An AI-based software project schedule and supervision system

Abstract

The present invention discloses an AI-based software project schedule and supervision system, which relates to the field of artificial intelligence technology. It includes optimizing the initial schedule of the project by combining reinforcement learning and simulated annealing algorithm, comprehensively considering multi-dimensional factors among tasks, and generating a preliminary schedule plan; dynamically adjusting the initial schedule of the project based on genetic algorithm, reordering the priority and execution order of tasks to minimize project delays and resource waste; using multi-modal data fusion to predict risks of project progress, and identifying potential risks in the initial schedule by integrating various data sources; finally simulating and optimizing the initial schedule plan through a generative adversarial network, generating multiple candidate schedule plans and selecting the optimal schedule plan. The present invention uses a generative adversarial network to generate multiple candidate plans, combines non-dominated sorting and Monte Carlo simulation to screen the optimal schedule, and ensures that the project achieves an overall optimal balance among the construction period, cost and risks.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to a software project schedule and supervision system based on AI. Background Art

[0002] In recent years, artificial intelligence technology has been gradually introduced into the field of software project management to optimize schedule and resource allocation. Traditional methods rely on the Critical Path Method (CPM) and Program Evaluation and Review Technique (PERT), generating schedule plans through static task dependencies and fixed resource allocations. With the expansion of complex project scales, dynamic scheduling algorithms based on machine learning (such as reinforcement learning and genetic algorithms) have been applied to multi-objective optimization to address issues such as dynamic adjustment of task priorities and resource competition conflicts. In addition, graph neural networks and multi-modal data fusion technologies have gradually been used to model complex dependencies between tasks and risk prediction. However, existing technologies mostly focus on a single optimization dimension (such as time or resources), lacking a dynamic collaborative optimization mechanism for multi-dimensional factors (time, resources, risks), and having limitations in real-time anomaly detection and the robustness of global solutions.

[0003] The main deficiencies of current technologies are reflected in two aspects: Firstly, traditional reinforcement learning algorithms are prone to falling into local optima in schedule optimization and do not effectively combine heuristic algorithms (such as simulated annealing) to balance exploration and exploitation capabilities, resulting in insufficient adaptability of schedule plans to dynamic environments; Secondly, existing resource scheduling models are mostly based on static competition graph analysis and fail to real-time model the dynamic resource sharing relationships between tasks through adaptive network algorithms, causing low resource utilization and loss of execution efficiency. For example, although the adjustment strategy based on genetic algorithms can optimize priorities, it lacks a closed-loop feedback with the anomaly detection module and is difficult to respond promptly to sudden delays or resource bottlenecks. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a software project schedule and supervision system based on AI to solve the problems of insufficient comprehensive optimization and real-time adjustment efficiency of multi-dimensional factors in software project scheduling in a dynamic environment.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] The present invention provides a software project schedule and supervision system based on AI, which includes a schedule optimization module that optimizes the initial schedule of the project by combining reinforcement learning and simulated annealing algorithms, comprehensively considering multi-dimensional factors between tasks, and generating a preliminary schedule plan;

[0008] The resource scheduling module optimizes resource scheduling using an adaptive network algorithm, analyzes the resource competition situation among tasks in real time, dynamically adjusts the resource allocation plan, and maximizes the execution efficiency of the project;

[0009] The anomaly detection module tracks the project progress and performs anomaly detection through unsupervised learning algorithms to identify potential task delays and resource bottlenecks in advance;

[0010] The priority adjustment module dynamically adjusts the initial schedule of the project based on the genetic algorithm, reorders the priorities and execution sequences of tasks, and minimizes project delays and resource waste;

[0011] The risk prediction module uses multi-modal data fusion to predict the risks of the project progress, and identifies potential risks in the initial schedule by integrating various data sources;

[0012] The solution optimization module finally simulates and optimizes the initial schedule plan through a generative adversarial network, generates multiple candidate schedule plans, and selects the optimal schedule plan.

[0013] As a preferred solution of the AI-based software project schedule and supervision system described in the present invention, wherein: the steps of generating the preliminary schedule plan are as follows.

[0014] Receive the requirement information and resource data related to the software project, clean and standardize them, and generate a structured project requirement database;

[0015] Extract multi-dimensional features for each task, perform feature normalization and standardization processing on different types of features, and output multi-dimensional factors of time, resources, and risks of the project;

[0016] Use the Q-learning reinforcement learning algorithm, take task scheduling as the agent action, define the state space and action space of the project, and establish a scheduling decision model;

[0017] Continuously adjust the task schedule through interaction with the environment, construct a reward function, and select the optimal scheduling action;

[0018] On the basis of reinforcement learning, use the simulated annealing algorithm to further optimize the schedule plan;

[0019] Generate a neighborhood plan by adjusting the task order and resource allocation, and use the acceptance criterion to update the latest schedule plan to avoid local optimality;

[0020] Comprehensively consider multi-dimensional factors, further comprehensively evaluate the updated schedule plan and adjust the plan in real time, and finally generate a preliminary schedule plan.

[0021] As a preferred solution of the AI-based software project schedule and supervision system described in the present invention, wherein: the state space of the project includes the start time, end time, resource usage, and task dependencies of the tasks, and the action space of the project consists of the schedule selections of the tasks. For each state, the agent selects the execution order of the tasks and resource allocation;

[0022] The steps for selecting the optimal scheduling action are as follows:

[0023] Based on the state space of the project, a reward function is constructed by giving positive rewards for completion timeliness and resource utilization efficiency, and negative rewards for delays and resource waste;

[0024] The agent executes an action in the current state and adjusts the schedule according to the reward value and state transition feedback by the environment;

[0025] Through the Q-learning reinforcement learning algorithm, the Q value is updated in combination with the reward value feedback by the environment, the optimal scheduling action is selected, and the schedule is continuously optimized.

[0026] As a preferred solution of the AI-based software project schedule and supervision system described in the present invention, wherein: the steps for maximizing the execution efficiency of the project are as follows:

[0027] Extract the task resource requirements in the preliminary schedule, analyze the resource competition relationships between tasks, and construct a resource competition graph;

[0028] Use a pre-trained graph neural network to model the resource sharing and dependencies between tasks, and combine with the deep reinforcement learning algorithm to optimize the task resource allocation strategy in real time;

[0029] Real-time monitor the task resource usage, combine with the pre-trained long short-term memory network to predict the task progress, and adjust the resource allocation, and dynamically adjust the resources according to the progress feedback and prediction results.

