Intelligent scheduling and collaboration system and method for city reconstruction full life cycle

Abstract

The invention belongs to the technical field of smart city construction and intelligent construction management, and particularly relates to an intelligent scheduling and cooperation system and method for the whole life cycle of city reconstruction. The system comprises an edge sensing layer, a digital twinning network module, a space-time brain module, a block chain collaborative trust layer and an XR collaborative interaction layer. The system collects multi-source data of a construction site in real time through a 5G edge device, realizes fusion modeling of BIM, GIS and TIN models by using a graph convolutional neural network, and performs multi-target scheduling optimization in combination with meta reinforcement learning and a dynamic graph neural network; and meanwhile, construction task verification and fund payment linkage is realized through a block chain smart contract, and immersive cooperation and visual acceptance are provided in cooperation with an AR / VR platform. According to the method, the problems of model splitting, scheduling response lagging, low data collaboration efficiency, insufficient credibility of the performance process and the like in the urban reconstruction project are solved, and the intelligence, transparency and automation level of the construction process is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of smart city construction and intelligent construction management, and in particular to an intelligent scheduling and coordination system and method for the entire life cycle of urban reconstruction. Background Art

[0002] With the rapid advancement of urban renewal and infrastructure renovation projects, traditional construction management models face a series of challenges, including fragmented multi-source data, delayed decision-making responses, and difficulty verifying execution processes. To achieve efficient operation of urban reconstruction projects under complex working conditions, smart construction and digital supervision methods have gradually become a development trend. In existing technologies, although BIM (Building Information Modeling), GIS (Geographic Information System), and IoT sensing devices have been introduced, and a digital management framework for the entire construction process has been initially established, there are still obvious deficiencies in the following aspects:

[0003] 1. Limited model fusion capabilities: Currently, BIM, GIS, and TIN models are maintained independently, lacking effective topological fusion and semantic alignment mechanisms. This makes it difficult to support cross-system task collaboration within a unified space. Model boundary data conflicts frequently occur, especially during the registration of 3D map modeling and design drawings in older urban areas. Fusion errors are particularly prominent, resulting in mismatched construction plans and wasted material resources.

[0004] 2. Slow scheduling response and lack of intelligence: Existing construction scheduling often relies on static planning, lacking the ability to respond in real time to dynamic construction environments (such as weather changes, personnel behavior, and traffic congestion). Even when some AI models are introduced, their algorithms are generally optimized for a single task and fail to formulate strategic trade-offs across multiple objectives (such as progress, carbon emissions, and safety). This results in insufficient system robustness and an inability to adapt to complex site changes.

[0005] 3. Low transparency in data collaboration and contract fulfillment: Due to data barriers between construction stakeholders, task progress and acceptance processes rely heavily on manual confirmation and paper records, which can lead to information lags, tampering risks, and frequent disputes. While some projects have attempted to incorporate blockchain technology to upload data, most only implement "logging" functionality and lack a trusted trigger mechanism that integrates with actual project status (such as LiDAR acceptance point clouds).

[0006] 4. Inefficient human-computer interaction: Traditional visualization platforms are disconnected from terminal operations, making it difficult for workers to access information such as drawings, tasks, and risk alerts through intuitive interfaces. Some AR systems also suffer from low spatial registration accuracy and high command interaction latency, limiting their practicality in complex construction sites.

[0007] 5. Lack of support for green and low-carbon construction: Most scheduling systems fail to incorporate carbon emissions as a constraint variable in resource allocation optimization. Equipment carbon emission profiles and route planning mechanisms are lacking, hindering green construction. Furthermore, AI strategy models lack interpretability, making it difficult for construction management units to conduct causal audits and compliance tracing of scheduling results.

[0008] In view of the shortcomings of the above-mentioned existing technologies, there is an urgent need for an intelligent scheduling and coordination system for the entire life cycle of urban reconstruction. It can integrate multi-source model data to build a unified digital twin environment; it has adaptive multi-task optimization scheduling capabilities; at the same time, through a trusted execution verification mechanism and efficient human-computer interaction methods, it can significantly improve construction efficiency, management transparency and performance credibility, and support the continuous evolution of green and low-carbon goals and model migration capabilities. Summary of the Invention

[0009] In order to solve the technical bottleneck problems commonly found in existing urban reconstruction projects, such as model data fragmentation, delayed scheduling response, low collaboration efficiency, and insufficient credibility of task execution, the present invention proposes an intelligent scheduling and collaboration system and method that integrates multi-source perception, digital twin modeling, graph neural network intelligent decision-making, blockchain performance verification mechanism, and XR immersive interactive collaboration. The system takes "data-driven, model unification, intelligent optimization, and trusted execution" as its core logic. By building an intelligent, modular, and traceable closed-loop system covering the entire process of urban reconstruction, it realizes real-time perception and standardized fusion expression of construction site data, supports cross-model topology unification, dynamic scheduling decisions for multi-objective tasks, and cooperates with on-chain task verification and immersive collaborative feedback mechanisms to significantly improve the safety, efficiency, transparency, and sustainability of urban infrastructure renewal. Based on this, the present invention proposes the following technical solutions.

[0010] In one possible implementation, an intelligent scheduling and coordination system for the entire life cycle of urban reconstruction is provided. The system adopts a hierarchical modular structure and includes the following functional layers:

[0011] 1. The edge perception layer, comprising a cluster of 5G IoT devices and edge computing nodes deployed at the construction site, collects 3D point clouds, equipment operating status, environmental parameters, and personnel trajectory information. The sampling frequency is no less than 10Hz, and the point cloud data is pre-processed using a Voxel Grid filter, compressing the voxel size to 0.05m. Asynchronous data transmission and timestamp alignment are achieved through the ROS2 communication framework, keeping communication latency to under 30ms. The edge nodes integrate a Kalman filter module for IMU trajectory data denoising and posture correction.

[0012] 2. The digital twin meshing module is used to integrate BIM (Building Information Model), GIS (Geographic Information System), and TIN (Triangulated Network) models to construct a unified topology. This module uses a graph convolutional neural network (GCN) to process the topology structure. The node feature dimension is 64, including spatial coordinates, semantic labels, and physical attributes. GCN adopts a three-layer network architecture. The front end enhances semantic perception capabilities through the Transformer substructure, and the back end uses a softmax weighted fusion mechanism to resolve model boundary conflicts. The model edge generation rule is based on the Euclidean distance threshold d < 0.3m. The RBF kernel bandwidth σ in the topology alignment loss is determined by grid search in the interval [0.1, 1.0] with a step size of 0.1.

[0013] 3. The spatiotemporal brain module implements multi-task scheduling decisions based on meta-reinforcement learning and a dynamic graph neural network (DGNN). The policy network is built on the RLlib training framework, with a state space dimension of 23 and a PPO algorithm as the training strategy. During the pre-training phase, historical scheduling trajectories are loaded for imitation learning, and during the fine-tuning phase, policy adaptation is achieved through online reinforcement learning. The DGNN model incorporates a graph attention mechanism (GAT) to enhance edge feature representation. During inference, an early-exit mechanism is used to return high-confidence results early, optimizing average response time to 130ms.

[0014] 4. The blockchain collaborative trust layer, based on the Hyperledger Fabric consortium chain architecture, integrates Oracle nodes to verify construction completion status. LiDAR modules capture on-site point clouds and compare them against the BIM model. When 97% of points have a matching error of less than 2.5 cm and the overall discrepancy rate is less than 3%, a trusted acceptance certificate is generated, triggering the chain code to execute the payment logic and recording the hash and GPS timestamp. To enhance trustworthiness, a dual-factor acceptance mechanism is introduced: design model hash verification and AI error prediction comparison.

