A construction management system based on whole-process digitalization collaboration

Abstract

This invention belongs to the technical field of construction management systems. It discloses a construction management system based on full-process digital collaboration, comprising an edge-aware collaboration module, a data fusion collaboration module, an intelligent decision-making and governance module, a knowledge learning and evolution module, and a cross-level collaborative control module. Utilizing improved multi-source heterogeneous data fusion and precise spatiotemporal matching, domain-specific fine-tuning of MLLM and federated learning joint optimization, and the principles of a construction full-process knowledge graph and closed-loop problem-solving algorithms, through configuring dynamic spatiotemporal weight factors, multimodal data fine-tuning and multi-project federated training, unstructured data mining, and closed-loop management processes, it significantly improves the efficiency of data collaborative flow throughout the construction process, algorithm recognition accuracy and generalization ability, and knowledge reuse efficiency. This shortens collaborative response time, reduces the recurrence rate of similar risk defects, and solves the technical problems of isolated data, poor algorithm adaptability, scattered knowledge storage, and lack of full-process problem traceability in existing technologies.

Description

Technical Field

[0001] This invention belongs to the technical field of construction management systems, and in particular relates to a construction management system based on full-process digital collaboration. Background Technology

[0002] Construction project management is characterized by its broad process scope (covering all stages of surveying, design, construction, testing, and acceptance), diverse participants (collaboration among construction, design, construction, and supervision parties), and complex data types (including structured sensor data, unstructured drawings / videos, and semi-structured reports and logs). Furthermore, each stage is interconnected, creating an urgent need for data flow. With the surge in demand for large-scale infrastructure projects and multi-project management across regions, traditional decentralized management models and single-stage digital systems are no longer sufficient to meet the needs of collaborative control throughout the entire process. There is an urgent need for a comprehensive digital collaborative construction management system that can break down data barriers between stages, adapt to the characteristics of construction scenarios, and support multi-project collaborative optimization. This system will enable the transformation of construction management from "decentralized control" to "global collaboration," and from "experience-driven" to "data-driven."

[0003] There are still two major issues that need to be addressed in the existing technology:

[0004] (1) Data is isolated at each stage and collaborative response is lagging behind. Existing digital systems mostly focus on a single construction stage (such as only covering the construction process or testing), lacking a full-process data integration mechanism. Design change information cannot be quickly synchronized to the construction and testing stages, and on-site measured data is difficult to be fed back to the design stage, resulting in low efficiency of cross-stage data flow and lagging collaborative decision-making, which in turn leads to problems such as construction rework and schedule delays.

[0005] (2) The general algorithm has poor adaptability and lacks the ability to optimize multiple projects. The artificial intelligence algorithms carried by the existing system are mostly general models and have not been customized for the professional characteristics of construction scenarios (such as construction equipment, safety specifications, and non-standard documents), resulting in insufficient accuracy in target recognition, document parsing, and risk warning. At the same time, the algorithm models are mostly trained based on single-project data, which cannot achieve multi-project knowledge sharing and joint model optimization. The generalization ability is weak and it is difficult to adapt to the engineering construction needs of different types and scenarios. Summary of the Invention

[0006] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a construction management system based on full-process digital collaboration. Utilizing improved multi-source heterogeneous data fusion and precise spatiotemporal matching, domain-specific fine-tuning of MLLM and federated learning for joint optimization, and the principles of a construction process knowledge graph and closed-loop problem-solving algorithms, this system achieves significant improvements in the efficiency of data collaborative flow throughout the construction process, algorithm recognition accuracy and generalization ability, and knowledge reuse efficiency through the configuration of dynamic spatiotemporal weight factors, multimodal data fine-tuning and multi-project federated training, unstructured data mining, and closed-loop management processes. This shortens collaborative response time, reduces the recurrence rate of similar risk defects, and solves the technical problems of isolated data, poor algorithm adaptability, scattered knowledge storage, and lack of full-process problem traceability in existing technologies.

[0007] The technical solution adopted in this invention is as follows: a construction management system based on full-process digital collaboration, including an edge-aware collaboration module, a data fusion collaboration module, an intelligent decision-making and governance module, a knowledge learning and evolution module, and a cross-level collaborative control module.

[0008] The edge-sensing collaboration module is deployed at the construction project site. It collects multi-source sensing data from each stage of the entire construction process and performs local preprocessing to construct a spatiotemporal data sub-matrix for each construction stage. Simultaneously, it executes local edge-collaborative early warning and command response operations. The various stages of the entire construction process in the edge-sensing collaboration module include the measurement stage, design stage, construction stage, inspection stage, and acceptance stage. The multi-source sensing data includes BeiDou high-precision positioning data and 3D laser scanning data from the measurement stage; digitized drawing data and BIM model foundation data from the design stage; IoT sensor data, high-definition monitoring video data, and personnel / equipment positioning data from the construction stage; engineering quality inspection data and panoramic image data from the inspection stage; and completion acceptance data and measured data from the acceptance stage. The preprocessing includes data standardization, coarse noise filtering, and structured transformation.

[0009] The edge-aware collaboration module also includes a local edge AI agent and a project-level basic model unit. The local edge AI agent embeds construction scenario-based computer vision algorithms and data preprocessing algorithms to achieve real-time anomaly identification and local early warning of on-site data. The project-level basic model unit is equipped with a lightweight BIM model and a digital twin basic model adapted to the construction project to achieve preliminary association between on-site perception data and the basic model.

[0010] As a preferred technical solution of this scheme, the data fusion and collaboration module receives spatiotemporal data sub-matrices from each construction stage, embeds an improved multi-source heterogeneous data fusion algorithm and a spatiotemporal precise matching algorithm to complete the integration and spatiotemporal alignment of the entire process data, embeds the ICP iterative nearest point algorithm to perform spatiotemporal precise matching, and performs hierarchical encrypted storage of the data.

[0011] The data fusion and collaboration module is specifically used to perform the following operations:

[0012] It receives spatiotemporal data sub-matrices for each construction stage, and configures dynamic spatiotemporal weight factors for different stages and types of sensing data based on the stage characteristics, data types and data credibility of the entire construction process. The weight of the on-site measured data in the construction stage is higher than that of the simulated data in the design stage, and the weight of the monitoring data at the supervision end is higher than that of the data collected at the construction end.

