A municipal road intelligent construction and collaborative management method based on digital twinning

Abstract

The application discloses a kind of municipal road intelligent construction and collaborative management method based on digital twinning, belong to digital twinning field, comprising: S1.BIM, construction monitoring, environment and equipment data constitute unified data twin body;S2.real-time synchronization is realized by time series mapping virtual model and entity;S3.adopt multi-agent collaborative optimization algorithm to carry out whole cycle control;S4.embed risk evolution prediction model to carry out structure and material risk assessment;S5.based on scene layered distribution multi-task;S6.through two-way communication, realize progress self-checking and closed-loop control;S7.multiple stage data fusion realizes model consistency update;S8.establishes whole life cycle carbon emission and energy consumption evaluation system;S9.output comprehensive decision report, realize planning, construction and operation integrated management.A beneficial effect: improve construction efficiency and decision accuracy.

Description

Technical Field

[0001] This invention relates to the field of digital twins, and more specifically, to a method for intelligent construction and collaborative management of municipal roads based on digital twins. Background Technology

[0002] With the digital transformation of urban infrastructure construction, municipal road construction and management are gradually transitioning from experience-driven to data-driven approaches. Early technologies relied heavily on single BIM models or sensor monitoring systems, achieving structural visualization and localized monitoring, but lacking the ability for virtual-physical interaction and full-cycle collaboration. Subsequently, the introduction of the Internet of Things, cloud computing, and big data analytics led to initial integration of construction monitoring and remote management; however, issues such as data silos, static model updates, and inefficient multi-tasking scheduling persist, resulting in information lag and decision-making delays. In recent years, the introduction of digital twin technology has promoted virtual-physical integrated management, but traditional solutions largely focus on static mapping and one-way monitoring, failing to achieve closed-loop control for real-time dynamic synchronization, intelligent collaborative decision-making, and risk prediction. This invention addresses these shortcomings by employing innovative methods such as multi-source heterogeneous data modeling, twin synchronization mechanisms, multi-agent collaborative optimization, and risk evolution prediction. This significantly improves the system's real-time performance, autonomy, and refined management level, enabling a shift from passive monitoring to proactive optimization and driving the development of municipal road construction towards intelligent, green, and full-lifecycle management. Summary of the Invention

[0003] The purpose of this invention is to provide a method for intelligent construction and collaborative management of municipal roads based on digital twins, in order to solve the problems mentioned in the background: In recent years, the introduction of digital twin technology has promoted the integration of virtual and real management, but traditional solutions are mostly focused on static mapping and one-way monitoring, which have failed to achieve closed-loop control of real-time dynamic synchronization, intelligent collaborative decision-making and risk prediction.

[0004] Technical Solution: A method for intelligent construction and collaborative management of municipal roads based on digital twins includes the following steps: S1. Establish a multi-source heterogeneous data twin modeling system for municipal road engineering, and construct a unified data twin composed of BIM model, construction monitoring data, environmental data and equipment data; the unified data twin achieves dynamic mapping of structural components, construction stages and real-time status through semantic layering, attribute encoding and multi-dimensional coordinate unification during the modeling process, forming a high-fidelity virtual-real integrated model that can be directly called by the full-cycle management system. S2. Based on time-series mapping, a twin synchronization mechanism is used to achieve bidirectional dynamic matching and state synchronization between the virtual model and the actual entity. Through timestamp alignment, data delay compensation and asynchronous stream merging algorithms, the data update delay between the virtual and the real is less than 1 second, ensuring the timeliness consistency between the twin model and the actual project in the three-dimensional space of progress, location and state. S3. Employ a multi-agent collaborative optimization algorithm to construct a full-cycle intelligent collaborative control module covering design, construction, testing, and operation and maintenance. Each agent participates in task allocation, resource scheduling, and anomaly detection as a role-based digital node, and continuously optimizes construction decisions through a reinforcement learning self-evolution mechanism. S4. Embed a risk evolution prediction model in the digital twin to conduct advance assessment of road structure deformation, material performance degradation and environmental load response; utilize the coupled analysis of historical data and real-time twin status to achieve dynamic identification and trend prediction of structural deterioration and safety risks. S5. An adaptive task allocation mechanism based on scene layering coordinates intelligent operation tasks of multiple trades and equipment; through scene partitioning and mapping, equipment capacity assessment and resource load balancing, it generates the optimal task path and updates the construction progress in real time. S6. Through two-way communication between the digital twin at the construction site and the remote decision-making platform, the system implements self-verification of construction progress and closed-loop control of abnormal states; the system uses a difference detection algorithm to determine the deviation of the twin state and triggers automatic diagnosis and adjustment strategies to achieve dynamic correction. S7. A multi-stage data fusion and reconstruction method is adopted to perform consistency correction and structural self-updating on the data collected in the twin model; a hybrid modeling method of graph convolution and autoregression is used to achieve semantic unification and dynamic reconstruction between different stages and different data sources. S8. Based on the operating status of the digital twin, establish a dynamic assessment system for carbon emissions and energy consumption throughout the entire life cycle of the road; through the coupled calculation of equipment energy consumption records, material carbon coefficients and construction plans, achieve refined quantification of carbon emissions during the construction and operation periods; S9. Output a comprehensive decision-making report on intelligent construction and collaborative management of municipal roads, realizing an intelligent control closed loop integrating planning, construction and operation and maintenance.

