A construction engineering digital twin progress management and control method and system

Abstract

The application relates to the technical field of building informatization and intelligent construction, and discloses a building engineering digital twin progress control method and system, wherein the building engineering digital twin progress control method comprises the following steps: collecting multi-source heterogeneous data in a building engineering construction process; performing fusion processing on an original data set by adopting a space-time alignment network; updating a digital twin body by adopting an incremental state estimation method; predicting future progress by adopting a model combining a fusion graph neural network and a time sequence attention mechanism; identifying progress deviation and analyzing root causes by adopting a multilevel detection mechanism; generating an optimized resource scheduling scheme by adopting a reinforcement learning method; and constructing an intelligent decision support system to provide visual display and interaction functions; by combining the fusion graph neural network and the time sequence attention mechanism, the progress prediction model capable of simultaneously capturing task topological dependence and time sequence evolution characteristics is constructed, and the accuracy of progress prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of building information technology and intelligent construction technology, and more specifically, to a method and system for digital twin progress control of building engineering projects. Background Technology

[0002] Construction progress control is a core aspect of project management, directly impacting the achievement of project timeline, cost, and quality objectives. With accelerating urbanization, construction projects are expanding in scale and becoming increasingly complex, particularly with the rise of complex projects such as super high-rise buildings and large-scale complexes, placing higher demands on construction progress control.

[0003] Traditional construction project progress management relies primarily on manual inspections, paper records, and periodic reports, which suffers from problems such as delayed information collection, poor data accuracy, and coarse-grained control. Managers struggle to grasp the real-time status of the construction site, and progress deviations are often only discovered when problems have become severe, missing the optimal intervention window and leading to delays and cost overruns. In recent years, Building Information Modeling (BIM) and Internet of Things (IoT) technologies have been applied in the construction industry, providing new tools for progress management; however, existing digital progress management solutions still have many shortcomings.

[0004] Existing technologies for data fusion primarily employ simple data splicing or rule matching methods, which cannot effectively handle differences in timestamps, spatial coordinate systems, and sampling frequencies among multi-source heterogeneous data, resulting in poor fused data quality. In schedule prediction, existing methods either only consider temporal characteristics while ignoring task topology relationships, or only consider topological structure while ignoring dynamic evolution characteristics, leading to limited prediction accuracy and generalization ability. Regarding digital twin updates, existing systems generally employ a full-update strategy, resulting in high computational complexity and difficulty in achieving real-time updates for large-scale projects. In schedule deviation identification, existing methods mainly rely on task-level or process-level schedule comparisons, providing coarse-grained identification and lacking effective root cause analysis capabilities. In resource scheduling, existing systems primarily provide visualization and simple conflict detection functions, lacking intelligent optimization capabilities and struggling to cope with dynamic changes and multiple constraints.

[0005] Therefore, there is an urgent need for a construction project progress management method that can achieve deep fusion of multi-source data, high-precision progress prediction, real-time digital twin updates, refined deviation detection, and intelligent resource scheduling, in order to solve the problems of traditional methods and existing technologies and improve the level of intelligence in construction project progress management. Summary of the Invention

[0006] This invention provides a digital twin-based progress control method and system for construction projects, which solves the technical problems of insufficient data fusion depth, low prediction accuracy, low update efficiency, coarse deviation identification, and weak resource scheduling capability in related technologies.

[0007] This invention provides a digital twin-based progress control method for construction projects, comprising:

[0008] S1, based on the project's BIM model and the IoT sensing system deployed at the construction site, collects multi-source heterogeneous data during the construction process to obtain the raw data set;

[0009] S2, a spatiotemporal alignment network is used to fuse the original dataset to obtain spatiotemporally consistent and semantically unified fused data;

[0010] S3, based on fused data, uses an incremental state estimation method to update the digital twin and obtain the current state of the digital twin;

[0011] S4, based on the state of the digital twin, uses a model that integrates graph neural networks and temporal attention mechanisms to predict future progress and obtain the progress prediction results for each task.

[0012] S5, based on the status and progress prediction results of the digital twin, adopts a multi-level detection mechanism to identify progress deviations and analyze the root causes, and obtains deviation detection results and root cause analysis reports;

[0013] S6. Based on the deviation detection results and root cause analysis report, reinforcement learning method is used to generate an optimized resource scheduling scheme and obtain dynamic scheduling decision.

[0014] S7, based on dynamic scheduling decisions, builds an intelligent decision support system, providing visualization and interactive functions, and providing decision support interfaces and analysis reports.

[0015] In a preferred embodiment, S1 includes:

[0016] Component information is extracted based on the BIM model. Geometric information, attribute information and construction plan information of components are parsed from the BIM model file. The mapping relationship between component identifiers and spatial location, construction tasks and planned time is established to obtain the component information database.

[0017] Deploy an IoT sensing network to collect real-time data. Deploy RFID readers, cameras, laser scanners, environmental sensors, and equipment monitoring devices at the construction site to collect data on component location, personnel activity, equipment operation, and environmental parameters in real time, and obtain IoT sensing data streams.

[0018] We attach quality assessment indicators to the data records. Based on four dimensions—completeness, timeliness, accuracy, and consistency—we use a weighted summation method to calculate the comprehensive quality score, resulting in a dataset with quality labels.

[0019] In a preferred embodiment, S2 includes:

[0020] An adaptive time window is constructed for time alignment. The time window length is adaptively adjusted according to the data change rate to obtain the time alignment parameters, taking into account the differences in sampling frequency of different data sources.

[0021] Weighted fusion is performed within the time window. Weights are calculated based on time distance and data quality. The observations are then weighted and averaged to obtain the fused time-aligned data.

[0022] A unified engineering coordinate system is established for spatial alignment. Using the BIM model coordinate system as a reference, the spatial coordinates of each data source are transformed to the unified coordinate system using the homogeneous coordinate transformation method to obtain spatially aligned data.

[0023] Knowledge graph technology is used to organize and integrate data, and the integrated data is used to construct a knowledge graph, where nodes represent entities and edges represent relationships, thus obtaining a construction progress knowledge graph.

[0024] In a preferred embodiment, S3 includes:

[0025] Identify the set of components that have changed, compare the differences between the new data and the state of the digital twin at the previous moment, identify the components that have changed, and obtain the set of changed components;

[0026] The state of the changing component is updated by fusing the observed data and the predicted state using the Kalman filter method, and the state vector of the changing component is updated to obtain the updated component state.

[0027] For unchanged components, a simplified update is performed, either by directly inheriting the state from the previous moment or by making a linear prediction based on the construction plan, to obtain the state of the unchanged components.

[0028] The update priority of components is dynamically adjusted. The update priority is calculated based on the importance and change frequency of the components, and the priority ranking is obtained.

[0029] In a preferred embodiment, S4 includes:

[0030] Construct a construction task dependency graph by modeling the construction task network as a directed graph, where nodes represent construction tasks and directed edges represent the dependencies between tasks, thus obtaining the task dependency graph.

[0031] A graph convolutional network is used to learn the task node embedding. The neighbor node information is aggregated through multi-layer graph convolution operations to obtain the task embedding vector.

[0032] A temporal attention mechanism is used to capture progress evolution features, extract historical state sequences, calculate the importance weight of different historical moments to the current prediction, and obtain a weighted historical state representation.

[0033] The system integrates spatial and temporal features for progress prediction. It concatenates task embedding vectors and historical state representations, inputs them into the prediction layer, and outputs future progress prediction values ​​to obtain the progress prediction results.

[0034] In a preferred embodiment, S5 includes:

[0035] Rule-based rapid deviation screening compares actual progress with planned progress. When the deviation exceeds a preset threshold, it is marked as a potential deviation task, and a list of potential deviation tasks is obtained.

[0036] Anomaly detection based on statistical learning is used to analyze the time series characteristics of progress deviations, and the local anomaly factor algorithm is used to detect anomaly patterns, identify the true deviations, and obtain a list of anomalous tasks.

[0037] A causal graph model is constructed to conduct root cause analysis. Based on domain knowledge and historical data, a causal graph of the factors affecting construction progress is constructed to obtain the causal graph model.

[0038] The causal effects of each factor on schedule deviations are calculated. Based on the cause-effect diagram and observation data, the causal effects of each potential cause on schedule deviations are calculated, and the causal effects are ranked.

[0039] In a preferred embodiment, S6 includes:

[0040] Define the state space of the resource scheduling problem. The state includes the progress status of each task, the availability and allocation of resources, and the environmental state, resulting in a state vector.

[0041] Define a multi-objective reward function that comprehensively considers multiple objectives, including schedule objectives, cost constraints, quality requirements, and safety standards. Use a weighted method to calculate schedule progress rewards, cost penalties, delay penalties, and risk penalties to obtain the total reward value.

[0042] A digital twin simulation environment is constructed, and a reinforcement learning training environment is built based on the digital twin to simulate the execution effect of resource scheduling decisions, thereby obtaining a simulation environment model.

[0043] A deep reinforcement learning scheduling policy is trained, and a policy optimization algorithm is used to train the scheduling policy network. The agent interacts with the simulation environment to learn the optimal scheduling policy, and a well-trained policy model is obtained.

[0044] In a preferred embodiment, S7 includes:

[0045] A 3D visualization digital twin interface is constructed, and a 3D graphics engine is used to render the digital twin to display the 3D model and construction status of the building. The components are displayed differently according to the construction status to obtain the visualization interface.

[0046] Implement the progress timeline playback function, allowing users to view the construction status at any point in history or automatically play the progress evolution process to obtain the timeline playback function;

[0047] To achieve intelligent question answering based on natural language, users can input questions via text and voice, and the system will understand the intent of the question and return the answer, thus providing intelligent question answering functionality;

[0048] Deploy a cloud-edge-device collaborative architecture, where the edge device is responsible for data collection and real-time response, the cloud device is responsible for complex calculations and model training, and the device provides interactive interfaces for different scenarios, thus obtaining the collaborative architecture.

[0049] In a preferred embodiment, the Kalman filtering method includes:

[0050] Based on the state of the previous moment and the construction plan, the current state is predicted, and the predicted state vector is calculated using the state transition matrix and the control input matrix.

[0051] The Kalman gain is calculated based on the prediction error covariance matrix, the observation matrix, and the observation noise covariance matrix.

[0052] The state is updated by fusing observation data, and the state correction is calculated based on the observation residuals and the Kalman gain matrix to obtain the updated state vector.