[0030] As a preferred solution of the AI-based software project schedule and supervision system described in the present invention, wherein: the steps for identifying potential task delays and resource bottlenecks in advance are as follows:

[0031] During the process of dynamically adjusting resources, record the task execution status data in real time, track the task progress, resource consumption, and the deviation between the estimated and actual completion times, and establish a dynamically updated project progress database;

[0032] Extract time series features, resource consumption features, and task dependencies from the project progress database, perform normalization and standardization processing, and output a feature dataset;

[0033] Use the Isolation Forest algorithm to train the feature dataset, identify potential patterns of abnormal progress and resource consumption, and construct a detection model for task delays and resource bottlenecks;

[0034] At each task progress update, use the detection model of task delays and resource bottlenecks to evaluate the anomaly score of the task in real time and identify the risks of delays and resource bottlenecks.

[0035] As a preferred solution of the AI-based software project schedule scheduling and supervision system described in the present invention, wherein: minimizing project delays and resource waste, the specific steps are as follows,

[0036] Create a genetic algorithm environment, determine the individual coding method, and generate an initial population;

[0037] Each individual expresses task sequencing and resource allocation through gene coding;

[0038] Based on project progress, resource utilization rate, delay penalty, and task priority, evaluate the quality of each scheduling plan and define a fitness function;

[0039] Use the tournament selection method to select individuals with higher fitness from the current initial population as parent individuals to participate in crossover and mutation operations;

[0040] Generate new scheduling plans by exchanging the genes of parent individuals through crossover operations;

[0041] Evaluate the fitness of each generation of individuals and select the optimal plan;

[0042] According to the project progress, monitor the task execution situation in real time, dynamically adjust the scheduling plan, and if delays or resource bottlenecks occur, use the genetic algorithm to further optimize the scheduling.

[0043] As a preferred solution of the AI-based software project schedule scheduling and supervision system described in the present invention, wherein: evaluating the quality of each scheduling plan and defining a fitness function, the specific steps are as follows,

[0044] Collect the task data of the project and perform cleaning and standardization processing;

[0045] Extract numerical features from the processed task data as the pre-input of the fitness function;

[0046] Calculate the project progress based on the proportion of completed tasks and the total number of tasks as the first weight coefficient of the fitness function;

[0047] Calculate the ratio of actual resource consumption to total resources to evaluate the resource utilization rate as the second weight coefficient of the fitness function;

[0048] By calculating the difference between the actual completion time and the planned completion time of each task, a penalty value is assigned to the delayed part of each task, which serves as the third weight coefficient of the fitness function;

[0049] According to the priority and actual completion degree of the tasks, a priority score is calculated, which serves as the fourth weight coefficient of the fitness function;

[0050] All the weight coefficients are weighted and aggregated to generate the final fitness function.

[0051] As a preferred solution of the AI-based software project schedule arrangement and supervision system described in the present invention, wherein: the potential risks in the initial schedule are identified by integrating various data sources, and the specific steps are as follows,

[0052] The time series features of the task execution status are extracted by a pre-trained long short-term memory network, the resource dependency relationship is analyzed by a pre-trained graph neural network, and the project progress features are extracted by a self-attention mechanism;

[0053] A multi-modal fusion neural network is used to weight and merge various features and their dependencies to generate a comprehensive feature representation;

[0054] A multi-layer perceptron is used to receive the comprehensive feature representation, and perform non-linear transformation and processing to output the risk scores of tasks and projects;

[0055] The output process is trained using historical project data, and the multi-layer perceptron layer is optimized by a cross-entropy loss function;

[0056] The task progress is tracked in real time, and the delay risk and resource bottlenecks are detected in combination with the predicted risk scores.

[0057] As a preferred solution of the AI-based software project schedule arrangement and supervision system described in the present invention, wherein: the time series features of the task execution status are extracted by a pre-trained long short-term memory network, and the specific steps are as follows,

[0058] Based on the time series data of the historical task execution status, after standardization and sliding window segmentation, a self-supervised pre-training task is constructed by randomly masking some time steps;

[0059] Taking the masked time step prediction as a self-supervised task, a bidirectional long short-term memory network is constructed, the masked time series is input, and the numerical prediction of the masked position is output;

[0060] Pre-training is carried out through the MSE loss function and the Adam optimizer to learn the temporal law of the task status;

[0061] Remove the regression layer of the pre-trained long short-term memory network, freeze the parameters of the first layer and fine-tune the last layer, and extract the hidden state of the last layer as the time series feature of the current task.

[0062] As a preferred solution of the AI-based software project schedule and supervision system described in the present invention, wherein: the steps of generating multiple candidate schedule plans and selecting the optimal schedule plan are as follows.

[0063] Use the generator to generate the probability matrix of task order and resource allocation, and the discriminator evaluates the feasibility of the initial schedule plan and predicts multi-objective indicators, and outputs multiple candidate schedule plans.

[0064] Perform non-dominated sorting on the candidate solutions through the NSGA-II algorithm, and retain the Pareto front solution set with the best comprehensive performance in terms of construction period, cost, resource utilization rate and risk score.

[0065] Adopt Monte Carlo simulation to inject task time fluctuations and resource changes, and calculate the robustness score and resource flexibility index of the candidate solutions.

[0066] Combine mixed integer programming and visualization interface to select the final schedule plan based on multi-objective weights and constraint penalty coefficients.

[0067] The beneficial effects of the present invention are as follows: By combining reinforcement learning and simulated annealing algorithms to optimize the initial schedule, a preliminary plan that takes into account timeliness and resource efficiency is generated, providing a reliable basis for subsequent dynamic adjustment. The adaptive network algorithm is used to analyze the resource competition relationship between tasks in real time, dynamically adjust the resource allocation strategy, and maximize the project execution efficiency and resource utilization rate. Unsupervised learning algorithms are used to track the project progress and detect anomalies, identify potential task delays and resource bottlenecks in advance, and reduce project risks. Based on the genetic algorithm, dynamically adjust the task priorities and execution order to minimize project delays and resource waste. Through multi-modal data fusion technology, integrate various data sources, accurately identify potential risks in the initial schedule, and enhance the robustness of the plan. Use the generative adversarial network to generate multiple candidate solutions, and combine non-dominated sorting and Monte Carlo simulation to screen the optimal schedule to ensure that the project achieves an optimal balance among construction period, cost and risk. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0069] Figure 1 It is a module diagram of the AI-based software project schedule and supervision system in the embodiment.