[0015] 5. The XR collaborative interaction layer supports AR navigation and VR sandbox functions to create an immersive collaborative environment. Spatial registration is completed using AprilTag combined with SLAM (positioning accuracy ≤ 2cm). The VR sandbox supports multi-role collaboration and process playback. Construction drawings, risk warnings, and other information are overlaid on the user's field of view in real time via graphics and text. Command control uses a voice and gesture fusion method and is equipped with an intent recognition module based on a graph neural network, with an accuracy rate exceeding 95%. Collaborative tasks support multi-person status synchronization, with a maximum interaction delay of no more than 100ms.

[0016] In one possible implementation, the system further includes:

[0017] The carbon emission optimization module is used to establish a mapping table between construction equipment and carbon emission factors, and dynamically generate low-carbon scheduling paths based on task paths, equipment types, and environmental conditions;

[0018] The causal analysis module uses SHAP values ​​to perform causal attribution analysis on the input variables in the policy network to form a task causal graph;

[0019] The federated upgrade module uses a differential privacy mechanism to protect data privacy, enabling secure aggregation of policy parameters and personalized optimization of local models across multiple projects. The differential privacy parameter ε is controlled between 0.5 and 1.

[0020] The present invention dynamically adjusts the ε value (0.5–1.0) through scene classification and training rounds to prioritize privacy protection under high-sensitivity data and improve model effectiveness under low-sensitivity data, taking into account both security and performance.

[0021] In one possible implementation, the edge weight update of the DGNN model is controlled by the following coupling function:

[0022]

[0023] Where Wt is the edge weight matrix at time t, L is the joint loss function, α is the learning rate, β is the coupling function weight parameter, and Φ(St,Tt) is the spatiotemporal coupling factor, which is calculated as:

[0024]

[0025] Among them, S t is the current state tensor, is the predicted state, T t is the timestamp, λ is the time decay coefficient, the default value is 0.5, and it is adjusted to 0.8 when a high-risk event is detected.

[0026] In one possible implementation, the loss function for topological alignment of the GCN model is defined as:

[0027]

[0028] where f θ is the node feature extraction function, γ is the balance parameter, MMD (maximum mean difference) is used to measure the distribution consistency, and the kernel function bandwidth σ is determined by grid search in the interval [0.1, 1.0] to optimize the training accuracy.

[0029] To further ensure the reproducibility of the selection of parameter σ, the present invention divides the interval [0.1, 1.0] into 10 candidate values ​​with a step size of 0.1. For each candidate σ, the node classification accuracy is trained and evaluated on a fixed random seed and the same validation set. Finally, the σ with the highest average accuracy on the validation set is selected as the kernel bandwidth.

[0030] In a possible implementation, a method for implementing the above functions is provided, which includes the following steps:

[0031] S101: Use 5G edge devices to collect multimodal data from construction sites and perform heterogeneous encapsulation through ROS2;

[0032] S102: Use GCN to perform topological fusion of BIM, GIS, and TIN models to build a digital twin graph structure;

[0033] S103: Build a policy network based on meta-reinforcement learning, and use DGNN to complete multi-objective reasoning and policy push;

[0034] S104: Triggering LiDAR scanning and comparing it with the BIM model. When the set difference threshold is met, the blockchain contract is automatically executed.

[0035] S105: Provides visual navigation, immersive acceptance, and remote collaboration support through the XR platform;

[0036] S106: Implement strategy migration and local model update through federated learning mechanism.

[0037] In a possible implementation, a computer-readable storage medium is provided, storing program instructions. When the program is executed by a processor, the device performs the method described in each of the above steps.

[0038] Based on the above technical solutions, the intelligent scheduling and coordination system proposed in this invention for the entire life cycle of urban reconstruction effectively solves the following common problems in urban reconstruction by building a five-layer architecture (perception, modeling, decision-making, verification, and interaction) that integrates AI, BIM, GIS, blockchain, and XR technologies:

[0039] 1. Model fragmentation and difficulty in data fusion: A GCN+Transformer fusion mechanism is proposed to achieve collaborative and consistent expression of heterogeneous model space and semantics;

[0040] 2. Scheduling policy response lag: DGNN introduces graph attention mechanism and early-exit to optimize the policy reasoning process and shorten the policy response time;

[0041] 3. Lack of verifiability in acceptance and performance: Combining LiDAR point cloud comparison with a blockchain-based two-factor acceptance mechanism ensures trustworthy execution and transparent financial linkage.

[0042] 4. Inefficient human-machine collaboration: Build an immersive intelligent interactive environment by embedding semantic recognition and collaborative state synchronization in AR / VR.

[0043] 5. Weak policy transferability and interpretability: Integrate federated learning and causal analysis modules to enhance policy adaptability and AI auditability;

[0044] 6. Green and low-carbon goals are difficult to achieve: The built-in dynamic optimization module for carbon emission paths supports carbon emission control as a constraint condition in scheduling decisions.

[0045] This invention not only has a complete "perception-fusion-scheduling-execution-verification" technical closed loop, but also significantly improves the level of intelligence, credibility and low carbonization through structural optimization and algorithm nesting, and has significant engineering application value.

[0046] Experimental environment and plan

[0047] Hardware: Intel Xeon Gold 6248R @ 3.0 GHz, 64 GB RAM, NVIDIA Tesla V100 GPU

[0048] Software environment: Ubuntu 20.04 + TensorFlow 2.6 + PyTorch 1.10 + blockchain simulation platform Hyperledger Fabric 2.3

[0049] Comparison plan:

[0050] Traditional digital twin + single GCNN inference architecture (Baseline)

[0051] The "space-time brain" multi-model weighted fusion + edge reasoning + blockchain parallel verification in this invention (Proposed)

[0052] Table 1: Comparison of key performance indicators

[0053]

[0054]

[0055] Result Analysis

[0056] 1. Inference Performance

[0057] After deploying lightweight model parallel inference on edge nodes, the average latency dropped from the baseline of 200ms to 120ms, a 40% reduction, significantly improving real-time performance and meeting the strict low-latency requirements of large-scale urban transformation decisions.

[0058] The peak throughput increased from 50 times / second to 82 times / second, a 64% increase. In high-concurrency scenarios, it can support more parallel decision requests, ensuring that the system continues to operate efficiently in complex environments.

[0059] 2. Prediction Accuracy

[0060] By weightedly fusing the graph convolutional network (GCN) with the dynamic graph neural network (DGNN) and introducing multi-source data, the model's event prediction accuracy was improved from 86.3% to 93.2%, an increase of 6.9%, effectively reducing missed alerts and false alarms and enhancing the reliability of the system.

[0061] 3. Energy consumption performance

[0062] With the help of model pruning and edge inference optimization, the average power consumption of a single node dropped from 142W to 96W, a reduction of about 32.4%, significantly reducing the energy burden and providing lower operation and maintenance costs for large-scale deployment.

[0063] 4. Trusted Verification Efficiency

[0064] After adopting parallel packaging and lightweight consensus algorithms, the blockchain transaction confirmation delay was reduced from 110ms to 65ms, a decrease of about 40.9%. While ensuring that data cannot be tampered with and is traceable, it no longer becomes a system bottleneck.

[0065] 5. End-to-end responsive experience

[0066] The overall end-to-end latency from on-site perception to XR interactive feedback has been reduced from 410ms to 250ms, a reduction of 39%. Users have almost no perceived lag in their operations in the AR / VR interface, greatly improving the interaction smoothness and user experience.

[0067] In summary, through technical optimization in multi-model fusion, collaborative scheduling, model pruning and parallel verification, this system has achieved significant improvements in multiple dimensions such as latency, throughput, accuracy, energy consumption and reliability, and has effectively met the stringent performance requirements of intelligent scheduling throughout the entire life cycle of urban reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solution of the present invention, the following further details the specific embodiments of the present invention in conjunction with the accompanying drawings. The drawings involved in the present invention include but are not limited to:

[0069] Figure 1 System overall architecture diagram

[0070] The diagram illustrates the overall hierarchical structure of the intelligent scheduling and collaboration system for the entire life cycle of urban reconstruction in the present invention, including the edge perception layer, digital twin network module, spatiotemporal brain module, blockchain collaborative trust layer and XR collaborative interaction layer. Each layer in turn constitutes a closed-loop system architecture from perception acquisition, model fusion, intelligent reasoning, execution verification to immersive collaboration.

