Full-process project management system and method based on large language model

Abstract

The invention discloses a whole-process project management system and method based on a large language model, relates to the technical field of construction engineering management, and solves the problem that full-life-cycle data of a construction project is difficult to analyze through a constructed GPT system model. And a three-party benefit balance model is difficult to construct to optimize three-party benefit conflicts in construction engineering. Comprising the following modules: a data acquisition and preprocessing module used for collecting project data; the intelligent engineering decision module is used for extracting a six-tuple entity from the unstructured data and identifying key milestone turning through semantic similarity analysis; the construction energy efficiency evaluation module is used for calculating an innovation fusion coefficient and a project fusion coefficient; the three-party collaborative revenue calculation module is used for defining a three-party joint revenue function; the dynamic equilibrium risk quantification module is used for locally calculating gradients in three parties and outputting an optimal combination; and the risk coping and monitoring integrated module is used for generating a coping strategy through the GPT system model.

Description

Technical Field

[0001] The present invention belongs to the technical field of construction project management, and specifically relates to a system and method for whole-process project management based on a large language model. Background Art

[0002] With the continuous development of technology, the engineering construction industry is undergoing digital transformation, but project management still faces many challenges. Whole-process project management refers to the integrated management of the entire life cycle of engineering projects from pre-project planning, design, construction, completion acceptance to operation and maintenance.

[0003] In traditional whole-process project management systems and methods, the data of customers, designers, and constructors are scattered in independent systems, resulting in a relatively high distortion rate of demand transmission. Risk identification often relies on manual experience, the average time from problem occurrence to solution formulation is too long, and decision-making is highly subjective. It is difficult to solve the problem of analyzing the whole-life cycle data in construction projects, and it is also difficult to build a tripartite interest balance model to optimize multi-party interest conflicts, lacking the joint analysis of BIM models and various text data at different levels.

[0004] For this reason, a system and method for whole-process project management based on a large language model have emerged. By analyzing the whole-life cycle data in the project of the GPT system model based on the Transformer architecture, a dynamic game equilibrium algorithm is adopted to build an intelligent decision-making center throughout the whole process of the project. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; for this purpose, the present invention proposes a system and method for whole-process project management based on a large language model to solve the following technical problems:

[0006] In traditional whole-process project management systems and methods, the data of customers, designers, and constructors are scattered in independent systems, resulting in a relatively high distortion rate of demand transmission. Risk identification often relies on manual experience, the average time from problem occurrence to solution formulation is too long, and decision-making is highly subjective. It is difficult to solve the problem of analyzing the whole-life cycle data in construction projects, and it is also difficult to build a tripartite interest balance model to optimize multi-party interest conflicts, lacking the joint analysis of BIM models and various text data at different levels.

[0007] To solve the above problems, the first aspect of the present invention provides a system for whole-process project management based on a large language model, including the following modules:

[0008] Data collection and preprocessing module: Collect various data within the whole life cycle of the project and perform preprocessing operations on the data;

[0009] Intelligent Engineering Decision-making Module: Annotate and encode the preprocessed data, and select the GPT system model based on the Transformer architecture to analyze the data;

[0010] Construction Energy Efficiency Evaluation Module: Normalize the ROI, parse the uploaded BIM model through the GPT system model, calculate the innovation integration coefficient, integrate BIM and ergonomics data, and calculate the process integration coefficient and project integration coefficient;

[0011] Tripartite Collaborative Revenue Calculation Module: Define the tripartite joint revenue function, and construct customer, design institute, and construction subcontractor sub-functions through ROI, innovation integration coefficient, project integration coefficient, and corresponding data;

[0012] Dynamic Equilibrium Risk Quantification Module: Input the corresponding values of three weight coefficients into the tripartite joint revenue function, calculate the total revenue value, find the weight combination that optimizes the joint revenue function through the iterative gradient descent method, use it as the input of the Monte Carlo simulation, output the corresponding revenue value, and use the revenue value as the input data of KDE for analysis, and output the KDE risk interval;

[0013] Risk Response and Monitoring Integration Module: Format the KDE risk interval data, generate response strategies through the GPT system model, and monitor and automatically trigger strategy updates in real time.

[0014] Preferably, collecting various data during the whole life cycle of the project and performing preprocessing operations on the data include the following steps:

[0015] Starting from the project initiation stage, collect the requirement documents, design documents, project plan documents, and acceptance documents proposed by customers and stakeholders. At the same time, collect email records, meeting minutes, and instant messaging records among construction subcontractor members and between construction subcontractors and stakeholders;

[0016] Collect the reports on the project progress submitted by the construction subcontractor every week and the reports submitted when the project reaches key milestone nodes; collect test data, including the data generated during unit testing, integration testing, and system testing, and collect the records of quality inspections of project achievements during the project implementation process; perform data cleaning operations on the collected data to unify the data date format, numerical format, and standardize the text format.