[0053] In a preferred embodiment, a digital twin schedule control system for construction projects is used to execute the aforementioned digital twin schedule control method for construction projects, including:

[0054] The multi-source data acquisition module, based on the project's BIM model and the IoT sensing system deployed at the construction site, collects multi-source heterogeneous data during the construction process to obtain the raw data set;

[0055] The spatiotemporal alignment fusion module uses a spatiotemporal alignment network to fuse the original dataset, resulting in spatiotemporally consistent and semantically unified fused data.

[0056] The digital twin update module updates the digital twin based on fused data using an incremental state estimation method to obtain the current state of the digital twin.

[0057] The progress prediction module, based on the state of the digital twin, uses a model that integrates graph neural networks and temporal attention mechanisms to predict future progress and obtain the progress prediction results for each task.

[0058] The deviation detection and root cause analysis module, based on the digital twin status and schedule prediction results, uses a multi-level detection mechanism to identify schedule deviations and analyze their root causes, and obtains deviation detection results and root cause analysis reports.

[0059] The resource scheduling optimization module, based on deviation detection results and root cause analysis reports, uses reinforcement learning to generate optimized resource scheduling schemes and obtain dynamic scheduling decisions.

[0060] The intelligent decision support module, based on dynamic scheduling decisions, constructs an intelligent decision support system, providing visualization and interactive functions, and delivering a decision support interface and analysis reports.

[0061] The beneficial effects of this invention are as follows: By constructing a spatiotemporal aligned network and employing an adaptive time window mechanism and spatial semantic mapping strategy, deep fusion of multi-source heterogeneous data is achieved, effectively solving the inconsistency problem of different data sources in the time, spatial, and semantic dimensions; an uncertainty quantification mechanism is introduced to probabilistically represent low-quality data, propagating and updating uncertainty information in subsequent processing, thereby improving the robustness of data fusion; knowledge graph technology is used to organize fused data, supporting efficient graph query and graph reasoning operations, providing a high-quality data foundation for subsequent intelligent analysis; compared with existing simple data splicing methods, the data fusion scheme of this invention can more accurately reflect the real state of the construction site, providing reliable data support for progress control.

[0062] This invention constructs a progress prediction model that simultaneously captures task topological dependencies and time-series evolution features by fusing graph neural networks and temporal attention mechanisms, thereby improving the accuracy of progress prediction. The graph convolutional network, through multi-layer information aggregation, integrates information from multi-hop neighbors into the representation of each task, accurately modeling the complex dependencies between tasks. The temporal attention mechanism automatically learns the importance of different historical moments to the current prediction, effectively capturing the evolutionary patterns of progress. The Monte Carlo Dropout method is used for uncertainty estimation, providing confidence assessment for the prediction results and helping managers identify high-risk tasks. Practical applications show that the progress prediction accuracy of this invention is improved compared to traditional methods, providing a scientific basis for progress control decisions.

[0063] This invention proposes a lightweight update strategy based on incremental state estimation. It performs complex calculations only on components that have changed, and uses simplified updates for unchanged components, thereby reducing computational complexity. This enables large-scale digital twins to meet the real-time requirement of updating once per minute, ensuring a high degree of synchronization between the digital twin and the physical site, and providing a foundation for real-time decision-making. Attached Figure Description

[0064] Figure 1This is a flowchart of a digital twin progress control method for construction projects according to the present invention;

[0065] Figure 2 This is a module diagram of a digital twin progress control system for construction projects according to the present invention. Detailed Implementation

[0066] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0067] At least one embodiment of the present invention discloses a digital twin progress control method for construction projects, such as... Figure 1 As shown, it includes:

[0068] S1, based on the project's BIM model and the IoT sensing system deployed at the construction site, collects multi-source heterogeneous data during the construction process to obtain the raw data set;

[0069] S11. Extract component information based on the BIM model. The geometric, attribute, and construction plan information of components are parsed from the project's BIM model files. A mapping relationship is established between the component's unique identifier and its 3D spatial location, construction task, and planned time, resulting in a component information database. Specifically, for each component, its globally unique identifier (GUID), geometric center coordinates, bounding box dimensions, floor, discipline, associated task number, planned start time, planned completion time, and list of preceding tasks are extracted and stored as a structured data table.

[0070] S12, Deploy an IoT sensing network to collect real-time data. 128 RFID readers, 256 high-definition cameras, 16 laser scanners, 64 environmental sensors, 8 sets of tower crane status monitoring equipment, and 12 sets of construction elevator operation monitoring equipment are deployed at key locations on the construction site. This collects data in real-time on component locations, personnel activities, equipment operation, environmental parameters, etc., resulting in an IoT sensing data stream. The RFID system tracks the location and installation status of prefabricated components. RFID tags are affixed to each component, and the time and location are automatically recorded when a component passes through the reader's coverage area. The video surveillance system captures real-time images of the construction site, sampling at 25 frames per second. The laser scanner periodically performs 3D scanning of completed areas to obtain actual geometric data. Environmental sensors monitor meteorological parameters such as temperature, humidity, wind speed, and rainfall, sampling once per minute. The equipment monitoring system records the operating status, load, location, and fault information of tower cranes and construction elevators, sampling once per second.

[0071] S13. Obtain business data from the construction management system. Through an interface, obtain structured and semi-structured data from the project's construction management system, including construction logs, resource scheduling records, quality inspection reports, safety inspection records, material arrival records, and design change notices, to obtain a construction management dataset. Construction logs record daily work content, completed work volume, and encountered problems for each work team, stored in text format. Resource scheduling records include personnel allocation, equipment scheduling, and material delivery information, stored in tabular format. Quality inspection reports record inspection results, pass rates, and rectification requirements for each inspection batch. Safety inspection records include on-site safety inspection information, discovered safety hazards, rectification measures, and responsible persons, stored in structured tabular format. Material arrival records include material name, specifications, quantity, supplier, arrival time, and quality certification documents, stored in tabular and attachment formats. Design change notices include design modification content, reasons for changes, scope of impact, and approval process information, stored in document and drawing formats.

[0072] S14. Obtain external meteorological and environmental data. By calling the meteorological service API interface, obtain historical meteorological data and 7-day weather forecast data for the project location, including parameters such as temperature, humidity, wind speed, rainfall, and visibility, with a time resolution of hourly, to obtain a meteorological and environmental dataset.

[0073] S15. Preprocessing of the collected raw data. Statistical anomaly detection based on the 3σ criterion and box plot method is used to identify outliers in the numerical data. Data points deviating from the mean by more than three standard deviations are marked as anomalies, resulting in an anomaly data label set. For video data, motion detection based on optical flow and object detection based on deep learning are used to filter invalid and blurry frames. For text data, natural language processing techniques are used for word segmentation, entity recognition, and key information extraction.

[0074] S16, Addressing Missing Data. Based on the missing data pattern and rate, different imputation strategies are employed to obtain a complete dataset. For time-series data with completely random missing data and a missing rate below 5%, linear interpolation is used for imputation, calculating a weighted average of observations before and after the missing time. For data with a missing rate between 5% and 20%, K-nearest neighbor imputation based on similar samples is used, finding the K most similar samples to the missing sample in other features, and taking the mean of their corresponding feature values ​​as the imputation value. For data with a missing rate exceeding 20%, it is marked as low-quality data, and its uncertainty is recorded, reducing its weight in subsequent fusion. For discrete data such as construction status, mode imputation or prediction imputation based on decision trees is used.

[0075] S17 adds quality assessment metrics to each data record. Based on four dimensions—completeness, timeliness, accuracy, and consistency—a comprehensive quality score is calculated, resulting in a quality-labeled dataset. Completeness is calculated based on the completeness of data fields; timeliness is calculated based on the interval between data collection time and the current time; accuracy is calculated based on the historical reliability and calibration status of sensors; and consistency is calculated based on the cross-validation results of the data with other data sources. The comprehensive quality score is calculated using a weighted summation method, as detailed below:

[0076] The data records are assessed for completeness by checking whether all required fields are fully filled and calculating the ratio of filled fields to the total number of fields to obtain a completeness score, ranging from 0 to 1. A higher completeness score results in a higher score. The data records are also assessed for timeliness by obtaining the data collection timestamp and the current timestamp, calculating the time interval between them, and determining the timeliness score based on the length of the time interval. A shorter time interval indicates fresher data and a higher timeliness score, ranging from 0 to 1. The data records are then assessed for accuracy by querying the historical reliability records and the most recent calibration time of the sensor from which the data was collected. An accuracy score is calculated based on the sensor's historical failure rate and calibration timeliness; a more reliable sensor and more timely calibration result in a higher accuracy score, ranging from 0 to 1. Finally, the data records are assessed for consistency by cross-validating the data with identical or related data collected from other data sources and calculating the degree of consistency between the data values. Higher consistency indicates more reliable data and a higher consistency score, ranging from 0 to 1.

[0077] After obtaining the scores for each of the four dimensions, a weighted sum is performed to calculate the overall quality score: The completeness score is multiplied by a weighting factor of 0.3 to obtain the completeness weighted score; the timeliness score is multiplied by a weighting factor of 0.25 to obtain the timeliness weighted score; the accuracy score is multiplied by a weighting factor of 0.25 to obtain the accuracy weighted score; and the consistency score is multiplied by a weighting factor of 0.2 to obtain the consistency weighted score. The four weighted scores are then summed: the completeness weighted score, the timeliness weighted score, the accuracy weighted score, and the consistency weighted score, to obtain the overall quality score. The overall quality score ranges from 0 to 1, with a higher score indicating better data quality. The weighting factors reflect the importance of each dimension to data quality. Completeness has the highest weight of 0.3, indicating that data completeness is the primary factor in quality assessment; timeliness and accuracy both have weights of 0.25, indicating that data freshness and reliability are equally important; and consistency has a weight of 0.2, indicating that cross-validation results are also an important quality indicator. This weighted summation method enables a comprehensive quantitative assessment of data quality.

[0078] Through the above sub-steps, a cleaned multi-source heterogeneous dataset is obtained, including a component information database, an IoT sensing data stream, a construction management dataset, a meteorological environment dataset, and quality assessment indicators for each data point, which serve as input for the next step.

[0079] S2, a spatiotemporal alignment network is used to fuse the original dataset to obtain spatiotemporally consistent and semantically unified fused data;

[0080] S21. Construct an adaptive time window for time alignment. Considering the sampling frequency differences of different data sources, an adaptive time window is designed for each data source to obtain time alignment parameters. For high-frequency data such as equipment monitoring data, the sampling frequency is once per second, and the time window length is set to 60 seconds, with downsampling performed within the window. For low-frequency data such as construction logs, the sampling frequency is once per day, and the time window length is set to 3 days, with interpolation upsampling performed within the window. The length of the time window is adaptively adjusted according to the rate of data change; a shorter window is used for rapidly changing data, and a longer window is used for slowly changing data. Specifically, all data values ​​of data source i within the historical time period are extracted to form a time series set; the variance of this time series is calculated, as the variance reflects the degree of data fluctuation; a larger variance indicates faster data change; the time window length is calculated based on the variance value, using an adjustment coefficient divided by the square root of the variance to make the window length inversely proportional to the variance. The adjustment coefficient is set to 300 seconds based on the actual situation; thus, the adaptive time window length for data source i is obtained. In this way, fast-changing data sources are given shorter time windows, while slow-changing data sources are given longer time windows, thus achieving adaptive adjustment of the time window.

