A project management system and method based on generative adversarial network

Abstract

The present invention discloses a project management system and method based on a generative adversarial network, the system comprising a data acquisition unit, a feature extraction unit, a schedule prediction unit and a scheme generation unit; wherein the feature extraction unit adopts a long short-term memory network and a multi-head graph attention network to extract temporal features and spatial features respectively; the schedule prediction unit generates a prediction sequence through a generator based on a generative adversarial network architecture, and its rationality is evaluated by a discriminator; the scheme generation unit generates a schedule adjustment scheme based on a critical path analysis, resource optimization and buffer time allocation method; the present invention can effectively improve the accuracy and reliability of project schedule prediction, and provide a scientific basis for project management decision-making.

Description

Technical Field

[0001] The present invention belongs to the technical field of project management, and in particular relates to a project management system and method based on a generative adversarial network. Background Art

[0002] Traditional project management systems and methods are often unsatisfactory when dealing with large-scale, highly uncertain projects. For example, in terms of schedule management, existing project management systems mainly use deterministic methods such as the critical path method (CPM) and the program evaluation and review technique (PERT) to manage project schedules. These methods are usually based on empirical estimates of task durations, and determine the critical path of the project through network diagram analysis, and then formulate corresponding schedules.

[0003] In the actual project execution process, the existing project management system is difficult to effectively capture the dynamic changes in the project execution process, and cannot fully consider the impact of complex dependencies between multiple tasks on the progress. For highly uncertain project environments, the prediction accuracy is poor; the existing system often treats resource constraints as static conditions, ignoring the dynamic changes in resource occupancy, and it is difficult to achieve coordinated optimization among multiple resources; in addition, the existing system often lacks systematic analysis and quantitative evaluation of project risks, and functions such as progress management, resource management, and risk management are often independent of each other, and cannot provide project managers with a variety of progress adjustment plans.

[0004] These problems restrict the efficiency and effectiveness of project management. When faced with a complex and changing project environment, the existing system is difficult to provide satisfactory solutions. Therefore, there is an urgent need to develop a new generation of intelligent project management systems and methods. Summary of the invention

[0005] In view of the above problems, the present invention proposes a project management system and method based on a generative adversarial network. By introducing a generative adversarial network and utilizing artificial intelligence technology, the scientific nature of progress prediction and management in project management is improved.

[0006] In a first aspect, the present invention provides a project management system based on a generative adversarial network, the system comprising: a data acquisition unit, a feature extraction unit, a progress prediction unit and a solution generation unit.

[0007] The data collection unit is used to collect project progress data, task dependency data and resource occupancy data.

[0008] The feature extraction unit is connected to the data collection unit and is used to extract temporal features and spatial features from the collected data; the feature extraction unit includes: a temporal feature extraction module for processing project progress data and resource occupancy data; and a spatial feature extraction module for processing task dependency data.

[0009] The progress prediction unit includes a generative adversarial network engine, which generates a project progress prediction sequence based on the temporal features and spatial features.

[0010] The scheme generation unit is connected to the output end of the progress prediction unit and is used to generate a progress adjustment scheme based on the project progress prediction sequence; the scheme generation unit includes a critical path analysis module, a resource optimization module, a buffer time allocation module and a scheme integration module connected in sequence; the critical path analysis module is used to identify the key task sequence that affects the overall progress of the project; the resource optimization module is used to optimize the task completion time under resource constraints; the buffer time allocation module is used to allocate buffer time for key tasks based on task uncertainty; and the scheme integration module is used to integrate and form a final progress adjustment scheme.

[0011] Furthermore, the time series feature extraction subunit uses a long short-term memory network LSTM to extract time series features:

[0012] ;

[0013] The expressions of the input gate, forget gate, and output gate are:

[0014] , , ;

[0015] Candidate Memory and current memory The expression is:

[0016] , , ;

[0017] in, express The input gate state vector at time , express The forget gate state vector at time , express The output gate state vector at time , express The hidden state vector at time t, express The input vector at time instant, , , , Respectively represent the weight matrices of the corresponding gating units; , , , Respectively represent the bias vectors of the corresponding gating units; represents the sigmoid activation function, represents the hyperbolic tangent activation function, represents the Hadamard product, i.e. element-by-element multiplication; Represents the time series of project milestone completion, Indicates the resource occupation sequence.

[0018] Furthermore, the spatial feature extraction subunit uses a multi-head graph attention network to extract features of task dependency data, and the expression is:

[0019] , where the expression of attention weight is:

[0020] , ;

[0021] in, Representation Node The updated feature representation is represents the number of attention heads, Representation Node The neighborhood set of

[0022] Indicates Attention head nodes For Node The attention weight, Indicates The transformation matrix of the attention head, Indicates The attention vector of the attention head, represents the rectified linear unit activation function with leakage, is the intermediate calculation result of the attention coefficient, which means that Under the attention head, the node For Node The unnormalized attention score of .

[0023] The expression is: , =0.01, used to avoid the death of neurons caused by the traditional ReLU being completely set to zero in the negative value area.

