, is the path depth; S304. Construct multiple traceability paths, sort them according to the influence range, transfer speed, and intersection degree with the critical path, and preferentially output the propagation chain results.

[0031] Specifically, in another preferred embodiment, the project global control module is used to reconstruct the edge weights of the project critical path based on the output results of each module, form dynamic task priorities, and output an optimized scheduling path.

[0032] Specifically, the project global control module includes the following steps: S401. Receive three types of results from the tensor analysis module, the deviation traceability module, and the resource prediction module, including risk path nodes, impact task chains, and resource scheduling lists; S402. Construct a path frequency matrix and a risk edge weight mapping table, adjust the edge weights of tasks whose recent deviation frequencies exceed the threshold (such as three consecutive periods). The edge weight reconstruction formula is as follows: ; where is the memory factor, represents the deviation frequency of edge i→j; S403. Use a bilateral cross-attention mechanism to fuse and learn the current task graph and the historical successful graph, and dynamically generate a task scheduling priority sequence; S404. Use the A* search algorithm to find the shortest risk avoidance path in the reconstructed graph, and output a scheduling result with hop point identification, including fields such as task ID, path order, and priority identification, for direct invocation and execution by the project management platform.

[0033] The data acquisition module includes a structured data acquisition unit, an unstructured text transcoding unit, an unstructured speech transcribing unit, and a data normalization sub-module; The structured data acquisition unit adopts a time-tagging partition mechanism based on Kafka streams to capture the log data in ERP, PMIS, and WBS systems in the form of event blocks, and generates a multi-level timestamp index; the unstructured text transcoding unit extracts entity pair relationships in project emails, meeting minutes, and contract documents through a sentence and word aggregation network based on the Transformer architecture, and uniformly maps them to a nested triple vector group; The unstructured speech transcribing unit uses a deep attention audio recognition network to transcribe engineering meeting recordings into a segment structure and segment them into task-related semantic segments; the data normalization sub-module performs unified time domain padding and vector scale normalization on all acquired content, and compresses it to a fixed-dimensional nested vector standard.

[0034] Specifically, the Kafka log event block uses "task number + operation type" as the trigger granularity. After each log enters, an event chain is constructed with a five-level timestamp (system time, submission time, approval time, execution time, synchronization time). The unstructured text transcoding unit uses the RoBERTa-base model to fine-tune on project corpora, and extracts triples (entity A, relationship R, entity B) in combination with a regular expression-based trigger word library (such as "adjust resources", "change schedule"). The speech transcribing unit adopts a Conformer model + ASR-TaggingHead structure, and the speech slices are segmented at 15 seconds / frame. The transcribed text is timestamp-mapped to the task node timeline through a semantic alignment model. The data normalization adopts a vector structure with a unified length of 512 dimensions, of which 128 dimensions are structured nested fields and 384 dimensions are text representations.

[0035] The data normalization sub-module adopts a heterogeneous vector rescaling mechanism based on an improved BatchNorm strategy, and combines a reversible linear normalization function based on entropy weights, so that the vector dimension distributions generated by each data source still retain the main component covariance structure unchanged after multi-source mixed input. The compression process passes through a multi-head attention screening mechanism, and only retains the vector components with the top 10% largest clustering entropy value contributions, and the compression accuracy is not less than 93%.

[0036] Specifically, the improved BatchNorm strategy sets different mean-variance sliding windows for each type of data source (structured / text / speech), and introduces a cross-source balancing parameter α for vector recombination. The normalization function uses the entropy weight method to weight the information entropy of different dimensions to generate an invertible mapping matrix, ensuring that the main covariance pattern is still retained after compression. The multi-head attention mechanism uses 4-head Transformer attention to screen out high-contribution factors in the 128-dimensional clustering projection space, retaining the top 10% vector components, and the measured nested compression error remains below 7%.

[0037] The task coupling analysis module represents the coupling relationship between task entities based on a sixth-order coupling tensor structure. The six dimensions are task ID, estimated duration, resource type, historical deviation rate, external interference index, and responsible unit number. The tensor graph decomposition algorithm is used to decompose the potential dependencies between tasks into multiple observable directed graphs, and the main subgraph sets dynamic edge weights according to the deviation propagation frequency.