[0081] S22, weighted fusion is performed within the time window. For the target time t, all observations from each data source within their time window are collected. Weights are calculated based on time distance and data quality, and a weighted average is performed to obtain the fused time-aligned data value. The weight calculation process is as follows: Obtain the observations and quality scores of data source i at historical times; calculate the time difference between the historical time and the target time, take its absolute value, and square it; divide the squared time difference by twice the square of the time scale parameter, which is set to one-third of the time window length; then take the negative value of the above calculation result and use it as the exponent of the natural exponential function to calculate the decay factor based on time distance; multiply this decay factor by the quality score of the observation to obtain the weight of the historical time observation to the target time. This weight calculation method gives greater weight to observations that are closer in time and have higher quality scores.

[0082] The calculation process for the fused time-aligned data value is as follows: Determine the time window range of data source i at the target time t, and collect observations from all historical moments within the window; for each historical moment within the window, multiply its weight by the corresponding original observation to obtain a weighted observation value; then sum all weighted observation values ​​to obtain a total weighted observation value; simultaneously, sum the weights of all historical moments within the window to obtain a total weight; divide the total weighted observation value by the total weight to obtain the fused time-aligned data value. This weighted averaging method effectively fuses multiple observations within the time window.

[0083] S23. Establish a unified engineering coordinate system for spatial alignment. Using the project's BIM model coordinate system as a reference, establish a unified engineering coordinate system, transforming the spatial coordinates of various data sources to this unified coordinate system to obtain spatially aligned data. For component locations detected by RFID readers, a triangulation algorithm is used to calculate the component's 3D coordinates, which are then transformed to the engineering coordinate system using a coordinate transformation matrix. For target locations captured by video surveillance, camera calibration and perspective transformation are used to convert image coordinates to world coordinates. For point cloud data acquired by laser scanning, the ICP algorithm is used to register the point cloud to the BIM model coordinate system. The coordinate transformation process employs a homogeneous coordinate transformation method, with the following specific steps: The three components of the original 3D coordinates (x, y, z) are combined with a constant 1 to form a 4D homogeneous coordinate vector; a 4×4 homogeneous coordinate transformation matrix is ​​prepared, containing rotation and translation parameters. Several control points are selected in both the engineering coordinate system and the data source coordinate system, and the elements of the transformation matrix are calculated using a control point matching algorithm; the transformation matrix is ​​multiplied by the original homogeneous coordinate vector. Following the matrix multiplication rules, each row of the transformation matrix is ​​multiplied by the original coordinate vector to obtain the corresponding components of the transformed homogeneous coordinate vector; the first three components are extracted from the transformed homogeneous coordinate vector and used as the transformed 3D coordinates x′, y′, and z′, respectively. This homogeneous coordinate transformation method achieves the conversion of spatial coordinates from different data sources to a unified engineering coordinate system.

[0084] S24. Spatial semantic mapping is used to associate physical objects with BIM components. For physical objects detected by sensors, the spatial distance and semantic similarity between them and each component in the BIM model are calculated. The component with the highest comprehensive score is selected for association, resulting in an object-component mapping relationship. Spatial distance is calculated using Euclidean distance, and semantic similarity is calculated based on the matching degree between object type and component type. The comprehensive score calculation process is as follows: Calculate the spatial distance score by extracting the spatial position vectors of the physical object and component j, and calculating the Euclidean distance between the two position vectors, which is the square root of the sum of the squares of the differences in the corresponding coordinate components; divide the square of the Euclidean distance by twice the square of the spatial scale parameter, which is set to 5 meters; take the negative value of the above calculation result and use it as the exponent of the natural exponential function to calculate the similarity score based on spatial distance, with higher scores for closer distances; then multiply the spatial distance score by a weighting coefficient of 0.6 to obtain the weighted score based on spatial distance. The semantic similarity score is calculated by comparing the physical object type and component j type using predefined type matching rules. A perfect match results in a similarity score of 1, a partial match results in a similarity score of 0.5, and a mismatch results in a similarity score of 0. The semantic similarity score is then multiplied by a weighting coefficient of 0.4 to obtain a weighted semantic score. The weighted score based on spatial distance is then added to the weighted semantic score to obtain the overall score for component j. This calculation process is repeated for all components in the BIM model, and the component with the highest overall score is selected to be associated with the physical object, establishing an object-component mapping relationship.

[0085] S25. Construct an ontology for the construction engineering domain to achieve semantic fusion. Define the core concepts and relationships within the construction engineering domain, including entity types such as components, tasks, resources, states, and events, as well as relationship types such as containment, dependency, allocation, and triggering, to obtain the domain ontology model. Map data from various data sources to ontology concepts to achieve semantic unification. For example, map "beam" in the BIM model, "precast beam component" in IoT data, and "beam installation" in construction logs to the ontology concept of "beam component"; map "not started," "pending construction," and "planned" expressing construction status from different data sources to the "not under construction" status.

[0086] S26 employs knowledge graph technology to organize and integrate data. Using an ontology model as a framework, the integrated data is constructed into a knowledge graph. Nodes represent entities, edges represent relationships, and the attributes of nodes and edges represent feature information, resulting in a construction progress knowledge graph. The knowledge graph is stored in triples, with each triple containing a subject, predicate, and object. For example, the triple "Component C001, installed on, floor 20" indicates that component C001 is installed on the 20th floor; the triple "Task T056, depends on, task T055" indicates that task T056 depends on the completion of task T055. The knowledge graph supports efficient graph query and graph reasoning operations.

[0087] S27 introduces an uncertainty quantification mechanism to handle low-quality data. For data with a quality score below 0.6, instead of using the numerical value directly, its uncertainty is represented by a probability distribution, resulting in fused data with uncertainty annotations. A Bayesian framework is used to represent the data as a normal distribution, with the mean being the observed or imputed value, and the variance calculated based on the quality score. The variance calculation process is as follows: A baseline variance is determined based on the data type, reflecting the inherent uncertainty of that type of data; the quality score is obtained, ranging from 0 to 1, with higher scores indicating better quality; 1 is subtracted from the quality score to obtain the degree of quality defect; the degree of quality defect is then squared to amplify the impact of quality differences; the square of the baseline variance is multiplied by the square of the degree of quality defect to obtain the variance of the data. In this way, data with lower quality scores have larger variances, indicating higher uncertainty; data with higher quality scores have smaller variances, indicating lower uncertainty. After representing the data as a normal distribution with a defined mean and variance, uncertainty can be propagated and quantified in subsequent processing.

[0088] Through the above sub-steps, a spatiotemporally aligned and semantically unified fusion dataset is obtained, which is organized in the form of a knowledge graph and accompanied by uncertainty annotations, serving as the input for the next step.

[0089] S3, based on fused data, uses an incremental state estimation method to update the digital twin and obtain the current state of the digital twin;

[0090] S31. Initialize the digital twin model. Based on the BIM model, construct the geometric model of the digital twin, representing the three-dimensional geometry and spatial location of 160,000 components; construct a physical model based on component attributes, representing the physical properties of the components such as material, weight, and strength; construct a behavioral model based on construction technology, representing the dynamic process and resource consumption patterns of construction activities; construct a rule model based on construction specifications, representing constraints such as safety constraints and quality standards, thus obtaining the initial digital twin. The state representation of the digital twin is a sparse vector set, with each component corresponding to a state vector containing information such as location, installation progress, quality status, and associated resources.

[0091] S32, Identify the set of components that have changed. When new fused data is received, compare the differences between the new data and the state of the digital twin at the previous time step to identify the components that have changed, thus obtaining the set of changed components. A change detection algorithm is used to calculate the degree of difference between the observed data and the predicted state for each component. The specific calculation process is as follows: Obtain the observed data vector of component i at the current time t and its state vector at the previous time step. The observed data vector contains the current information of the component extracted from the fused data, and the state vector contains the historical information of the component recorded in the digital twin. Calculate the difference between the corresponding components of the two vectors, subtracting each component separately. Square all component differences and sum them. Then take the square root of the sum to obtain the Euclidean distance, which represents the amount of change in component i. The larger the amount of change, the more significant the change in the component's state.

[0092] The process for determining whether a component has changed is as follows: A change detection threshold is set for each component based on its type and measurement accuracy. For large components such as beams and columns, a larger threshold is set due to greater measurement error; for small components such as doors, windows, and accessories, a smaller threshold is set due to higher measurement accuracy. The change in component i is compared with its corresponding change detection threshold. If the change exceeds the threshold, component i is determined to have changed at time t and added to the set of changed components. If the change does not exceed the threshold, component i is determined not to have changed and is not added to the set of changed components. In this way, all components that have changed at the current time are identified, forming a set of changed components.

[0093] S33, Update the state of the changed component. The Kalman filter method is used to fuse observed data and predicted states to update the state vector of the changed component, obtaining the updated component state. The state update process includes the following steps:

[0094] Predict the current state based on the previous state and construction plan. Obtain the state vector of component i at the previous time step, which contains information such as the component's position and progress; prepare a state transition matrix, which describes the natural evolution of the component's state over time; perform matrix multiplication between the state transition matrix and the previous state vector to obtain the natural evolution part of the state; simultaneously obtain the control input vector at the current time step, which contains construction plan and resource allocation information; prepare a control input matrix, which describes the influence of the control input on the state; perform matrix multiplication between the control input matrix and the control input vector to obtain the control influence part; add the natural evolution part of the state to the control influence part to obtain the predicted state vector of component i at time t.

[0095] Calculate the prediction error covariance. Obtain the error covariance matrix of component i at the previous time step, which describes the uncertainty of the state estimation; perform matrix multiplication on the state transition matrix and the error covariance matrix; then perform matrix multiplication on the result and the transpose of the state transition matrix; add the process noise covariance matrix, which describes the uncertainty in the state evolution process; obtain the prediction error covariance matrix of component i at time t.

