Intelligent project management method and system based on multi-modal big data and adaptive large model
By employing an intelligent project management approach that integrates multi-modal big data and adaptive large models, and constructing a dynamic adaptive large model framework, we have solved the problems of data silos and delayed dynamic decision-making in traditional project management. This approach enables risk warning and resource optimization, thereby improving the automation and efficiency of project management.
Patent Information
- Application Number
- CN202510619517.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-14
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional project management suffers from problems such as data silos, delayed dynamic decision-making, inefficient risk prediction, and insufficient resource optimization, making it difficult to achieve real-time perception of environmental changes and make intelligent decisions.
By employing multimodal big data and adaptive large models, integrating multi-source heterogeneous data, analyzing risk signals through graph neural networks and semantic models, constructing a dynamic adaptive large model framework, and combining multi-objective optimization functions and reinforcement learning modules, dynamic scheduling of resource allocation is achieved, and decision-making is assisted through digital twin simulation and interactive natural language interface.
It significantly improves the risk warning capability and dynamic decision-making efficiency of project management, optimizes resource utilization, reduces idle waste, improves the accuracy of risk prediction and the efficiency of resource allocation, and realizes full-process automated management.
Smart Images

Figure CN123118003A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of project management technology, specifically an intelligent project management method and system based on multimodal big data and adaptive large models. Background Technology
[0002] Intelligent project management throughout its entire lifecycle refers to the process of intelligent, automated, and refined management of a project from initiation, planning, execution, monitoring to closure, utilizing advanced information technologies such as artificial intelligence, big data, the Internet of Things, and cloud computing. By integrating various project management tools and intelligent analysis platforms, it achieves real-time perception, intelligent prediction, dynamic optimization, and precise decision-making regarding key elements such as project resources, schedule, cost, quality, and risk. The system can automatically identify potential risks, intelligently allocate tasks, optimize resource allocation, predict project trends, and provide visualized reports and intelligent suggestions, significantly improving the efficiency, transparency, and success rate of project management while reducing management costs. It is an inevitable trend in the development of modern project management.
[0003] However, traditional project management relies on human experience and static models, which has the following pain points:
[0004] 1. Data silos: The multi-source heterogeneous data (schedule, cost, resources, risks, etc.) involved in the project are scattered and isolated, making them difficult to integrate and analyze;
[0005] 2. Lagging dynamic decision-making: Relying on periodic reports and feedback, it is unable to perceive environmental changes in real time and adjust strategies accordingly;
[0006] 3. Inefficient risk prediction: Static models based on historical data are difficult to cope with sudden risks (such as supply chain disruptions and technological bottlenecks);
[0007] 4. Insufficient resource optimization: The allocation of resources such as manpower and equipment relies on experience, which can easily lead to waste or overload. Summary of the Invention
[0008] The purpose of this invention is to provide an intelligent project management method and system based on multimodal big data and adaptive large models in order to solve the problems mentioned above.
[0009] The technical solution adopted in this invention is as follows: an intelligent project management method based on multimodal big data and adaptive large models, the method comprising the following steps:
[0010] S1: Integrates multi-source heterogeneous data, collects structured data, unstructured text, audio and video recordings, and environmental sensor data, covering indicators throughout the entire project lifecycle;
[0011] S2: Perform multimodal feature extraction and standardization on the collected data, model task dependencies through graph neural networks, and use semantic models to parse risk signals in the text;
[0012] S3: Construct a dynamic adaptive large model framework, integrate incremental pre-training mechanism and reinforcement learning module, and fine-tune model parameters in real time to respond to dynamic changes in the project.
[0013] S4: Design a multi-objective optimization function that encodes cost, schedule and quality objectives into a multi-dimensional reward vector and balances resource allocation conflicts through dynamic weight adjustment;
[0014] S5: Deploy a cross-modal embedding layer and use a multi-head attention mechanism to align heterogeneous data features to achieve joint representation of task progress, team sentiment, and physical environment parameters;
[0015] S6: Introducing elastic weight solidification technology, combining historical project memory and real-time data stream to prevent the model from forgetting key historical experiences during incremental training;
[0016] S7: Establish an interactive natural language interface to support project managers in querying risks in the form of instructions and generating interpretable intervention suggestions through causal reasoning;
[0017] S8: Construct a digital twin simulation environment to simulate the impact of different decisions on the project outcome and provide visualized simulation results to assist in dynamic strategy adjustments;
[0018] S9: Implement a closed-loop feedback mechanism to feed the actual execution results back to the adaptive large model, and continuously optimize prediction accuracy and resource scheduling efficiency.
