Civil engineering structure crack prediction method based on bim and deep learning

Abstract

The application discloses a kind of civil engineering structure crack prediction methods based on BIM and deep learning.The method first parses BIM data to construct semantic fusion mechanical dual graph, sets physical and virtual nodes to maintain mechanical connectivity, and embeds component attributes into graph nodes.Then, map sparse sensor time series data to graph structure, use physical information neural network and dynamic equation constraints to invert global stiffness field and locate potential damage subgraph.For this subgraph, construct two kinds of hypothesis models of continuous damage and discontinuous fracture respectively, calculate the comprehensive residual energy of both on observed data.Finally, introduce semantic topological penalty factor based on component reinforcement ratio calculation, compare energy difference using Bayes criterion to determine crack state, and quantitatively calculate equivalent crack depth.The application can accurately distinguish material degradation and structural crack under unlabeled and sparse sensing conditions, and realize dynamic closed-loop updating of monitoring model.

Description

Technical Field

[0001] This invention relates to the field of civil engineering structural health monitoring technology, specifically a method for predicting cracks in civil engineering structures based on BIM and deep learning. Background Technology

[0002] In the field of civil engineering, real-time health monitoring of large building structures is a key measure to ensure structural safety and durability. With the development of IoT technology, sensor network-based structural health monitoring technology has been widely used. However, due to economic costs and construction conditions, the number of sensors deployed in actual projects is often limited and sparsely distributed, making it difficult to cover the entire area of ​​a large building structure. Existing monitoring data processing methods are mostly based on simple geometric interpolation or statistical laws to infer the structural response in areas where no sensors are deployed, often ignoring the actual physical connections and mechanical transmission paths between building components. Although Building Information Modeling (BIM) contains rich geometric and material information, in existing monitoring systems, BIM data is usually only used as a static display model and has not been effectively transformed into the mechanical topological priors required for dynamic monitoring. This results in the spatial mapping of sparse sensor data lacking physical support, making it difficult to accurately reflect the overall stress state of the structure.

[0003] Furthermore, with the introduction of artificial intelligence technology, damage identification methods based on deep learning have gradually become a research hotspot. These methods typically rely on massive amounts of sample data with damage labels for model training. However, in actual civil engineering, obtaining crack damage labels for real structures is extremely difficult, and synthetic data generated through finite element simulation differs in distribution from the real environment. While traditional pure data-driven deep learning models perform excellently in fitting nonlinear relationships, due to the lack of constraints from physical mechanisms, their outputs often do not satisfy the basic structural dynamic equilibrium equations. This results in weak generalization ability when facing conditions not present in the training set, and the physical parameters obtained through inversion lack interpretability, making them difficult to directly use for structural safety assessment.

[0004] In the specific assessment stage of damage identification, existing monitoring methods typically infer damage by monitoring changes in modal parameters or a decrease in the stiffness matrix value. However, a decrease in structural stiffness may stem from the overall degradation of material properties (such as concrete aging and a decrease in the elastic modulus) or from the geometric fracture of local components (such as sectional failure caused by cracks). Existing techniques struggle to effectively distinguish between these two fundamentally different damage modes based solely on vibration data in the absence of detailed on-site inspections. In particular, they neglect the inhibitory effect of design semantics such as component reinforcement ratio and component type on crack propagation, and lack a mechanism to probabilistically verify different damage hypotheses by incorporating prior knowledge of structural design, which can easily lead to misjudgments or omissions. Summary of the Invention

[0005] The first aspect of this invention provides a method for predicting cracks in civil engineering structures based on BIM and deep learning. This method utilizes parametric data from Building Information Modeling (BIM) to establish physical priors, and combines this with deep learning algorithms to process sparse sensor data, thereby inverting the structural stiffness field and determining the crack state even without labeled samples.

[0006] First, a semantic fusion mechanical dual graph based on BIM is constructed. Building information model data of the target monitoring object is acquired, and its geometric information, material properties, and component connection relationships are analyzed. Node discretization is performed, retrieving component entities containing coordinates of the sparse sensor group, and defining their geometric centers or sensor attachment points as physical nodes. For beam or column components not covered by the sparse sensor group, virtual nodes are generated by discretizing along the component axis at a preset step size. Force transfer connections are established, creating connecting edges between components with topological connections, and between adjacent nodes discretized within the same component. The initial weights of the connecting edges are calculated, specifically based on the arithmetic mean of the design elastic moduli of the components at both ends of the connection to determine the equivalent elastic modulus, combined with the equivalent transfer area and the Euclidean distance between nodes. A weighted adjacency matrix is ​​generated based on the node set and edge set to establish the topological structure of the semantic fusion mechanical dual graph. Simultaneously, non-geometric semantic attributes of the component entities are extracted, generating semantic feature vectors containing the one-heat encoded vector of the component type, normalized design elastic modulus, normalized reinforcement ratio, and normalized section moment of inertia, which are then embedded into the corresponding graph nodes.

[0007] Secondly, data mapping and dynamic equation initialization of the sparse sensor array are performed. The initial mass matrix is ​​calculated using the discrete volume element volume and material density from the geometric information. Based on the Rayleigh damping model, the initial damping matrix is ​​calculated using the initial mass matrix and the initial global stiffness matrix generated based on BIM design parameters, thereby constructing the structural dynamic equilibrium equations. An acceleration selection matrix is ​​constructed to establish a linear mapping between the accelerometer output sequence and the global node acceleration vector. A global strain observation matrix containing the derivatives of finite element shape functions is constructed to establish a linear mapping between the strain gauge output sequence and the global node displacement vector. The time-series data acquired by the sparse sensor array is mapped to the corresponding physical nodes in the dual-graph structure.

[0008] Next, a physical constraint-based inverse reconstruction of the stiffness field is performed. A physical information neural network is constructed, comprising parameter estimation and state estimation branches. The parameter estimation branch takes the adjacency matrix and node semantic feature matrix of the dual graph as input, outputs a stiffness correction coefficient vector, and assembles the estimated value of the global stiffness matrix using the finite element stiffness ensemble principle. The state estimation branch takes the time coordinate and graph embedding features as input, outputs the displacement estimates of the global nodes, and calculates the velocity and acceleration estimates using an automatic differentiation algorithm. A total loss function is constructed, including a correlation term for matching measured data, a physical residual term for satisfying the dynamic equilibrium equations, and a stiffness smoothing regularization term for maintaining the continuity of topologically adjacent elements. The stiffness correction coefficient vector is optimized by minimizing the total loss function, and potential damage subgraphs are identified based on the converged global stiffness matrix estimates.

