Intelligent construction site large model working system and method

Abstract

The invention relates to the technical field of intelligent construction sites, in particular to an intelligent construction site large model working system and method, and the system comprises a basic large model which is used for understanding a natural language instruction of a user, providing semantic analysis, multi-modal data processing and context memory functions, and adapting to the requirements of the construction industry; the Agent is used for analyzing the instruction and calling a function module to execute a task; the industry expert consultation module is used for generating professional consultation replies through a retrieval knowledge base; the intelligent data analysis module is used for collecting construction data, performing statistical analysis and trend prediction and generating a data analysis report; the intelligent service evaluation module generates a service report in combination with an evaluation algorithm and assists in decision making; the application program serves as an interaction entrance, receives input and displays the visual report. According to the invention, the technical problems of insufficient professional knowledge understanding, low data analysis and decision support efficiency and limited inter-module cooperation capability in existing intelligent construction site management are effectively solved, and the intelligent and precise level of construction management is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of smart construction sites, and in particular to a smart construction site large model working system and method. Background Art

[0002] With the rapid development of artificial intelligence technology, especially natural language processing and machine learning, big models have shown great potential in the field of smart construction sites, and are able to process large amounts of data and provide intelligent decision support through deep learning algorithms.

[0003] However, the current general large model has the following problems in the actual application of the construction industry: lack of industry customization, unable to accurately understand construction terms, construction processes and safety specifications; low data service integration, difficult to achieve real-time automated decision support; insufficient real-time and interactivity, limited performance in multiple rounds of dialogue and instant response, and difficult to maintain contextual coherence. These problems limit the application effect of the model in the field of smart construction sites, and it is urgently needed to optimize it according to the characteristics of the industry, integrate professional knowledge and large model technology, so as to improve construction efficiency, safety and management quality. Summary of the invention

[0004] The present invention provides a smart construction site large model working system and method, thereby effectively solving the problems pointed out in the background technology.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: A smart construction site large model working system, comprising: The basic large model is used to understand the natural language instructions input by users, provide semantic parsing, multimodal data processing and context memory functions, and adapt to the professional needs of the construction industry through fine-tuning; Agent, which is used to receive natural language instructions from users, parse the instructions, call corresponding functional modules based on task intent, and perform task scheduling between modules; The industry expert consultation module is used to process users' industry consultation requests and generate expert-level consultation responses by searching the industry knowledge base; Intelligent data analysis module, which is used to collect construction process data at the construction site, conduct statistical analysis and trend forecasting in combination with the construction database, and generate construction data analysis reports; The intelligent service evaluation module is used to respond to users' requests for the generation of construction data analysis reports and auxiliary decision-making, and to generate service evaluation reports in combination with a priori evaluation algorithms; The application, as the entry point for users to interact with the system, is used to receive user input, display analysis results, and generate visual data reports.

[0006] Furthermore, the industry expert consulting module, intelligent data analysis module and intelligent service evaluation module all include a data processing layer and an API layer. The data processing layer corresponds to the industry knowledge base, construction database and evaluation algorithm library respectively, and the API layer corresponds to internal APIs, external APIs and decision support APIs respectively, which are used to exchange data with external systems.

[0007] Furthermore, the Agent calls functions and APIs of different target modules through Function Call according to different task types.

[0008] Furthermore, the technical implementation of the industry expert consultation module includes: Based on the knowledge base of the construction industry, including technical specifications, project cases and expert experience, as the source of expertise for the model; Parse user input queries through natural language processing technology, understand query intent and retrieve relevant information from the knowledge base; Combined with machine learning algorithms, it dynamically recommends the best expert answers and solutions based on user questions.

[0009] Furthermore, the technical implementation of the intelligent data analysis module includes: Deploy a variety of sensors and monitoring equipment to collect real-time environmental data and equipment operation data on the construction site; Combined with big data analysis technology, real-time monitoring, progress forecasting and early warning of construction data can be carried out to assist managers in optimizing resource allocation.