[0017] Preferably, annotating and encoding the preprocessed data and selecting the GPT system model based on the Transformer architecture to analyze the data include the following steps:

[0018] S301: Annotate and encode the preprocessed full life cycle data, select the GPT system model based on the Transformer architecture, perform domain adaptive pre-training on the model in the construction industry corpus, dynamically retrieve the knowledge graph, and associate the input data with the processing solutions of similar historical projects;

[0019] S302: Extract six-tuple entities from the text based on the BiLSTM-CRF model, calculate the probability distribution of risks through the probabilistic graph network, and output the optimization solution in combination with reinforcement learning;

[0020] S303: Align the semantics of text descriptions, tabular data, and BIM component attributes through the transmembrane state attention layer, and dynamically allocate modal weights according to the data feasibility;

[0021] S304: Generate a data traceability chain for each decision, including associated original data fragments, original text of cited specification clauses, and success rate statistics of historical similar cases.

[0022] Preferably, calculating the innovation integration coefficient includes the following steps:

[0023] The client provides a statistical table of the return on investment of historical projects, including the original ROI values of each project, and performs min-max normalization calculation on it; the GPT system model is used to parse the BIM models uploaded by the design institute to obtain the application area of each material and the total building area, determine the part of innovative materials in the project, respectively count the total application area of each innovative material in the project, divide it by the total building area and multiply by 100% to obtain the area coefficient of the innovative material, and set the corresponding weight coefficient according to the material type, and sum the product of the area coefficient and the weight coefficient to generate the innovation integration coefficient;

[0024] Innovation integration coefficient calculation formula:

[0025]

[0026] Among them, is the innovation integration coefficient, and are respectively the total area and the weight coefficient of the th innovative material, is the total building area, is the total number of innovative material types.

[0027] Preferably, calculating the process integration coefficient and the project integration coefficient includes the following steps:

[0028] Obtain the man-hour data and process parameters of each process from the work efficiency database provided by the construction party, extract the construction process list from the corresponding BIM model, match it with the work efficiency database, and mark the unconventional processes; extract the penalty coefficients of the corresponding unconventional processes from the work efficiency database and conduct statistics on the data from the meteorological department, and assign values according to the degree of damage to obtain the environmental damage coefficient; obtain the work efficiency basic coefficient by subtracting the difference between the actual man-hours and the standard man-hours from 1 and dividing it by the standard man-hours, correct the work efficiency basic coefficient by multiplying it with the penalty coefficients of unconventional processes using the multiplication rule, and then introduce the environmental damage coefficient through the cumulative deduction method to obtain the process integration coefficient;

[0029] Calculation formula for the process integration coefficient:

[0030] Among them, is the process integration coefficient, is the actual man-hours, is the standard man-hours, and are the penalty coefficients of two unconventional processes involved in this process, and are the two environmental damage coefficients involved in this process;

[0031] Statistical the construction period of each process and the total construction period of the project, take the ratio of the construction period of each process to the total construction period as the corresponding process weight coefficient, and multiply each process integration coefficient by the corresponding process weight coefficient and sum them up to obtain the project integration coefficient;

[0032]

[0033] Among them, is the project integration coefficient, is the th process integration coefficient, is the corresponding th process weight coefficient, is the total number of processes.

[0034] Preferably, define the tripartite joint revenue function, and construct the sub-functions of the customer, the design institute, and the construction party through ROI, the innovation integration coefficient, the project integration coefficient, and the corresponding data, including the following steps:

[0035] Define the tripartite joint revenue function:

[0036]

[0037] Among them, 、 and are the weight coefficients, satisfying + + = 1, is the customer benefit sub - function, x is the independent variable, is the design institute benefit sub - function, y is the independent variable, is the construction party loss sub - function, z is the independent variable;

[0038] The said customer benefit sub - function:

[0039]

[0040] Among them, is the customer benefit sub - function, is the actual return on investment, is the benchmark return on investment;

[0041] The said design institute benefit sub - function:

[0042]

[0043] Among them, is the design institute benefit sub - function, is the number of patented technologies involved in this project, is the total number of patented technologies, is the innovation integration coefficient corresponding to this project, is the green certification score of this project;

[0044] The said construction party loss sub - function:

[0045]

[0046] Among them, is the construction party loss sub - function, is the actual cost, is the target cost, is the number of non - conventional processes used in the project, is the total number of processes used in the project, is the project integration coefficient.

