Bridge construction digital simulation collaboration method and system

Abstract

The application provides a bridge construction digital simulation cooperation method and system, comprising: acquiring real-time construction process data, using a preset construction state monitor to perform early abnormality identification, acquiring an early abnormality data set, using a pre-constructed candidate construction regulation strategy generation network to analyze and decide the early abnormality data set, generating a candidate construction regulation strategy set, performing cross-scale strategy transmission and adaptive adjustment on the candidate construction regulation strategy set based on a preset multi-scale construction mode association network, generating an initial construction regulation strategy set, using a construction site regulation model based on a cellular automaton to simulate and optimize the initial construction regulation strategy set, generating a reinforced construction regulation strategy, and finally realizing efficient and accurate regulation of the bridge construction process.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and more specifically, to a digital simulation and collaborative method and system for bridge construction. Background Technology

[0002] The construction of modern large-scale bridge projects is a complex system engineering project involving multiple disciplines, multiple stakeholders, long cycles, and complex environments. Under the wave of digitalization, although building information modeling, the Internet of Things, and project management software have been widely used, the current technology system still has certain limitations in dealing with uncertainties in the construction process and in achieving cross-level collaboration and forward-looking decision-making.

[0003] On the one hand, traditional methods mostly focus on threshold alarms for single physical parameters or rely on static comparisons of schedules. Various data generated during construction are often isolated and lack effective means of fusion and analysis, resulting in a one-sided and reactive monitoring of construction anomalies. On the other hand, when faced with construction anomalies or design changes, traditional methods mostly rely on the personal experience and ad-hoc coordination of the responsible persons of each participating party. The decision-making process often revolves around a single problem and lacks consideration of multiple perspectives, resulting in an incomplete construction control decision. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a collaborative method for digital simulation of bridge construction, the method comprising:

[0005] Step A1: Obtain real-time construction process data, use a preset construction status monitor to identify early anomalies, and obtain an early anomaly dataset;

[0006] Step A2: The early abnormal dataset is analyzed and decided upon using a pre-constructed candidate construction control strategy generation network to generate a candidate construction control strategy set;

[0007] Step A3: Based on the preset multi-scale construction mode association network, perform cross-scale strategy transfer and adaptive adjustment on the candidate construction control strategy set to generate the initial construction control strategy set;

[0008] Step A4: Use a cellular automata-based construction site control model to simulate and optimize the initial construction control strategy set, and generate enhanced construction control strategies.

[0009] As a further aspect of the present invention, real-time construction process data is obtained, and an early anomaly identification is performed using a preset construction status monitor to obtain an early anomaly dataset, including:

[0010] The real-time construction process data includes at least physical status data, construction progress data, resource status data, and construction environment data;

[0011] The manifold learning algorithm is used to map real-time construction process data to the manifold space corresponding to the construction status monitor, thereby generating real-time status nodes.

[0012] Connect the real-time status nodes with the corresponding historical status nodes to generate an actual construction trajectory sequence;

[0013] The actual construction trajectory sequence is input into a preset spatiotemporal graph convolutional neural network for trajectory prediction, generating a predicted construction trajectory sequence.

[0014] The predicted construction trajectory sequence is compared and analyzed with the healthy construction mode cluster contained in the manifold space to obtain the corresponding comparison and analysis results.

[0015] If the comparison and analysis results meet the preset warning conditions, early anomaly detection will be performed, and a corresponding early anomaly dataset will be generated.

[0016] The early anomaly dataset contains at least the anomaly patterns and the contribution distribution of the anomaly pattern causes.

[0017] As a further aspect of the present invention, the method further includes:

[0018] A high-dimensional construction phase space is constructed based on historical construction process data. A manifold learning algorithm is used to perform nonlinear dimensionality reduction on the high-dimensional construction phase space to generate the corresponding manifold space.

[0019] The manifold contains multiple health pattern clusters, which are represented as continuous clusters formed by the aggregation of state nodes corresponding to construction stages in which no anomalies have occurred in historical data.

[0020] The spatiotemporal graph convolutional neural network is used to store the dependencies between data in various dimensions within historical or real-time construction process data, and predicts the probability distribution of state trajectories at multiple future time steps based on the input real-time construction trajectory sequence, generating a corresponding predicted construction trajectory sequence.

[0021] As a further aspect of the present invention, a pre-constructed candidate construction control strategy generation network is used to analyze and make decisions on the early abnormal dataset to generate a candidate construction control strategy set, including:

[0022] The candidate construction control strategy generation network performs counterfactual intervention on the causal variables that lead to abnormal patterns based on early abnormal datasets, generating a counterfactual intervention event set;

[0023] The candidate construction control strategy generation network takes each counterfactual intervention event in the counterfactual intervention event set as the game issue and real-time construction process data as the state space.

[0024] Each game agent combines its corresponding action space and payoff function to engage in multiple rounds of game to obtain a consensus action sequence.

[0025] The simulation executes the consensus action sequence, obtains the corresponding reward vector, and merges all the reward vectors with the corresponding consensus action sequence into a set of policy reward pairs.

[0026] The set of strategy payoff pairs is subjected to refined Bayesian equilibrium and Pareto front analysis to generate a set of candidate construction control strategies.

[0027] As a further aspect of the present invention, the candidate construction control strategy generation network includes:

[0028] The candidate construction control strategy generation network consists of multiple game agents, each of which comprises a strategy network and a value network.

[0029] A course learning strategy is used to train the candidate construction control strategy generation network, and a combination of inverse reinforcement learning and adversarial learning is used simultaneously to infer the implicit reward function contained in the training dataset.

[0030] The parameters of the candidate construction control strategy generation network are fine-tuned based on the implicit reward function, and a causal discovery algorithm is used during the training process to construct a construction constraint graph.

[0031] The construction constraint graph is used to constrain the game actions generated by the game agent to ensure that they do not violate the corresponding domain prior knowledge.

[0032] As a further aspect of the present invention, a preliminary construction control strategy set is generated by performing cross-scale strategy transfer and adaptive adjustment on the candidate construction control strategy set based on a preset multi-scale construction mode association network, including:

[0033] The multi-scale construction mode association network is represented as a multi-layer Bayesian network containing macro-level, meso-level, and micro-level layers.

[0034] Among them, the macro layer is used to characterize variables at the construction project level, the meso layer is used to characterize variables at the engineering part level, and the micro layer is used to describe variables at the component or process level.

[0035] The candidate construction control strategy set is mapped to macro-level nodes in a multi-scale construction mode association network to generate corresponding construction intervention nodes.

[0036] The propagation algorithm is run starting from the construction intervention node to obtain the corresponding set of intervention targets;

[0037] The set of intervention targets is used to describe the specific variables and their target values ​​that need to be adjusted at the meso- and micro-levels.

[0038] A corresponding initial construction control strategy set is generated based on the set of intervention targets.

[0039] As a further aspect of the present invention, a construction site control model based on cellular automata is used to simulate and optimize the initial construction control strategy set, generating an enhanced construction control strategy, including:

[0040] The construction site control model is represented by discretizing the construction site into a spatiotemporal cellular automaton model based on cellular automata, where each cell of the model represents a spatial unit.

[0041] The status attributes of the spatial unit include at least construction progress data, resource status data, and construction environment data;

[0042] The initial construction control strategy set is injected into the construction site control model to initialize the state of each cell.