[0024] Furthermore, the generative adversarial network engine uses a multi-layer perceptron MLP to predict the project progress sequence, and the generative adversarial network engine includes a discriminator:

[0025] , where weight initialization is initialization: , the loss functions include:

[0026] , the constraint violation penalty is calculated as:

[0027] , ,in, represents the predicted project schedule sequence, Indicates that the condition Under the noise The generated result is represents the feature concatenation operation, Represents the time series characteristics, Represents spatial features, Indicates The input dimension of the layer, represents normal distribution, BCE represents binary cross entropy loss function, Represents the discriminator for the real data The judgment result of Represents the discriminator's judgment result on the generated data; represents the weight coefficient of the constraint violation penalty term, represents the constraint violation penalty term, represents the minimum time interval constraint, , They represent the upper and lower bounds of the progress respectively.

[0028] Furthermore, a schedule adjustment plan is generated based on the project schedule forecast sequence, including:

[0029] The Critical Path Analysis module calculates the critical task sequence: ,in, Representative tasks, Representation Task The earliest start time, Representation Task The expected duration of Representation Task The latest start time of Represents the set of tasks on the critical path.

[0030] The resource optimization module uses a heuristic algorithm to optimize task completion time under resource constraints:

[0031] , ,in, represents the objective function, i.e., the critical path length, express The set of tasks being executed at any time. Representation Task About Resources The demand, Representation Resources capacity constraints;

[0032] The buffer time allocation module allocates buffer time for critical tasks based on task uncertainty : ,

[0033] in, is the buffer coefficient, Representation Task The uncertainty measure of is estimated by the variance of the forecast series.

[0034] The solution integration module integrates the above optimization results to form the final schedule adjustment plan:

[0035] ,in, and Represent the optimized tasks respectively The earliest and latest start times, Represents the final set of schedule adjustment plans.

[0036] In a second aspect, the present invention provides a project management method based on a generative adversarial network, the project management method is implemented based on the system described in the first aspect, and the project management method comprises the following steps:

[0037] Step S1, collecting project progress data, task dependency data and resource occupancy data.

[0038] Step S2, extracting temporal features and spatial features from the collected data, processing project progress data and resource occupancy data to extract temporal features, and processing task dependency data to extract spatial features.

[0039] Step S3, generating a project progress prediction sequence based on the temporal features and spatial features through a generative adversarial network engine.

[0040] Step S4, generate a schedule adjustment plan based on the project schedule forecast sequence, identify the key task sequence that affects the overall project progress through critical path analysis, optimize the task completion time under resource constraints, allocate buffer time for key tasks based on task uncertainty, and integrate to form the final schedule adjustment plan.

[0041] Furthermore, the timing characteristics The extraction uses the long short-term memory network LSTM: ;

[0042] The expressions of the input gate, forget gate, and output gate are:

[0043] , , ;

[0044] Candidate Memory and current memory The expression is:

[0045] , , ;

[0046] in, express The input gate state vector at time , express The forget gate state vector at time , express The output gate state vector at time , express The hidden state vector at time t, express The input vector at time instant, , , , Respectively represent the weight matrices of the corresponding gating units; , , , Respectively represent the bias vectors of the corresponding gating units; represents the sigmoid activation function, represents the hyperbolic tangent activation function, represents the Hadamard product, i.e. element-by-element multiplication; Represents the time series of project milestone completion, Indicates the resource occupation sequence.

[0047] Furthermore, a multi-head graph attention network is used to extract the features of task dependency data, and the expression is:

[0048] , where the expression of attention weight is:

[0049] , ;

[0050] in, Represents spatial features, which are nodes The updated feature representation is represents the number of attention heads, Representation Node The neighborhood set of

[0051] Indicates Attention head nodes For Node The attention weight, Indicates The transformation matrix of the attention head, Indicates The attention vector of the attention head, represents the rectified linear unit activation function with leakage, is the intermediate calculation result of the attention coefficient, which means that Under the attention head, the node For Node The unnormalized attention score of

[0052] The expression is: , =0.01, used to avoid the death of neurons caused by the traditional ReLU being completely set to zero in the negative value area.

[0053] Furthermore, the generative adversarial network engine uses a multi-layer perceptron MLP to predict the project progress sequence, and the generative adversarial network engine includes a discriminator:

[0054] , where weight initialization is initialization: , the loss functions include:

[0055] , the constraint violation penalty is calculated as:

[0056] , ,in, represents the predicted project schedule sequence, Indicates that the condition Under the noise The generated result is represents the feature concatenation operation, Represents the time series characteristics, Represents spatial features, Indicates The input dimension of the layer, represents normal distribution, BCE represents binary cross entropy loss function, Represents the discriminator for the real data The judgment result of Represents the discriminator's judgment result on the generated data; represents the weight coefficient of the constraint violation penalty term, represents the constraint violation penalty term, represents the minimum time interval constraint, , They represent the upper and lower bounds of the progress respectively.