[0070] Figure 2 It is a schematic diagram of the function of the anomaly detection module in the embodiment. Specific implementation manners

[0071] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the specific implementation manners of the present invention will be described in detail below with reference to the accompanying drawings of the specification.

[0072] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0073] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure or characteristic that may be included in at least one implementation manner of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that exclude each other from other embodiments.

[0074] Referring to Figure 1 and Figure 2 , which is an embodiment of the present invention. This embodiment provides an AI-based software project schedule and supervision system, including the following steps:

[0075] A schedule optimization module that optimizes the initial schedule of the project by combining reinforcement learning and simulated annealing algorithm, comprehensively considers multi-dimensional factors between tasks, and generates a preliminary schedule plan.

[0076] Specifically, it includes the following steps:

[0077] Receive the requirement information and resource data related to the software project, clean and standardize them, and generate a structured project requirement database;

[0078] Specifically, collect the requirement information related to the project from project management tools, documents and various interfaces, including detailed descriptions of tasks (including workload, priority, dependencies); start and end dates of tasks; resource data (including developers, hardware resources, software tools, etc.).

[0079] Data cleaning includes: cleaning the collected raw data (requirement information and resource data related to the software project) to ensure that the data is complete and non-duplicate; performing text preprocessing on the descriptions of tasks, including removing redundant information, standardizing formats, unifying date formats, etc.; checking whether the dependencies and constraints between tasks are consistent, and eliminating contradictions in the raw data.

[0080] Standardization and structuring include: using standard methods to convert descriptions of tasks, resource information, etc. into structured data formats to ensure fast reading and processing of raw data; numericalizing information such as resource allocation, task priorities, workloads, etc. and converting them into structured data that can be directly processed by machine learning algorithms.

[0081] Extract multi-dimensional features for each task, perform feature normalization and standardization on different types of features, and output multi-dimensional factors of project time, resources, and risks.

[0082] It should be noted that for the time features of tasks (such as start time and end time), the min-max normalization method is used to scale their value ranges to the interval [0,1] to eliminate differences in the time scales of different tasks and ensure that the time features of all tasks are compared under the same standard. For the resource requirements and risk factors of tasks, the Z-score standardization method is adopted to calculate the mean and standard deviation of each feature and convert the feature values into a standard normal distribution with a mean of 0 and a standard deviation of 1 to ensure the consistency of the scales of the feature values.

[0083] Use the Q-learning reinforcement learning algorithm, take task scheduling as the agent's action, define the state space of the project and the action space of the project, and establish a scheduling decision model.

[0084] Among them, the state space of the project includes the start time, end time, resource usage of tasks, and task dependencies; the action space of the project consists of the scheduling choices of tasks. For each state, the agent selects the execution order of tasks and resource allocation.

[0085] It should be noted that in the reinforcement learning framework, "state" refers to the complete set of information about the project execution environment at a certain moment, which reflects the progress of the current task and the resource allocation situation. The state is composed of multiple dimensions, including the time progress of tasks, the usage of resources, the dependencies between tasks, and potential risk indicators, etc. When the project progresses to a specific moment, this information will be captured and form a state snapshot, which not only contains the states of completed tasks but also reflects the detailed information of ongoing tasks and tasks to be started. Based on this state, the agent will evaluate the current environmental conditions and select the optimal action according to the preset goals and constraints, such as adjusting the execution order of tasks or reallocating resources. As the project progresses, the state will be continuously updated, and the agent will also make corresponding decisions according to the new state, thus forming a dynamic feedback loop to ensure the efficient and orderly progress of the project.

[0086] Example illustration: When it is detected that "the front - end development task is over - allocated (status dimension 1) and the dependent back - end interface is delayed (status dimension 3)", the current state will trigger the agent to generate a series of schedule optimization actions. For example, adjust the task order, prioritize the processing of independent modules (action 1), and at the same time re - allocate the full - stack engineer resources (action 2). The generation of these actions is based on a comprehensive assessment of the current state, ensuring more reasonable resource utilization and more efficient task execution. Through this dynamic response mechanism, the agent can optimize the project schedule in real - time according to environmental changes, thus effectively coping with problems such as uneven resource allocation and task - dependency delays, and promoting the smooth progress of the project.

[0087] Furthermore, the start time and end time of the task: can be represented by specific time points on the timeline.

[0088] Resource usage: can be represented by resource occupancy rates (such as manpower, equipment, budget).

[0089] Task - dependency relationship: can be represented by a directed acyclic graph (DAG) to show the dependency relationship between tasks.

[0090] Continuously adjust the task schedule by interacting with the environment, construct a reward function, and select the optimal scheduling action;

[0091] It should be noted that in reinforcement learning, task scheduling, as the action of the agent, means that the agent affects the state of the project by choosing different scheduling schemes (i.e., actions), and then evaluates the effect of the actions through the reward function to optimize the overall scheduling scheme.

[0092] Specifically, comprehensively consider the multi - dimensional factors in the state space of the project (i.e., the start time, end time, resource usage, and task - dependency relationship of the task), give positive rewards for timeliness and resource - utilization efficiency, and negative rewards for delays and resource waste to construct the reward function; the agent executes an action (selects a scheduling scheme) in the current state, and adjusts the scheduling scheme according to the reward value and state transition feedback by the environment (the process in which the environment transfers from the current state to the next state after the agent executes a certain action); through the Q - learning reinforcement learning algorithm, update the Q - value combined with the reward value feedback by the environment, select the optimal scheduling action, and continuously optimize the scheduling scheme.

[0093] On the basis of reinforcement learning, use the simulated annealing algorithm to further optimize the scheduling scheme;

[0094] Preferably, the simulated annealing algorithm can effectively avoid local optimal solutions through its global search ability, improving the global optimization level of the scheduling plan. Due to its randomness, the simulated annealing algorithm allows a certain degree of non-optimal solutions during the search process, enabling it to break out of the trap of local optimality and gradually approach the global optimal solution. This can not only enhance the flexibility of task execution order and resource allocation but also improve the utilization efficiency of resources, ensuring that the scheduling plan has higher adaptability in the face of dynamic changes.