[0043] The schedulable execution units at the construction site are modeled as scheduling agents. Each scheduling agent combines its corresponding set of cells with the real-time acquired actual disturbance signals to perform strategy simulation execution.

[0044] Based on the simulation results, the parameters of the initial construction control strategy set are fine-tuned to generate an enhanced construction control strategy.

[0045] Furthermore, embodiments of the present invention also provide a digital simulation and collaborative system for bridge construction, comprising:

[0046] An anomaly monitoring module is used to acquire real-time construction process data and, in conjunction with a preset construction status monitor, to identify early anomalies and acquire an early anomaly dataset.

[0047] The strategy generation module is used to analyze and make decisions on the early abnormal dataset based on the candidate construction control strategy generation network, and generate a candidate construction control strategy set.

[0048] The strategy conversion module is used to adaptively adjust the candidate construction control strategy set according to the multi-scale construction mode association network to generate an initial construction control strategy set.

[0049] The strategy optimization module is used to simulate and optimize the initial construction control strategy set using a construction site control model, and generate enhanced construction control strategies.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] By acquiring real-time construction process data and using a pre-set construction status monitor for early anomaly identification, an early anomaly dataset is obtained. This step combines manifold learning and spatiotemporal graph neural networks to achieve probabilistic prediction of the construction status evolution trend, thereby improving the efficiency and accuracy of construction anomaly identification.

[0052] The early abnormal dataset is analyzed and decided by a pre-constructed candidate construction control strategy generation network to generate a set of candidate construction control strategies. This step generates a series of candidate strategies that have undergone conflict simulation and benefit balancing through multi-party game simulation and counterfactual inference, thereby improving the technical feasibility and implementation acceptability of the final generated strategy.

[0053] Based on a pre-defined multi-scale construction mode association network, the candidate construction control strategy set is transmitted and adaptively adjusted across scales to generate an initial construction control strategy set. This step reduces the contradictions between drawings and instructions caused by information transmission deviations or different understandings among different disciplines by performing causal mining and cross-scale model synchronization, further improving the reliability of the final generated strategy.

[0054] A cellular automata-based construction site control model is used to simulate and optimize the initial construction control strategy set, generating an enhanced construction control strategy. This step, through dynamic simulation and multi-agent coordination based on cellular automata, realizes the simulation of real-time resource flow and process interaction during construction, further enhancing the reliability and executability of the enhanced construction control strategy. Attached Figure Description

[0055] Figure 1 This is a flowchart of the steps of a digital simulation and collaborative method for bridge construction according to the present invention;

[0056] Figure 2 This is a schematic diagram of a digital simulation and collaborative system for bridge construction according to the present invention. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart of the steps of a digital simulation and collaborative method for bridge construction according to the present invention. The following is a detailed description of this digital simulation and collaborative method for bridge construction.

[0058] Step A1: Obtain real-time construction process data, use a preset construction status monitor to identify early anomalies, and obtain an early anomaly dataset.

[0059] Specifically, a manifold learning algorithm is used to map real-time construction process data to the manifold space corresponding to the construction status monitor, generating real-time status nodes. These real-time status nodes are then connected to corresponding historical status nodes to generate an actual construction trajectory sequence. This actual construction trajectory sequence is input into a preset spatiotemporal graph convolutional neural network for trajectory prediction, generating a predicted construction trajectory sequence. The predicted construction trajectory sequence is then compared and analyzed with the healthy construction mode clusters contained in the manifold space to obtain corresponding comparison and analysis results. If the comparison and analysis results meet preset warning conditions, early anomaly detection is performed, generating a corresponding early anomaly dataset. The early anomaly dataset contains at least anomaly patterns and the contribution distribution of anomaly pattern causes.

[0060] Understandably, the real-time construction process data includes at least physical state data, construction progress data, resource status data, and construction environment data. Specifically, the following four types of real-time construction data are collected in real time from three dimensions: construction site, management process, and natural environment. The first type is physical state data, which comes from sensor networks installed on key structural parts such as the main beam control section, cable tower anchorage area, and large temporary structures. It includes at least time-series signals such as strain, stress, displacement, tilt angle, vibration frequency, and acceleration. This type of data is used to characterize the structural mechanical and geometric features during bridge construction.

[0061] The second category is construction progress data, which is dynamically extracted from project management software such as BIM 4D / 5D platforms to reflect the project's progress over time. This data includes at least the planned start / end time, actual start / end time, percentage of completion, and changes in total float time and free float time calculated based on the critical path method for all processes.

[0062] The third category is resource status data, which is used to characterize the status of construction material resources, human resources, and equipment resources. The data characterizing the status of construction material resources includes at least the inventory, consumption rate, and arrival and inspection records of key materials such as prestressed steel strands, high-strength bolts, and special concrete. The data characterizing the status of human resources includes at least the number of registered construction workers for each trade, attendance rate, and working hours records. The data characterizing the status of equipment resources includes at least the operating status, working hours, fault codes, maintenance records, and real-time location information of large key equipment such as bridge erecting machines, large hoisting equipment, and concrete mixing plants.

[0063] The fourth category is construction environment data, which includes at least temperature, humidity, wind speed, wind direction, rainfall, atmospheric pressure, as well as indicators such as noise and dust.

[0064] After performing data preprocessing on the above four types of data, including at least data cleaning, missing value imputation, and spatiotemporal alignment, the data is parsed and aligned according to a preset data pattern. This standardizes the raw data of different frequencies and units into dimensionless feature vectors with the same timestamp. Specifically, a data pattern with dimensions, a reasonable range, and data characteristics is predefined for each type of state variable. For continuous variables such as physical monitoring data, Z-score standardization based on historical health data statistics is used, i.e., (real-time measurement value - historical mean) / historical standard deviation, converting the readings into dimensionless values ​​with a mean of 0 and a standard deviation of 1. For example, based on the historical mean of 20 MPa and the standard deviation of 5 MPa for a certain cross-section stress value, "+25.6 MPa" is converted to "+1.12". For progress-related proportional variables such as the percentage of process completion, the original percentage is directly used. For resource-related counts or intensity variables such as equipment utilization, maximum-minimum normalization is used to compress them to the [0,1] interval. For example, if the current utilization rate of a certain equipment is 75%, and its preset theoretical maximum utilization rate is 90%, then the standardized value is 0.83. For environmental variables, corresponding standardization processing is performed according to their influence patterns. For example, temperature is standardized using deviation based on standard temperature, and rainfall is mapped using piecewise functions such as 0 for no rain, 0.3 for light rain, and 0.7 for moderate rain. All variables processed by the above data model are then timestamped and concatenated into a multidimensional feature vector in a predefined order. This feature vector is defined as the actual state vector, and the actual state vector is mapped to a high-dimensional construction phase space to form a data point representing the global state of the construction process at a specific moment.

[0065] It should be noted that the construction of the manifold space is represented by constructing a high-dimensional construction phase space based on historical construction process data, and using a manifold learning algorithm to perform nonlinear dimensionality reduction on the high-dimensional construction phase space to generate the corresponding manifold space. The manifold space contains multiple healthy mode clusters, which are represented by continuous clusters formed by the aggregation of state nodes corresponding to construction stages that have not shown abnormalities in historical data.

[0066] Understandably, since the high-dimensional construction phase space contains a large amount of noise, redundancy, and nonlinear coupling relationships, the complexity of direct analysis within this phase space is quite high. Therefore, nonlinear manifold learning technology is introduced. By assuming that the construction process data is dominated by only a few key endogenous or exogenous driving factors, the high-dimensional construction phase space is compressed into a low-dimensional manifold space.