[0038] Specifically, the six-dimensional tensor uses a float32 data structure, and the tensor dimension range is: [task ID] × [1 - 100 days] × [resource type one-hot 32 dimensions] × [deviation rate 0 - 1] × [interference index normalized value 0 - 1] × [responsible person ID mapping vector 64 dimensions]. The tensor construction is based on the project historical task trajectory data to construct an interaction sequence. Use Tensorly (PyTorch) to implement HOSVD decomposition. Consider a task dependency as a "potential diffusion path" according to the frequency threshold > 3, and the dynamic edge weights are updated after each round of propagation.

[0039] The tensor core in the tensor graph decomposition algorithm uses a low-rank core tensor with a truncated rank r = 6. The embedding space constructs an edge weight perturbation adjustment mechanism based on the graph Laplace eigen-distribution, applies a regularization constraint that maximizes the edge connection stability to the coupling tensor offset that appears in the multi-period task scheduling, and uses the residual gradient to guide the convergence to a steady-state influence subgraph in the iterative optimization.

[0040] Specifically, the coupling tensor has a six-dimensional tensor structure, where the six dimensions respectively represent the task ID, time period, resource type, impact factor, interference index, and historical deviation level. The data types of each dimension are as follows: the task ID is an integer index, the resource type uses a one-hot encoded vector, the interference index is a normalized continuous value (range 0-1), and the historical deviation level is a categorical value (0-3 represents slight to severe). Tensor decomposition uses the HOSVD algorithm in the Tensorly (based on PyTorch) library for low-rank kernel extraction, and the truncation rank is set to r = 6. The task relationship graph is formed by constructing an association matrix for each dimension and then performing threshold pruning to form a graph structure. The graph Laplace eigenvectors are used to estimate local perturbations, and combined with residual guidance to form a stable embedding. The "potential diffusion paths" in the tensor are automatically marked when the path frequency is greater than a set threshold (such as 0.25) for input to the propagation model.

[0041] The steady-state influence subgraph extracts the set of core risk propagation paths through a centrality screening mechanism, maps the set of core risk propagation paths to the priority scheduling graph in the resource allocation prediction module, and constructs a weighted impedance matrix based on the mutual interference score between tasks. The weighted impedance matrix guides the reordering of priorities among similar tasks in the resource scheduling order.

[0042] Specifically, the steady-state influence subgraph uses a combination of betweenness centrality and PageRank for path screening. After calculating the centrality scores in the graph, the nodes are sorted and the top 20% of the nodes are selected to form the set of core risk propagation paths. The mutual interference score matrix is calculated jointly using the historical conflict frequency and resource competition degree between tasks, and the numerical range is normalized to [0,1]. The weighted impedance matrix is based on the mutual interference, and models the penalty for the scheduling order between nodes (such as if the mutual interference > 0.7, then it is preferentially scheduled later), guiding the reordering logic of similar tasks in the resource allocation prediction module to generate a dynamic priority graph.

[0043] The risk clustering and identification module uses the graph-structured variational autoencoder mechanism VGAE-GNN. The encoder consists of two layers of graph convolutional units, and a gated fusion unit based on the high-order reconstruction error distribution is appended after the node representation. The node vectors from the task influence subgraph are mapped to a high-dimensional clustering embedding space, and the clustering output is guided by the K-L divergence to fall into the potential risk hot spot distribution area. The hot spot risk distribution forms a connected subgraph through the DBSCAN density threshold clustering strategy, and each subgraph is matched with the isomorphic risk paths in the historical deviation graph.

[0044] Specifically, in the VGAE-GNN structure, the encoder consists of two layers of GCN (Graph Convolutional Neural Network) units. The output dimension of the first layer is 64, and the second layer is 32. The ReLU activation function is used. The node vectors are passed into the DBSCAN clustering module after gated fusion. In DBSCAN, the clustering radius eps = 0.5 and the minimum number of samples minPts = 5 are set. The embedding space adjusts the gated output with the high-order reconstruction error as the weight, and the Kullback-Leibler divergence is used to optimize the difference between the latent variable distribution and the risk hot spot distribution, so that the final clustering result falls into the high-risk density area. The similarity between each clustering connected subgraph and the risk path structure recorded in the historical graph is calculated (if the structural isomorphism rate is greater than 0.8, it is determined to be a match), which is used for subsequent tracking and early warning.

[0045] The gated fusion unit introduces an edge weight masking mechanism based on the inversion of edge perturbation probability, performs reverse weight suppression on the instability of edge connections in the task impact graph, and adopts the maximum entropy compression strategy to reduce the distortion propagation probability from the high-noise area to the clustering space, so that the matching accuracy of risk hot spots in the final clustering graph is more than 10% higher than that of the existing risk identification algorithms.

