A digital twin-based intelligent control method and system for engineering management
By constructing a digital twin graph and combining it with graph convolutional networks and reinforcement learning, equipment control commands and scheduling parameters are generated. This solves the problem that existing systems cannot automatically calculate optimal scheduling parameters, realizes dynamic analysis and reverse closed-loop automated control of engineering management, and improves the intelligence level of the system.
Patent Information
- Application Number
- CN202611438897.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing digital twin engineering management systems lack algorithm-driven reverse closed-loop control capabilities. They cannot automatically calculate and issue optimal equipment servo control commands and resource scheduling parameters based on the real-time situation at the engineering site, resulting in a serious lag in on-site intervention response and failing to achieve automated and intelligent closed-loop engineering management.
By constructing an initial digital twin graph, combining graph convolutional networks and reinforcement learning strategies, evolutionary feature extraction and strategy evaluation are performed to generate a resource scheduling strategy matrix. Then, a greedy algorithm is used to allocate equipment control commands and scheduling parameters with micro-delay peak-shifting parameters, thereby realizing automated intervention and closed-loop control of equipment status.
It enables dynamic analysis and reverse closed-loop automated intervention of the engineering site environment, automatically amplifies the feature weights of nodes with attention scores above the threshold, quickly converges to the safety scheduling strategy, smooths the power load curve of the engineering site, avoids the buffer breakdown problem of traditional early warning, and realizes the automated and intelligent management and control of the system.
Smart Images

Figure CN123118029A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing and intelligent engineering management technology, specifically to an intelligent control method and system for engineering management based on digital twins. Background Technology
[0002] Large-scale engineering projects involve multi-disciplinary collaboration, multi-equipment scheduling, and complex physical environment evolution. Their construction and operation processes are highly dynamic, characterized by multi-variable coupling and uncertain risks. With the promotion of the Internet of Things and 3D modeling technology, digital twin technology has been gradually introduced into the field of engineering management, attempting to achieve visualization and intelligent control of the entire lifecycle of a project by mapping the state of physical entities.
[0003] Existing digital twin engineering management systems primarily focus on the time-series overlay and static 3D display of multi-source sensor data, typically relying on fixed static thresholds to trigger early warnings. However, engineering site environments are complex and ever-changing, and current technologies generally lack algorithm-driven reverse closed-loop control capabilities. When the system predicts a bottleneck in project progress or a potential structural safety hazard, it often only reaches the level of open-loop monitoring and large-screen alarms at the front end. When dealing with sudden risks or resource conflicts, it still heavily relies on manual analysis and intervention based on human experience. The system itself cannot automatically calculate and directly issue optimal equipment servo control commands and resource scheduling parameters based on the real-time evolution of the site situation. This decision-making gap leads to a severe lag in on-site intervention response, failing to truly achieve automated and intelligent closed-loop engineering management.
[0004] To address this, a digital twin-based intelligent control method and system for engineering management is proposed. Summary of the Invention
[0005] The purpose of this invention is to provide an intelligent control method and system for engineering management based on digital twins. By combining graph convolutional network evolution prediction and reinforcement learning strategy evaluation, it can realize dynamic analysis of the engineering site environment and automated intervention in reverse closed loop.
[0006] To achieve the above objectives, the present invention provides the following technical solution: A digital twin-based intelligent control method for engineering management includes: Acquire the equipment timing operation data and process association topology data at the engineering site, generate an initial digital twin map and extract the node timing feature matrix and spatial adjacency matrix; Evolutionary features are extracted from the temporal feature matrix and spatial adjacency matrix of the nodes based on graph convolutional networks. Attention scores are constructed using physical spatial distance and process dependency strength, and the spatial adjacency matrix is masked and modulated to amplify the aggregation weight of nodes with attention scores higher than a preset threshold, generating a resource node evolution state sequence. The difference between the resource node evolution state sequence and a preset engineering scheduling constraint threshold is calculated, and out-of-bounds residual variables are extracted and density clustered according to the spatial coordinate set to generate a resource conflict distribution matrix that includes spatial conflicts and equipment power consumption status. Based on a reinforcement learning network, a strategy evaluation is performed on the resource conflict distribution matrix. When the minimum prediction spatial distance between nodes is less than or equal to a preset safety threshold, a nonlinearly exponentially growing dynamic safety penalty scalar is triggered as a negative reward to generate a resource scheduling strategy matrix. The resource scheduling strategy matrix is analyzed to extract the peak power consumption of concurrent devices at startup. Based on a greedy algorithm, an incremental delay trigger time is allocated to generate device control instructions and scheduling parameters containing micro-delay peak-shifting parameters.
[0007] Preferably, the process of generating an initial digital twin graph and extracting the node temporal feature matrix and spatial adjacency matrix therein includes: The device timing operation data is mapped to node attribute information in the initial digital twin graph, and the node timing feature matrix is integrated and constructed. The topological relationships in the process association topology data are parsed and mapped to node edges in the initial digital twin graph to construct a spatial adjacency matrix.
[0008] Preferably, the process of constructing attention scores using physical spatial distance and process dependency strength to perform mask modulation on the spatial adjacency matrix includes: Extract the physical spatial distance and process dependency strength between each node in the initial digital twin graph; perform a weighted calculation on the physical spatial distance and the process dependency strength to obtain the attention score between each node; generate a mask matrix based on the attention score, and multiply the mask matrix with the spatial adjacency matrix.
[0009] Preferably, the process of generating the resource conflict distribution matrix specifically includes: The difference between the state values in the evolution state sequence of the resource nodes and the engineering scheduling constraint threshold is calculated, and the part of the difference exceeding the threshold is extracted as an out-of-bounds residual variable; the spatial coordinate set of the node corresponding to the out-of-bounds residual variable in the initial digital twin map is obtained; density clustering is performed on the out-of-bounds residual variable based on the spatial coordinate set, and a resource conflict distribution matrix containing spatial conflict and device power consumption status is output according to the clustering result.