[0019] In a preferred embodiment, in step S1, when integrating multi-source heterogeneous data, it is necessary to connect to project management platforms such as JIRA and Trello through ETL tools such as Apache NiFi to collect structured data such as task progress tables and resource consumption tables. Unstructured text such as emails and Slack chat logs need to be extracted for key content using OCR and speech-to-text technology. Environmental sensor data such as temperature, humidity, and equipment status are collected in real time through the MQTT protocol. The data covers all stages of the project lifecycle, including project requirement documents, execution logs, monitoring indicators, and closing reports. Anomalies are filtered through a verification mechanism, such as invalid data with task progress exceeding 100% or negative resource usage. Data storage adopts a distributed database, and the sharding strategy is divided according to project stages to ensure high-concurrency read and write performance.
[0020] In a preferred embodiment, in step S2, during the multimodal feature extraction stage, structured data is modeled into a task dependency graph using a graph neural network. Node attributes are fused with task priority, resource occupancy rate, and text sentiment analysis results, and edge weights represent the coupling strength between tasks. Unstructured text is extracted using a BERT model to extract semantic features, identify risk keywords such as "delay" and "insufficient resources," and combine them with the TF-IDF algorithm to quantify the risk signal strength. A threshold of 0.75 or higher is set to trigger an early warning.
[0021] In a preferred embodiment, in step S3, the dynamic adaptive large model is based on the Transformer architecture, the incremental pre-training adopts a hybrid sampling strategy, the real-time data stream is updated once per hour, and the historical memory bank is stored according to scenario classification; the model fine-tuning frequency is set to once per hour, the learning rate adopts a Cosine decay strategy, the initial value is set to 3e-5, and the minimum value is 1e-6.
[0022] In a preferred embodiment, in step S4, the multi-objective optimization function encodes cost deviation, schedule delay, and quality score into a three-dimensional reward vector, and the weight coefficients are dynamically adjusted by the project manager through an interactive interface; the Pareto front search algorithm filters the non-dominated solution set and eliminates invalid strategies with resource utilization exceeding 90% or schedule delay exceeding 7 days.
[0023] In a preferred embodiment, in step S5, the cross-modal embedding layer adopts a multi-head attention mechanism with a fixed number of heads of 8, mapping the task progress values, team sentiment analysis results, and sensor time-series data to a 512-dimensional shared vector space; the pre-training of the embedding layer adopts contrastive learning, with positive samples being multimodal data pairs of the same task and negative samples being randomly combined across tasks, the loss function maximizing the cosine similarity of positive sample pairs, and the threshold being set to 0.9; the training parameters include a learning rate of 1e-4 and a batch size of 128, and the iteration stops when the loss function drops below 0.05.
[0024] In a preferred embodiment, in step S6, firstly, the parameter importance score is obtained by multiple backpropagation calculations on historical project data, marking the parameters that play a decisive role in predicting historical tasks; subsequently, a regularization term is introduced into the objective function of incremental training to limit drastic changes in these key parameters; for example, for model parameters θ, the loss function is defined as a combination of the original task loss and the historical parameter protection term, the latter being implemented through a weighted squared penalty term; through alternating sampling of real-time data streams and historical memory banks, the model can capture the dynamic features of the current project while maintaining the stability of long-term knowledge; this mechanism effectively balances the model's ability to learn new scenarios with the need to retain historical experience, significantly improving decision robustness in complex scenarios such as resource conflicts or sudden changes in demand;
[0025] The formula for the regularized loss function of elastic weight fixation is:
[0026]
[0027] Among them, L current Let θ be the loss function for the current task. i * indicates the optimal value of the parameters after training on the historical task, Fi is the parameter importance weight calculated based on the Fisher information matrix, and λ is the tradeoff coefficient; this formula prevents catastrophic forgetting by constraining the degree to which the key parameter θi deviates from the historical optimal value.