[0013] An improved multi-source heterogeneous data fusion algorithm is used to perform weighted fusion of the spatiotemporal data sub-matrices at each stage;

[0014] By embedding the ICP iterative nearest point algorithm, precise spatiotemporal matching of BIM model data, Beidou positioning data, construction area GIS data and unified spatiotemporal data matrix of the entire construction process is achieved, ensuring data collaboration and consistency across stages and spaces;

[0015] The attribute-based encryption (ABE) algorithm is used to perform hierarchical encryption storage of the unified spatiotemporal data matrix throughout the construction process. Storage levels are divided according to the sensitivity of the data, and different data access permissions are configured for different participants.

[0016] As a preferred technical solution of this scheme, the intelligent decision-making and governance module constructs a dynamic digital twin collaborative model for the entire construction process based on a unified spatiotemporal data matrix. It embeds a domain-specific fine-tuned multimodal large language model and a construction scenario-based computer vision algorithm to realize intelligent analysis and decision-making, multi-participant collaborative management, and closed-loop problem solving for the entire construction process, and generates collaborative control instructions for the entire construction process.

[0017] The intelligent decision-making and governance module specifically includes a full-process digital twin collaborative unit, an intelligent analysis and decision-making unit, a multi-participant collaborative management unit, and a closed-loop problem-solving unit. The collaborative work of each unit enables full-process control of construction.

[0018] As a preferred technical solution of this scheme, the knowledge learning and evolution module relies on the unified spatiotemporal data matrix of the entire construction process to construct a knowledge graph of the entire construction process, embeds the federated learning full-process model training algorithm, realizes the joint optimization of multi-project algorithm models under the premise of ensuring data privacy, and completes the structured storage and reuse push of construction knowledge.

[0019] The knowledge learning and evolution module is specifically used to perform the following operations:

[0020] Based on a domain-specific fine-tuned multimodal large language model, entity extraction, relation mining, and semantic analysis are performed on unstructured data in a unified spatiotemporal data matrix throughout the construction process. A knowledge graph covering all stages and specialties of the entire construction process is constructed. After being fine-tuned by construction domain-annotated images and professional documents, the multimodal large language model has the ability to answer construction-related questions, parse unstructured documents, and perform cross-modal data association analysis.

[0021] The full-process model training algorithm embedded with federated learning uses the stage spatiotemporal data sub-matrices of each construction project as federated training samples. Without disclosing sensitive project data, it realizes multi-project joint optimization of construction scenario-based computer vision algorithm, multimodal large language model and risk warning model. The full-process model training algorithm of federated learning configures stage-based model training weights according to the stage characteristics of the entire construction process.

[0022] By constructing a knowledge reuse and push unit, excellent solutions, management experience, and process standards accumulated throughout the construction process are linked with a knowledge graph. Based on the current construction project's stage, type, and on-site issues, knowledge can be accurately pushed and reused.

[0023] As a preferred technical solution of this scheme, the cross-level collaborative management and control module is configured with a federated learning mechanism and a full-process policy execution mechanism to realize bidirectional data flow, functional collaboration and unified implementation of management rules among the edge-aware collaborative module, data fusion collaborative module, intelligent decision-making governance module and knowledge learning evolution module. It receives the management and control instructions from the intelligent decision-making governance module and sends them to each execution end, while collecting execution feedback data and sending it back to each module.

[0024] The federated learning mechanism in the cross-level collaborative management and control module enables the push and update of the optimized algorithm model from the knowledge learning and evolution module to the edge perception collaboration module and the intelligent decision-making governance module. The local edge AI agent of the edge perception collaboration module optimizes the local early warning rules based on the updated algorithm model. The full-process policy execution mechanism embeds the technical standards, management specifications, safety regulations, and customized project management rules of the construction industry into the edge perception collaboration module, data fusion collaboration module, intelligent decision-making governance module, and knowledge learning and evolution module to ensure that the management behavior of each construction stage and each participant conforms to unified standards, while monitoring the rule execution in real time.

[0025] As a preferred technical solution of this scheme, the system further includes an execution feedback acquisition unit, which is used to collect the execution result data of each execution end to the control command, compare the execution result data with the expected control target of the intelligent decision-making and governance module, and calculate the deviation value; when the deviation value continues to exceed the preset threshold, the deviation information is sent back to the data fusion and collaboration module and the knowledge learning and evolution module. The data fusion and collaboration module adjusts the dynamic spatiotemporal weight factor of the improved multi-source heterogeneous data fusion algorithm according to the deviation information, and the knowledge learning and evolution module optimizes the training parameters of the federated learning full-process model training algorithm according to the deviation information.

[0026] The beneficial effects of the present invention after adopting the above structure are as follows:

[0027] (1) By using an improved multi-source heterogeneous data fusion algorithm and precise spatiotemporal matching, and by configuring dynamic spatiotemporal weight factors for the entire construction process and setting cross-stage data collaborative verification rules, the efficient integration and spatiotemporal alignment of multi-stage and multi-type data are achieved, the efficiency of data collaborative flow in the entire construction process is improved, and the collaborative response time for design changes and construction adjustments is shortened; the technical problems of isolated data and delayed collaboration in each stage in the existing technology are solved.

[0028] (2) By combining domain-specific fine-tuning of multimodal large language model (MLLM) with federated learning, the model is fine-tuned using multimodal data (professional documents, labeled images, sensor logs) in the construction field and federated training with desensitized data from multiple projects. This improves the accuracy of construction target recognition and quality defect detection, and enhances the model's cross-project generalization and adaptation capabilities. It solves the technical problems of poor adaptability of general algorithms to construction scenarios and weak generalization capabilities of single-project training models in existing technologies, and achieves deep matching between the algorithm and construction professional scenarios.