[0010] Preferably, the process of evaluating the resource conflict distribution matrix based on a reinforcement learning network specifically includes: The resource conflict distribution matrix is input into the reinforcement learning network to perform policy evaluation; during the evaluation process, node trajectories are monitored in real time to calculate the minimum prediction spatial distance between nodes; when the minimum prediction spatial distance between nodes is less than or equal to the safety threshold, the ratio of the safety threshold to the minimum prediction spatial distance between nodes is calculated; using the natural constant as the base and the ratio as the exponent, an exponential operation is performed to obtain a dynamic safety penalty scalar that grows non-linearly, and the dynamic safety penalty scalar is added as a negative reward to the feedback of the reinforcement learning network.
[0011] Preferably, the process of generating equipment control commands and scheduling parameters containing micro-delay peak-shaving parameters specifically includes: The resource scheduling strategy matrix is parsed to identify the corresponding concurrent devices, and the peak power consumption at startup of each concurrent device is extracted. Using a greedy algorithm, incremental delay trigger times are allocated to each concurrent device according to the power consumption distribution. The allocated incremental delay trigger times are used as peak shaving parameters and written into the device instructions to generate device control instructions and scheduling parameters containing micro-delay peak shaving parameters.
[0012] Preferably, after generating the device control instructions and scheduling parameters containing the micro-delay peak-shaving parameters, the method further includes: The equipment control commands and scheduling parameters containing micro-delay peak-shaving parameters are sent to the corresponding equipment at the engineering site for control; the latest equipment timing operation data after control is collected and synchronously updated to the initial digital twin map to form a closed loop.
[0013] A digital twin-based intelligent control system for engineering management includes: The graph generation and extraction module acquires the equipment time-series operation data and process-related topology data at the engineering site, generates an initial digital twin graph, and extracts the node time-series feature matrix and spatial adjacency matrix from it. Graph Convolutional Mask Modulation Module: Based on the graph convolutional network, the module extracts evolutionary features from the temporal feature matrix and spatial adjacency matrix of the nodes. It constructs attention scores using physical spatial distance and process dependency strength to perform mask modulation on the spatial adjacency matrix, amplifies the aggregation weight of nodes with attention scores higher than a preset threshold, and generates a sequence of resource node evolution states. Difference clustering module: Calculates the difference between the evolution state sequence of the resource nodes and the preset engineering scheduling constraint threshold, extracts the out-of-bounds residual variables, and performs density clustering according to the spatial coordinate set to generate a resource conflict distribution matrix; Reinforcement learning evaluation module: Based on the reinforcement learning network, the resource conflict distribution matrix is evaluated for policy. When the minimum prediction spatial distance between nodes is less than or equal to the preset safety threshold, a nonlinear exponentially growing dynamic safety penalty scalar is triggered as a negative reward to generate a resource scheduling policy matrix. Peak shaving instruction generation module: parses the resource scheduling strategy matrix, extracts the peak power consumption of concurrent devices at startup, allocates incremental delay trigger times based on a greedy algorithm, and generates device control instructions and scheduling parameters containing micro-delay peak shaving parameters.
[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs an initial digital twin graph and introduces a graph convolutional network with attention mask modulation, which can simultaneously quantify and fuse physical spatial distance and process dependency strength. This enables the model to automatically amplify the feature weights of nodes with attention scores higher than a preset threshold and strongly coupled nodes when extracting evolutionary features, thus overcoming the shortcomings of traditional static 3D models in capturing complex spatial and temporal evolution patterns.
[0015] 2. This invention innovatively introduces a nonlinearly exponentially growing dynamic safety penalty scalar into reinforcement learning networks. When the spatial spacing is less than or equal to the safety threshold, the exponential negative reward with a base of the natural constant can quickly converge the policy search space, forcing the system to output a safer scheduling policy and avoiding the buffer breakdown problem that is easily caused by traditional linear warnings.
[0016] 3. This invention seamlessly transforms the front-end digital twin early warning results into equipment control commands. It uses a greedy algorithm to perform low-delay peak power consumption scheduling for concurrent devices and directly sends control commands to the field equipment. This closed-loop mechanism of "prediction-decision-control-state update" replaces manual intervention based on experience, smoothing the power load curve at the engineering site. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the steps of an intelligent control method for engineering management based on digital twins according to the present invention. Figure 2 This is a diagram illustrating the closed-loop interaction process of digital twin map updating and device micro-delay peak-shifting control in this invention. Figure 3 This is a schematic diagram of the structure of an intelligent control system for engineering management based on digital twins according to the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Example 1: Please see Figures 1 to 2 This invention provides an intelligent control method for engineering management based on digital twins, the technical solution of which is as follows: A digital twin-based intelligent control method for engineering management includes: Acquire the equipment timing operation data and process association topology data at the engineering site, generate an initial digital twin map and extract the node timing feature matrix and spatial adjacency matrix; Evolutionary features are extracted from the temporal feature matrix and spatial adjacency matrix of the nodes based on graph convolutional networks. Attention scores are constructed using physical spatial distance and process dependency strength, and the spatial adjacency matrix is masked and modulated to amplify the aggregation weight of nodes with attention scores higher than a preset threshold, generating a resource node evolution state sequence. The difference between the resource node evolution state sequence and a preset engineering scheduling constraint threshold is calculated, and out-of-bounds residual variables are extracted and density clustered according to the spatial coordinate set to generate a resource conflict distribution matrix that includes spatial conflicts and equipment power consumption status. Based on a reinforcement learning network, a strategy evaluation is performed on the resource conflict distribution matrix. When the minimum prediction spatial distance between nodes is less than or equal to a preset safety threshold, a nonlinearly exponentially growing dynamic safety penalty scalar is triggered as a negative reward to generate a resource scheduling strategy matrix. The resource scheduling strategy matrix is analyzed to extract the peak power consumption of concurrent devices at startup. Based on a greedy algorithm, an incremental delay trigger time is allocated to generate device control instructions and scheduling parameters containing micro-delay peak-shifting parameters.