[0047] Preferably, the dynamic equilibrium risk quantification module includes:

[0048] Input the values corresponding to the three weight coefficients into the tripartite joint revenue function, calculate an overall revenue value through the joint revenue function. Each participant calculates the partial derivative of the revenue function with respect to its own weight based on its local data. The central server collects the partial derivatives of all parties, performs weighted aggregation, and the weight allocation rule is the ratio of the data volume of this participant to the total data volume multiplied by the credibility score. Through the iterative gradient descent method, find the weight combination (A, B, C) that makes the joint revenue function reach the optimum, that is, the equilibrium point, until the iteration exceeds 500 times;

[0049] Use the equilibrium point (A, B, C) as the input for Monte Carlo simulation, run 10,000 simulations, generate a set of revenue values for each simulation, and the output result is 10,000 sets of simulated revenue values and risk indicators; Use the 10,000 sets of revenue values as the input data for KDE, automatically calculate the optimal bandwidth using the Silverman criterion, estimate the probability density function using the selected KDE algorithm and bandwidth, find the point corresponding to the cumulative probability reaching 5%, and give a risk interval of a 95% confidence interval. Use matplotlib to draw the KDE curve and mark the risk interval in the figure.

[0050] Preferably, the risk response and monitoring integration module includes:

[0051] Format and clean the risk interval data output by kernel density estimation, establish a data transmission channel, input the processed risk interval data into the GPT system model based on the Transformer architecture to automatically generate coping strategies; and design a user interface, including a risk interval display area, a coping strategy generation button, and a historical record viewing function, add a real-time monitoring module, analyze the monitoring data, identify potential risk points and change trends, and when it detects that the project status changes, the system automatically triggers the generation and update of coping strategies and sends notifications to users through the user interface.

[0052] The second aspect of the present invention provides a method for the whole-process project management based on a large language model, including the following steps:

[0053] S1: Collect various data within the entire life cycle of the project and perform preprocessing operations on the data;

[0054] S2: Annotate and encode the preprocessed data, and select the GPT system model based on the Transformer architecture to analyze the data;

[0055] S3: Normalize the ROI, parse the uploaded BIM model through the GPT system model, calculate the innovation fusion coefficient, fuse BIM and ergonomic data, and calculate the process fusion coefficient and the project fusion coefficient;

[0056] S4: Define the tripartite joint revenue function, and construct the sub-functions of the customer, design institute, and construction party through ROI, innovation integration coefficient, project integration coefficient, and corresponding data.

[0057] S5: Input the corresponding values of the three weight coefficients into the tripartite joint revenue function, calculate the total revenue value, and find the weight combination that optimizes the joint revenue function through the iterative gradient descent method. Use this as the input for the Monte Carlo simulation, output the corresponding revenue value, and use the revenue value as the input data for KDE analysis to output the KDE risk interval.

[0058] S6: Format the KDE risk interval data, generate coping strategies through the GPT system model, and monitor and automatically trigger strategy updates in real time.

[0059] Advantages of the present invention

[0060] The present invention automatically analyzes unstructured documents such as contracts and emails through the GPT system model, greatly improving the entity recognition accuracy. Through the dynamic game equilibrium algorithm under the federated learning framework, the calculation rate of the collaborative optimization of the interests of the three parties, namely the customer, design institute, and construction party, is also significantly improved compared to the genetic algorithm, achieving a substantial acceleration of the project goal setting and adjustment process. At the same time, the risk quantification engine integrating kernel density estimation and probabilistic graphical models can effectively identify hidden risks that are difficult to capture by traditional methods, greatly improving the risk warning coverage rate. The implementation monitoring module realizes high-speed risk status updates through stream data processing, and combines data twin simulation to verify the feasibility of the strategy, greatly enhancing the effectiveness of the risk response method. Brief description of the drawings

[0061] Figure 1 It is a schematic diagram of the module process of the present invention;

[0062] Figure 2 It is a schematic diagram of the method process of the present invention. Detailed implementation manners

[0063] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0064] Please refer to Figure 1 As shown, the present invention is a system for the whole-process project management based on a large language model, including the following modules:

[0065] Data Acquisition and Preprocessing Module: Collect various data within the entire project life cycle and perform preprocessing operations on the data;

[0066] Intelligent Engineering Decision-making Module: Annotate and encode the preprocessed data, select the GPT system model based on the Transformer architecture to analyze the data. The GPT system model is a language model based on artificial intelligence technology. Through pre-training on a large-scale corpus, it learns the statistical laws of language and generates coherent and natural text. The core is based on the Transformer architecture in deep learning, and it is pre-trained through unsupervised learning and fine-tuned on specific tasks;

[0067] Construction Energy Efficiency Evaluation Module: Normalize the ROI, parse the uploaded BIM model through the GPT system model, calculate the innovation integration coefficient, integrate BIM and ergonomic data, calculate the process integration coefficient and the project integration coefficient, and perform normalization processing on the ROI data through min-max normalization to eliminate the differences in dimensions and numerical ranges between different projects and make the data under a unified standard;

[0068] Tripartite Collaborative Revenue Calculation Module: Define the tripartite joint revenue function and construct the sub-functions of the customer, design institute, and construction party through the ROI, innovation integration coefficient, project integration coefficient, and corresponding data;

[0069] Dynamic Equilibrium Risk Quantification Module: Input the corresponding values of three weight coefficients into the tripartite joint revenue function, calculate the overall revenue value, find the weight combination that optimizes the joint revenue function through the iterative gradient descent method, use it as the input for Monte Carlo simulation, output the corresponding revenue value, and use the revenue value as the input data for KDE analysis to output the KDE risk interval;

[0070] Risk Response and Monitoring Integration Module: Format the KDE risk interval data, generate response strategies through the GPT system model, and monitor and automatically trigger strategy updates in real time.