[0005] Preferably, the S1 multi-source heterogeneous data twin modeling system includes the following steps: S1-1. Perform multi-level semantic decomposition on the BIM model to extract structural components, material types, node connections, and construction stage attributes; during the decomposition process, hierarchical organization is carried out based on three-dimensional geometric topology to ensure that the contextual association and spatial logic of each component are consistent. S1-2. Reconstruct the spatial-temporal coordinates of the construction monitoring data, perform coordinate calibration using multi-node sensor array data, eliminate dynamic measurement noise through Kalman filtering, and form a traceable dynamic attribute matrix; S1-3. The environmental data and equipment data are feature-encoded, and double-aligned and mapped according to timestamps, geographic coordinates and BIM semantic nodes, so that the equipment status, construction climate conditions and material performance data are kept synchronized in real time in the unified twin data body; S1-4. The unified twin data body is constructed by a high-dimensional feature fusion algorithm, and the feature mutual information entropy optimization algorithm is used to reduce redundancy, so as to realize the adaptive interoperability of multi-source heterogeneous data under the unified coordinate, time and semantic level.

[0006] Preferably, the S1-1 multi-layer semantic decomposition further includes the following steps: S1-1-1. A structural graph convolutional neural network is used to extract features of the topological relationships between BIM components and to establish a node-edge weight graph to identify key connection parts and highly sensitive structural units. S1-1-2. Perform dynamic weight allocation according to node importance, and establish a reconstructable semantic hierarchical network by using the PageRank-based hierarchical importance ranking method to achieve reversible mapping of information at different hierarchical levels; S1-1-3. Redundant nodes are removed and information is aggregated in the semantic hierarchical network, with the removal rate controlled at 20%-30%, to form a lightweight interactive twin model.

[0007] Preferably, the S3 multi-agent collaborative optimization algorithm is constructed based on a dynamic weighted game mechanism, where each agent represents a design, construction, inspection, or operation and maintenance role; each agent calculates behavioral conflicts and concession conditions through a task payoff matrix, and selects the optimal strategy through game learning in a real-time simulation space; when the payoff convergence rate is lower than a set threshold, the system automatically triggers a reinforcement learning backoff mechanism to redistribute the weight parameters, thereby achieving coordinated optimization and dynamic balance of multiple roles throughout the entire cycle.

[0008] Preferably, the S3 multi-agent cooperative optimization algorithm includes the following steps: S3-1. Construct the role behavior state space and benefit matrix, and define the task inputs, outputs and constraints for each role; S3-2. Calculate task priority based on time decay function, and assign dynamic time weights to construction progress, equipment utilization and risk level; S3-3. Employ reinforcement learning strategies to iteratively optimize the collaboration paths of each role, and continuously refine the policy function based on feedback from historical task data; S3-4. Based on the real-time task completion rate, the global state is fed back, and the system automatically updates the reward signal, ultimately achieving global collaborative convergence of multiple agents under multi-dimensional constraints.

[0009] Preferably, the S4 risk evolution prediction model adopts a multi-layer prediction structure based on spatiotemporal convolution and graph attention mechanisms, and constructs a risk propagation path map by fusing material aging data, structural strain data and meteorological load parameters; The risk evolution prediction model identifies anomalies through spatial clustering in the early stages of structural state changes, and combines a time window sliding prediction algorithm to form a dynamically adjustable risk response range, thereby achieving early warning of structural damage and adaptive correction of construction parameters.

[0010] Preferably, the S5 scene-layered adaptive task allocation mechanism includes the following steps: S5-1. Divide the construction site into foundation operation layer, structural layer and equipment layer, and construct a hierarchical topology diagram based on the site spatial layout; S5-2. Perform task partitioning and coding for each layer, including task number, resource requirements, execution time limit and constraints. S5-3. Calculate the task coupling degree based on the multi-layer task status monitoring matrix, and generate a coupling weight matrix according to task dependencies, equipment capabilities and time overlap. S5-4. Based on the results of the task coupling degree, perform adaptive task matching and resource scheduling, and achieve optimal resource allocation for multi-task parallel operations by minimizing the weighted load cost function.

[0011] Preferably, the S5-3 task coupling degree calculation further includes the following steps: S5-3-1. Extract the spatiotemporal correlation features between task nodes through dual-channel tensor decomposition, and construct the task interaction matrix in the dual domains of the task time axis and spatial coordinate axis; S5-3-2. A nonlinear synergy factor model is used to calculate the dynamic mutual influence between tasks, and the synchronicity of tasks is quantified by high-order synergy coefficients. S5-3-3. Construct a task collaboration network graph, cluster highly coupled tasks, and generate a synchronization scheduling benchmark based on the similarity matrix to achieve dynamic decoupling of multiple tasks and global progress optimization.