[0010] Furthermore, the technical implementation of the intelligent service evaluation module includes: The analytic hierarchy process is used to assign weights to the evaluation indicators; Integrate various data on the construction site and provide a basis for site management through data analysis and mining; An intelligent evaluation model is designed to perform different lock-scope object analyses on the relevant data of the login roles according to different login roles, and various analysis and evaluation results are calculated.

[0011] A method for working with a large model of a smart construction site, the method comprising: Receive natural language instructions input by users, perform semantic analysis through the smart construction site model, and identify the user's task intent; Based on the identified task intentions, instructions are decomposed and the execution order of different tasks is coordinated; In response to consultation requests from the construction industry, we use the big model to retrieve expertise in the construction field, including technical specifications, project cases and expert experience, and generate professional consulting results related to the directive; Collect environmental data and equipment operation data at the construction site and analyze them in combination with historical construction data, including real-time monitoring of construction progress, trend prediction, and potential risk warning; Conduct service evaluation based on construction analysis results, generate construction service evaluation results through weighted analysis methods, and propose optimization suggestions based on relevant data; The analysis results, evaluation reports and optimization suggestions are output in a visual way, supporting users to view them in real time and make further decisions.

[0012] Furthermore, the smart construction site big model is trained by an industry training set, and the industry training set is constructed, including: Collect structured and unstructured data, pre-process the collected data, classify and label the unstructured data, and determine the model learning objectives; Expand the amount of training data through data enhancement technology, integrate multimodal data, and build a unified feature space; Based on industry data, industry terms and rules are structured through entity extraction and relationship extraction to build an industry knowledge graph.

[0013] Furthermore, a large-scale pre-trained language model is used as a baseline, and label flipping is performed to enhance the data in the industry training set, including: Use the BERT model to train a label classifier to classify the labels of the original data; Generate new enhanced samples using the T5 model, generate masks by randomly deleting part of the input content, and use the T5 model to predict part of the mask content to form a new sample set; Use the label classifier to screen the generated enhanced samples, evaluate the sample quality based on the confidence level, and select high-quality samples; The enhanced samples are combined with the original samples to retrain the classifier and related modules of the large model.

[0014] Furthermore, the large model of the smart construction site is trained in parallel, including: According to the structure and parameter quantity of the large model, the various layers of the model are reasonably divided into multiple parts and allocated to different computing devices; Divide the training data into multiple batches, and assign each batch to different computing devices so that each device can independently obtain enough data for training; On each of the computing devices, forward propagation calculations are independently performed based on the segmented model and data, each device calculates its assigned part, and transmits the result to the computing device of the next layer; On each of the computing devices, independently perform back propagation calculation based on the segmented model and data, generate gradient information of each part, and transmit the gradient information to the computing device of the upper layer; After each of the computing devices completes the back-propagation calculation, the gradient information calculated by each device is aggregated to generate the gradient of the overall model and update the model parameters.

[0015] The technical solution of the present invention can achieve the following technical effects: It effectively solves the technical problems of insufficient understanding of professional knowledge, inefficient data analysis and decision support, and limited collaboration capabilities between modules in the existing smart construction site management, and improves the intelligence and precision of construction management. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 It is a structural diagram of a large-scale model working system of a smart construction site; Figure 2 It is a flowchart of a working method of a large model of a smart construction site; Figure 3 Schematic diagram of the process of building industry training sets; Figure 4 A schematic diagram of the process of enhancing the data in the industry training set; Figure 5 Schematic diagram of the process of parallel training of the smart construction site large model. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0020] Embodiment 1 like Figure 1 As shown, the present invention provides a smart construction site large model working system, the system comprising: The basic large model is used to understand the natural language instructions input by users, provide semantic parsing, multimodal data processing and context memory functions, and adapt to the professional needs of the construction industry through fine-tuning; Specifically, the present invention selects the Qianwen Online Big Model as the basic model. The model has been trained on large-scale text data in the pre-training stage and has strong language understanding and generation capabilities. The model uses natural language processing technology to perform word segmentation, semantic analysis and intent recognition on the instructions input by the user to ensure that the instructions can be accurately understood. In response to the needs of the construction industry, the basic big model supports the processing of multimodal data, including data input in the form of text, images, videos, etc., and uses multimodal features for comprehensive analysis and understanding, thereby improving the adaptability to complex inputs; in addition, the basic big model has a context memory function, and supports smooth parsing of multiple rounds of dialogues by storing and associating user interaction history, ensuring that context consistency can be maintained under continuous instructions. Through fine-tuning and optimization on specific data sets for the construction industry, the model further enhances the ability to understand professional terms, construction processes and industry specifications, and provides basic support for efficient instruction parsing and intelligent response in construction industry scenarios. The basic big model is only responsible for parsing and understanding user instructions, and does not involve specific task execution or the calling of other functional modules.