[0071] Specifically, through API automatic scraping, connect to the OA system, such as DingTalk and Enterprise WeChat, to obtain meeting minutes, acceptance documents and other unstructured files, which are submitted through the Web interface, and the quality inspection data of Internet of Things devices are transmitted in real time through RFID sensors; use the text encoder of the GPT-4 architecture to process unstructured data to generate semantic embedding vectors, and at the same time encode structured test data and quality records through the TabTransformer model to extract table features, and parse the IFC file of the BIM model based on the graph neural network to generate component-level spatial relationship embeddings; calculate the innovation integration coefficient and the project integration coefficient respectively; define the basic form of the tripartite joint benefit function, and construct the customer benefit sub-function, the design institute benefit sub-function and the construction party loss sub-function respectively, and the three functions are related to return on investment, technological innovation and cost deviation respectively; use the estimated bandwidth to select the kernel density estimation model, input the data into various models, select the one with the best performance, and output the risk interval; use gRPC bidirectional stream transmission, TLS 1.3 encrypted channel, data verification hash value, and the model outputs the corresponding strategy.

[0072] In one embodiment of the present invention, collecting various data during the entire life cycle of the project and performing preprocessing operations on the data include the following steps:

[0073] Starting from the project initiation phase, collect requirement documents, design documents, project plan documents, and acceptance documents proposed by customers and stakeholders. At the same time, collect email records, meeting minutes, and instant messaging records among construction party members and between the construction party and stakeholders;

[0074] Collect reports on the project progress submitted by the construction party on a weekly basis and reports submitted when the project reaches key milestone nodes; collect test data, including data generated during unit testing, integration testing, and system testing, and collect records of quality inspections of project results during the project implementation process; perform data cleaning operations on the collected data to unify the data date format and numerical format and standardize the text format.

[0075] Specifically, it automatically grabs through the API, connects to the OA system, such as DingTalk and WeCom, to obtain meeting minutes, acceptance documents and other unstructured files which are submitted through the Web interface. The quality inspection data of IoT devices is transmitted in real time through RFID sensors. Configure the Exchange Server monitoring rules to automatically archive emails containing keywords, such as changes and extensions. Connect to instant messaging tools through the SDK, extract group chat records and mark the senders and timestamps. Use ASR technology to transcribe audio files and manually reproduce and mark the resolution items. Perform formatting operations on the data, including date formatting and numerical formatting. And perform cleaning operations on unstructured text, extract key information and then perform quality verification, and transport the cleaned data to the GPT system model through the gRPC protocol.

[0076] In one of the embodiments of the present invention, the preprocessed data is annotated and encoded, and a GPT system model based on the Transformer architecture is selected to analyze the data, including the following steps:

[0077] S301: Annotate and encode the full life cycle data after preprocessing, select a GPT system model based on the Transformer architecture, perform domain adaptive pre-training on the model in the construction industry corpus, dynamically retrieve the knowledge graph, and associate the input data with the processing solutions of similar historical projects;

[0078] S302: Extract six-tuple entities from the text based on the BiLSTM-CRF model, calculate the probability distribution of risks through the probabilistic graph network, and output an optimization solution in combination with reinforcement learning;

[0079] S303: Align the semantics of text descriptions, table data and BIM component attributes through the transmembrane state attention layer, and dynamically allocate modal weights according to the data feasibility;

[0080] S304: Generate a data traceability chain for each decision output, including the associated original data segments, the original text of the cited specification clauses, and the success rate statistics of historical similar cases.

[0081] Specifically, a text encoder adopting the GPT-4 architecture processes unstructured data to generate semantic embedding vectors. The TabTransformer model encodes structured test data and quality records to extract table features. And the IFC file of the BIM model is parsed based on a graph neural network to generate component-level spatial relationship embeddings. In the building industry material library, which includes specification texts, construction method manuals, and historical cases, domain adaptive pre-training of the model is carried out, and the knowledge graph is dynamically retrieved to associate the input data with the processing solutions of similar historical projects. The task-specific output heads include an entity recognition head, a risk prediction head, and a strategy generation head. Semantic alignment of text descriptions, tabular data, and BIM component attributes is achieved through cross-modal attention layers, and a data traceability chain is generated for each decision output, including associated original data segments, the original text of cited specification clauses, and the success rate statistics of historical similar cases.

