A structural stability assessment and early warning method for multi-source data fusion

Abstract

This invention discloses a method for structural stability assessment and early warning based on multi-source data fusion, relating to the field of building structural stability assessment technology. The method includes the following steps: obtaining the building structure type, determining design parameters, and constructing a digital building model; simulating typical working conditions on the digital model, acquiring simulation data, and fusing the data using a Bayesian network algorithm to form a data matrix; extracting core factors using principal component analysis, including real-time and historical data, and constructing a training dataset based on historical data; constructing a stability assessment model using factor analysis, training the model with the training dataset, inputting real-time data, and outputting stability assessment results; comparing the assessment results with thresholds to obtain graded early warning results. Based on the graded early warning results, multi-source data fusion-based structural stability assessment and early warning are achieved.

Description

Technical Field

[0001] This invention relates to the field of structural stability assessment and early warning, specifically to a method for structural stability assessment and early warning based on multi-source data fusion. Background Technology

[0002] As the core load-bearing framework of engineering facilities, the stability of building structures is a crucial prerequisite for ensuring the safety of people's lives, property, and the public interest. Establishing a scientific and reliable building structure stability assessment system is a core element in extending structural service life, reducing operation and maintenance costs, and preventing safety risks, and it has irreplaceable significance for engineering safety management.

[0003] Existing methods for assessing the stability of building structures primarily employ a single-source data acquisition approach combined with simplified model calculations. The data acquisition phase typically utilizes a single-parameter monitoring mode, collecting vibration data from the building structure via accelerometers. The assessment model construction phase employs a single algorithm model, namely a basic linear regression model that linearly correlates historical vibration data with the building structure's stability, and then predicts real-time structural stability to complete the stability assessment. However, existing methods for assessing the stability of building structures, relying solely on single-source data, cannot reflect the true state of the structure under the coupling influence of multiple factors, making it difficult to meet the precise stability assessment requirements of engineering projects. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a structural stability assessment and early warning method based on multi-source data fusion. This method solves the problem that single-source data alone cannot reflect the true state of a structure under the coupling influence of multiple factors, making it difficult to meet the precise stability assessment requirements of engineering projects.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for structural stability assessment and early warning of multi-source data fusion, comprising the following steps: Step S1: Obtain the building structure type, obtain the building structure design parameters based on the building structure type, and construct a digital building model based on the building structure design parameters; Step S2: Simulate typical working conditions based on the digital building model to obtain building structure simulation data. Use a Bayesian network algorithm to perform heterogeneous data fusion on the building structure simulation data to obtain a building structure data matrix. Step S3: Use principal component analysis to extract the core factors of the building structure data matrix to obtain the core factors of the building structure. The core factors of the building structure include real-time data and historical data of the core factors of the building structure. Construct a training dataset based on the historical data of the core factors of the building structure. Step S4: Based on the Stacking ensemble learning framework, construct a building structure stability assessment model. Train the building structure stability assessment model using the training dataset of the building structure core factors to obtain a trained building structure stability assessment model. Input the real-time data of the building structure core factors into the trained building structure stability assessment model and output the building structure stability assessment result. Step S5: Compare the building structure stability assessment results with preset thresholds to obtain graded early warning results.

[0006] Preferably, obtaining the building structure type and obtaining building structure design parameters based on the building structure type includes: First, identify and obtain the building structure type of the building to be evaluated. This information is used to guide the focus of data collection. For frame structures, the focus should be on the cross-sectional properties of beams and columns and the node connection parameters, while for shear wall structures, the focus should be on the wall thickness, reinforcement and boundary constraints. Subsequently, based on the structural type, four core building structural design parameters were collected, specifically including geometric dimensional parameters G, such as member length, section height and width, derived from design drawings or BIM models; and material property parameters. Parameters such as elastic modulus, Poisson's ratio, and yield strength are derived from material test reports or design specifications; characteristic parameters of connection nodes. Parameters such as nodal stiffness, damping, and ultimate bearing capacity can be obtained through nodal testing or numerical fitting; boundary condition parameters... For example, the type of support, the degree of freedom of constraint, and the foundation stiffness are determined based on the geological survey and design documents.

[0007] Preferably, constructing a digital building model based on the aforementioned building structural design parameters includes: A digital building model is constructed based on the building structural design parameters using parametric modeling methods. Its mathematical expression is:

[0008] in, It is a digital building model; function This refers to the process of converting design parameters into a physical mapping of a computable model through the interface between CAD and finite element analysis software. The specific rules are as follows: geometric dimension parameters G are converted into the geometric topology of the model; material property parameters... Assign material constitutive relations to the model; assign property parameters to the connection nodes. Define the connection method between components; define the boundary condition parameters Set as support constraints for the model; Digital building model It is a digital benchmark model that is physically equivalent to the actual structure, possessing two main characteristics: its geometry and topology are consistent with the actual structure; and its mechanical parameters conform to design specifications, allowing it to be used for subsequent S2 stage working condition simulations; design parameters yes The input basis, It is the physical representation of the parameters. If any parameter is missing, This will fail to accurately reflect the structural dynamics; parameter accuracy directly determines... Reliability.

[0009] Preferably, based on the building structure type, the digital building model is simulated for typical working conditions to obtain building structure simulation data, including: Building structure data simulation requires a detailed digital building model. The simulation covers four dimensions: physical response, environmental load, morphological changes, and acoustic emission characteristics. The specific simulation content and technical standards are as follows: Structural dynamic response simulation in a highly detailed digital building model Virtual monitoring points are set up in key parts. The vibration response of the structure under typical working conditions is simulated by finite element analysis. The vibration acceleration time history data is simulated and output. The sampling frequency is ≥100Hz. The natural frequency and damping ratio of the structure in the range of 0.5-10Hz are extracted by Fourier transform. The layout of virtual monitoring points must meet the requirements of spatial modal analysis. For example, at least 2 bidirectional monitoring points are set up for every 10 floors of a high-rise building model. Environmental load simulation in a detailed digital building model A temperature field is applied to the interior and exterior surfaces of the building to simulate the temperature gradient effect; wind speed and direction time series data are input into the model boundary conditions, with particular attention paid to the coupling relationship between vortex-induced vibration frequency and the natural frequency of the structure; the ground motion time history is used as input to simulate the peak ground acceleration, and the analysis threshold is set to 0.005g. Morphological change simulation, based on continuum mechanics, monitors the overall deformation of the structure and the initiation and development of macroscopic cracks by calculating the displacement and strain fields of the model; its output is field variables, such as the maximum principal strain of key parts and inter-story drift angle, which are used to evaluate the stiffness degradation and geometric stability of the structure. Acoustic emission characteristic simulation, based on the material damage constitutive model and acoustic principles, simulates the stress wave signals released during microscopic damage. Its output is a sequence of data characterizing acoustic emission events, such as event count rate, energy release, and amplitude distribution. These parameters are active indicators of ongoing damage and can provide early warning of microscopic damage before macroscopic cracks appear.