[0012] Preferably, the S7 multi-stage data fusion and reconstruction method uses an autoregressive-graph convolution hybrid model to perform dynamic consistency correction; Within the data update cycle, the adaptive structural reconstruction of the twin model is achieved through the time difference integral strategy and the dynamic balancing algorithm of node reconstruction weights, maintaining the topological consistency and data fidelity of the twin in the multi-stage information evolution.

[0013] Compared with the prior art, the advantages of this invention are: (1) Introduce a multi-source heterogeneous data twin modeling system to achieve deep semantic fusion of BIM model, monitoring data, environment and equipment data, and break through the fragmentation problem of traditional single BIM or IoT modeling.

[0014] (2) By adopting the twin synchronization mechanism of time series mapping, a virtual and real two-way dynamic alignment channel is constructed, which significantly reduces data delay and model bias.

[0015] (3) Establish a multi-agent collaborative optimization algorithm to achieve adaptive coordination between design, construction, testing and operation and maintenance through dynamic weight game mechanism, and break through the decision-making bottleneck of single-center scheduling.

[0016] (4) Embed the risk evolution prediction model into the twin system, and identify potential risks of structure and material in advance through spatiotemporal convolution and graph attention structure to improve the predictive foresight.

[0017] (5) Construct a scenario-based hierarchical adaptive task allocation mechanism to realize multi-level dynamic scheduling based on processes, equipment and environment, and solve the resource conflict problem of multi-job collaboration.

[0018] (6) Introduce dual-channel tensor decomposition and nonlinear synergistic factor model to calculate task coupling degree, finely characterize the spatiotemporal correlation between tasks, and improve the accuracy of construction organization optimization.

[0019] (7) A multi-stage data fusion and reconstruction model is adopted to achieve continuous consistency and adaptive evolution of the twin model, which surpasses the static update mechanism.

[0020] (8) Integrate carbon emission and energy consumption dynamic assessment modules into the twin system, introduce ecological benefit indicators into the entire road construction process, and form a closed loop of intelligent assessment throughout the entire life cycle.

[0021] (9) Establish a virtual-real linkage construction progress self-verification and abnormal closed-loop control mechanism to support the system to automatically detect deviations and trigger adjustments, and realize unattended intelligent management and control. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of a digital twin-based intelligent construction and collaborative management method for municipal roads according to the present invention. Detailed Implementation

[0023] For examples, please refer to Figure 1 A digital twin-based intelligent construction and collaborative management method for municipal roads includes the following steps: S1. Establish a multi-source heterogeneous data twin modeling system for municipal road engineering, and construct a unified data twin composed of BIM model, construction monitoring data, environmental data and equipment data; during the modeling process, the unified data twin realizes dynamic mapping of structural components, construction stages and real-time status through semantic layering, attribute encoding and multi-dimensional coordinate unification, forming a high-fidelity virtual-real integrated model that can be directly called by the full-cycle management system; S2. Based on time-series mapping, a twin synchronization mechanism is used to achieve bidirectional dynamic matching and state synchronization between the virtual model and the actual entity. Through timestamp alignment, data delay compensation and asynchronous stream merging algorithms, the data update delay between the virtual and the real is less than 1 second, ensuring the timeliness consistency between the twin model and the actual project in the three-dimensional space of progress, location and state. S3. Employ a multi-agent collaborative optimization algorithm to construct a full-cycle intelligent collaborative control module covering design, construction, testing, and operation and maintenance. Each agent participates in task allocation, resource scheduling, and anomaly detection as a role-based digital node, and continuously optimizes construction decisions through a reinforcement learning self-evolution mechanism. S4. Embed a risk evolution prediction model in the digital twin to conduct advance assessment of road structure deformation, material performance degradation and environmental load response; utilize the coupled analysis of historical data and real-time twin status to achieve dynamic identification and trend prediction of structural deterioration and safety risks. S5. An adaptive task allocation mechanism based on scene layering coordinates intelligent operation tasks of multiple trades and equipment; through scene partitioning and mapping, equipment capacity assessment and resource load balancing, it generates the optimal task path and updates the construction progress in real time. S6. Through two-way communication between the digital twin at the construction site and the remote decision-making platform, the system implements self-verification of construction progress and closed-loop control of abnormal states; the system uses a difference detection algorithm to determine the deviation of the twin state and triggers automatic diagnosis and adjustment strategies to achieve dynamic correction. S7. A multi-stage data fusion and reconstruction method is adopted to perform consistency correction and structural self-updating on the data collected in the twin model; a hybrid modeling method of graph convolution and autoregression is used to achieve semantic unification and dynamic reconstruction between different stages and different data sources. S8. Based on the operational status of the digital twin, establish a dynamic assessment system for carbon emissions and energy consumption throughout the entire life cycle of roads; through the coupled calculation of equipment energy consumption records, material carbon coefficients and construction plans, achieve refined quantification of carbon emissions during the construction and operation periods; S9. Output a comprehensive decision-making report on intelligent construction and collaborative management of municipal roads, realizing an intelligent control closed loop integrating planning, construction and operation and maintenance.