[0029] (3) By using the knowledge graph construction and closed-loop problem-solving algorithm of the entire construction process, a professional knowledge graph is constructed through unstructured data entity extraction and relationship mining. Combined with the closed-loop management process of "identification-dispatch-rectification-acceptance-closure", the recurrence rate of similar safety risks and quality defects is reduced and the efficiency of construction knowledge reuse is improved. It solves the technical problems of scattered construction knowledge storage, difficulty in reuse, and lack of full-process traceability in problem handling in the existing technology, and realizes the structured accumulation of knowledge and precise control of problems. Attached Figure Description

[0030] The accompanying drawings are provided to further understand the present invention and form part of the specification. They are used together with the embodiments of the present invention to explain the invention and do not constitute a limitation thereof.

[0031] Figure 1 This is a block diagram of the overall module of the construction management system based on full-process digital collaboration proposed in this invention;

[0032] Figure 2 This is a flowchart of the multi-source heterogeneous data fusion and spatiotemporal precise matching of the construction management system based on full-process digital collaboration proposed in this invention;

[0033] Figure 3 This is a flowchart illustrating the intelligent decision-making governance and collaborative management of the entire construction process based on a fully digital collaborative construction management system proposed in this invention. Detailed Implementation

[0034] Example 1, as Figures 1-3 As shown, a construction management system based on full-process digital collaboration includes an edge-aware collaboration module, a data fusion collaboration module, an intelligent decision-making and governance module, a knowledge learning and evolution module, and a cross-level collaborative control module. Each module achieves bidirectional data flow through a standardized API data interface. The cross-level collaborative control module is configured with a federated learning mechanism and a full-process policy execution mechanism, binding each module into an organic whole to achieve digital collaborative control of the entire construction process.

[0035] The edge perception collaboration module is deployed at the construction site of each construction project. It uses multiple types of perception terminals to collect and preprocess multi-source perception data throughout the construction process, constructs a stage-specific spatiotemporal data sub-matrix, and simultaneously completes local edge collaborative early warning and command execution.

[0036] The data fusion and collaboration module is deployed on an enterprise-level cloud platform. It receives the phased spatiotemporal data sub-matrices of each project and constructs a unified spatiotemporal data matrix for the entire construction process through an improved multi-source heterogeneous data fusion algorithm and a spatiotemporal precise matching algorithm, thereby achieving the integration, alignment and secure storage of data throughout the entire process.

[0037] The intelligent decision-making and governance module is deployed on an enterprise-level cloud platform. It relies on a unified spatiotemporal data matrix for the entire construction process to build a dynamic digital twin model, enabling intelligent analysis and decision-making, collaborative management of multiple stakeholders, and closed-loop problem solving throughout the construction process, and generating control instructions.

[0038] The knowledge learning and evolution module is deployed on an enterprise-level cloud platform to build a knowledge graph of the entire construction process. Through federated learning, it achieves joint optimization of multi-project algorithm models and completes the structured storage and reuse of knowledge.

[0039] The cross-level collaborative management and control module enables collaborative linkage between the edge perception collaborative module, the data fusion collaborative module, the intelligent decision-making and governance module, and the knowledge learning and evolution module. It issues control and control instructions to each execution terminal and collects execution feedback data to send back to each module, thereby realizing closed-loop management of the system.

[0040] Example 2: Based on the above examples, this example details the specific implementation of the edge-aware collaboration module:

[0041] The edge-aware collaboration module includes multiple types of sensing terminals, a data preprocessing submodule, a local edge AI agent, and project-level basic model units, adapting to the data acquisition needs of the entire process of construction surveying, design, construction, testing, and acceptance.

[0042] (1) Configuration of multiple types of sensing terminals: During the measurement phase, deploy Beidou high-precision positioning instrument (centimeter-level RTK positioning) and 3D laser scanner to collect spatial data of terrain and control points; during the design phase, deploy drawing digitization acquisition terminal and BIM model lightweight terminal to collect drawing digitization data and BIM model basic data; during the construction phase, deploy IoT sensors (collecting equipment operating parameters and environmental parameters), 4-megapixel high-definition monitoring camera (with infrared night vision and anti-fog function), and smart safety helmet (collecting personnel positioning and operation behavior data); during the inspection phase, deploy engineering quality inspection data acquisition instrument and panoramic imaging equipment to collect inspection data and engineering real-scene images; during the acceptance phase, deploy actual measurement data acquisition instrument and data digitization terminal to collect completion acceptance data and actual measurement data.

[0043] (2) Data Preprocessing Submodule: The collected multi-source sensing data are sequentially processed using data standardization, coarse noise filtering, and structure transformation. Data standardization employs min-max standardization to map all data to... Intervals are used to eliminate dimensional differences; coarse noise filtering employs... The criteria remove outlier data that deviates from the mean by three times the standard deviation; unstructured data (videos, images, documents) are transformed into structured data to extract core feature information, and finally a spatiotemporal data submatrix of the construction stage is constructed according to time and space dimensions.

[0044] (3) Local edge AI agent: embeds construction scenario-based computer vision algorithm and data preprocessing algorithm to perform real-time analysis of on-site video and image data, identify safety / quality problems (construction personnel not wearing safety helmets, not wearing reflective vests, abnormal equipment operating parameters, lack of edge protection), realize local real-time early warning, and push the early warning information through the on-site sound and light alarm and the mobile APP of the management personnel; at the same time, it receives the control instructions (equipment shutdown, construction area control) issued by the upper level and directly sends them to the on-site execution end.

[0045] (4) Project-level basic model unit: equipped with a lightweight BIM model and digital twin basic model adapted to this project, and preliminarily associate the perception data collected on site with the model (mapping personnel / equipment positioning data to the spatial location of the BIM model) to achieve preliminary visualization of the on-site working conditions.

[0046] Example 3: Based on the above examples, this example details the specific implementation of the data fusion and collaboration module, focusing on the implementation of the improved multi-source heterogeneous data fusion algorithm and the spatiotemporal precise matching algorithm.

[0047] (1) Construction of a full-process data platform: A classified database is built according to the five stages of construction surveying, design, construction, testing and acceptance. A distributed storage architecture is adopted to support the unified storage of structured, semi-structured and unstructured data. A unique spatiotemporal identifier is established for the data of each stage to realize the rapid retrieval of data.