[0020] The process of generating an initial digital twin graph and extracting the node temporal feature matrix and spatial adjacency matrix includes: The device timing operation data is mapped to node attribute information in the initial digital twin graph, and the node timing feature matrix is integrated and constructed. The topological relationships in the process association topology data are parsed and mapped to node edges in the initial digital twin graph to construct a spatial adjacency matrix.
[0021] By deploying sensor arrays on key equipment at the engineering site, dynamic monitoring parameters of the equipment during operation are collected in real time as the time-series operation data of the equipment. Digital twin nodes corresponding one-to-one with the on-site equipment are established in virtual space. A feature encoding network is used to map the real-time collected time-series operation data of the equipment into high-dimensional feature vectors. Specifically, the feature encoding network adopts a multilayer perceptron structure, including an input layer, three hidden layers, and an output layer. Adjacent network layers are connected through a fully connected manner and nonlinearly mapped using the ReLU activation function, thus upscaling the low-dimensional original equipment operation data to a fixed-dimensional high-dimensional feature vector. This feature vector is then used as the node attribute information in the initial digital twin map and assigned to the corresponding digital twin node in real time. A time sliding window mechanism is introduced, setting a fixed time window length and sliding step size. Based on a unified timestamp, the feature vectors of all digital twin nodes within the same time window are aligned and combined according to the node sequence number to integrate and construct a node time-series feature matrix reflecting the dynamic characteristics changes of the equipment. By reading the building information model and construction schedule of the project, a graph analysis algorithm is used to extract the process association topology data between each work stage and corresponding equipment. The process association topology data contains the dependencies between preceding and subsequent processes and the requirements for spatial collaborative operations. A connection channel is established between digital twin nodes with dependencies or collaborative relationships, and it is mapped to the directed node edges in the initial digital twin graph from the preceding node to the following node. All node edges in the initial digital twin graph are traversed, and the edge relationships are matrixed using a graph theory matrix generation algorithm. That is, it is determined whether there is a node edge between node pairs. If there is, a weight of 1 is assigned to the corresponding element position in the matrix; otherwise, it is set to zero. In this way, a directed spatial adjacency matrix reflecting the engineering space and process structure is constructed.
[0022] By encoding the node features of the equipment operating parameters collected in real time on site and deeply analyzing the process topology of the project schedule, a high-fidelity mapping of the physical reality node temporal feature matrix and spatial adjacency matrix is obtained. This enables the accurate quantitative expression of the coupling relationship between the operating status of engineering equipment and spatial processes in the virtual digital space, solving the problem that the digital twin engineering management system mainly focuses on the temporal superposition and static three-dimensional display of multi-source sensor data.
[0023] The process of masking the spatial adjacency matrix by constructing attention scores using physical spatial distance and process dependency strength includes: Extract the physical spatial distance and process dependency strength between each node in the initial digital twin graph; perform a weighted calculation on the physical spatial distance and the process dependency strength to obtain the attention score between each node; generate a mask matrix based on the attention score, and multiply the mask matrix with the spatial adjacency matrix.
[0024] By invoking the 3D spatial positioning system deployed on-site, the system analyzes and acquires the dynamic 3D coordinate data of each engineering equipment node in real time. The Euclidean distance formula is used to calculate the physical spatial distance between any two nodes in real time. Simultaneously, the system connects to the on-site construction scheduling engine to capture the pre-executed task constraints and material flow frequency between nodes in real time. A pre-set process association calculation model is used to weight and quantify the task constraint level and flow frequency, specifically employing a linear weighting formula. ,in The normalized task constraint level, This represents the normalized material flow frequency. and To satisfy the preset weight coefficients that sum to 1, they can be predetermined based on the analytic hierarchy process (AHP) to output normalized process dependency strength with values between zero and one. A graph convolutional network is constructed, whose network structure includes multiple stacked graph convolutional layers. Each graph convolutional layer sequentially includes a feature aggregation module for gathering neighbor node information, a linear transformation module for feature dimension mapping, and a non-linear activation module using the ReLU function. Learnable weight parameters from the linear transformation module of the graph convolutional network are introduced to construct the attention calculation formula: ,in This represents the initial attention score between nodes. It represents the real-time physical distance and uses an exponential decay function to make it negatively correlated with attention. Indicates the intensity of process dependence. and The weight allocation coefficients are dynamically updated based on the loss gradient during the backpropagation process of network training. After calculating the initial scores of all node pairs, the scores of each node's neighbor nodes are mapped to a global probability distribution using a normalized exponential function to obtain the final attention scores between nodes. The final attention scores of all node pairs are mapped and filled one by one according to the row and column indices of the node numbers. A mask matrix with the same dimension as the spatial adjacency matrix is generated based on the final attention scores. Before feature aggregation in the graph convolutional network, the corresponding elements of the mask matrix and the spatial adjacency matrix are multiplied respectively to obtain the modulated spatial adjacency matrix used for feature aggregation. Then, the spatial adjacency matrix after mask modulation is input into the feature aggregation module, and graph convolution is performed layer by layer forward propagation in combination with the node temporal feature matrix. The resource node evolution state sequence is obtained by extracting the output of the last layer of the network. Specifically, the graph convolutional network combines with a temporal convolutional network to form a spatiotemporal graph model. The graph convolutional network is configured with three stacked graph convolutional layers for spatial topological feature aggregation, and the temporal convolutional network employs a residual block structure including dilated causal convolutions to capture long-term sequence dependencies. The input time step is set to the past 60 consecutive time periods, and the prediction time window is the prediction time step corresponding to the next 10 to 30 minutes. The final output dimension of the spatiotemporal graph model is... The three-dimensional tensor, in which For the number of resource nodes, To predict the number of time steps, The output feature dimension includes the predicted values of the node's three-dimensional spatial displacement and the predicted state of power consumption. During the model training phase, the historically collected node spatial displacement and power consumption data are used as the true labels, and the mean squared error is used as the loss function for supervised training.