[0024] Specifically, a closed-loop data interaction chain is formed between the S2 twin synchronization mechanism and the S4 risk evolution prediction model. In the S2 stage, the digital twin synchronizes construction status parameters in real time, including structural displacement, stress-strain, temperature field, and equipment load signals. This data stream is preprocessed by edge computing nodes and then input into the S4 risk evolution prediction model. The risk model uses a graph attention network to aggregate the weights of multi-dimensional features, generating a risk level matrix and a trend change vector. This vector is smoothed by a time sliding window function and returned to the S2 twin synchronization module as a dynamic correction signal to adjust on-site construction parameters, including formwork support force, pouring rate, and equipment operation sequence. When the system detects that the risk level has risen above a threshold, it automatically triggers the construction parameter reconfiguration module to re-optimize the progress planning of the corresponding construction area, achieving a dynamic safety closed loop driven by both virtual and real factors.

[0025] Specifically, to achieve synchronization performance with a data latency of less than 1 second, the system deploys no fewer than 12 edge computing nodes at the construction site. Each node has a service radius of no more than 80 meters, and the nodes form a low-latency communication link with the main control cloud server through a 5G industrial private network. The data sampling frequency range is 50 Hz–200 Hz, adaptively scheduled according to equipment type; highly dynamic equipment (such as pavers and hoisting equipment) maintains a sampling rate of 200 Hz, while static environmental sensors maintain a sampling rate of 50 Hz. Each edge node embeds a timestamp synchronization module and a buffer control chip, ensuring that the data latency compensation error is less than 0.2 seconds through a time alignment protocol (PTP). The main control node uses an FPGA acceleration array to execute an asynchronous stream merging algorithm, ensuring that the total latency of the entire system in the virtual and real data transmission, parsing, and feedback stages does not exceed 1 second.

[0026] Specifically, the S8 carbon emission and energy consumption dynamic assessment system adopts a carbon emission calculation method based on a multi-source regression weighted model. The system collects real-time data on construction equipment operating power, fuel consumption, and working time, combining this data with material usage and transportation route data to form an energy consumption record matrix. Material carbon coefficients are automatically calculated based on the ISO 14067 standard and local building material carbon factor databases, while equipment carbon emission coefficients are obtained through a measured correction method. The assessment model uses the following calculation expression:

[0027] in For equipment power, For efficiency, As a carbon emission factor, For material usage, The carbon coefficient of the material.

[0028] The assessment accuracy threshold is set at ±2%. When the real-time carbon emission assessment result deviates from the prediction model by more than the threshold, the system automatically executes an energy consumption recorrection mechanism to dynamically optimize equipment operation strategies and material delivery routes. This system can simultaneously output carbon emissions, carbon intensity, and carbon balance trend curves for both the construction and operation phases, achieving full-process carbon footprint visualization and energy efficiency optimization control.

[0029] The S1 multi-source heterogeneous data twin modeling system includes the following steps: S1-1. Perform multi-level semantic decomposition on the BIM model to extract structural components, material types, node connections, and construction stage attributes; during the decomposition process, hierarchical organization is carried out based on three-dimensional geometric topology to ensure that the contextual association and spatial logic of each component are consistent. S1-2. Reconstruct the spatial-temporal coordinates of the construction monitoring data, perform coordinate calibration using multi-node sensor array data, eliminate dynamic measurement noise through Kalman filtering, and form a traceable dynamic attribute matrix. S1-3. Encode environmental data and equipment data with features, and perform dual alignment mapping based on timestamps, geographic coordinates and BIM semantic nodes to keep equipment status, construction climate conditions and material performance data synchronized in real time in a unified twin data body; S1-4. A unified twin data volume is constructed through a high-dimensional feature fusion algorithm, and a feature mutual information entropy optimization algorithm is used to reduce redundancy, so as to realize adaptive interoperability of multi-source heterogeneous data under a unified coordinate, time and semantic level.

[0030] Specifically, the inputs to the "high-dimensional feature fusion algorithm" in S1-4 are BIM semantic node feature vectors, time-series tensors of construction monitoring data, environmental parameter matrices, and equipment operating status matrices. Each type of input feature is first mapped to a unified 512-dimensional feature space through a linear transformation to ensure dimensional consistency and comparability of multi-source data. After mapping, a multi-head attention mechanism is used to perform cross-domain feature fusion, generating a high-dimensional fusion tensor. The output is the main feature matrix of the unified twin data volume, with a dimension of 512×N (where N is the number of data nodes).

[0031] In the mutual information entropy optimization stage, features with redundant correlation exceeding 0.85 are identified by calculating the mutual information entropy difference between each feature vector, and dimensionality reduction is performed. The target dimension for dimensionality reduction is fixed at 256 dimensions to ensure the repeatability of the optimization process. The algorithm is trained for 500 iterations with a fixed learning rate of 0.0005. The optimization objective function adopts a weighted combination of maximizing mutual information and minimizing reconstruction error, with a weight ratio of 0.6:0.4.