[0096] Calculate the Kalman gain. Prepare the observation matrix, which describes the mapping relationship between the state vector and the observed data. Perform matrix multiplication of the prediction error covariance matrix and the transpose of the observation matrix to obtain intermediate result A. Perform matrix multiplication of the observation matrix and the prediction error covariance matrix, and then perform matrix multiplication with the transpose of the observation matrix to obtain intermediate result B. Add intermediate result B to the observation noise covariance matrix, which is set according to the data quality score; the lower the quality, the greater the noise. Perform matrix inversion on the sum to obtain intermediate result C. Perform matrix multiplication of intermediate result A and intermediate result C to obtain the Kalman gain matrix. The Kalman gain determines the weight of the observed data and the predicted state in the fusion process.

[0097] The state is updated by fusing observation data. The observation data vector of component i at time t is obtained; the observation matrix and the predicted state vector are multiplied to obtain the predicted observation value; the difference between the actual observation value and the predicted observation value is calculated to obtain the observation residual; the Kalman gain matrix and the observation residual are multiplied to obtain the state correction; the predicted state vector and the state correction are added to obtain the updated state vector of component i at time t.

[0098] Update the error covariance. Prepare the identity matrix; perform matrix multiplication between the Kalman gain matrix and the observation matrix; subtract the result of the multiplication from the identity matrix to obtain the intermediate matrix; perform matrix multiplication between the intermediate matrix and the prediction error covariance matrix to obtain the updated error covariance matrix of component i at time t. The updated error covariance matrix reflects the reduction in uncertainty of state estimation after fusing observation data.

[0099] S34. Simplified update of unchanged components. For components not in the set of changed components, the state of the previous time step is directly inherited or a simple linear prediction is made based on the construction plan, without complex filtering calculations, to obtain the state of the unchanged component. The simplified update process is as follows: Obtain the state vector of component i at the previous time step; determine the state change rate vector of component i according to the construction plan, which describes the rate of change of the component state according to the plan; calculate the time interval between the current time step and the previous time step; perform a scalar-vector multiplication operation on the time interval and the state change rate vector to obtain the state change amount; add the state vector of the previous time step to the state change amount to obtain the state vector of component i at the current time step. This simplified update method avoids complex Kalman filtering calculations, reduces the amount of computation, and is suitable for components whose state has not changed.

[0100] S35, dynamically adjust component update priorities. Based on the importance and change frequency of components, the update priority is calculated, allocating more computing resources and a higher update frequency to high-priority components, resulting in a priority ranking. Component importance is comprehensively assessed based on factors such as whether it is on the critical path, in a high-risk area, and is a critical structural component; change frequency is calculated statistically based on historical change records. The priority calculation process is as follows: First, determine if component i is on the critical path. If it is, the critical path indicator value is 1; otherwise, it is 0. Multiply the critical path indicator value by the critical path weight coefficient of 0.5 to obtain the critical path score. Second, determine if component i is in a high-risk area. If it is, the risk area indicator value is 1; otherwise, it is 0. Multiply the risk area indicator value by the risk area weight coefficient of 0.3 to obtain the risk area score. Third, calculate the change frequency of component i from historical change records, i.e., the number of changes per unit time, and normalize the change frequency to the range of 0 to 1. Multiply the normalized change frequency by the change frequency weight coefficient of 0.2 to obtain the change frequency score. Finally, add the critical path score, risk area score, and change frequency score to obtain the priority score for component i. After calculating the priority scores for all components, sort them from highest to lowest score to obtain the priority ranking. With limited resources, prioritize updating high-priority components to ensure that critical and high-risk components receive timely and accurate status updates.

[0101] S36, Calculate the reduction in computational complexity of the digital twin update. Statistically analyze the ratio of changed components to the total number of components, evaluate the efficiency improvement of the incremental update strategy, and obtain a computational complexity analysis report. In this embodiment, each update involves an average of approximately 2800 changed components, accounting for 1.75% of the total 160,000 components. Compared to a full update, the computational load is reduced by approximately 98%, enabling the digital twin to achieve the real-time requirement of updating once per minute.

[0102] S37 records the amount and type of change of the changing components. For each changing component, it records the specific details of its state change, including changes in location, schedule, and quality status, and classifies them into types such as normal change, abnormal change, and abrupt change, thus obtaining a change record database. The determination of the change type is based on the magnitude and rate of change. Normal change refers to changes that conform to the construction plan and historical patterns; abnormal change refers to changes that deviate from expectations but are within a reasonable range; and abrupt change refers to drastic changes that exceed the reasonable range.

[0103] Through the above sub-steps, we obtain the updated digital twin state, the set of changed components and their changes, and a computational complexity analysis report, which serve as input for the next step.

[0104] In some embodiments, a deep learning-based state prediction model is used to replace the traditional linear prediction method to improve prediction accuracy. Specifically, an LSTM (Long Short-Term Memory) model is trained, taking the historical state sequence of components and construction plan information as input, and outputting the predicted value of the future state. The model structure includes an input layer, two LSTM hidden layers, and a fully connected output layer. The input layer receives state vectors from the past 10 time steps, each state vector having a dimension of 32, containing information such as location, progress, and resources. The first LSTM layer contains 64 hidden units, the second LSTM layer contains 32 hidden units, and the output layer is a fully connected layer with a dimension of 32, representing the predicted state vector. The model uses the mean squared error loss function, is trained using the Adam optimizer, with a learning rate of 0.001, a batch size of 128, and 100 training epochs. The training data comes from the construction records of historical projects, containing approximately 500,000 state sequence samples. After training, the model's prediction error on the test set is reduced by approximately 35% compared to the linear prediction method, improving the accuracy of state prediction and thus enhancing the quality of digital twin updates. The purpose of this alternative embodiment is to reduce prediction errors and improve the consistency between the digital twin and the physical site by making more accurate state predictions when computing resources are sufficient.

[0105] S4, based on the state of the digital twin, uses a model that integrates graph neural networks and temporal attention mechanisms to predict future progress and obtain the progress prediction results for each task.

[0106] S41, Construct the construction task dependency graph. The project's construction task network is modeled as a directed graph, where nodes represent construction tasks and directed edges represent dependencies between tasks, resulting in a task dependency graph. In this embodiment, the project contains 1850 construction tasks, with 3420 dependencies between them. Task dependencies are extracted from the construction plan, including predecessor-successor relationships, overlap relationships, etc. The adjacency matrix construction process is as follows: Create an 1850-row, 1850-column matrix, initially with all elements set to 0; traverse all task dependencies. For each dependency, if task j depends on task i (i.e., task i must be completed or start before task j), set the element in the i-th row and j-th column of the adjacency matrix to 1; if task j does not depend on task i, the corresponding element remains 0; the complete adjacency matrix is ​​obtained, where elements with a value of 1 indicate a dependency, and elements with a value of 0 indicate no dependency. The adjacency matrix clearly expresses the topological dependency structure between tasks.

[0107] S42, Construct the task node feature vector. Extract multi-dimensional features for each task node, including planned progress, actual progress, progress deviation, resource allocation, historical completion rate, environmental factors, etc., to obtain the task feature matrix. Specifically, for task i at time t, the feature vector contains the following components: planned progress percentage, actual progress percentage, progress deviation value, number of allocated personnel, number of allocated equipment, material arrival rate, average completion rate over the past 7 days, current weather condition code, task duration, remaining task time, etc., for a total of 32 features. Feature values ​​are normalized; numerical features use min-max normalization, and categorical features use one-hot encoding.

[0108] S43 employs a graph convolutional network to learn task node embeddings. The task dependency graph and node features are input into the graph convolutional network, which aggregates neighbor node information through multiple layers of graph convolutional operations to learn the embedding representation of each node, thus obtaining the task embedding vector. The computation process of the graph convolutional layer is as follows: The original adjacency matrix is ​​preprocessed by adding 1 to the diagonal, which is equivalent to adding a self-loop edge pointing to each node, resulting in an adjacency matrix with added self-loops. The degree matrix of the adjacency matrix with added self-loops is calculated. The degree matrix is ​​a diagonal matrix, and the i-th element on the diagonal equals the degree of the i-th node, i.e., the sum of the elements in the i-th row of the adjacency matrix. The negative 1 / 2 power of the degree matrix is ​​calculated, which is the square root of the reciprocal of each diagonal element of the degree matrix. The negative 1 / 2 power of the degree matrix is ​​multiplied by the adjacency matrix with added self-loops, and then multiplied again by the negative 1 / 2 power of the degree matrix to obtain a normalized adjacency matrix. The normalized adjacency matrix is ​​multiplied by the node feature matrix of the current layer to aggregate the features of neighboring nodes. The aggregated result is then multiplied by the weight matrix of the current layer to perform feature transformation. The ReLU activation function is applied to the transformation result, setting negative values ​​to zero and retaining positive values ​​to obtain the node feature matrix of the next layer. The network consists of three graph convolutional layers. The first and second layers have a hidden dimension of 64, while the third layer has a hidden dimension of 32. Through multi-layer graph convolutional operations, the features of each node not only contain its own information but also incorporate information from its multi-hop neighbors, forming a rich task embedding representation.

[0109] S44 employs a temporal attention mechanism to capture progress evolution features. For each task, its state sequence over the past T time points is extracted and input into the temporal attention layer. The importance weights of different historical time points to the current prediction are calculated to obtain a weighted historical state representation. In this embodiment, T is set to 14, meaning that historical data from the past 14 days is used. Attention weights are calculated using a scaled dot product attention mechanism, with the following steps: The query vector is obtained from the current state of task i through a linear transformation, which multiplies the current state vector by the query weight matrix. For each historical moment, the key vector is obtained from the state of task i at that historical moment through a linear transformation, which multiplies the historical state vector by the key weight matrix. Next, the dot product of the query vector and the key vector at each historical moment is calculated, i.e., the corresponding elements are multiplied and summed. The dot product result is then divided by the square root of the key vector dimension to scale and stabilize the gradient. Softmax normalization is applied to the scaled dot product results for all historical moments. A natural exponent is calculated for each scaled dot product result, and all natural exponent values ​​are summed. Each natural exponent value is then divided by the sum to obtain the normalized attention weight. The sum of the attention weights is 1; a larger weight indicates that the corresponding historical moment is more important to the current prediction.

[0110] The weighted historical state representation is calculated as follows: For each historical moment, the state of task i at that moment is transformed linearly to obtain a value vector, which involves multiplying the historical state vector by the value weight matrix. The attention weight for each historical moment is then multiplied by the corresponding value vector using a scalar-vector multiplication to obtain a weighted value vector. Finally, the weighted value vectors from all historical moments are summed to obtain the weighted historical state representation vector of task i at the current moment. This vector integrates historical information from the past 14 days and automatically highlights historical moments that are more important for the current prediction.