[0010] Preferably, the building structure simulation data is fused using a Bayesian network algorithm to obtain a building structure data matrix including: Static probabilistic fusion is performed based on Bayesian networks, which are used to characterize the inherent, physical-mechanism-based static probabilistic dependencies between the aforementioned features. Mutual information I(X;Y) is used as a statistical measure to calculate the dependency strength between all feature pairs, not just individual variables. The formula for calculating mutual information is:

[0011] Here, X and Y represent any two characteristic variables. The larger the value, the closer the relationship between the two variables. The optimal network topology is determined by a search algorithm to make it conform to engineering physical constraints such as "temperature change affects strain, and then affects frequency". After determining the network structure, the conditional probability distribution of each node is learned. The core advantage of Bayesian networks lies in decomposing the high-dimensional joint probability distribution into a product of a series of conditional probabilities:

[0012] in, It is a high-dimensional joint probability distribution, representing all variables. arrive The probability of taking a set of preset values ​​at the same time serves to decompose the complex high-dimensional joint probability calculation into a product of a series of simple local conditional probabilities. It is the i-th feature variable; It represents The set of parent nodes is all nodes that directly point to... The nodes represent the variables; by injecting observation data, the network performs probabilistic inference and calculates the posterior distribution of all variables; The output of probabilistic inference is extracted to form a preliminary fusion feature matrix. This includes continuous features based on posterior expectation, classification features based on probability confidence, and statistical features based on uncertainty measures. Each feature column of the matrix incorporates probabilistic representations of multi-source information. Then, dynamic feature enhancement is performed based on the attention mechanism. Since Bayesian networks represent static dependencies, while actual data interactions are dynamic, an attention mechanism is introduced for enhancement. This mechanism dynamically calculates the mutual influence weights of different feature modalities at a specific time. The output of the attention mechanism is the enhanced feature matrix after dynamic weighting and residual connections. It effectively separates effects from different sources, and the specific process is as follows: The dynamic intermodal attention weights are calculated to adaptively capture complex spatiotemporal dependencies between data, as shown in the following formula:

[0013] in, It is the feature vector of the target mode; is the feature vector of the source mode; d is the feature dimension, which is the optimal dimension determined on the validation set through cross-validation, and is set to 64. These are attention weights; multimodal offset generation is achieved by generating an offset tensor through convolution operations, as shown in the following formula:

[0014] in, It is the output offset tensor, which is used to identify "interference" or "offset" patterns that are mixed in with the target modal features but are caused by other source modes; It is a one-dimensional convolution operation; The kernel size is 3 to capture local temporal patterns; the number of kernels is 32, the stride is 1, and the output offset is concatenated with the target modal feature residual, as shown in the following formula:

[0015] in, These are enhanced features, whose function is to separate the estimated "interference" effect from the original features, thereby enhancing the main features of the target mode. It is the offset tensor; It is the preliminary fusion feature matrix obtained after the initial fusion of Bayesian networks; Finally, the enhanced features are subjected to nonlinear transformation and dimensionality reduction through a fully connected layer to generate the final output of step S2, the building structure data matrix M, which is expressed as follows:

[0016] Where W is the weight matrix, obtained through backpropagation optimization training; b is the bias term, obtained through backpropagation optimization training. ( ) is the ReLU activation function; this matrix M∈ It is a high-level feature representation that has been deeply integrated and refined. It contains complementary and consistent information from multiple sources and will be used as the direct input for step S3.

[0017] Preferably, principal component analysis is used to extract core structural factors from the building structure data matrix, resulting in the following core structural factors: The standardized data is projected onto the principal component space to obtain the core factor matrix of the building structure, as shown in the following formula:

[0018] in, It is the core factor matrix of building structure. ∈ Each column is a sequence of principal component scores. These principal components are linear combinations of the original multi-source simulated variables, which can retain the information that contributes most to the variance of the original data with the lowest possible dimensionality.

[0019] Preferably, the core building structure factors include real-time data and historical data of the core building structure factors. A training dataset is constructed based on the historical data of the core building structure factors, and the feature differences obtained include: To construct the training dataset, a digital twin model that is consistent with the initial state of the real structure must first be established. This model utilizes a highly detailed digital architectural model. The model is obtained by revising the initial baseline monitoring data, thereby effectively incorporating real-world factors such as construction errors and material property deviations, making its simulation data highly reliable. Subsequently, based on this digital twin model Simulations are performed to generate historical data on core structural factors. This historical data undergoes standardization preprocessing, consistent with real-time data, before being used to construct training samples to eliminate dimensional influences and ensure uniform feature scales. To effectively capture the dynamic evolution of structural states, a sliding time window method is used to construct training samples, avoiding the simplistic treatment of time points as independent samples. With a window length of L, a training sample can be represented as... Its corresponding This label represents the structural state assessment result at the end of the window. It is obtained through structural mechanics calculations or performance indicators defined by specifications. training set ={ , } Used for model training in step S4; In practical applications, real-time monitoring modules deployed on the building structure continuously collect the latest time-series data. This real-time monitoring data undergoes standardized preprocessing that is completely consistent with historical data. Then, a pre-trained principal component analysis projection matrix is ​​used to project the real-time data onto the same feature space as the historical core factors, generating a core factor sequence representing the current state. This latest sequence is then organized into real-time data of the building structure's core factors according to a predetermined window length L. .

[0020] Preferably, a building structure stability assessment model is constructed based on the Stacking ensemble learning framework. The model is trained using the core building structure factor training dataset to obtain the trained building structure stability assessment model, which includes: The core of the model building is the Stacking ensemble framework, which consists of two layers: the first layer comprises multiple heterogeneous base models, and the second layer is a meta-model. The base models include CatBoost, XGBoost, GBDT, and ExtraTrees, each independently learning the nonlinear relationship between core structural factors and stability. The CatBoost model excels at handling categorical features and avoiding prediction bias, employing OrderedBoosting. The XGBoost model is based on gradient boosting decision trees and prevents overfitting through regularization. The GBDT model iteratively builds weak learners. The ExtraTrees model increases diversity through random feature selection. The meta-model uses LogisticRegression to fuse the output probabilities of the base models. The mathematical expression for Stacking can be simplified to the following formula:

[0021] Where F represents the core factor vector of the building structure; These are the prediction functions of each base model, outputting the instability probability. LR is the mapping function of the meta-model, which generates the building structure instability probability through a linear combination of these probabilities. ; The training process uses a training dataset, first addressing the data imbalance problem, and then performing hyperparameter optimization and model training.