[0009] Subsequently, topological variation generation based on BIM semantics is performed. For the potentially damaged subgraph, a first hypothetical model that maintains the graph's topological structure is generated. Simultaneously, specific connection edges are identified to generate a mutated dual graph as a second hypothetical model. The logic for identifying specific connection edges is as follows: the stress tensor of each component element in the subgraph is calculated using the derivative of the finite element shape function, and the maximum principal tensile stress is extracted. The semantic vulnerability factor is determined by combining the node semantic feature vector. The crack tendency score is calculated by weighted summation of the maximum principal tensile stress and the semantic vulnerability factor, and the edge with the highest score is selected as the specific connection edge.

[0010] Next, adversarial residual verification and Bayesian confidence discrimination are performed. The combined residual energy of the first and second hypothetical models on the time-series data is calculated, and this energy is a weighted average of the data consistency residual and the physical equilibrium residual. A semantic topological penalty factor is calculated based on the normalized reinforcement ratio and component type coefficient of the component corresponding to a specific connection edge, used to characterize the prior cost of removing that connection edge. The Bayesian discriminant index is calculated, which is the difference between the combined residual energy of the first hypothetical model and the combined residual energy of the second hypothetical model, minus the semantic topological penalty factor. If the Bayesian discriminant index is greater than zero, the crack is considered established.

[0011] Finally, the crack status output and model closed-loop update are performed. If a crack is determined, the target stiffness correction coefficient corresponding to a specific connection edge is determined. Based on the cubic proportionality between the moment of inertia and the geometric height of the section, the equivalent crack depth is calculated using the target stiffness correction coefficient and the member section height. The weights corresponding to the specific connection edge are reset to zero in the weighted adjacency matrix of the semantic fusion mechanical dual graph, and the crack depth is recorded in the attributes of the disconnected node. If a crack is not determined but material damage exists, the design elastic modulus value in the semantic feature vector of the corresponding node is updated according to the estimated value of the global stiffness matrix. Based on the updated semantic fusion mechanical dual graph, the reference global stiffness matrix for the next monitoring time is reassembled to reset the structural dynamic equilibrium equations.

[0012] A second aspect of this invention provides a civil engineering structure crack prediction system based on BIM and deep learning. The system includes a processor, a memory, and a communication interface. The memory stores computer program instructions, and the processor executes these instructions to implement the steps described in the first aspect.

[0013] This invention provides a method for predicting cracks in civil engineering structures based on BIM and deep learning. It has the following beneficial effects: 1. This invention solves the problem of disconnected mechanical transmission paths caused by sparse sensor distribution by constructing a semantic fusion mechanical dual graph containing physical and virtual nodes. By parsing BIM data, static semantic information such as component material properties and reinforcement ratios are embedded into the graph structure, enabling deep learning models to use prior physical knowledge of the building structure for reasoning. This method maintains the integrity of the mechanical transmission path even with a limited number and uneven distribution of sensors, enabling prediction of structural responses in uncovered areas and inversion of the global stiffness field.

[0014] 2. This invention utilizes a physical information neural network combined with structural dynamic equilibrium equations to achieve unsupervised physical parameter inversion under conditions lacking crack label data. By incorporating the imbalance quantities of the dynamic equations as physical residuals into the loss function, the displacement, velocity, and stiffness parameters output by the network are constrained to satisfy physical conservation laws. This overcomes the dependence of purely data-driven methods on a large number of labeled samples, ensures that the inverted stiffness matrix has a clear physical meaning, and improves the model's generalization ability in practical engineering applications.

[0015] 3. This invention effectively distinguishes between two physical modes: material performance degradation and structural crack fracture, by establishing a competing hypothesis model of continuous damage and discontinuous fracture, combined with a Bayesian confidence discrimination mechanism. This method introduces a topological penalty factor based on BIM semantics as a priori cost to quantitatively constrain the ease of crack formation. By comparing the residual energy of different hypothesis models' interpretation of observed data with the prior cost, it can accurately determine the existence of cracks at the probabilistic level, avoiding misjudgments caused by relying solely on a single indicator of stiffness reduction. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the method for predicting cracks in civil engineering structures based on BIM and deep learning provided in this embodiment of the invention. Figure 2 This is a functional module block diagram of the civil engineering structure crack prediction system provided in the embodiments of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see the appendix Figure 1 This invention provides a method for predicting cracks in civil engineering structures based on BIM and deep learning, comprising the following steps: Step S100: Construct a semantic fusion mechanical dual graph based on BIM; Obtain the building information model data of the target monitoring object, analyze the geometric information, material properties and component connection relationships in the data, and transform it into a semantic fusion mechanical dual graph structure. .

[0019] Specifically, node set This includes physical nodes and virtual nodes. Physical nodes are mapped to the actual locations of the deployed sparse sensor array and the component entities in the BIM model; virtual nodes are generated based on the geometric centers or connection points of BIM components and are distributed in areas not covered by the sparse sensor array to maintain the connectivity of the mechanical transfer paths. Edge sets The force transmission path between components is characterized, and the initial weights of the edges are set based on the component connection stiffness and geometric adjacency relationships in the BIM. At the same time, the component type, design elastic modulus, and reinforcement ratio information are extracted from the BIM, and semantic attribute vectors are generated and embedded into the corresponding graph nodes.

[0020] Step S200: Sparse sensor group data mapping and dynamic equation initialization; Semantic fusion mechanical dual graph structure Construct the structural dynamic equilibrium equations: ; in: , , They are respectively The displacement, velocity, and acceleration vectors of all nodes at any given time; The initial mass matrix, The initial damping matrix was calculated based on BIM design parameters and remained constant during the monitoring process. Let be the time-varying global stiffness matrix to be solved; This is the external load vector.

[0021] Acquire time-series data collected by a sparsely distributed group of sensors, and map this time-series data to a semantic fusion mechanical dual graph structure. The corresponding physical nodes. For virtual nodes without deployed sensors, their state variables are used as variables to be solved in the initial stage.

[0022] Step S300: Inverse reconstruction of stiffness field based on physical constraints; Construct a physical information neural network and integrate semantics with the mechanical dual graph structure. Using sparse sensor observation data as input, the output is an estimate of the global stiffness matrix. .

[0023] A loss function incorporating physical residuals is constructed, and the stiffness parameters are optimized by minimizing this loss function. The physical residuals characterize the estimated values. The imbalance generated after substituting into the dynamic equation in step S200. This is detected after the network converges. If any region exists with a stiffness value lower than a preset threshold, that region is marked as a potential damage submap. Step S400: Generation of topological variations based on BIM semantics; For potential damage subgraphs Two hypothesis models are generated for state verification: 1. First assumption model (continuous medium damage model): Keeping the graph topology unchanged, the low stiffness parameters obtained from the inversion in step S300 are directly used as the stiffness properties of this region.