[0046] Specifically, in the edge weight masking mechanism, the perturbation probability is expressed by the following formula: ; where is the original edge weight, represents the instability perturbation probability of edge (i, j), which is calculated from the historical volatility. The maximum entropy compression loss function is defined as follows: ; where is the risk probability output of node i, is the high-order noise distribution, is the trade-off parameter (empirically set to 0.1). The comparative experiment results show that the clustering accuracy of VGAE-GNN is improved by about 13.4% compared with the traditional KMeans, and by about 11.7% compared with the static graph clustering, effectively improving the accuracy of risk clustering hot spots.

[0047] The knowledge graph decision module collaboratively executes the graph edge reconstruction mechanism and the policy edge weight adjustment mechanism. The graph edge reconstruction mechanism adopts an edge embedding update algorithm based on the third-order path frequency regularization to reconstruct the connection weights of key nodes based on the progress deviation chain path; the policy edge weight adjustment mechanism adopts a bilateral cross-attention mechanism to guide the fine-tuning of the edge weights of each policy edge in the project control graph, automatically reducing the transfer probability value of the edges related to the failed path and simultaneously increasing the alternative path when the risk surges, so as to update the output path of the final project global control module from the static path graph to the dynamic path graph calculated based on the current situation.

[0048] Specifically, the graph edge reconstruction mechanism first performs frequency statistics on all third-order task paths. If a certain path appears more than 3 times in the historical deviation records, it is regarded as a potential instability path, and then the edge weight of this path is adjusted. In the policy edge adjustment mechanism, the cross-attention mechanism includes two independent Attention matrices, which act on the current risk graph and the historical success graph respectively, and form a probability adjustment matrix after synthesis. Among them is the dynamic weight parameter (such as = 0.6). Based on this matrix, the optimal control path graph in the current project situation is generated through the Dijkstra or A* search algorithm, and is encapsulated and output in JSON format (such as path node list, edge weight value, path cost), for the API of project scheduling platforms such as OraclePrimavera to call, to implement dynamic path writing and control command issuance.

[0049] The above shows and describes the basic principles, main features and advantages of the present invention; those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present invention, and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed; the scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An information-based project management system based on big data analysis, characterized in that, The described information-based project management system based on big data analysis includes a data acquisition module, a time series data modeling module, a task coupling analysis module, a risk clustering and identification module, a resource allocation prediction module, a progress deviation tracing module, a knowledge graph decision-making module, and a project global control module; The data acquisition module parallelly acquires structured data and unstructured data based on an asynchronous pipeline trigger logic, and encodes them through a unified nested vector standard; The time series data modeling module constructs a multi-dimensional project time series model based on an autoregressive residual network and a multi-scale sliding window mechanism; the task coupling analysis module fuses the task execution trajectories output by the multi-dimensional project time series model with the static dependency matrix in the project task library to form a high-dimensional interaction tensor structure, and performs tensor graph decomposition on the high-dimensional interaction tensor to generate a task influence directed graph; the risk clustering and identification module takes the task influence directed graph as the graph input, generates a multi-dimensional clustering embedding through a variational graph autoencoder, and maps it to the clustering distribution space to identify potential risk points; the resource allocation prediction module nests a resource sensitivity network based on a graph attention mechanism in the clustering distribution space to construct a dynamic priority queue for resource allocation; the progress deviation tracing module maps the identified risk points and resource scheduling offsets based on a deviation propagation chain graph modeling strategy, and identifies the deviation causal chain through multi-hop path accumulation; the knowledge graph decision-making module takes the risk clustering output and the progress deviation chain as inputs, performs cross-graph edge weight tensor reconstruction in a unified graph semantic layer, and corrects the resource allocation priority and path selection strategy through a bilateral attention mechanism; the project global control module receives the correction result in real time, dynamically rewrites the critical path, non-critical path, and resource arrangement table by adopting a hierarchical boundary value strategy, and constructs a closed-loop self-correcting information-based project management system based on big data analysis.