[0071] Figure 2 Data collection and preprocessing flow chart

[0072] The diagram illustrates the data sources and preprocessing process collected by this system in the edge perception stage, including input sources such as BIM / GIS models, sensor arrays, historical work orders, and meteorological APIs. After data cleaning, feature extraction, and format conversion, it is used for subsequent digital twin modeling.

[0073] Figure 3 Digital twin mesh module structure diagram

[0074] This diagram illustrates the process by which BIM / GIS input data, after undergoing topology construction and attribute mapping, enters a three-layer graph convolutional network (GCN) for structural and semantic fusion. Each GCN layer sequentially extracts node spatial features, semantic labels, and physical attributes, ultimately outputting a unified topological structure.

[0075] Figure 4 Edge perception layer structure diagram

[0076] This diagram illustrates the perception and packaging process of multimodal data at a construction site. This includes point cloud, meteorological, and personnel trajectory data collected by the LiDAR module, environmental sensor module, and worker location tags. The data is packaged using ROS2 and denoised using a Kalman filter. It is then processed in a Jetson edge computing node and uploaded to the dispatch center.

[0077] Figure 5 Spatiotemporal brain module AI reasoning structure diagram

[0078] This demonstration demonstrates a meta-policy network built using RLlib as a training platform, employing the Propagation of Positive Opportunities (PPO) algorithm to train a multi-task scheduling model, and deploying it within a TensorRT-accelerated dynamic graph neural network (DGNN). The diagram includes graph structure propagation, attention mechanisms, an early-exit decision mechanism, and response control modules, ensuring inference latency of less than 200 milliseconds.

[0079] Figure 6 Blockchain collaborative trust mechanism diagram

[0080] After the construction task is completed, point cloud data is obtained through LiDAR scanning and compared with the ICP difference rate of the BIM model. After verification by the Oracle node, a trusted acceptance certificate is generated, which automatically triggers the smart contract to execute payment and record the task hash and timestamp. Finally, the result is stored on the chain to realize the acceptance-payment linkage closed loop.

[0081] Figure 7 XR collaborative interaction flow chart

[0082] The demonstration shows workers wearing AR devices entering the construction site, completing space registration through AprilTag and SLAM, and overlaying construction drawings and risk warnings on the AR interface. The system uses graph neural networks to recognize gestures and voice commands to implement operational intention analysis. Task feedback is recorded via the Fabric chain, and the VR sandbox supports multi-role collaboration and process playback. DETAILED DESCRIPTION

[0083] To better understand the technical solutions of this invention, the following detailed description of the intelligent scheduling and coordination system for the entire lifecycle of urban reconstruction, combined with specific implementation scenarios, is provided. This invention can be applied to construction scheduling and management scenarios in a variety of complex environments, such as urban rail transit reconstruction, municipal road reconstruction, and the demolition and renovation of older buildings. Through a five-stage technical process encompassing perception, modeling, reasoning, verification, and coordination, it forms a closed-loop intelligent management system.

[0084] In specific applications, such as Figure 1 As shown, the following technical solutions can be adopted, including an intelligent scheduling and coordination system for the entire life cycle of urban reconstruction composed of the following modules: an edge perception layer, including a 5G IoT device cluster, used to collect three-dimensional point clouds, environmental data and personnel trajectories at the construction site, and realize multi-source heterogeneous data alignment through ROS2; a digital twin weaving module, which uses a graph convolutional neural network (GCN) to topologically align BIM, GIS and TIN models to build a dynamic multi-scale digital twin; a spatiotemporal brain module, based on a meta-reinforcement learning framework, which executes multi-task decisions on construction scheduling, traffic diversion and carbon emission control; a blockchain collaborative trust layer, relying on the Hyperledger Fabric architecture, verifies the construction status through Oracle nodes and triggers smart contracts; an XR collaborative interaction layer, which supports AR navigation and VR sandbox functions, and provides immersive collaboration and visual acceptance.

[0085] The system realizes the functional coordination of each module through a layered architecture, such as Figure 4As shown, the edge perception layer collects multimodal data from the construction site in real time, synchronizes it through ROS2 encapsulation, and uses Kalman filtering to remove IMU noise data; the digital twin weaving module processes model topology information through a three-layer GCN structure, integrates spatial coordinates, semantic labels and physical properties, combines Transformer to enhance semantic perception, and adopts a softmax weighting mechanism to resolve model boundary conflicts; the spatiotemporal brain module trains the PPO policy network based on RLlib, combines it with the dynamic graph neural network (DGNN) for multi-target scheduling reasoning, and optimizes the reasoning delay to 130ms; the blockchain collaborative trust layer integrates the LiDAR point cloud comparison mechanism, and the chain code triggers payment and records the hash and timestamp when the error meets the threshold; the XR collaborative interaction layer realizes spatial registration through AprilTag and SLAM, and the command interaction is based on the intention recognition module of the graph neural network to support multi-user collaboration.

[0086] The system has achieved technological breakthroughs in many aspects: First, the GCN+Transformer structure significantly improves the fusion accuracy and topological consistency of BIM, GIS and TIN models, solving the problem of model fragmentation; second, DGNN and early-exit mechanism accelerate scheduling response and effectively cope with complex dynamic changes on site; third, combining the dual-factor acceptance of blockchain and LiDAR point cloud to build a trusted execution verification mechanism to improve the transparency of the performance process; fourth, the XR interactive platform that integrates AR / VR and semantic recognition enhances the efficiency of human-computer collaboration; in addition, the system also has the advantages of green and low-carbon scheduling capabilities and the portability and explainability of AI strategies, and has high engineering practicality and scalability.

[0087] In terms of material selection, edge sensing equipment can be replaced with a combination of industrial-grade cameras and lidars with dust and water resistance according to the construction environment; the GCN network structure can be replaced with a variant based on GraphSAGE to enhance model migration capabilities; the policy network algorithm can also be replaced from PPO to SAC to enhance continuous motion control performance; in addition to Hyperledger Fabric, the blockchain platform can also adopt an Ethereum-based consortium chain framework; in the XR module, the spatial registration method can use a deep learning-based point cloud matching method to improve positioning accuracy; the overall system also supports modular deployment to adapt to the flexible configuration and expansion needs in different urban reconstruction scenarios.

[0088] like Figure 2 The data collection and preprocessing flow chart of the system of the present invention is shown. This process covers multiple input channels and processing nodes, providing a unified, complete and high-quality data foundation for subsequent modeling and scheduling reasoning work.

[0089] During data collection, there are four types of data sources. The first is BIM models and GIS information, which provide basic spatial topology and structural semantics. The second is the sensor array, which includes on-site LiDAR scanning to acquire point cloud data, IMU sensors to record trajectories, and UWB technology to achieve worker positioning, enabling real-time perception of on-site conditions. The third is historical work orders, which include scheduling logs, task execution records, and operation and maintenance work orders, reflecting past work status. The fourth is the weather API, which can obtain meteorological parameters such as temperature, humidity, and wind speed.

[0090] These input data will enter the preprocessing stage and undergo operations such as cleaning, feature extraction, and format conversion in sequence. In the cleaning stage, outliers, null values, and duplicate information will be removed to ensure data quality; in the feature extraction process, the component numbers, attribute labels, etc. in the model will be associated and bound with the sensor data; format conversion is to convert data from different sources and in different forms, such as converting the original BIM data into a graph structure input format, and converting point cloud data into a sparse tensor form. At the same time, spatial data from different sources will be unified into the local coordinate system of the construction site. In addition, the entire preprocessing process is also equipped with a data cache and incremental update mechanism to support the daily incremental upload of data for continuous construction periods, thereby improving the integrity and real-time performance of the data. Finally, after processing, an input data stream with a unified structure and clear semantics is obtained, which can be directly used by subsequent graph neural network models and scheduling strategy models.