[0082] In one embodiment of the present invention, normalizing the ROI and parsing the uploaded BIM model through the GPT system model to calculate the innovation fusion coefficient includes the following steps:

[0083] The client provides a statistical table of the return on investment of historical projects, including the original ROI values of each project, and performs min-max normalization calculation on it; the GPT system model parses the BIM model uploaded by the design institute to obtain the application area of each material and the total building area, determines the part of innovative materials in the project, separately calculates the sum of the application areas of each innovative material in the project, divides it by the total building area and multiplies by 100% to obtain the area coefficient of the innovative material, and sets the corresponding weight coefficient according to the material type, and sums the product of the area coefficient and the weight coefficient to generate the innovation fusion coefficient;

[0084] Innovation fusion coefficient calculation formula:

[0085]

[0086] Where, is the innovation fusion coefficient, and are respectively the total area sum and the weight coefficient of the th innovative material, is the total building area, is the total number of innovative material types.

[0087] Specifically, for the input data, normalization processing is performed on the data through the min-max normalization calculation formula to eliminate the differences in dimension and numerical range between different items, so that the data is under a unified standard. The normalized ROI is used for subsequent weight calculation. The BIM model is parsed to extract material data, and whether the material name is in the innovation library is checked to identify the innovative materials for the project. The total application area of each innovative material in the project is respectively counted, divided by the total building area and then multiplied by 100% to obtain the area coefficient of the innovative material, and the corresponding weight coefficient is set according to the material type. The sum of the products of the area coefficient and the weight coefficient is used to generate the innovation integration coefficient.

[0088] The corresponding weight coefficients are set according to the material type as shown in Table 1:

[0089] Table 1

[0090] Material type Weight coefficient Judgment basis Patented materials 1.2 Patent validity period ≥ 5 Self-developed without patent 0.8 Enterprise white paper certification Introduced technology 0.5 Record-filing of technology transfer agreement

[0091] In one embodiment of the present invention, for fusing the BIM and ergonomic data and calculating the process fusion coefficient and the project fusion coefficient, the following steps are included:

[0092] Obtain the man-hour data and process parameters of each process from the ergonomic database provided by the construction party, extract the construction process list from the corresponding BIM model, match it with the ergonomic database, and mark the non-conventional processes; extract the penalty coefficients of the corresponding non-conventional processes from the ergonomic database and perform statistics on the data from the meteorological department, and assign values according to the degree of damage impact to obtain the environmental damage coefficient; obtain the ergonomic basic coefficient by subtracting the difference between the actual man-hours and the standard man-hours from 1 and then dividing by the standard man-hours, correct the ergonomic basic coefficient by applying the penalty coefficients of non-conventional processes using the multiplication rule, and then introduce the environmental damage coefficient through the cumulative deduction method to obtain the process fusion coefficient;

[0093] Process fusion coefficient calculation formula:

[0094] Among them, is the process fusion coefficient, is the actual man-hours, is the standard man-hours, and are the penalty coefficients of two non-conventional processes involved in this process, and are the environmental damage coefficients of two processes involved in this process;

[0095] Statistical the duration time of each process and the total duration time of the project, use the ratio value of the duration time of each process to the total duration time as the corresponding process weight coefficient, and sum up the products of each process fusion coefficient and the corresponding process weight coefficient to obtain the project fusion coefficient;

[0096]

[0097] Among them, is the project integration coefficient, is the integration coefficient of the th process, is the weight coefficient of the corresponding th process, is the total number of processes.

[0098] Specifically, data is obtained from the construction efficiency database of the construction party, including process name, standard working hours, process type, and penalty coefficient. The output form is set as a table form. The corresponding BIM process list is parsed, the IFC model is extracted, the process names of BIM and the efficiency database are matched using regular expressions, an environmental damage coefficient table is set from the meteorological data, and the coefficient is recalculated every time 20% of the project volume is completed to adjust the subsequent process parameters.

[0099] Set the environmental damage coefficient table according to the meteorological data as shown in Table 2:

[0100] Table 2

[0101] Meteorological conditions Influence coefficient Judgment threshold Blizzard (≥ 10 mm) 0.5 Continuous snowfall ≥ 1 hour Heavy rain (> 50 mm) 0.3 Continuous rainfall ≥ 2 hours High temperature (> 35 °C) 0.15 Daily maximum temperature ≥ 35 °C Low temperature (< 10 °C) 0.1 Daily minimum temperature < 10 °C

[0102] In one embodiment of the present invention, the tripartite joint revenue function is defined, and the customer, designer, and construction party sub-functions are constructed through ROI, innovation integration coefficient, project integration coefficient, and corresponding data, including the following steps:

[0103] Define the tripartite joint revenue function:

[0104]

[0105] Among them, , and are weight coefficients, satisfying + + = 1, is the customer revenue sub-function, x is the independent variable, is the designer revenue sub-function, y is the independent variable, is the construction party loss sub-function, z is the independent variable;

[0106] The customer revenue sub-function:

[0107]

[0108] Among them, is the customer revenue sub-function, is the actual return on investment, is the benchmark return on investment;

[0109] The sub - function of the design institute's income:

[0110]

[0111] Among them, is the sub - function of the design institute's income, is the number of patented technologies involved in this project, is the total number of patented technologies, is the innovation integration coefficient corresponding to this project, is the green certification score of this project;

[0112] The sub - function of the constructor's loss:

[0113]

[0114] Among them, is the sub - function of the constructor's loss, is the actual cost, is the target cost, is the number of unconventional processes used in the project, is the total number of processes used in the project, is the project integration coefficient.