[0057] Furthermore, a schedule adjustment plan is generated based on the project schedule forecast sequence, including:

[0058] The Critical Path Analysis module calculates the critical task sequence: ,in, Representative tasks, Representation Task The earliest start time, Representation Task The expected duration of Representation Task The latest start time of Represents the set of tasks on the critical path;

[0059] The resource optimization module uses a heuristic algorithm to optimize task completion time under resource constraints:

[0060] , ,in, represents the objective function, i.e., the critical path length, express The set of tasks being executed at any time. Representation Task About Resources The demand, Representation Resources capacity constraints;

[0061] The buffer time allocation module allocates buffer time for critical tasks based on task uncertainty : ,

[0062] in, is the buffer coefficient, Representation Task The uncertainty measure of is estimated by the variance of the forecast series;

[0063] The solution integration module integrates the above optimization results to form the final schedule adjustment plan:

[0064] ,in, and Represent the optimized tasks respectively The earliest and latest start times, Represents the final set of schedule adjustment plans.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] The present invention uses the long short-term memory network LSTM and the multi-head graph attention network to extract temporal features and spatial features respectively, and combines the generative adversarial network to predict the project progress. It can not only fully mine the deep information contained in the project progress data, task dependency data and resource occupancy data, but also improve the accuracy and stability of the prediction results through adversarial training. The system uses He initialization and LeakyReLU activation function to effectively avoid the gradient vanishing problem in the deep neural network training process. By introducing the constraint violation penalty term, it ensures that the generated prediction sequence meets the actual engineering constraints such as the minimum time interval and the upper and lower bounds of the progress.

[0067] On this basis, the system further identifies key tasks through critical path analysis, uses heuristic algorithms to optimize task completion time under resource constraints, and reasonably allocates buffer time based on task uncertainty, ultimately forming a scientific and reliable progress adjustment plan. The functional modules of the present invention work closely together to form a complete analysis and prediction chain from data collection to solution generation, which not only improves the level of intelligent project management, but also provides project managers with practical decision-making support tools. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 The present invention is a schematic diagram of the composition of a project management system based on a generative adversarial network.

[0069] Figure 2 This is a flow chart of a project management method based on a generative adversarial network according to the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] Example 1

[0072] like Figure 1 As shown, it is a schematic diagram of the composition of a project management system based on a generative adversarial network of the present invention, and the system includes: a data acquisition unit, a feature extraction unit, a progress prediction unit and a solution generation unit.

[0073] The data collection unit is used to collect project progress data, task dependency data and resource occupancy data.

[0074] Data collection supports multiple formats: project progress data supports the import of progress plans of mainstream project management software such as Microsoft Project and OraclePrimavera; task dependency data supports the import of network diagrams, adjacency matrices and other representation methods; resource occupancy data supports the collection and integration of multi-dimensional resource data such as manpower, equipment, and materials.

[0075] The data preprocessing process includes the following steps: (1) Data cleaning: missing values ​​are interpolated based on time series correlation, outliers are corrected using the moving median method, time series data are standardized using Z-score, and spatial feature data are normalized using Min-Max; (2) Data enhancement: training samples are expanded by adding Gaussian noise (μ=0, σ=0.01) and random time offset (±5%); (3) Feature engineering: derived features are constructed, including indicators such as task completion rate, resource utilization rate, and progress deviation rate, and sliding window statistics (window size is 7 days) such as mean, variance, trend, etc. are calculated. To ensure data quality, the system also establishes a data consistency check mechanism, which automatically identifies and marks abnormal data points that require manual review through the rule engine.

[0076] The data collection unit adopts specific preprocessing strategies when collecting different types of data: for project progress data, the system supports incremental data collection and can automatically extract updated progress information from the project management software on a regular basis (such as daily or weekly); for task dependency data, the system can handle complex dependency types, including the four basic dependencies of finish-start (FS), start-start (SS), finish-finish (FF) and start-finish (SF), and can handle dependencies with lead or lag; for resource occupancy data, the system supports the automated collection and integration of multi-dimensional resource data, including attribute information such as the professional skill level of human resources, the work efficiency of equipment, and the availability of materials.

[0077] The feature extraction unit is connected to the data collection unit and is used to extract temporal features and spatial features from the collected data; the feature extraction unit includes: a temporal feature extraction module for processing project progress data and resource occupancy data; and a spatial feature extraction module for processing task dependency data.

[0078] The progress prediction unit includes a generative adversarial network engine, which generates a project progress prediction sequence based on the temporal features and spatial features.

[0079] The scheme generation unit is connected to the output end of the progress prediction unit and is used to generate a progress adjustment scheme based on the project progress prediction sequence; the scheme generation unit includes a critical path analysis module, a resource optimization module, a buffer time allocation module and a scheme integration module connected in sequence; the critical path analysis module is used to identify the key task sequence that affects the overall progress of the project; the resource optimization module is used to optimize the task completion time under resource constraints; the buffer time allocation module is used to allocate buffer time for key tasks based on task uncertainty; and the scheme integration module is used to integrate and form a final progress adjustment scheme.

[0080] Supplementary information on the overall system architecture: The system adopts a modular design, and each unit interacts with data through a standard interface; the data acquisition unit can connect to a variety of data sources, including project management software APIs, databases, and file systems; the system as a whole adopts a layered architecture to facilitate subsequent expansion and maintenance.