[0032] S1-1 multi-level semantic decomposition also includes the following steps: S1-1-1. A structural graph convolutional neural network is used to extract features of the topological relationships between BIM components and to establish a node-edge weight graph to identify key connection parts and highly sensitive structural units. Specifically, the Graph Convolutional Neural Network (GCN) model in S1-1-1 consists of 4 convolutional layers and 1 fully connected layer, with 64, 128, 128, and 256 convolutional kernel nodes respectively, and ReLU activation function. The learning rate is set to 0.0001-0.001, and the Adam optimizer is used for parameter updates, with a batch size of 64 and 800 iterations. The model input samples come from a joint dataset of measured data and simulated component samples from municipal road BIM, with measured samples accounting for 70% and simulated samples accounting for 30%. The training data covers typical structural units such as bridge decks, roadbeds, drainage ditches, and guardrails, with a total of approximately 8,000 samples.

[0033] S1-1-2. Perform dynamic weight allocation according to node importance, and establish a reconstructable semantic hierarchical network by using the PageRank-based hierarchical importance ranking method to achieve reversible mapping of information at different hierarchical levels; Specifically, the node importance ranking module in S1-1-2 is trained based on the node feature embedding results extracted by the aforementioned GCN, using PageRank parameter α=0.85, 100 iterations, and node connection weights as the initial distribution vector.

[0034] S1-1-3. Redundant nodes are removed and information is aggregated in the semantic hierarchical network, with the removal rate controlled at 20%-30%, forming a lightweight and interactive twin model.

[0035] Specifically, the information aggregation and redundancy removal model in S1-1-3 adopts a joint strategy of layer normalization and batch normalization, with the removal threshold fixed at nodes whose information entropy is less than 0.3. The redundancy removal model has been validated with 1000 sets of BIM samples, with an average information retention rate of 91.7%, ensuring the semantic integrity and topological consistency of the lightweight base model.

[0036] The S3 multi-agent collaborative optimization algorithm is built on a dynamic weighted game mechanism, where each agent represents a design, construction, inspection, or operation and maintenance role. Each agent calculates behavioral conflicts and concession conditions through a task payoff matrix and selects the optimal strategy through game learning in a real-time simulation space. When the payoff convergence rate is lower than a set threshold, the system automatically triggers a reinforcement learning backoff mechanism to redistribute the weight parameters, achieving coordinated optimization and dynamic balance of multiple roles throughout the entire lifecycle.

[0037] Specifically, the communication between the agents adopts a hybrid communication structure based on message queues, shared memory, and event-driven mechanisms. Message queues are used for transmitting asynchronous task instructions and status signals, shared memory is used for fast reading and writing of high-frequency numerical parameters, and the event-driven mechanism is used to trigger policy updates and conflict responses, enabling synchronous feedback of decisions made by design, construction, testing, and maintenance roles with millisecond-level latency.

[0038] In constructing the payoff matrix, a multidimensional weighted payoff model is used, and the payoff parameters include: (1) Task risk weight: Automatically generated based on the safety level of the construction area, environmental complexity, and operational hazard coefficient; (2) Resource consumption cost: calculated by weighting energy consumption, material loss, human input and machinery usage time; (3) Schedule penalty coefficient: The priority of tasks is dynamically adjusted based on the plan deviation rate; (4) Quality and safety constraints: impose penalty factors on low-quality or high-risk behaviors.

[0039] The payoff matrix is ​​updated in real time through a rolling time window, providing continuous feedback for game decision-making.

[0040] Furthermore, to address algorithm stability, a trigger condition for the reinforcement learning backoff mechanism is defined: when the return convergence rate falls below a set threshold (between 0.01 and 0.05) for three consecutive training epochs, the system automatically enters a backoff state, reinitializing the policy function and weight matrix and adjusting the learning rate, thereby ensuring the verifiability and convergence stability of the model training.

[0041] The S3 multi-agent cooperative optimization algorithm includes the following steps: S3-1. Construct the role behavior state space and benefit matrix, and define the task inputs, outputs and constraints for each role; S3-2. Calculate task priority based on time decay function, and assign dynamic time weights to construction progress, equipment utilization and risk level; S3-3. Employ reinforcement learning strategies to iteratively optimize the collaboration paths of each role, and continuously refine the policy function based on feedback from historical task data; S3-4. Based on the real-time task completion rate, the global state is fed back, and the system automatically updates the reward signal, ultimately achieving global collaborative convergence of multiple agents under multi-dimensional constraints.

[0042] Specifically, the benefit matrix constructed in section 3-1 sets initial parameters based on the task characteristics of different roles. The design agent focuses on the feasibility and safety of the solution, the construction agent focuses on schedule and resource optimization, the detection agent focuses on quality and deviation correction, and the operation and maintenance agent focuses on long-term stability and cost control. The matrix parameters are jointly calibrated using historical engineering data and an expert experience database.

[0043] In S3-2, the time decay function adopts an exponential decay model. Where k is the dynamic task adjustment coefficient. The task weight at time t is used to calculate the priority of the current task in the agent's reward matrix; t represents the initial weight of the task, i.e., the baseline importance of the task before execution begins; t represents the elapsed time of the task, usually in seconds, minutes, or hours, calculated based on the system sampling period; e represents the base of the natural logarithm, a mathematical constant, approximately equal to 2.71828, used for exponential decay calculations.