[0095] Generate a neighborhood plan by adjusting the task order and resource allocation, and use the acceptance criterion to update the latest scheduling plan to avoid local optimality;

[0096] Among them, the acceptance criterion refers to the standard used to determine whether to accept a new solution when generating a neighborhood solution. In the simulated annealing algorithm, the role of the acceptance criterion is particularly important.

[0097] Specifically, the acceptance criterion is used to judge whether to accept the generated neighborhood plan (i.e., the adjusted scheduling plan), even if the objective function value (such as project delay time, resource waste, etc.) of this new plan is not better than the current plan. The core role of the acceptance criterion is to avoid falling into local optimal solutions during the neighborhood search process, increase the breadth of the search space, and enhance the opportunity for global optimization.

[0098] Comprehensively consider multi-dimensional factors, further comprehensively evaluate the updated scheduling plan and adjust the plan in real time, and finally generate a preliminary scheduling plan;

[0099] It should be noted that when comprehensively considering multi-dimensional factors such as the time, resources, and risks of the project, it is first necessary to define the evaluation indicators for each dimension, as follows: Time factor: By monitoring the gap between the actual completion time and the scheduled time of the task, evaluate the deviation of the task progress; if the task is delayed, the overall progress of the project will also be affected, so the time factor has an important impact on scheduling optimization; Resource factor: Includes the allocation efficiency of resources to ensure that each task can obtain the required resources (such as manpower, equipment, etc.); The resource utilization efficiency directly affects the execution speed of tasks and the overall efficiency of the project; Risk factor: Evaluate potential risks that may cause the project progress to deviate, such as resource bottlenecks, task dependency conflicts, unexpected problems, etc.; Through risk analysis, identify high-risk tasks and take priority treatment measures.

[0100] Furthermore, by quantifying the time factor, resource factor, and risk factor, using the weighted average method or other optimization methods, establish a comprehensive evaluation model, and output the evaluation results of the three dimensions of time, resources, and risks as a combined score. This combined score reflects the comprehensive performance of the current scheduling plan and provides a basis for subsequent optimization.

[0101] The resource scheduling module optimizes resource scheduling using an adaptive network algorithm, analyzes the resource competition situation among tasks in real time, dynamically adjusts the resource allocation plan, and maximizes the execution efficiency of the project.

[0102] Specifically, it includes the following steps:

[0103] Extract the task resource requirements in the preliminary schedule plan, analyze the resource competition relationships among tasks, and construct a resource competition graph;

[0104] It should be noted that the preliminary schedule plan is usually presented in the form of Excel, CSV format, or Gantt charts in project management tools, etc. If it is an electronic spreadsheet (such as Excel), the tasks and their resource requirements can be directly extracted from the spreadsheet through a data extraction tool (such as the Pandas library in Python). If it is a Gantt chart or a schedule plan in a project management tool, the resource requirement information of the tasks needs to be obtained through an API interface (for example, using the Microsoft Project API or Asana API).

[0105] For example, assume that Task A requires 3 developers, Task B requires 2 test devices, Task C requires 1 test device and 2 engineers, and Task D requires 1 test device. Then, analyze the resource competition relationships among the tasks. For example, Task A and Task B overlap in time and both require test devices, so there is a resource competition between them. Further construct a resource competition graph with Tasks A, B, C, and D as nodes, and the edges between them represent their resource competition relationships. In this way, it can be clearly identified which tasks have resource conflicts in the same time period, providing a basis for subsequent resource scheduling and optimization.

[0106] It should also be noted that the concept of task resource requirements comes from the resource requirement information clearly recorded in the preliminary schedule plan. Whether it is through data extraction from electronic spreadsheets (such as Excel, CSV) or obtaining the task resource allocation details through project management tools (such as Gantt charts or API interfaces), specific definitions and quantitative information on the resources required for each task (such as manpower, equipment, time, etc.) are provided.

[0107] Use a pre-trained graph neural network to model the resource sharing and dependency relationships among tasks, and combine a deep reinforcement learning algorithm to optimize the resource allocation strategy of tasks in real time, improving resource utilization efficiency;

[0108] Specifically, the specific steps for pre-training a graph neural network (GNN) are as follows: First, construct a task relationship graph using historical data in a structured project requirements database, where nodes represent tasks and edges represent resource sharing or dependency relationships between tasks. Next, perform multi-dimensional feature extraction on each task, including factors such as time, resources, and risks, and through feature normalization and standardization processing, ensure the unity of the input data. Then, use a graph convolutional network (GCN) as the GNN model, and through multi-layer information transfer and aggregation, learn the complex relationship patterns between tasks. The training objective is to enable the GNN to accurately model the resource sharing and dependency relationships between tasks.

[0109] Furthermore, during the training process, use the Adam optimizer, set the learning rate to 0.001, and use the mean squared error (MSE) as the loss function. Gradually optimize the GNN model parameters through backpropagation and gradient descent. After training is completed, save the pre-trained GNN model and use it in subsequent steps for real-time modeling of the relationships between tasks.

[0110] Monitor the task resource usage in real time, combine with a long short-term memory network (LSTM) to predict the task progress, and adjust the resource allocation, dynamically adjusting the resources according to the progress feedback and prediction results.

[0111] Specifically, by deploying a resource monitoring agent, collect the CPU, memory, disk I / O, and network bandwidth occupancy rates of task nodes at a second-level frequency and input them into a pre-trained long short-term memory network (LSTM) model (the input features include historical resource sequences, task progress curves, and external interference factors). The pre-trained long short-term memory network (LSTM) model outputs the progress deviation probability within the next 1 hour and the predicted value of the key resource bottleneck; based on the difference between the real-time progress and the prediction result, combined with a preset resource elasticity threshold (such as the CPU continuously exceeding 80% for more than 5 minutes), use a dynamic priority scheduling algorithm (such as an improved strategy based on weighted shortest job first) to automatically generate resource reallocation instructions, and at the same time feedback the adjusted actual effect data to the LSTM model for online parameter fine-tuning to form a closed-loop optimization.

[0112] Anomaly detection module, track the project progress, and perform anomaly detection through unsupervised learning algorithms to identify potential task delays and resource bottlenecks in advance.