[0081] The time series feature extraction subunit uses the long short-term memory network LSTM to extract time series features:

[0082] ;

[0083] The expressions of the input gate, forget gate, and output gate are:

[0084] , , ;

[0085] Candidate Memory and current memory The expression is:

[0086] , , ;

[0087] in, express The input gate state vector at time , express The forget gate state vector at time , express The output gate state vector at time , express The hidden state vector at time t, express The input vector at time instant, , , , Respectively represent the weight matrices of the corresponding gating units; , , , Respectively represent the bias vectors of the corresponding gating units; represents the sigmoid activation function, represents the hyperbolic tangent activation function, represents the Hadamard product, i.e. element-by-element multiplication; Represents the time series of project milestone completion, Indicates the resource occupation sequence.

[0088] When extracting time series features, the specific implementation of the LSTM network can adopt a bidirectional LSTM structure to make full use of historical and future information. The network depth can be dynamically adjusted according to the project scale. Generally, a 2-3-layer stacking structure is adopted. In order to improve the robustness of feature extraction, a Dropout layer can be added between LSTM layers. The typical dropout rate is 0.2-0.3.

[0089] The optimization strategies for the LSTM network include: using the Adam optimizer for parameter update, setting the initial learning rate to 0.001, and using a learning rate decay strategy, where the learning rate is reduced to 0.1 times the original after every 50 epochs; to prevent overfitting, in addition to using the Dropout layer, L2 regularization is also introduced, and the regularization coefficient is set to 0.0001; the batch size is dynamically adjusted according to the data scale, generally between 32 and 128; the early stopping strategy is used during training, and training is stopped when the loss function on the validation set has not improved for five consecutive epochs.

[0090] The spatial feature extraction subunit uses a multi-head graph attention network to extract the features of task dependency data, and the expression is:

[0091] , where the expression of attention weight is:

[0092] , ;

[0093] in, Representation Node The updated feature representation is represents the number of attention heads, Representation Node The neighborhood set of

[0094] Indicates Attention head nodes For Node The attention weight, Indicates The transformation matrix of the attention head, Indicates The attention vector of the attention head, represents the rectified linear unit activation function with leakage, is the intermediate calculation result of the attention coefficient, which means that Under the attention head, the node For Node The unnormalized attention score of

[0095] The expression is: , =0.01, used to avoid the death of neurons caused by the traditional ReLU being completely set to zero in the negative value area.

[0096] In the multi-head attention mechanism, the number of attention heads K is usually set to 4 or 8. Each attention head can focus on different aspects of the task dependency network, such as direct predecessor relationships, indirect dependencies, etc. Residual connections can be introduced to alleviate the difficulty of deep network training.

[0097] The following strategies are used for network structure optimization:

[0098] (1) Attention mechanism optimization: Based on multi-head attention, a gating mechanism is introduced. The gating function To control the flow of information, is a learnable gating weight matrix;

[0099] (2) Feature fusion optimization: The feature gating unit (FGU) is used to achieve adaptive fusion of temporal features and spatial features. The fusion weight is:

[0100] , the final feature is expressed as:

[0101] ;

[0102] (3) Residual connection optimization: Introduce cross-layer residual connections and dense connections to improve the training effect of deep networks. The specific implementation is as follows: ,in is the nonlinear transformation function of the lth layer.

[0103] The generative adversarial network engine uses a multi-layer perceptron MLP to predict the project progress sequence. The generative adversarial network engine contains a discriminator:

[0104] , where weight initialization is initialization: ; The training strategy of the generative adversarial network can adopt the WGAN-GP framework. The specific implementation of the WGAN-GP framework includes the following key settings: the gradient penalty coefficient is set to 10, the generator and the discriminator both use a 5-layer feedforward neural network structure, and the number of neurons in each layer is [512, 256, 128, 64, 32], and the LeakyReLU activation function is used in all layers except the last layer. The last layer of the generator uses the Tanh activation function, and the discriminator uses a linear output. During the training process, the generator is updated once every 5 discriminator updates, using the Adam optimizer, with the learning rate set to 0.0001, β1=0.5, and β2=0.9. In order to improve the quality of the generated results, a historical buffer pool mechanism is introduced during training to store the 100 most recently generated samples for discriminator training.

[0105] The loss functions include:

[0106] , the constraint violation penalty is calculated as:

[0107] , ,in, represents the predicted project schedule sequence, Indicates that the condition Under the noise The generated result is represents the feature concatenation operation, Represents the time series characteristics, Represents spatial features, Indicates The input dimension of the layer, represents normal distribution, BCE represents binary cross entropy loss function, Represents the discriminator for the real data The judgment result of Represents the discriminator's judgment result on the generated data; represents the weight coefficient of the constraint violation penalty term, represents the constraint violation penalty term, represents the minimum time interval constraint, , They represent the upper and lower bounds of the progress respectively.

