A BIM high-rise suspended structure monitoring method and system based on machine learning

Abstract

The present invention relates to the technical field of building structure monitoring, and discloses a BIM high-rise suspended structure monitoring method and system based on machine learning. Specifically, a BIM high-rise suspended structure monitoring method based on machine learning includes: constructing a physical constraint variational causal reasoning framework to integrate BIM model data and sensor monitoring data; training a dynamic physical causal graph learning module to identify the causal relationship network in the high-rise suspended structure; constructing and training a hierarchical attention perception module to capture structural abnormality patterns; developing a physical constraint neural ordinary differential equation module to achieve accurate prediction of the dynamic behavior of the high-rise structure; and establishing a two-way real-time mapping between monitoring data and BIM model elements. The present invention solves the technical problem of "weak correlation" characteristic monitoring of high-rise suspended structures by integrating physical constraint variational causal reasoning, dynamic physical causal graph learning, hierarchical attention perception, physical constraint neural ordinary differential equations and BIM semantic mapping.

Description

Technical Field

[0001] The present invention relates to the technical field of building structure monitoring, and more specifically, to a BIM high-rise suspended structure monitoring method and system based on machine learning. Background Art

[0002] High-rise suspended structures are a new structural system in modern architecture. They achieve large spans and spacious spaces by supporting lower floors with top trusses and vertical steel tie rods. This structural form is increasingly used in buildings such as commercial complexes and cultural venues. However, high-rise suspended structures exhibit a typical "weak correlation" characteristic, with loads transmitted between floors primarily through steel tie rods, making their behavior complex and difficult to predict.

[0003] Existing structural monitoring technologies mainly include physical model-based methods and data-driven methods. Physical model-based methods rely on precise structural models and are difficult to adapt to the complexity and uncertainty of high-rise suspended structures. Although data-driven methods can learn patterns from data, they lack physical interpretation capabilities and perform poorly in cases of sparse or abnormal data. In addition, existing monitoring systems generally have the following problems: first, it is difficult to capture the "weak correlation" characteristics between floors in high-rise suspended structures; second, the monitoring results lack effective integration with the BIM model, making data interpretation difficult; third, it is impossible to achieve predictive risk warnings, and most alarms are issued after the fact.

[0004] Therefore, there is an urgent need to develop an intelligent monitoring method and system that can effectively handle the "weak correlation" characteristics of high-rise suspended structures, integrate physical knowledge with data-driven methods, and be deeply integrated with the BIM system. Summary of the Invention

[0005] The present invention provides a BIM high-rise suspended structure monitoring method and system based on machine learning, which solves the technical problem in related technologies that it is difficult to effectively monitor the "weak correlation" characteristics of high-rise suspended structures.

[0006] The present invention provides a BIM high-rise suspended structure monitoring method based on machine learning, comprising:

[0007] Construct a physically constrained variational causal inference framework to integrate BIM model data with sensor monitoring data;

[0008] Based on the fusion results of BIM model data and sensor monitoring data, a dynamic physical causal graph learning module is trained to identify the causal relationship network in high-rise suspended structures;

[0009] Using the identified causal network, we construct and train a hierarchical attention perception module to capture structural abnormal patterns;

[0010] Based on the captured structural anomaly patterns, a physical constraint neural ordinary differential equation module is developed to accurately predict the dynamic behavior of high-rise structures;

[0011] Based on the predicted structural dynamic behavior, the BIM semantic mapping and visualization module is implemented to establish a two-way real-time mapping between monitoring data and BIM model elements.

[0012] Furthermore, the steps of constructing a physical constrained variational causal inference framework include:

[0013] Extract the geometric model, material properties and component relationships of high-rise suspended structures from the BIM system;

[0014] Deploy a multimodal sensor network to collect real-time monitoring data of the structure;

[0015] Construct physical constraints based on structural dynamics equations;

[0016] Construct a physically constrained variational causal inference framework, combining structural dynamics equations, Bayesian causal reasoning, and deep learning through a unified model.

[0017] Furthermore, the step of training the dynamic physical causal graph learning module includes:

[0018] Initialize the dynamic causal graph, use sensor monitoring points as nodes, and build the initial topological structure;

[0019] Using the physically constrained Bayesian network learning algorithm, the dynamic causal relationship between nodes is learned based on time series monitoring data;

[0020] Applying structural sparsity regularization, the causal graph is optimized based on the "weak correlation" characteristics of high-rise suspension structures;

[0021] Dynamically update the causal graph structure and periodically update the causal network as the structural state and external environment change.