[0021] Agent, which is used to receive natural language instructions from users, parse the instructions, call corresponding functional modules based on task intent, and perform task scheduling between modules; Specifically, the Agent first performs word segmentation, semantic analysis and intent recognition on the natural language instructions input by the user to determine the operation type and target module required for the instruction, such as industry consultation, data analysis or service evaluation. After completing the task intent analysis, the Agent dynamically assigns tasks to corresponding functional modules based on the current system status and task requirements. Through the task scheduling logic, the Agent can coordinate the task execution order and data transmission between different modules to ensure that the modules work together to complete complex tasks. At the same time, the Agent monitors the task execution process, receives the intermediate results or completion feedback returned by the module in real time, and adjusts the subsequent task flow.

[0022] The industry expert consultation module is used to process users' industry consultation requests and generate expert-level consultation responses by searching the industry knowledge base; Intelligent data analysis module, which is used to collect construction process data at the construction site, conduct statistical analysis and trend forecasting in combination with the construction database, and generate construction data analysis reports; The intelligent service evaluation module is used to respond to users' requests for the generation of construction data analysis reports and auxiliary decision-making, and to generate service evaluation reports in combination with a priori evaluation algorithms; The application, as the entry point for users to interact with the system, is used to receive user input, display analysis results, and generate visual data reports.

[0023] Specifically, the application receives the user's natural language instructions or other interactive operations through the user interface and passes them to the system for processing. The user interface supports multiple input methods, such as text input, voice commands, etc., to ensure the convenience and diversity of interaction. After the task is completed, the application is responsible for presenting the analysis results returned by the system to the user in a clear and intuitive manner. The display methods include text reports, charts, trend analysis charts, and multi-dimensional visual data views, which are convenient for users to understand and use. In addition, the application supports the generation of structured or personalized visual data reports, such as construction progress reports, safety risk analysis reports, etc., to meet the specific needs of users in construction management.

[0024] Through the present invention, the technical problems of insufficient understanding of professional knowledge, low efficiency of data analysis and decision support, and limited collaboration capabilities between modules in the existing smart construction site management are effectively solved, and the intelligence and precision of construction management are improved.

[0025] As a preferred embodiment of the above-mentioned embodiments, the industry expert consultation module, the intelligent data analysis module and the intelligent service evaluation module all include a data processing layer and an API layer. The data processing layer corresponds to the industry knowledge base, the construction database and the evaluation algorithm library respectively, and the API layer corresponds to internal APIs, external APIs and decision support APIs respectively, and is used to exchange data with external systems.