[0067] Specifically, utilizing a large amount of full-process data from historical successful projects and digital construction plan simulation data corresponding to this construction project, an unsupervised learning strategy is adopted to set the hyperparameters of the Uniform Manifold Approximation and Projection (UMAP) manifold model. For example, the number of neighboring points is set between 15 and 50 to accommodate different degrees of local clustering in the construction state; the minimum distance is set between 0.1 and 0.5 to ensure that healthy patterns form distinguishable clusters; and the target dimension is fixed at 2 or 3 for easy visualization and interpretation. In the high-dimensional construction phase space, a nearest-neighbor-based fuzzy topological representation is constructed for each data point, where the edge weights are defined by conditional probabilities, i.e., for a point... and High-dimensional probability Represented as yes The possibility of neighbors, and ,in for and The distance between, It is the distance to the nearest neighbor. For example, assuming there exists a scale parameter in the high-dimensional construction phase space. Three points, assuming ,for Its nearest neighbor is The nearest neighbor is The nearest neighbor is ,but The nearest neighbor distances for the three points are 0.173, 0.173, and 1.056, respectively. Then, for each point, a binary search is used to determine its corresponding scale parameter. For example, by solving The corresponding scale parameter is obtained, and the corresponding probability value is calculated based on the obtained parameter; in the low-dimensional manifold space, the corresponding probability is... ,in The coordinates are in the low-dimensional manifold space. Preset hyperparameters; to minimize The cross-entropy loss between the two distributions is used as the training objective, and the algorithm iteratively adjusts the coordinates in the low-dimensional manifold space by employing stochastic gradient descent. This minimizes the loss value; training ends when the maximum number of training rounds is reached, or when the absolute value of the rate of change of the loss function value for multiple consecutive rounds is lower than a preset threshold.

[0068] The trained Uniform Manifold Approximation and Projection (UMAP) manifold model can naturally form several compact and continuous healthy mode clusters after historical healthy construction states are mapped to a low-dimensional space. For example, there are clusters such as the stable state of standard segment prefabrication hoisting, the symmetrical equilibrium state of cantilever casting, and the resource buffer scheduling state during the rainy season. The shape and distribution of each cluster reflect the dynamic characteristics of the construction system under this healthy mode. The low-dimensional space containing multiple healthy mode clusters is the manifold space. An incremental manifold learning strategy is adopted, that is, a dynamic update time window or data volume threshold is set. When the amount of newly flowing real-time data reaches the preset threshold, the local micro-update or global retraining of the UMAP manifold model is triggered to cope with the mode drift caused by stage transition or process change during construction.

[0069] As one possible implementation, firstly, the actual state vector They are simultaneously fed into the following two parallel processing channels: In channel one, the current time... Actual state vector As input, a nonlinear dimensionality reduction mapping is performed based on the pre-trained UMAP manifold model to generate real-time state nodes. Simultaneously obtain its coordinates in the manifold space And match the real-time status node with the historical mapping point sequence. Connect to form the actual state trajectory In passage two, With the previous consecutive The state sequence at each time step Connect them to form a spatiotemporal state diagram. Spatiotemporal graph convolutional neural networks with Input, output future State prediction sequence at each time step The predicted state vector in the state prediction sequence A predicted state trajectory is obtained by sequentially mapping through the UMAP manifold model. For example, assuming the current time t is "20xx-0x-20 10:00:00", data such as main beam stress, tower crane utilization rate, and concrete pouring progress are collected and formed into a vector. Assuming a time step is 5 minutes long, the channel first maps it to a point in the manifold space. and compare that point with , Connecting the points forms the actual trajectory line; channel two is based on... Based on the data from the previous 11 time steps, the state for the next 30 minutes, i.e., K=6, is predicted, and 6 prediction points are mapped to form a prediction trajectory line pointing to the future.

[0070] It should be noted that the spatiotemporal graph convolutional neural network is used to store the dependencies between data in various dimensions within historical or real-time construction process data, and predicts the probability distribution of state trajectories for multiple future time steps based on the input real-time construction trajectory sequence, generating the corresponding predicted construction trajectory sequence. Specifically, firstly, based on domain prior knowledge, construction objects such as components and work surfaces, and resource entities such as equipment and work teams are abstracted into entity nodes. The attribute feature vector corresponding to each node is its corresponding multi-dimensional state data, such as stress, progress, and resource intensity.

[0071] Next, the connection edges between nodes are established in the following three ways: First, structural edges based on physical topology, such as the membership relationship between adjacent components, support connections, and equipment and working surfaces; second, temporal edges based on process logic, which can be determined by the work breakdown structure and network plan diagram, such as the predecessor-successor relationship; and third, data-driven influence edges, which can be identified by calculating the mutual information between different node state sequences in historical data or by using causal discovery algorithms to identify node pairs with dependencies or causal relationships, such as finding a strong correlation edge between the "ambient temperature node" and the "concrete component stress node", thereby generating a spatiotemporal graph structure to reflect the internal relationships of the construction system.

[0072] Next, the spatiotemporal graph convolutional neural network uses the aforementioned spatiotemporal graph structure as its backbone, and incorporates the spatiotemporal state graphs of the past T consecutive time steps. As input, after multiple layers of spatiotemporal convolution and pooling operations, the output is a complete graph state prediction for the next K time steps. The multiple layers of spatiotemporal convolution include spatial convolutional layers and temporal convolutional layers. The spatial convolutional layers use Graph Attention Network (GAT) or Chebyshev Graph Convolution to aggregate information from neighboring nodes at each time step, thereby capturing spatial interdependencies. The temporal convolutional layers can use Gated Recurrent Unit (GRU) or LSTM to extract temporal features from the node's own time series and capture dynamic evolution patterns.

[0073] It should be added that the training data for the spatiotemporal graph convolutional neural network comes from the full-process data of historical successful projects and the digital construction plan simulation data corresponding to this construction project. The training employs a quantile regression strategy. The last layer of the spatiotemporal graph convolutional neural network outputs predicted values ​​for three specific quantiles in parallel, such as the 0.1, 0.5, and 0.9 quantiles, thus directly obtaining the prediction interval. The loss function can be expressed as... ,in, This is represented as quantile loss, used to ensure the accuracy of predicted quantiles; For graph Laplace regularization, it is used to encourage smooth and continuous evolution of neighboring nodes in the hidden state space; is the corresponding regularization coefficient; the training process adopts a joint strategy combining mini-batch stochastic gradient descent, Adam optimizer and learning rate decay, with the goal of minimizing the total loss on the validation set, until the total loss no longer decreases within a certain number of training steps or reaches the preset maximum number of training rounds, at which point the training ends.

[0074] calculate The nearest cluster of healthy patterns in the manifold space The geometric relationships, specifically, are calculated using the dynamic time warping algorithm. and Shape distance between two trajectories Simultaneously calculate the predicted trajectory endpoint. To the cluster center Mahalanobis distance ,right and These two distances are used to calculate the first-order difference (velocity) and the second-order difference (acceleration) within a time window. and If the velocity and acceleration corresponding to any indicator continuously exceed a preset threshold, it is judged as a trend deviation; at the same time, the width of the 95% confidence interval of the spatiotemporal graph convolutional neural network for the predicted values ​​of key variables is monitored. Real-time calculation The rate of change relative to its recent baseline level ,like If the threshold for sudden occurrence is exceeded, it is determined to be an uncertain anomaly.