[0028] The formula for calculating the importance weights of parameters based on the Fisher information matrix is as follows:
[0029]
[0030] Specifically, N data samples (xn, yn) are sampled from the historical memory bank, and the Fisher information content of parameter θi is calculated to quantify its contribution to the prediction of historical tasks. The larger the Fi value, the more critical θi is to historical knowledge, and stronger protection needs to be applied in incremental training. These two formulas together construct a parameter dynamic protection mechanism, which enables the model to balance stability and adaptability during continuous learning.
[0031] In a preferred embodiment, in step S7, the interactive natural language interface is based on a fine-tuned GPT-4 model and supports commands such as "predict high-risk tasks next week". The intent recognition module determines the query type through keyword matching and context analysis. The causal reasoning module adopts a structural causal model to construct an event causal chain such as "requirement change → code submission decrease → test delay" and generates intervention suggestions such as "allocate backup test resources in advance".
[0032] In a preferred embodiment, in step S8, the digital twin simulation environment is built based on the Unity engine. The input parameters include resource adjustment strategies, and 1,000 decision paths are generated through Monte Carlo simulation. The simulation results output dynamic Gantt charts and resource heatmaps. Key indicators such as cost savings rate and project time reduction ratio are compared through line charts. The simulation frequency supports real-time updates, refreshed every 15 minutes. The simulation error rate is monitored through the root mean square error (RMSE), and model recalibration is triggered when it exceeds 5%.
[0033] In step S9, the closed-loop feedback mechanism compares the actual project cost, delivery quality and other data with the predicted values, calculates the mean absolute error (MAE) and root mean square error (RMSE), and sets the thresholds to 3% and 5% respectively; the online learning pipeline adopts a sliding window mechanism, retaining only the data from the most recent 3 months for training, and archiving historical data to cold storage.
[0034] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:
[0035] 1. This invention significantly improves the risk warning capability and dynamic decision-making efficiency of project management by integrating multimodal data and an adaptive learning mechanism. The system integrates multi-dimensional information such as task progress, team communication, and environmental parameters, and continuously optimizes the model using incremental training technology. This enables it to identify potential risks such as task delays or resource bottlenecks in advance and generate targeted intervention suggestions. The application of elastic weight solidification technology allows the model to retain historical experience while absorbing new data, avoiding the forgetting of key knowledge due to frequent updates, thereby enhancing the predictive stability for unexpected events. For example, the system can quickly adjust strategies when demand changes or supply chain disruptions occur, reducing the frequency of manual intervention.
[0036] 2. This invention enhances the rationality of resource allocation and global coordination capabilities through reinforcement learning and multi-objective optimization. The dynamic weight adjustment mechanism balances cost, schedule, and quality objectives, optimizing the efficiency of manpower and equipment utilization and reducing resource idleness or overload. Digital twin simulation and natural language interaction further lower the decision-making threshold, allowing project managers to intuitively assess the impact of different strategies and continuously calibrate model accuracy through closed-loop feedback. Ultimately, it achieves full-process automation from data integration to decision execution, improving project execution efficiency and success rate.
[0037] 3. In this invention, by implementing this invention, artificial intelligence is applied to the field of project management, resulting in improved technical effects: 1. Improved risk prediction accuracy: Through multimodal data correlation analysis, the accuracy of risk warning 14 days in advance is ≥85%. 2. Optimized resource utilization: Dynamic scheduling algorithms reduce idle resource waste, lowering average costs by 12%–18%. 3. Increased automation: 80% of routine decisions (such as task assignment and schedule synchronization) are completed autonomously by the system. Attached Figure Description
[0038] Figure 1 This is a schematic diagram illustrating the process principle of the present invention. Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0040] Example:
[0041] Reference Figure 1 A smart project management method based on multimodal big data and adaptive large models, the method includes the following steps:
[0042] S1: Integrates multi-source heterogeneous data, collecting structured data, unstructured text, audio and video recordings, and environmental sensor data, covering indicators throughout the entire project lifecycle.
[0043] S2: Perform multimodal feature extraction and standardization on the collected data, model task dependencies through graph neural networks, and use semantic models to analyze risk signals in the text.