[0115] Specifically, set the initial weights of the three parties to equal values, all 0.333. Define the basic form of the combined income function of the three parties, and construct the customer income sub - function, the design institute income sub - function, and the constructor loss sub - function respectively. The three functions are respectively related to return on investment, technological innovation, and cost deviation. Construct the customer income sub - function through the historical project financial statements, allocate the grade coefficients corresponding to each grade in the LEED certification grade mapping table, construct the design institute income sub - function, and construct the constructor loss sub - function through two factors of cost and non - normal processes; use the historical project data to train the neural network and then output the values of A, B, and C.

[0116] The grade coefficients corresponding to each grade in the LEED certification grade mapping table are shown in Table 3:

[0117] Table 3

[0118] LEED certification level Level coefficient Certification level 0.4 Silver level 0.6 Gold level 0.8 Platinum level 1

[0119] In one of the embodiments of the present invention, the dynamic equilibrium risk quantification module includes:

[0120] Input the values corresponding to the three weight coefficients into the tripartite joint revenue function, calculate an overall revenue value through the joint revenue function. Each participant calculates the partial derivative of the revenue function with respect to its own weight based on its local data. The central server collects the partial derivatives of all parties and performs weighted aggregation. The weight allocation rule is the ratio of the data volume of this participant to the total data volume multiplied by the credibility score. Through the iterative gradient descent method, find the weight combination (A, B, C) that makes the joint revenue function reach the optimum, that is, the equilibrium point, until the iteration exceeds 500 times;

[0121] Use the equilibrium point (A, B, C) as the input for Monte Carlo simulation, run 10,000 simulations. Each simulation generates a set of revenue values, and the output result is 10,000 sets of simulated revenue values and risk metrics; Use the 10,000 sets of revenue values as the input data for KDE, automatically calculate the optimal bandwidth using the Silverman criterion, use the selected KDE algorithm and bandwidth to estimate the probability density function, find the point corresponding to the cumulative probability reaching 5%, and then give a risk interval with a 95% confidence interval. Use matplotlib to plot the KDE curve and mark the risk interval in the graph.

[0122] Specifically, initialize the parameters, input the initial weights of the three parties, all of which are 0.333, set the learning rate to 0.01, and the maximum number of iterations to 500. The central server performs weighted aggregation, and the ratio of the data volume of the corresponding participant to the total data volume multiplied by the credibility score to obtain the corresponding final weight, and perform parameter update. The termination condition is ≥500 times; Input the equilibrium solution simulation, calculate the standard deviation and mean deviation of the data, multiply the deviation by a constant, usually 1.06, to estimate the bandwidth, use the estimated bandwidth to select the kernel density estimation model, input the data into various models, and select the one with the best performance, and output the risk interval.

[0123] In one embodiment of the present invention, the risk response and monitoring integration module includes:

[0124] Format and clean the risk interval data output by the kernel density estimation, establish a data transmission channel, input the processed risk interval data into the GPT system model based on the Transformer architecture to automatically generate response strategies; and design a user interface, including a risk interval display area, a response strategy generation button, and a historical record viewing function, add a real-time monitoring module, analyze the monitoring data, identify potential risk points and change trends. When it detects that the project status changes, the system automatically triggers the generation and update of the response strategy and sends a notification to the user through the user interface.

[0125] Specifically, use gRPC bi-directional streaming, TLS 1.3 encrypted channels, data verification hash values, and the model outputs corresponding strategies. For example, increasing night construction shortens the construction period by 3 - 5 days, and increases the labor cost by 12%. Starting precast components shortens the construction period by 7 days and increases the material cost by 8%.

[0126] Please refer to Figure 2 As described above, the present invention is a method for whole-process project management based on a large language model, including the following steps:

[0127] S1: Collect various data within the whole life cycle of the project and perform preprocessing operations on the data;

[0128] S2: Annotate and encode the preprocessed data, and select a GPT system model based on the Transformer architecture to analyze the data;

[0129] S3: Normalize the ROI, parse the uploaded BIM model through the GPT system model, calculate the innovation integration coefficient, integrate BIM and ergonomic data, and calculate the process integration coefficient and the project integration coefficient;

[0130] S4: Define a three-party joint benefit function, and construct sub-functions for the client, the design institute, and the construction party through the ROI, the innovation integration coefficient, the project integration coefficient, and the corresponding data;

[0131] S5: Input the corresponding values of three weight coefficients into the three-party joint benefit function, calculate the overall benefit value, find the optimal weight combination that makes the joint benefit function optimal through the iterative gradient descent method, use it as the input for Monte Carlo simulation, output the corresponding benefit value, and use the benefit value as the input data for KDE analysis to output the KDE risk interval;

[0132] S6: Format the KDE risk interval data, generate coping strategies through the GPT system model, and monitor and automatically trigger strategy updates in real time.