[0025] By introducing real-time spatial positioning analysis and task scheduling quantization techniques to calculate attention weights and perform matrix multiplication, a masked modulated spatial adjacency matrix with risk and coupling scaling was obtained. This achieved the effect of automatically amplifying the feature weights of nodes with attention scores higher than a preset threshold and strong coupling, solving the problem that digital twin systems cannot automatically perceive multivariate coupling risks and lack algorithm-driven closed-loop control.
[0026] The process of generating the resource conflict distribution matrix specifically includes: The difference between the state values in the evolution state sequence of the resource nodes and the engineering scheduling constraint threshold is calculated, and the part of the difference exceeding the threshold is extracted as an out-of-bounds residual variable; the spatial coordinate set of the node corresponding to the out-of-bounds residual variable in the initial digital twin map is obtained; density clustering is performed on the out-of-bounds residual variable based on the spatial coordinate set, and a resource conflict distribution matrix containing spatial conflict and device power consumption status is output according to the clustering result.
[0027] The system receives the resource node evolution state sequence output by the graph convolutional network in real time, extracts the predicted three-dimensional spatial position and corresponding predicted power consumption state value of each device node within the future prediction time window explicitly mapped in the sequence, and synchronously connects to the construction management database of the project site to dynamically read the maximum power load (power upper limit) and minimum safe distance (distance lower limit) allowed for the current construction stage, and uses them as the project scheduling constraint threshold. It traverses the state sequence and extracts out-of-bounds residual variables: for power state, the difference between the predicted power and the maximum power load upper limit is extracted as the power out-of-bounds residual; for spatial position, the predicted distance between node pairs is calculated, and the difference between the predicted distance and the minimum safe distance lower limit is extracted as the distance out-of-bounds residual. The system obtains the spatial coordinate set of the nodes corresponding to the two types of out-of-bounds residual variables in the initial digital twin graph. Using the coordinate points in the spatial coordinate set as data samples, the core distance and reachability distance between each coordinate point are calculated using a density-based spatial clustering algorithm. Based on the dynamically calculated neighborhood radius and minimum number of contained points, a connected region composed of dense data points is searched, and nodes that are spatially close and densely cross the boundary are divided into the same cluster to complete density clustering. The neighborhood search radius of the density clustering is adaptively and dynamically adjusted based on the device bounding box expansion rate. Specifically, when performing density clustering based on a spatial coordinate set, a fixed neighborhood search radius is not used. Instead, based on the device type corresponding to the out-of-bounds residual variable, its maximum operating speed and robotic arm extension range are extracted to calculate and construct a dynamically expanded three-dimensional bounding box. Specifically, the formula is used... Calculate the dynamic neighborhood search radius ,in Let the radius be the radius of the static foundation bounding box of the equipment. For maximum operating speed, For the robotic arm's extension range, and These are the preset velocity expansion coefficient and spatial extension coefficient; velocity expansion coefficient The preferred value range for the spatial extension coefficient is 1.2-1.5. The preferred value range is 1.1-1.3; the extreme boundary of the expanded bounding box is used as the dynamic neighborhood search radius of density clustering to adaptively delineate the regions of nodes that cross the boundary; by adaptively adjusting the neighborhood search radius of density clustering, clusters with boundaries that better fit the actual physical interference range are obtained, thus achieving accurate delineation of potential spatial collision regions. During the clustering process, the mapping relationship of the device unique identifier corresponding to each node is retained; the centroid coordinates and boundary range of each cluster are extracted as spatial conflict features, and the retained device unique identifier is mapped back to the corresponding out-of-bounds residual variable to extract the current power consumption value of the corresponding node as the device power consumption status feature. After processing isolated noise points, the spatial conflict features and the device power consumption status features are matrix-concatenated according to the cluster number, and a resource conflict distribution matrix is generated and output based on the clustering results.
[0028] By introducing dynamic state difference calculation and density clustering technology based on spatial coordinate sets, a resource conflict distribution matrix containing spatial conflicts and equipment power consumption status was obtained. This enabled the accurate location and quantification of resource cross-interference and local overload areas on the engineering site, solving the problem that when dealing with sudden risks or resource conflicts, manual analysis and intervention based on human experience are highly dependent, and the system cannot automatically calculate and analyze based on the situation.
[0029] The process of evaluating the resource conflict distribution matrix based on a reinforcement learning network specifically includes: The resource conflict distribution matrix is input into the reinforcement learning network to perform policy evaluation; during the evaluation process, node trajectories are monitored in real time to calculate the minimum prediction spatial distance between nodes; when the minimum prediction spatial distance between nodes is less than or equal to the safety threshold, the ratio of the safety threshold to the minimum prediction spatial distance between nodes is calculated; using the natural constant as the base and the ratio as the exponent, an exponential operation is performed to obtain a dynamic safety penalty scalar that grows non-linearly, and the dynamic safety penalty scalar is added as a negative reward to the feedback of the reinforcement learning network.