[0044] S3: Construct a dynamic adaptive large model framework, integrate incremental pre-training mechanism and reinforcement learning module, and fine-tune model parameters in real time to respond to dynamic changes in the project.
[0045] S4: Design a multi-objective optimization function that encodes cost, schedule, and quality objectives into a multi-dimensional reward vector and balances resource allocation conflicts through dynamic weight adjustment.
[0046] S5: Deploy a cross-modal embedding layer and use a multi-head attention mechanism to align heterogeneous data features to achieve joint representation of task progress, team sentiment, and physical environment parameters.
[0047] S6: Introduces elastic weight solidification technology, combining historical project memory and real-time data stream to prevent the model from forgetting key historical experiences during incremental training.
[0048] S7: Establish an interactive natural language interface to support project managers in querying risks in the form of instructions and generating interpretable intervention suggestions through causal reasoning.
[0049] S8: Construct a digital twin simulation environment to simulate the impact of different decisions on the project outcome and provide visualized simulation results to assist in dynamic strategy adjustments.
[0050] S9: Implement a closed-loop feedback mechanism to feed the actual execution results back to the adaptive large model, and continuously optimize prediction accuracy and resource scheduling efficiency.
[0051] In step S1, when integrating multi-source heterogeneous data, ETL tools such as Apache NiFi are used to connect to project management platforms such as JIRA and Trello. Structured data, such as task progress tables and resource consumption tables, are collected. Unstructured text, such as emails and Slack chat logs, requires OCR and speech-to-text technology to extract key content. Environmental sensor data, including temperature, humidity, and equipment status, is collected in real-time via the MQTT protocol. The data covers all stages of the project lifecycle, including project requirement documents, execution logs, monitoring metrics, and closing reports. Anomalies, such as invalid data with task progress exceeding 100% or negative resource usage, are filtered through a validation mechanism. Data storage uses a distributed database, with sharding strategies divided according to project stages to ensure high-concurrency read / write performance.
[0052] In step S2, during the multimodal feature extraction stage, structured data is modeled using a graph neural network to create a task dependency graph. Node attributes are fused with task priority, resource occupancy, and text sentiment analysis results, with edge weights representing the coupling strength between tasks. Unstructured text is extracted using the BERT model to identify risk keywords such as "delay" and "insufficient resources," and the TF-IDF algorithm is used to quantify the risk signal strength, with a threshold of 0.75 triggering an early warning. Time-series data, such as sensor readings, are calculated using a sliding window to determine the mean, variance, and trend slope, and standardized to a unified dimension of [0,1]. The output dimension of the feature fusion layer is fixed at 256 dimensions, aligned with the input layer of the downstream model to ensure semantic consistency across modal data.
[0053] In step S3, the dynamic adaptive large model is based on the Transformer architecture. Incremental pre-training employs a hybrid sampling strategy, with real-time data streams updated hourly. Historical memory is stored categorized by scenario, such as "supply chain disruption" and "demand change" cases. The model fine-tuning frequency is set to once per hour, and the learning rate uses a Cosine decay strategy, with an initial value of 3e-5 and a minimum value of 1e-6. In the reinforcement learning module, the Actor network outputs resource allocation strategies such as the ratio of manpower increases / decreases (±20%) and priority weight adjustments (0.1–0.9). The Critic network evaluates the expected returns for the next 5 days based on the Bellman equation. Training requires an NVIDIA A100 GPU cluster, with batch size dynamically adjusted from 32 to 128, depending on data throughput.
[0054] In step S4, the multi-objective optimization function encodes cost deviation, schedule delay, and quality score into a three-dimensional reward vector. The weight coefficients are dynamically adjusted by the project manager through an interactive interface; for example, the cost weight w1 defaults to 0.5 and automatically increases to 0.8 after a 10% cost overrun. The Pareto front search algorithm filters out non-dominated solutions, eliminating invalid strategies with resource utilization exceeding 90% or schedule delays exceeding 7 days. During optimization, server load thresholds are monitored in real time; a degradation strategy is triggered when CPU utilization is ≥85%, pausing non-critical task computation. The gradient descent optimizer uses AdamW, with a fixed learning rate of 2e-4 and an upper limit of 1000 iterations or when the loss converges to 1e-3.