[0108] Generate a schedule adjustment plan based on the project schedule forecast sequence, including:

[0109] The Critical Path Analysis module calculates the critical task sequence: ,in, Representative tasks, Representation Task The earliest start time, Representation Task The expected duration of Representation Task The latest start time of Represents the set of tasks on the critical path;

[0110] In addition, the critical path analysis can also use an improved critical path algorithm: first use depth-first search to identify all possible paths, then calculate the total duration of each path, while considering the dependencies between tasks and resource constraints. Resource optimization uses an improved genetic algorithm: the population size is set to 100, the number of iterations is 500, the crossover probability is 0.8, the mutation probability is 0.1, and the tournament selection strategy is adopted. Buffer time allocation adopts a risk-based dynamic allocation strategy: the buffer coefficient b ranges from [0.1, 0.3] and is dynamically adjusted according to the historical delay of the task and the stability of resources; the uncertainty measure σ_i of the task is calculated by analyzing the variance of the completion time of the task in multiple forecasts, and is weighted in combination with the subjective evaluation of the project manager. When integrating the solutions, the system will generate multiple alternative solutions, and comprehensively score the solutions from three dimensions: schedule optimization degree, resource utilization efficiency, and risk controllability, and select the solution with the highest score as the final output.

[0111] The resource optimization module uses a heuristic algorithm to optimize task completion time under resource constraints:

[0112] , ,in, represents the objective function, i.e., the critical path length, express The set of tasks being executed at any time. Representation Task About Resources The demand, Representation Resources capacity constraints;

[0113] The buffer time allocation module allocates buffer time for critical tasks based on task uncertainty : ,

[0114] in, is the buffer coefficient, Representation Task The uncertainty measure of is estimated by the variance of the forecast series;

[0115] The solution integration module integrates the above optimization results to form the final schedule adjustment plan:

[0116] ,in, and Represent the optimized tasks respectively The earliest and latest start times, Represents the final set of schedule adjustment plans.

[0117] The training strategy of the generative adversarial network can adopt the WGAN-GP framework to improve training stability. The update ratio of the discriminator and the generator can be set to 5:1.

[0118] The system provides standardized interface specifications: (1) Data interface: supports REST API and WebSocket, the data format uses JSON, supports batch operations and incremental synchronization; (2) Visualization interface: based on Web component design, provides visualization components such as project Gantt chart, resource load chart, progress deviation chart, etc., supports custom chart styles and interaction methods; (3) Alarm interface: set multi-level alarm rules (general, important, urgent), support multiple alarm methods such as email, SMS, system notification, etc., and the alarm threshold can be dynamically adjusted through the configuration file; (4) Permission interface: adopts RBAC model to manage user permissions, supports fine-grained data access control and operation audit. The interface design follows the OpenAPI specification, provides detailed API documentation and call examples, and facilitates third-party system integration.

[0119] Taking a large-scale software development project as an example, the project includes multiple stages such as requirements analysis, design, development, and testing, involving 50 specific tasks and lasting 12 months.

[0120] The actual application process of the system includes data collection: collecting task progress data from the JIRA system, collecting resource usage data from the human resources system, and extracting task dependencies through project documents; feature extraction: the LSTM network extracts the historical completion status of each task, and the graph attention network analyzes the complex dependencies between tasks; progress prediction: the generator predicts the task completion status in the next three months, and the discriminator ensures that the prediction results are in line with the actual rules of the project; plan generation: identify the test phase as the critical path, optimize the allocation of test resources, allocate additional buffer time for high-risk modules, and generate adjustment plans with specific time nodes.

[0121] Finally, the system successfully predicted that the project might be delayed by 2 weeks, and provided a specific optimization plan by adding testers and adjusting the test strategy. After the implementation of the plan, the project was completed on schedule. This example shows the application value of the system in actual projects, especially its advantages in schedule prediction and optimization of complex projects. The modular design of the system also allows it to be flexibly configured and adjusted according to the characteristics of different projects.

[0122] Example 2

[0123] like Figure 2 As shown, it is a flow chart of a project management method based on a generative adversarial network of the present invention, and the method comprises the following steps:

[0124] Step S1, collecting project progress data, task dependency data and resource occupancy data.

[0125] Step S2, extracting temporal features and spatial features from the collected data, processing project progress data and resource occupancy data to extract temporal features, and processing task dependency data to extract spatial features.

[0126] Step S3, generating a project progress prediction sequence based on the temporal features and spatial features through a generative adversarial network engine.

[0127] Step S4, generate a schedule adjustment plan based on the project schedule forecast sequence, identify the key task sequence that affects the overall project progress through critical path analysis, optimize the task completion time under resource constraints, allocate buffer time for key tasks based on task uncertainty, and integrate to form the final schedule adjustment plan.

[0128] Resource scheduling uses a multi-objective optimization method: (1) Resource smoothing: by minimizing the variance of resource usage Achieve resource load balancing; (2) Cost optimization: Establish a cost function ,in Representation Resources Complete the task The unit cost, Represents the amount of resource allocation; (3) Multi-objective trade-off: The Pareto optimal method is used to deal with the triple constraints of schedule, resources, and cost, and the optimal solution set is solved by the NSGA-II algorithm (population size = 200, number of iterations = 1000). Resource scheduling also takes into account the dynamic availability of resources, estimates the future availability of resources through a prediction model, and sets task priorities according to the degree of urgency to achieve dynamic allocation and adjustment of resources.