[0044] In S3-3, the reinforcement learning module performs policy gradient updates after each round of the game and backpropagates the payoff difference from the previous round. When the average payoff gain is lower than the 0.02 threshold, a backoff mechanism is automatically triggered to perform a secondary policy search to prevent local optima traps.

[0045] In S3-4, the global state feedback process is implemented through message queue broadcasting, and the reward signal update adopts the time window-based moving average method to ensure convergence smoothness in a multi-role, multi-constraint environment, ultimately achieving global optimal collaborative control in a dynamic digital twin environment.

[0046] The S4 risk evolution prediction model adopts a multi-layer prediction structure based on spatiotemporal convolution and graph attention mechanisms, and constructs a risk propagation path map by integrating material aging data, structural strain data and meteorological load parameters. The risk evolution prediction model spatially clusters and identifies anomalies in the early stages of structural state changes, and combines a time window sliding prediction algorithm to form a dynamically adjustable risk response range, thereby achieving early warning of structural damage and adaptive correction of construction parameters.

[0047] Specifically, the risk propagation path diagram consists of approximately 50 to 500 nodes, each corresponding to a unit or component of the road structure. Node attributes include: material type (such as asphalt, concrete), historical and real-time strain rate, stress load, temperature and humidity conditions, and construction stage information; the number of edges between nodes is approximately 1.5 to 3 times the number of nodes, and edge attributes include component connection relationships, load transfer paths, and stress correlation coefficients between adjacent units. The model input is the real-time feature vector of each node in the unified data twin (256 dimensions, including material, strain, load, environmental parameters, etc.), and the output is the risk value of the corresponding node in the future prediction time window (a single scalar or 5-level risk score). The time window sliding prediction algorithm has a window length of 10 minutes and an update frequency of once per minute. It performs time-series prediction of node risk through continuous sliding windows, thereby achieving continuous dynamic updates of risk values. Risk warning output is presented in the form of levels. Each node is assigned a risk level of 1 to 5, or a corresponding red, yellow and green status classification. The system can trigger visual warnings based on thresholds and feed back high-risk nodes to the unified data twin of S1-4 and the multi-agent collaborative optimization module of S3 to adjust construction parameters and resource scheduling strategies in reverse, so as to achieve closed-loop risk control.

[0048] The S5 scene-layered adaptive task allocation mechanism includes the following steps: S5-1. Divide the construction site into foundation operation layer, structural layer and equipment layer, and construct a hierarchical topology diagram based on the site spatial layout; Specifically, the hierarchical topology diagram is generated through node division rules: each node corresponds to a spatial unit or work unit of the construction scenario. The node division is based on spatial coordinate thresholds (such as components or work points within a range of 10 to 50 meters being grouped into the same node) and the construction process sequence (such as structural layer work can only begin after the foundation is completed). Edges between nodes are defined as task dependencies or physical spatial proximity, forming a multi-level hierarchical structure to ensure that the task logic sequence and spatial mapping between the foundation layer, structural layer, and equipment layer are consistent.

[0049] S5-2. Perform task partitioning and coding for each layer, including task number, resource requirements, execution time limit and constraints. Specifically, each task node code further includes topological relationship attributes, such as the preceding task node number, the spatial coordinates of adjacent nodes, and the construction stage identifier, for use in subsequent task coupling calculations.

[0050] S5-3. Calculate the task coupling degree based on the multi-layer task status monitoring matrix, and generate a coupling weight matrix according to task dependencies, equipment capabilities and time overlap. S5-4. Based on the results of task coupling, perform adaptive task matching and resource scheduling, and achieve optimal resource allocation for multi-task parallel operations by minimizing the weighted load cost function.

[0051] The S5-3 task coupling calculation also includes the following steps: S5-3-1. Extract the spatiotemporal correlation features between task nodes through dual-channel tensor decomposition, and construct the task interaction matrix in the dual domains of the task time axis and spatial coordinate axis; Specifically, the dual-channel tensor decomposition implementation parameters are as follows: the task interaction matrix tensor order is set to 3 (task × time × space), each channel has 256 dimensions, the number of iterations is 50~100, and the convergence threshold is 1e-4; during the tensor decomposition process, the temporal correlation and spatial proximity features of the task are extracted by alternating least squares method to achieve dimensionality reduction of multi-dimensional task information and extraction of associated features.

[0052] S5-3-2. A nonlinear synergy factor model is used to calculate the dynamic mutual influence between tasks, and the synchronicity of tasks is quantified by high-order synergy coefficients. Specifically, the nonlinear synergy factor model takes the decomposed node feature vectors as input and outputs a dynamic influence weight matrix between tasks. It quantifies task synchronicity through higher-order synergy coefficients to assess potential conflicts or collaborative relationships between tasks, thus aiding in the generation of a coupling weight matrix. S5-3-3. Construct a task collaboration network graph, cluster highly coupled tasks, and generate a synchronization scheduling benchmark based on the similarity matrix to achieve dynamic decoupling of multiple tasks and global progress optimization.

[0053] Specifically, after the task collaboration network graph is constructed, the central node of the highly coupled task group is determined by clustering algorithm (such as K-means or spectral clustering), a synchronization scheduling benchmark is generated based on the similarity matrix, and the optimal execution order of multiple tasks is calculated by combining the shortest path scheduling algorithm to achieve global progress optimization. This enables multi-type and multi-node tasks to run in parallel efficiently in time and space, minimizing resource conflicts and construction delays.