[0055] In step S5, the cross-modal embedding layer employs a multi-head attention mechanism with a fixed number of 8 heads, mapping task progress values, team sentiment analysis results, and sensor time-series data to a 512-dimensional shared vector space. Pre-training of the embedding layer uses contrastive learning; positive samples are multimodal data pairs from the same task, while negative samples are randomly combined across tasks. The loss function maximizes the cosine similarity of positive sample pairs, with a threshold of 0.9. Training parameters include a learning rate of 1e-4 and a batch size of 128. Iteration stops when the loss function drops below 0.05. After feature alignment, the output vector is compressed to 128 dimensions through a fully connected layer for direct use by downstream models.
[0056] In step S6, during the incremental pre-training phase, the system uses elastic weight solidification technology to evaluate the importance of model parameters, calculates the influence weight of each parameter on historical tasks, and dynamically adjusts the training loss function accordingly. Specifically, the system uses the Fisher information matrix of historical data to quantify parameter importance, locks in the weight range of key parameters, and ensures that it retains the ability to remember historical scenarios while absorbing new data. During training, a course-based learning strategy is adopted, prioritizing training from high-frequency, routine scenarios (such as task allocation and progress synchronization) and gradually introducing low-frequency, complex scenarios (such as supply chain disruption response), thereby improving the model's adaptability to unexpected events.
[0057] First, parameter importance scores are calculated through multiple backpropagations on historical project data, identifying parameters that play a decisive role in predicting historical tasks. Then, a regularization term is introduced into the objective function of incremental training to limit drastic changes in these key parameters. For example, for model parameters θ, the loss function is defined as a combination of the original task loss and a historical parameter protection term, the latter achieved through a weighted squared penalty term. By alternating sampling of real-time data streams and historical memory, the model can capture the dynamic features of the current project while maintaining stability with long-term knowledge. This mechanism effectively balances the model's ability to learn new scenarios with the need to retain historical experience, significantly improving decision robustness in complex scenarios such as resource conflicts or sudden changes in demand.
[0058] The formula for the regularized loss function of elastic weight fixation is:
[0059]
[0060] Among them, L current Let θ be the loss function for the current task. i * indicates the optimal parameter value after training on a historical task, Fi is the parameter importance weight calculated based on the Fisher information matrix, and λ is the tradeoff coefficient. This formula prevents catastrophic forgetting by constraining the degree to which the key parameter θi deviates from the historical optimal value.
[0061] The formula for calculating the importance weights of parameters based on the Fisher information matrix is as follows:
[0062]
[0063] This process involves sampling N data samples (xn, yn) from the historical memory bank, calculating the Fisher information content of parameter θi, and quantifying its contribution to the prediction of historical tasks. A larger Fi value indicates that θi is more critical to historical knowledge, requiring stronger protection during incremental training. These two formulas together construct a dynamic parameter protection mechanism, enabling the model to balance stability and adaptability during continuous learning.
[0064] In step S7, the interactive natural language interface, based on a fine-tuned GPT-4 model, supports commands such as "predict high-risk tasks next week." The intent recognition module determines the query type through keyword matching and context analysis. The causal reasoning module uses a structural causal model to construct event causal chains such as "requirement change → decreased code submissions → test delay," and generates intervention suggestions such as "allocate backup test resources in advance." The output results include a confidence score, with a threshold of 80% or higher marked as highly reliable. It also displays related log snippets and historical similar cases, such as "a similar change in Q2 2024 caused a 3-day delay."
[0065] In step S8, the digital twin simulation environment is built based on the Unity engine. Input parameters include resource adjustment strategies such as increasing the number of developers by 10% and raising the priority weight to 0.8. 1000 decision paths are generated through Monte Carlo simulation. The simulation results output a dynamic Gantt chart and a resource heatmap. Key indicators such as cost savings rate and project time reduction ratio are compared using line charts. The simulation frequency supports real-time updates, refreshing every 15 minutes. The simulation error rate is monitored using the root mean square error (RMSE), and model recalibration is triggered when it exceeds 5%. The visualization charts support interactive zooming, and critical path tasks are highlighted in red.