[0048] (2) Execution of the improved multi-source heterogeneous data fusion algorithm:

[0049] To address the shortcomings of traditional multi-source heterogeneous data fusion algorithms in adapting to construction scenarios, a five-step improvement process is implemented to achieve efficient data fusion across the entire process. The specific improvement steps, combined with actual implementation, are as follows:

[0050] Step t1: Reconstruct and improve the weighting system, abandoning the traditional fixed weight allocation method. Based on the stage characteristics, data types, and data reliability of the entire construction process, configure dynamic spatiotemporal weight factors. Clearly define that the weight of on-site measured data in the construction stage is higher than that of simulated data in the design stage, and the weight of inspection data from the supervision end is higher than that of data collected from the construction end. Specifically, according to the stage characteristics of the entire construction process, set the weight of on-site measured data in the construction stage to 0.8-0.9, the weight of simulated data in the design stage to 0.5-0.6, the weight of supervision data in the inspection / acceptance stage to 0.9-0.95, and the weight of construction data to 0.7-0.8. Through the reconstruction of the weighting system, achieve deep adaptation of weights to the construction scenario.

[0051] Step t2: Improved hierarchical preprocessing of heterogeneous data. To address the issues of heterogeneous dimensions and incompatible features in multi-source construction data, this step replaces the traditional single preprocessing mode. It performs three-layer hierarchical preprocessing on the spatiotemporal data sub-matrices of each stage. First, it normalizes and removes redundancy from homogeneous data of the same stage and type. Then, it performs feature mapping and dimension unification on heterogeneous data of different types in the same stage. Finally, it performs spatiotemporal label completion processing on data of different stages and cross-types (relying on the spatiotemporal identifiers of the full-process data platform), laying the data foundation for subsequent weighted fusion.

[0052] Step t3: Improve the construction process logic verification. Add a process logic verification step. Combine the construction process specifications and process logic of building engineering, set process collaboration verification rules for cross-stage and cross-type data, and simultaneously set the collaboration verification threshold of design model parameters and actual measured parameters on the construction site to 3%. When the deviation between the two is ≤3%, it is judged as valid. When the deviation is >3%, the data review process is triggered. The design, construction and supervision parties confirm the validity of the data, remove invalid and erroneous data, and retain only the valid data that conforms to the process logic.

[0053] After data validation is completed, a validity check is performed to determine whether the data is valid. If invalid, return to step t2 to re-execute the heterogeneous data hierarchical preprocessing improvement until the data is valid.

[0054] Step t4: Spatiotemporal incremental fusion improvement, replacing the traditional full data fusion mode. Based on the unique spatiotemporal identifier of the spatiotemporal data sub-matrix of each construction stage, incremental weighted fusion driven by the spatiotemporal identifier is performed on the verified valid data. Only the core data that changes during the entire construction process (design change data, construction progress data, and detection exceeding the standard data) are fused and uploaded. The fusion cycle of full data is 24 hours, and the fusion cycle of incremental data is 5 minutes, which greatly improves the fusion efficiency.

[0055] Step t5: Iterative optimization and improvement of fused data. Cross-stage cross-validation is performed on the initially fused data. Based on the dynamic spatiotemporal weight factor, the spatiotemporal data sub-matrices of each stage are weighted and fused. The weighted fusion calculation formula is as follows: ;in To create a unified spatiotemporal data matrix for the entire construction process, For the first Dynamic spatiotemporal weighting factors for each construction stage For the first The spatiotemporal data sub-matrices of each construction stage are used to ultimately construct a unified spatiotemporal data matrix covering the entire construction process;

[0056] After fusion is completed, the fusion effect is judged to determine whether the fused data is valid. If it is invalid, return to step t1 to readjust the weights and perform fusion again until the fused data is valid.

[0057] (3) Execution of the precise spatiotemporal matching algorithm: The ICP iterative nearest point algorithm is embedded. Based on the spatial coordinates of the BIM model, the spatial coordinates of the Beidou positioning data, construction area GIS data, and on-site perception data are registered with the BIM model. The Euclidean distance between point clouds is minimized through iterative calculation. The Euclidean distance calculation formula is: ;in, For point With point European distance between them , , For point The three-dimensional coordinates , , For point The three-dimensional coordinates enable precise spatiotemporal matching of data, with registration errors controlled within 5cm, thus solving the problem of spatiotemporal misalignment of data.

[0058] After spatiotemporal registration is completed, error judgment is performed to determine whether the spatiotemporal registration error is ≤5cm. If it is not satisfied, return to step t5 to re-execute the data fusion iteration optimization and improvement until the registration error is ≤5cm.

[0059] (4) Data security encryption: The attribute encryption ABE algorithm is adopted to divide the construction data into three levels: public data, internal shared data, and core sensitive data. Public data (construction progress disclosure data) is open to all participants, internal shared data (construction process data) is open to the construction, construction and supervision units, and core sensitive data (BIM core design and cost data) is only open to designated personnel of the design unit and the construction unit. At the same time, data operation permissions (view, edit and download) are configured for each participant to ensure data security.

[0060] Example 4, based on the above examples, details the specific implementation of the intelligent decision-making and governance module. The intelligent decision-making and governance module specifically includes a full-process digital twin collaborative unit, an intelligent analysis and decision-making unit, a multi-participant collaborative management unit, and a closed-loop problem-solving unit. These units work collaboratively to achieve full-process control of construction. The core is the implementation of a full-process dynamic digital twin collaborative model and a closed-loop problem-solving algorithm, while simultaneously covering the full-link control of four core issues throughout the construction process: safety risks, quality defects, schedule deviations, and resource allocation imbalances.

[0061] (1) Full-process digital twin collaborative unit execution: Integrating BIM technology with a unified spatiotemporal data matrix for the entire construction process, a full-process dynamic digital twin collaborative model covering all stages of measurement, design, construction, testing, and acceptance is constructed. The models at each stage are seamlessly connected, and the update frequency of the model is consistent with the data collection frequency at the site (5 minutes / time).