[0026] Specifically, the data processing layer corresponds to the industry knowledge base, construction database and evaluation algorithm library respectively. Among them, the data processing layer of the industry expert consultation module is responsible for retrieving and processing relevant data in the industry knowledge base and generating industry consultation replies for users; the data processing layer of the intelligent data analysis module is responsible for extracting and analyzing relevant data of the construction site from the construction database, completing statistical analysis and trend prediction; the data processing layer of the intelligent service evaluation module is responsible for calling the evaluation algorithm library, performing service evaluation on construction data and generating a comprehensive service report. The API layer is used to support the functional implementation of the module and interact with external systems. The API layer of the industry expert consultation module includes internal APIs for calling data from the industry knowledge base; the API layer of the intelligent data analysis module includes external APIs for docking with external data sources such as sensor systems and construction management systems; the API layer of the intelligent service evaluation module includes decision support APIs for exchanging evaluation results and decision suggestions with the user decision system. Through the division of labor and cooperation between the data processing layer and the API layer, each module can efficiently complete data acquisition, analysis and processing, and result return.

[0027] As a preferred embodiment of the above embodiment, the Agent calls functions and APIs of different target modules through Function Call according to different task types.

[0028] Specifically, after the natural language instructions input by the user are parsed, the Agent identifies the task type of the instruction and matches the corresponding target module. For example, for industry consulting tasks, the Agent calls the relevant functions and APIs of the industry expert consulting module through Function Call to retrieve the industry knowledge base and generate expert-level consulting responses; for construction data analysis tasks, the Agent calls the API of the intelligent data analysis module through Function Call to obtain relevant data in the construction database and complete statistical analysis and trend prediction; for service evaluation tasks, the Agent calls the functions and APIs of the intelligent service evaluation module through Function Call to generate a service evaluation report based on construction data and evaluation algorithms. Function Call is responsible for dynamically assigning tasks to the corresponding modules and receiving the results returned by the modules after the tasks are completed. The Agent achieves efficient scheduling and execution of tasks through flexible calls to the functions and APIs of each module.

[0029] As a preferred embodiment of the above, the technical implementation of the industry expert consultation module includes: Based on the knowledge base of the construction industry, including technical specifications, project cases and expert experience, as the source of expertise for the model; Parse user input queries through natural language processing technology, understand query intent and retrieve relevant information from the knowledge base; Combined with machine learning algorithms, it dynamically recommends the best expert answers and solutions based on user questions.

[0030] Specifically, the module first parses the query input by the user through natural language processing technology, extracts key content and understands the user's query intent. Then, through the matching algorithm, it retrieves information highly relevant to the query content in the industry knowledge base, including construction specifications, case analysis or expert advice. Based on the search results, it combines machine learning algorithms to sort and optimize relevant information, and dynamically generates expert answers and solutions that best match the user's questions. For example, for construction quality management issues raised by users, the module can parse the scenarios described by users, extract corresponding technical specifications and cases from the knowledge base, and recommend the best solutions through model evaluation. At the same time, the module supports the dynamic update of the industry knowledge base, combining new industry information to continuously improve the accuracy and reliability of consulting services. Finally, the module will feedback the generated results to the user in a structured or textual form.

[0031] As a preferred embodiment of the above, the technical implementation of the intelligent data analysis module includes: Deploy a variety of sensors and monitoring equipment to collect real-time environmental data and equipment operation data on the construction site; Combined with big data analysis technology, real-time monitoring, progress forecasting and early warning of construction data can be carried out to assist managers in optimizing resource allocation.

[0032] Specifically, the module accesses the data acquisition system through the sensor network, integrates and preprocesses the real-time collected data, including deduplication, outlier detection and data formatting, to ensure the quality and consistency of the data. In the data analysis stage, combined with big data analysis technology, the intelligent data analysis module monitors and analyzes the key data in the construction process in real time. For the construction progress, the module compares and analyzes the current construction status with the planned target through the progress prediction algorithm, generates the future progress prediction value, and promptly finds deviations and provides optimization suggestions. For safety warnings, the module analyzes sensor data in real time through the data anomaly detection model, identifies potential risk factors and triggers safety alarms. In addition, the module can also assist managers in optimizing the allocation of resources such as manpower, equipment and materials, and improve construction efficiency by analyzing resource consumption and equipment operation status. At the same time, the analysis results can be converted into charts or reports for reference and decision-making by managers, further improving the intelligent management level of the construction site. In the intelligent data analysis function, the following formula is used to represent the prediction model of construction progress: ,in, Pt Indicates the current construction progress. Tt Indicates the planned construction progress. α is an adjustment factor used to control the speed of progress updates.