[0024] 2. Second Hypothesis Model (Discontinuous Fragmentation Model): Based on the semantic attribute vectors of nodes within the subgraph, identify connecting edges and reset the weight of these connecting edges to zero to generate a mutated dual graph. The stiffness parameters of the relevant nodes are then reset to their initial undamaged state.

[0025] Step S500: Adversarial residual verification and Bayesian confidence assessment; Calculate the interpretation residuals of the first and second hypothetical models for the sparse sensor observation data, respectively. Calculate the first physical residual energy corresponding to the first hypothetical model and the second physical residual energy corresponding to the second hypothetical model. Calculate the semantic topology penalty factor based on the semantic attributes of the cut edges.

[0026] Compare the first physical residual energy with the second physical residual energy. If the sum of the second physical residual energy and the semantic topology penalty factor is less than the first physical residual energy, the crack is determined to be established.

[0027] Step S600: Crack state output and model closed-loop update; If the crack is determined to be established, output the location of the cut edge and the equivalent crack depth. (This refers to the mutated dual graph.) Update the baseline model for the next monitoring time.

[0028] Step S100 constructs a BIM-based semantic fusion mechanical dual graph as the initialization stage of the entire monitoring method. Step S100 aims to establish the mapping foundation between the physical world and the digital computing space. This step transforms the static Building Information Model (BIM) into graph-structured data containing physical attributes and semantic features by parsing it, providing topological support for subsequent deep learning physical inversion.

[0029] Specifically, this step includes sub-steps S110 to S140.

[0030] Sub-step S110: BIM data parsing and node discretization; Parse the IFC format file of the target monitoring object to extract building component entity objects. Construct a node set based on the extracted component entities. This process includes physical node mapping and virtual node interpolation.

[0031] Physical node mapping: Obtain the 3D coordinates of the installed sparse sensor array in the global coordinate system, retrieve the component entities in the BIM model that contain these coordinates, and define the geometric center of the component or the attachment point of the sparse sensor array as a physical node. .

[0032] Virtual node interpolation: Identifies areas in the BIM model not covered by sparse sensor arrays. For beam and column members with lengths exceeding a preset threshold, discretization is performed along their axes at preset step sizes to generate internal virtual nodes; for member connections (such as beam-column joints), if no sparse sensor array is deployed, boundary virtual nodes are generated at the connection.

[0033] The generated virtual node is denoted as A node set is defined as follows: To avoid index confusion, let any node be denoted as (in Record each node spatial coordinates .

[0034] Sub-step S120: Establish force transmission connection edges and initialize weights; Based on node set Constructing edge sets Establish connecting edges. The rules include: 1. Inter-component connection: When two different components have a topological connection or share a geometric surface in BIM, a connection edge is established between the boundary nodes belonging to the two components respectively; 2. Intra-component connection: For adjacent nodes generated by discretization within the same component, establish connection edges to characterize the stress continuity within the component.

[0035] Calculate the initial weights of the connecting edges using the following formula: in: For nodes With nodes The initial weights of the connecting edges represent the stiffness and strength of the force transmission. The equivalent elastic modulus is taken as the arithmetic mean of the design elastic moduli of the components at both ends of the connection. The equivalent transmission area characterizes the size (cross-sectional area or contact area) of the force transmission contact surface. For nodes and The Euclidean distance between them, i.e., the spatial geometric length. A weighted adjacency matrix is ​​constructed based on the calculated weights. .

[0036] A weighted adjacency matrix is ​​constructed based on the calculated weights. .

[0037] Sub-step S130: Semantic feature vector encoding and embedding; Extract the non-geometric semantic attributes of each component entity and construct nodes. semantic feature vector in, For nodes The semantic feature vector. This is the one-hot encoding vector for the component type. The normalized value of the elastic modulus of the component is used for design. This is a normalized value for the reinforcement ratio of the component. This is the normalized value of the moment of inertia of the component cross section. The semantic feature vector of the node Embedded into the corresponding graph nodes, serving as the initial node features of the graph neural network.

[0038] Sub-step S140: Initial physics matrix assembly; Assemble the initial mass matrix based on BIM parameters. With the initial damping matrix .

[0039] Initial mass matrix It is a diagonal matrix, and its diagonal elements The calculation is as follows: in: The initial mass matrix The first on the diagonal 1 element (corresponding node) ). For nodes The material density of the component. For nodes The volume of the corresponding discrete volume unit. For nodes Additional mass at the location. Initial damping matrix. Calculation based on Rayleigh damping model 2 in Let be the initial damping matrix. This is the initial mass matrix. This is the initial global stiffness matrix generated based on BIM design parameters. and This is the mass proportion factor for Rayleigh Damping.

[0040] Step S200: Sparse Sensor Group Data Mapping and Dynamic Equation Initialization. After constructing the static semantic fusion mechanical dual graph in step S100, step S200 mainly involves data acquisition and time synchronization of the collected sparse sensor group data, and constructing dynamic equilibrium equations to constrain the subsequent physical inversion process based on the initial physical matrix assembled in step S140.

[0041] This step includes sub-steps S210 to S230.

[0042] Sub-step S210: Multi-source sensor data acquisition and time synchronization; The control data acquisition module acquires time-series data from a sparse sensor array deployed on the structural surface. The sparse sensor array specifically includes an accelerometer for monitoring vibration response and strain gauges for monitoring local deformation. The types of data collected include acceleration response sequences output by accelerometers. and the strain response sequence output by the strain gauge .

[0043] The acquired raw data is preprocessed, including removing DC components and bandpass filtering. An interpolation resampling algorithm is used to unify the data from sparse sensor groups with different sampling frequencies onto the same time axis t, so that the time steps Δt of the observation data from all channels are synchronized, achieving time synchronization.

[0044] Sub-step S220: Sparse observation vector construction and spatial mapping; Define the global nodal displacement vector of the structural system Where N is the total number of nodes, and d is the number of degrees of freedom for each node. An observation equation is constructed to describe the mapping relationship between sparse sensor group data and global state variables: 1. Acceleration Mapping: Constructing the Acceleration Selection Matrix ,in Let Z be the total number of channels for the accelerometers in the sparse sensor group. If the Z-th accelerometer measures the acceleration of the J-th degree of freedom, then... The elements (Z, J) are 1, and the rest are 0. The acceleration observation equation is: ; in, The measured observation vector of the accelerometer at any given time. A matrix is ​​selected for acceleration to extract sensor position data from the global state. The true acceleration vector of the global nodes at time t.