[0075] For example, the most relevant health pattern cluster corresponding to a certain predicted trajectory is the standard cantilever casting steady state. Calculations show the shape and distance of the predicted trajectory over the next 30 minutes. The pressure is increasing at an acceleration of 0.1 units per minute, and the confidence interval width of the predicted stress variable at the root of the main beam suddenly expands by 50% within 10 minutes. At the same time, the corresponding thresholds for trend deviation and uncertainty anomaly are triggered. At this time, a warning message for trend anomaly is generated, indicating that the current construction status is accelerating away from the stable mode and the reliability of the system prediction is decreasing.

[0076] It should be noted that the above The nearest cluster of healthy patterns in the manifold space The discrimination method is expressed as calculating the end point of the predicted trajectory. to cluster Mahalanobis distance of multivariate Gaussian distribution This distance is used to measure the degree of deviation of the predicted point from the overall cluster distribution; extraction with cluster The typical trajectory is calculated by taking the first derivative (velocity vector) and second derivative (acceleration vector) at key points, calculating the Euclidean distance between them, and using this distance as the dynamic characteristic distance. This is used to assess whether the dynamic behaviors of the two entities in state evolution are similar; based on domain knowledge... , and Weights are assigned, and the three distances are weighted and fused according to the assigned weights to obtain a comprehensive nearest neighbor distance. The smaller the comprehensive nearest neighbor distance, the better. The closer to a cluster of healthy patterns in the manifold space.

[0077] After the warning is triggered, the parameters of the spatiotemporal graph convolutional neural network are fixed, and the spatiotemporal state graph is... As a sample, the interpreter of SHAP, such as GraphSHAP or an interpreter of gradient-based approximation methods, is invoked to compute... The SHAP value of each input node feature, such as "stress value of beam segment 3" and "efficiency of concrete pump truck A", is calculated. This SHAP value quantifies the contribution of each feature to pushing the spatiotemporal graph convolutional neural network output from the baseline to the current anomaly prediction. All input features are sorted in descending order of absolute SHAP value, and the Top features are selected. N features are selected as leading root cause candidates, and combined with corresponding domain knowledge, these features are mapped to specific engineering entities and events. For example, SHAP analysis shows that the three features contributing most to the abnormal increase in the predicted stress value at the root of the main beam are: {node feature: current fluctuation index of prestressing tensioning equipment A, SHAP: +0.42}, {node feature: estimated 24-hour strength of concrete in beam segment 3, SHAP: -0.38}, and {edge feature: influence coefficient of ambient temperature on formwork deformation, SHAP: +0.15}. These can be translated using the corresponding domain knowledge as: This warning is mainly attributed to: unstable operation of prestressing tensioning equipment A, contributing 42%; lower-than-expected concrete strength development in beam segment 3, contributing 38%; and the additional impact of ambient temperature changes on the formwork system, contributing 15%.

[0078] Finally, the output is an early anomaly dataset containing at least the alert ID and timestamp, alert type and level, associated health patterns, root cause analysis list, and associated entity list. The alert ID and timestamp represent the unique identifier of the alert and the trigger time; the alert type and level are, for example, {Trend Anomaly, Level 1} and {Uncertainty Anomaly, Level 1}, where the level represents the number of alert types triggered. If only one type of alert (Trend Deviation or Uncertainty Anomaly) is triggered, the level is 1; if both are triggered, the level is 2; the associated health patterns represent the names of the deviated health pattern clusters; the root cause analysis list contains the Top N dominant root cause candidates and their contribution; and the associated entity list represents the affected components, equipment, and process numbers.

[0079] Step A2: The early abnormal dataset is analyzed and decided upon using a pre-built candidate construction control strategy generation network to generate a set of candidate construction control strategies.

[0080] Specifically, the candidate construction control strategy generation network performs counterfactual intervention on the causal variables that lead to abnormal patterns based on the early abnormal dataset, generating a counterfactual intervention event set. The candidate construction control strategy generation network uses each counterfactual intervention event in the counterfactual intervention event set as a game issue and real-time construction process data as the state space. Each game agent combines the corresponding action space and payoff function to conduct multiple rounds of game to obtain a consensus action sequence.

[0081] It should be noted that the candidate construction control strategy generation network consists of multiple game agents, each of which consists of a policy network and a value network. The candidate construction control strategy generation network is trained using a course learning strategy, and simultaneously, a combination of inverse reinforcement learning and adversarial learning is used to infer the implicit reward function contained in the training dataset.

[0082] The parameters of the candidate construction control strategy generation network are fine-tuned based on the implicit reward function, and a causal discovery algorithm is used during training to construct a construction constraint graph. The construction constraint graph is used to constrain the game actions generated by the game agent to not violate the corresponding domain prior knowledge.

[0083] Specifically, based on the construction project contract and organizational structure, key decision-making participants are defined, such as design agents, general contractor agents, key subcontractors (e.g., steel structure, prestressed concrete), and supervision agents. Each agent is defined as a game agent with autonomous decision-making capabilities. The action space of each agent is represented by enumerating its executable, discrete, or parameterized decision actions based on its responsibilities and authority. For example, the action space of the design agent might include: {issuing design change orders, issuing design change notices, issuing technical approval forms, maintaining the original design}, while the action space of the general contractor agent might include: {applying for a construction period extension of X days, applying for a fee of Y yuan, adjusting resource allocation, initiating a rush work plan}. Simultaneously, a payout function is learned from historical project data using inverse reinforcement learning. This involves analyzing numerous historical cases, examining the actions taken by each agent under specific project conditions, and the resulting comprehensive outcomes such as profit, schedule bonuses / penalties, quality evaluation, and cooperation relationship scores. Inverse reinforcement learning is then used to deduce the implicit weight preferences of each agent for each dimension of the outcome. For example, the payout function of the construction agent can be represented as: The weight Inferred from historical behavioral data; simultaneously, pre-defined basic rules for multi-party negotiation, such as adopting an issue-based alternating offer agreement, and the rules at least include: a maximum number of negotiation rounds, such as... The order of each round of speaking, the format of proposals including positions on the current issue and supporting conditions, and the criteria for reaching consensus, such as all parties voting in favor of a joint action plan.

[0084] The training data for the game theory agent consists of complete decision records from historical successful construction projects. Through natural language processing and process mining techniques, these complete decision records are transformed into structured trajectories {project status}. Joint action by all parties Subsequent status Final project results Furthermore, each trajectory is labeled with a difficulty level, such as unilateral decision-making level, two-party coordination level, and multi-party complex game level. The training process of the game agent is divided into the following three stages: the single agent basic learning stage, which is used to train each game agent to master the basic operations in the domain corresponding to its agent role, without training the value network. This stage adopts the behavior cloning method in supervised learning, with the goal of maximizing the log-likelihood of the policy network on the state-action pairs contained in historical successful construction project cases. This enables the policy network of the game agent to replicate the unilateral response of the expert in the corresponding state. For example, the construction agent learns from a large number of progress delay state samples that when the delay is caused by material shortage, it should prioritize initiating an emergency material procurement application.