[0033] As a preferred embodiment of the above, the technical implementation of the intelligent service evaluation module includes: The analytic hierarchy process is used to assign weights to the evaluation indicators; Integrate various data on the construction site and provide a basis for site management through data analysis and mining; Design an intelligent evaluation model to perform different lock-scope object analyses on the relevant data of the login roles according to different login roles, and calculate various analysis and evaluation results.

[0034] Specifically, the module first decomposes the evaluation index system, takes indicators such as construction quality, resource utilization efficiency, and safety risk control as core evaluation dimensions, and determines the weights of each dimension through expert scoring or historical data analysis, and constructs a multi-level index weight system; in the data analysis stage, the module integrates multi-source data of the construction site, including construction progress, equipment operation status, personnel management information and environmental monitoring data, and discovers key influencing factors through data mining technology to provide managers with scientific decision-making basis, for example, analyzing the correlation between equipment utilization and progress deviation, identifying the optimization direction of resource allocation, or providing improvement suggestions for the construction site in combination with safety risk data; in addition, the module designs an intelligent evaluation model that can dynamically adjust the analysis scope according to the permissions and concerns of the login role. For example, for the project manager account, the module focuses on analyzing the construction progress and resource allocation data; for the safety manager account, the safety risk and hidden danger analysis data are locked. After the module classifies and locks the data of different roles, it combines the weighted scoring model to calculate accurate analysis and evaluation results. This embodiment uses a weighted scoring formula to comprehensively evaluate the construction quality: ,in, Q is the comprehensive evaluation score, wi It is i The weight of the evaluation index, qi It is i The score of the evaluation index.

[0035] Embodiment 2 Based on the same inventive concept as the smart construction site large model working system in the aforementioned embodiment, the present invention also provides a smart construction site large model working method, such as Figure 2 As shown, the method includes: S1: Receive natural language instructions input by users, perform semantic analysis through the smart construction site model, and identify the user's task intention; S2: Based on the identified task intent, decompose the instructions and coordinate the execution order of different tasks; S3: For consultation requests from the construction industry, the big model retrieves expertise in the construction field, including technical specifications, project cases and expert experience, and generates professional consultation results related to the directive; S4: Collect environmental data and equipment operation data at the construction site, and analyze them in combination with historical construction data, including real-time monitoring of construction progress, trend prediction, and potential risk warning; S5: Carry out service evaluation based on the construction analysis results, generate evaluation results of construction services through weighted analysis methods, and propose optimization suggestions based on relevant data; S6: Output analysis results, evaluation reports and optimization suggestions in a visual way to support users to view in real time and make further decisions.

[0036] The above working method in the present invention can be effectively implemented through the smart construction site large model working system, and the technical effects that can be achieved are as described in the above embodiments, which will not be repeated here.

[0037] As a preferred embodiment of the above, Figure 3 As shown in the figure, the smart construction site model is trained through the industry training set to build the industry training set, including: A10: Collect structured and unstructured data, pre-process the collected data, classify and label the unstructured data, and determine the model learning objectives; A20: Expand the amount of training data through data enhancement technology, integrate multimodal data, and build a unified feature space; A30: Based on industry data, industry terms and rules are structured through entity extraction and relationship extraction to build an industry knowledge graph.