[0111] S45, merging spatial and temporal features for progress prediction. The task embedding vector obtained from graph convolution and the historical state representation obtained from temporal attention are concatenated and input into a fully connected prediction layer. The output is the progress prediction value for multiple future time points, yielding the progress prediction result. The prediction layer contains two fully connected network layers: the first layer has a dimension of 64, and the second layer has a dimension equal to the prediction time step multiplied by the progress dimension. In this embodiment, the progress is predicted for the next 7 days, with each day's progress represented by a scalar as the percentage of completion; therefore, the output dimension is 7. The prediction process is as follows: First, obtain the output embedding vector of task i at the current time after the third layer of graph convolution. This vector contains the spatial topological features of the task. Simultaneously, obtain the weighted historical state representation vector of task i at the current time, which contains the temporal evolution features of the task. Then, concatenate the two vectors, arranging their elements sequentially to form a longer vector that integrates both spatial and temporal information. Next, perform matrix multiplication with the weight matrix of the first fully connected layer, and add the bias vector of the first layer to complete the linear transformation of the first layer. Apply the ReLU activation function to the linear transformation result, introducing non-linearity, setting negative values ​​to zero, and retaining positive values. Then, perform matrix multiplication with the weight matrix of the second fully connected layer, and add the bias vector of the second layer to complete the linear transformation of the second layer. This yields the predicted progress vector of task i for the next 7 days, containing 7 elements representing the percentage of completion predicted for days 1 to 7. In this way, the model utilizes both the dependencies between tasks and the historical evolution patterns of the tasks themselves, achieving accurate progress prediction.

[0112] S46, Training the Progress Prediction Model. The model parameters are trained using historical project data. The loss function is the mean squared error between the predicted and actual progress. The Adam optimizer is used for optimization to obtain the trained prediction model. The training data comes from three completed similar projects, containing approximately 120,000 task progress records. The loss function calculation process is as follows: For each task i and each future time step k, calculate the difference between the predicted and actual progress; square the difference to obtain the squared error; sum the squared errors of all tasks and all future time steps to obtain the total squared error; divide the total squared error by the product of the number of tasks and the number of future time steps, i.e., divide by N and multiply by 7, to obtain the mean squared error as the loss function value. The smaller the loss function value, the more accurate the model prediction.

[0113] The optimizer learning rate was set to 0.001, the batch size to 64, and the number of training epochs to 200. An early stopping strategy was employed, stopping training when the validation set loss did not decrease for 10 consecutive epochs. After training, the model achieved a mean absolute error of 3.2% and a root mean square error of 4.8% on the test set, representing an approximately 42% improvement in prediction accuracy compared to traditional time series prediction methods.

[0114] S47. Real-time progress prediction is performed using a trained model. The current digital twin state and historical progress data are input into the model, forward propagation calculation is performed, and the predicted progress values ​​for each task over the next 7 days are output, resulting in a progress prediction matrix. The prediction results contain the progress values ​​for 1850 tasks over the next 7 days, forming an 1850×7 matrix.

[0115] S48. Estimating the uncertainty of the forecast. The Monte Carlo Dropout method is employed, maintaining the activation of the Dropout layer during forecasting and performing multiple forward propagations to obtain the distribution of the forecast results. The forecast mean and variance are calculated to obtain the forecast result with uncertainty estimation. Specifically, 50 forward propagations are performed, each using a different Dropout mask, resulting in 50 forecast values. The mean of these 50 values ​​is calculated as the final forecast, and the standard deviation is calculated as the uncertainty measure. Predictions with high uncertainty indicate to managers that the task's progress is subject to significant variability and requires close monitoring.

[0116] Through the above sub-steps, we obtain the progress prediction results, prediction uncertainty estimates, and model performance evaluation reports for each task, which serve as inputs for the next step.

[0117] In some embodiments, a graph attention network is used instead of a graph convolutional network to further enhance the model's ability to model task dependencies. The graph attention network learns adaptive attention weights for each edge instead of using fixed normalization coefficients. The attention weight calculation process is as follows: The feature vectors of task i and task j are multiplied by the weight matrix to perform feature transformation; the two transformed vectors are concatenated to form a longer vector; then, the concatenated vector is multiplied by the learnable parameter vector of the attention mechanism using a dot product operation, i.e., element-wise multiplication and summation; finally, the Leaky ReLU activation function is applied to the dot product result. This activation function remains unchanged for positive values ​​and multiplies negative values ​​by a small negative slope coefficient to obtain the attention score of the edge from task i to task j.

[0118] The normalized attention weights are calculated as follows: For all neighboring nodes of task i, calculate the attention score of the edge from task i to each neighboring node; apply the natural exponential function to all attention scores; then sum all the natural exponential values; divide the sum by the natural exponential values ​​of the attention scores from task i to task j to obtain the normalized attention weights. Normalization ensures that the sum of the attention weights from task i to all neighboring nodes is 1.

[0119] The node update process is as follows: For each neighbor node j of task i, obtain the feature vector of task j in the current layer; perform matrix multiplication operation on the feature vector and the weight matrix of the current layer to perform feature transformation; then perform scalar and vector multiplication operation on the transformed feature vector and the normalized attention weights to obtain the weighted feature vector; then sum the weighted feature vectors of all neighbor nodes to obtain the aggregated feature; apply the activation function to the aggregated feature to obtain the feature vector of task i in the next layer.

[0120] Graph attention networks can automatically learn which dependencies are more important for progress prediction, assigning higher weights to strong dependencies and lower weights to weak dependencies. In experiments, the model using graph attention networks further improved prediction accuracy by approximately 8% compared to graph convolutional networks, especially performing better when handling complex dependencies. The purpose of this alternative embodiment is to improve the accuracy and interpretability of progress prediction by more accurately modeling the strength of dependencies between tasks through an adaptive attention mechanism.

[0121] S5, based on the status and progress prediction results of the digital twin, adopts a multi-level detection mechanism to identify progress deviations and analyze the root causes, and obtains deviation detection results and root cause analysis reports;

[0122] S51, Rule-based rapid deviation screening. The actual progress of each task is compared with the planned progress, and the progress deviation value is calculated. When the deviation exceeds a preset threshold, it is marked as a potential deviation task, resulting in a list of potential deviation tasks. The progress deviation calculation process is as follows: Obtain the actual progress percentage of task i at time t. This value is extracted from the digital twin state and reflects the actual completion level of the task. Obtain the planned progress percentage of task i at time t. This value is extracted from the construction plan baseline and reflects the completion level that the task should achieve according to the plan. Subtract the planned progress percentage from the actual progress percentage to obtain the progress deviation value. If the deviation value is positive, it indicates that the task is ahead of schedule; if the deviation value is negative, it indicates that the task is behind schedule; if the deviation value is zero, it indicates that the task is in line with the plan.

[0123] The deviation threshold is set according to the importance of the task. For tasks on the critical path, the threshold is set to 5%, and for non-critical tasks, the threshold is set to 10%. The marking process is as follows: Calculate the absolute value of the schedule deviation, eliminating the influence of the positive or negative sign; compare the absolute value with the deviation threshold corresponding to the task; if the absolute value exceeds the threshold, it indicates a significant deviation, and the deviation mark for task i is set to 1, marking it as a potential deviation task; if the absolute value does not exceed the threshold, it indicates that the deviation is within an acceptable range, and the deviation mark for task i is set to 0, not marking it as a potential deviation. Through this rapid screening, tasks that may have problems are initially identified. In this embodiment, 87 tasks are currently marked as potential deviations.

[0124] S52, Anomaly Detection Based on Statistical Learning. For potential deviation tasks, the time series characteristics of their progress deviations are analyzed. The Local Anomaly Factor (LAM) algorithm is used to detect abnormal patterns, filter normal fluctuations, identify true deviations, and obtain a list of abnormal tasks. Statistical features of the progress deviation time series are extracted, including mean, standard deviation, trend slope, and autocorrelation coefficient. The LAM algorithm calculates the degree of density deviation of each sample point relative to its neighborhood. The specific process is as follows: For sample i, the k nearest samples are found in the feature space to form the k-nearest neighbor set of sample i. In this embodiment, k is set to 10. The local reachability density of sample i is calculated. This density reflects the density of the samples around sample i; the higher the density, the denser the surrounding samples. Then, for each nearest neighbor sample j in the k-nearest neighbor set, its local reachability density is calculated. The ratio of the local reachability density of nearest neighbor sample j to the local reachability density of sample i is calculated. The density ratios of all nearest neighbor samples are summed. The summation result is divided by the number of elements in the k-nearest neighbor set to obtain the local anomaly factor of sample i. The local anomaly factor reflects the degree of density deviation of sample i relative to its neighborhood. If the density of samples around sample i is significantly higher than the density of sample i itself, the local anomaly factor is greater than 1, indicating that sample i is an anomaly. If the density of sample i is comparable to that of the surrounding samples, the local anomaly factor is close to 1, indicating that sample i is a normal point.

[0125] In this embodiment, the local anomaly factor threshold is set to 1.5. Potentially biased tasks with a local anomaly factor greater than 1.5 are identified as true deviations; tasks with a local anomaly factor less than or equal to 1.5 are judged as normal fluctuations and are not considered true deviations. After anomaly detection, 52 out of 87 potentially biased tasks were identified as true deviations.

[0126] S53 is a deep learning-based deviation type identification system. For confirmed deviation tasks, its multi-dimensional features are extracted and input into a trained classification model to identify the deviation type and severity, yielding the deviation classification result. Features include progress deviation value, deviation duration, resource allocation, environmental conditions, and historical deviation records, totaling 48 dimensions. The classification model employs a multilayer perceptron, consisting of an input layer, two hidden layers, and an output layer. The hidden layer dimensions are 64 and 32, respectively, with ReLU activation. The output layer uses softmax activation to output the probability distribution of the deviation type. Deviation types include six categories: insufficient personnel, equipment failure, material shortage, design changes, weather impact, and organizational coordination problems. Severity is categorized into three levels: minor, moderate, and severe. The model was trained on historical data, achieving a classification accuracy of 89%.

[0127] S54. Construct a causal graph model for root cause analysis. Based on domain knowledge and historical data, construct a causal graph of factors affecting construction progress. Nodes represent variables, and directed edges represent causal relationships, resulting in a causal graph model. The causal graph includes nodes such as progress, personnel, equipment, materials, weather, and management, as well as causal relationship edges between them. For example, "number of personnel" points to "construction efficiency," and "construction efficiency" points to "progress," indicating that the number of personnel affects progress by influencing construction efficiency.