[0085] In the fixed-partner collaborative learning phase, the training environment is a simplified environment consisting of a game-theoretic agent and a fixed expert partner. For example, when training a construction agent, its partner is fixed as a design expert agent abstracted from historical best collaborative records. The expert agent's behavioral strategy is predefined and frozen to simulate the typical reaction pattern of an ideal, experienced designer. The input to the game-theoretic agent's value network is the current collaborative state and its own candidate actions, and the output is a scalar representing the long-term benefit that the agent estimates can bring by taking this action in this state. The initial value of the value network is inferred from historical data using maximum entropy inverse reinforcement learning. The initial reward function, or value function, is used at this stage as a discriminator to distinguish between the expert sequence formed by the baseline state-policy pairs actually used in historical successful construction projects and the apprentice sequence formed by the actual state-policy pairs generated by the game agents. The value function is trained using a maximum entropy inverse reinforcement learning framework, aiming to minimize the binary cross-entropy between the expert and apprentice sequences. The policy network is trained at this stage using a proximal policy optimization algorithm, aiming to maximize the cumulative reward obtained from its own value network while fixing the policy network parameters of the expert partners. Through this stage of training, the value network of each game agent is initialized.

[0086] In the multi-agent full game learning phase, the parameters of any game agent are no longer frozen. A multi-agent proximal policy optimization algorithm is used for training. Specifically, the training environment in this phase simulates a complete bridge construction cycle. Each game agent makes decisions based on the global state vector of the training environment. A course-based learning strategy is adopted. Initially, the game agents interact in a simplified scenario with single-process decision-making. Subsequently, the complexity of the scenario is gradually increased, such as multiple process intersections, resource conflicts, and sudden weather events. Finally, game training is conducted in a complete scenario covering various abnormal modes. In each round of training, each agent generates actions based on the current policy network and simulates the corresponding project state transitions and the payoffs of each game agent using the corresponding environment simulator and each agent's value function. The update of the policy network in this phase depends on a combined loss function, the main body of which... Part of the optimization is a pruning objective function for near-end policy optimization. This function ensures training stability by limiting the step size of policy updates and maximizes the expected value of the advantage function calculated by the value network. It also introduces a causal regularization term, which is dynamically generated by a parallel causal discovery module. This module continuously learns the causal graph between variables from the interaction data generated in this stage and uses this causal graph as a construction constraint graph. At the same time, it penalizes the probability of actions that violate the identified causal constraints output by the policy network, such as penalizing the causal constraint that "excessive increase in cement usage will exponentially increase the risk of temperature rise and cracking". The value network is trained using the Huber loss function, with the goal of minimizing its temporal difference error in predicting the long-term return of the state, until the average return of each game agent fluctuates less than a preset threshold or reaches a preset maximum number of training steps over multiple consecutive training cycles.

[0087] Furthermore, the consensus action sequence is executed in the simulation to obtain the corresponding payoff vector. All payoff vectors and their corresponding consensus action sequences are merged into a set of strategy payoff pairs. The set of strategy payoff pairs is then subjected to refined Bayesian equilibrium and Pareto front analysis to generate a set of candidate construction control strategies.

[0088] As one possible implementation, firstly, one or more causal variables that contribute most to the anomalous pattern are extracted from the root cause analysis list contained in the early anomaly dataset; then, combining a pre-set construction measures knowledge base and a construction constraint map constructed in the pre-training phase, a set of technically feasible and logically related counterfactual intervention target values ​​are generated for each causal variable, thus forming a counterfactual intervention event set. For example, the intervention event generated for the variable "low rate of development of elastic modulus" is: "It is recommended to upgrade the concrete design grade of segments 20-25 from C50 to C55"; next, based on the current project's "real-time construction process"... The data, namely a global state snapshot containing all physical states, progress, resources, and environmental indicators, serves as the initial state space for the game. For each event in the counterfactual intervention event set, an independent multi-agent game environment is instantiated, and this event becomes the core issue of this round of the game. In this environment, game agents representing each participant are activated. Each agent, based on its pre-trained policy network, selects action proposals from its corresponding action space based on the current state space and the game issue, and evaluates the long-term value of various action choices according to the payoff function corresponding to its value network. For example, suppose the action space of the designer agent contains [actions], and its payoff function is [function]. Subsequently, the agents engage in multiple rounds of negotiation and game-playing. In each round, the policy network of each agent dynamically adjusts its bid based on the latest negotiation situation and its own payoff function assessment. After multiple rounds of interaction, the process eventually converges to a consensus action sequence to which all agents agree. This sequence details the specific, time-ordered actions that each party must take in response to the intervention event, starting from the current moment. For example, the consensus sequence reached after three rounds of negotiation might be: [T0: The designer issues a C55 design change order], [T0+2h: The client approves the change], [T0+4h: The construction company signs a C55 concrete supply contract and adjusts the working hours for the subsequent three processes], [T0+6h: The supervisor approves the new process plan]. If a consensus cannot be reached, the negotiation breaks down, and manual intervention is required. This involves transmitting the data corresponding to the abnormal mode where a consensus cannot be reached to the user terminal for real experts to make decisions.

[0089] For each consensus action sequence, it is input into a high-fidelity construction process simulator for simulation. The simulator simulates the complete project dynamics from the implementation of the sequence to the next decision point, and calculates the final key indicator results, such as changes in construction period, cost changes, and risk values. Based on the key indicator results, the payoff functions of each game agent are called to calculate their respective payoff scalars, thus obtaining a payoff vector. The consensus action sequence and its corresponding payoff vector are defined as a strategy payoff pair. For example, for the above consensus sequence, the simulation results show that the total cost increases by 150,000 yuan, the construction period is shortened by 1.5 days, and the deviation risk is reduced by 55%. The payoff vectors calculated by the payoff functions of each game agent are: [Designer: +0.25, Client: +0.10, Construction: +0.15, Supervisor: +0.05]. After traversing all intervention events in the counterfactual intervention event set, a set of strategy payoff pairs containing multiple possible solutions is formed.

[0090] It should be noted that the construction process simulator is a high-fidelity, multi-agent simulation platform based on ABM (Agent-Based Modeling) and discrete event simulation, which includes the following four agent modules: a physical process agent module, used to simulate the physical effects of key construction processes such as concrete pouring and prestressing tensioning, and to calculate physical state changes such as stress and deformation by embedding finite element analysis models or empirical formulas derived from historical data analysis; a schedule management agent module, which simulates the dependencies and progress of processes based on process logic networks such as CPM / PERT, and handles process insertion, delay, or parallelization; a resource management agent module, used to simulate the dynamic allocation, consumption, and flow of labor, equipment, and materials; and an external environment agent module, used to simulate external random factors such as weather. The agent modules interact through an event bus. The simulator is trained using real historical project data as the training dataset, with the goal of minimizing the mean square error between the simulator's output and the corresponding parameters of the real historical project data, until a preset maximum number of training steps is reached or the change in the mean square error value over multiple consecutive training cycles is lower than a preset threshold.

[0091] The strategy payoff set is analyzed and the optimal strategy is selected. Specifically, firstly, the refined Bayesian equilibrium solution algorithm is used to analyze the entire game scenario and select stable equilibrium strategy combinations from the set, that is, strategies in which no player can obtain higher payoffs by unilaterally changing their own actions. Then, Pareto front analysis is performed on the equilibrium strategy combinations, that is, the evaluation is carried out from the perspective of the overall project objectives such as comprehensive cost, total project duration, and overall quality, and Pareto optimal solutions are selected that are better than other strategies on at least one objective and not inferior to other strategies on other objectives. For example, a strategy that increases cost by 1% but shortens the project duration by 2 days is not absolutely superior to a strategy that keeps cost unchanged but shortens the project duration by 1 day, and may both be on the Pareto front.