[0054] The S7 multi-stage data fusion and reconstruction method uses an autoregressive-graph convolution hybrid model to perform dynamic consistency correction. Within the data update cycle, the adaptive structural reconstruction of the twin model is achieved through the time difference integral strategy and the dynamic balancing algorithm of node reconstruction weights, maintaining the topological consistency and data fidelity of the twin in the multi-stage information evolution.

[0055] Specifically, the autoregressive-graph convolutional hybrid model (AR-GCN hybrid model) is configured as follows: the autoregressive (AR) layer is set to order 2-3 to capture the dynamic changes in continuous time steps; the graph convolutional layer is set to 3-4 layers, with 64-128 convolutional kernels per layer and a stride of 1, to extract the spatial dependencies between nodes on the Siamese model topology graph; each layer uses the ReLU activation function, and residual connections are used to improve the stability of multi-layer feature fusion.

[0056] In the time difference integration strategy, the data sampling period is set to Δt=0.5 seconds. The feature value of each node is dynamically balanced by integrating the difference between the current sample and the previous sample, combined with the node importance weight. The node weight dynamic balancing algorithm achieves adaptive allocation of information contribution between nodes through normalization processing and gradient adjustment, ensuring that the twin maintains structural consistency and data fidelity in multi-stage information updates.

[0057] The model validation process includes training and testing using historical construction monitoring datasets (containing construction progress, structural strain, and equipment status data) to evaluate the model's prediction accuracy. The predicted output is compared with the actual observation data to calculate metrics including root mean square error (RMSE) and mean absolute error (MAE) to quantify the accuracy and reliability of multi-stage reconstruction and to use for iterative optimization of model parameters.

[0058] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent construction and collaborative management of municipal roads based on digital twins, characterized in that, The aforementioned intelligent construction and collaborative management method for municipal roads based on digital twins includes the following steps: S1. Establish a multi-source heterogeneous data twin modeling system for municipal road engineering, and construct a unified data twin composed of BIM model, construction monitoring data, environmental data and equipment data; the unified data twin achieves dynamic mapping of structural components, construction stages and real-time status through semantic layering, attribute encoding and multi-dimensional coordinate unification during the modeling process, forming a high-fidelity virtual-real integrated model that can be directly called by the full-cycle management system. S2. Based on time-series mapping, a twin synchronization mechanism is used to achieve bidirectional dynamic matching and state synchronization between the virtual model and the actual entity. Through timestamp alignment, data delay compensation and asynchronous stream merging algorithms, the data update delay between the virtual and the real is less than 1 second, ensuring the timeliness consistency between the twin model and the actual project in the three-dimensional space of progress, location and state. S3. Employ a multi-agent collaborative optimization algorithm to construct a full-cycle intelligent collaborative control module covering design, construction, testing, and operation and maintenance. Each agent participates in task allocation, resource scheduling, and anomaly detection as a role-based digital node, and continuously optimizes construction decisions through a reinforcement learning self-evolution mechanism. S4. Embed a risk evolution prediction model in the digital twin to conduct advance assessment of road structure deformation, material performance degradation and environmental load response; utilize the coupled analysis of historical data and real-time twin status to achieve dynamic identification and trend prediction of structural deterioration and safety risks. S5. An adaptive task allocation mechanism based on scene layering coordinates intelligent operation tasks of multiple trades and equipment; through scene partitioning and mapping, equipment capacity assessment and resource load balancing, it generates the optimal task path and updates the construction progress in real time. The S5 scene-layered adaptive task allocation mechanism includes the following steps: S5-1. Divide the construction site into foundation operation layer, structural layer and equipment layer, and construct a hierarchical topology diagram based on the site spatial layout; S5-2. Perform task partitioning and coding for each layer, including task number, resource requirements, execution time limit and constraints. S5-3. Calculate the task coupling degree based on the multi-layer task status monitoring matrix, and generate a coupling weight matrix according to task dependencies, equipment capabilities and time overlap. The S5-3 task coupling degree calculation also includes the following steps: S5-3-1. Extract the spatiotemporal correlation features between task nodes through dual-channel tensor decomposition, and construct the task interaction matrix in the dual domains of the task time axis and spatial coordinate axis; S5-3-2. A nonlinear synergy factor model is used to calculate the dynamic mutual influence between tasks, and the synchronicity of tasks is quantified by high-order synergy coefficients. S5-3-3. Construct a task collaboration network graph, cluster highly coupled tasks, and generate a synchronization scheduling benchmark based on the similarity matrix to achieve dynamic decoupling of multiple tasks and global progress optimization. S5-4. Based on the results of the task coupling degree, perform adaptive task matching and resource scheduling, and achieve optimal resource allocation for multi-task parallel operations by minimizing the weighted load cost function; S6. Through two-way communication between the digital twin at the construction site and the remote decision-making platform, the system implements self-verification of construction progress and closed-loop control of abnormal states; the system uses a difference detection algorithm to determine the deviation of the twin state and triggers automatic diagnosis and adjustment strategies to achieve dynamic correction. S7. A multi-stage data fusion and reconstruction method is adopted to perform consistency correction and structural self-updating on the data collected in the twin model; a hybrid modeling method of graph convolution and autoregression is used to achieve semantic unification and dynamic reconstruction between different stages and different data sources. S8. Based on the operating status of the digital twin, establish a dynamic assessment system for carbon emissions and energy consumption throughout the entire life cycle of the road; through the coupled calculation of equipment energy consumption records, material carbon coefficients and construction plans, achieve refined quantification of carbon emissions during the construction and operation periods; S9. Output a comprehensive decision-making report on intelligent construction and collaborative management of municipal roads, realizing an intelligent control closed loop integrating planning, construction and operation and maintenance.