[0091] As an exemplary embodiment, Figure 3 As shown in the figure, the graph convolutional neural network (GCN) node features used in the digital twin weaving module include spatial coordinates, semantic labels and physical properties. The GCN network structure is three-layer, and a weighted fusion algorithm is used at the model boundary to handle topological conflicts.

[0092] This module constructs a unified topological graph based on the spatial and semantic differences between BIM, GIS and TIN models in urban reconstruction and construction. Each graph node carries three types of features: spatial coordinates (such as X, Y, Z coordinates), semantic labels (such as "column", "wall", "ground", etc.) and physical properties (such as material strength, thermal conductivity, etc.), and features are extracted and propagated through three layers of GCN. The first two layers complete the structured encoding of the node feature vector, and the third layer is used for semantic aggregation of the entire graph and topological alignment. In response to the topological conflicts at the junction of the models, the system introduces a weighted fusion algorithm, which calculates the fusion priority of adjacent nodes based on the softmax normalized feature similarity to form a transition structure to avoid geometric offset or semantic mismatch.

[0093] This module significantly improves the fusion quality and topological consistency between heterogeneous models, resolving frequent boundary conflicts between traditional BIM, GIS, and TIN models. It is particularly suitable for scenarios in complex 3D urban areas with numerous overlapping models. Through a weighted fusion algorithm, it preserves the semantic information of the original model while achieving spatial transitions, enhancing the coherence of the model's overall structure and providing a more accurate digital twin foundation for subsequent scheduling and visualization.

[0094] In terms of node feature composition, in addition to spatial coordinates, semantic labels, and physical attributes, timestamps or construction phase identifiers can also be introduced to support dynamically updated model version management; the GCN structure can be replaced with a Graph Attention Network (GAT) to improve the selectivity of feature aggregation; the weighted fusion algorithm can be replaced based on Euclidean distance, cosine similarity, or a fusion score based on neural network learning; in addition, the topological alignment processing flow can also be encapsulated as a modular plug-in to support docking and customized development with different modeling platforms (such as Revit, ArcGIS, etc.).

[0095] As an exemplary embodiment, Figure 5 As shown in the figure, the meta-strategy network constructed by the spatiotemporal brain module is based on the RLlib training framework, with a state space dimension of 23 and a PPO training algorithm. During the inference process, the dynamic graph neural network (DGNN) deployed by TensorRT is called, and the inference response time does not exceed 200 milliseconds.

[0096] This module executes multi-objective optimization tasks such as construction scheduling, traffic diversion, and carbon emission control by constructing a meta-policy network with generalization capabilities. The system uses a 23-dimensional state space to represent dynamic information of the construction site, including equipment location, operating status, personnel distribution, weather conditions, road traffic grade, etc. The meta-policy network is built on RLlib and uses the PPO (Proximal Policy Optimization) algorithm for iterative training. In the pre-training phase, historical working condition trajectories are used for imitation learning, and online reinforcement learning is introduced for policy adaptation in the fine-tuning phase. In the inference phase, the DGNN model is accelerated through the TensorRT engine, and the graph attention mechanism (GAT) is introduced to enhance the modeling ability of the influence relationship between nodes and edges in the graph. The early-exit mechanism is used to output decisions in advance under high confidence conditions, thereby controlling the average inference delay to less than 200 milliseconds.

[0097] This module boasts high real-time performance and strong generalization capabilities, enabling it to adapt to the dynamic changes in urban reconstruction construction site environments and rapidly respond to scheduling needs. Using a reinforcement learning training framework combining RLlib and the PPO algorithm, it effectively addresses conflicts and optimizes the balance between multi-objective strategies. The introduction of TensorRT to deploy DGNN significantly reduces inference time, meeting the real-time decision-making requirements under complex working conditions. Graph structure modeling and attention mechanisms enhance the rationality and explainability of scheduling decisions, improving the stability and security of the strategy.

[0098] The state space dimensions can be adjusted according to specific construction scenarios. For example, dimensions such as carbon emission load and equipment energy consumption level can be added to enhance green scheduling capabilities. The PPO algorithm can be replaced with other reinforcement learning algorithms such as A3C or SAC to adapt to the different needs of continuous or discrete action spaces. The DGNN inference module can be replaced with a framework based on ONNX or TVM for cross-platform deployment. In addition to TensorRT, inference acceleration methods can also use CUDA kernel optimization, XLA compiler, and other methods to improve processing performance. The early-exit mechanism threshold can be set based on different construction tasks to balance response speed and decision-making accuracy.

[0099] As an exemplary embodiment, Figure 6 As shown, the Oracle node in the blockchain collaborative trust layer compares the difference rate between the LiDAR scan point cloud and the design model. When the difference is less than 5%, a trusted proof is generated, and the chain code automatically executes the payment operation and records the hash value and GPS timestamp.

[0100] This module is designed to achieve the linkage between automated verification of construction status and payment execution. After the construction is completed, the Oracle node calls the LiDAR sensor to perform a three-dimensional point cloud scan of the target area to obtain a high-precision on-site model. The system compares the measured point cloud with the original design model and uses a minimum error registration algorithm (such as ICP) to calculate the difference rate. When the overall difference rate between the measured point cloud and the design model is less than 5%, and the key structure matching degree reaches the set threshold (such as 95% point-to-point overlap), the system determines that the task is completed and meets the requirements. At this point, the Oracle node generates a trusted acceptance certificate, and the chain code automatically executes the payment logic, while recording the operation hash value and the corresponding GPS timestamp to ensure that the verification process is traceable and tamper-resistant.

[0101] This module effectively enhances the objectivity and automation of construction task acceptance, avoiding the errors and disputes associated with traditional reliance on manual inspections and paper reports. Combining the immutable nature of blockchain with LiDAR's high-precision comparison mechanism, it achieves trusted linkage and transparency in the contract fulfillment process, enhancing confidence among all parties involved in the project's execution status. Automatically triggering the payment process reduces manual review costs and shortens cycle times, improving overall construction efficiency.

[0102] The difference rate calculation can use various geometric evaluation indicators such as point cloud Hausdorff distance and Chamfer distance, and the threshold can be flexibly adjusted according to different construction accuracy requirements; LiDAR equipment can be replaced with a multi-modal camera system with depth sensing capabilities (such as RGB-D cameras) to reduce costs; Oracle nodes can be designed to support multiple data source inputs (such as environmental sensors, image recognition results, etc.) to enhance the robustness of the acceptance strategy; the payment chain code execution logic can be customized according to the contract details, including phased payments, dynamic progress ratio settlement, etc.; the timestamp mechanism can also incorporate satellite synchronization verification and multi-node consensus time to enhance anti-counterfeiting capabilities.

[0103] As an exemplary embodiment, Figure 7 As shown, the AR navigation function of the XR collaborative interaction layer includes: AprilTag combined with SLAM to realize spatial registration, construction information is superimposed on the real scene in the form of graphics and text, and workers upload progress information through gestures and trigger blockchain records.

[0104] The AR navigation module is based on computer vision and graph neural network recognition technology, providing construction workers with an augmented reality operation interface. First, the system uses AprilTag tags combined with the SLAM (Simultaneous Localization and Mapping) algorithm to achieve spatial positioning and environmental mapping, ensuring that AR content can be accurately anchored to designated components or areas in the real scene. Construction task information, risk warnings, drawing nodes and other content are presented in the user's field of view through graphic overlays. Users control the interaction process through natural gestures (such as sliding, clicking, pointing, etc.), and the system uses an intention recognition module based on graph neural network training for gesture analysis, with a recognition accuracy rate of over 95%. After workers upload construction progress information (such as task completion, problem reporting, etc.), the system will upload the relevant data to the chain through a predefined smart contract to form a traceable progress record.