[0133] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. 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 method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A system for full-process project management based on large language models, characterized in that, It includes the following modules: Data collection and preprocessing module: Collect various data within the entire project life cycle and perform preprocessing operations on the data; Intelligent engineering decision-making module: Annotate and encode the preprocessed data, and select the GPT system model based on the Transformer architecture to analyze the data; Construction energy efficiency evaluation module: Normalize the ROI, parse the uploaded BIM model through the GPT system model, calculate the innovation integration coefficient, integrate BIM and ergonomic data, calculate the process integration coefficient and the project integration coefficient; Tripartite collaborative benefit calculation module: Define the tripartite joint benefit function, and construct the sub-functions of the customer, design institute, and construction party through the ROI, innovation integration coefficient, project integration coefficient, and corresponding data; Dynamic equilibrium risk quantification module: Input the corresponding values of three weight coefficients into the tripartite joint benefit function, calculate the overall benefit value, find the weight combination that optimizes the joint benefit function through the iterative gradient descent method, use it as the input of the Monte Carlo simulation, output the corresponding benefit value, and use the benefit value as the input data of the KDE for analysis, and output the KDE risk interval; Risk response and monitoring integration module: Format the KDE risk interval data, generate response strategies through the GPT system model, and monitor and automatically trigger strategy updates in real time.

2. The system for full-process project management based on a large language model according to claim 1, wherein The collection of various data within the entire project life cycle and the preprocessing operations on the data include the following steps: Starting from the project initiation phase, collect the requirement documents, design documents, project plan documents, and acceptance documents proposed by customers and stakeholders. At the same time, collect the email records, meeting minutes, and instant messaging records among construction party members and between the construction party and stakeholders; Collect the reports on the project progress submitted by the construction party on a weekly basis and the reports submitted when the project reaches key milestone nodes; collect test data, including the data generated during unit testing, integration testing, and system testing, and collect the records of quality inspections of the project results during the project implementation process; perform data cleaning operations on the collected data to unify the data date format, numerical format, and standardize the text format.

3. The system for full-process project management based on a large language model according to claim 1, characterized in that, The annotation and encoding of the preprocessed data, and the selection of the GPT system model based on the Transformer architecture to analyze the data include the following steps: S301: Annotate and encode the preprocessed full life cycle data, select the GPT system model based on the Transformer architecture, perform domain adaptive pre-training on the model in the construction industry corpus, dynamically retrieve the knowledge graph, and associate the input data with the processing solutions of similar historical projects; S302: Extract six-tuple entities from the text based on the BiLSTM-CRF model, calculate the probability distribution of risks through the probabilistic graph network, and output the optimization solution in combination with reinforcement learning; S303: Align the semantics of text descriptions, tabular data, and BIM component attributes through the cross-modal attention layer, and dynamically allocate modal weights according to the data feasibility. S304: Generate a data traceability chain for each decision output, including associated original data segments, the original text of cited specification clauses, and the success rate statistics of historical similar cases.

4. A system for full-process project management based on a large language model according to claim 1, characterized in that, The calculation of the innovation integration coefficient includes the following steps: The client provides a statistical table of the return on investment of historical projects, including the original ROI values of each project, and performs min-max normalization calculation on it; the GPT system model is used to parse the BIM model uploaded by the design institute to obtain the application area of each material and the total building area, determine the innovative material part in the project, respectively count the total application area of each innovative material in the project, divide by the total building area and multiply by 100% to obtain the area coefficient of the innovative material, and set the corresponding weight coefficient according to the material type, and sum the product of the area coefficient and the weight coefficient to generate the innovation integration coefficient; Innovation integration coefficient calculation formula: Among them, is the innovation integration coefficient, and are respectively the total area sum and the weight coefficient of the th innovative material, is the total building area, is the total number of innovative material types.