[0038] Specifically, first, data collection and integration are carried out to collect structured data such as construction site management data, sensor data, equipment operation data, and unstructured data such as construction specifications, industry standards, policies and regulations. The data sources include sensor equipment, construction management systems, industry reports, academic literature, industry databases (such as building materials and construction equipment databases), as well as expert interviews and experience summaries. At the same time, multimodal data (text, images, videos, time series, etc.) are integrated to form a preliminary training data set; then, the collected data is cleaned and preprocessed, including deduplication, missing value processing, outlier detection, format standardization, and time series alignment to ensure data consistency and quality. For unstructured data, natural language processing technology is used to extract key information, such as parsing safety rules from construction specifications and extracting key information such as construction methods from technical documents. After preprocessing, data annotation and enhancement are performed. By classifying and labeling the data, for example, safety specification data is annotated as the "safety management" category and construction cases are annotated as the "construction method" category. At the same time, the training data set is expanded through data enhancement techniques, such as text replacement, time series perturbation, and multimodal data fusion, to improve the diversity and generalization ability of data; then, based on unstructured industry data, an industry knowledge graph is constructed, and through entity extraction and relationship extraction techniques, key entities such as "construction personnel", "construction materials", "equipment operation" and their relationships (such as "construction personnel wearing safety helmets") are identified and associated, and information from different data sources is integrated to form a unified structured knowledge graph to provide support for the reasoning and knowledge application of the large model; finally, the constructed training set and large model are evaluated and optimized, and the performance of the model is tested through a variety of evaluation indicators (such as precision, recall rate, F1 score, AUC, etc.) to ensure its accuracy and real-time performance in tasks such as construction site monitoring, progress prediction, and safety warning. At the same time, the knowledge base is regularly updated through automation technology (such as crawlers to capture the latest regulations) and expert feedback mechanisms to ensure the timeliness and integrity of the data. The above method provides comprehensive and reliable industry data support for the training of the smart construction site large model.

[0039] As a preferred embodiment of the above, Figure 4 As shown in the figure, a large-scale pre-trained language model is used as a baseline, and label flipping is performed to enhance the data in the industry training set, including: B10: Use the BERT model to train a label classifier to classify the labels of the original data; B20: Generate new enhanced samples using the T5 model, generate a mask by randomly deleting part of the input content, and predict part of the mask content through the T5 model to form a new sample set; B30: Use the label classifier to screen the generated enhanced samples, evaluate the sample quality based on the confidence level, and select high-quality samples; B40: Combine the enhanced samples with the original samples to retrain the classifier and related modules of the large model.

[0040] Specifically, first, a label classifier is trained using the BERT model to accurately identify the category label of each data sample by classifying the labels of the original industry data. This step provides a basic guarantee for subsequent data enhancement and screening, ensuring that the label classification of the enhanced samples meets the specific standards of industry data; a mask area is generated by randomly deleting part of the original data, and the mask part is predicted by the T5 model to generate the completed content, thereby forming a new set of enhanced samples. For example, for construction text data, key phrases or sentence fragments can be randomly deleted, and the T5 model will predict these mask contents to generate reasonable new samples; the generated enhanced samples are quality screened using the trained BERT label classifier, and the confidence of the classification results of the samples is evaluated to retain high-quality enhanced samples, and samples with low quality or that do not match the original data labels are eliminated to ensure the accuracy and effectiveness of the enhanced data; the screened high-quality enhanced samples are merged with the original sample set to construct an extended training data set containing the original and enhanced samples, and the classifier and related modules of the large model are retrained using this data set, thereby improving the generalization ability of the model and its adaptability in industry scenarios, and further enhancing its ability to handle complex tasks.

[0041] As a preferred embodiment of the above, Figure 5 As shown in the figure, the large model of the smart construction site is trained in parallel, including: C10: According to the structure and parameter quantity of the large model, the various layers of the model are reasonably divided into multiple parts and allocated to different computing devices; C20: Divide the training data into multiple batches, and assign each batch to different computing devices, so that each device can independently obtain enough data for training; C30: On each computing device, forward propagation calculations are performed independently based on the segmented model and data. Each device calculates its assigned part and passes the result to the computing device of the next layer. C40: On each computing device, independently perform back propagation calculation based on the segmented model and data, generate gradient information for each part, and pass the gradient information to the computing device of the upper layer; C50: After each computing device completes the back-propagation calculation, the gradient information calculated by each device is aggregated to generate the gradient of the overall model and update the model parameters.