2. The method for intelligent construction and collaborative management of municipal roads based on digital twins according to claim 1, characterized in that, The S1 multi-source heterogeneous data twin modeling system includes the following steps: S1-1. Perform multi-level semantic decomposition on the BIM model to extract structural components, material types, node connections, and construction stage attributes; during the decomposition process, hierarchical organization is carried out based on three-dimensional geometric topology to ensure that the contextual association and spatial logic of each component are consistent. S1-2. Reconstruct the spatial-temporal coordinates of the construction monitoring data, perform coordinate calibration using multi-node sensor array data, eliminate dynamic measurement noise through Kalman filtering, and form a traceable dynamic attribute matrix; S1-3. The environmental data and equipment data are feature-encoded, and double-aligned and mapped according to timestamps, geographic coordinates and BIM semantic nodes, so that the equipment status, construction climate conditions and material performance data are kept synchronized in real time in the unified data twin. S1-4. The unified data twin is constructed by a high-dimensional feature fusion algorithm, and the feature mutual information entropy optimization algorithm is used to reduce redundancy, so as to realize the adaptive interoperability of multi-source heterogeneous data under the unified coordinate, time and semantic level.

3. The method for intelligent construction and collaborative management of municipal roads based on digital twins according to claim 2, characterized in that, The S1-1 multi-layer semantic decomposition further includes the following steps: S1-1-1. A structural graph convolutional neural network is used to extract features of the topological relationships between BIM components and to establish a node-edge weight graph to identify key connection parts and highly sensitive structural units. S1-1-2. Perform dynamic weight allocation according to node importance, and establish a reconstructable semantic hierarchical network by using the PageRank-based hierarchical importance ranking method to achieve reversible mapping of information at different hierarchical levels; S1-1-3. Redundant nodes are removed and information is aggregated in the semantic hierarchical network, with the removal rate controlled at 20%-30%, to form a lightweight interactive twin model.

4. The method for intelligent construction and collaborative management of municipal roads based on digital twins according to claim 1, characterized in that, The S3 multi-agent collaborative optimization algorithm is constructed based on a dynamic weighted game mechanism, where each agent represents a design, construction, inspection, or operation and maintenance role. Each agent calculates behavioral conflicts and concession conditions through a task payoff matrix and selects the optimal strategy through game learning in a real-time simulation space. When the payoff convergence rate is lower than a set threshold, the system automatically triggers a reinforcement learning backoff mechanism to redistribute the weight parameters, thereby achieving coordinated optimization and dynamic balance of multiple roles throughout the entire cycle.

5. A method for intelligent construction and collaborative management of municipal roads based on digital twins according to claim 4, characterized in that, The S3 multi-agent cooperative optimization algorithm includes the following steps: S3-1. Construct the role behavior state space and benefit matrix, and define the task inputs, outputs and constraints for each role; S3-2. Calculate task priority based on time decay function, and assign dynamic time weights to construction progress, equipment utilization and risk level; S3-3. Employ reinforcement learning strategies to iteratively optimize the collaborative paths of each role, and continuously refine the policy function based on feedback from historical task data; S3-4. Based on the real-time task completion rate, the global state is fed back, and the system automatically updates the reward signal, ultimately achieving global collaborative convergence of multiple agents under multi-dimensional constraints.

6. The method for intelligent construction and collaborative management of municipal roads based on digital twins according to claim 1, characterized in that, The S4 risk evolution prediction model adopts a multi-layer prediction structure based on spatiotemporal convolution and graph attention mechanisms, and constructs a risk propagation path map by fusing material aging data, structural strain data and meteorological load parameters. The risk evolution prediction model identifies anomalies through spatial clustering in the early stages of structural state changes, and combines a time window sliding prediction algorithm to form a dynamically adjustable risk response range, thereby achieving early warning of structural damage and adaptive correction of construction parameters.

7. The method for intelligent construction and collaborative management of municipal roads based on digital twins according to claim 1, characterized in that, The S7 multi-stage data fusion and reconstruction method uses an autoregressive-graph convolution hybrid model to perform dynamic consistency correction. Within the data update cycle, the adaptive structural reconstruction of the twin model is achieved through the time difference integral strategy and the dynamic balancing algorithm of node reconstruction weights, maintaining the topological consistency and data fidelity of the twin in the multi-stage information evolution.

Images