[0129] The timing characteristics The extraction uses the long short-term memory network LSTM: ;

[0130] The expressions of the input gate, forget gate, and output gate are:

[0131] , , ;

[0132] Candidate Memory and current memory The expression is:

[0133] , , ;

[0134] in, express The input gate state vector at time , express The forget gate state vector at time , express The output gate state vector at time , express The hidden state vector at time t, express The input vector at time instant, , , , Respectively represent the weight matrices of the corresponding gating units; , , , Respectively represent the bias vectors of the corresponding gating units; represents the sigmoid activation function, represents the hyperbolic tangent activation function, represents the Hadamard product, i.e. element-by-element multiplication; Represents the time series of project milestone completion, Indicates the resource occupation sequence.

[0135] The multi-head graph attention network is used to extract the features of task dependency data, and the expression is:

[0136] , where the expression of attention weight is:

[0137] , ;

[0138] in, Represents spatial features, which are nodes The updated feature representation is represents the number of attention heads, Representation Node The neighborhood set of

[0139] Indicates Attention head nodes For Node The attention weight, Indicates The transformation matrix of the attention head, Indicates The attention vector of the attention head, represents the rectified linear unit activation function with leakage, is the intermediate calculation result of the attention coefficient, which means that Under the attention head, the node For Node The unnormalized attention score of

[0140] The expression is: , =0.01, used to avoid the death of neurons caused by the traditional ReLU being completely set to zero in the negative value area.

[0141] The generative adversarial network engine uses a multi-layer perceptron MLP to predict the project progress sequence. The generative adversarial network engine includes a discriminator:

[0142] , where weight initialization is initialization: , the loss functions include:

[0143] , the constraint violation penalty is calculated as:

[0144] , ,in, represents the predicted project schedule sequence, Indicates that the condition Under the noise The generated result is represents the feature concatenation operation, Represents the time series characteristics, Represents spatial features, Indicates The input dimension of the layer, represents normal distribution, BCE represents binary cross entropy loss function, Represents the discriminator for the real data The judgment result of Represents the discriminator's judgment result on the generated data; represents the weight coefficient of the constraint violation penalty term, represents the constraint violation penalty term, represents the minimum time interval constraint, , They represent the upper and lower bounds of the progress respectively.

[0145] Generate a schedule adjustment plan based on the project schedule forecast sequence, including:

[0146] The Critical Path Analysis module calculates the critical task sequence: ,in, Representative tasks, Representation Task The earliest start time, Representation Task The expected duration of Representation Task The latest start time of Represents the set of tasks on the critical path;

[0147] The resource optimization module uses a heuristic algorithm to optimize task completion time under resource constraints:

[0148] , ,in, represents the objective function, i.e., the critical path length, express The set of tasks being executed at any time. Representation Task About Resources The demand, Representation Resources capacity constraints;

[0149] The buffer time allocation module allocates buffer time for critical tasks based on task uncertainty : ,

[0150] in, is the buffer coefficient, Representation Task The uncertainty measure of is estimated by the variance of the forecast series;

[0151] The solution integration module integrates the above optimization results to form the final schedule adjustment plan:

[0152] ,in, and Represent the optimized tasks respectively The earliest and latest start times, Represents the final set of schedule adjustment plans.

[0153] Taking a large software development project as a specific case, the implementation process of this method is explained in detail:

[0154] Data collection phase (step S1): The initial planning period of the software development project is 12 months, including 50 specific tasks. The data collection process is as follows:

[0155] The project progress data is exported from the JIRA system, including the task completion records for the past 6 months, including the planned start time, actual start time, planned completion time, actual completion time, completion rate, etc. The data collection frequency is once a day, and a total of 180 days of data are collected.

[0156] Task dependency data Task dependencies were extracted from the project design documents, and an adjacency matrix was constructed to represent the task network. A total of 127 dependencies were identified, including 64 finish-start (FS) dependencies, 32 start-start (SS) dependencies, 21 finish-finish (FF) dependencies, and 10 start-finish (SF) dependencies.

[0157] Resource usage data is collected from the human resources system, including the professional skills, work experience, historical performance, and other information of the 35 members of the development team; and the usage of 12 test servers and 8 test environments is collected from the asset management system.

[0158] Feature extraction stage (step S2): Time series feature extraction, building a bidirectional LSTM network with a network depth of 3 layers and the number of neurons in each layer being [128, 64, 32] respectively. Taking 7 days as the time window, extract time series features such as task completion rate and resource utilization rate. For each task, calculate the statistical features of its historical completion status, such as the average number of delay days and delay probability. LSTM training adopts the mini-batch method, with the batch size set to 64 and the number of training rounds being 200.

[0159] For spatial feature extraction, an 8-head attention network is configured, and the output dimension of each attention head is 32. The criticality of each task is analyzed through the task dependency network, the node centrality index is calculated, and the key task chain of the test phase is identified, which includes 12 strongly related test tasks.