[0092] Finally, the top N Pareto optimal equilibrium strategy combinations, along with their corresponding action sequences, expected global effects, and benefit matrices for each party, are encapsulated into several candidate construction control strategies. For example, strategy A {Event: Increase concrete grade, Action sequence: ..., Expected effect: [Cost +0.8%, Construction period -1.5 days, Deviation risk -55%], Benefit matrix: [Designer: +0.25, Client: +0.10, Construction: +0.15, Supervisor: +0.05]}; all candidate construction control strategies are encapsulated into a candidate construction control strategy set for output.

[0093] Step A3: Based on the preset multi-scale construction mode association network, cross-scale strategy transmission and adaptive adjustment are performed on the candidate construction control strategy set to generate the initial construction control strategy set.

[0094] Specifically, the candidate construction control strategy set is mapped to macro-level nodes in a multi-scale construction mode association network, generating corresponding construction intervention nodes. A propagation algorithm is run starting from the construction intervention nodes to obtain the corresponding intervention target set. The intervention target set is used to describe the specific variables and their target values ​​that need to be adjusted at the meso-level and micro-level. Based on the intervention target set, the corresponding initial construction control strategy set is generated.

[0095] It should be noted that the multi-scale construction mode association network is represented as a multi-layer Bayesian network containing macro-level, meso-level, and micro-level layers. The macro-level layer characterizes variables at the construction project level, such as the total project duration deviation, total project cost performance index, and overall quality risk level; the nodes in the macro-level layer reflect the top-level objectives and health status of the construction project. The meso-level layer characterizes variables at the engineering component level, such as the cumulative settlement of the main pier, the height difference at the closure joint of the steel box girder, and the uniformity index of the stay cable force. The micro-level layer describes variables at the component or process level, i.e., specific component attributes. Or process parameters, such as the 28-day compressive strength of the concrete in the 5th segment of pier No. 3, the real-time value of the tension of prestressed cable P15, and the daily completion efficiency of the rebar tying team B, etc., the nodes of the micro-level are directly executable and measurable operation endpoints; the edges in the multi-scale construction mode association network are used to characterize the causal relationship between variables. These relationships are established through learning from historical data and injecting domain knowledge. For example, an edge from the micro-level node "concrete placement temperature" to the meso-level node "early cracking risk of the main beam" has a conditional probability distribution that describes the intensity of the influence of temperature on the risk.

[0096] As one possible implementation, firstly, the action sequence of each candidate construction control strategy in the candidate construction control strategy set is analyzed to identify its intervention intention on macro-level nodes. For example, the action sequence corresponding to a certain candidate construction control strategy is converted into a corresponding semantic expression using a large language model. Suppose that the semantic expression obtained by converting the action sequence is "the designer agrees to upgrade the concrete grade of the main beam from C50 to the more expensive C55, thereby increasing the total project cost and shortening the total project duration". Combining the keyword extraction technology of the large language model, "total project duration" and "total project cost" are extracted, thereby analyzing the two macro-level intervention intentions of total project duration and total project cost. In the macro-level, a construction intervention node of "increasing the concrete grade of the main beam" is established, and edge connections are established between this construction intervention node and the existing nodes of total project duration and total project cost in the macro-level.

[0097] Next, starting from this construction intervention node, a causal propagation algorithm is run to calculate and obtain the corresponding set of intervention targets. Specifically, based on the backward reasoning and do-calculus of Bayesian networks, the new states or target values ​​that the relevant variables at the meso- and micro-levels must reach to achieve this macro-intervention target are derived from the macro-intervention target. This propagation process proceeds along the causal edges in the multi-scale construction mode association network, combining the conditional probability distribution between nodes and the current project state for propagation reasoning. For example, assuming the macro-intervention target is to shorten the total construction period, its do-calculus representation is do(total construction period deviation = Given that the increased concrete grade has been adopted, the propagation algorithm will calculate how much the prefabrication cycle of the main beam segment at the meso-level node needs to be shortened to achieve the macro-level construction period target. For example, it might calculate that the cycle should be shortened from 7 days / segment to 6 days / segment. Further calculations will be made to determine how the micro-level nodes need to be adjusted to achieve the meso-level cycle shortening. For example, it might calculate that the concrete steam curing time needs to be reduced from 48 hours to 36 hours, and the concrete strength requirement at the time of formwork removal needs to be increased from 75% to 85% of the design strength. Finally, an intervention target set will be output, which clearly defines the specific variables that need to be adjusted at the meso- and micro-levels and their target values ​​or target ranges. For example: {Meso-level: [Main beam segment prefabrication cycle, target value ≤ 6 days], Micro-level: {[C55 concrete 48-hour strength, target value ≥ 40 MPa], [Constant temperature time in curing shed, target value ≥ 30 hours]}}.

[0098] Next, based on the generated set of intervention targets, the fine-grained models of each discipline are driven to undergo constrained automated adjustments, thereby generating a set of directly executable initial construction control strategies. Specifically, each item in the set of intervention targets is distributed to the corresponding professional calculation model, such as the structural analysis model, BIM model, and construction simulation model. Each model runs a local optimizer or parametric generator, combining all the design specifications, physical laws, and structural constraints within its model. These design specifications, physical laws, and structural constraints are set based on historical data and prior knowledge of the domain to maximize the achievement of the target value. For example, after the BIM software receives the micro-level target of "reducing the reinforcement ratio of a certain area from 2.0% to 1.8%", its internal parametric rule engine will automatically recalculate and adjust the diameter, spacing, or arrangement of the reinforcing bars while keeping the geometric dimensions and connection relationships of the components unchanged, generating a new and compliant reinforcing bar layout drawing. If a target cannot be achieved under the existing constraints, such as the strength target not being achievable by adjusting the curing time without changing the material, the conflict will be marked and the corresponding candidate construction control strategy will be removed.

[0099] Then, the adjustment results generated by all professional calculation models are extracted and uniformly mapped into a project digital twin knowledge graph. The nodes of this graph are entities such as components, processes, and resources, while the edges are used to represent relationships such as spatial connections, temporal order, and logical dependencies. Multi-round conflict detection is performed in the project digital twin knowledge graph. For example, a fast interference check algorithm based on bounding box hierarchical trees is called to scan all adjusted components and identify geometric spatial collision conflicts such as "the newly arranged 32mm diameter main reinforcement and the already positioned prestressed corrugated pipe physically interfere in three-dimensional space". The adjusted process network is checked through temporal logical reasoning to identify logical temporal conflicts such as "the concrete curing time compressed to meet the construction schedule contradicts the strength development time window required by the structural model calculation". The resource allocation scheme is verified through linear programming to identify resource overload conflicts such as "the tensioning equipment is simultaneously allocated to two far apart work surfaces in the same period".

[0100] If a conflict is detected, firstly, the corresponding automated negotiation and resolution mechanism is activated according to the conflict type to generate an automated resolution plan. For example, for geometric collision conflicts, the corresponding professional calculation model is invoked for a second iteration, and the component layout is fine-tuned while maintaining the intervention objective. For logical timing conflicts, the constraint programming solver is invoked to re-optimize the process logic and duration. For resource conflicts, dynamic reallocation is performed based on the priority of the resource allocation formula. All automated resolution plans will go through the above conflict detection and elimination cycle again until no new conflicts are detected or the iteration limit is reached. Conflicts are then eliminated on the intervention target set according to the generated automated resolution plan. For complex conflicts that cannot be automatically resolved, such as those involving modifications to safety specification red lines, a structured conflict report will be generated, clearly listing the conflict location, stakeholders, impact analysis, and recommended manual handling plans for relevant experts to make manual decisions.