[0066] In step S9, the closed-loop feedback mechanism compares actual project costs, delivery quality, and other data with predicted values to calculate the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), with thresholds set at 3% and 5%, respectively. The online learning pipeline uses a sliding window mechanism, retaining only data from the most recent three months for training, while historical data is archived to cold storage. The feedback frequency is set to execute automatically at 2 AM daily. Key parameters, such as the regularization coefficient λ, are dynamically adjusted based on error indicators, doubling when the error increases by 10%. The resource scheduling strategy iterates every 24 hours, aiming for a weekly improvement in prediction accuracy of ≥2%; if this target is not met, an incremental training process is triggered.
[0067] An intelligent project management system based on multimodal big data and adaptive large model is proposed. When the system is in use, it runs the aforementioned intelligent project management method based on multimodal big data and adaptive large model.
[0068] Working principle:
[0069] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0070] The foregoing description enables those skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. An intelligent project management method based on multimodal big data and adaptive large models, characterized by: The method includes the following steps: S1: Integrates multi-source heterogeneous data, collects structured data, unstructured text, audio and video recordings, and environmental sensor data, covering indicators throughout the entire project lifecycle; S2: Perform multimodal feature extraction and standardization on the collected data, model task dependencies through graph neural networks, and use semantic models to parse risk signals in the text; S3: Construct a dynamic adaptive large model framework, integrate incremental pre-training mechanism and reinforcement learning module, and fine-tune model parameters in real time to respond to dynamic changes in the project. S4: Design a multi-objective optimization function that encodes cost, schedule and quality objectives into a multi-dimensional reward vector and balances resource allocation conflicts through dynamic weight adjustment; S5: Deploy a cross-modal embedding layer and use a multi-head attention mechanism to align heterogeneous data features to achieve joint representation of task progress, team sentiment, and physical environment parameters; S6: Introducing elastic weight solidification technology, combining historical project memory and real-time data stream to prevent the model from forgetting key historical experiences during incremental training; S7: Establish an interactive natural language interface to support project managers in querying risks in the form of instructions and generating interpretable intervention suggestions through causal reasoning; S8: Construct a digital twin simulation environment to simulate the impact of different decisions on the project outcome and provide visualized simulation results to assist in dynamic strategy adjustments; S9: Implement a closed-loop feedback mechanism to feed the actual execution results back to the adaptive large model, and continuously optimize prediction accuracy and resource scheduling efficiency.
2. The intelligent project management method based on multimodal big data and adaptive large model as described in claim 1, characterized in that: In step S1, when integrating multi-source heterogeneous data, it is necessary to connect to the JIRA and Trello project management platforms through the Apache NiFiETL tool. The structured data collected includes task progress tables and resource consumption tables. Unstructured text, including emails and Slack chat logs, needs to be extracted for key content using OCR and speech-to-text technology. Environmental sensor data, such as temperature, humidity, and equipment status parameters, are collected in real time through the MQTT protocol. The data covers all stages of the project lifecycle, including project requirement documents, execution logs, monitoring indicators, and closing reports, and outliers are filtered through a verification mechanism.
3. The intelligent project management method based on multimodal big data and adaptive large model as described in claim 1, characterized in that: In step S2, during the multimodal feature extraction stage, structured data is modeled into a task dependency graph using a graph neural network. Node attributes are fused with task priority, resource occupancy rate, and text sentiment analysis results, and edge weights represent the coupling strength between tasks. Unstructured text is extracted using the BERT model to extract semantic features, and the risk signal strength is quantified using the TF-IDF algorithm. The threshold is set to trigger an early warning if it is above 0.
75.
4. The intelligent project management method based on multimodal big data and adaptive large model as described in claim 1, characterized in that: In step S3, the dynamic adaptive large model is based on the Transformer architecture, the incremental pre-training adopts a hybrid sampling strategy, the real-time data stream is updated once per hour, and the historical memory bank is stored according to scenario classification; the model fine-tuning frequency is set to once per hour, the learning rate adopts the Cosine decay strategy, the initial value is set to 3e-5, and the minimum value is 1e-6.