5. A system for full-process project management based on a large language model according to claim 1, characterized in that, The calculation of the process integration coefficient and the project integration coefficient includes the following steps: Obtain the man-hour data and process parameters of each process from the labor efficiency database provided by the construction party, extract the construction process list from the corresponding BIM model, match it with the labor efficiency database, and mark the unconventional processes; extract the penalty coefficient of the corresponding unconventional process from the labor efficiency database and perform statistics on the data from the meteorological department, and assign values according to the degree of damage to obtain the environmental damage coefficient; obtain the labor efficiency basic coefficient by subtracting the difference between the actual man-hours and the standard man-hours divided by the standard man-hours from 1, correct the labor efficiency basic coefficient by applying the penalty coefficient of the unconventional process using the multiplication rule, and then introduce the environmental damage coefficient through the cumulative deduction method to obtain the process integration coefficient; Calculation formula for process integration coefficient: Among them, is the process integration coefficient, is the actual working hours, is the standard working hours, and are the penalty coefficients of two non-conventional processes involved in this process, and are the two environmental damage coefficients involved in this process; Statistical the construction period of each process and the total construction period of the project, use the ratio of the construction period of each process to the total construction period as the corresponding process weight coefficient, and sum up the product of each process integration coefficient and the corresponding process weight coefficient to obtain the project integration coefficient; Among them, is the project integration coefficient, is the integration coefficient of the th process, is the corresponding weight coefficient of the th process, is the total number of processes.

6. The system for whole-process project management based on a large language model according to claim 1, wherein The definition of the tripartite joint benefit function, and the construction of the client, design institute, and construction party sub-functions through ROI, innovation integration coefficient, project integration coefficient, and corresponding data, includes the following steps: Define the tripartite joint benefit function: Among them, , and are weight coefficients, satisfying + + = 1, is the customer benefit sub-function, x is the independent variable, is the design institute benefit sub-function, y is the independent variable, is the construction party loss sub-function, z is the independent variable; The client benefit sub-function: Among them, is the customer benefit sub-function, is the actual rate of return on investment, is the benchmark rate of return on investment; The design institute benefit sub-function: Among them, is the revenue sub-function of the design institute, is the number of patented technologies involved in this project, is the total number of patented technologies, is the innovation integration coefficient corresponding to this project, is the green certification score of this project; The construction party loss sub-function: Among them, is the loss sub-function of the construction party, is the actual cost, is the target cost, is the number of non-conventional processes used in the project, is the total number of processes used in the project, is the project integration coefficient.

7. A system for full-process project management based on a large language model according to claim 1, characterized in that, The dynamic equilibrium risk quantification module includes: Input the values corresponding to the three weight coefficients into the tripartite joint benefit function, calculate an overall benefit value through the joint benefit function, each participating party calculates the partial derivative of the benefit function with respect to its own weight based on its local data, the central server collects the partial derivatives of each party, performs weighted aggregation, and the weight distribution rule is the ratio of the data volume of this participating party to the total data volume multiplied by the credibility score. Through the iterative gradient descent method, find the weight combination (A, B, C) that makes the joint benefit function reach the optimal, that is, the equilibrium point, until the iteration exceeds 500 times; Use the equilibrium points (A, B, C) as the input for Monte Carlo simulation, run 10,000 simulations. Each simulation generates a set of return values, and the output results are 10,000 sets of simulated return values and risk metrics. Use the 10,000 sets of return values as the input data for KDE, automatically calculate the optimal bandwidth using the Silverman criterion, estimate the probability density function using the selected KDE algorithm and bandwidth, find the point corresponding to a cumulative probability of 5%, and then give a risk interval with a 95% confidence interval. Use matplotlib to plot the KDE curve and label the risk interval in the graph.

8. A system for full-process project management based on a large language model according to claim 1, characterized in that, The risk response and monitoring integration module includes: Format and clean the risk interval data output by kernel density estimation, establish a data transmission channel, input the processed risk interval data into the GPT system model based on the Transformer architecture to automatically generate response strategies. Design a user interface, including a risk interval display area, a response strategy generation button, and a historical record viewing function. Add a real-time monitoring module to analyze the monitoring data, identify potential risk points and change trends. When a change in the project status is detected, the system automatically triggers the generation and update of response strategies and sends notifications to users through the user interface.

9. A method for whole-process project management based on large language models, characterized in that, It includes the following steps: S1: Collect various data during the entire project life cycle and perform preprocessing operations on the data. S2: Annotate and encode the preprocessed data, and select the GPT system model based on the Transformer architecture to analyze the data. S3: Normalize the ROI, parse the uploaded BIM model through the GPT system model, calculate the innovation integration coefficient, fuse BIM and ergonomics data, calculate the process integration coefficient and the project integration coefficient. S4: Define the tripartite joint return function, and construct the sub-functions of the client, design institute, and construction party through the ROI, innovation integration coefficient, and project integration coefficient and the corresponding data. S5: Input the corresponding values of three weight coefficients into the tripartite joint return function, calculate the overall return value, use the iterative gradient descent method to find the weight combination that optimizes the joint return function, use it as the input for Monte Carlo simulation, output the corresponding return values, use the return values as the input data for KDE analysis, and output the KDE risk interval. S6: Format the KDE risk interval data, generate response strategies through the GPT system model, and the real-time monitoring automatically triggers strategy updates.

Images