[0101] Finally, the set of intervention targets that achieve conflict resolution is output as the initial set of construction control strategies.

[0102] Step A4: The initial construction control strategy set is simulated and optimized using a cellular automata-based construction site control model to generate enhanced construction control strategies.

[0103] Specifically, the initial construction control strategy set is injected into the construction site control model to initialize the state of each cell. The schedulable execution unit of the construction site is modeled as a scheduling agent. Each scheduling agent combines the corresponding cell set with the real-time acquired actual disturbance signal to perform strategy simulation execution. Based on the simulation execution results, the parameters of the initial construction control strategy set are fine-tuned to generate an enhanced construction control strategy.

[0104] It should be noted that the construction site control model is represented by discretizing the construction site into a spatiotemporal cellular automaton model based on cellular automata. Each cell of the model represents a spatial unit, and the state attributes of the spatial unit include at least construction progress data, resource status data, and construction environment data.

[0105] As one possible implementation, firstly, the two-dimensional or three-dimensional space of the actual construction site is discretized into a regular grid, such as a 1m×1m×1m three-dimensional space as a grid cell. Each grid cell is defined as a unit cell, and the state attributes of each unit cell include at least: construction progress data, such as the current process type, completion percentage, and planned or actual start and end time; resource status data, such as the work group number, equipment ID, and material inventory of the work group occupying the unit cell; construction environment data, such as temperature, humidity, wind speed, and light intensity; the local evolution rules contained in the construction site control model define how the state is updated over time. These rules are formulated based on the analysis of historical data and prior knowledge of the domain. For example, for a unit cell to change from the "rebar tying in progress" state to the "formwork installation in progress" state, its own progress must reach 100%, and at least one of the four adjacent units cell must be in the state of "formwork ready".

[0106] Next, the initial construction control strategy set is injected into the construction site control model to initialize the state of each cell. Specifically, each detailed instruction in the initial construction control strategy set is converted into the expected state of the corresponding spatial cell at a specific time step. For example, the spatial region specified in the instruction is mapped to the corresponding cell ID set {101,102,...,110}, the time information is converted into the time step sequence of the model, and the resource requirements are written into the resource state attribute of the corresponding cell at the corresponding time step. After the initialization is completed, the construction site control model will present a theoretically conflict-free ideal construction state starting from the current moment.

[0107] Next, the schedulable execution units at the construction site are modeled as scheduling agents. These schedulable execution units are entity resource units such as "Tower Crane A Operation Team", "Third Rebar Binding Team", and "Concrete Pump Truck B". Each scheduling agent is bound to a set of cells corresponding to the work area it is responsible for, so as to efficiently complete the cell state transition tasks assigned to it. For example, "Tower Crane A Agent" is responsible for all cells that need to be hoisted, and its decision-making logic includes selecting the hoisting sequence, path planning, and coordination and avoidance.

[0108] Then, based on the actual disturbance signals transmitted from the IoT sensor network, such as the current actual position of tower crane A deviating from the plan by 10 meters, weather radar predicting heavy rainfall in 15 minutes, and the arrival of the steel bar delivery truck delayed by 30 minutes, the state of the affected cells is updated in real time. The construction site control model advances the simulation with a fixed time step. Specifically, at the beginning of each simulation step, the cells marked as ready in the construction site control model will generate specific construction task requirements. The ready state means that its state meets the preset process triggering conditions, such as the completion of the previous process and the availability of resources. The virtual scheduling center module of the model scans all ready cells in real time and aggregates tasks of the same type and spatiotemporally close into task packages. Each task package is formalized as a tender notice, which includes the task type, spatial location coordinates, expected time window, resource requirement list, and dynamic priority based on critical path calculation. The virtual scheduling center module sends the tender notice to all scheduling agents with matching capabilities through a global broadcast channel.

[0109] Upon receiving the tender notice, the scheduling agent initiates its local bidding decision-making process in parallel. First, it performs a feasibility check, including its own resource availability, time window compatibility, and mobility reachability. If feasible, it calls the bidding function. Calculate the comprehensive bid price, where, It is expressed as the migration cost from the current location to the task location, and the corresponding time or energy consumption can be estimated through a real-time traffic model; This represents the expected waiting cost caused by the current task queue. This is expressed as the task execution cost predicted based on historical data and execution efficiency. } represents the task priority adjustment item, with each task priority corresponding to a specific adjustment item value, which can be set based on historical data and prior knowledge; next, for each bidding announcement, the bidding score of each matching scheduling agent is calculated, and the calculation of the bidding score can be expressed as... ,in To preset weights, The current load rate of the agent. The historical task completion rate of the scheduling agent is represented by the bidder. The scheduling agent with the highest bid score is selected as the winner. After receiving the task package, the winning scheduling agent will update its state attributes and drive its corresponding cell to run the simulation according to the local evolution rules.

[0110] It should be noted that during the simulation, the parameters of the initial construction control strategy set are fine-tuned based on the simulation results. Specifically, the simulation progress is compared with the baseline progress set in the initial construction control strategy set. When the difference between the two exceeds the preset performance fluctuation threshold, it is determined that a performance deviation has occurred. Starting from the current simulation state, with the optimization objectives of minimizing deviations from the core goals of the initial construction control strategy, such as the shortest total construction period, and eliminating conflicts, the adjustable parameters in the initial construction control strategy set, such as the start time of the work process, the amount of resources allocated, and the priority of the agent, are quickly and locally replanned. For example, if the simulation finds that the rebar binding may be delayed due to truck delays, the optimization analysis determines that the start time of the "subsequent formwork installation" task will be postponed by 30 minutes, and a waiting instruction will be sent to the "formwork team" agent at the same time. The fine-tuned parameters are immediately injected as new constraints into the subsequent simulation steps to guide the agent to adjust its behavior, thereby realizing real-time conflict resolution and parameter fine-tuning of the initial construction control strategy set during the simulation.

[0111] Finally, the simulation process ends when the preset simulation endpoint is reached. If all intervention objectives of the initial construction control strategy set or the strategies of all scheduling agents do not change within multiple consecutive simulation cycles, the adjusted initial construction control strategy set will be output as an enhanced construction control strategy after the simulation is completed.

[0112] Figure 2 This is a schematic diagram of a digital simulation and collaborative system for bridge construction according to the present invention.

[0113] Specifically, a digital simulation and collaborative system for bridge construction includes:

[0114] An anomaly monitoring module is used to acquire real-time construction process data and, in conjunction with a preset construction status monitor, to identify early anomalies and acquire an early anomaly dataset.

[0115] The strategy generation module is used to analyze and make decisions on the early abnormal dataset based on the candidate construction control strategy generation network, and generate a candidate construction control strategy set.

[0116] The strategy conversion module is used to adaptively adjust the candidate construction control strategy set according to the multi-scale construction mode association network to generate an initial construction control strategy set.

[0117] The strategy optimization module is used to simulate and optimize the initial construction control strategy set using a construction site control model, and generate enhanced construction control strategies.