[0105] This module significantly improves the efficiency of human-machine collaboration and the intuitiveness of information interaction. Spatial registration technology ensures accurate virtual-real alignment and avoids information drift. Graphical overlays provide clear task guidance and risk warnings, improving operational accuracy and safety. Gesture interaction reduces reliance on traditional terminal devices, freeing construction workers' hands and enhancing operational flexibility. A blockchain recording mechanism ensures that uploaded data cannot be tampered with and is instantly synchronized, providing a real, transparent, and reliable task tracking mechanism for construction management.

[0106] The spatial registration method can be replaced with a visual-inertial navigation (VIO) mechanism based on RGB-D point cloud or ARKit / ARCore to enhance adaptability; the graphic and text overlay information can be expanded to multimodal content such as three-dimensional graphics and voice broadcast to enhance immersion; the gesture recognition module can introduce infrared depth cameras or millimeter-wave radars to enhance detection accuracy; in addition to gesture operations, various human-computer interaction methods such as voice recognition and head motion tracking can also be integrated; the blockchain recording mechanism can be configured as a multi-level trigger strategy based on task type and role permissions to achieve more fine-grained data governance.

[0107] As an exemplary embodiment, the system further includes: a carbon emission optimization module, which is used to establish a carbon emission mapping relationship for construction equipment and realize dynamic planning of carbon emission paths; a causal analysis module, which uses SHAP values ​​to analyze the impact of state variables in multi-objective decision-making; and a federated upgrade module, which uses a differential privacy mechanism to realize cross-project policy migration and model localization optimization.

[0108] The carbon emission optimization module builds a carbon emission factor database, correlating the carbon emission characteristics of different construction equipment (such as energy consumption per unit of operating time and emissions) with factors such as the task, route, and climate. It then uses graph search and path planning algorithms to generate low-carbon scheduling routes. This module introduces carbon emission constraints into task scheduling as one of the optimization objectives.

[0109] The causal analysis module introduces the SHAP (SHapley Additive exPlanations) algorithm to perform causal impact assessment on the state variables (such as equipment location, load, weather, etc.) input into the reinforcement learning strategy network, calculate the contribution of each variable to the task completion probability and decision changes, draw a causal graph, and provide explainable support for the scheduling strategy.

[0110] The federated upgrade module employs a differential privacy mechanism to securely aggregate policy network parameters across multiple projects without exposing specific construction data. It then pushes optimization results to the local DGNN model, enabling cross-project knowledge transfer and localized adaptation. This module sets the privacy protection strength parameter ε to a range of 0.5 to 1, and uses noise perturbation to ensure data sensitivity.

[0111] This extension module integrates three key technical areas: green construction, AI auditing, and intelligent adaptation. The carbon emission optimization module achieves green scheduling goals and improves the system's responsiveness to the "dual carbon" policy. The causal analysis module enhances model transparency and interpretability, helping construction companies review scheduling logic and enforce compliance. The federated upgrade module supports continuous model evolution across different project scenarios, avoiding repetitive training costs, improving algorithm generalization and deployment flexibility, and ensuring data privacy and security.

[0112] The carbon emission factor can be expanded based on local environmental protection standards or internal evaluation indicators of the construction company; in addition to AI and Dijkstra, the path planning algorithm can use reinforcement learning or graph optimization methods to achieve dynamic path replanning; the causal analysis method can use LIME, Counterfactuals and other methods to supplement SHAP for joint interpretation; the federated learning framework can be replaced with architectures such as Flower and FedAvg that support multiple communication protocols, and the differential privacy mechanism can also adjust the noise intensity and gradient clipping strategy according to the sensitivity level.

[0113] As an exemplary embodiment, the edge weight update of the dynamic graph neural network (DGNN) satisfies the following formula:

[0114]

[0115] Among them, Wt is the edge weight matrix at time t, L is the joint loss function, α is the learning rate, β is the coupling function weight parameter, and the spatiotemporal coupling factor Φ(S t ,T t ) is calculated as:

[0116]

[0117] Among them, S t is the current state tensor, is the predicted state, T t The default value of the time decay coefficient λ is 0.5 and it is adjusted to 0.8 when a high-risk event is detected.

[0118] The DGNN model introduces the dynamic coupling effect of time and state differences through the above-mentioned edge weight update mechanism, realizing the adaptive adjustment of the network structure as the construction status changes. The joint loss function L combines the scheduling accuracy and carbon emission cost to ensure that the network controls resource consumption while maintaining the quality of task completion. The coupling factor Φ(S t ,T t ) reflects the deviation between the current state and the predicted state, amplifying the weight adjustment when the state changes dramatically or when a risk event occurs. The time decay term ensures the model's limited memory of historical information. This mechanism enhances DGNN's ability to model spatiotemporal variations in complex and dynamic construction environments.

[0119] By adopting this update mechanism, the DGNN can perceive multiple dynamic factors in the construction scene, such as task status, equipment movement, and risk events, in real time, improving the model's response speed and prediction accuracy. The state error term introduced by the coupling factor ensures the scheduling strategy's sensitivity to construction deviations, while time decay control prevents excessive perturbations and model instability, effectively improving the model's robustness and adaptability over long-term operation.

[0120] The coupling function form can be adjusted based on the characteristics of different construction tasks. For example, cosine similarity, KL divergence and other indicators can be introduced to replace Euclidean distance to improve semantic alignment capabilities. The time decay coefficient can be set as a task sensitivity function for dynamic adjustment, giving high-priority tasks a faster response speed. A constrained regularization term can be added to the joint loss function to control the balance of resource scheduling. If the data frequency in the construction scenario is high, the state tensor update frequency can be combined with the decay factor to form an adaptive step adjustment strategy.

[0121] As an exemplary embodiment, the loss function for GCN topology alignment is defined as:

[0122]

[0123] Among them, f θ is the feature extraction function, MMD is the maximum mean discrepancy, γ is the balance parameter, and the bandwidth parameter σ of the RBF kernel function is determined in the interval [0.1, 1.0] by grid search.

[0124] This loss function is designed to guide the GCN model to achieve alignment and consistent fusion of BIM and GIS models in the feature space. The first term measures the distance between node feature vectors using Euclidean distance, ensuring geometric and semantic consistency at the single-point level. The second MMD term measures the structural deviation between the two models in the latent representation space using the overall distribution difference metric, thereby optimizing global consistency. The balance coefficient γ is used to weigh the contribution between local feature alignment and overall distribution fusion. The RBF kernel function is used to calculate the kernel embedding map in MMD, and its bandwidth parameter σ is determined through grid search to obtain optimal distribution matching performance.

[0125] This loss function effectively improves the topological alignment accuracy of BIM and GIS models, resolving issues such as misaligned model geometry and inconsistent semantic labels. Local point-pair feature differences control detail errors, while global MMD distribution matching enhances overall spatial coordination. This improves the topological stability and task execution reliability of the entire digital twin model, providing solid data support for subsequent scheduling and visualization operations.

[0126] The feature extraction function fθ can be implemented using graph neural network architectures such as GCN, GAT, or GraphSAGE based on different task scenarios; the MMD indicator can be replaced by more complex distribution distance functions such as Sinkhorn distance and Wasserstein distance; the RBF kernel function can also be replaced by a multi-core combination strategy or a deep learning-driven kernel function adaptation mechanism; the γ value can be dynamically adjusted based on the performance of the validation set, and can also be set as a function of the number of training iterations to enhance the model's early convergence stability and later fusion accuracy.

[0127] In one embodiment, a smart scheduling and coordination method for the entire life cycle of urban reconstruction includes the following steps:

[0128] S101: 3D point clouds, equipment operating status, and environmental parameters of the construction site are collected through 5G edge devices and synchronized using ROS2 encapsulation.