[0128] S55, Calculate the causal effect of each factor on the schedule deviation. For each deviation task, based on the cause-effect diagram and observation data, the do-calculus method is used to calculate the causal effect of each potential cause factor on the schedule deviation, obtaining a causal effect ranking. The causal effect calculation process is as follows: Select a potential cause factor, denoted as factor X; Imagine intervening in factor X, increasing its value from its current value by a certain amount; Based on the cause-effect diagram and observation data, calculate the expected value of the schedule deviation of task i under the intervention condition, which is obtained by probability-weighted averaging over all possible scenarios; Simultaneously, calculate the expected value of the schedule deviation of task i under the non-intervention condition, i.e., when factor X remains at its current value; Subtract the expected value before intervention from the expected value after intervention to obtain the causal effect of factor X on the schedule deviation of task i. The causal effect reflects the actual degree of influence of changing factor X on the schedule deviation; a positive value indicates that increasing factor X will increase the schedule deviation, and a negative value indicates that increasing factor X will decrease the schedule deviation.

[0129] Causal effects are estimated from observed data using propensity score matching or instrumental variable methods. The above calculations are repeated for all potential causal factors to obtain the causal effect value for each factor. The larger the absolute value of the causal effect, the more significant the factor's impact on schedule deviations. The root causes with the greatest impact on schedule deviations are identified by sorting the causal effects from largest to smallest absolute value.

[0130] S56 generates targeted corrective action recommendations. Based on the identified deviation types and root cause analysis results, corresponding corrective measures are matched to generate specific corrective action recommendation schemes, resulting in a list of corrective action recommendations. A mapping rule library between deviation types and corrective action measures is established. For deviations caused by insufficient personnel, it is recommended to increase staffing or adjust work shifts; for deviations caused by equipment failure, it is recommended to arrange emergency repairs or allocate alternative equipment; for deviations caused by material shortages, it is recommended to expedite material procurement or adjust the construction sequence to avoid tasks with material shortages; for deviations caused by weather effects, it is recommended to adjust outdoor work plans or add protective measures; for deviations caused by organizational and coordination problems, it is recommended to hold coordination meetings or adjust task interfaces. Each recommendation includes information such as a description of the measure, expected effects, implementation difficulty, and required resources.

[0131] S57, Assess the impact of deviations on the overall schedule. Based on task dependencies and critical path analysis, calculate the impact of the deviation task on subsequent tasks and the overall project duration, and obtain an impact assessment report. The critical path method is used to calculate the total float and free float of the deviation task. Tasks with zero total float are located on the critical path, and their delays will directly lead to a delay in the overall project duration. Tasks with larger total float have a certain buffer, and their delays within the float range do not affect the overall project duration. The impact calculation process is as follows: Obtain the delay time of task i, i.e., the difference between the actual completion time and the planned completion time; obtain the total float of task i, which represents the maximum time the task can be delayed without affecting the overall project duration; then add 1 to the total float to avoid division by zero; divide the delay time by the result of the total float plus 1 to obtain the float consumption ratio, the larger the ratio, the more severe the delay; simultaneously count the number of subsequent tasks of task i, the more subsequent tasks, the wider the impact range; add 1 to the number of subsequent tasks; multiply the float consumption ratio by the result of the number of subsequent tasks plus 1 to obtain the impact degree of the deviation of task i. The greater the impact, the greater the threat that deviation in the task poses to the overall schedule, and the more it needs to be addressed.

[0132] Through the above sub-steps, deviation detection results and root cause analysis reports are obtained. Deviation detection results include a list of schedule deviation tasks, deviation types, and severity. Root cause analysis reports include root cause analysis results, a list of corrective action recommendations, and an impact assessment report, which serve as input for the next step.

[0133] In some embodiments, a deep reinforcement learning-based root cause diagnosis method is used to replace traditional causal inference methods, improving the accuracy and efficiency of root cause identification. Root cause diagnosis is modeled as a sequential decision problem, where the agent interacts with the environment to gradually narrow down the range of possible causes, ultimately locating the root cause. The state space contains currently known symptom information and ruled-out factors, the action space contains the selection of the next factor to check, and the reward function is designed based on whether the root cause is found and the cost of checking. A deep Q-network is used to train the diagnostic strategy, enabling the agent to find the root cause with the fewest checking steps. The training data comes from historical deviation cases, containing approximately 8000 deviation events and their root cause annotations. After training, the model achieves a root cause identification accuracy of 92% on the test set, with an average of 3.2 checking steps, representing an efficiency improvement of approximately 65% ​​compared to exhaustive search methods. The purpose of this alternative embodiment is to quickly locate the root cause with incomplete information through an intelligent diagnostic strategy, reducing diagnostic time and cost.

[0134] S6. Based on the deviation detection results and root cause analysis report, reinforcement learning method is used to generate an optimized resource scheduling scheme and obtain dynamic scheduling decision.

[0135] S61 defines the state space for the resource scheduling problem. The state includes information such as the current progress status of each task, the availability and allocation of various resources, and the environmental conditions, resulting in a state vector. Specifically, the state vector contains the progress percentages of 1850 tasks, the allocation of 320 workers, the location and load of 8 tower cranes, the operating status of 12 construction elevators, the inventory of various materials, and current weather conditions, totaling approximately 2500 features. These features are normalized, and discrete features are encoded using one-hot encoding.

[0136] S62 defines the action space for the resource scheduling problem. Actions include decisions on allocating personnel, equipment, and materials to each task, as well as decisions on adjusting task start times and durations, resulting in an action vector. Due to the extremely high dimensionality of the action space, a hierarchical decision-making strategy is adopted, selecting the tasks to be adjusted and choosing specific resource allocation schemes. The action vector includes task selection indexes, personnel allocation increments, equipment allocation schemes, material delivery plans, time adjustment amounts, etc., totaling approximately 150 dimensions.

[0137] S63 defines a multi-objective reward function. Taking into account multiple objectives such as schedule goals, cost constraints, quality requirements, and safety standards, a reward function is designed, resulting in a reward calculation formula. The reward function calculation process is as follows: Calculate the progress reward at time t based on the amount of work completed at that time; the more work completed, the higher the reward. Calculate the cost penalty at time t based on resource usage costs; the higher the cost, the greater the penalty. Next, calculate the delay penalty at time t based on the degree of deviation from the plan; the more severe the delay, the greater the penalty. Then, calculate the risk penalty at time t based on safety hazards and quality problems; the higher the risk, the greater the penalty. Multiply the progress reward by a weighting coefficient of 0.4, the cost penalty by a weighting coefficient of 0.25, the delay penalty by a weighting coefficient of 0.25, and the risk penalty by a weighting coefficient of 0.1. Each weighting coefficient is set according to the project objectives; in this embodiment, the schedule goal has the highest weight. Subtract the weighted cost penalty, delay penalty, and risk penalty from the weighted progress reward to obtain the total reward value at time t. A higher total reward value indicates a better scheduling scheme.

[0138] S64, Constructing a Digital Twin Simulation Environment. A reinforcement learning training environment is built based on the digital twin. This environment simulates the execution effect of resource scheduling decisions, updates the construction status according to the scheduling scheme, and returns rewards, resulting in a simulation environment model. The environment's state transition function is constructed based on the physical laws and empirical models of the construction process, considering the impact of resources on construction efficiency, dependencies between tasks, and interference from environmental factors. The environment supports rapid simulation, capable of simulating several days of construction in seconds.

[0139] S65 trains a deep reinforcement learning scheduling policy. A proximal policy optimization algorithm is used to train the scheduling policy network. The agent interacts with the simulation environment, learning the optimal scheduling policy through trial and error to obtain a trained policy model. The policy network adopts an Actor-Critic architecture; the Actor network outputs the probability distribution of actions, and the Critic network evaluates the value of states. Both networks use a multilayer perceptron structure with three hidden layers of 512, 256, and 128 dimensions, respectively. During training, the agent performs actions in the simulation environment, collecting experience samples of states, actions, rewards, and the next state, storing them in an experience replay buffer. After collecting a certain number of samples, the policy is updated by sampling from the buffer. Training is accelerated using a parallel environment, running 16 simulation environment instances simultaneously. The total training steps are 5 million, taking approximately 48 hours.

[0140] S66 introduces environmental disturbances to enhance policy robustness. During training, disturbances such as equipment failure, material delays, and sudden weather changes are randomly injected, enabling the policy to learn to cope with uncertainty and obtain a robust scheduling strategy. The probability and intensity of the disturbance events are set based on historical statistical data: the probability of equipment failure is 2% per day, the probability of material delay is 5% per week, and the probability of severe weather is 10% per month. By training in a disturbed environment, the policy's generalization ability and robustness are improved.

[0141] S67 generates an optimized scheduling plan for the current state. The current digital twin state is input into the trained policy network, forward propagation is performed, and resource scheduling decisions are output, resulting in the optimized scheduling plan. The scheduling plan includes a detailed resource allocation plan for the next three days, specifying the number of personnel, equipment used, and materials delivered for each task each day. The plan is presented in tabular and Gantt chart formats for easy viewing and execution by management.

[0142] S68. Evaluate the expected effects of the scheduling scheme. Execute the scheduling scheme in a simulation environment to simulate its implementation effects, calculate the expected schedule improvement, cost changes, risk level, and other indicators, and obtain a scheme evaluation report. The simulation is run 100 times, with different random disturbances injected each time, and the average effect and confidence interval are statistically analyzed. The evaluation results show that after adopting the optimized scheduling scheme, it is expected that delayed tasks on the critical path can be completed 2.3 days earlier, the total cost will increase by approximately 1.8%, and the risk level will remain within an acceptable range.

[0143] Through the above sub-steps, we obtain an optimized resource scheduling scheme, the expected progress improvement effect, and a robustness evaluation report of the scheduling scheme, which serve as inputs for the next step.

[0144] In some embodiments, a multi-agent reinforcement learning approach is employed to replace the single-agent approach, improving scheduling performance in complex scenarios. Different types of resource scheduling are modeled as collaborative decision-making problems involving multiple agents. Personnel scheduling agents, equipment scheduling agents, and material scheduling agents are each responsible for optimizing the allocation of their respective resources, reaching a globally optimal solution through communication and negotiation. A multi-agent deep deterministic policy gradient algorithm is used for training. Each agent observes the local state and the actions of other agents, outputting its own action decision. Agents communicate through an attention mechanism, learning which agents' decisions are more important to them. During training, agents must optimize their own goals while also considering collaboration with other agents. After training, the multi-agent system demonstrates approximately 18% better scheduling performance under complex constraints compared to a single agent, particularly excelling in scenarios with frequent resource conflicts. The purpose of this alternative embodiment is to better handle large-scale, complex resource scheduling problems through distributed decision-making and agent collaboration, improving the quality and computational efficiency of scheduling schemes.