[0062] This unit provides a visual spatial foundation and data benchmark support for the management of four core issues: For safety risks, it accurately marks high-risk work areas, edge protection points, and special equipment operating ranges in the model, achieving spatial positioning and visual display of risk points; for quality defects, it uses the design BIM model as a benchmark, synchronizing on-site measured data and inspection data to achieve accurate comparison between the model and the actual on-site entity; for schedule deviations, it links the construction schedule plan with model components, synchronizing the actual on-site construction progress to achieve visual comparison between the actual progress and the planned progress; and for resource allocation imbalances, it synchronizes the real-time location and quantity information of personnel, equipment, and materials in the model, achieving global visualization of the spatial distribution and occupancy of construction resources.

[0063] After design change information is entered, it can be synchronized to the corresponding location in the full-process dynamic digital twin collaborative model in real time and automatically pushed to relevant management personnel in the construction and testing stages; the actual working condition data of the construction site can be mapped to the model in real time to realize the visualized collaborative management and control of the entire construction process. Management personnel can operate the digital twin model through PC and mobile APP to realize remote virtual inspection.

[0064] (2) Intelligent analysis and decision-making unit execution: The core application of the construction scenario-based computer vision algorithm and the domain-specific fine-tuned multimodal large language model is used to complete the intelligent identification, analysis and root cause determination of the four core problems. Among them, the construction scenario-based computer vision algorithm is implemented according to the four-step execution process mentioned in the invention, and the improved YOLOv8 algorithm completes three customized improvements. The specific execution is as follows:

[0065] Step s1: Data preprocessing: The multi-source sensing data collected by the edge perception collaboration module is denoised, enhanced, and scaled. Noise (light interference, dust blurring) in the construction scene is removed by Gaussian filtering, and histogram equalization is used to enhance the image. All image data are uniformly scaled to a size of 1024×1024 pixels. At the same time, based on the spatiotemporal identifier of the unified spatiotemporal data matrix of the entire construction process, the corresponding construction stage and regional coordinate information are matched to the multi-source sensing data after denoising, enhancement, and scale normalization to generate a standardized construction scene perception dataset.

[0066] Step s2: Construction scene feature extraction: A feature extraction network is built using an improved YOLOv8 algorithm, completing three improvements, specifically:

[0067] Step s21 Backbone Network Improvement: Receive the standardized construction scene perception dataset generated in step s1 of claim 9, embed the lightweight channel-spatial attention mechanism (CBAM-Lite) of construction scene into the C2f module of YOLOv8, strengthen the material features of construction targets (steel, concrete) through channel attention branches, suppress background interference features (dust, shadows, dense scaffold textures) through spatial attention branches, improve the feature extraction efficiency of small targets (rebars, bolts, edge protection buckles), extract the feature maps of construction targets and defect foundations through preliminary convolution operations, and pass them to step s22;

[0068] Step s22 Neck Network Improvement: Receive the basic feature maps of construction targets and defects generated in step s21, adjust the fusion scale of the original FPN+PAN structure of YOLOv8, add a 1×1 convolutional layer to adapt to the multi-scale target detection requirements of construction scenarios, retain 8x downsampling feature fusion for large targets (cranes, tower cranes), and add a 4x downsampling feature fusion branch for small and medium targets (safety helmets, rebar misalignment) to solve the problem of feature loss of multi-scale targets in construction scenarios, complete multi-scale feature fusion enhancement, generate multi-scale fused feature maps of construction targets and defects, and pass them to step s23;

[0069] Step s23 Loss Function Improvement: Receive the multi-scale fusion feature map of construction targets and defects generated in step s22, replace the original CIoU loss function of YOLOv8 with the construction defect weighted EIoU loss function, assign 1.5 times the weight to construction safety and quality defects with high risk and low proportion (concrete micro-cracks, high risk of formwork instability), and assign 1.0 times the weight to conventional appearance defects, thereby improving the detection accuracy of high-risk construction defects. After loss calculation and feature optimization, a set of feature vectors of construction targets and defects is finally generated and synchronously transmitted to step s3 of claim 9.

[0070] Step s3: Target and Defect Identification and Classification: Construct a construction scenario-specific classifier, input the above feature vectors into the classifier, match the core construction targets (workers, cranes, scaffolding, formwork supports) and construction safety and quality defect feature library (concrete micro-cracks, rebar misalignment, missing edge protection, not wearing safety helmets) preset by the intelligent analysis and decision-making unit, and output the target / defect category, on-site spatial location coordinates, and identification confidence level (confidence threshold set to 0.85).

[0071] Step s4: Verification of recognition results: Spatial matching verification is performed between the recognition results and the full-process dynamic digital twin collaborative model of the full-process digital twin collaborative unit. If the identified defect location does not match the structural features of the construction area in the model (confidence level < 0.85), it is judged as a misidentification and removed, and only valid recognition results that pass the verification are retained. After the recognition results are verified, the validity is judged to determine whether the recognition results are valid. If invalid, return to step s2 to re-execute the construction scene feature extraction until the recognition results are valid.

[0072] Based on the effective identification results of construction targets and defects output by the computer vision algorithm in the construction scenario, the intelligent analysis and decision-making unit combines a domain-specific fine-tuned multimodal large language model to generate construction process problems, complete overall identification and aggregation, and then conduct in-depth analysis and root cause determination. Construction process problems refer to various abnormal problems that affect the core objectives of project safety management, quality compliance, schedule performance, and efficient resource utilization throughout all stages of construction surveying, design, construction, inspection, and acceptance. Construction process problems include: safety risks, quality defects, schedule deviations, and resource allocation imbalances. The specific analysis and execution logic of the intelligent analysis and decision-making unit for the above-mentioned construction process problems is as follows:

[0073] In response to safety risks, the intelligent analysis and decision-making unit identifies risky behaviors such as not wearing a safety helmet, lack of edge protection, and improper operation of special equipment, and analyzes the risk level, scope of impact, and root causes of safety risk-related issues throughout the entire construction process.

[0074] In response to quality defects, the intelligent analysis and decision-making unit identifies physical quality problems such as concrete cracks, rebar misalignment, and formwork deformation. It compares the defects with the current national building engineering design specifications and the project's established design requirements to determine the degree of exceedance of the defects in the entire construction process and analyzes the technological and management root causes of the quality defects in the entire construction process.