[0022] Furthermore, the steps of constructing and training the hierarchical attention perception module include:

[0023] Construct a three-level hierarchical attention network, including a temporal attention layer, a spatial attention layer, and a component attention layer;

[0024] Temporal attention layer training uses a multi-head self-attention mechanism to identify key time points for monitoring data at different time scales;

[0025] Spatial attention layer training uses a graph attention network to learn the spatial correlation between different monitoring points based on the spatial distribution and structural topological relationship of sensors;

[0026] Component attention layer training assigns dynamic attention weights to different structural components based on the component type and functional importance information in the BIM model.

[0027] Furthermore, the steps of developing the physical constraint neural ordinary differential equation module include:

[0028] A neural ordinary differential equation model is established to represent the time evolution of the structural state as a combination of a neural network and a physical model;

[0029] Design a hybrid loss function to optimize data fitting accuracy and consistency with physical laws;

[0030] Adopting adaptive step-size numerical integration method to solve neural ordinary differential equations and dynamically adjust the calculation step size;

[0031] By integrating the results of multiple numerical solution methods, the model's robustness to noise and abnormal data is improved.

[0032] Furthermore, the steps of implementing the BIM semantic mapping and visualization module include:

[0033] Establish BIM semantic mapping functions to map monitoring data, causal relationships, and uncertainty information to BIM model elements;

[0034] Develop a layered visualization interface to support multi-scale and multi-dimensional structural state display;

[0035] Implement adaptive computing resource allocation, dynamically adjusting computing resource allocation based on the output of dynamic causal graphs and hierarchical attention;

[0036] Build a predictive maintenance decision support system to automatically generate maintenance recommendations based on model prediction results.

[0037] Furthermore, the structural sparsity regularization is implemented through a graph structure loss function, which includes a probability logarithm term and a graph structure regularization term to promote the sparsity of the graph.

[0038] Furthermore, the three-level hierarchical attention network combines temporal attention, spatial attention and component attention in a cascade manner, and performs attention calculations on the input monitoring data in sequence.

[0039] Furthermore, the neural ordinary differential equation model combines neural network representation with physical constraints to achieve the prediction of the dynamic evolution of structural states over time.

[0040] The present invention provides a BIM high-rise suspended structure monitoring system based on machine learning, which is used to execute the above-mentioned BIM high-rise suspended structure monitoring method based on machine learning, including:

[0041] Physically constrained variational causal reasoning module, used to integrate BIM model data with sensor monitoring data to form a unified data processing foundation;

[0042] A dynamic physical causal graph learning module for identifying causal networks in high-rise suspended structures;

[0043] A hierarchical attention perception module to capture structural abnormality patterns from multiple temporal and spatial scales;

[0044] Physically constrained neural ordinary differential equations module for accurate prediction of the dynamic behavior of high-rise suspended structures;

[0045] BIM semantic mapping and visualization module is used to establish a two-way real-time mapping between monitoring data and BIM model elements.

[0046] The beneficial effects of the present invention are as follows: by integrating technologies such as physical constraint variational causal reasoning, dynamic physical causal graph learning, hierarchical attention perception, physical constraint neural ordinary differential equations and BIM semantic mapping, the technical difficulty of monitoring the "weak correlation" characteristics of high-rise suspended structures is solved;

[0047] It has achieved predictive risk warning and can identify potential risks in advance; improved causal interpretation capabilities and can automatically identify and track multi-step causal chains; optimized computing resource utilization and reduced computing load; enhanced the adaptability of "weakly correlated" structural characteristics and improved the accuracy in identifying weak correlation patterns of suspended structures; achieved deep integration of BIM and monitoring systems and reduced the workload of data interpretation; and possessed self-evolution capabilities and could automatically adjust focus and prediction models as the project progressed. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of a BIM high-rise suspended structure monitoring method based on machine learning in the present invention. DETAILED DESCRIPTION

[0049] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0050] At least one embodiment of the present invention discloses a BIM high-rise suspended structure monitoring method based on machine learning, such as Figure 1 As shown, including:

[0051] Step 1: Build a physical constraint variational causal reasoning framework to integrate BIM model data with sensor monitoring data;

[0052] Step 1.1, extract BIM data;

[0053] The geometric model, material properties and component relationships of the high-rise suspended structure, including information such as steel tie rod positions, connection node parameters, and floor layout, are extracted from the BIM system to generate a digital representation of the structure.

[0054] Step 1.2, deploy the sensor network;

[0055] Deploy a multimodal sensor network to collect real-time monitoring data of the structure, including physical quantities such as displacement, strain, inclination, and temperature, to form the original monitoring data stream.