[0022] Preferably, the real-time data of the core building structure factors are input into the trained building structure stability assessment model, and the output building structure stability assessment results include: Input processing: Real-time data is directly input into the pre-trained building structure stability assessment model; Prediction and Decision Making: The building structure stability assessment model outputs the probability of building structure instability. ∈[0,1], the closer the value is to 1, the higher the risk of instability.

[0023] Preferably, comparing the building structure stability assessment results with preset thresholds to obtain graded early warning results includes: This step is based on the structural instability probability obtained in step S4. The results are compared with preset thresholds to generate tiered early warning results; The decision rule is formalized as follows:

[0024] in, The threshold is the probability of structural instability output by the ensemble model, obtained by stacking the results of the fused base models. It is a dynamically adjustable parameter. Setting it low allows for the capture of potential risks while reducing underreporting. Setting it at a high level ensures the reliability of high-risk warnings; For different warning levels, corresponding tiered response plans are automatically activated: For the Stable level, routine automated monitoring and periodic report generation are maintained without special intervention; for the Low Risk level, a detailed diagnostic report is automatically generated, and engineers are prompted to verify the validity of the data, initiate preliminary analysis procedures, and arrange for on-site manual inspections in the near future; for the Medium Risk level, the depth and frequency of data analysis are automatically increased, while an expert team is notified for remote consultation and assessment, and contingency plans are developed; for the High Risk level, the highest level alarm is immediately triggered, relevant emergency management departments are automatically notified, emergency plans are activated, and an expert team is dispatched to the site for final assessment and disaster relief decision-making.

[0025] This invention provides a method for structural stability assessment and early warning of multi-source data fusion, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) The structural stability assessment and early warning method of multi-source data fusion constructs the probabilistic dependency relationship of building structure simulation data through Bayesian network algorithm, quantifies the joint probability distribution of building structure simulation data of different modalities, effectively integrates complementary information, improves data consistency compared with simple fusion, and can handle data missing or conflict through probabilistic reasoning, significantly improving the consistency and reliability of data input, laying a high-quality data foundation for subsequent building structure stability assessment models.

[0026] (2) This multi-source data fusion structural stability assessment and early warning method innovatively introduces an attention mechanism on the basis of static Bayesian fusion, realizing adaptive calculation of dynamic interaction weights between different data modalities. This mechanism can accurately identify and separate interference signals caused by environmental changes and other factors, and generate enhanced building structural features, thereby ensuring that the core features can more realistically reflect the mechanical state of the structure itself and improving the accuracy of the state assessment.

[0027] (3) A structural stability assessment and early warning method based on multi-source data fusion: This method constructs a stability assessment model using the Stacking ensemble learning framework. By integrating the predictive advantages of multiple heterogeneous base models such as CatBoost, XGBoost, GBDT, and ExtraTrees, and using the LogisticRegression meta-model for decision fusion, it can learn complex nonlinear coupling relationships from multiple core factors such as vibration, temperature, and load. Compared with traditional single models, this ensemble strategy significantly improves the assessment accuracy. Due to its strong generalization ability, it can be adapted to different building types such as frames, shear walls, and steel structures, effectively solving the problems of poor adaptability and insufficient accuracy of traditional models.

[0028] (4) A multi-source data fusion method for structural stability assessment and early warning. By constructing a multimodal feature template library based on the identification results of real visible cracks, and using this as a benchmark for counterfactual comparison analysis, the difference between the features of unknown areas and the features of typical cracks can be accurately quantified. This cross-modal difference analysis logic based on real defects can effectively filter out irrelevant interference signals and directly anchor highly suspicious areas of hidden cracks, thereby achieving early, accurate, and efficient identification of potential structural damage. Attached Figure Description

[0029] Figure 1 This is a flowchart of a structural stability assessment and early warning method for multi-source data fusion proposed in this invention.

[0030] Figure 2 The hierarchical diagram of the data matrix is ​​obtained for the structural stability assessment and early warning method of multi-source data fusion proposed in this invention.

[0031] Figure 3 This is a hierarchical diagram of graded early warning obtained in the structural stability assessment and early warning method of multi-source data fusion proposed in this invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figure 1-3 This invention provides a technical solution: a method for assessing and warning the structural stability of multi-source data fusion. Specifically, the method for assessing and warning the structural stability of multi-source data fusion is provided below. Please refer to [link / reference]. Figure 1 The method includes the following steps: Step S1: Obtain the building structure type, obtain the building structure design parameters based on the building structure type, and construct a digital building model based on the building structure design parameters.

[0034] This step aims to build a high-precision digital benchmark model that can interface with multi-source monitoring data, which is a prerequisite for subsequent data fusion and intelligent assessment.

[0035] First, identify and obtain the building structure type of the building to be evaluated (such as frame structure, shear wall structure, steel structure, masonry structure, or hybrid structure). This information will guide the focus of data collection. For example, for frame structures, the focus should be on the cross-sectional properties of beams and columns and the parameters of node connections, while for shear wall structures, the focus should be on the wall thickness, reinforcement, and boundary constraints.

[0036] Subsequently, based on the structural type, four core building structural design parameters were collected, specifically including geometric dimensional parameters G, such as component length, section height and width, which are derived from design drawings or BIM models; and material property parameters. Parameters such as elastic modulus, Poisson's ratio, and yield strength are derived from material test reports or design specifications; characteristic parameters of connection nodes. Parameters such as nodal stiffness, damping, and ultimate bearing capacity can be obtained through nodal testing or numerical fitting; boundary condition parameters... Specific parameters, such as support type, constrained degrees of freedom, and foundation stiffness, are determined based on geological survey and design documents.

[0037] Based on the above building structural design parameters, a digital building model is constructed using parametric modeling methods. Its mathematical expression is:

[0038] in, It is a digital building model; function This refers to the process of converting design parameters into a physical mapping of a computable model through the interface between CAD and finite element analysis software. The specific rules are as follows: geometric dimension parameters G are converted into the geometric topology of the model; material property parameters... Assign material constitutive relations to the model; assign property parameters to the connection nodes. Define the connection method between components; define the boundary condition parameters Set as the support constraint for the model.