[0030] A reinforcement learning network model is constructed, comprising a state space, an action space, and a reward function. The state space vector includes the three-dimensional coordinates, velocity vector, and instantaneous power consumption of each node within the current and prediction time window. The action space is defined as a set of discrete actions including maintaining the original state, deceleration buffering, acceleration execution, and pause / hovering. The complete reward function is a scalar weighted combination of basic operational efficiency reward, energy consumption penalty, and the aforementioned dynamic safety penalty, with the specific weighting formula as follows: ,in This is the total reward value after a single step of execution. The basic operational efficiency bonus is calculated based on the reduction in process time. This is the energy consumption penalty calculated based on the difference between the device's rated power consumption and the actual power consumption. Let be the dynamic security penalty scalar that exhibits nonlinear exponential growth; , , These are the corresponding fixed weighting coefficients, preset to ensure absolute construction safety. , , This assigns the highest weight and priority to the security penalty; the reinforcement learning network is specifically trained using the proximal policy optimization algorithm, and when the average cumulative reward value fluctuation over 100 consecutive rounds is less than 1%, it is determined that the training has converged. After the dynamically generated resource conflict distribution matrix is flattened, it is used as the real-time state observation sequence of the current engineering environment and input into the input layer of the reinforcement learning network. The policy network and value network inside the network are invoked to traverse and calculate the expected reward value of all potential resource scheduling actions in the current state in real time based on the device spatial distribution characteristics and power consumption state data dynamically represented in the resource conflict distribution matrix, thereby completing the automated evaluation for different scheduling strategies. When the reinforcement learning network uses a virtual simulation environment to extrapolate the effect of the next action intervention, it dynamically extracts the expected motion velocity vector and direction parameters of each engineering equipment node within the virtual time step; it extrapolates the future three-dimensional coordinates of each equipment node based on a time-series prediction algorithm, and monitors the trajectory evolution curves of all dynamic nodes and static structures within a set future time window in real time; on each tiny time slice of trajectory extrapolation, it uses a spatial geometric distance algorithm to traverse and calculate the shortest straight-line distance between the physical bounding boxes of any two nodes, and extracts the minimum value within this extrapolation window as the minimum predicted spatial distance between the nodes. The collision avoidance boundary parameters set in the engineering safety management specifications are read in real time as a dynamically updated safety threshold. In each simulation time step, the minimum predicted spatial distance between nodes calculated in real time is compared with the safety threshold. A condition trigger is set. Once the minimum predicted spatial distance between nodes is detected to shrink to the safety threshold, the risk calculation module is immediately triggered. The safety threshold is used as the numerator and the minimum predicted spatial distance between nodes is used as the denominator for division, and the ratio reflecting the current local spatial compression degree is output. The underlying mathematical calculation engine is invoked to construct the exponential calculation formula: Penalty scalar = exp(safety threshold / (minimum prediction spatial spacing + ... )),in To prevent division by zero errors, a preset minimum constant is used, and a maximum penalty limit is set for the extreme state of physical interference with a prediction gap of 0. A dynamic safety penalty scalar with a nonlinear increase in value as the spatial approximation is obtained by solving. In the reverse parameter update stage of the reinforcement learning network, this dynamic safety penalty scalar is multiplied by -1 to convert it into a negative reward value, and then added to the overall reward feedback function of the current state action pair. The dynamic safety penalty scalar introduces a risk multiplier based on the equipment's kinetic potential difference for joint adjustment; after calculating the exponential operation result with the natural constant as the base, the relative motion velocity of the two nodes that generate spatial approximation and the instantaneous mass of the equipment are extracted in real time, and the kinetic potential difference between the two is calculated; using the logistic function, specifically the formula... ,in As a risk multiplier, For kinetic potential difference, To penalize the scaling upper limit parameter, To penalize the surge rate coefficient, This refers to the adjustment coefficient (i.e., the kinetic energy tolerance threshold); the specific methods for determining the values of each adjustment parameter are as follows: The preferred value range is 2.0-5.0, based on the pre-calibrated mechanical damage intensity level of the engineering equipment; The preferred value range is 0.1-0.5, which is used to control the sensitivity of the penalty term as kinetic energy increases; The specific value is dynamically determined based on the statistical mean of the maximum non-collision kinetic energy fluctuation in the historical safe operation database of similar equipment on site; the kinetic energy potential difference is mapped to a risk multiplier greater than 1 and multiplied with the original exponential result to generate the final dynamic safety penalty scalar; by introducing the equipment kinetic energy potential difference to construct the risk multiplier, a dual penalty scalar that integrates spatial approximation degree and mechanical damage potential is obtained, which realizes the ultimate suppression of the collision risk of high-energy-consuming and high-speed equipment.
[0031] By introducing trajectory extrapolation monitoring and an exponential penalty mechanism based on natural constants, a dynamic safety penalty scalar with nonlinear growth was obtained and incorporated as a negative reward into model training. This enabled the reinforcement learning search space to converge quickly under critical dangerous conditions and output a safer scheduling strategy, solving the problem of response lag caused by the inability to automatically calculate and analyze the optimal scheduling parameters based on the on-site situation.
[0032] The process of generating equipment control commands and scheduling parameters containing micro-delay peak-shaving parameters specifically includes: The resource scheduling strategy matrix is parsed to identify the corresponding concurrent devices, and the peak power consumption at startup of each concurrent device is extracted. Using a greedy algorithm, incremental delay trigger times are allocated to each concurrent device according to the power consumption distribution. The allocated incremental delay trigger times are used as peak shaving parameters and written into the device instructions to generate device control instructions and scheduling parameters containing micro-delay peak shaving parameters.