[0056] Step 1.3, construct physical constraints;

[0057] Physical constraints are constructed based on structural dynamics equations to express the mechanical behavior of high-rise suspended structures as mathematical models, including:

[0058] Structural equilibrium equation:

[0059] ;

[0060] in is the stiffness matrix; is the displacement vector; is the external force vector;

[0061] Material constitutive relations:

[0062] ;

[0063] in is stress; is the material matrix; For strain;

[0064] Geometric compatibility conditions:

[0065] ;

[0066] in For strain; is the strain-displacement matrix; is the displacement vector;

[0067] Step 1.4, establish the PC-VCI framework;

[0068] A physically constrained variational causal inference (PC-VCI) framework is constructed and uniformly expressed through the following formula:

[0069] ;

[0070] in is the probability density function; is a cause-effect diagram; is a latent variable; are model parameters; is the observation data; For physical constraints; is the conditional probability separator;

[0071] This framework unifies structural dynamics equations, Bayesian causal reasoning, and deep learning in a single mathematical model to achieve comprehensive inference of the state of high-rise suspended structures.

[0072] The specific implementation of the PC-VCI framework includes three core units:

[0073] Variational encoder unit, physical constraint unit, and probabilistic graphical reasoning unit.

[0074] The variational encoder unit consists of a multi-layer feedforward neural network that transforms the observed data Mapping to latent space Z;

[0075] The physical constraint unit converts the mechanical equations into constraint terms through the Lagrange multiplier method;

[0076] The probabilistic graphical inference unit uses the Monte Carlo variational inference algorithm to estimate the posterior distribution.

[0077] The entire framework achieves a deep integration of mechanical knowledge and data-driven approaches by alternately optimizing the parameters of these three units. In high-rise suspension structure applications, the framework specifically addresses the temperature-dependent nature of steel tie rods by introducing temperature-stress relationships as physical constraints, improving the ability to separate temperature effects from structural issues.

[0078] Step 2: Based on the fusion results of BIM model data and sensor monitoring data, a dynamic physical causal graph learning module is trained to identify the causal relationship network in the high-rise suspended structure;

[0079] Step 2.1, initialize the dynamic causal graph;

[0080] The initial topology is constructed using sensor monitoring points as nodes. For high-rise suspended structures, these nodes include tension rod stress sensors on each floor, displacement sensors at connection nodes, and temperature sensors. The initial topology is determined based on the component connection relationships in the BIM model.

[0081] Step 2.2, learning dynamic causal relationships;

[0082] Using the physical constraint Bayesian network learning algorithm, the dynamic causal relationship between nodes is learned based on time series monitoring data. The formula is as follows:

[0083] ;

[0084] in is the probability function; for The causal graph structure of the moment; for The causal graph structure of the moment; for Observation data at the moment; For physical constraints; is the conditional probability separator;

[0085] The algorithm overcomes the limitations of purely data-driven methods by combining monitoring data and physical knowledge, thereby improving the accuracy of causal discovery.

[0086] The specific implementation of the physical constraint Bayesian network learning algorithm is based on the Structural Equation Modeling (SEM) framework, which includes a graph structure search module and a parameter learning module. The graph structure search module uses a scoring-based approach to define the scoring function. , indicating that the given data and physical constraints Φ. This module first generates a set of candidate graphs and then uses a greedy search algorithm to select the optimal graph structure. The parameter learning module optimizes model parameters using an expectation-maximization algorithm with physical constraints based on the principle of maximum likelihood estimation. In practical high-rise suspended structure monitoring, this algorithm maintains the sparse connectivity of the graph structure, making it particularly suitable for capturing weak correlations between tie rods on different floors of a suspended structure while also satisfying mechanical equilibrium constraints.

[0087] Step 2.3, optimize sparsity;

[0088] Applying structural sparsity regularization, the causal graph is optimized based on the “weak correlation” characteristics of high-rise suspension structures. The optimization objective function is:

[0089] ;

[0090] in is the loss function of the graph; is the natural logarithm function; is the probability function; is the observation data; is a cause-effect diagram; For physical constraints; is the sparsity balance parameter; is the L1 norm of the graph structure; is the conditional probability separator;

[0091] This step is specifically designed for the characteristic that each floor in the suspended structure is connected by independent tension rods, and can accurately capture the "weak correlation" causal relationship between floors.

[0092] Alternatively, in some embodiments, the structural sparsity regularization may adopt a more complex group sparsity regularization method, which is expressed as:

[0093] ;

[0094] in The loss function of the graph after adopting a more complex group sparsity regularization method; is the natural logarithm function; is the probability function; is the observation data; is a cause-effect diagram; For physical constraints; is the L1 norm balance parameter; is the L1 norm of the graph structure; is the group norm balance parameter; is the group norm of the graph structure; is the conditional probability separator;

[0095] This method is particularly suitable for groups of components with similar functions in high-rise suspended structures (such as multiple tie rods on the same floor), and can better capture the association patterns between components with similar functions.