[0075] In response to schedule deviations, the intelligent analysis and decision-making unit compares the actual construction progress at the construction site with the approved construction schedule plan, calculates the lag time of each construction process and each construction area, and analyzes the core causes of schedule deviation problems throughout the entire construction process. The core causes include, but are not limited to, insufficient on-site workers, delayed supply of engineering materials, design changes and adjustments, and the impact of extreme weather.

[0076] In response to resource allocation imbalances, the intelligent analysis and decision-making unit analyzes abnormal resource allocation issues such as idle on-site workers, idle special equipment, and stockpiling or shortage of engineering materials, calculates the utilization efficiency deviation of various construction resources, and generates resource allocation optimization directions corresponding to the full-process construction problems of resource allocation imbalances.

[0077] After completing the in-depth analysis and root cause determination of all the problems in the entire construction process, the intelligent analysis and decision-making unit will generate corresponding safety risk warnings, quality control analysis, schedule coordination and scheduling and resource optimization allocation suggestions. The intelligent analysis and decision-making unit will then transmit the above safety risk warnings, quality control analysis, schedule coordination and scheduling and resource optimization allocation suggestions to the multi-participant collaborative management unit and the closed-loop problem-solving unit.

[0078] (3) Multi-participant collaborative management unit execution: Design role-stage-permission three-dimensional mapping rules, configure differentiated permissions for construction unit, design unit, construction unit and supervision unit (design unit can only edit BIM model data in the design stage, supervision unit can view and review test data in the construction / testing stage); realize full-process online collaborative management and control for four core issues: realize online circulation, approval and receipt of hidden danger rectification notice for safety risk warning and quality defect rectification; realize online collaborative approval of schedule adjustment plan and construction period extension application for schedule deviation; realize online approval and multi-party linkage confirmation of personnel, equipment and material scheduling application for resource allocation imbalance; at the same time, it covers the regular collaborative approval process such as progress payment application and design change application submitted by construction unit, and the core information (application amount, change content, scope of impact) is automatically parsed by multimodal big language model and pushed to the corresponding approval node, improving approval efficiency by more than 50%;

[0079] The collaborative approval results, design changes, and construction adjustment requirements are synchronized to the closed-loop problem-solving unit, ultimately generating a collaborative conclusion. After the collaborative approval process is completed, the results are judged to determine whether the collaborative approval has passed. If it has not passed, it returns to the intelligent analysis and decision-making unit to regenerate control suggestions and optimize approval materials until the collaborative approval is passed.

[0080] (4) Execution of closed-loop problem-solving algorithm: Combining the analysis results generated by the intelligent analysis and decision-making unit and the collaborative conclusions generated by the multi-participant collaborative management unit, the four major categories of problems found in the entire construction process, namely safety risks, quality defects, schedule deviations and resource allocation imbalances, are managed in a closed-loop manner according to the process of "identification-filing-dispatch-rectification-acceptance-closure".

[0081] The algorithm designs differentiated rectification priorities and rectification time limits based on the stage characteristics and problem types of the entire construction process: safety risk issues in the structural construction stage are of first priority, with a rectification time limit of 2 hours; quality defects related to the main structure are of second priority, with a rectification time limit of 24 hours; schedule deviations and resource allocation imbalances are of third priority, with a rectification time limit of 72 hours; and routine appearance quality defects in the decoration and finishing stage are of fourth priority, with a rectification time limit of 72 hours.

[0082] The system automatically generates a unique work order for each issue, binding the responsible personnel and rectification deadline. After rectification, acceptance documents and post-rectification image data are uploaded for online acceptance by the supervision unit and the construction unit. Once the acceptance is passed, the work order is closed. Issues that are not rectified on time will continuously send reminders to management personnel, achieving full-process traceability and control of issues. The results of issue handling and the optimization needs of full-process control are fed back to the intelligent analysis and decision-making unit, forming a closed loop of control needs. After the issue rectification and acceptance are completed, a closed-loop judgment is performed to determine whether the issue rectification is closed. If not, the rectification tracking and acceptance process is re-executed until the issue rectification is closed.

[0083] (5) Generation of collaborative control instructions for the entire construction process: After the work results of the full-process digital twin collaborative unit, intelligent analysis and decision-making unit, multi-participant collaborative management unit and closed-loop problem-solving unit are uniformly integrated by the intelligent decision-making and governance module, collaborative control instructions for the entire construction process are generated for four major categories of problems: safety risks, quality defects, schedule deviations and resource allocation imbalances. These instructions include the control object, control requirements, execution subject and completion time limit, and are pushed to the cross-level collaborative control module.

[0084] Example 5: Based on the above examples, this example details the specific implementation of the knowledge learning and evolution module. It utilizes a unified spatiotemporal data matrix for the entire construction process to construct a knowledge graph for the entire construction process, implement the federated learning model training algorithm, and achieve structured storage and reuse of construction knowledge. The specific implementation is as follows:

[0085] (1) Construction of a knowledge graph for the entire construction process: Based on a multimodal large language model with domain-specific fine-tuning, entity extraction (construction technology, material type, quality standard, problem type), relationship mining (association between “concrete pouring” and “curing technology”, attribution between “reinforcement misalignment” and “quality defect”) and semantic analysis are performed on unstructured data (design documents, inspection reports, safety specifications, construction logs, rectification plans) in the unified spatiotemporal data matrix of the entire construction process. This results in the construction of a knowledge graph covering various disciplines such as building structure, water supply and drainage, HVAC, and electrical engineering, as well as the stages of measurement, design, construction, inspection, and acceptance. The knowledge graph has more than 100,000 nodes and more than 500,000 edges, achieving structured storage of construction knowledge.

[0086] After the knowledge graph is built, an integrity check is performed to determine whether the knowledge graph has been built. If it has not been built, the entity extraction, relationship mining and semantic analysis steps are re-executed until the knowledge graph is built.

[0087] (2) Execution of the full-process model training algorithm for federated learning:

[0088] Sample selection: The spatiotemporal data submatrices of each construction project stage are used as federated training samples. Each sample is anonymized to remove sensitive information (project name, cost, core design).