[0039] Digital building model It is a digital benchmark model that is physically equivalent to the actual structure, possessing two main characteristics: its geometric topology is consistent with the actual structure (e.g., beam and column positions and node construction match the construction drawings); and its mechanical parameters conform to design specifications (e.g., material elastic modulus and stiffness coefficient are assigned design values), making it suitable for subsequent S2 stage load case simulation. Design parameters yes The input basis, It is the physical representation of the parameters; if any parameter is missing, This will fail to accurately reflect the structural dynamics; parameter accuracy directly determines... Reliability.

[0040] To improve the consistency between the digital building model and the actual structure, a sensitivity-based model update algorithm is introduced to correct the digital building model and generate a refined digital building model. The core of the update function U is an iterative optimization process aimed at minimizing the error between the simulated response and the design target response. Its mathematical expression is clearly defined as follows:

[0041] in, It is a model In parameters Simulated response under the following conditions; It is the design target response; C is the parameter vector to be corrected, obtained through sensitivity analysis; C is the parameter... The set of physical constraints; It is a preset error threshold. =5%. The specific implementation of this process uses a two-step method: "sensitivity screening + gradient descent optimization". Sensitivity screening, using a formula Calculation parameters Response to the i-th target Sensitivity, to filter out | Significant parameters ≥ 0.6 constitute the parameter vector θ to be corrected; iterative optimization is performed by using gradient descent to iteratively adjust the parameter vector θ until the convergence condition is met. At this point, the iteration stops and the final detailed digital building model is output. .

[0042] This step, as the foundational link in the entire multi-source data fusion structural stability assessment and early warning process, plays a crucial role in systematically acquiring building structure types, collecting building structure design parameters, and constructing a refined digital building model. This provides an accurate and complete digital representation of the structure and a foundation for data fusion for subsequent steps. This step not only ensures that subsequent heterogeneous data (such as real-time monitoring data, historical data, and environmental data) can accurately interface with the model, but also improves the mechanical consistency between the digital model and the actual structure through sensitivity-based model updates. It enhances the model's compatibility and analytical capabilities with subsequent multi-source simulation data, directly supporting data fusion, feature extraction, model training, and early warning decisions in stages S2 to S5. It is the fundamental guarantee for the reliability, scientific rigor, and engineering practicality of the entire assessment system. By constructing a high-fidelity digital model, this step provides an accurate spatiotemporal reference and physical background for the multi-source data in step S2, ensuring the consistency of the mechanical behavior of the digital model with the actual structure. This provides a reliable physical basis for the working condition simulation in stage S2, thereby supporting the accurate alignment of subsequent simulation data with real data, and is the fundamental guarantee for the reliability of the entire assessment system.

[0043] Step S2: Simulate typical working conditions based on the digital building model to obtain building structure simulation data. Use a Bayesian network algorithm to perform heterogeneous data fusion on the building structure simulation data to obtain a building structure data matrix.

[0044] This step aims to obtain a detailed digital building model through finite element simulation. The response under various external influences provides a source for data fusion. Its core logic is to apply typical working conditions (input), and through finite element simulation (processing), obtain the physical response of the structure in four dimensions (output).

[0045] The input for multi-dimensional physics simulation is a typical working condition, referring to preset external loads and environmental conditions, such as seismic wave input, wind load, and temperature field changes. These are the driving factors for the simulation; they are processed through finite element simulation. The refined model obtained in step S1 is then used. We applied the above operating conditions to perform numerical simulations; the output physical responses in four dimensions were then obtained. After the simulation calculations, we obtained the structure's response data in the following four dimensions, which together constitute a comprehensive picture of the structure's health status: The four dimensions include structural dynamic response, such as the structure's vibration frequency, damping ratio, and acceleration time history; environmental loads, such as wind pressure distribution on the structural surface and internal temperature field; morphological changes, such as strain distribution, displacement, and deflection in key parts; and acoustic emission characteristics, such as the event count rate and energy release simulating material damage.

[0046] Through this explicit input-output relationship, we obtained a comprehensive multi-source heterogeneous simulation dataset.

[0047] The simulation strategy is not generalized but strictly customized based on the specific building structure type identified in step S1 to ensure that the simulated physical processes are consistent with the main mechanical behaviors of the real structure. For frame structures, the simulation will focus on the nonlinear response of key force transmission components such as beam-column joints under horizontal loads, monitoring shear strain and steel stress in the joint area, and using inter-story drift angle as the core evaluation index. In contrast, for shear wall structures, the simulation focuses on the overall lateral resistance performance of the wall, requiring detailed analysis of bending and shear deformation at the bottom of the wall limbs and the energy dissipation mechanism of the coupling beams, with key indicators being the top displacement and overall deformation of the structure. For large-span steel structures, the core of the simulation will shift to wind-induced vibration effects and temperature stress analysis. Monitoring points should be concentrated at support nodes, member joints, and the point of maximum mid-span deflection, focusing on obtaining the wind pressure distribution of the roof, the stress ratio of key members, and the overall vibration modes of the structure. Through this differentiated simulation design, the subsequently extracted simulation data can be highly focused on the key stability characteristics of various structures. The specific simulation content and technical standards are as follows: Structural dynamic response simulation in a highly detailed digital building model Virtual monitoring points are set at key components (beam-slab joints, shear walls, etc.) to simulate the vibration response of the structure under typical working conditions (such as seismic wave input and wind load) through finite element analysis. The simulation outputs vibration acceleration time history data (unit: m / s²). 2 The sampling frequency is ≥100Hz, and the natural frequency and damping ratio of the structure in the range of 0.5-10Hz are extracted by Fourier transform. The layout of virtual measuring points must meet the requirements of spatial modal analysis; for example, at least two bidirectional monitoring points should be set for every 10 floors in a high-rise building model. Environmental load simulation in a detailed digital building model A temperature field is applied to the interior and exterior surfaces of the building to simulate the temperature gradient effect (such as the temperature difference between the sunlit and shaded sides); wind speed and direction time series data (range 0-60m / s) are input into the model boundary conditions, with particular attention paid to the coupling relationship between vortex-induced vibration frequency and the structure's natural frequency; the ground motion time history is used as input to simulate peak ground acceleration (PGA), with the analysis threshold set to 0.005g; Morphological change simulation, based on continuum mechanics, monitors the overall deformation of a structure and the initiation and development of macroscopic cracks by calculating the displacement and strain fields of the model. Its outputs are field variables, such as the maximum principal strain at key locations (used to identify macroscopic cracks ≥0.02mm in width) and inter-story drift angles, used to assess the stiffness degradation and geometric stability of the structure. Acoustic emission characteristic simulation, based on the material damage constitutive model and acoustic principles, simulates the stress wave signals released during micro-damage (microcrack propagation). Its output is a sequence of data characterizing acoustic emission events, such as event count rate, energy release, and amplitude distribution. These parameters are active indicators of ongoing damage, providing early warning of micro-damage before macro-cracks appear.