[0145] S7, based on dynamic scheduling decision-making, builds an intelligent decision support system, providing visualization and interactive functions, and providing decision support interface and analysis reports;

[0146] S71 constructs a 3D visualized digital twin interface. Utilizing the WebGL 3D graphics engine to render the digital twin, it displays the building's 3D model and construction status in a browser, resulting in a visualized interface. The interface supports mouse and touch operations, allowing users to rotate, zoom, and pan the view to examine any angle and detail of the building. Components are colored according to their construction status: unstarted components are gray, in progress are yellow, completed components are green, and delayed components are red. Clicking on a component displays its detailed information, including component number, assigned task, planned progress, actual progress, and responsible work team.

[0147] The S72 module implements a progress timeline playback function. A timeline control is added to the bottom of the interface, allowing users to drag the timeline to view the construction status at any historical point in time, or click the play button to automatically play the progress evolution process, thus achieving timeline playback. The timeline ranges from the start of the project to the current moment, with time granularity selectable by day, week, or month. Playback speed is adjustable, supporting 1x, 2x, 5x, and 10x speeds. During playback, the color and position of components dynamically change over time, intuitively displaying the progress of construction. Users can analyze historical progress changes through the playback function to identify key events and turning points.

[0148] S73 visualizes the future progress forecast. The predicted progress output from step S4 is overlaid onto the 3D model, displaying the predicted future state in a semi-transparent or dashed line format, resulting in a forecast visualization effect. Users can switch to a specific future point in time to view the predicted construction status. Predicted components are rendered semi-transparently to distinguish them from actual components. Components with high forecast uncertainty are marked with dashed borders to indicate that the forecast involves significant variables. Forecast visualization helps managers identify potential problems in advance and develop preventative measures.

[0149] S74 integrates various charts to display progress information. The sidebar integrates Gantt charts, network diagrams, S-curves, heatmaps, and other charts to display progress information from different perspectives, providing a multi-dimensional progress overview. Gantt charts display the planned time, actual time, and schedule deviation for each task, supporting filtering and grouping by specialty, floor, and region; network diagrams show the dependencies and critical paths between tasks, with the critical path highlighted in red; S-curves show the cumulative completed work volume over time, comparing the planned curve with the actual curve; heatmaps show the progress completion rate for different regions or floors, with color intensity indicating the degree of completion. Users can interact between the 3D view and the charts; clicking on a task in a chart automatically locates the corresponding component in the 3D view.

[0150] The S75 implements intelligent question answering based on natural language. Integrating a natural language processing module, users can input questions via text or voice. The system understands the intent of the question and returns an answer, providing intelligent question answering functionality. It employs a pre-trained BERT model for intent recognition and entity extraction, converting natural language questions into structured queries. Supported question types include status queries, cause analysis, predictive queries, and suggestion consultations. For example, if a user inputs "What is the progress of the beam installation on the 20th floor?", the system identifies the query intent, the floor entity "20th floor", and the component type "beam", retrieves relevant information from the knowledge graph, and generates the answer "There are 128 beam components on the 20th floor, 96 of which have been completed, a completion rate of 75%, in line with the planned schedule." For complex questions, the system conducts multiple rounds of dialogue to gradually clarify the user's needs.

[0151] The S76 provides scenario simulation and solution comparison functions. Users can adjust resource configuration, modify construction sequence, and set environmental conditions in the interface. The system calculates simulation results based on a progress prediction model, displaying the effects of different solutions, thus providing the scenario simulation function. Users select the task to be simulated, adjust parameters such as personnel configuration, equipment allocation, and start time, and click the "Simulate" button. The system runs the prediction model in the background, calculates indicators such as progress, cost, and risk under the given solution, and displays the simulation results on the interface. Users can create multiple solutions, and the system generates a solution comparison table, listing the performance of each solution on different indicators to help users select the optimal solution.

[0152] S77, Constructing a Schedule Management Knowledge Base. This involves mining experiential knowledge from historical project data, using association rule mining algorithms to extract key factors influencing schedule, employing cluster analysis to identify common schedule problem patterns, and using causal discovery algorithms to learn the causal relationships between factors, resulting in the knowledge base. Knowledge is stored in the form of rules, cases, and patterns. For example, the rule "If it rains for three consecutive days and the task is exterior wall construction, the probability of schedule delay is 85%"; the case "Project A used a manual handling solution when the tower crane malfunctioned, increasing costs by 12% but avoiding schedule delays." The knowledge base supports keyword and semantic searches, allowing users to query relevant knowledge and refer to historical experience when encountering problems. The system also supports administrators manually adding knowledge, enabling continuous knowledge accumulation.

[0153] The S78 employs a cloud-edge-device collaborative architecture. Edge devices are deployed at the construction site, running lightweight models to handle data collection, preliminary processing, and real-time response (response time in the milliseconds). The cloud device, deployed in the data center, runs a complete model, handling complex calculations, big data analysis, and model training, boasting powerful computing capabilities. Devices include PCs, mobile devices, and VR devices, providing interactive interfaces for different scenarios, thus achieving the collaborative architecture. Edge devices and the cloud are connected via a 5G network for real-time data synchronization. For urgent queries and simple predictions, the edge devices respond directly; for complex analysis and optimization tasks, they are submitted to the cloud for processing. A mobile app is provided for on-site management personnel, allowing them to view progress, receive alerts, and make approval decisions anytime, anywhere. VR devices offer an immersive experience; managers wearing VR headsets can virtually tour the construction site and gain a more intuitive understanding of the progress status.

[0154] S79 integrates blockchain technology to achieve trusted data storage. Key progress data, decision records, and change information are stored on the blockchain using a consortium blockchain architecture. All participating parties act as nodes in the consensus process, resulting in a blockchain-based data storage system. Each record includes a timestamp, data hash, operator identity, and digital signature, ensuring data immutability. When progress data is updated, the system automatically calculates the data hash value, packages it along with the timestamp and operator information into a transaction, broadcasts it to the blockchain network, and writes it into a block after verification by the consensus algorithm. Subsequently, anyone can verify the authenticity and integrity of the data, supporting full-process traceability. Blockchain-based data storage provides a reliable basis for resolving engineering disputes and determining liability.

[0155] Through the above sub-steps, a visualized progress control interface, an intelligent question-and-answer system, a scenario simulation tool, a progress control knowledge base, a cloud-edge-device collaborative architecture, and a blockchain evidence storage system are obtained, providing managers with comprehensive decision support.

[0156] In some embodiments, augmented reality (AR) technology is used instead of virtual reality (VR) technology to provide a more convenient on-site visualization experience. Managers use AR glasses or the AR function of their mobile phones to overlay digital twin information onto the actual construction site. Through image recognition and spatial positioning technology, the system identifies the components and locations on site and overlays corresponding virtual information such as progress information, quality status, and safety reminders onto the real scene. For example, when a manager looks at a beam, the AR interface automatically displays the beam's number, installation progress, quality inspection results, and responsible work team. AR technology seamlessly integrates digital information with the physical site, allowing managers to obtain rich information without switching perspectives, thus improving the efficiency and accuracy of on-site management. The purpose of this alternative embodiment is to achieve the direct presentation of digital twin information on the physical site through augmented reality technology, lowering the barrier to entry and enhancing the user experience.

[0157] A digital twin progress control system for construction projects, such as Figure 2 As shown, a digital twin-based schedule control method for construction projects, as described above, includes:

[0158] The multi-source data acquisition module, based on the project's BIM model and the IoT sensing system deployed at the construction site, collects multi-source heterogeneous data during the construction process to obtain the raw data set;

[0159] The spatiotemporal alignment fusion module uses a spatiotemporal alignment network to fuse the original dataset, resulting in spatiotemporally consistent and semantically unified fused data.

[0160] The digital twin update module updates the digital twin based on fused data using an incremental state estimation method to obtain the current state of the digital twin.

[0161] The progress prediction module, based on the state of the digital twin, uses a model that integrates graph neural networks and temporal attention mechanisms to predict future progress and obtain the progress prediction results for each task.

[0162] The deviation detection and root cause analysis module, based on the digital twin status and schedule prediction results, uses a multi-level detection mechanism to identify schedule deviations and analyze their root causes, and obtains deviation detection results and root cause analysis reports.

[0163] The resource scheduling optimization module, based on deviation detection results and root cause analysis reports, uses reinforcement learning to generate optimized resource scheduling schemes and obtain dynamic scheduling decisions.

[0164] The intelligent decision support module, based on dynamic scheduling decisions, constructs an intelligent decision support system, providing visualization and interactive functions, and delivering a decision support interface and analysis reports.

[0165] In one embodiment of the present invention, a specific example is provided:

[0166] This invention focuses on the application of schedule control in super high-rise complex construction projects. Using a super high-rise complex project in a city center as a case study, the project faced typical challenges during construction, including frequent cross-disciplinary operations, limited vertical transportation resources, and difficulties in schedule control. By deploying the digital twin schedule control system of this invention, refined and intelligent management of the construction schedule was achieved.

[0167] The following table shows some of the data collected by the system when the project was in its 18th month of construction:

[0168]

[0169] After performing spatiotemporal alignment and fusion processing on the collected data, the system constructed a digital twin containing 160,000 components. Using an incremental update strategy, the system updates the digital twin's state every minute, with each update involving approximately 2,800 components and a computation time of about 45 seconds, meeting real-time requirements.

[0170] In the third week of the 18th month, the system identified 52 tasks with schedule deviations through a multi-level deviation detection mechanism, including 8 tasks on the critical path. Root cause analysis was performed on task T0856, which involved the concrete pouring of the 20th-floor core tube. The planned progress was 90%, the actual progress was 85%, a delay of 2 days. System analysis revealed that the main cause of the delay was insufficient concrete supply, with a causality effect value of 0.68, followed by low staff attendance, with a causality effect value of 0.32. The system's corrective action recommendations included: coordinating with the concrete mixing plant to increase supply, scheduling nighttime pouring shifts, and allocating additional personnel from adjacent floors for support.

[0171] The resource scheduling module based on reinforcement learning generated an optimized plan, suggesting adding 5 workers to task T0856 within the next 3 days, coordinating with the concrete mixing plant to supply an additional 60 cubic meters of concrete daily, and adjusting the operation priority of tower crane No. 3 to prioritize this task. The system evaluated this plan in a simulation environment, estimating that it could make up for the delay in task T0856 within 2 days, with a total cost increase of approximately 8,000 yuan, and the risk level is manageable.