[0096] In another implementation, a priori graph structure initialization method based on expert knowledge can be employed, incorporating the experience of structural engineers to provide a better initial state for the dynamic causal diagram. Specifically, the influence matrix calculated by structural analysis software can be used as prior knowledge to reduce the search space for causal discovery and accelerate algorithm convergence. This approach is suitable for scenarios where a clear understanding of the structural system is available.

[0097] Step 2.4, dynamically update the causal graph structure;

[0098] As the structural state and external environment change, the causal network is updated periodically to ensure that the model can adapt to the dynamic changes of high-rise suspended structures in different construction and usage stages.

[0099] Step 3: Using the identified causal network, we construct and train a hierarchical attention perception module to capture structural abnormal patterns.

[0100] Specifically, step 3 includes the following sub-steps:

[0101] Step 3.1, construct a three-level hierarchical attention network;

[0102] The three-level hierarchical attention network includes a temporal attention layer , spatial attention layer and component attention layer , formally expressed as:

[0103] ;

[0104] in is the hierarchical attention function; To input monitoring data; is the temporal attention layer function; is the spatial attention layer function; is the component attention layer function;

[0105] The network is specifically designed for the local to global anomaly propagation patterns that may occur in high-rise suspended structures, and is capable of capturing structural anomalies across scales.

[0106] The implementation of the three-level hierarchical attention network utilizes a multi-branch parallel computing architecture, consisting of a feature extraction backbone network and three parallel attention branches. The feature extraction backbone network uses a one-dimensional convolutional neural network to extract temporal features, combined with a graph convolutional network to capture spatial topological relationships. The three attention branches implement attention mechanisms in different dimensions: the temporal attention branch uses a bidirectional long short-term memory (BiLSTM) to capture long-term dependencies; the spatial attention branch uses a graph attention network (GAT) to handle irregular spatial topology; and the component attention branch uses a multi-layer perceptron (MLP) to process component attribute information. The outputs of each branch are integrated through an adaptive fusion module to generate the final attention weights. In monitoring high-rise suspended structures, this network specifically focuses on steel tie rod connection nodes, which are often critical areas of deformation anomalies.

[0107] Step 3.2, temporal attention layer training;

[0108] For monitoring data of different time scales, a multi-head self-attention mechanism is used to identify key time points. For the input sequence:

[0109] ;

[0110] in represents a collection of time series monitoring data, 、 、 Represents the first, second, and Monitoring data of time steps, is the total length of the time series.

[0111] Calculate attention weights:

[0112] ;

[0113] in for The attention weight of the moment; is a soft maximization function; is the query weight matrix; is the key weight matrix; for Input data at the moment; Input data for attention; is the transpose operator; is the square root function; is the feature dimension.

[0114] This layer is able to identify early signals of abnormal structural behavior at different time scales (minutes, hours, days).

[0115] Step 3.3, spatial attention layer training;

[0116] Using the graph attention network, we can learn the spatial correlation between different monitoring points based on the spatial distribution and structural topological relationship of sensors. and its neighbors , calculate the attention coefficient:

[0117] ;

[0118] in For nodes and The attention coefficient between is a soft maximization function; is the activation function of the rectified linear unit with leakage; is the learnable parameter vector; is the transpose operator; is the learnable parameter matrix; For nodes The eigenvector of For nodes The eigenvector of is the vector concatenation operator.

[0119] This layer is particularly suitable for capturing the “weakly correlated” spatial relationships between floors in a suspended structure.

[0120] Step 3.4, component attention layer training;

[0121] According to the component type, functional importance and other information in the BIM model, dynamic attention weights are assigned to different structural components. The component attention weight is calculated as:

[0122] ;

[0123] in For components The attention weight of is the activation function; is the learnable parameter matrix of component attention; For components The eigenvector of Components extracted from the BIM model The attribute vector of .

[0124] This layer ensures that higher attention is paid to the critical components of the structure (such as connection nodes and steel ties).

[0125] For example, in practical applications, different attention weight calculation methods can be used for different types of monitoring scenarios. For construction phase monitoring, a dynamic gated attention mechanism can be used, which is expressed as:

[0126] ;

[0127] in To adopt the dynamic gated attention mechanism The attention weight of For components The gating coefficient of is the activation function; is the learnable parameter matrix of component attention; For components The eigenvector of Components extracted from the BIM model The attribute vector of .

[0128] This method can automatically adjust the attention paid to different components according to the construction stage, and give higher weight when the component status is changed during construction (such as installation and prestressing).

[0129] In some implementations, an uncertainty-based attention allocation strategy can also be introduced. This strategy estimates the uncertainty of the monitoring data through a Bayesian neural network and allocates more attention to high-uncertainty areas:

[0130] ;

[0131] in To allocate more attention to high uncertainty areas, The attention weight of is the activation function; is the learnable parameter matrix of component attention; For components The eigenvector of Components extracted from the BIM model Attribute vector of ; For components uncertainty in the estimate.