[0113] Specifically, it includes the following steps:

[0114] During the process of dynamically adjusting resources, record the task execution status data in real time, track the task progress, resource consumption, and the deviation between the estimated and actual completion times, and establish a dynamically updated project progress database;

[0115] Among them, the task execution status data includes the basic information of the task (such as task identifier, type), progress-related information (progress percentage, completion status of key milestones), resource consumption information (CPU, memory, disk I / O, network bandwidth occupancy), time information (start, estimated and actual completion time), resource allocation details (allocation and usage status of various resources), task dependency relationship status (completion status of dependent tasks), and exception information (errors, exceptions, and resource competition conflicts).

[0116] Extract time series features, resource consumption features, and task dependency relationships from the project progress data, and perform normalization (such as linear normalization) and standardization processing (Z-score standardization);

[0117] Specifically, the following are the specific methods for extracting time series features, resource consumption features, and task dependency relationships: For time series feature extraction, it will be extracted through the time interval calculation method, sliding window analysis method, and trend fitting method; for resource consumption feature extraction, it will be extracted through statistical analysis method and correlation analysis method; for task dependency relationship extraction, it will be extracted through database query method, topological sorting method, and weight assignment method.

[0118] Use the Isolation Forest algorithm to train the processed task progress data, identify potential patterns of progress anomalies and resource consumption anomalies, and construct a detection model for task delays and resource bottlenecks;

[0119] Specifically, divide the feature dataset into a training set and a test set, usually divided according to a certain ratio (such as 8:2); then, initialize the relevant parameters of the Isolation Forest algorithm, such as the number of trees, subsample size, etc., where the selection of relevant parameters can be adjusted and optimized according to the data characteristics and actual requirements; then, input the training set data into the Isolation Forest algorithm for detection model training, and the Isolation Forest algorithm will automatically learn the normal patterns in the training set data, construct multiple decision trees by randomly selecting features and split values, and calculate the anomaly score of each data point; after training, use the test set data to verify and evaluate the trained detection model, calculate evaluation metrics such as the accuracy, recall rate, and F1 value of the detection model, and fine-tune the detection model parameters according to the evaluation results until the performance of the detection model reaches a satisfactory effect, forming a detection model for task delays and resource bottlenecks.

[0120] Evaluate the anomaly score of the task in real time every time the task progress is updated, and identify the risks of delays and resource bottlenecks.

[0121] Give priority to adjusting the module, dynamically adjust the initial schedule of the project based on the genetic algorithm, reorder the priorities and execution orders of tasks, and minimize project delays and resource waste.

[0122] Specifically, it includes the following steps:

[0123] Create a genetic algorithm environment, determine the individual coding method, and generate an initial population;

[0124] It should be noted that when creating a genetic algorithm environment, it is first necessary to clarify the specific problem to be solved, such as project schedule optimization problem, determine the objective function (such as minimizing project delay time, minimizing resource waste, etc.) and various constraints (such as task dependencies, resource limitations, etc.). This is the basis for all subsequent operations. Only after clearly understanding the nature and requirements of the problem can coding and population generation be better carried out.

[0125] Each individual expresses task sequencing and resource allocation through gene coding;

[0126] Based on project progress, resource utilization rate, delay penalty, and task priority, evaluate the quality of each scheduling plan and define the fitness function. The specific steps are as follows:

[0127] Collect the task data of the project and perform cleaning and standardization processing;

[0128] Specifically, after collecting the task data of the project, first clean the data, including deleting duplicate records (such as duplicate task IDs or exactly the same task descriptions), filling in missing values (such as using forward filling or interpolation method when the task start time is missing), and correcting outliers (such as adjusting according to historical data or domain experience when the task duration exceeds the preset threshold); then perform standardization processing on the cleaned data, including unifying the time format (such as converting the task start time to the ISO 8601 standard), normalizing numerical fields (such as mapping the task priority to a value between 0 and 1), and encoding categorical variables (such as converting the task status 'in progress' 'completed' to One-Hot encoding), and finally generate a structured task dataset to provide high-quality input for subsequent analysis.

[0129] Extract numerical features from the processed task data as the pre-input of the fitness function;

[0130] Calculate the project progress based on the proportion of completed tasks and the total number of tasks as the first weight coefficient of the fitness function;

[0131] Calculate the ratio of actual resource consumption to total resources to evaluate the resource utilization rate as the second weight coefficient of the fitness function;

[0132] Calculate the difference between the actual completion time and the planned completion time of the task, and assign a penalty value to the delayed part of each task as the third weight coefficient of the fitness function;

[0133] Specifically, the process of assigning penalty values to each task delay part is as follows: First, calculate the difference between the actual completion time and the planned completion time of the task to obtain the delay duration (for example, if task A is planned to be completed in 5 days and actually takes 7 days, the delay duration is 2 days); then, according to the preset penalty rule (such as the penalty value for each day of delay is 0.1), multiply the delay duration by the penalty coefficient to obtain the penalty value for each task (for example, if task A is delayed for 2 days, the penalty value is 0.2); finally, use the penalty value as the third weight coefficient of the fitness function to evaluate the overall performance of the task scheduling plan. The higher the penalty value, the lower the score of the plan.

[0134] Calculate the priority score according to the priority and actual completion degree of the task, as the fourth weight coefficient of the fitness function;

[0135] Perform weighted aggregation on all weight coefficients to generate the final fitness function.

[0136] Preferably, the fitness function comprehensively considers four key dimensions: project progress, resource utilization rate, delay penalty, and task priority. It avoids the imbalance problems caused by only focusing on a single factor in traditional methods, such as only paying attention to progress while ignoring resource waste, or only emphasizing resource utilization but resulting in project delays. By comprehensively considering these four dimensions, it can more comprehensively and objectively evaluate the quality of each scheduling plan, ensuring that the selected plan performs well in multiple aspects.

[0137] Use the tournament selection method to select individuals with higher fitness from the current initial population as parent individuals to participate in crossover and mutation operations;

[0138] Generate new scheduling plans by exchanging the genes of parent individuals through crossover operations;

[0139] Specifically, randomly select two parent individuals from the current population. These two parent individuals should meet certain selection strategies, such as selection based on fitness values (individuals with higher fitness values have a greater probability of being selected), to ensure that the selected parent individuals have good gene characteristics. Assume that the two selected parent individuals are Parent1 and Parent2, which represent two different task scheduling plans respectively. The gene encoding of each individual represents the arrangement order of tasks or the encoding information of tasks and other related attributes.