[0129] S102: Use GCN to perform topological fusion of BIM, GIS, and TIN models to generate a unified digital twin structure;

[0130] S103: Meta-reinforcement learning-based policy network training, using DGNN for multi-objective reasoning, with response latency less than or equal to 200ms;

[0131] S104: Calling LiDAR to scan the task area, detecting the model difference rate, and automatically triggering the blockchain smart contract when the threshold condition is met;

[0132] S105: Provide construction navigation and collaborative acceptance functions through AR / VR platforms;

[0133] S106: When the scenario matching conditions are met, the federated learning mechanism is used to complete the strategy migration and model update.

[0134] This method builds a complete closed loop of intelligent construction management. First, in S101, high-frequency 5G equipment is used to achieve all-round perception of the on-site status. The ROS2 framework ensures the synchronization of heterogeneous data and maintains timestamp alignment. In S102, the three types of modeling data are fused through the GCN neural network to establish a spatially consistent digital twin graph structure. S103 trains the meta-strategy network based on RLlib, and calls the DGNN deployed with TensorRT for real-time multi-target task reasoning. In the S104 stage, the difference rate is verified by comparing the LiDAR point cloud with the design model, and the blockchain contract is triggered to ensure the credibility of the execution status. S105 provides an XR-supported human-computer interaction interface, allowing workers to visually obtain task information and completion progress; finally, in S106, the federated learning mechanism is used to achieve model migration and localization adjustment between multiple construction projects.

[0135] This approach achieves closed-loop automated management from data collection, model construction, policy reasoning, execution verification, to collaborative interaction, enhancing the intelligence and controllability of the construction process. GCN integration enhances the quality of digital twin models, while DGNN policy optimization achieves a balanced response to multiple task objectives. Blockchain mechanisms enhance transparency and trust in the acceptance process, while the XR platform improves operational efficiency and task collaboration. Federated learning mechanisms strengthen the model's cross-project adaptability and protect data privacy.

[0136] In addition to 5G devices, data collection methods can also be combined with Wi-Fi 6 or LoRaWAN technology to achieve low-power communication; in the modeling process, GCN can be replaced with a dynamic graph structure neural network or a multi-scale network structure to enhance the fusion of models at different levels; the policy network training algorithm can also use DDPG, SAC and other algorithms to improve the policy convergence in the continuous control space; the blockchain platform can choose consortium chain, public chain or hybrid chain structure according to the project scale and the relationship between the participants; the federated learning communication framework can adopt FLUTE, FedProx and other methods to enhance robustness and heterogeneous collaboration capabilities.

[0137] In the following embodiments, a computer-readable storage medium is provided, on which instructions are stored. When the instructions are executed by a processor, the computing device performs all the steps of the above method, including:

[0138] Collect multimodal data from the construction site and perform heterogeneous packaging through ROS2;

[0139] Use GCN to integrate BIM, GIS, and TIN models to build a digital twin graph structure;

[0140] Multi-objective task reasoning based on meta-reinforcement learning and DGNN;

[0141] Trigger the comparison between LiDAR scans and BIM models, and automatically execute blockchain smart contracts based on the difference rate;

[0142] Leverage XR platforms for navigation, acceptance, and multi-user collaboration;

[0143] Apply the federated learning mechanism to achieve strategy migration and model update.

[0144] This computer-readable storage medium can be a physical or virtual medium such as an SSD, eMMC, TF card, or cloud storage platform. The embedded software program includes components such as a multi-threaded control module, a graph neural network inference engine, a blockchain interface program, XR interface rendering logic, and a federated learning communication protocol stack. After the processor reads and executes this program, it can implement end-to-end automated task processes in the urban reconstruction and construction management system. Key modules include edge data stream encapsulation, graph structure learning, real-time policy deployment, smart contract-based verification and payment linkage, and efficient task collaboration among multiple users.

[0145] This computer-readable storage medium integrates full-process management logic and intelligent algorithm models to provide standardized deployment capabilities and flexible application interfaces for urban construction intelligent systems. It is widely adaptable to various construction control platforms (such as edge gateways, mobile terminals, and cloud servers), significantly reducing system integration complexity and improving execution efficiency, ensuring the system's multiple application benefits, including intelligence, trustworthiness, and low carbon.

[0146] The program modules in this medium can be encapsulated in Docker or Kubernetes containers to achieve cross-platform deployment; the processor can be an embedded ARM chip, X86 server or FPGA acceleration platform; the graph neural network module can support the ONNX format to be compatible with third-party AI engines; the blockchain interface can be compatible with Hyperledger Fabric, Ethereum or other mainstream smart contract platforms; the program can also be connected to third-party BIM / GIS platforms through API to achieve task linkage and platform integration applications.

[0147] Application Example 1:

[0148] In a typical implementation, the intelligent scheduling and coordination system described in this invention was applied to the reconstruction of a multifunctional, multi-level transportation hub in the old urban area of ​​a provincial capital city. This project involved simultaneous construction of multiple sections, including subway station expansion, viaduct structural reinforcement, bus transfer channel integration, and intelligent reconstruction of the surrounding road system. The construction site was characterized by its narrow space, compressed construction schedule, and frequent multi-party collaboration, making it a highly representative project.

[0149] During the initial construction phase, the system deployed an edge perception layer network based on 5G communication modules. This network, combined with LiDAR, thermal imaging equipment, environmental sensors, and vision units, enabled multimodal, real-time acquisition of the work area's three-dimensional structure, equipment operating status, personnel behavior, and on-site meteorological data. All perception nodes were heterogeneously packaged using ROS2 middleware, with a unified data synchronization standard implemented through a timestamp mechanism. Edge computing nodes utilized the NVIDIA Jetson AGX Xavier platform, with containerized deployment including data filtering, target tracking, and point cloud preprocessing modules, ensuring that perception data was initially calculated and uploaded to the dispatch center within 30 milliseconds.

[0150] Next, the system enters the model construction stage. The construction party provides the BIM three-dimensional component model, the traffic management platform provides the GIS traffic plot model, and the survey unit provides the TIN terrain grid. The system uses the graph convolutional neural network (GCN) to extract the spatial position, semantic label and physical attributes of the above three types of models as node features, and constructs the corresponding topological map. In the three-layer GCN processing structure, the system uses the softmax fusion algorithm to resolve model boundary conflicts and enhances semantic consistency by introducing the Transformer semantic attention mechanism. The system further dynamically generates cross-model edge structures based on the Euclidean distance (<0.3m) and semantic cosine similarity (>0.85), and uses the maximum mean difference (MMD) as the alignment loss to measure the global feature distribution of the heterogeneous model fusion, and finally constructs a digital twin graph of the transportation hub with multi-scale structural consistency.

[0151] After the model is built, the system models various tasks such as construction scheduling, equipment operation, carbon emission balance, and safety management as a joint multi-objective Markov decision process. The state tensor includes 23 dimensions such as real-time equipment utilization, operation risk level, path occupancy, energy consumption intensity, and traffic impact index. The system uses the PPO reinforcement learning algorithm to train the meta-policy network on the RLlib platform and loads the historical construction case experience pool for pre-training and fine-tuning. The policy reasoning link is executed by the dynamic graph neural network (DGNN) module deployed in the edge computing node. This module introduces a graph attention mechanism to enhance the ability to model inter-edge dependencies. The reasoning process is compiled and deployed through TensorRT, achieving the optimal task sequence generation and control parameter push in <200ms.

[0152] During the construction task execution phase, the system deployed a blockchain collaborative trust layer based on the Hyperledger Fabric consortium chain. Each key construction node is bound to independent task contract logic, defining completion conditions, verifying data sources, and funding triggering logic. When a component is completed, the system uses the LiDAR installed on the equipment to perform a high-precision point cloud scan of the target area. This is then compared to the BIM design model using the ICP registration algorithm. If the matching points are ≥95% and the maximum deviation is <2.5cm, an Oracle node issues a "Trusted Acceptance Certificate." The system executes the chaincode according to the contract terms, automatically triggering funding disbursement and recording the task ID, contract hash, and GPS timestamp to ensure traceability and verifiability of data throughout the entire process.