[0132] This approach is particularly suitable for situations where sensor noise is inconsistent or data is missing in certain areas, and can improve the robustness of the monitoring system to uncertainty.

[0133] Step 4: Based on the captured structural anomaly patterns, a physical constraint neural ordinary differential equation module is developed to accurately predict the dynamic behavior of high-rise structures;

[0134] Step 4.1, establish the neural ODE model;

[0135] A neural ordinary differential equation model is established to express the time evolution of the structural state as:

[0136] ;

[0137] in Structural state About time The derivative of for The structural state vector at time t; The parameters are Neural network function; is the physical constraint balance parameter; is the physical model term function; is the time variable; are the neural network parameters.

[0138] The model explicitly introduces physical constraints to ensure that the prediction results conform to the laws of structural mechanics.

[0139] The specific implementation of the neural ordinary differential equation model adopts a hybrid architecture, including a physics-independent neural network part and physical model term functions . The neural network part adopts a multi-layer residual network architecture, which includes three hidden layers with jump connections, 64 neurons in each layer, and uses the GELU activation function. The physical constraint part is based on a simplified structural dynamics equation, and specifically introduces physical quantities such as the axial stiffness of the tie rod and the influence of temperature changes for high-rise suspension structures. The model training uses the Adam optimizer, the learning rate is set to 0.001, the batch size is 64, and the number of training iterations is 1000 rounds. For numerical integration solutions, the Dormand-Prince method with adaptive step size is adopted, and the tolerance is set to 1e-5. In practical applications, the model can effectively capture the nonlinear dynamic response of high-rise suspension structures under conditions such as temperature changes and load redistribution.

[0140] Step 4.2, design a hybrid loss function;

[0141] Design a hybrid loss function to optimize both data fitting accuracy and consistency with physical laws:

[0142] ;

[0143] in is the total loss function; Fit a loss function to the data; weighing parameters for physical losses; is the physical constraint loss function.

[0144] The physical constraint loss is specifically defined as:

[0145] ;

[0146] in is the physical constraint loss function; is the square of the L2 norm; is the physical model term function; for The structural state vector at time t; Structural state About time The derivative of is the time variable.

[0147] Physical constraint losses ensure that model predictions are consistent with both observed data and the laws of structural mechanics.

[0148] Step 4.3, adaptive solution;

[0149] Adopting adaptive step-size numerical integration method to solve the neural ordinary differential equation, dynamically adjusting the calculation step size, and improving the calculation efficiency while ensuring the accuracy. , the solution process is expressed as:

[0150] ;

[0151] in for The structural state at the moment; for The structural state at the moment; For arrive The definite integral operator of ; The parameters are Neural network function; for The structural state vector at time t; is the physical constraint balance parameter; is the physical model term function; is the time differential element; is the time variable; is the starting time; is the end time; are the neural network parameters.

[0152] This method is particularly suitable for dealing with the multi-time-scale dynamic responses that may occur in high-rise suspended structures.

[0153] Alternatively, different numerical solution strategies can be employed based on specific application requirements and computing resource constraints. For scenarios requiring high precision but with sufficient computing resources, high-order Runge-Kutta methods, such as the RK45 (Dormand-Prince) algorithm, can be used. For applications with high real-time requirements, simplified or improved Euler methods can be employed, sacrificing a certain degree of accuracy in exchange for computational speed.

[0154] For example, when running on an edge computing device, an approximate solution based on Taylor expansion can be used:

[0155] ;

[0156] in for The approximate structural state at the moment; for The structural state at the moment; is the time step; The parameters are Neural network function; is the coefficient of the second-order term; is the neural network function derivative with respect to time; is the time variable; are the neural network parameters.

[0157] This method has low computational overhead and is suitable for monitoring terminal devices with limited resources.

[0158] In another implementation, for applications where long-term behavior is of particular interest, an inverse stochastic differential equation solution method can be used, which is more suitable for capturing rare events and long-tail distribution phenomena that may exist in high-rise suspended structures:

[0159] ;

[0160] in for The structural state at the moment; End time structural state; For arrive The definite integral operator of ; The parameters are Neural network function; is the integration variable The structural state at the moment; is the physical constraint balance parameter; is the physical model term function; is the differential element of the integral variable; is the noise factor; for The Wiener process of moments; is the end time; is the time variable; is the integral variable; are the neural network parameters.

[0161] This method can more effectively explore the possible evolution paths of the system under different initial conditions and is suitable for risk assessment and extreme event analysis.