[0045] 2. Strain Mapping: Constructing the strain-displacement transformation relationship. Based on the derivative of the finite element shape function, calculate the strain-displacement matrix B for each component element. For the component element containing the k-th strain gauge, its observed strain is related to the strain of that element. Displacement of global nodes at any time The relationship is Assemble the global strain observation matrix. The strain observation equation is: ; in, The measured observation vector of the strain gauge at any given time. The global strain observation matrix contains the strain-displacement transformation relationship. The true displacement vector of all nodes at any given time. The data are combined into a system observation dataset, which is used for subsequent constraints on the displacement and acceleration fields.

[0046] Sub-step S230: Constraint formalization of the dynamic equilibrium equations; Based on the initial mass matrix assembled in step S140 With the initial damping matrix Establish the discrete dynamic equilibrium equations for the structure. These equations are used to calculate the physical residual vector. : ; in, The physical residual vector at time t represents the unbalance quantity in the dynamic equation. Initial mass matrix (constant). Initial damping matrix (constant). : To be solved Time-varying global stiffness matrix. The force is obtained through actual measurement by a force sensor or set based on known working conditions. for Displacement estimates of all nodes at time step for The estimated velocity of all nodes at any given time. for The acceleration vectors of all nodes at any given time are used as state variables to be solved during the physics inversion process, and must satisfy the following conditions: and The micro-classification constraints.

[0047] Step 300: Inverse reconstruction of the stiffness field based on physical constraints; Using the discrete dynamic equilibrium equations of the structure established in step S230 and the sparse sensor observation data constructed in step S220, step S300 uses a physical information neural network to process the sparse sensor observation data under the condition of lacking crack labels, invert the global stiffness distribution of the target monitoring object and locate the abnormal parameter areas.

[0048] This step specifically includes sub-steps S310 to S340.

[0049] Sub-step S310: Construct a physical information generation network model; Construct a deep neural network model that includes a state estimation branch and a parameter estimation branch.

[0050] Parameter estimation branch: semantic fusion of mechanical dual graph structure The adjacency matrix A and the node semantic feature matrix S are taken as input. Spatial topological features are processed through graph convolutional layers to generate implicit graph embedding features, and a stiffness correction coefficient vector that remains constant within the current monitoring time window is output based on these graph embedding features. ,in, Stiffness correction coefficient vector, output of the network parameter estimation branch. The total number of component units. .

[0051] State estimation branch: based on time coordinates The graph embedding features are used as input. A fully connected layer or a recurrent neural network layer is employed to output the global physical nodes and virtual nodes. Displacement estimate at time 1 Using automatic differentiation algorithms, for Regarding time coordinates Find the first and second derivatives to calculate the velocity estimate. and acceleration estimates .

[0052] Sub-step S320: Parametric assembly of the global stiffness matrix; Stiffness correction coefficient vector based on network output Assemble the global stiffness matrix using the finite element stiffness integration principle. .

[0053] For each component element in the graph structure Calculate its corrected element stiffness matrix. : ; in, Component unit The corrected element stiffness matrix. Component unit The stiffness correction factor has a value range of (0,1). Component unit The initial design element stiffness matrix.

[0054] Perform a global assembly operation: ; in, The assembled global stiffness matrix. Total number of units. The element coordinate transformation matrix is ​​used to map the stiffness matrix in the local coordinate system of the element to the global coordinate system. Its dimensions depend on the number of degrees of freedom of the element and the total number of degrees of freedom of the system. Transpose of the unit coordinate transformation matrix.

[0055] Sub-step S330: Construction and calculation of the physical constraint loss function; Define the total loss function This includes data consistency terms, physical residual terms, and stiffness smoothing regularization terms: ; in, This is the total loss function of the physical information neural network. This represents the data consistency loss item. This is the physical residual loss term. Stiffness smoothing regularization loss term. The preset weighting coefficients for each loss term can be set as follows: The specific values ​​are adjusted based on the data noise level.

[0056] 1. Data consistency items The constraint network output is matched with the measured data at the location of the sparse sensor group.

[0057] ; in, This represents the number of time steps for the training batch.

[0058] 2. Physical residuals The state variables of the entire domain are constrained to satisfy the dynamic equilibrium equations.

[0059] ; 3. Stiffness smoothing regularization term : The continuity of constraint stiffness coefficients between topologically adjacent elements.

[0060] ; in, To be in the graph structure with the unit A set of adjacent units that share a common node. Adjacent units The stiffness correction factor.

[0061] Sub-step S340: Model optimization and potential damage sub-map extraction; Minimize using gradient-based optimization algorithms Update the network weights until the loss function converges.

[0062] Obtain the converged stiffness correction coefficient vector Traverse all component elements and identify those that satisfy the following conditions. The unit, in which The preset stiffness degradation threshold is used. The identified elements and their connection nodes are defined as potential damage sub-maps. .

[0063] Step 400: Generation of topological variations based on BIM semantics; For the potential damage sub-map identified in step S300 This involves constructing two competing hypothetical models with fundamentally different physical properties, so that subsequent Bayesian confidence-based discriminant analysis can differentiate between material performance degradation and structural cracking. This step specifically includes sub-steps S410 to S430.

[0064] Sub-step S410: Constructing a continuum damage hypothesis model We construct the first hypothetical model, denoted as the continuous medium damage model. In this model, the semantic fusion mechanical dual graph structure is preserved. The topology remains unchanged, and the connections between all nodes are maintained. .

[0065] Define the global stiffness matrix of the model. The stiffness matrix obtained by inversion convergence in step S340 is used directly. As At this point, the potential damage subgraph The mechanical response characteristics within the component are entirely characterized by the reduction of the stiffness parameters of the component elements.

[0066] Sub-step S420: Semantic rule-based crack path optimization; In the potential damage subgraph Internally, crack tendency scores for connection edges are calculated based on stress state and BIM semantic attributes. To determine the target cutting edge.

[0067] First, calculate the potential damage subgraph. The stress tensor of each component element. For the potential damage subplot. Any connecting edge within For the corresponding component element, obtain its local nodal displacement vector at the current moment. The element stress vector is calculated using the geometric and physical equations of the finite element method. : ; in: Component unit The stress vector. Component unit The elastic constitutive matrix. Component unit The strain-displacement matrix. Component unit The local nodal displacement vector.

[0068] based on The maximum principal tensile stress of this element is denoted as . If the calculation result is compressive stress, then let .