[0118] An electronic device, comprising:

[0119] At least one processor; and at least one memory communicatively connected to the processor; wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to perform the method proposed in Embodiment 1 of the present invention.

[0120] The following is a detailed introduction to the various components of the electronic device:

[0121] In this context, the processor is the control center of the electronic device. It can be a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement Embodiment 1 of this invention, such as one or more digital signal processors (DSPs) or one or more field-programmable gate arrays (FPGAs).

[0122] The processor can perform various functions of an electronic device by running or executing software programs stored in memory and by calling data stored in memory.

[0123] The memory is used to store the software program that executes the solution of the present invention, and the execution is controlled by the processor. For specific implementation methods, please refer to the above method embodiments, which will not be repeated here.

[0124] The memory can be a real-only memory (ROM) or other type of static storage device capable of storing static information and instructions, a random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only (CD-ROM), or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through an interface circuit of an electronic device; this embodiment of the invention does not specifically limit this.

[0125] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via limited means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0126] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0127] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0128] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A digital simulation and collaborative method for bridge construction, characterized in that, It includes the following steps: Step A1: Obtain real-time construction process data, use a preset construction status monitor to identify early anomalies, and obtain an early anomaly dataset; Step A2: The early abnormal dataset is analyzed and decided upon using a pre-constructed candidate construction control strategy generation network to generate a candidate construction control strategy set; Step A3: Based on the preset multi-scale construction mode association network, perform cross-scale strategy transfer and adaptive adjustment on the candidate construction control strategy set to generate the initial construction control strategy set; Step A4: Use a cellular automata-based construction site control model to simulate and optimize the initial construction control strategy set, and generate enhanced construction control strategies.

2. The bridge construction digital simulation collaborative method according to claim 1, characterized in that, Acquire real-time construction process data, use a pre-set construction status monitor to identify early anomalies, and obtain an early anomaly dataset, including: The real-time construction process data includes at least physical status data, construction progress data, resource status data, and construction environment data; The manifold learning algorithm is used to map real-time construction process data to the manifold space corresponding to the construction status monitor, thereby generating real-time status nodes. Connect the real-time status nodes with the corresponding historical status nodes to generate an actual construction trajectory sequence; The actual construction trajectory sequence is input into a preset spatiotemporal graph convolutional neural network for trajectory prediction, generating a predicted construction trajectory sequence. The predicted construction trajectory sequence is compared and analyzed with the healthy construction mode cluster contained in the manifold space to obtain the corresponding comparison and analysis results. If the comparison and analysis results meet the preset warning conditions, early anomaly detection will be performed, and a corresponding early anomaly dataset will be generated. The early anomaly dataset contains at least the anomaly patterns and the contribution distribution of the anomaly pattern causes.

3. The bridge construction digital simulation and collaborative method according to claim 2, characterized in that, The method further includes: A high-dimensional construction phase space is constructed based on historical construction process data. A manifold learning algorithm is used to perform nonlinear dimensionality reduction on the high-dimensional construction phase space to generate the corresponding manifold space. The manifold contains multiple health pattern clusters, which are represented as continuous clusters formed by the aggregation of state nodes corresponding to construction stages in which no anomalies have occurred in historical data. The spatiotemporal graph convolutional neural network is used to store the dependencies between data in various dimensions within historical or real-time construction process data, and predicts the probability distribution of state trajectories at multiple future time steps based on the input real-time construction trajectory sequence, generating a corresponding predicted construction trajectory sequence.

4. The bridge construction digital simulation and collaborative method according to claim 1, characterized in that, A pre-built candidate construction control strategy generation network is used to analyze and make decisions on the early anomaly dataset, generating a candidate construction control strategy set, including: The candidate construction control strategy generation network performs counterfactual intervention on the causal variables that lead to abnormal patterns based on early abnormal datasets, generating a counterfactual intervention event set; The candidate construction control strategy generation network takes each counterfactual intervention event in the counterfactual intervention event set as the game issue and real-time construction process data as the state space. Each game agent combines its corresponding action space and payoff function to engage in multiple rounds of game to obtain a consensus action sequence. The simulation executes the consensus action sequence, obtains the corresponding reward vector, and merges all the reward vectors with the corresponding consensus action sequence into a set of policy reward pairs. The set of strategy payoff pairs is subjected to refined Bayesian equilibrium and Pareto front analysis to generate a set of candidate construction control strategies.

5. The bridge construction digital simulation collaborative method according to claim 4, characterized in that, The candidate construction control strategy generation network includes: The candidate construction control strategy generation network consists of multiple game agents, each of which comprises a strategy network and a value network. A course learning strategy is used to train the candidate construction control strategy generation network, and a combination of inverse reinforcement learning and adversarial learning is used simultaneously to infer the implicit reward function contained in the training dataset. The parameters of the candidate construction control strategy generation network are fine-tuned based on the implicit reward function, and a causal discovery algorithm is used during the training process to construct a construction constraint graph. The construction constraint graph is used to constrain the game actions generated by the game agent to ensure that they do not violate the corresponding domain prior knowledge.

6. The digital simulation and collaborative method for bridge construction according to claim 1, characterized in that, Based on a pre-defined multi-scale construction mode association network, cross-scale strategy transfer and adaptive adjustment are performed on the candidate construction control strategy set to generate an initial construction control strategy set, including: The multi-scale construction mode association network is represented as a multi-layer Bayesian network containing macro-level, meso-level, and micro-level layers. Among them, the macro layer is used to characterize variables at the construction project level, the meso layer is used to characterize variables at the engineering part level, and the micro layer is used to describe variables at the component or process level. The candidate construction control strategy set is mapped to macro-level nodes in a multi-scale construction mode association network to generate corresponding construction intervention nodes. The propagation algorithm is run starting from the construction intervention node to obtain the corresponding set of intervention targets; The set of intervention targets is used to describe the specific variables and their target values ​​that need to be adjusted at the meso- and micro-levels. A corresponding initial construction control strategy set is generated based on the set of intervention targets.

7. The digital simulation and collaborative method for bridge construction according to claim 1, characterized in that, A cellular automata-based construction site control model is used to simulate and optimize the initial construction control strategy set, generating enhanced construction control strategies, including: The construction site control model is represented by discretizing the construction site into a spatiotemporal cellular automaton model based on cellular automata, where each cell of the model represents a spatial unit. The status attributes of the spatial unit include at least construction progress data, resource status data, and construction environment data; The initial construction control strategy set is injected into the construction site control model to initialize the state of each cell. The schedulable execution units at the construction site are modeled as scheduling agents. Each scheduling agent combines its corresponding set of cells with the real-time acquired actual disturbance signals to perform strategy simulation execution. Based on the simulation results, the parameters of the initial construction control strategy set are fine-tuned to generate an enhanced construction control strategy.

8. A digital simulation and collaborative system for bridge construction, used to implement the method described in any one of claims 1 to 7, characterized in that, include: An anomaly monitoring module is used to acquire real-time construction process data and, in conjunction with a preset construction status monitor, to identify early anomalies and acquire an early anomaly dataset. The strategy generation module is used to analyze and make decisions on the early abnormal dataset based on the candidate construction control strategy generation network, and generate a candidate construction control strategy set. The strategy conversion module is used to adaptively adjust the candidate construction control strategy set according to the multi-scale construction mode association network to generate an initial construction control strategy set. The strategy optimization module is used to simulate and optimize the initial construction control strategy set using a construction site control model, and generate enhanced construction control strategies.

Images