[0162] Step 4.4, enhance robustness;

[0163] By integrating the results of multiple numerical solution methods, the model's robustness to noise and abnormal data is improved. The integrated prediction result is calculated as:

[0164] ;

[0165] in for The integrated prediction results at the moment; is the symbol for summation; For the The weight of the method; For the Numerical solution method in The prediction results at the moment; is the total number of numerical solution methods.

[0166] This effectively reduces the numerical instability that may be caused by a single solution method and improves the accuracy of predicting the complex dynamic behavior of high-rise suspended structures.

[0167] Step 5: Based on the predicted structural dynamic behavior, implement the BIM semantic mapping and visualization module to establish a two-way real-time mapping between monitoring data and BIM model elements;

[0168] Step 5.1, establish semantic mapping;

[0169] Establish a BIM semantic mapping function to map monitoring data, causal relationships, and uncertainty information to BIM model elements:

[0170] ;

[0171] in It is the BIM semantic mapping function; is the state estimation data; For causal data; is the uncertainty measurement data; It is a BIM model element; To map direction symbols

[0172] The BIM semantic mapping function transforms abstract monitoring results into three-dimensional visual representations through semantic association.

[0173] The specific implementation of the BIM semantic mapping function is based on a multi-level semantic indexing architecture, including three levels: element-level indexing, component-level indexing, and system-level indexing. The bottom-level element-level index establishes a direct correspondence between monitoring data points and geometric elements in the BIM model; the middle-level component-level index organizes related elements into meaningful structural components, such as steel tie rods, connection nodes, etc.; the top-level system-level index constructs a relationship network between components, corresponding to the causal graph structure. The mapping process uses a graph-based search algorithm. First, the element-level correspondence is determined through spatial position matching, then the component-level mapping is established based on the component classification information of the BIM model, and finally the system-level mapping is constructed using topological relationships. In high-rise suspended structure applications, this function particularly strengthens the semantic understanding of key components in the suspension system (such as tie rods and connection nodes), so that the monitoring results can accurately correspond to the structural entities.

[0174] Step 5.2, develop the visualization interface;

[0175] Develop a layered visualization interface that supports multi-scale and multi-dimensional structural state display. The interface includes:

[0176] Global view: Visualize the overall structural health status, using color coding to indicate the degree of abnormality;

[0177] Local view: Detailed status information of key nodes and components, including deformation, stress and warning level;

[0178] Time view: historical evolution of structural status and future forecast trends, supporting timeline interaction.

[0179] The visual interface enables engineers to intuitively understand the complex status information of high-rise suspension structures.

[0180] In some implementations, different visualization strategies can be employed based on usage scenarios and user needs. For example, for mobile devices used on-site, simplified 2D representations and symbol encoding can be employed to reduce rendering overhead while retaining key information. For large-screen displays in control centers, immersive 3D visualization can be employed, combined with ambient lighting and material rendering, to provide richer information presentation.

[0181] Alternatively, on-site visualization can be implemented using augmented reality (AR) technology. Using mobile device cameras to identify structural locations, monitoring data and warning information can be directly overlaid onto the actual structural image, creating a "see-through" effect and allowing on-site personnel to intuitively understand the status of invisible components. This approach is particularly suitable for inspection and maintenance scenarios.

[0182] For collaborative teamwork, a cloud-based shared visualization platform can be used, allowing different professionals (such as structural engineers, construction teams, and owners) to simultaneously view the same monitoring information, displaying different levels and types of data based on their needs. This approach supports remote collaboration and multi-level decision-making, improving emergency response efficiency.

[0183] Step 5.3, adaptive resource allocation;

[0184] Implement adaptive computing resource allocation. Based on the output of dynamic causal graph and hierarchical attention, dynamically adjust computing resource allocation and optimize computing efficiency:

[0185] ;

[0186] in Assigned to monitoring points computing resources; is the total computing resources; is an exponential function; For monitoring points Attention score; For the sum symbol; For monitoring points Attention score; is the total number of monitoring points.

[0187] This mechanism ensures that limited computing resources are allocated preferentially to high-risk areas, reducing computing overhead while maintaining monitoring accuracy.

[0188] Step 5.4, build decision support;

[0189] Build a predictive maintenance decision support system that automatically generates maintenance recommendations based on model prediction results, including:

[0190] Prioritization: Sort components according to risk assessment and identify areas that require priority inspection;

[0191] Intervention recommendations: Provide solutions to potential problems based on causal analysis;

[0192] Action Plan: Generates a specific inspection and maintenance schedule.

[0193] The system converts complex monitoring data into directly executable decision recommendations, improving the operation and maintenance efficiency of high-rise suspended structures.