[0089] Phased training weight configuration: The training weight is set to 0.4 for the construction phase, 0.3 for the inspection / acceptance phase, and 0.3 for the measurement / design phase, which aligns with the core needs of construction management;

[0090] Federated training execution: A federated averaging algorithm is used. Each construction project acts as a local training node. After completing model training locally, only the model parameters are uploaded to the central training node of the enterprise-level cloud platform. The central training node performs a weighted average of the model parameters from each local node. The weighted average calculation formula is as follows: ;in, To globally optimize model parameters, This represents the total number of local training nodes. For the first The percentage of samples from each local node. For the first The model parameters of each local node are used to generate a globally optimized model. Then, the parameters of the globally optimized model are pushed to each local node to update the model. Throughout the training process, the original data of each project is kept locally to ensure data privacy.

[0091] Model optimization: Through multiple rounds of federated training, the recognition accuracy of the construction scenario-based computer vision algorithm has been improved to over 92%, and the accuracy of the multimodal large language model in answering construction-related questions has been improved to over 88%.

[0092] After each round of federated training, it is determined whether the model accuracy meets the standard. If it does not meet the standard, the process returns to reconfigure the staged training weights and executes a new round of federated training until the model accuracy meets the standard.

[0093] (3) Knowledge reuse and push: The excellent solutions, management experience, and process standards accumulated in the entire construction process are linked with the knowledge graph to establish a "problem type-rectification plan-process standard" linkage; when a construction project has specific safety risks or quality defects, the system accurately retrieves the corresponding solutions and process standards from the knowledge graph according to the current construction project stage, type and on-site problems, and pushes them to the mobile APP of the on-site construction management personnel to realize the accurate reuse and rapid implementation of construction knowledge;

[0094] After the knowledge push is completed, on-site matching and verification are performed to determine whether the knowledge push matches the on-site needs. If they do not match, the process returns to re-associate the knowledge graph nodes and optimize the push link until the knowledge push matches the on-site needs.

[0095] Example 6: Based on the above examples, this example details the specific implementation of the cross-level collaborative management and control module, focusing on the implementation of the federated learning mechanism, the full-process policy execution mechanism, and closed-loop adaptive learning.

[0096] (1) Federated learning mechanism: After the knowledge learning evolution module completes the joint optimization of the multi-project model, it pushes the globally optimized algorithm model parameters to the edge perception collaboration module and intelligent decision governance module of each project. The local edge AI agent of the edge perception collaboration module optimizes the local early warning rules according to the updated algorithm model (improving the sensitivity of identifying "crane overload"); the intelligent decision governance module optimizes the risk early warning and quality control rules according to the updated model, realizing the full system iterative optimization of the algorithm model.

[0097] (2) Full-process policy implementation mechanism: The customized management rules for the project (on-site inspection system, quality acceptance process, collaborative approval rules) are embedded into the edge perception collaboration module, data fusion collaboration module, intelligent decision-making governance module, and knowledge learning evolution module; quantitative monitoring indicators (completion rate of safety hazard rectification, timeliness rate of collaborative approval, and data fusion efficiency) are set for each rule. For behaviors that are not implemented in accordance with the rules (failure to complete safety hazard rectification on time, overdue approval process), the system will automatically trigger an early warning and incorporate the relevant information into the performance evaluation of each participant to ensure that the management rules are uniformly implemented in each module and each construction stage.

[0098] (3) Issuance of collaborative control instructions for the entire construction process: After receiving the collaborative control instructions for the entire construction process generated by the intelligent decision-making and governance module, the cross-level collaborative control module accurately issues the instructions to the edge perception collaborative module and each execution end (construction team, equipment operation end, and supervision end) on the construction site according to the executing subject and control area, so as to realize the seamless and rapid flow of control instructions.

[0099] (4) Execution feedback acquisition unit full process: The execution feedback acquisition unit completes the acquisition of instruction execution results, deviation calculation, deviation information feedback and algorithm parameter optimization according to the technical requirements of the invention, specifically as follows:

[0100] Execution result collection: Real-time collection of execution result data of the collaborative management and control instructions for the entire construction process from each execution terminal, including quantitative data (instruction completion status, completion quality, and completion time limit);

[0101] Deviation calculation: The collected execution result data is compared with the expected control targets preset by the intelligent decision-making and governance module to calculate the deviation value. The deviation value is calculated according to the relative deviation formula. When the deviation value continuously exceeds the preset threshold (10%), the deviation information is triggered to be returned.

[0102] After the deviation calculation is completed, a threshold judgment is performed to determine whether the execution deviation is less than the preset threshold (10%). If it is greater than the preset threshold, the deviation information feedback and parameter optimization process is triggered.

[0103] Deviation information feedback: When the deviation value continues to exceed the preset threshold (10%), the deviation information (including deviation type, deviation value, and construction stage / area involved) will be synchronously fed back to the data fusion and collaboration module and the knowledge learning and evolution module.

[0104] Algorithm parameter optimization: The data fusion and collaboration module adjusts the dynamic spatiotemporal weight factor of the improved multi-source heterogeneous data fusion algorithm based on the deviation information (increasing the weight level of construction progress data and quality inspection data). The knowledge learning and evolution module optimizes the training parameters of the federated learning full-process model training algorithm based on the deviation information (adjusting the training weight of each construction stage and optimizing the sample selection rules), realizing the closed-loop adaptive learning of the system and improving the system's control accuracy and adaptability to construction scenarios.

Claims

1. A construction management system based on full-process digital collaboration, characterized in that: It includes an edge-aware collaboration module, a data fusion collaboration module, an intelligent decision-making and governance module, a knowledge learning and evolution module, and a cross-level collaborative management and control module; The edge perception collaboration module is deployed at the construction project site, collects multi-source perception data from each stage of the entire construction process and performs local preprocessing, constructs a spatiotemporal data sub-matrix for the construction stage, and simultaneously executes local edge collaboration early warning and command response operations. The data fusion and collaboration module receives spatiotemporal data sub-matrices from each construction stage, embeds an improved multi-source heterogeneous data fusion algorithm and a spatiotemporal precise matching algorithm, and constructs a unified spatiotemporal data matrix for the entire construction process. The intelligent decision-making and governance module constructs a dynamic digital twin collaborative model for the entire construction process based on a unified spatiotemporal data matrix. It embeds a domain-specific fine-tuned multimodal large language model and a construction scenario-based computer vision algorithm to generate collaborative control instructions for the entire construction process. The knowledge learning and evolution module relies on a unified spatiotemporal data matrix for the entire construction process to construct a knowledge graph of the entire construction process, thereby completing the structured storage and reuse of construction knowledge. The cross-level collaborative management and control module is equipped with a federated learning mechanism and a full-process policy execution mechanism. It receives collaborative management and control instructions for the entire construction process, distributes them to the edge perception collaborative module and various execution terminals on site, and collects instruction execution feedback data to send back to the intelligent decision-making governance module and the data fusion collaborative module.