[0033] The resource scheduling strategy matrix is represented by device node numbers, with columns indicating future time steps. The elements within the matrix are action codes, such as 0 for hold, 1 for deceleration and buffering, and 2 for pause and hover. The system maps the extracted action codes to specific control messages that can be directly executed by the underlying PLC or servo controller through a preset mapping table that maps action codes to device control commands. The system receives the resource scheduling strategy matrix output by the reinforcement learning network in real time, calls the matrix parsing algorithm to scan the action codes and execution timestamp information in the matrix row by row, and filters out the device nodes whose time difference is less than the preset synchronous start tolerance range by dynamically comparing the planned start timestamps of each device in the instruction sequence, identifies and marks them as the corresponding concurrent devices, and links the device parameter database and historical dynamic operation logs at the engineering site to query across tables in real time according to the unique identifier of the device to extract the high current transient power consumption value generated by each concurrent device at the moment of initial power-on or load start-up, and extracts it as the start-up peak power consumption. The extracted peak power consumption of each concurrent device is used as the sorting weight index of the greedy algorithm. All concurrent devices are dynamically sorted in descending order of power consumption value to construct a real-time power consumption distribution sequence. The basic trigger time node is initialized, and the power consumption distribution sequence is traversed. Under the safety constraint of the instantaneous withstand threshold of the current power grid, a zero-delay initial trigger time is allocated to the highest-ranking high-power concurrent device. The power consumption superposition risk of adjacent devices is evaluated sequentially downwards along the descending sequence. At the same time, the cumulative superposition total power in the entire time domain period is checked in conjunction with the existing base load of the microgrid to see if it exceeds the limit. The greedy algorithm is used to select a local optimal strategy, and a delay step based on the transient response period of the power grid is added to each subsequent concurrent device in turn. The incremental delay trigger time is dynamically allocated to ensure that the total concurrent startup power consumption at any time does not exceed the limit. The incremental delay trigger time is dynamically interpolated based on the transient power recovery curve of the engineering microgrid. Specifically, when allocating the delay trigger time sequentially using a greedy algorithm, the transient power fluctuation characteristics of the microgrid after the previous concurrent device starts under load are read in real time. Using a spline interpolation algorithm, with discrete sampling timestamps as independent variables and the corresponding transient sampling values of grid voltage and frequency as dependent variables, the recovery curves of grid voltage and frequency are fitted, and the first derivative of the curve is changed from negative to positive, and the voltage and frequency re-enter the set tolerance range. The moment when the target time is continuously maintained is taken as the steady-state moment, and this moment is used as the dynamic delay step, which is accumulated into the trigger time of the next device. Based on the dynamic interpolation delay step of the power grid transient power recovery curve, a peak-shaving trigger sequence that perfectly matches the on-site power supply capacity is obtained, realizing the smooth soft start of multiple concurrent high-power devices. The spline interpolation algorithm specifically adopts cubic spline interpolation. The determination condition of the steady-state moment is: the first derivative of the fitted curve turns from negative to positive, and the power grid voltage and frequency remain within ±5% tolerance of the rated value for 3 consecutive seconds. The system calls the communication control protocol template of the underlying equipment at the engineering site, converts and encodes the dynamically allocated incremental delay trigger time into a data type compatible with the underlying controller's peak-shaving parameters, and positions the concurrent start control message originally generated by the positioning system. Using message reassembly and byte insertion technology, the determined delay trigger time is written into the control instructions of the corresponding devices, so that the concurrent devices start at different trigger times. Combined with other control information such as spatial location allocation parsed from the resource scheduling strategy matrix, the system packages and executes instruction compilation to generate device control instructions and scheduling parameters containing micro-delay peak-shaving parameters, which are ready to be issued.
[0034] By introducing time-series matrix analysis to identify concurrent devices and combining it with a greedy algorithm to dynamically allocate incremental delay triggering technology, device control commands and scheduling parameters containing micro-delay peak-shaving parameters were obtained. This achieved the effect of smoothing the power load curve at the engineering site and completing the seamless transition from front-end early warning to back-end control. It solved the problem of relying heavily on human experience for command and intervention when dealing with resource conflicts and the inability to automatically issue optimal device servo control commands and scheduling parameters.
[0035] After generating the device control commands and scheduling parameters containing the micro-delay peak-shaving parameters, the process further includes: The equipment control commands and scheduling parameters containing micro-delay peak-shaving parameters are sent to the corresponding equipment at the engineering site for control; the latest equipment timing operation data after control is collected and synchronously updated to the initial digital twin map to form a closed loop.
[0036] By invoking the communication interface between the IoT gateway and the industrial control bus, and based on the TLS / SSL network security transmission protocol or the AES symmetric encryption algorithm, a secure transmission encrypted channel is established between the control system and the programmable logic controllers of each device on the engineering site. Through this encrypted communication channel, the generated device control commands and scheduling parameters containing micro-delay peak-shaving parameters are distributed and routed in real time to the corresponding target device nodes according to the underlying network transmission protocol. After receiving the commands, the controller of the target device on the site calls the built-in parsing program to unpack the data frame, extract the trigger time node, and drive the underlying relays or servo drivers to execute actions sequentially according to the allocated micro-delay timing sequence, thereby completing the direct automated intervention and control of the physical equipment on the engineering site. At the synchronous moment when the target equipment on site performs scheduling and control actions, the high-frequency sensor array and data acquisition terminal mounted on the equipment are woken up in real time; through streaming data subscription and acquisition technology, the dynamic current waveform, spatial position displacement and real-time power consumption changes of the equipment during the staggered start-up and subsequent operation are continuously monitored, the latest physical characteristic indicators generated after control intervention are dynamically captured, packaged and integrated into the latest equipment time sequence operation data according to the timestamp order and uploaded to the management and control platform; The control platform receives the latest equipment time-series operation data in real time and uses the parsed unique device identifier to lock the corresponding digital twin node in the virtual space in real time. It calls the data mapping model, specifically adopting a direct mapping rule based on the precise alignment of device attribute fields and timestamps, to replace the historical attribute values in the twin node with the latest operation status parameters, and triggers a real-time status refresh of the graph topology. Specifically, based on the latest spatial position displacement contained in the latest equipment time-series operation data, it recalculates the real-time physical spatial distance between the corresponding digital twin nodes, and synchronously modifies the edge weights of the spatial adjacency matrix in the initial digital twin graph according to the updated physical spatial distance, completing the synchronous update of the initial digital twin graph. The updated graph data then serves as the input source for a new round of time-series feature extraction, forming a continuously iterative control loop between the physical site status and virtual space decision-making.
[0037] Example 2: This embodiment provides an overall application and implementation process of a digital twin-based intelligent engineering management and control method and system in a specific engineering scenario. Large-scale intelligent super high-rise building construction sites typically feature narrow working areas at heights, frequent overlapping operations of multiple large luffing-arm tower cranes, and frequent start-ups and shutdowns of high-power concrete pump trucks and material hoists.