[0069] Calculate crack tendency score : ; in, For component units Crack tendency score for the corresponding connecting edge. For component units The maximum principal tensile stress (compressive stress is set to zero). Component unit The semantic vulnerability factor, based on the graph (what does "graph" refer to here? It refers to the dual graph structure of semantic fusion mechanics). Or potential damage subgraph Semantic feature vectors embedded in nodes Identify the component material type. If the material property is identified as a brittle material (such as plain concrete or masonry), set... The first preset value; if the material is identified as ductile (such as steel), set... The second preset value is less than the first preset value. Weighting coefficients for crack tendency calculation.

[0070] Traversal Select all edges The edge with the largest value is used as the target cutting edge. .

[0071] Sub-step S430: Construct a model of discontinuous topological variation assumptions; Construct a second hypothetical model, denoted as the discontinuous fracture model. .

[0072] Performing topological mutation operations: in semantic fusion mechanical dual graph structures Remove target cut edge Generate a mutated dual graph The new edge set For the original edge set Remove The set after.

[0073] Based on mutation dual graph Assemble the global stiffness matrix For potential damage subgraphs All component elements within the unit are reset to their initial design stiffness parameters. The calculation formula is as follows: ; in, This is the global stiffness matrix of the second hypothetical model (discontinuous fracture model). To remove the target cut edge The new edge set afterwards. This is the coordinate transformation matrix.

[0074] Step 500: Adversarial residual verification and Bayesian confidence assessment; After obtaining two competing hypotheses, the continuous medium damage model and the discontinuous fracture model, step S500, based on the Bayesian model selection principle, achieves a physical determination of structural cracks by quantitatively comparing the interpretability of the two models to the observation data and combining the probability cost of BIM priors.

[0075] This step specifically includes sub-steps S510 to S540.

[0076] Sub-step S510: Assume model-driven state field reconstruction; Based on the first hypothesis model Second Hypothesis Model Given the stiffness boundary conditions, solve for the corresponding state fields.

[0077] For the first hypothesis model The stiffness matrix determined in step S410 Set as a constant and input it to the state estimation branch of the physical information neural network. Fix the stiffness correction coefficients in the network, and only iteratively update the weights of the state estimation branch until the data consistency loss function converges, thus obtaining the first reconstructed state field. .

[0078] For the second hypothesis model The stiffness matrix determined in step S430 Set as a constant. In the mutated dual graph Under the given topology, the state variables that minimize data consistency loss are solved using the state estimation branch to obtain the second reconstructed state field. .

[0079] Sub-step S520: Calculate the energy of the fitted residual; Calculate the combined residual energy of the two hypothetical models respectively.

[0080] Calculate the combined residual energy of the first-hypothesis model. The calculation formula is as follows: ; in, This represents the combined residual energy of the first hypothetical model (continuous medium damage model). This represents the reconstructed state field (displacement) under the first assumption model. The physical residual vector under the first assumption model. This is the weighting coefficient of the physical residual in energy calculation. and These are the measured acceleration vector and strain vector, respectively. and The observation matrix constructed in step S220; To be and The physical residual vector is obtained by substituting it into the dynamic equilibrium equation.

[0081] Similarly, using and Calculate the combined residual energy of the second hypothesis model. .

[0082] Sub-step S530: Calculation of semantic topology penalty factor; Calculate the topology penalty factor based on BIM semantics Based on the target cutting edge determined in step S420 The semantic attributes are calculated using the following formula: ; in, is the topological penalty factor based on BIM semantic computation, representing the prior cost of cutting edges. This is the baseline penalty coefficient. This is a reinforcement enhancement factor that adjusts the effect of reinforcement ratio on fracture cost. The normalized reinforcement ratio of the component corresponding to the target cut edge. This is the component type coefficient. For column and main beam components, The value range is [1.2, 1.5]; for secondary beams and infill wall components, The value range is [0.8, 1.0].

[0083] For the first hypothesis model Set its corresponding topology penalty value to zero.

[0084] Sub-step S540: Bayesian confidence judgment and crack detection; Calculate the Bayesian discriminant index : ; in, Bayesian discriminant index.

[0085] based on Execution decision: 1. If The crack is confirmed. The output indicates the existence of a crack, and the second hypothesis model is applied. and its corresponding stiffness matrix Mark the current state of the structure.

[0086] 2. If The crack was determined not to be established. The output result is "no crack," indicating that the current stiffness change is due to material property degradation, and the first hypothesis model is updated. and its corresponding stiffness matrix Mark the current state of the structure.

[0087] Step 600: Crack state output and model closed-loop update; In response to the Bayesian confidence score result from step S540, and to enable visualization of the monitoring results and continuous iteration of the system, step S600 performs crack state quantification calculation and simultaneously updates the semantic fusion mechanical dual graph structure. Based on the fundamental data, a closed-loop monitoring system is formed. This specifically includes sub-steps S610 to S640.

[0088] Sub-step S610: Quantitative calculation of damage parameters; Based on the discrimination result output in step S540, the damage quantification module is called to calculate the physical damage index of the component.

[0089] If the determination result indicates the presence of cracks, calculate the equivalent crack depth of the component. The stiffness correction coefficients after convergence are obtained based on the inversion in step S340. The calculation is performed using the geometric relationship between the moment of inertia and the height of the cross section. For rectangular cross-section members or irregularly shaped members that are transformed into equivalent rectangular cross-sections through geometric parameters, the calculation formula is as follows: ; in, The equivalent crack depth of the component. This is the cross-sectional height of the component (or the equivalent rectangular cross-sectional height). This is the stiffness correction coefficient after network convergence (a general stiffness correction coefficient used to characterize the reduction rate of the moment of inertia or the reduction rate of the elastic modulus, depending on the damage type; in this case, its physical meaning is the ratio of the effective moment of inertia to the initial moment of inertia). Under the crack damage mode, assuming the elastic modulus remains constant, this coefficient represents the effective moment of inertia. With the initial section moment of inertia The ratio; This formula is based on the moment of inertia of a rectangular cross section. With height The cube proportionality relationship ( The structure converts stiffness loss into an equivalent loss of cross-sectional geometric height.

[0090] If the judgment result is material damage, calculate the material damage index. : ; in, The material damage index characterizes the proportion of degradation in the elastic modulus of a material, where... Characterizing the effective elastic modulus under material damage modes With design elastic modulus The ratio of . This is the stiffness correction coefficient after network convergence (a general stiffness correction coefficient, used to characterize the reduction rate of the moment of inertia or the reduction rate of the elastic modulus of the cross section, depending on the type of damage; in this case, its physical meaning is the ratio of the effective elastic modulus to the design elastic modulus).