2. The construction management system based on full-process digital collaboration according to claim 1, characterized in that: The data fusion and collaboration module is specifically used to perform the following operations: receive spatiotemporal data sub-matrices for each construction stage, configure dynamic spatiotemporal weight factors for the sensing data; after processing by an improved multi-source heterogeneous data fusion algorithm, complete the integration and spatiotemporal alignment of the entire process data, and construct a unified spatiotemporal data matrix for the entire construction process.

3. A construction management system based on full-process digital collaboration as described in claim 2, characterized in that: The intelligent decision-making and governance module specifically includes a full-process digital twin collaborative unit, an intelligent analysis and decision-making unit, a multi-participant collaborative management unit, and a closed-loop problem-solving unit. Each unit is linked at a hierarchical level and functions in synergy, sequentially completing visual modeling, intelligent analysis and decision-making, multi-party collaborative management, and closed-loop problem handling, ultimately jointly generating collaborative control instructions for the entire construction process.

4. A construction management system based on full-process digital collaboration as described in claim 3, characterized in that: The full-process digital twin collaborative unit integrates a unified spatiotemporal data matrix for the entire construction process, and constructs a full-process dynamic digital twin collaborative model covering all stages of the construction process. The models at each stage are seamlessly connected and updated in real time with the on-site data. Design changes and construction adjustments are mapped to the full-process dynamic digital twin collaborative model in real time.

5. A construction management system based on full-process digital collaboration as described in claim 4, characterized in that: The intelligent analysis and decision-making unit presets construction targets and a safety and quality defect feature library, and relies on construction scenario-based computer vision algorithms to perform target recognition and defect detection on multi-source perception data, generating analysis results.

6. A construction management system based on full-process digital collaboration as described in claim 5, characterized in that: The intelligent analysis and decision-making unit combines a domain-specific fine-tuned multimodal large language model to perform semantic analysis and root cause determination on the analysis results, generate and identify problems throughout the construction process, and generate safety risk warnings, quality control analysis, schedule coordination and scheduling, and resource optimization allocation suggestions.

7. A construction management system based on full-process digital collaboration as described in claim 6, characterized in that: The multi-participant collaborative management unit automatically parses the core information of the collaborative application document through a multimodal large language model and pushes it to the corresponding approval node. Its collaborative approval results, design changes and construction adjustment requirements are synchronized to the closed-loop problem-solving unit, and finally a collaborative conclusion is generated.

8. A construction management system based on full-process digital collaboration as described in claim 7, characterized in that: The closed-loop problem-solving unit embeds a full-process closed-loop problem-solving algorithm. The handling results of problems throughout the construction process and the optimization requirements for full-process control are fed back to the intelligent analysis and decision-making unit, forming a closed loop of control requirements.

9. A construction management system based on full-process digital collaboration as described in claim 8, characterized in that: The construction scenario-based computer vision algorithm specifically performs the following steps: Step s1: Data preprocessing. Receive multi-source sensing data collected by the edge sensing collaboration module, perform noise reduction, enhancement, and scale normalization on the multi-source sensing data, match the spatiotemporal identifier of the unified spatiotemporal data matrix for the entire construction process, supplement the processed multi-source sensing data with construction stage and regional coordinate information, and generate a standardized construction scene sensing dataset. Step s2: Construction scene feature extraction. Based on the standardized construction scene perception dataset generated in step s1, a feature extraction network is built using the improved YOLOv8 algorithm. Convolution operation is performed on the standardized construction scene perception dataset to extract the shallow texture features and deep semantic features of construction targets and safety and quality defects in the data, and generate a set of feature vectors for construction targets and defects. Step s3: Target and defect identification and classification. Receive the construction target and defect feature vector set generated in step s2, construct a construction scenario-specific classifier, input the construction target and defect feature vector set into the construction scenario-specific classifier, match the construction target and quality defect feature library, and output the preliminary identification results of construction targets and defects. Step s4: Verify the identification results. Receive the preliminary identification results of construction targets and defects generated in step s3. Perform spatial matching verification between the preliminary identification results of construction targets and defects and the full-process dynamic digital twin collaborative model. Eliminate misidentification results that do not conform to the structural characteristics of the full-process dynamic digital twin collaborative model. Retain the valid identification results of construction targets and defects that have passed the verification and transmit them synchronously to the intelligent analysis and decision-making unit.

10. A construction management system based on full-process digital collaboration according to claim 9, characterized in that: The feature extraction network built using the improved YOLOv8 algorithm is specifically implemented through the following steps: Step s21: Backbone network improvement. Receive the standardized construction scene perception dataset generated in step s1. Embed a lightweight channel-spatial attention mechanism for construction scenes into the C2f module in YOLOv8 to enhance the features of construction targets and suppress background interference. Extract the basic feature maps of construction targets and defects through preliminary convolution operations and pass them to step s22. Step s22: Improve the neck network, receive the basic feature map of construction target and defect generated in step s21, optimize the original FPN+PAN fusion structure of YOLOv8, add a 1×1 convolutional layer to adapt to multi-scale target detection in construction scenarios, complete multi-scale feature fusion enhancement, generate multi-scale fusion feature map of construction target and defect, and pass it to step s23. Step s23: Loss function improvement. Receive the multi-scale fusion feature map of construction targets and defects generated in step s22, replace the original CIoU loss function of YOLOv8 with the construction defect weighted EIoU loss function, configure differentiated weights for construction safety and quality defects, and after loss calculation and feature optimization, finally generate a set of feature vectors of construction targets and defects, which is synchronously transmitted to step s3.

Images