[0042] Specifically, the parallel training of the large model of the smart construction site adopts a distributed training method based on a multi-class parallel synchronization framework, and designs a comprehensive training framework of tensor parallelism, data parallelism, and pipeline parallelism to cope with the training needs of large models with hundreds of billions of parameters, thereby achieving a more efficient and stable training process. Specifically, tensor parallelism divides each tensor into multiple blocks, each shard is assigned to a different GPU device for independent processing, and the shard results are synchronized at the end of each training step. This horizontal splitting method effectively reduces the video memory pressure of a single GPU and improves computing efficiency. In addition, the zero-redundancy optimizer (ZeRO) divides and distributes model parameters to different devices, and only reconstructs the complete tensor during forward propagation or backward propagation. Combined with pipeline parallelism and zero-redundancy technology, it further reduces communication overhead and memory usage, and optimizes video memory usage; in terms of data parallelism, the complete model copy is copied to multiple computing devices, and the training data is divided into multiple batches, each batch is assigned to different devices for independent calculation. After each training step, the model parameters are updated synchronously to make full use of device resources and accelerate the training process. Pipeline parallelism is based on the concept of pipeline, dividing the model calculation process into multiple stages, and each stage is assigned to different devices for parallel calculation. For example, the layers of the model are divided vertically, and each GPU is only responsible for calculating the forward propagation and back propagation of a part of the layers. By processing different batch data at different stages of the pipeline, the training throughput and speed are significantly improved. Combined with multiple types of parallel operations, the training process of the smart construction site large model includes steps such as model segmentation, data segmentation, forward propagation, back propagation, and gradient aggregation. First, the layers or parameters of the large model are divided into multiple parts and reasonably allocated to different devices to ensure balanced computing load; secondly, the training data is divided into multiple batches and allocated to each device, so that each device can independently obtain sufficient data for training, while reducing data transmission between devices. During the training process, each device independently processes the assigned model slices and data in the forward propagation stage and passes the results to the next layer of devices; in the back propagation stage, each device independently calculates the gradient information and passes it to the previous layer; finally, the calculation results of all devices are globally integrated through gradient aggregation to generate the overall model gradient and complete the model parameter update. Through this parallel training method, the large model of the smart construction site can fully utilize the computing power of multiple computing devices, effectively reduce the resource consumption of a single device, improve training efficiency, and provide support for the efficient training of complex large-scale models.

[0043] Although the present application has been described in conjunction with specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present application. Accordingly, this specification and the accompanying drawings are merely exemplary illustrations of the present application as defined herein, and are deemed to have covered any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalent technologies, the present application is intended to include these modifications and variations.

Claims

1. A smart construction site large model working system, characterized in that: include: The basic large model is used to understand the natural language instructions input by users, provide semantic parsing, multimodal data processing and context memory functions, and adapt to the professional needs of the construction industry through fine-tuning; Agent, which is used to receive natural language instructions from users, parse the instructions, call corresponding functional modules based on task intent, and perform task scheduling between modules; The industry expert consultation module is used to process users' industry consultation requests and generate expert-level consultation responses by searching the industry knowledge base; Intelligent data analysis module, which is used to collect construction process data at the construction site, conduct statistical analysis and trend forecasting in combination with the construction database, and generate construction data analysis reports; The intelligent service evaluation module is used to respond to users' requests for the generation of construction data analysis reports and auxiliary decision-making, and to generate service evaluation reports in combination with a priori evaluation algorithms; The application, as the entry point for users to interact with the system, is used to receive user input, display analysis results, and generate visual data reports.

2. The smart construction site large model working system according to claim 1 is characterized in that: The industry expert consulting module, intelligent data analysis module and intelligent service evaluation module all include a data processing layer and an API layer. The data processing layer corresponds to the industry knowledge base, construction database and evaluation algorithm library respectively, and the API layer corresponds to internal APIs, external APIs and decision support APIs respectively, and is used to exchange data with external systems.