5. The intelligent project management method based on multimodal big data and adaptive large model as described in claim 1, characterized in that: In step S4, the multi-objective optimization function encodes cost deviation, schedule delay, and quality score into a three-dimensional reward vector, and the weight coefficients are dynamically adjusted by the project manager through an interactive interface; the Pareto front search algorithm filters the non-dominated solution set and eliminates invalid strategies with resource utilization exceeding 90% or schedule delay exceeding 7 days.
6. The intelligent project management method based on multimodal big data and adaptive large model as described in claim 1, characterized in that: In step S5, the cross-modal embedding layer adopts a multi-head attention mechanism with a fixed number of heads of 8, mapping the task progress values, team sentiment analysis results, and sensor time-series data to a 512-dimensional shared vector space. The pre-training of the embedding layer adopts contrastive learning, with positive samples being multimodal data pairs of the same task and negative samples being randomly combined across tasks. The loss function maximizes the cosine similarity of positive sample pairs, with a threshold set to 0.
9. The training parameters include a learning rate of 1e-4 and a batch size of 128, and the iteration stops when the loss function drops below 0.
05.
7. The intelligent project management method based on multimodal big data and adaptive large model as described in claim 1, characterized in that: In step S6, firstly, the parameter importance score is obtained by multiple backpropagation calculations on historical project data, marking the parameters that play a decisive role in predicting historical tasks; then, a regularization term is introduced into the objective function of incremental training to limit the drastic changes of these key parameters; through alternating sampling of real-time data streams and historical memory banks, the model can capture the dynamic features of the current project while maintaining the stability of long-term knowledge; this mechanism effectively balances the model's ability to learn new scenarios with the need to retain historical experience, significantly improving the robustness of decision-making in complex scenarios of resource conflicts or sudden changes in demand; The formula for the regularized loss function of elastic weight fixation is: Among them, L current Let θ be the loss function for the current task. i * indicates the optimal value of the parameters after training on the historical task, Fi is the parameter importance weight calculated based on the Fisher information matrix, and λ is the tradeoff coefficient; this formula prevents catastrophic forgetting by constraining the degree to which the key parameter θi deviates from the historical optimal value. The formula for calculating the importance weights of parameters based on the Fisher information matrix is as follows: Specifically, N data samples (xn, yn) are sampled from the historical memory bank, and the Fisher information content of parameter θi is calculated to quantify its contribution to the prediction of historical tasks. The larger the Fi value, the more critical θi is to historical knowledge, and stronger protection needs to be applied in incremental training. These two formulas together construct a parameter dynamic protection mechanism, which enables the model to balance stability and adaptability during continuous learning.
8. The intelligent project management method based on multimodal big data and adaptive large model as described in claim 1, characterized in that: In step S7, the interactive natural language interface is based on a fine-tuned GPT-4 model and supports commands including "predict high-risk tasks next week". The intent recognition module determines the query type through keyword matching and context analysis. The causal reasoning module adopts a structural causal model to construct an event causal chain including "requirement change → code submission decrease → test delay" and generates intervention suggestions including "allocate backup test resources in advance".
9. The intelligent project management method based on multimodal big data and adaptive large model as described in claim 1, characterized in that: In step S8, the digital twin simulation environment is built based on the Unity engine. The input parameters include resource adjustment strategies, and 1,000 decision paths are generated through Monte Carlo simulation. The simulation results output dynamic Gantt charts and resource heatmaps. Key indicators include cost savings rate and project time compression ratio, which are compared through line charts. The simulation frequency supports real-time updates, refreshing every 15 minutes. The simulation error rate is monitored through the root mean square error (RMSE), and model recalibration is triggered when it exceeds 5%. In step S9, the closed-loop feedback mechanism compares the actual project cost and delivery quality data with the predicted values, calculates the mean absolute error (MAE) and root mean square error (RMSE), and sets the thresholds to 3% and 5%, respectively. The online learning pipeline adopts a sliding window mechanism, retaining only the data from the most recent 3 months for training, and archiving historical data to cold storage.
10. An intelligent project management system based on multimodal big data and adaptive large models, characterized in that: When in use, the system operates the intelligent project management method based on multimodal big data and adaptive large model as described in any one of claims 1 to 9.