[0172] Project managers viewed the construction status of the 20th floor through a 3D visualization interface. They noticed that the components in the core tube area were colored yellow, indicating progress but with a risk of delay. Clicking on the relevant component displayed detailed information and corrective action suggestions. Managers used the scenario simulation function to compare the effects of different resource allocation schemes and ultimately adopted the system-recommended optimization plan.

[0173] Within three days of implementing the solution, the system continuously monitored the progress of task T0856. Actual data showed that the increased concrete supply improved pouring efficiency, raising the task progress from 85% to 98%, successfully making up for the delay. The system recorded this case in its knowledge base for future reference in similar situations.

[0174] Through the digital twin progress control method of this invention, the progress prediction accuracy of this super high-rise complex project reached 91% and the timely identification rate of progress deviations reached 95% throughout the entire construction process. The critical path delay events were reduced by 62%, and the resource utilization efficiency was improved by 28%. The project was finally completed and delivered on schedule, achieving certain management results.

[0175] This invention achieves intelligent full-process management and control of construction project progress through innovative technologies such as multi-source heterogeneous data fusion, graph neural network progress prediction, incremental digital twin updates, multi-level deviation detection, and reinforcement learning resource scheduling. Compared with traditional progress management methods, this invention has the following advantages: higher data fusion depth, effectively handling spatiotemporal inconsistencies; higher progress prediction accuracy, simultaneously capturing task topology relationships and temporal characteristics; higher digital twin update efficiency, supporting real-time management and control of large-scale projects; more refined deviation identification, quickly locating root causes; more intelligent resource scheduling, capable of coping with dynamic changes and uncertainties; and more comprehensive decision support, providing multi-dimensional visualization and intelligent analysis. This invention provides important technical support for the digital transformation of construction projects and has significant practical value in promoting the intelligent development of the construction industry.

[0176] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for digital twin-based progress control in construction projects, characterized in that, include: S1, based on the project's BIM model and the IoT sensing system deployed at the construction site, collects multi-source heterogeneous data during the construction process to obtain a raw data set; wherein, S1 includes: extracting component information based on the BIM model, parsing the geometric information, attribute information, and construction plan information of components from the BIM model file, establishing a mapping relationship between component identifiers and spatial location, construction tasks, and planned time to obtain a component information database; deploying an IoT sensing network to collect real-time data, deploying RFID readers, cameras, laser scanners, environmental sensors, and equipment monitoring devices at the construction site to collect real-time data on component location, personnel activities, equipment operation, and environmental parameters to obtain an IoT sensing data stream; adding quality assessment indicators to the data records, calculating a comprehensive quality score based on four dimensions of data completeness, timeliness, accuracy, and consistency using a weighted summation method to obtain a dataset with quality labels; S2 employs a spatiotemporal alignment network to fuse the original dataset, resulting in spatiotemporally consistent and semantically unified fused data. S2 includes: constructing an adaptive time window for time alignment; adjusting the time window length adaptively based on the data change rate to obtain time alignment parameters, taking into account the sampling frequency differences of different data sources; performing weighted fusion within the time window; calculating weights based on time distance and data quality; and performing a weighted average of the observations to obtain fused time-aligned data; establishing a unified engineering coordinate system for spatial alignment; using the BIM model coordinate system as a reference, transforming the spatial coordinates of each data source to the unified coordinate system using a homogeneous coordinate transformation method to obtain spatially aligned data; and using knowledge graph technology to organize the fused data, constructing a knowledge graph from the fused data, where nodes represent entities and edges represent relationships, to obtain a construction progress knowledge graph. S3, based on fused data, uses an incremental state estimation method to update the digital twin, obtaining the current state of the digital twin; wherein, S3 includes: identifying the set of changed components, comparing the difference between the new data and the state of the digital twin at the previous moment, identifying the changed components, and obtaining the set of changed components; updating the state of the changed components by fusing the observed data and predicted state using a Kalman filter method, updating the state vector of the changed components, and obtaining the updated component state; simplifying the update of unchanged components by directly inheriting the state at the previous moment or making linear predictions based on the construction plan, and obtaining the state of the unchanged components; dynamically adjusting the component update priority by calculating the update priority based on the importance and change frequency of the components, and obtaining the priority ranking; S4, based on the state of the digital twin, uses a model that integrates graph neural networks and temporal attention mechanisms to predict future progress, obtaining the progress prediction results for each task; wherein, S4 includes: constructing a construction task dependency graph, modeling the construction task network as a directed graph, where nodes represent construction tasks and directed edges represent dependencies between tasks, obtaining a task dependency graph; using a graph convolutional network to learn task node embeddings, aggregating neighbor node information through multi-layer graph convolution operations to obtain task embedding vectors; using a temporal attention mechanism to capture progress evolution features, extracting historical state sequences, calculating the importance weights of different historical moments to the current prediction, obtaining a weighted historical state representation; fusing spatial and temporal features for progress prediction, concatenating the task embedding vectors and historical state representations, inputting them into the prediction layer, and outputting future progress prediction values ​​to obtain the progress prediction results; S5, based on the digital twin status and schedule prediction results, employs a multi-level detection mechanism to identify schedule deviations and analyze their root causes, obtaining deviation detection results and a root cause analysis report; wherein, S5 includes: rule-based rapid deviation screening, comparing actual progress with planned progress, marking potential deviation tasks when deviations exceed preset thresholds, and obtaining a list of potential deviation tasks; statistical learning-based anomaly detection, analyzing the time series characteristics of schedule deviations, using a local anomaly factor algorithm to detect anomaly patterns, identifying true deviations, and obtaining a list of anomaly tasks; constructing a causal graph model for root cause analysis, building a causal graph of factors affecting construction progress based on domain knowledge and historical data, and obtaining a causal graph model; calculating the causal effect of each factor on schedule deviations, calculating the causal effect of each potential cause on schedule deviations based on the causal graph and observation data, and obtaining a causal effect ranking; S6. Based on the deviation detection results and root cause analysis report, reinforcement learning method is used to generate an optimized resource scheduling scheme and obtain dynamic scheduling decision. S7, based on dynamic scheduling decisions, builds an intelligent decision support system, providing visualization and interactive functions, and providing decision support interfaces and analysis reports.

2. The method for digital twin progress control of construction projects according to claim 1, characterized in that, The Kalman filtering method includes: Based on the state of the previous moment and the construction plan, the current state is predicted, and the predicted state vector is calculated using the state transition matrix and the control input matrix. The Kalman gain is calculated based on the prediction error covariance matrix, the observation matrix, and the observation noise covariance matrix. The state is updated by fusing observation data, and the state correction is calculated based on the observation residuals and the Kalman gain matrix to obtain the updated state vector.

3. The method for digital twin progress control of construction projects according to claim 1, characterized in that, S6 includes: Define the state space of the resource scheduling problem. The state includes the progress status of each task, the availability and allocation of resources, and the environmental state, resulting in a state vector. Define a multi-objective reward function that comprehensively considers multiple objectives, including schedule objectives, cost constraints, quality requirements, and safety standards. Use a weighted method to calculate schedule progress rewards, cost penalties, delay penalties, and risk penalties to obtain the total reward value. A digital twin simulation environment is constructed, and a reinforcement learning training environment is built based on the digital twin to simulate the execution effect of resource scheduling decisions, thereby obtaining a simulation environment model. A deep reinforcement learning scheduling policy is trained, and a policy optimization algorithm is used to train the scheduling policy network. The agent interacts with the simulation environment to learn the optimal scheduling policy, and a well-trained policy model is obtained.

4. The method for digital twin progress control of construction projects according to claim 1, characterized in that, S7 includes: A three-dimensional visualization digital twin interface is constructed, and a three-dimensional graphics engine is used to render the digital twin to display the three-dimensional model and construction status of the building. The components are displayed differently according to the construction status to obtain the visualization interface. Implement the progress timeline playback function, allowing users to view the construction status at any point in history or automatically play the progress evolution process to obtain the timeline playback function; To achieve intelligent question answering based on natural language, users can input questions via text and voice, and the system will understand the intent of the question and return the answer, thus providing intelligent question answering functionality; Deploy a cloud-edge-device collaborative architecture, where the edge device is responsible for data collection and real-time response, the cloud device is responsible for complex calculations and model training, and the device provides interactive interfaces for different scenarios, thus obtaining the collaborative architecture.

5. The method for digital twin progress control of construction projects according to claim 1, characterized in that, The additional quality assessment indicators for data records in S1 include: To assess the completeness of data records, check whether each required field in the data record is completely filled, and calculate the ratio of the number of filled fields to the total number of fields to obtain a completeness score. The timeliness of data records is assessed by obtaining the data collection timestamp and the current timestamp, calculating the time interval between the two, and determining the timeliness score based on the length of the time interval. The accuracy of the data records is assessed by querying the historical reliability records of the sensor from which the data was obtained and the date of the most recent calibration. An accuracy score is calculated based on the sensor’s historical failure rate and calibration timeliness. To assess the consistency of data records, cross-validate the data with the same or related data collected from other data sources, calculate the degree of consistency between data values, and obtain a consistency score. The comprehensive quality score is obtained by multiplying the completeness score by a weighting factor of 0.3, the timeliness score by a weighting factor of 0.25, the accuracy score by a weighting factor of 0.25, and the consistency score by a weighting factor of 0.

2.

6. A digital twin progress control system for construction projects, characterized in that, A method for implementing a digital twin progress control method for construction projects as described in any one of claims 1-5, comprising: The multi-source data acquisition module, based on the project's BIM model and the IoT sensing system deployed at the construction site, collects multi-source heterogeneous data during the construction process to obtain the raw data set; The spatiotemporal alignment fusion module uses a spatiotemporal alignment network to fuse the original dataset, resulting in spatiotemporally consistent and semantically unified fused data. The digital twin update module updates the digital twin based on fused data using an incremental state estimation method to obtain the current state of the digital twin. The progress prediction module, based on the state of the digital twin, uses a model that integrates graph neural networks and temporal attention mechanisms to predict future progress and obtain the progress prediction results for each task. The deviation detection and root cause analysis module, based on the digital twin status and schedule prediction results, uses a multi-level detection mechanism to identify schedule deviations and analyze their root causes, and obtains deviation detection results and root cause analysis reports. The resource scheduling optimization module, based on deviation detection results and root cause analysis reports, uses reinforcement learning to generate optimized resource scheduling schemes and obtain dynamic scheduling decisions. The intelligent decision support module, based on dynamic scheduling decisions, constructs an intelligent decision support system, providing visualization and interactive functions, and delivering a decision support interface and analysis reports.

Images