[0160] Progress prediction stage (step S3): The generator network structure is a 5-layer feedforward neural network [512->256->128->64->32], and the discriminator network structure is a 5-layer feedforward neural network [32->64->128->256->1].

[0161] The training process consists of three stages: the first stage (rounds 1-100): the generator learns the basic task timing pattern; the second stage (rounds 101-300): the discriminator is introduced for adversarial training; the third stage (rounds 301-500): fine-tuning the model parameters to improve prediction accuracy; the prediction results show that the project may be delayed by 2 weeks during the testing phase.

[0162] Solution generation phase (step S4): Identify the main risk points in the test phase: integration test and performance test tasks, calculate the task slack time on the critical path, and find that the slack time of the performance test task is close to 0. Resource optimization: Before optimization: 8 test team members, 2 test environments, optimization suggestions: add 2 performance test engineers, configure 1 dedicated performance test environment in parallel; expected effect: can reduce performance test time by 30%. Buffer time allocation: 5 days of buffer time are allocated for performance test tasks (based on historical variance analysis), 3 days of buffer time are allocated for integration test tasks, and 2 days of buffer time are allocated for system test tasks.

[0163] Final solution integration: adjust the test team structure and size, optimize the order of test task execution, start performance testing in advance, increase daily morning meetings to track progress, and set milestone checks at three key time nodes.

[0164] The effects of the implementation of the plan include: short-term effect: the testing phase was completed as scheduled, and the expected 2-week delay did not occur; medium-term effect: testing efficiency increased by 25%, and defect detection rate increased by 20%; long-term effect: the plan was solidified as an organizational-level process asset for subsequent project management optimization.

[0165] This case fully demonstrates the application process and effect of this method in actual projects. Through systematic data collection, feature extraction, predictive analysis and solution generation, it successfully solved the risk of project delay and established reusable project management optimization experience. Especially in dealing with complex task dependencies and resource constraints, this method has shown strong practical value.

[0166] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A project management system based on generative adversarial network, characterized in that: The system comprises: a data collection unit, a feature extraction unit, a progress prediction unit and a solution generation unit; The data collection unit is used to collect project progress data, task dependency data and resource occupancy data; The feature extraction unit is connected to the data collection unit and is used to extract time series features and space features from the collected data; the feature extraction unit includes: a time series feature extraction module for processing project progress data and resource occupancy data; a space feature extraction module for processing task dependency data; The progress prediction unit includes a generative adversarial network engine, which generates a project progress prediction sequence based on the temporal features and spatial features; The scheme generation unit is connected to the output end of the progress prediction unit and is used to generate a progress adjustment scheme based on the project progress prediction sequence; the scheme generation unit includes a critical path analysis module, a resource optimization module, a buffer time allocation module and a scheme integration module connected in sequence; the critical path analysis module is used to identify the key task sequence that affects the overall progress of the project; the resource optimization module is used to optimize the task completion time under resource constraints; the buffer time allocation module is used to allocate buffer time for key tasks based on task uncertainty; and the scheme integration module is used to integrate and form a final progress adjustment scheme; The generative adversarial network engine uses a multi-layer perceptron MLP to predict the project progress sequence. The generative adversarial network engine includes a discriminator: , where weight initialization is initialization: , the loss functions include: , the constraint violation penalty is calculated as: , ,in, represents the predicted project schedule sequence, Indicates that the condition Under the noise The generated result is represents the feature concatenation operation, Represents the time series characteristics, Indicates The input dimension of the layer, represents normal distribution, BCE represents binary cross entropy loss function, Represents the discriminator for the real data The judgment result of Represents the discriminator's judgment result on the generated data; represents the weight coefficient of the constraint violation penalty term, represents the constraint violation penalty term, represents the minimum time interval constraint, , They represent the upper and lower bounds of the progress respectively.

2. A project management system based on generative adversarial network according to claim 1, characterized in that: The time series feature extraction subunit uses a long short-term memory network LSTM to extract time series features: ； The expressions of the input gate, forget gate, and output gate are: ， ， ； Candidate Memory and current memory The expression is: ， ， ； in, express The input gate state vector at time , express The forget gate state vector at time , express The output gate state vector at time , express The hidden state vector at time t, express The input vector at time instant, , , , Respectively represent the weight matrices of the corresponding gating units; , , , Respectively represent the bias vectors of the corresponding gating units; represents the sigmoid activation function, represents the hyperbolic tangent activation function, represents the Hadamard product, i.e. element-by-element multiplication; Represents the time series of project milestone completion, Indicates the resource occupation sequence.

3. A project management system based on generative adversarial network according to claim 2, characterized in that: The spatial feature extraction subunit uses a multi-head graph attention network to extract the features of task dependency data, and the expression is: , where the expression of attention weight is: ， ； in, Represents spatial features, which are nodes The updated feature representation is represents the number of attention heads, Representation Node The neighborhood set of Indicates Attention head nodes For Node The attention weight, Indicates The transformation matrix of the attention head, Indicates The attention vector of the attention head, represents the rectified linear unit activation function with leakage, is the intermediate calculation result of the attention coefficient, which means that Under the attention head, the node For Node The unnormalized attention score of The expression is: , =0.01, used to avoid the death of neurons caused by the traditional ReLU being completely set to zero in the negative value area.