[0048] In the simulation, the active micro-damage activity revealed by acoustic emission characteristics is the direct cause of the appearance and propagation of macroscopic cracks in subsequent morphological changes. The combination of these two factors enables monitoring of the entire damage evolution process, from microscopic processes to macroscopic results.

[0049] The aforementioned four-dimensional simulations generated multiple sets of time-series data with varying physical meanings and dimensions, namely, building structure simulation data. To transform these heterogeneous data into a unified input suitable for machine learning models, a fusion strategy capable of understanding their inherent physical relationships is required. This method employs a hierarchical fusion framework, realizing the transformation from statistical features extracted from the original data to high-quality fused features.

[0050] First, statistical features with clear physical meaning are extracted from the original simulation sequences of various dimensions. For example, features such as natural frequency and damping ratio are extracted from the structural dynamic response; features such as maximum principal strain and average displacement are extracted from morphological changes; features such as temperature gradient and average wind pressure are extracted from environmental loads; and features such as event count rate and energy release rate are extracted from acoustic emission characteristics.

[0051] All extracted features constitute the initial feature set, which is then standardized using Z-score to eliminate the influence of dimensions, laying the foundation for subsequent fusion.

[0052] Subsequently, static probabilistic fusion is performed based on a Bayesian network, which is used to characterize the inherent, physical-mechanism-based static probabilistic dependencies among the aforementioned features. Mutual information I(X;Y) is used as a statistical measure to calculate the dependency strength between all feature pairs, not just individual variables (such as temperature and strain). The formula for calculating mutual information is:

[0053] Here, X and Y represent any two feature variables. The larger the value, the stronger the relationship between the two variables. The optimal network topology is determined by a search algorithm (such as particle swarm optimization) to ensure that it conforms to engineering physical constraints such as "temperature changes affect strain, and in turn affect frequency".

[0054] After determining the network structure, the conditional probability distribution of each node is learned. The core advantage of Bayesian networks lies in decomposing the high-dimensional joint probability distribution into a product of a series of conditional probabilities:

[0055] in, It is a high-dimensional joint probability distribution, representing all variables. arrive The probability of taking a set of preset values ​​at the same time serves to decompose the complex high-dimensional joint probability calculation into a product of a series of simple local conditional probabilities. It is the i-th feature variable; It represents The set of parent nodes is all nodes that directly point to... The nodes represent the variables. By injecting observed data (i.e., extracted feature values), the network performs probabilistic inference and calculates the posterior distribution of all variables.

[0056] The outputs of probabilistic inference (such as the posterior expectation and probability values ​​of variables) are extracted to form a preliminary fusion feature matrix. This includes continuous features based on posterior expectation, classification features based on probability confidence, and statistical features based on uncertainty measures. Each feature column of the matrix incorporates probabilistic representations of multi-source information.

[0057] Then, dynamic feature enhancement is performed based on an attention mechanism. Since Bayesian networks represent static dependencies, while actual data interactions are dynamic, an attention mechanism is introduced for enhancement. This mechanism dynamically calculates the mutual influence weights of different feature modes at specific times. For example, when the wind speed approaches the critical eddy shedding frequency, the attention weight of the "wind speed" feature on the "structure frequency" feature is automatically increased. The output of the attention mechanism is the enhanced feature matrix after dynamic weighting and residual connections. It effectively separates effects from different sources. The specific process is as follows: First, the CrossmodalTransformer module is introduced to calculate dynamic intermodal attention weights, in order to adaptively capture complex spatiotemporal dependencies between data, as shown in the following formula:

[0058] in, It is the feature vector of the target mode; is the feature vector of the source mode; d is the feature dimension, which is the optimal dimension determined on the validation set through cross-validation, and is set to 64. These are attention weights. Multimodal offset generation uses convolution operations to generate an offset tensor, as shown in the following formula:

[0059] in, It is the output offset tensor, which is used to identify "interference" or "offset" patterns that are mixed in with the target modal features but are caused by other source modes; It is a one-dimensional convolution operation; The kernel size is 3 to capture local temporal patterns; the number of kernels is 32, the stride is 1, and the output offset is concatenated with the target modal feature residual, as shown in the following formula:

[0060] in, These are enhanced features, whose function is to separate the estimated "interference" effect from the original features, thereby enhancing the main features of the target mode. It is the offset tensor; It is the preliminary fusion feature matrix obtained after the initial fusion of Bayesian networks.

[0061] Finally, the enhanced features are subjected to nonlinear transformation and dimensionality reduction through a fully connected layer to generate the final output of step S2, the building structure data matrix M, which is expressed as follows:

[0062] Where W is the weight matrix, obtained through backpropagation optimization training; b is the bias term, obtained through backpropagation optimization training. ( ) is the ReLU activation function. This matrix M∈ It is a high-level feature representation that has been deeply integrated and refined. It contains complementary and consistent information from multiple sources and will serve as the direct input for step S3 (core factor extraction).

[0063] This step, through the collection and deep fusion processing of simulation data from a refined digital building model, constructs a high-quality building structure data matrix. It transforms multimodal simulation data into feature representations with clear physical meaning and statistical robustness, providing a standardized and highly reliable input data foundation for subsequent assessments. This significantly improves the integrity and consistency of the data, eliminating the limitations of a single data source and the uncertainties caused by sensor heterogeneity. Furthermore, it provides a standardized and highly reliable input data foundation for subsequent health status assessments, damage diagnosis, and early warning analysis, enabling subsequent models to more accurately capture the evolution patterns of structural states and support the reliability of the decision-making process.

[0064] Step S3: Use principal component analysis to extract core building structure factors from the building structure data matrix to obtain core building structure factors. The core building structure factors include real-time data and historical data of core building structure factors. Construct a training dataset based on the historical data of core building structure factors.