[0038] Reference Figure 3During the full-scale construction phase of the super high-rise building, multiple large luffing jib tower cranes (Tower Crane A, Tower Crane B, etc.), multiple construction elevators, and several automated ground material transfer vehicles were deployed on site. The map generation and extraction module was activated first, acquiring real-time equipment timing data (tower crane slewing angle, lifting height, luffing trolley position, instantaneous hook load, and current waveforms of various motors, etc.) and process-related topology data of the aforementioned on-site equipment through an IoT gateway. This module generated an initial digital twin map of the construction site in the system's virtual digital space, mapping entities such as Tower Crane A and Tower Crane B into a high-dimensional node timing feature matrix. Simultaneously, considering the strong process dependency relationship where Tower Crane A provides pre-assembly materials for Tower Crane B, the module resolved this as directed edges, constructing a spatial adjacency matrix containing all on-site scheduling node spaces and process coupling relationships, thus enabling a precise quantitative expression of the complex on-site operational status in the virtual dimension.
[0039] As the main structure grows taller, the slewing radii of tower cranes A and B overlap significantly on the high-altitude work surface. The system uses a graph convolutional mask modulation module to predict the evolution of this dynamic situation. This module not only receives the aforementioned feature matrix but also extracts in real time the physical spatial distance between the current boom tips of tower cranes A and B, as well as the intensity of their process dependency. Based on the shortening of the physical spatial distance and the peak of the process dependency intensity, the module automatically calculates an extremely high attention score for the "tower crane A-tower crane B" node pair and generates a mask matrix which is multiplied by the original spatial adjacency matrix. During the feature aggregation process of the graph convolutional network, this mask modulation mechanism forcibly amplifies the feature weights of spatial collisions and process delays between tower cranes A and B, successfully outputting a sequence of resource node evolution states that includes the evolution trend of the motion trajectories of multiple tower cranes within the future prediction time window.
[0040] After acquiring the evolutionary state sequence of tower cranes and material handling equipment, the difference clustering module connects in real time and calculates the difference between the predicted state values and the system's preset engineering scheduling constraint thresholds. The system detects that at a certain predicted time in the future, the predicted distance between the luffing trolley of tower crane A and the hoisting wire rope of tower crane B will shrink to a safe critical range, generating an out-of-bounds residual variable. The module extracts the three-dimensional spatial coordinate set corresponding to this out-of-bounds residual variable and performs density-based spatial clustering. During the clustering process, the system does not use a fixed search radius, but adaptively constructs a dynamically expanding equipment bounding box based on the swing speed of the tower crane hook and the size of the giant steel structure. Through this dynamic density clustering, the system accurately divides the slewing area of tower crane A, the hoisting area of tower crane B, and the waiting area of the AGV directly below into the same spatial conflict cluster, and outputs a "resource conflict distribution matrix" containing the conflict range and the expected power consumption of each device at this time, providing a precise target for subsequent strategic intervention.
[0041] Upon receiving the resource conflict distribution matrix, the reinforcement learning evaluation module rapidly initiated a self-game of scheduling strategies within the background simulation engine. This module attempted various interventions. When evaluating the strategy of "crane A maintaining its original speed while crane B accelerates and decelerates," system trajectory extrapolation revealed that the minimum predicted spatial distance between the two cranes was extremely close to the safety threshold within a tiny time slice. At this point, the underlying conditional trigger responded instantly. The system, using the natural constant e as the base and the ratio of the safety threshold to the predicted distance as the exponent, triggered a non-linearly exponentially growing "dynamic safety penalty scalar." The system, considering the enormous kinetic potential difference generated by the hook carrying the giant steel component, transformed this exponential penalty into an extremely large negative reward value, directly feeding it back to the reinforcement learning network. Under this stringent extreme risk suppression, the reinforcement learning network quickly abandoned this high-risk strategy, ultimately converging rapidly and generating the optimal resource scheduling strategy matrix: instructing crane A to pre-set its angle and begin deceleration, while simultaneously instructing crane B to pause its lifting and remain hovered, resolving this complex spatial pressure crisis in the safest and smoothest way.
[0042] After the reinforcement learning network outputs the aforementioned safe scheduling actions, the execution of commands at the physical site faces another potential problem. The system analyzes the resource scheduling strategy matrix and determines that tower crane A needs to restart its slewing motor after deceleration and buffering. Simultaneously, multiple high-power construction elevators on the ground receive a full-load material lifting signal, leading the system to classify these devices as "concurrent devices." The peak-shifting command generation module intervenes, extracting the peak power consumption of the tower crane's slewing motor and the construction elevators' main units, and constructing a descending sequence. Using a greedy algorithm, the system allocates zero-delay active action time to tower crane A while dynamically allocating incrementally delayed trigger times to the multiple construction elevators, ensuring the stability of the microgrid's frequency and voltage recovery curves. Subsequently, the system encapsulates these micro-delay peak-shifting parameters into the underlying device control command messages.
[0043] Finally, the system directly sends composite scheduling commands carrying micro-delay peak-shaving parameters and deceleration avoidance parameters to the PLCs of the tower cranes and the servo controllers of the construction elevators on site via an AES-secure encrypted IoT channel. Without human intervention, the tower cranes and elevators automatically complete peak-shaving startup and space avoidance according to the command sequence. The sensor array then collects and uploads the stable current waveform after intervention and the new spatial coordinates after the crisis is resolved, synchronously updating the initial digital twin map, perfectly forming a closed loop of "prediction-decision-peak-shaving control-status refresh" without human intervention.
[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A digital twin-based intelligent control method for engineering management, characterized in that, include: Acquire the equipment timing operation data and process association topology data at the engineering site, generate an initial digital twin map and extract the node timing feature matrix and spatial adjacency matrix; Evolutionary features are extracted from the temporal feature matrix and spatial adjacency matrix of the nodes based on graph convolutional networks. Attention scores are constructed using physical spatial distance and process dependency strength, and the spatial adjacency matrix is masked and modulated to amplify the aggregation weight of nodes with attention scores higher than a preset threshold. The difference between the evolutionary state sequence of the resource nodes and the preset engineering scheduling constraint threshold is calculated, and the out-of-bounds residual variables are extracted and density clustered according to the spatial coordinate set to generate a resource conflict distribution matrix that includes spatial conflicts and equipment power consumption status. Based on a reinforcement learning network, a strategy evaluation is performed on the resource conflict distribution matrix. When the minimum prediction spatial distance between nodes is less than or equal to a preset safety threshold, a nonlinearly exponentially growing dynamic safety penalty scalar is triggered as a negative reward to generate a resource scheduling strategy matrix. The resource scheduling strategy matrix is analyzed to extract the peak power consumption of concurrent devices at startup. Based on a greedy algorithm, an incremental delay trigger time is allocated to generate device control instructions and scheduling parameters containing micro-delay peak-shifting parameters.