[0091] Sub-step S620: Topology and attribute update of semantic fusion mechanical dual graph Update the semantic fusion mechanical dual graph structure based on the quantization results. adjacency matrix With node attribute matrix Generate a baseline model for the next monitoring time. .

[0092] If the determination result indicates the existence of a crack, perform a topology update. Locate the target cut-off edge determined in step S420. Adjacency matrix The corresponding row and column elements are set to 0, physically representing a broken connection. Simultaneously, crack feature data, including crack depth, is written into the attribute vector of the broken node. And the timestamp of the occurrence.

[0093] If the determination result is material damage, perform attribute parameter updates. Maintain the adjacency matrix. Without changing the parameters, extract the semantic feature vector of the corresponding node and update the material elastic modulus value therein. : ; in, This is the updated elastic modulus of the material. This is the initial design elastic modulus of the component.

[0094] Sub-step S630: Visual mapping of the BIM model; Establish a semantic fusion mechanical dual graph structure The bidirectional index with the 3D BIM model maps the calculated damage state to the visual attributes of the BIM component.

[0095] Use the globally unique identifier (GlobalId) from the IFC standard as the index key. Configure the component's display properties according to the damage type: 1. Crack Visualization: For associated components marked as cracks, a 3D crack geometry is constructed in BIM space based on the spatial location data of the cut edges. The width of this geometry is set to equal the component cross-sectional width, and the height is equal to... It is then assigned a first preset color attribute (such as red) to identify structural fractures.

[0096] 2. Material Damage Visualization: For associated components marked as having material damage, a damage index is established. A linear mapping relationship with RGB color values. The material color attributes of the component surface are modified to the mapped RGB values, and color gradients are used to characterize the degree of material degradation.

[0097] Sub-step S640: Reset the baseline of the dynamic equations; Based on the updated semantic fusion mechanical dual graph structure The system matrix of the reconstructed structural dynamics equations is used as the physical prior for the next monitoring cycle.

[0098] Based on the updated adjacency matrix The defined topological relationships and the updated node attributes The material parameters stored in the matrix are used to reassemble the global mass matrix using a finite element stiffness ensemble algorithm. and global reference stiffness matrix At that moment In the monitoring process, the system will use Instead of the initial stiffness matrix, a physical inversion is performed for the new damage relative to this baseline.

[0099] Please see the appendix Figure 2 This invention provides a structural crack monitoring system based on physical information and BIM semantics for the above method. At the hardware level, the system includes at least one processor, memory, communication interface and input / output device. At the logical function level, it is divided into a semantic fusion mechanical modeling module, a multi-source sensing observation module, a physical field reverse reconstruction module, a topological variation generation module, a Bayesian adversarial verification module and a closed-loop feedback update module.

[0100] The semantic fusion mechanics modeling module, serving as the system's input processing end, is responsible for establishing the digital foundational graph required for monitoring. This module includes an IFC parsing unit and a graph construction unit. The IFC parsing unit is configured to load Building Information Modeling (BIM) files in Industrial Foundation Class (IFC) format via a file reading interface, and utilize a preset parsing algorithm to extract geometric topological data (including node coordinates and connection relationships) and physical property data (including elastic modulus, moment of inertia, and reinforcement ratio) of building components. The graph construction unit is configured to establish a semantic fusion mechanics dual graph structure based on the extracted geometric topological data. Generate an adjacency matrix A to characterize the physical connections between components, and generate a node feature matrix S to store the physical attributes of the components.

[0101] The multi-source sensing and observation module, serving as the system's physical interface, is responsible for real-time connection to the physical world and data mapping. This module includes a multi-channel data acquisition interface and a coordinate mapping unit. The data acquisition interface connects via wired or wireless network to a sparse sensor array (including accelerometers and strain gauges) distributed across the building structure. It receives analog or digital signals output from the sparse sensor array, performs analog-to-digital conversion and time synchronization, and generates an acceleration vector sequence. and strain vector sequence The coordinate mapping unit stores the transformation matrix between the physical installation coordinates of the sparse sensor group and the BIM model coordinate system, which is used to construct the observation matrix H. This matrix defines the linear mapping relationship between the dual graph node state vectors and the observation data of the sparse sensor group.

[0102] The physics field inverse reconstruction module, as the core computing unit of the system, is responsible for performing deep learning-based physics inversion calculations. This module is deployed on a high-performance computing unit (such as a GPU or TPU) and includes a neural network inference unit, a physical residual calculation unit, and an anomaly filtering unit. The neural network inference unit is loaded with a spatiotemporal graph neural network model based on physical information, taking dual graph data (A, S) and time coordinate t as input, and outputting a global displacement field estimate. and stiffness correction coefficient vector The physical residual calculation unit is configured to construct a loss function based on the structural dynamic equilibrium equations and calculate the physical residuals of the network output values. The anomaly filtering unit is configured to traverse the stiffness correction coefficient vector. Identify component elements with coefficients less than a preset threshold and output a potential damage submap. Index information.

[0103] The topology variation generation module, serving as the system's hypothesis generation end, is responsible for modeling the physical mechanisms of potential damage. This module includes a stress calculation unit and a topology variation execution unit. The stress calculation unit is configured to use the derivatives of finite element shape functions to calculate the potential damage submap based on displacement field estimates. The stress tensor of each component element is calculated, and the principal tensile stress perpendicular to the component axis is extracted. The topology variation execution unit combines the principal tensile stress value with the material property factors in BIM semantics to calculate the crack tendency score and determine the target cut-off edge. This module further generates two sets of data models based on the target cutting edge: a continuous medium damage model that maintains the adjacency matrix. And a discontinuous fracture model that sets the corresponding elements of the target cut edge to zero. .

[0104] The Bayesian adversarial verification module, acting as the system's decision-making endpoint, is responsible for evaluating the credibility of different physical hypotheses. This module includes a residual energy calculation unit and a probabilistic discrimination unit. The residual energy calculation unit is configured to process the two hypothetical models (…) and Substituting the stiffness matrix of the component into the dynamic equation, the residual energy of its comprehensive fit to the observed data is calculated. The probabilistic discriminant element is configured to read the reinforcement ratio and importance coefficient of the component from the BIM data, and calculate the semantic topological penalty factor characterizing the cost of topological failure. The relationship between residual energy difference and penalty factor is compared based on Bayesian discrimination criterion, and the crack judgment result is output.