[0065] First, the building structure data matrix M is preprocessed. The m-dimensional simulation output parameters, containing N time points, are standardized using Z-scores to eliminate dimensional differences. The calculation formula is as follows:

[0066] in, It is a standardized matrix The value of the i-th time point and the j-th parameter serves to eliminate dimensions; It is a building structure data matrix, from the output of S2, with dimensions of N×m, where N is the time point, i.e. the number of samples, and m is the number of monitoring parameters (feature dimension). It is the original value of the j-th monitoring parameter at the i-th time point in matrix M; It is the mean of the j-th monitoring parameter at all time points; It is the standard deviation of the j-th monitoring parameter at all time points. The standardized matrix. Meeting the requirements of zero mean and unit variance lays the foundation for subsequent factor extraction.

[0067] Next, the covariance matrix of the standardized data is calculated using the following formula:

[0068] Where, Σ∈ It is a symmetric positive semi-definite matrix; It is a matrix The transpose of the array results in a dimension of m×N; diagonal elements Represents the variance of the i-th type of sensor data, with off-diagonal elements. (i j) Characterize the covariance relationship between different monitoring parameters (such as vibration frequency and temperature change); N-1 is used to correct the natural bias that occurs when using sample data instead of population data, and to ensure that the estimate is unbiased.

[0069] The eigenvalue decomposition of the covariance matrix is ​​shown in the following formula:

[0070] in, The eigenvalues ​​are arranged in descending order. ≥ ≥…≥ A diagonal matrix consisting of (≥0) ; It is an eigenvector matrix, column vector For the corresponding eigenvalues The eigenvectors. Each eigenvalue This reflects the amount of original data variance that can be explained by the direction of this principal component. To determine the number k of core structural factors, the cumulative contribution rate of the first k principal components is calculated using the following formula:

[0071] in, It is the cumulative variance contribution rate of the first k principal components; It is the i-th largest eigenvalue; This represents the planned number of principal components (core factors of the building structure) to be retained. A cumulative contribution rate threshold of 95% is set, and the smallest integer k is chosen such that… ≥95%. The eigenvectors corresponding to the first k eigenvalues ​​form the projection matrix Q∈ The column vectors represent the core structural factor loads that influence the structural state of the building. Projecting the standardized data onto the principal component space yields the core structural factor matrix, as shown in the following formula:

[0072] in, It is the core factor matrix of building structure. ∈ Each column is a sequence of principal component (PC) scores. These principal components are linear combinations of the original multi-source simulated variables, capable of preserving the information that contributes most to the variance in the original data with the lowest possible dimensionality. The specific physical meaning of these core factors needs to be determined by analyzing the relationship between their corresponding loading vectors (i.e., eigenvectors in the projection matrix) and the original physical variables.

[0073] To construct the training dataset, a digital twin model that is consistent with the initial state of the real structure must first be established. This model utilizes a highly detailed digital architectural model. The model is obtained by modifying the initial baseline monitoring data, thereby effectively incorporating real-world factors such as construction errors and material property deviations, making its simulation data highly reliable.

[0074] Subsequently, based on this digital twin model Simulations are performed to generate historical data on core structural factors. Before constructing training samples, this historical data undergoes standardization preprocessing (e.g., Z-score standardization) consistent with real-time data to eliminate the influence of dimensions and ensure uniform feature scale. To effectively capture the dynamic evolution of structural states, a sliding time window method is used to construct training samples, avoiding the simplistic treatment of time points as independent samples. With a window length of L, a training sample can be represented as... Its corresponding tags This represents the structural state assessment result at the end of the window. This label is obtained through structural mechanics calculations or performance indicators defined by specifications. Finally, the training set... ={ , This is used for model training in step S4. The core building structure factors extracted through principal component analysis have orthogonal properties, which can effectively eliminate redundancy between multi-source data while retaining more than 95% of the variance information of the original data.

[0075] In practical applications, real-time monitoring modules (such as accelerometers, strain gauges, and temperature and humidity sensors) deployed on the building structure continuously collect the latest time-series data. This real-time monitoring data undergoes standardized preprocessing identical to historical data, and the principal component analysis (PCA) projection matrix trained in this step is used to project the real-time data into the same feature space as the historical core factors, generating a core factor sequence representing the current state. This latest sequence is then organized into real-time data of the building structure's core factors according to a predetermined window length L. This information can be directly input into the stability assessment model trained in step S4 to calculate the current stability assessment result, thereby completing the online early warning.

[0076] This step uses principal component analysis to reduce the dimensionality of the standardized multi-source building monitoring structural data matrix, extracting core structural factors that characterize the structural state. The key function of this step is to transform high-dimensional heterogeneous sensor data into low-dimensional orthogonal feature representations, eliminating redundancy and dimensional differences between multi-source data while retaining more than 95% of the original data variance information. The extracted core structural factors provide highly condensed and physically meaningful input features for the subsequent stability assessment model. The training dataset built based on historical building structural data directly supports the parameter learning of the S4 stage assessment model, significantly improving the model's sensitivity to structural state changes and its generalization ability, laying a data foundation for achieving accurate structural health early warning.

[0077] Step S4: Based on the Stacking ensemble learning framework, construct a building structure stability assessment model. Train the building structure stability assessment model using the training dataset of the core building structure factors to obtain a trained building structure stability assessment model. Input the real-time data of the core building structure factors into the trained building structure stability assessment model and output the building structure stability assessment result.

[0078] This step directly utilizes the real-time data of the core building structure factors that have been preprocessed and extracted in step S3. As input features, repetitive data preprocessing and factor extraction processes are omitted. The model employs an ensemble learning strategy based on the Stacking method, improving prediction accuracy and generalization ability through multi-layer model fusion. The entire process, including model building, training, and evaluation, aims to achieve automated and highly accurate stability assessment.

[0079] The core of the model construction is the Stacking ensemble framework, which consists of two layers: the first layer comprises multiple heterogeneous base models, and the second layer is a meta-model. The base models include CatBoost, XGBoost, GBDT, and ExtraTrees, each independently learning the nonlinear relationship between core structural factors and stability. The CatBoost model excels at handling categorical features and avoiding prediction bias, using OrderedBoosting; the XGBoost model is based on gradient boosting decision trees and prevents overfitting through regularization; the GBDT model iteratively builds weak learners; and the ExtraTrees model increases diversity through random feature selection. The meta-model uses LogisticRegression to fuse the output probabilities of the base models. The mathematical expression for Stacking can be simplified to the following formula:

[0080] Where F represents the core factor vector of the building structure; These are the prediction functions of each base model, outputting the instability probability. LR is the mapping function of the meta-model, which generates the building structure instability probability through a linear combination of these probabilities. (Range 0 to 1). This structure leverages the advantages of different algorithms, avoids the limitations of a single model, and improves robustness.