[0140] Furthermore, determine the type and position of the crossover point of the crossover operation. Common types of crossover operations include single-point crossover, multi-point crossover, uniform crossover, etc. Taking single-point crossover as an example, randomly select a crossover point position in the gene sequences of Parent1 and Parent2. This position divides the gene sequence into two parts.

[0141] For example, if the gene encoding of Parent1 is [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], the gene encoding of Parent2 is [10, 9, 8, 7, 6, 5, 4, 3, 2, 1], and the randomly selected crossover point position is the 5th gene position, then Parent1 is split into two parts: [1, 2, 3, 4, 5] and [6, 7, 8, 9, 10], and Parent2 is correspondingly split into [10, 9, 8, 7, 6] and [5, 4, 3, 2, 1].

[0142] Next, a gene exchange operation is performed. The latter half of the genes of Parent1 is exchanged with the first half of the genes of Parent2, and at the same time, the latter half of the genes of Parent2 is exchanged with the first half of the genes of Parent1, thereby generating two new individuals, namely offspring individuals. According to the above example, the gene encoding of the generated offspring individual Child1 is [1, 2, 3, 4, 5, 5, 4, 3, 2, 1], and the gene encoding of the offspring individual Child2 is [10, 9, 8, 7, 6, 6, 7, 8, 9, 10].

[0143] It should be noted that after generating the offspring individuals, it is necessary to check and repair the genes of the offspring individuals. Since the crossover operation may cause the gene encoding to not meet the constraints of the problem, such as the destruction of dependencies between tasks and unreasonable resource allocation. For the gene parts that do not meet the constraint conditions, adjustments are made according to specific constraint conditions and repair strategies. For example, if there is a sequential dependency between tasks, and the execution order of a certain task after crossover violates the dependency, then the order of this task and its subsequent related tasks can be adjusted to repair this violation of the constraint, ensuring that the generated offspring individuals are legal and effective scheduling plans at the gene level.

[0144] Finally, the newly generated legal offspring individuals are added to the population, replacing the individuals with lower fitness values in the original population to keep the population size unchanged and update the genetic diversity of the population. In this way, through the crossover operation, the genes of the parent individuals are successfully exchanged, generating a new scheduling plan, providing more possibilities for the subsequent iterative optimization of the genetic algorithm.

[0145] Perform fitness evaluation on each generation of individuals and select the optimal solution;

[0146] Monitor the task execution situation in real time according to the project progress, dynamically adjust the scheduling plan, and if there are delays or resource bottlenecks, use the genetic algorithm to further optimize the scheduling.

[0147] The risk prediction module uses multi-modal data fusion to predict the risks of the project progress and identifies potential risks in the initial scheduling by integrating various data sources.

[0148] Specifically, it includes the following steps:

[0149] Extract the time series features of the task execution status through a pre-trained long short-term memory network, analyze the resource dependency relationship using a pre-trained graph neural network, and extract the project progress features through a self-attention mechanism;

[0150] Among them, the task execution status refers to the actual dynamic performance of the task in the project, including the progress in the time dimension (such as the completion time points of each stage), the actual completion degree (compared with the plan), the dependency relationship with other tasks (such as whether the previous task is completed), and the resource consumption situation (whether the use of manpower and material resources is reasonable). These task execution status data are the basis for subsequent feature extraction (such as LSTM and GCN modeling), and directly affect the prediction ability of task delay risk and resource bottlenecks.

[0151] Extract the time series features of the task execution status through a pre-trained long short-term memory network. The specific steps are as follows:

[0152] Based on the time series data of historical task execution status, after standardization (Min-Max) and sliding window segmentation (window length 30 days, step size 1 day), randomly mask part of the time (such as masking 15% of the time steps) to construct a self-supervised pre-training task;

[0153] Take the masked time step prediction as a self-supervised task, construct a bidirectional long short-term memory network, input the masked time series, and output the numerical prediction of the masked position;

[0154] Perform pre-training through the MSE loss function and the Adam optimizer (learning rate 1e-3) to learn the temporal law of the task state;

[0155] Remove the regression layer of the pre-trained long short-term memory network, freeze the parameters of the first layer and fine-tune the last layer (learning rate 1e-4), and extract the last layer hidden state (dimension 256) as the time series feature of the current task.

[0156] Use a multi-modal fusion neural network to weighted merge various features and their dependencies to generate a comprehensive feature representation;

[0157] Adopt a multi-layer perceptron (MLP) to receive the comprehensive feature representation, perform non-linear transformation and processing, and output the risk scores of the task and the project, indicating potential delays;

[0158] Preferably, the MLP can learn the complex relationships between input features by stacking multiple fully connected layers and non-linear activation functions (such as ReLU and Sigmoid), and is suitable for processing the high-dimensional features after fusion.

[0159] For example, by constructing a multi-layer perceptron (MLP) with two hidden layers, where the number of neurons in each layer is 128 and 64 respectively, a non-linear mapping is achieved through the ReLU activation function.

[0160] Use historical project data to train the output process and optimize the multi-layer perceptron layer through the cross-entropy loss function;

[0161] Track the task progress in real-time, and detect delay risks and resource bottlenecks by combining the predicted risk scores;

[0162] The scheme optimization module finally simulates and optimizes the initial scheduling scheme through a generative adversarial network, generates multiple candidate scheduling schemes, and selects the optimal scheduling scheme.

[0163] Specifically, it includes the following steps:

[0164] Use the generator to generate the probability matrix of task sequence and resource allocation, and the discriminator evaluates the feasibility of the initial scheduling scheme and predicts multi-objective indicators, and outputs multiple candidate scheduling schemes;

[0165] Specifically, the architecture design of the generative adversarial network (GAN) is as follows:

[0166] Generator: Input: Task dependency graph embedding vector + Resource constraint condition encoding; Network structure: Transformer decoder layer (self-attention mechanism generates the probability distribution of task sequences); Output: Gene encoding of the candidate scheme (task sequence permutation matrix + resource allocation probability table).

[0167] Discriminator: Input: Gene encoding of the candidate scheme; Network structure: Graph convolution branch: Analyze resource dependency relationships, Temporal convolution branch: Predict project duration and cost; Output: Scheme feasibility probability (0 - 1) + Multi-objective prediction values (project duration / cost / risk).