[0105] Closed-loop feedback update module This module, serving as the system's output execution end, is responsible for result quantification and model iterative updates. It includes a damage quantization unit and a data synchronization unit. The damage quantization unit stores a stiffness geometry mapping algorithm, used to calculate the equivalent crack depth based on a stiffness correction coefficient when a crack is determined to exist, and to calculate the damage index when material damage is determined. The data synchronization unit is configured to update the adjacency matrix or node attribute data of the semantic fusion mechanical dual graph based on the determination results, and to generate an updated baseline stiffness matrix. This data will be used for the next monitoring cycle. Simultaneously, the unit will send the updated damage status data (including crack location and depth) to the visualization terminal, driving the 3D display update of the corresponding components in the BIM model.

[0106] Those skilled in the art will understand that the above modules can be implemented as independent hardware circuits, as software program modules executed by a processor, or a combination of both. The system's data bus is used to transmit matrix data and control instructions between the modules.

[0107] This invention provides a civil engineering structure crack prediction system based on BIM and deep learning, used to implement the above-mentioned method, including electronic equipment, which includes a processor, a memory, a communication interface and a system bus connecting the various components.

[0108] A processor is the computing execution unit of an electronic device. In this embodiment, the processor adopts a heterogeneous computing architecture, including a general-purpose processor (CPU) and a coprocessor (GPU or TPU).

[0109] The general-purpose processor is configured to execute the operating system's scheduling instructions and handle non-matrix logical operations, including: reading building information model data through the IFC parsing algorithm, executing graph theory algorithms to construct a semantic fusion mechanical dual graph G, and performing Bayesian logic discrimination.

[0110] The coprocessor is configured as a parallel computing unit specifically for performing high-dimensional matrix operations and gradient descent optimization of neural networks. Specifically, this includes: executing the forward and backward propagation of the physical information neural network in step S300, and calculating the global displacement field. ; Perform the assembly, inversion, and eigenvalue decomposition operations of the finite element stiffness matrix K in steps S400 and S600.

[0111] Memory and computer-readable storage media: Memory includes volatile memory and non-volatile memory.

[0112] Volatile memory (such as DRAM) serves as a system cache for temporarily storing dynamic data required by the processor during operation, including real-time sparse sensor observation sequences y(t), weight matrices during iteration, and intermediate state variables.

[0113] Non-volatile memory (such as NVMe SSD) serves as a computer-readable storage medium, storing non-transitory computer program instructions and static databases. The databases contain the original files of the Building Information Model (BIM), calibration parameter tables for sparse sensor arrays, and historical monitoring data logs.

[0114] When computer program instructions are loaded and executed by the processor, the processor performs the following steps: Perform semantic fusion mechanics modeling: parse the IFC file to extract component attributes and construct a dual graph containing the adjacency matrix A and the feature matrix S; Perform multi-source sensing mapping: Receive signals from a sparse sensor group and generate an observation matrix H using coordinate mapping relationships; Perform physical field reconstruction: run the physical information neural network and combine the dynamic equations to constrain the inversion of displacement field and stiffness coefficient; Perform topological mutation generation: calculate stress tensor and crack tendency to generate a continuous damage model. With fracture model ; Perform adversarial verification: Calculate and compare the residual energy and semantic topological penalty factor of the two models to determine the crack state; Perform closed-loop update: calculate damage quantification indicators, update the topology or attributes of the dual graph, and generate BIM 3D visualization data.

[0115] The communication interface includes physical network interfaces (such as RJ45 Ethernet ports or fiber optic interfaces) or wireless communication RF modules (such as 5G or WiFi 6 modules). The communication interface is configured to establish data channels between the electronic device and external sparse sensor network and remote clients. In this example, the communication interface receives raw acceleration and strain data packets from sparse sensor network nodes and writes them directly to a direct memory access (DMA) buffer in volatile memory.

[0116] In one distributed deployment implementation, the electronic device is implemented as a system comprising edge computing nodes and a cloud computing center. The edge computing nodes are deployed on-site at the building site and are responsible for data acquisition, filtering, and compression. The cloud computing center receives the preprocessed data via a broadband network, utilizes its high-performance GPU cluster to perform the physical inversion and Bayesian calculations described in steps S300 to S500, and returns the calculation results (crack location, depth) to the user terminal via a web service interface.

[0117] Those skilled in the art will understand that electronic devices may also include other well-known hardware structures such as power modules and heat dissipation components, which will not be described in detail here.

Claims

1. A method for predicting cracks in civil engineering structures based on BIM and deep learning, characterized in that, Includes the following steps: Step S100: Construct a semantic fusion mechanical dual graph based on BIM: Obtain the building information model data of the target monitoring object, parse the geometric information, material properties and component connection relationships in the building information model data, and transform the building information model data into a semantic fusion mechanical dual graph structure; Step S200, Sparse sensor group data mapping and dynamic equation initialization: Based on the semantic fusion mechanical dual graph structure, the initial physical matrix is ​​assembled using BIM design parameters to construct the structural dynamic equilibrium equation, and the time series data collected by the sparse sensor group is mapped to the corresponding physical nodes in the semantic fusion mechanical dual graph structure. Step S300: Inverse reconstruction of stiffness field based on physical constraints: Construct a physical information neural network, take the semantic fusion mechanical dual graph structure and the mapped time series data as input, and based on the physical residual constraints of the structural dynamic equilibrium equation, invert the estimated value of the global stiffness matrix, and identify potential damage subgraphs based on the estimated value of the global stiffness matrix. Step S400: Topological variation generation based on BIM semantics: For the potential damaged subgraph, generate a first hypothetical model that keeps the graph topology unchanged, and identify and remove the connecting edges in the potential damaged subgraph to generate a mutated dual graph. Step S500, Adversarial Residual Verification and Bayesian Confidence Determination: Calculate the residual energy of the first hypothesis model and the second hypothesis model for interpreting the time series data, respectively. Combine the semantic topological penalty factor calculated based on the semantic attributes of the connection edge, and determine whether the crack is established by comparing the Bayesian discriminant index. Step S600, Crack Status Output and Model Closed-Loop Update: When step S500 determines that a crack has been established, the equivalent crack depth is calculated based on the estimated value of the global stiffness matrix, the position of the connecting edge and the equivalent crack depth are output, and the mutated dual graph is updated to the benchmark model for the next monitoring time.