[0081] The training process uses a training dataset (containing core building structure factors and corresponding stability labels). First, the data imbalance problem is addressed, and then hyperparameter optimization and model training are performed.

[0082] To address data imbalance, the ADASYN algorithm is used to generate synthetic samples to increase the amount of data for the minority class (such as unstable samples). The key steps of ADASYN are as follows: First, calculate the number of samples G that need to be synthesized, using the following formula:

[0083] in, This is the total number of new samples that need to be synthesized for the minority class; It is the majority class, i.e., the stable number of samples; This represents the number of minority class samples, i.e., the unstable samples; β is the balance parameter, set to 1 to indicate perfect balance. For each minority class sample... The proportion of majority class samples among its P-nearest neighbors is calculated using the following formula:

[0084] in, Represents the minority class sample The density ratio of the majority of surrounding species; In the sample The number of samples belonging to the majority class among P nearest neighbors; P is the range of nearest neighbors, P=5. The weights are obtained by normalization, as shown in the following formula:

[0085] in, It is the normalized weight of the i-th minority class sample; It is all minority class samples The sum. Then, calculate the number of synthetic samples to be generated for each sample, using the following formula:

[0086] in, It is the specific number of new samples that need to be generated for the i-th minority class sample; This represents the total number of new samples needed to be generated for the minority class. Synthetic samples are generated through linear interpolation, using the following formula:

[0087] in, These are newly synthesized artificial minority class samples; These are the minority class samples for which new samples need to be generated; From the sample A minority class sample is randomly selected from the P nearest neighbors; λ is a random number in the range [0,1]. This ensures a more uniform distribution of the minority class sample and avoids the model being biased towards the majority class.

[0088] Hyperparameter optimization is performed using a grid search. Key parameters of each base model are tuned. The grid search iterates through all parameter combinations, and the parameters that perform best on the validation set are selected through five-fold cross-validation to maximize AUC and recall.

[0089] The base model is trained independently on the training set, outputting probability predictions. These probabilities are then used as new features to train the meta-model LR. Five-fold cross-validation is used during training to avoid overfitting. The meta-model is optimized by minimizing the cross-entropy loss function, calculated as follows:

[0090] in, It is the true label of the i-th sample; This represents the probability of structural instability predicted by the model for the i-th sample; N is the number of samples. Ultimately, the model parameters are fixed for subsequent evaluation.

[0091] After training, the building structure stability assessment model can be used to evaluate the stability of building structures in real time. It takes as input real-time monitored core factor data of the building structure, i.e., the k-dimensional factor feature vector obtained from step S3, the output being the probability of building structure instability. .

[0092] Input processing: Real-time data is directly input into the pre-trained building structure stability assessment model.

[0093] Prediction and Decision Making: The building structure stability assessment model outputs the probability of building structure instability. The value ∈ [0,1] indicates a higher risk of instability as it approaches 1. A binary classification decision is made based on a preset threshold, which can be adjusted according to actual needs. In high-risk scenarios, the threshold can be lowered to improve sensitivity.

[0094] This step constructs a building structure stability assessment model based on core building structure factors. It integrates base models such as CatBoost, XGBoost, GBDT, and ExtraTrees using a Stacking ensemble strategy, and employs LogisticRegression as the meta-model for probabilistic fusion, thereby efficiently learning nonlinear relationships and improving prediction accuracy. During training, the ADASYN algorithm is used to address data imbalance issues, and grid search is used to optimize hyperparameters, ensuring the model achieves high AUC and high recall on the test set. Finally, the model outputs the probability of building structure instability for real-time decision-making. The purpose of this step is to provide an automated and highly reliable assessment tool for engineering practice, accurately identifying instability risks and providing a core model foundation for subsequent empirical applications, real-time monitoring integration, or early warning mechanisms, thus supporting the intelligent and scientific management of the entire building structure health management process.

[0095] Step S5: Compare the building structure stability assessment results with preset thresholds to obtain graded early warning results.

[0096] This step is based on the structural instability probability obtained in step S4. (Range from 0 to 1, higher values ​​indicate greater instability risk), compared with a preset threshold to generate tiered early warning results. This process is implemented through mathematical decision rules: First, based on historical data and the validation set, a baseline binary classification threshold Tbase balancing false positives and false negatives is determined through ROC curve analysis, obtained by maximizing the Youden index. Based on this, combined with engineering experience and risk tolerance, intervals are divided, defining a tiered threshold vector T=[ , , For example, setting the low-risk range to [0, -Δ), the medium-risk range is [ -Δ, +Δ), the high-risk range is [ +Δ,1], and finally determined through verification and adjustment. , , The specific value, 0.3, 0.6, 0.8.

[0097] The decision rule is formalized as follows:

[0098] in, The threshold is the probability of structural instability output by the ensemble model, obtained by stacking the results of the fused base models. These are dynamically adjustable parameters, for example... Set it to a low value (0.3) to capture potential risks but reduce false negatives. Set it to a high value (0.8) to ensure the reliability of high-risk warnings.

[0099] This tiered early warning mechanism maps the continuous probability values ​​output by the model to discrete risk levels, providing standardized input for subsequent emergency response and decision support, thereby enhancing the initiative and accuracy of engineering safety management. The entire process relies on the model's output and empirical performance calibration to ensure that the early warning results accurately reflect the risk status of the structure.

[0100] For different warning levels, the corresponding tiered response plans will be automatically activated: for the Stable level ( <0.3, maintain routine automated monitoring and periodic report generation without special intervention; for LowRisk levels (0.3≤ For values ​​<0.6, a detailed diagnostic report is automatically generated, prompting engineers to verify the data validity, initiating a preliminary analysis procedure, and arranging an upcoming on-site inspection; for Medium Risk levels (0.6≤...),... <0.8, automatically increase the depth and frequency of data analysis (e.g., initiate more refined time-frequency analysis or damage localization algorithms), and simultaneously notify the expert team for remote consultation and assessment, and formulate contingency plans (e.g., prepare reinforcement materials and plans); for HighRisk level ( If the value is ≥0.8, the highest level alarm will be triggered immediately, automatically notifying the relevant emergency management departments to activate the emergency plan (such as personnel evacuation and traffic control) and dispatching an expert team to the site for final assessment and rescue decision-making.

[0101] These response plans are designed based on historical accident data and engineering management standards. The core of them is to accurately match automated data analysis, professional human judgment, and engineering response actions to form a closed-loop management system, thereby comprehensively improving the response speed and efficiency to major risks.