[0168] Perform non-dominated sorting on the candidate schemes through the NSGA-II algorithm, and retain the Pareto front solution set with the best comprehensive performance in terms of project duration, cost, resource utilization rate, and risk score;

[0169] Specifically, non-dominated sorting specifically includes the first-layer screening: Eliminate the schemes that violate hard constraints (resource overrun, dependency break); The second-layer sorting: Calculate the dominance relationship based on the objective vector and divide the Pareto levels.

[0170] Adopt Monte Carlo simulation to inject task time fluctuations and resource changes, and calculate the robustness score and resource elasticity index of the candidate schemes;

[0171] Specifically, the task timing, resource allocation and dependency data of candidate solutions are first loaded from the Pareto frontier solution set generated by the previous multi-objective optimization to build a simulation input template; then, an execution time fluctuation model is defined for each task (based on Beta distribution, parameters α=2, β=5), and availability change rules are configured for resource nodes (Poisson process trigger, λ=0.1 / hour), and a random scenario set containing task time offsets and resource interruption events is generated (1000 samplings by default); then, the candidate solutions are combined with random disturbance scenarios, and the task execution process is simulated through a parallel simulation engine to track progress deviations, resource conflicts and dynamic adjustment records in real time; based on the simulation results, the task delay ratio (number of delayed tasks / total number of tasks) of each solution in all disturbance scenarios is counted to calculate the robustness score, and the resource elasticity index is calculated according to the ratio of successful adjustments after resource conflicts (number of successful adjustments / total number of conflicts); finally, dynamic thresholds of robustness score ≥ 0.7 and resource elasticity index ≥ 0.6 are set to eliminate substandard solutions and retain high-stability candidate sets for subsequent decision-making processes.

[0172] Furthermore, the role of defining the execution time fluctuation model: In project scheduling, the execution time of tasks is often not fixed and may fluctuate due to various factors (such as insufficient resources, unexpected problems, etc.). In order to simulate the project execution process more realistically, it is necessary to define an execution time fluctuation model for each task. By defining the execution time fluctuation, it is possible to predict the possible time deviation of the task in actual execution, so as to more accurately evaluate the robustness of the scheduling plan. The specific steps are as follows:

[0173] Select distribution model: Beta distribution is selected as the execution time fluctuation model. Beta distribution is a continuous probability distribution suitable for describing random variables in a finite interval. Parameters α=2 and β=5 represent the shape of the distribution. Specifically, this parameter setting means that the task execution time tends to be completed in a shorter time, but there is also a certain degree of volatility.

[0174] Generate random time offset: Based on Beta distribution, a random time offset is generated for each task. This offset represents the deviation between the actual execution time of the task and the planned time. For example, if the planned time is 10 days, the generated offset may be +2 days or -1 day, indicating that the task may be completed earlier or later.

[0175] Configure resource node availability change rules: The availability of resources (such as manpower and equipment) may also change. The document uses the Poisson process to simulate the availability changes of resource nodes. The Poisson process is a model that describes the occurrence of random events over time. The parameter λ=0.1 / hour represents the average number of resource interruptions per hour.

[0176] Generate a set of random scenarios: Combine the time offset of the task and resource interruption events to generate multiple random scenarios (1000 samplings by default). Each scenario represents a possible project execution situation, including the time deviation of tasks and resource interruption situations.

[0177] It should also be noted that a random perturbation scenario refers to a specific scenario generated after introducing random factors (such as task time fluctuations, resource interruptions, etc.) during the simulation of project execution. Each scenario represents a possible project execution situation, including the actual execution time of tasks, resource availability, etc. By generating multiple random perturbation scenarios, the performance of the scheduling plan in different situations can be evaluated more comprehensively.

[0178] Combine mixed-integer programming with a visualization interface (Gantt chart / heat map), and select the final scheduling plan based on multi-objective weights and constraint penalty coefficients.

[0179] In summary, the present invention optimizes the initial scheduling by combining reinforcement learning and simulated annealing algorithm, generates a preliminary plan that takes into account timeliness and resource efficiency, and provides a reliable basis for subsequent dynamic adjustment. The adaptive network algorithm is used to analyze the resource competition relationship between tasks in real time, dynamically adjust the resource allocation strategy, and maximize the improvement of project execution efficiency and resource utilization rate. The unsupervised learning algorithm is used to track the project progress and detect anomalies, identify potential task delays and resource bottlenecks in advance, and reduce project risks. Based on the genetic algorithm, the task priorities and execution orders are dynamically adjusted to minimize project delays and resource waste. Through multi-modal data fusion technology, various data sources are integrated to accurately identify potential risks in the initial scheduling and enhance the robustness of the plan. The generative adversarial network is used to generate multiple candidate plans, and the non-dominated sorting and Monte Carlo simulation are combined to screen the optimal scheduling to ensure an overall optimal balance among project duration, cost, and risk.

[0180] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An AI-based software project scheduling and supervision system, characterized by: include, The scheduling optimization module uses a combination of reinforcement learning and simulated annealing algorithms to optimize the initial scheduling of projects, comprehensively considers multi-dimensional factors between tasks, and generates a preliminary scheduling plan; The resource scheduling module uses an adaptive network algorithm to optimize resource scheduling, analyze resource competition between tasks in real time, and dynamically adjust resource allocation plans to maximize project execution efficiency; Anomaly detection module tracks project progress and detects anomalies through unsupervised learning algorithms to identify potential task delays and resource bottlenecks in advance; Priority adjustment module, which dynamically adjusts the initial schedule of the project based on genetic algorithms, reorders the priority and execution order of tasks, and minimizes project delays and resource waste; The risk prediction module uses multimodal data fusion to predict the risk of project progress and identifies potential risks in the initial schedule by integrating various data sources; The solution optimization module uses the generative adversarial network to perform final simulation and optimization on the initial scheduling solution, generate multiple candidate scheduling solutions, and select the optimal scheduling solution; The specific steps of generating a preliminary scheduling plan are as follows: Receive demand information and resource data related to software projects, clean and standardize them, and generate a structured project demand database; Perform multi-dimensional feature extraction on each task, normalize and standardize different types of features, and output multi-dimensional factors of project time, resources, and risks; Use the Q-learning reinforcement learning algorithm to take task scheduling as agent action, define the state space and action space of the project, and establish a scheduling decision model; Continuously adjust task scheduling through interaction with the environment, build reward functions, and select the optimal scheduling action; Based on reinforcement learning, the simulated annealing algorithm is used to further optimize the scheduling plan; Generate neighborhood solutions by adjusting task order and resource allocation, and use acceptance criteria to update the latest scheduling solution to avoid local optimality; Taking into account multi-dimensional factors, further comprehensively evaluate the updated scheduling plan and adjust the plan in real time to finally generate a preliminary scheduling plan; The state space of the project includes the start time, end time, resource usage and task dependencies of the task, and the action space of the project consists of the scheduling selection of the task. For each state, the agent selects the execution order of the task and the resource allocation; The specific steps of selecting the optimal scheduling action are as follows: Based on the state space of the project, a reward function is constructed by giving positive rewards for timeliness and resource efficiency, and negative rewards for delays and resource waste; The agent performs actions in the current state and adjusts the scheduling plan based on the reward value and state transition fed back by the environment; Through the Q-learning reinforcement learning algorithm, the Q value is updated in combination with the reward value of environmental feedback, the optimal scheduling action is selected, and the scheduling plan is continuously optimized.