2. The method for predicting cracks in civil engineering structures based on BIM and deep learning according to claim 1, characterized in that, The semantic fusion mechanical dual graph consists of a node set and an edge set; the construction of the BIM-based semantic fusion mechanical dual graph in step S100 specifically includes: Perform node discretization to construct the node set: retrieve component entities containing sparse sensor group coordinates in the building information model data, and define the geometric center or sensor attachment point of the component entity as the physical node; identify beam or column components not covered by the sparse sensor group in the building information model data, discretize them along the component axis at a preset step size, and generate virtual nodes distributed inside the component. Establish force transmission connections to construct the edge set: For components with topological connections in the building information model data, establish inter-component connection edges between boundary nodes belonging to different components; for adjacent nodes generated by discretization within the same component, establish intra-component connection edges to characterize the stress continuity within the component. Assemble the semantic fusion mechanical dual graph structure: calculate the initial weight of the connecting edge based on the equivalent elastic modulus determined by the arithmetic mean of the elastic modulus of the design elastic moduli of the two end components, the equivalent transfer area, and the Euclidean distance between the nodes, and generate a weighted adjacency matrix based on the node set and the edge set, and use the weighted adjacency matrix to establish the topological structure of the semantic fusion mechanical dual graph.

3. The method for predicting cracks in civil engineering structures based on BIM and deep learning according to claim 1, characterized in that, Step S100 further includes extracting non-geometric semantic attributes of the component entity based on the material properties and the geometric information to generate a semantic feature vector, and embedding the semantic feature vector into the graph node corresponding to the semantic fusion mechanical dual graph structure, wherein the semantic feature vector includes the unique thermal encoding vector of the component type, the normalized design elastic modulus value, the normalized reinforcement ratio value, and the normalized section moment of inertia value. The construction of the structural dynamic equilibrium equation in step S200 specifically includes: calculating the initial mass matrix using the discrete volume element volume in the geometric information and the material density in the material properties; and calculating the initial damping matrix based on the Rayleigh damping model using the initial mass matrix and the initial global stiffness matrix generated based on the BIM design parameters.

4. The method for predicting cracks in civil engineering structures based on BIM and deep learning according to claim 1, characterized in that, In step S200, mapping the time-series data acquired by the sparse sensor group to the corresponding physical nodes in the semantic fusion mechanical dual graph structure specifically includes: Construct an acceleration selection matrix, and use the acceleration selection matrix to establish a linear mapping relationship between the output sequence of the accelerometers in the sparse sensor group and the global node acceleration vector in the structural dynamics equilibrium equation; A global strain observation matrix containing the derivatives of finite element shape functions is constructed, and the global strain observation matrix is ​​used to establish a linear mapping relationship between the strain gauge output sequence in the sparse sensor group and the global nodal displacement vector in the structural dynamic equilibrium equation.

5. The method for predicting cracks in civil engineering structures based on BIM and deep learning according to claim 1, characterized in that, The physical information neural network constructed in step S300 includes a parameter estimation branch and a state estimation branch; The parameter estimation branch takes the adjacency matrix and node semantic feature matrix of the semantic fusion mechanical dual graph structure as input, outputs a stiffness correction coefficient vector that remains constant within the current monitoring time window, and uses the finite element stiffness integration principle to assemble the estimated value of the global stiffness matrix based on the stiffness correction coefficient vector. The state estimation branch takes the time coordinate and the graph embedding features generated by the semantic fusion mechanical dual graph as input, outputs the displacement estimate of the global nodes, and uses an automatic differentiation algorithm to differentiate the displacement estimate with respect to time to calculate the velocity estimate and acceleration estimate.

6. The method for predicting cracks in civil engineering structures based on BIM and deep learning according to claim 5, characterized in that, In step S300, the stiffness correction coefficient vector is optimized by minimizing the total loss function, which includes: A data consistency term is used to constrain the response of the displacement estimate and acceleration estimate output by the state estimation branch at the location of the sparse sensor group to match the time series data; The physical residual term is used to constrain the estimated values ​​of the global stiffness matrix, the estimated values ​​of displacement, the estimated values ​​of velocity, and the estimated values ​​of acceleration to satisfy the structural dynamic equilibrium equations. A stiffness smoothing regularization term is used to constrain the continuity of the stiffness correction coefficient vector between topologically adjacent elements in the semantic fusion mechanical dual graph structure.

7. The method for predicting cracks in civil engineering structures based on BIM and deep learning according to claim 1, characterized in that, The specific steps of identifying the connecting edge in step S400 include: The stress tensor of each component element in the potential damage subplot is calculated using the derivative of the finite element shape function, and the maximum principal tensile stress is extracted from the stress tensor. The semantic vulnerability factor is determined based on the semantic feature vectors of the nodes in the potential damage subgraph; The crack tendency score is calculated by weighted summation of the maximum principal tensile stress and the semantic vulnerability factor, and the edge with the highest crack tendency score is selected as the connecting edge.

8. The method for predicting cracks in civil engineering structures based on BIM and deep learning according to claim 1, characterized in that, The adversarial residual verification and Bayesian confidence judgment in step S500 specifically include: Calculate the combined residual energy of the first hypothetical model and the second hypothetical model respectively. The combined residual energy is the weighted sum of the data consistency residual and the physical balance residual. The semantic topology penalty factor is calculated based on the normalized reinforcement ratio and component type coefficient of the component corresponding to the connection edge. The semantic topology penalty factor represents the prior cost of removing the connection edge. Calculate the Bayesian discriminant index, which is the difference between the comprehensive residual energy of the first hypothetical model and the comprehensive residual energy of the second hypothetical model, and then the semantic topology penalty factor. When the Bayesian discriminant index is greater than zero, the crack is determined to be established.

9. The method for predicting cracks in civil engineering structures based on BIM and deep learning according to claim 5, characterized in that, The calculation of the equivalent crack depth based on the estimated value of the global stiffness matrix in step S600 specifically includes: Obtain the converged stiffness correction coefficient vector from the parameter estimation branch, and determine the target stiffness correction coefficient corresponding to the connecting edge; Based on the cubic proportional relationship between the moment of inertia of a cross section and the geometric height of the cross section, the equivalent crack depth is calculated using the target stiffness correction coefficient and the cross section height of the component. The target stiffness correction coefficient represents the ratio of the effective moment of inertia of the cross section to the initial moment of inertia of the cross section under the crack damage mode.

10. The method for predicting cracks in civil engineering structures based on BIM and deep learning according to claim 1, characterized in that, The step S600 of updating the variant dual graph to the baseline model for the next monitoring time specifically includes: When a crack is determined to be established, the weight corresponding to the connecting edge is reset to zero in the weighted adjacency matrix of the semantic fusion mechanical dual graph, and the crack depth is recorded in the attribute of the disconnected node. When it is determined that the crack is not established but there is material damage, the design elastic modulus value in the semantic feature vector of the corresponding node is updated according to the estimated value of the global stiffness matrix. Based on the updated semantic fusion mechanical dual graph, the reference global stiffness matrix at the next monitoring time is reassembled to reset the structural dynamic equilibrium equations.

Images