[0153] At the human-machine collaboration level, the system integrates AR navigation and VR sandboxing to create an XR collaborative interactive environment. When workers enter the work area wearing HoloLens 2, the system uses AprilTag and SLAM mechanisms to achieve sub-centimeter spatial registration and automatically loads the current component's construction drawings, process instructions, and safety warnings. Task information is presented in real time in the user's field of view as an overlay of images and text. After the operation is completed, task feedback can be triggered through natural semantic confirmation or gesture interaction. This feedback event is automatically uploaded, stored, and synchronized to the VR sandbox. Managers can remotely replay and inspect the system, and conduct interactive adjustments and audit analysis on task errors, rhythm deviations, and other issues.

[0154] At the system security level, the system has designed multiple security protection measures for policy models and on-chain contracts. During the deployment process, the policy network introduces a SHA256 model signature mechanism and model permission control identifiers to ensure model version traceability and unauthorized access. In addition to basic encryption of on-chain data, key operations record node identities and two-factor timestamp tamper-proof logs. Furthermore, a fault-tolerant consensus mechanism is enabled across multiple nodes to ensure Fabric ledger consistency and chaincode transaction integrity.

[0155] Through the above integration, this system has realized the full-process closed-loop technical process of "multi-source perception-model fusion-strategy decision-making-trusted execution-interactive collaboration" in this transportation hub reconstruction project, significantly improving construction efficiency, scheduling response speed and multi-party collaboration transparency, and fully verifying the applicability, versatility and high reliability of the system of the present invention in complex urban reconstruction projects.

[0156] Application Example 2:

[0157] During the construction of an underground integrated pipeline corridor in a new urban district, the construction team implemented the system presented in this paper to perform real-time digital twin alignment modeling and construction conflict prediction for the access section. Initially, the design firm provided a BIM pipeline corridor model in IFC format and a GIS model of the urban underground in CityGML format to guide excavation routes and equipment layout.

[0158] The system first calls the GCN module to align the graph structures constructed by the two models mentioned above. The BIM model contains 3467 nodes (representing air ducts, water pipes, cable trays, etc.), and the GIS model contains 2184 underground feature nodes. The feature vector dimension of the node is 128, including spatial coordinates (x, y, z), semantic type encoding (one-hot), and component physical properties (such as inner diameter and material density). After the weights of the graph convolutional network are initialized, training and optimization are performed using the following loss function:

[0159]

[0160] The MMD (maximum mean difference) term is used to measure the overall embedding distribution difference of the model, and γ is set to 0.3. The kernel function uses the RBF kernel, and the bandwidth parameter σ is optimized through grid search between 0.1 and 1.0, and finally σ = 0.45.

[0161] After GCN training converged to a validation error below 0.05, the system automatically identified a 42mm center offset between a certain air duct (component ID: W-245) and an underground stormwater pipeline node. The system adjusted the GCN weight based on the geometric accuracy priority (weight 0.6), recommended an adjusted air duct path, and highlighted the conflict in real time using AR.

[0162] Ultimately, the adjusted path offset was controlled within ±10mm, avoiding the risk of on-site rework; the operation record was uploaded to the chain through the Oracle node, binding the component ID, change time and responsible person information to ensure the credibility of the data closed loop.

[0163] Application Example 3:

[0164] During the demolition and reconstruction of a certain elevated bridge, the system deployed edge sensing nodes and 5G edge boxes at the construction site, collecting real-time data from multiple sources, including rainfall, wind speed, equipment status, and operational emissions. At 3:24 PM on a specific day, the system detected local rainfall intensity reaching 14.3 mm / h and wind speed increasing to 10.1 m / s, approaching the preset risk threshold.

[0165] The system's spatiotemporal brain module immediately starts the dynamic scheduling reasoning process and constructs the Markov decision state tensor St, which contains the following indicators:

[0166] Crane utilization rate: 0.87

[0167] Vehicle traffic density: 0.52

[0168] PM2.5 concentration: 78 μg / m 3

[0169] Current rainfall: 14.3 mm / h

[0170] Carbon emission weight factor: default 0.5

[0171] At this point, the DGNN graph structure has a total of 1327 nodes and 2716 edges. The system uses the following weight update formula to calculate the scheduling priority weights between tasks:

[0172]

[0173] The coupling factor is calculated as follows:

[0174]

[0175] Among them, α=0.01, β=0.05, and the attenuation coefficient λ is automatically adjusted to 0.8 (increased from the default value of 0.5) based on the high-risk events detected by the system to enhance time sensitivity.

[0176] Calculations show that the emissions intensity of material transport task "MT-038" is 1.13 kg CO2 / km, but it has the highest transferability. The system completes strategy reconstruction within 47 milliseconds and switches the task from the original route T3 to the backup route T5, reducing the overall carbon emissions to 0.79 kg CO2 / km.

[0177] The relevant scheduling adjustment results are recorded through the Fabric chain smart contract, and AR navigation update instructions are sent to the front-line workers' terminals. The entire response process takes less than 5 seconds.

[0178] Application Example 4:

[0179] In a municipal underground space expansion project, due to the high degree of closure of the construction work surface and multiple risk factors, the project team used the XR collaborative interaction layer of the present invention to perform task distribution, construction navigation and real-time feedback.

[0180] Workers wearing HoloLens 2 smart glasses enter the work area. The system uses AprilTag QR codes placed on the wall for initial positioning and uses the SLAM module to perform spatial map corrections, with registration accuracy controlled to ±1.8cm.

[0181] In this XR collaborative interactive system, construction tasks are distributed graphically, using components as the smallest operational unit. For example, for the utility corridor node "JX-W322," the system retrieves the corresponding standard construction process and auxiliary information from the database based on the task number and pushes it to the frontline operator's terminal in a visual format.

[0182] During this node's operation, the construction process includes drilling, inserting casing, securing the structure, and backfilling. The required tools include an impact drill, an electric hammer, and a bracket and casing assembly. The impact drill is used for precise drilling of the concrete surface, the electric hammer assists in striking the foundation structure, and the bracket and casing assembly is used to position the pipes and ensure stability.

[0183] Given that there is a 10kV high-voltage cable buried near the work area, in order to ensure work safety and personnel protection, the system simultaneously provides safety reminder information, explicitly prohibiting the use of non-insulated power tools for any contact operations, and all construction work must comply with the power facility protection specifications.

[0184] This task information, including work order, tool matching, and safety risk alerts, is presented in real time to workers via AR graphics and text overlays. During the operation, workers trigger task feedback through dual confirmation using gesture recognition and voice commands. The system then generates a construction record, which is synchronized to the task management and blockchain logging modules, achieving a closed-loop management of task execution and data traceability.

[0185] During execution, the system calls the following data structure to bind the on-site operation to the task blockchain entry:

[0186] protobuf

[0187] message TaskExecution{

[0188] string user_id = 1;

[0189] string task_id = 2;

[0190] string component_id=3;

[0191] float completion_rate = 4;

[0192] repeated string media_evidence=5;

[0193] string hash_id = 6;

[0194] int64 timestamp = 7;

[0195] }

[0196] in:

[0197] ˋcompletion_rate=1.0ˋ(100%);

[0198] ˋmedia_evidenceˋ includes on-site photos and infrared temperature measurement screenshots;

[0199] ˋhash_idˋ is bound to the Fabric ledger through SHA256 calculation;

[0200] ˋtimestamp=2025-05-1209:31:52ˋ.

[0201] At the same time, the VR sandbox synchronously updates the twin status, and project commanders can remotely replay and interactively inspect the operation process in real time, greatly improving collaboration efficiency and construction visualization transparency.