2. The information-based project management system based on big data analysis according to claim 1, characterized in that The data acquisition module includes a structured data acquisition unit, an unstructured text transcoding unit, an unstructured speech transcription unit, and a data normalization sub-module; The structured data acquisition unit adopts a time tag partitioning mechanism based on Kafka streams to capture the log data in ERP, PMIS, and WBS systems in the form of event blocks, and generates a multi-level timestamp index; the unstructured text transcoding unit extracts the entity pair relationships in project emails, minutes, and contract documents through a sentence aggregation network based on the Transformer architecture, and uniformly maps them to a nested triple vector group; The unstructured speech transcription unit adopts a deep attention audio recognition network to transcribe the engineering meeting recordings into a segment structure and segment them into task-related semantic segments; the data normalization sub-module performs unified time domain padding and vector scale normalization on all acquired contents, and compresses them to a fixed-dimensional nested vector standard.

3. The information-based project management system based on big data analysis according to claim 2, wherein, The data normalization sub-module adopts a heterogeneous vector rescaling mechanism based on an improved BatchNorm strategy and combines a reversible linear normalization function based on entropy weights, so that the vector dimension distributions generated by each data source still retain the main component covariance structure unchanged after multi-source mixed input. The standard process of compressing to a fixed-dimensional nested vector uses a multi-head attention screening mechanism to retain only the vector components with the top 10% of the largest clustering entropy value contributions, and the compression accuracy is not less than 93%.

4. An information-based project management system based on big data analysis according to claim 1, characterized in that, The task coupling parsing module represents the coupling relationship between task entities based on a sixth-order coupling tensor structure. The multi-dimensions include task ID, expected duration, resource type, historical deviation rate, external interference index, and responsibility unit number. It uses a tensor graph decomposition algorithm to decompose the potential dependency relationships between tasks into multiple observable directed graphs, and the main sub-graphs set dynamic edge weights according to the deviation propagation frequency.

5. An information-based project management system based on big data analysis according to claim 4, characterized in that, The tensor kernel in the tensor graph decomposition algorithm adopts a low-rank kernel tensor with a truncated rank r = 6. The embedding space constructs an edge weight perturbation adjustment mechanism based on the graph Laplace eigen-distribution, applies a regularization constraint that maximizes the edge connection stability to the coupling tensor offset in multi-period task scheduling, and uses residual gradient guidance to converge to a steady-state influence sub-graph in iterative optimization.

6. An information-based project management system based on big data analysis according to claim 5, characterized in that The steady-state influence sub-graph extracts a set of core risk propagation paths through a centrality screening mechanism, maps the set of core risk propagation paths to the priority scheduling graph in the resource allocation prediction module, and constructs a weighted impedance matrix based on the mutual impedance score between tasks. The weighted impedance matrix guides the resource scheduling order to re-arrange priorities among tasks of the same type.

7. An information-based project management system based on big data analysis according to claim 1, characterized in that, The risk clustering and identification module adopts a graph-structured variational auto-encoder mechanism VGAE-GNN. The encoder consists of two layers of graph convolutional units, and a gated fusion unit based on the high-order reconstruction error distribution is appended after the node representation. It maps the node vectors from the task influence sub-graph to a high-dimensional clustering embedding space, and guides the clustering output to fall into the potential risk hot spot distribution area through the K-L divergence. The hot spot risk distribution forms a connected sub-graph through the DBSCAN density threshold clustering strategy, and each sub-graph matches the isomorphic risk paths in the historical deviation graph.

8. An information-based project management system based on big data analysis according to claim 7, characterized in that The gated fusion unit introduces an edge weight masking mechanism based on the inversion of edge perturbation probability to suppress the reverse weight of the instability of edge connections in the task influence graph, and adopts a maximum entropy compression strategy to reduce the distortion propagation probability from high-noise regions to the clustering space, so that the risk hot spot matching accuracy in the final clustering graph is more than 10% higher than that of existing risk identification algorithms.

9. An information-based project management system based on big data analysis according to claim 1, characterized in that, The knowledge graph decision-making module collaboratively executes using a graph edge reconstruction mechanism and a policy edge weight adjustment mechanism. The graph edge reconstruction mechanism uses an edge embedding update algorithm based on third-order path frequency regularization to reconstruct the connection weights of key nodes based on the progress deviation chain path. The policy edge weight adjustment mechanism uses a bilateral cross-attention mechanism to guide the fine-tuning of the edge weights of each policy edge in the project control graph, automatically lowering the transition probability value of the edges related to the failed path and simultaneously raising the alternative path in the state of rapid risk increase, so as to update the output path of the final project global control module from a static path graph to a dynamic path graph calculated based on the current situation.

Images