[0194] A BIM high-rise suspended structure monitoring system based on machine learning is used to execute the above-mentioned BIM high-rise suspended structure monitoring method based on machine learning, comprising:

[0195] Physically constrained variational causal reasoning module, used to integrate BIM model data with sensor monitoring data to form a unified data processing foundation;

[0196] A dynamic physical causal graph learning module for identifying causal networks in high-rise suspended structures;

[0197] A hierarchical attention perception module to capture structural abnormality patterns from multiple temporal and spatial scales;

[0198] Physically constrained neural ordinary differential equations module for accurate prediction of the dynamic behavior of high-rise suspended structures;

[0199] BIM semantic mapping and visualization module is used to establish a two-way real-time mapping between monitoring data and BIM model elements.

[0200] Here, the present invention provides an implementation example:

[0201] This application example targets a 28-story commercial and office complex. Floors 11-28 utilize a suspended structure, with top trusses and vertical steel tie rods supporting the lower floors. This building exhibits typical "weakly correlated" structural characteristics, with loads between floors primarily transmitted via steel tie rods, which range in diameter from 80 to 120 mm and in length from 3.6 to 4.2 m. Monitoring requirements cover both the construction and operational phases, focusing on changes in tie rod stress, deformation at connection points, and overall structural stability.

[0202] The main challenges faced by the project include:

[0203] The structure is sensitive to temperature changes, and thermal expansion and contraction of steel tie rods may lead to stress redistribution;

[0204] During the construction process, the load state is constantly changing, and the structural status needs to be evaluated in real time;

[0205] The irregular design makes it difficult for traditional monitoring methods to accurately assess the structural safety status.

[0206] This application example deploys 158 sensors, including 48 strain sensors (installed on steel tie rods), 36 displacement sensors (installed at key connection points), 32 tilt sensors, and 42 temperature sensors. All sensors are linked to the BIM model to form a semantically linked digital twin.

[0207] First, the structural geometry model and material property information are extracted from the project's BIM system to generate a digital representation of the structure. Some key component information extracted from the BIM model is shown in Table 1:

[0208] Table 1: Example of structural component information extracted from the BIM model

[0209]

[0210] Among them, Φ120, Φ100, and Φ80 respectively indicate that the diameters of the circular cross section are 120 mm, 100 mm, and 80 mm;

[0211] Subsequently, the real-time monitoring data of the structure is collected by combining with the deployed sensor network, and after preprocessing, it forms a monitoring data stream that can be used for analysis. Some examples of sensor data after preprocessing are shown in Table 2:

[0212] Table 2: Example of sensor monitoring data (sampling time: 2022-07-15 14:00-15:00)

[0213]

[0214] Based on structural mechanics theory, physical constraints for the suspension structure were constructed, including the relationship between the axial force and deformation of steel tie rods, the effect of temperature changes on tie rod stress, and node equilibrium equations. These physical constraints were converted into mathematical expressions and embedded into the PC-VCI framework. Table 3 shows some key physical constraints and their mathematical expressions.

[0215] Table 3: Examples of physical constraints and their mathematical expressions

[0216]

[0217] Through the above data and constraints, the PC-VCI framework achieves the overall inference of the structural state and establishes the causal network foundation between the monitoring variables.

[0218] Based on the results of the PC-VCI framework, an initial dynamic causal graph was constructed. The nodes of the graph correspond to sensor monitoring points, and the initial topology is determined based on the component connection relationships in the BIM model. Some key nodes and their initial connection relationships are shown in Table 4:

[0219] Table 4: Example of causal graph nodes and their physical meanings

[0220]

[0221] The system then applied a physically constrained Bayesian network learning algorithm based on three months of historical monitoring data (April 15, 2022, to July 15, 2022) to identify true causal relationships between nodes. The system identified the impact of temperature changes on the strain in steel tie rods, as well as weak correlations between tie rods on different floors. Table 5 shows some of the key causal relationships discovered and their strengths.

[0222] Table 5: Key causal relationships and their strengths discovered by the system

[0223]

[0224] Based on the established causal graph, the system constructed and trained a three-level hierarchical attention network. The network consists of a temporal attention layer, a spatial attention layer, and a component attention layer, which together form the attention mechanism. After model training, examples of the attention weights assigned to different structural components are shown in Table 6:

[0225] Table 6: Component attention weights assigned by the hierarchical attention network

[0226]

[0227] To accurately predict the dynamic behavior of the structure, the system established a physical constraint neural ordinary differential equation model. The performance comparison of this model under different prediction targets and time lengths is shown in Table 7:

[0228] Table 7: Prediction accuracy of the physical constraint neural ordinary differential equation model

[0229]

[0230] The system establishes a bidirectional mapping between monitoring data and BIM model elements through the BIM semantic mapping function. The mapping success rate statistics and related performance indicators are shown in Table 8:

[0231] Table 8: BIM semantic mapping performance indicators

[0232]