[0202] Application Example 5:

[0203] In a certain affordable housing reconstruction project, to improve the interpretability of the construction scheduling system and the transparency of the engineering decision-making process, the construction party activated the causal analysis module in the system of the present invention based on the "space-time brain module" and analyzed and explained the internal behavior of the AI ​​strategy based on the SHAP (SHapleyAdditive exPlanations) value.

[0204] During the 42nd construction period, the system needs to optimize the order of the following three scheduling subtasks:

[0205] 1. Formwork support (task ID: TP-201);

[0206] 2. Concrete pouring (TP-202);

[0207] 3. Material transportation (TP-203).

[0208] The current state tensor S of the scheduling model t The key dimensions include:

[0209] Concrete temperature: 22.5℃;

[0210] Concrete age: 8 hours;

[0211] Construction worker density: 0.68;

[0212] On-site light intensity: 4300lx;

[0213] CO2 emission threshold limit: 40kg / h.

[0214] Based on the deployed dynamic graph neural network (DGNN) inference model, the AI ​​system conducts a comprehensive assessment of the current construction status and outputs the optimal scheduling execution order: first execute the material transportation task (TP-203), followed by the formwork support task (TP-201) and the concrete pouring task (TP-202).

[0215] To perform causal tracing, as shown in Table 2, the system generates a SHAP explanation matrix for the decision sequence and outputs the following results:

[0216] Table 2: Some results of the system generating SHAP explanation matrix output for decision sequences

[0217] Feature Item SHAP value Contribution direction Age (8h) +0.42 Accelerated TP-202 <![CDATA[CO2 threshold limit]]> -0.61 Postponement of TP-203 Population density 0.68 +0.25 Accelerated TP-201

[0218] After visualizing the distribution of all SHAP values, the system automatically constructs a causal graph as shown below:

[0219] Node: task decision result;

[0220] Edges: SHAP explains causal edges;

[0221] Edge weight: w ij =|SHAP i -SHAP j |

[0222] Visual layout: Fruchterman-Reingold algorithm for automatic layout.

[0223] The graph results are submitted to the project supervision interface as structured JSON records, allowing supervisors to view the causal basis of the system in specific scheduling decisions and evaluate whether they are reasonable or have abnormal preference behavior.

[0224] If scheduling anomalies or inconsistencies in strategy weights are found, the system also allows supervisors to manually adjust the weights of influencing factors, trigger the strategy retraining process in real time, and leave a record in the supervisory channel ledger.

[0225] This intelligent scheduling and coordination system, designed for the entire lifecycle of urban reconstruction, utilizes a 5G edge perception layer to achieve millisecond-level data acquisition (≤200ms latency). It uses GCN-based multi-model topology alignment technology (BIM-GIS-TIN fusion error <2.5cm) to construct dynamic digital twins. It employs a meta-reinforcement learning framework (PPO algorithm) and a spatiotemporal coupled DGNN (λ = 0.5-0.8 dynamic adjustment) to achieve multi-objective joint optimization decision-making. The system innovatively combines LiDAR discrepancy rate verification (95% threshold) with a Hyperledger Fabric smart contract automatic triggering mechanism, and collaborates with SLAM + AprilTag's XR collaborative interaction (positioning accuracy 1.8cm), forming a closed-loop technical loop of "perception-modeling-decision-making-verification-interaction." Verified in actual projects, this system has increased construction scheduling response speed by 40%, reduced carbon emissions by 22%, and shortened the acceptance and payment cycle from the traditional 3-5 days to within 10 minutes. It has significantly solved industry pain points such as heterogeneous model alignment errors (>5cm), emergency response lags (>30s), and inefficient multi-party collaboration.

[0226] It should be understood that the various technical solutions disclosed in the present invention are not limited to the specific forms listed in the description and embodiments. Without departing from the core concept of the present invention, technicians can make various equivalent transformations and replacements to the structure, algorithm, module combination or parameter selection, which should all fall within the scope of protection of the present invention.

Claims

1. An intelligent scheduling and coordination system for the entire life cycle of urban reconstruction, characterized by: include: The edge perception layer, including a cluster of 5G IoT devices, is used to collect 3D point clouds, environmental data, and personnel trajectories at the construction site, and align multi-source heterogeneous data through ROS2; The digital twin weaving module is used to topologically align BIM, GIS, and TIN models based on graph convolutional neural networks (GCN) to build dynamic multi-scale digital twins; The spatiotemporal brain module implements multi-task decision-making for construction scheduling, traffic diversion, and carbon emission control based on a meta-reinforcement learning framework; The blockchain collaborative trust layer, based on the Hyperledger Fabric architecture, verifies the construction completion status through Oracle nodes and triggers smart contract execution; The XR collaborative interaction layer provides AR navigation and VR sandbox functions, supporting immersive collaboration and visual acceptance.

2. The system according to claim 1, wherein: The graph convolutional neural network (GCN) node features used in the digital twin mesh module include spatial coordinates, semantic labels and physical properties. The GCN adopts a three-layer network structure and uses a weighted fusion algorithm at the model boundary to handle topological conflicts.

3. The system according to claim 1, wherein: The meta-strategy network constructed by the spatiotemporal brain module is based on the RLlib training framework, with a state space dimension of 23 and a PPO training algorithm. During the inference process, the dynamic graph neural network (DGNN) deployed by TensorRT is called, and the inference response time does not exceed 200 milliseconds.

4. The system according to claim 1, wherein: In the blockchain collaborative trust layer, the Oracle node compares the difference rate between the LiDAR scan point cloud and the design model. When the difference is less than 5%, a trusted proof is generated. The chain code automatically executes the payment operation and records the hash value and GPS timestamp.

5. The system according to claim 1, wherein: The AR navigation function of the XR collaborative interaction layer includes: AprilTag combined with SLAM to realize spatial registration, construction information is superimposed on the real scene in the form of graphics and text, and workers upload progress information through gestures and trigger blockchain records.

6. The system according to any one of claims 1 to 5, characterized in that Also includes: Carbon emission optimization module, used to establish carbon emission mapping relationships for construction equipment and realize dynamic planning of carbon emission paths; Causal analysis module, which uses SHAP value to analyze the influence of state variables in multi-objective decision-making; The federated upgrade module uses a differential privacy mechanism to achieve cross-project policy migration and model localization optimization.

7. The system according to claim 3, wherein: The edge weight update of the dynamic graph neural network (DGNN) satisfies the following formula: Where Wt is the edge weight matrix at time t, L is the joint loss function, α is the learning rate, β is the coupling function weight parameter, and Φ(St,Tt) is the spatiotemporal coupling factor, which is calculated as: Among them, S t is the current state tensor, is the predicted state, T t is the timestamp, λ is the time decay coefficient, the default value is 0.5, and it is adjusted to 0.8 when a high-risk event is detected.

8. The system according to claim 2, wherein: The loss function of the GCN topology alignment is: Among them, f θ is the feature extraction function, MMD is the maximum mean difference, γ is the balance parameter, and the RBF kernel function bandwidth σ is determined in the interval [0.1, 1.0] by grid search.

9. An intelligent scheduling and coordination method for the entire life cycle of urban reconstruction, characterized by: The following steps are involved: S101 collects 3D point clouds, equipment operating status, and environmental parameters of the construction site through 5G edge devices and synchronizes them using ROS2 encapsulation; S102, using GCN to perform topological fusion of BIM, GIS, and TIN models to generate a unified digital twin structure; S103, based on meta-reinforcement learning training policy network, uses DGNN for multi-objective reasoning, and has a response delay of less than or equal to 200ms; S104, calling LiDAR to scan the task area, detecting the model difference rate, and automatically triggering the blockchain smart contract when the threshold condition is met; S105, provides construction navigation and collaborative acceptance functions through AR / VR platforms; S106: When the scenario matching conditions are met, the federated learning mechanism is used to complete the strategy migration and model update.

10. A computer-readable storage medium having instructions stored thereon, which, when executed by a processor, cause a computing device to execute all the steps of the method of claim 9.

Images