2. The intelligent control method for engineering management based on digital twins according to claim 1, characterized in that, The process of generating an initial digital twin graph and extracting the node temporal feature matrix and spatial adjacency matrix includes: The device timing operation data is mapped to node attribute information in the initial digital twin graph, and the node timing feature matrix is integrated and constructed. The topological relationships in the process association topology data are parsed and mapped to node edges in the initial digital twin graph to construct a spatial adjacency matrix.
3. The intelligent control method for engineering management based on digital twins according to claim 1, characterized in that, The process of masking the spatial adjacency matrix by constructing attention scores using physical spatial distance and process dependency strength includes: Extract the physical spatial distance and process dependency strength between each node in the initial digital twin graph; perform a weighted calculation on the physical spatial distance and the process dependency strength to obtain the attention score between each node; generate a mask matrix based on the attention score, and multiply the mask matrix with the spatial adjacency matrix.
4. The intelligent control method for engineering management based on digital twins according to claim 1, characterized in that, The process of generating the resource conflict distribution matrix specifically includes: The difference between the state values in the evolution state sequence of the resource nodes and the engineering scheduling constraint threshold is calculated, and the part of the difference exceeding the threshold is extracted as an out-of-bounds residual variable; the spatial coordinate set of the node corresponding to the out-of-bounds residual variable in the initial digital twin map is obtained; density clustering is performed on the out-of-bounds residual variable based on the spatial coordinate set, and a resource conflict distribution matrix containing spatial conflict and device power consumption status is output according to the clustering result.
5. The intelligent control method for engineering management based on digital twins according to claim 4, characterized in that, The process of evaluating the resource conflict distribution matrix based on a reinforcement learning network specifically includes: The resource conflict distribution matrix is input into the reinforcement learning network to perform policy evaluation; during the evaluation process, node trajectories are monitored in real time to calculate the minimum prediction spatial distance between nodes; when the minimum prediction spatial distance between nodes is less than or equal to the safety threshold, the ratio of the safety threshold to the minimum prediction spatial distance between nodes is calculated; using the natural constant as the base and the ratio as the exponent, an exponential operation is performed to obtain a dynamic safety penalty scalar that grows non-linearly, and the dynamic safety penalty scalar is added as a negative reward to the feedback of the reinforcement learning network.
6. The intelligent control method for engineering management based on digital twins according to claim 1, characterized in that, The process of generating equipment control commands and scheduling parameters containing micro-delay peak-shaving parameters specifically includes: The resource scheduling strategy matrix is parsed to identify the corresponding concurrent devices, and the peak power consumption at startup of each concurrent device is extracted. Using a greedy algorithm, incremental delay trigger times are allocated to each concurrent device according to the power consumption distribution. The allocated incremental delay trigger times are used as peak shaving parameters and written into the device instructions to generate device control instructions and scheduling parameters containing micro-delay peak shaving parameters.
7. The intelligent control method for engineering management based on digital twins according to claim 6, characterized in that, After generating the device control commands and scheduling parameters containing the micro-delay peak-shaving parameters, the process further includes: The equipment control commands and scheduling parameters containing micro-delay peak-shaving parameters are sent to the corresponding equipment at the engineering site for control; the latest equipment timing operation data after control is collected and synchronously updated to the initial digital twin map to form a closed loop.
8. A digital twin-based intelligent control system for engineering management, characterized in that, include: The graph generation and extraction module, graph convolution mask modulation module, difference clustering module, reinforcement learning evaluation module, and peak-shifting instruction generation module are connected in sequence. The graph generation and extraction module acquires the equipment time-series operation data and process-related topology data at the engineering site, generates an initial digital twin graph, and extracts the node time-series feature matrix and spatial adjacency matrix from it. Graph Convolutional Mask Modulation Module: The output of the graph generation and extraction module is connected to the input of the graph convolutional mask modulation module; Based on the graph convolutional network, the module extracts evolutionary features from the node temporal feature matrix and spatial adjacency matrix, constructs attention scores using physical spatial distance and process dependency strength, performs mask modulation on the spatial adjacency matrix, amplifies the aggregation weight of nodes with attention scores higher than a preset threshold, and generates a resource node evolution state sequence. Differential Clustering Module: The output of the graph convolutional mask modulation module is connected to the input of the differential clustering module; the difference between the resource node evolution state sequence and the preset engineering scheduling constraint threshold is calculated, the out-of-bounds residual variable is extracted and density clustered according to the spatial coordinate set to generate a resource conflict distribution matrix; Reinforcement learning evaluation module: The output of the difference clustering module is connected to the input of the reinforcement learning evaluation module; the resource conflict distribution matrix is evaluated based on the reinforcement learning network. When the minimum prediction spatial distance between nodes is less than or equal to the preset safety threshold, a nonlinear exponentially growing dynamic safety penalty scalar is triggered as a negative reward to generate a resource scheduling strategy matrix. Peak shaving instruction generation module: The output of the reinforcement learning evaluation module is connected to the input of the peak shaving instruction generation module; it parses the resource scheduling strategy matrix, extracts the peak power consumption of concurrent devices, allocates incremental delay trigger times based on a greedy algorithm, and generates device control instructions and scheduling parameters containing micro-delay peak shaving parameters; Furthermore, the output of the peak-shifting instruction generation module or the acquisition module is connected to the update input of the map generation and extraction module, forming a closed loop.