4. A project management system based on generative adversarial network according to claim 3, characterized in that: Generate a schedule adjustment plan based on the project schedule forecast sequence, including: The Critical Path Analysis module calculates the critical task sequence: ,in, Representative tasks, Indicates the task The earliest start time, Indicates the task The expected duration of Indicates the task The latest start time of Represents the set of tasks on the critical path; The resource optimization module uses a heuristic algorithm to optimize task completion time under resource constraints: , ,in, represents the objective function, i.e., the critical path length, express The set of tasks being executed at any time. Indicates the task About Resources The demand, Representation Resources capacity constraints; The buffer time allocation module allocates buffer time for critical tasks based on task uncertainty : , in, is the buffer coefficient, Indicates the task The uncertainty measure of is estimated by the variance of the forecast series; The solution integration module integrates the optimization results into the final schedule adjustment plan: ,in, and Represent the optimized tasks respectively The earliest and latest start times, Represents the final set of schedule adjustment plans.

5. A project management method based on a generative adversarial network, the project management method is implemented based on the system according to any one of claims 1 to 4, characterized in that: The project management method comprises the following steps: Step S1, collecting project progress data, task dependency data and resource occupancy data; Step S2, extracting temporal features and spatial features from the collected data, processing project progress data and resource occupancy data to extract temporal features, and processing task dependency data to extract spatial features; Step S3, generating a project progress prediction sequence based on the temporal features and spatial features through a generative adversarial network engine; Step S4, generating a schedule adjustment plan based on the project schedule forecast sequence, identifying the key task sequence that affects the overall project progress through critical path analysis, optimizing the task completion time under resource constraints, allocating buffer time for key tasks based on task uncertainty, and integrating to form a final schedule adjustment plan; The generative adversarial network engine uses a multi-layer perceptron MLP to predict the project progress sequence. The generative adversarial network engine includes a discriminator: , where weight initialization is initialization: , the loss functions include: , the constraint violation penalty is calculated as: , ,in, represents the predicted project schedule sequence, Indicates that the condition Under the noise The generated result is represents the feature concatenation operation, Represents the time series characteristics, Indicates The input dimension of the layer, represents normal distribution, BCE represents binary cross entropy loss function, Represents the discriminator for the real data The judgment result of Represents the discriminator's judgment result on the generated data; represents the weight coefficient of the constraint violation penalty term, represents the constraint violation penalty term, represents the minimum time interval constraint, , They represent the upper and lower bounds of the progress respectively.

6. A project management method based on generative adversarial network according to claim 5, characterized in that: The time series feature extraction subunit uses a long short-term memory network LSTM to extract time series features: ； The expressions of the input gate, forget gate, and output gate are: ， ， ； Candidate Memory and current memory The expression is: ， ， ； in, express The input gate state vector at time , express The forget gate state vector at time , express The output gate state vector at time , express The hidden state vector at time t, express The input vector at time instant, , , , Respectively represent the weight matrices of the corresponding gating units; , , , Respectively represent the bias vectors of the corresponding gating units; represents the sigmoid activation function, represents the hyperbolic tangent activation function, represents the Hadamard product, i.e. element-by-element multiplication; Represents the time series of project milestone completion, Indicates the resource occupation sequence.

7. A project management method based on generative adversarial network according to claim 6, characterized in that: The spatial feature extraction subunit uses a multi-head graph attention network to extract the features of task dependency data, and the expression is: , where the expression of attention weight is: ， ； in, Represents spatial features, which are nodes The updated feature representation is represents the number of attention heads, Representation Node The neighborhood set of Indicates Attention head nodes For Node The attention weight, Indicates The transformation matrix of the attention head, Indicates The attention vector of the attention head, represents the rectified linear unit activation function with leakage, is the intermediate calculation result of the attention coefficient, which means that Under the attention head, the node For Node The unnormalized attention score of The expression is: , =0.01, used to avoid the death of neurons caused by the traditional ReLU being completely set to zero in the negative value area.

8. A project management method based on generative adversarial network according to claim 7, characterized in that: Generate a schedule adjustment plan based on the project schedule forecast sequence, including: The Critical Path Analysis module calculates the critical task sequence: ,in, Representative tasks, Indicates the task The earliest start time, Indicates the task The expected duration of Indicates the task The latest start time of Represents the set of tasks on the critical path; The resource optimization module uses a heuristic algorithm to optimize task completion time under resource constraints: , ,in, represents the objective function, i.e., the critical path length, express The set of tasks being executed at any time. Indicates the task About Resources The demand, Representation Resources capacity constraints; The buffer time allocation module allocates buffer time for critical tasks based on task uncertainty : , in, is the buffer coefficient, Indicates the task The uncertainty measure of is estimated by the variance of the forecast series; The solution integration module integrates the optimization results into the final schedule adjustment plan: ,in, and Represent the optimized tasks respectively The earliest and latest start times, Represents the final set of schedule adjustment plans.

Images