2. The AI-based software project scheduling and monitoring system according to claim 1, characterized in that: The specific steps to maximize the efficiency of project execution are as follows: Extract the task resource requirements in the preliminary scheduling plan, analyze the resource competition relationship between tasks, and construct a resource competition diagram; Use pre-trained graph neural networks to model resource sharing and dependencies between tasks, and combine deep reinforcement learning algorithms to optimize the resource allocation strategy of tasks in real time; Monitor task resource usage in real time, predict task progress based on pre-trained long short-term memory networks, and adjust resource allocation, dynamically adjusting resources based on progress feedback and prediction results.

3. The AI-based software project scheduling and monitoring system according to claim 2, characterized in that: The specific steps for identifying potential task delays and resource bottlenecks in advance are as follows: In the process of dynamically adjusting resources, the task execution status data is recorded in real time, the task progress, resource consumption, deviation between the estimated and actual completion time are tracked, and a dynamically updated project progress database is established; Extract time series features, resource consumption features and task dependencies from the project progress database, perform normalization and standardization, and output feature data sets; Use the Isolation Forest algorithm to train the feature data set, identify potential patterns of abnormal progress and abnormal resource consumption, and build a detection model for task delays and resource bottlenecks; Each time the task progress is updated, the task abnormality score is evaluated in real time through the task delay and resource bottleneck detection model to identify the risks of delays and resource bottlenecks.

4. The AI-based software project scheduling and monitoring system according to claim 3, characterized in that: The specific steps to minimize project delays and resource waste are as follows: Create a genetic algorithm environment, determine the individual encoding method, and generate the initial population; Each individual expresses task sequencing and resource allocation through genetic coding; Evaluate the quality of each scheduling solution and define a fitness function based on project progress, resource utilization, delay penalties, and task priority; Use the tournament selection method to select individuals with higher fitness from the current initial population as parent individuals to participate in crossover and mutation operations; The genes of the parent individuals are exchanged through crossover operation to generate a new scheduling plan; Evaluate the fitness of each generation of individuals and select the best solution; Monitor task execution in real time based on project progress, dynamically adjust scheduling plans, and use genetic algorithms to further optimize scheduling if delays or resource bottlenecks occur.

5. The AI-based software project scheduling and monitoring system according to claim 4, characterized in that: The specific steps of evaluating the quality of each scheduling plan and defining the fitness function are as follows: Collect project task data, clean and standardize it; Extract numerical features from the processed task data as pre-input of the fitness function; According to the proportion of completed tasks and the total number of tasks, the project progress is calculated as the first weight coefficient of the fitness function; Calculate the ratio of actual resource consumption to total resources and evaluate resource utilization as the second weight coefficient of the fitness function; By calculating the difference between the actual completion time and the planned completion time of the task, a penalty value is assigned to each delayed part of the task as the third weight coefficient of the fitness function; According to the priority and actual completion of the task, the priority score is calculated as the fourth weight coefficient of the fitness function; All weight coefficients are weighted and summarized to generate the final fitness function.

6. The AI-based software project scheduling and monitoring system according to claim 5, characterized in that: The specific steps of identifying potential risks in initial scheduling by integrating various data sources are as follows: The time series features of task execution status are extracted through pre-trained long short-term memory networks, resource dependencies are analyzed using pre-trained graph neural networks, and project progress features are extracted through self-attention mechanisms; Use a multimodal fusion neural network to weight and merge various features and their dependencies to generate a comprehensive feature representation; A multi-layer perceptron is used to receive comprehensive feature representations, perform nonlinear transformation and processing, and output risk scores for tasks and projects; The output process is trained using historical project data and the multi-layer perceptron layer is optimized using the cross entropy loss function; Track task progress in real time and detect delay risks and resource bottlenecks using predicted risk scores.

7. The AI-based software project scheduling and monitoring system according to claim 6, characterized in that: The specific steps of extracting the time series features of the task execution status through the pre-trained long short-term memory network are as follows: Based on the time series data of historical task execution status, after normalization and sliding window segmentation, some time steps are randomly masked to construct a self-supervised pre-training task; Taking the masked time step prediction as a self-supervised task, a bidirectional long short-term memory network is constructed, which inputs the masked time series and outputs the numerical prediction of the masked position; Pre-training is performed through the MSE loss function and the Adam optimizer to learn the temporal rules of task states; Remove the regression layer of the pre-trained LSTM network, freeze the first layer parameters and fine-tune the last layer, and extract the hidden state of the last layer as the time series feature of the current task.

8. The AI-based software project scheduling and monitoring system according to claim 7, characterized in that: The specific steps of generating multiple candidate scheduling plans and selecting the optimal scheduling plan are as follows: The generator generates a probability matrix of task order and resource allocation, and the discriminator evaluates the feasibility of the initial scheduling plan and predicts multi-objective indicators, and outputs multiple candidate scheduling plans; The candidate solutions are non-dominatedly sorted using the NSGA-II algorithm, retaining the Pareto frontier solution set with the best combination of duration, cost, resource utilization and risk score; Monte Carlo simulation is used to inject task time fluctuations and resource changes, and the robustness scores and resource elasticity indexes of candidate solutions are calculated; Combine mixed integer programming with a visual interface to select the final scheduling solution based on multi-objective weights and constraint penalty coefficients.

Images