[0233] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A BIM high-rise suspended structure monitoring method based on machine learning, characterized in that: include: A physical constraint variational causal reasoning framework is constructed to integrate BIM model data with sensor monitoring data. The steps of constructing the physical constraint variational causal reasoning framework include: Extract the geometric model, material properties and component relationships of high-rise suspended structures from the BIM system; Deploy a multimodal sensor network to collect real-time monitoring data of the structure; Construct physical constraints based on structural dynamics equations; Construct a physically constrained variational causal inference framework that combines structural dynamics equations, Bayesian causal inference, and deep learning through a unified model; Based on the fusion results of BIM model data and sensor monitoring data, a dynamic physical causal graph learning module is trained to identify the causal relationship network in the high-rise suspended structure. The steps of training the dynamic physical causal graph learning module include: Initialize the dynamic causal graph, use sensor monitoring points as nodes, and build the initial topological structure; Using the physically constrained Bayesian network learning algorithm, the dynamic causal relationship between nodes is learned based on time series monitoring data; Applying structural sparsity regularization, we optimize the causal graph based on the "weak correlation" characteristics of high-rise suspension structures. Dynamically update the causal graph structure, and periodically update the causal relationship network as the structural state and external environment change; Using the identified causal network, we construct and train a hierarchical attention perception module to capture structural abnormal patterns; Based on the captured structural anomaly patterns, a physical constraint neural ordinary differential equation module is developed to accurately predict the dynamic behavior of the high-rise structure. The steps of developing the physical constraint neural ordinary differential equation module include: A neural ordinary differential equation model is established to represent the time evolution of the structural state as a combination of a neural network and a physical model; Design a hybrid loss function to optimize data fitting accuracy and consistency with physical laws; Adopting adaptive step-size numerical integration method to solve neural ordinary differential equations and dynamically adjust the calculation step size; By integrating the results of multiple numerical solution methods, the model's robustness to noise and abnormal data is improved; Based on the predicted structural dynamic behavior, the BIM semantic mapping and visualization module is implemented to establish a two-way real-time mapping between monitoring data and BIM model elements.

2. The BIM high-rise suspended structure monitoring method based on machine learning according to claim 1 is characterized in that: The steps of constructing and training the hierarchical attention perception module include: Construct a three-level hierarchical attention network, including a temporal attention layer, a spatial attention layer, and a component attention layer; Temporal attention layer training uses a multi-head self-attention mechanism to identify key time points for monitoring data at different time scales; Spatial attention layer training uses a graph attention network to learn the spatial correlation between different monitoring points based on the spatial distribution and structural topological relationship of sensors; Component attention layer training assigns dynamic attention weights to different structural components based on the component type and functional importance information in the BIM model.

3. The BIM high-rise suspended structure monitoring method based on machine learning according to claim 1 is characterized in that: The steps of implementing the BIM semantic mapping and visualization module include: Establish BIM semantic mapping functions to map monitoring data, causal relationships, and uncertainty information to BIM model elements; Develop a layered visualization interface to support multi-scale and multi-dimensional structural state display; Implement adaptive computing resource allocation, dynamically adjusting computing resource allocation based on the output of dynamic causal graphs and hierarchical attention; Build a predictive maintenance decision support system to automatically generate maintenance recommendations based on model prediction results.

4. The BIM high-rise suspended structure monitoring method based on machine learning according to claim 1 is characterized in that: The structural sparsity regularization is implemented through a graph structure loss function, which includes a probability logarithm term and a graph structure regularization term to promote the sparsity of the graph.

5. The BIM high-rise suspended structure monitoring method based on machine learning according to claim 2 is characterized in that: The three-level hierarchical attention network combines temporal attention, spatial attention and component attention in a cascade manner, and performs attention calculations on the input monitoring data in sequence.

6. The BIM high-rise suspended structure monitoring method based on machine learning according to claim 1 is characterized in that: The neural ordinary differential equation model combines neural network representation with physical constraints to achieve the prediction of the dynamic evolution of structural states over time.

7. A BIM high-rise suspended structure monitoring system based on machine learning, characterized in that: A method for monitoring a BIM high-rise suspended structure based on machine learning, for executing any one of claims 1 to 6, comprising: Physically constrained variational causal reasoning module, used to integrate BIM model data with sensor monitoring data to form a unified data processing foundation; A dynamic physical causal graph learning module for identifying causal networks in high-rise suspended structures; A hierarchical attention perception module to capture structural abnormality patterns from multiple temporal and spatial scales; Physically constrained neural ordinary differential equations module for accurate prediction of the dynamic behavior of high-rise suspended structures; BIM semantic mapping and visualization module is used to establish a two-way real-time mapping between monitoring data and BIM model elements.

Images