3. The smart construction site large model working system according to claim 1 is characterized in that: The Agent calls functions and APIs of different target modules through Function Call according to different task types.

4. The smart construction site large model working system according to claim 1 is characterized in that: The technical implementation of the industry expert consultation module includes: Based on the knowledge base of the construction industry, including technical specifications, project cases and expert experience, as the source of expertise for the model; Parse user input queries through natural language processing technology, understand query intent and retrieve relevant information from the knowledge base; Combined with machine learning algorithms, it dynamically recommends the best expert answers and solutions based on user questions.

5. The smart construction site large model working system according to claim 1 is characterized in that: The technical implementation of the intelligent data analysis module includes: Deploy a variety of sensors and monitoring equipment to collect real-time environmental data and equipment operation data on the construction site; Combined with big data analysis technology, real-time monitoring, progress forecasting and early warning of construction data can be carried out to assist managers in optimizing resource allocation.

6. The smart construction site large model working system according to claim 1 is characterized in that: The technical implementation of the intelligent service evaluation module includes: The analytic hierarchy process is used to assign weights to the evaluation indicators; Integrate various data on the construction site and provide a basis for site management through data analysis and mining; An intelligent evaluation model is designed to perform different lock-scope object analyses on the relevant data of the login roles according to different login roles, and various analysis and evaluation results are calculated.

7. A method for working with a large model of a smart construction site, characterized in that: The method comprises: Receive natural language instructions input by users, perform semantic analysis through the smart construction site model, and identify the user's task intent; Based on the identified task intentions, instructions are decomposed and the execution order of different tasks is coordinated; In response to consultation requests from the construction industry, we use the big model to retrieve expertise in the construction field, including technical specifications, project cases and expert experience, and generate professional consulting results related to the directive; Collect environmental data and equipment operation data at the construction site and analyze them in combination with historical construction data, including real-time monitoring of construction progress, trend prediction, and potential risk warning; Conduct service evaluation based on construction analysis results, generate construction service evaluation results through weighted analysis methods, and propose optimization suggestions based on relevant data; The analysis results, evaluation reports and optimization suggestions are output in a visual way, supporting users to view them in real time and make further decisions.

8. The smart construction site large model working method according to claim 7 is characterized in that: The smart construction site big model is trained by the industry training set, and the industry training set is constructed, including: Collect structured and unstructured data, pre-process the collected data, classify and label the unstructured data, and determine the model learning objectives; Expand the amount of training data through data enhancement technology, integrate multimodal data, and build a unified feature space; Based on industry data, industry terms and rules are structured through entity extraction and relationship extraction to build an industry knowledge graph.

9. The smart construction site large model working method according to claim 8, characterized in that: We use a large-scale pre-trained language model as a baseline and perform label flipping to enhance the data in the industry training set, including: Use the BERT model to train a label classifier to classify the labels of the original data; Generate new enhanced samples using the T5 model, generate masks by randomly deleting part of the input content, and use the T5 model to predict part of the mask content to form a new sample set; Use the label classifier to screen the generated enhanced samples, evaluate the sample quality based on the confidence level, and select high-quality samples; The enhanced samples are combined with the original samples to retrain the classifier and related modules of the large model.

10. The smart construction site large model working method according to claim 7, characterized in that: The large model of the smart construction site is trained in parallel, including: According to the structure and parameter quantity of the large model, the various layers of the model are reasonably divided into multiple parts and allocated to different computing devices; Divide the training data into multiple batches, and assign each batch to different computing devices so that each device can independently obtain enough data for training; On each of the computing devices, forward propagation calculations are independently performed based on the segmented model and data, each device calculates its assigned part, and transmits the result to the computing device of the next layer; On each of the computing devices, independently perform back propagation calculation based on the segmented model and data, generate gradient information of each part, and transmit the gradient information to the computing device of the upper layer; After each of the computing devices completes the back-propagation calculation, the gradient information calculated by each device is aggregated to generate the gradient of the overall model and update the model parameters.

Images