[0102] This step compares the probability values ​​output by the building structure stability assessment model with preset thresholds, thus mapping continuous probability to discrete risk levels. This provides clear and actionable graded early warning results for project management. The core function of this step is to transform complex probability outputs into intuitive risk classifications, enhancing the real-time nature and relevance of decision-making. It also provides standardized inputs for subsequent emergency response, resource allocation, and integration (such as automatically triggering monitoring measures or activating emergency plans), ensuring the proactiveness, accuracy, and automation of the risk management process, and ultimately improving overall safety and response efficiency.

[0103] This invention presents a multi-source data fusion method for structural stability assessment and early warning, achieving precise monitoring and early warning through a full-process technical architecture: First, the data collection scope is determined based on the building structure type, collecting building structure design parameters such as geometric dimensions, material properties, and node characteristics. A digital building model is constructed and corrected using a sensitivity update algorithm to generate a highly refined digital building model that closely matches the actual structure. Next, for typical working conditions such as earthquakes and strong winds, the dynamic response, environmental loads, and morphological changes of the simulated structure are simulated. A Bayesian network algorithm is used to construct the probabilistic dependencies of the multi-source data to achieve the fusion of building structure simulation data. Then, a CrossmodalTransformer attention mechanism is introduced to dynamically calculate modal weights and separate interference signals. A building structure data matrix is ​​constructed. Principal component analysis (PCA) is then used to reduce the dimensionality of the matrix, extracting core structural factors while retaining over 95% of the variance information. A training dataset is built based on historical data of these core structural factors. A stacking ensemble strategy is then employed to construct a building structure stability assessment model, integrating multi-base models such as CatBoost and XGBoost with a LogisticRegression meta-model. The ADASYN algorithm handles data imbalance, and grid search optimizes hyperparameters, achieving accurate stability assessments for different structural types. Finally, the assessed building structure instability probability is compared with preset thresholds to generate tiered early warning results and match differentiated response measures. This method improves data reliability through deep fusion of multi-source data, reduces computational costs through core structural factor extraction, and enhances assessment accuracy through the building structure stability assessment model. It provides scientific support for intelligent health management of building structures, effectively improving safety control capabilities and operational efficiency.

[0104] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

Claims

1. A method for structural stability assessment and early warning of multi-source data fusion, characterized in that, Includes the following steps: Step S1: Obtain the building structure type, obtain building structure design parameters based on the building structure type, and construct a digital building model based on the building structure design parameters, including: First, identify and obtain the building structure type of the building to be evaluated. This information is used to guide the focus of data collection. For frame structures, the focus should be on the cross-sectional properties of beams and columns and the node connection parameters, while for shear wall structures, the focus should be on the wall thickness, reinforcement and boundary constraints. Subsequently, based on the structure type, four core building structure design parameters were collected, including geometric dimension parameters G and material property parameters. Connection node characteristic parameters and boundary condition parameters ; A digital building model is constructed based on the building structural design parameters using parametric modeling methods. Its mathematical expression is: ； in, It is a digital building model; function This refers to the process of converting design parameters into a physical mapping of a computable model through the interface between CAD and finite element analysis software. The specific rules are as follows: geometric dimension parameters G are converted into the geometric topology of the model; material property parameters... Assign material constitutive relations to the model; assign property parameters to the connection nodes. Define the connection method between components; define the boundary condition parameters Set as support constraints for the model; Digital building model It is a digital benchmark model that is physically equivalent to the actual structure, possessing two main characteristics: its geometry and topology are consistent with the actual structure; and its mechanical parameters conform to design specifications, allowing it to be used for subsequent S2 stage working condition simulations; design parameters yes The input basis, It is the physical representation of the parameters. If any parameter is missing, This will fail to accurately reflect the structural dynamics; parameter accuracy directly determines... Reliability; Step S2: Simulate typical working conditions based on the digital building model to obtain building structure simulation data. Use a Bayesian network algorithm to perform heterogeneous data fusion on the building structure simulation data to obtain a building structure data matrix, including: Building structure data simulation requires a detailed digital building model. The simulation covers four dimensions: physical response, environmental load, morphological changes, and acoustic emission characteristics. The specific simulation content and technical standards are as follows: Structural dynamic response simulation in a highly detailed digital building model Virtual monitoring points are set up in key parts, and the vibration response of the structure under typical working conditions is simulated by finite element analysis. The vibration acceleration time history data is simulated and output, with a sampling frequency ≥100Hz. The natural frequency and damping ratio of the structure in the range of 0.5-10Hz are extracted by Fourier transform. The layout of virtual measuring points must meet the requirements of spatial modal analysis. Environmental load simulation in a detailed digital building model A temperature field is applied to the interior and exterior surfaces of the building to simulate the temperature gradient effect; wind speed and direction time series data are input into the model boundary conditions, with particular attention paid to the coupling relationship between vortex-induced vibration frequency and the natural frequency of the structure; the ground motion time history is used as input to simulate the peak ground acceleration, and the analysis threshold is set to 0.005g. Morphological change simulation, based on continuum mechanics, monitors the overall deformation of the structure and the initiation and development of macroscopic cracks by calculating the displacement and strain fields of the model; its output is field variables, which are used to evaluate the stiffness degradation and geometric stability of the structure. Acoustic emission characteristic simulation, based on the material damage constitutive model and acoustic principles, simulates the stress wave signals released during micro-damage; its output is a sequence of data characterizing acoustic emission events; these parameters are active indicators of damage in progress, and can provide early warning of micro-damage before macro-cracks appear. Step S3: Use principal component analysis to extract the core factors of the building structure data matrix to obtain the core factors of the building structure. The core factors of the building structure include real-time data and historical data of the core factors of the building structure. Construct a training dataset based on the historical data of the core factors of the building structure. Step S4: Based on the Stacking ensemble learning framework, construct a building structure stability assessment model. Train the building structure stability assessment model using the training dataset of the building structure core factors to obtain a trained building structure stability assessment model. Input the real-time data of the building structure core factors into the trained building structure stability assessment model and output the building structure stability assessment result. Step S5: Compare the building structure stability assessment results with preset thresholds to obtain graded early warning results.

2. The structural stability assessment and early warning method for multi-source data fusion according to claim 1, characterized in that, The heterogeneous data fusion of the building structure simulation data is performed using a Bayesian network algorithm to obtain a building structure data matrix, including: Static probabilistic fusion is performed based on Bayesian networks, which are used to characterize the inherent, physical-mechanism-based static probabilistic dependencies between the aforementioned features. Mutual information I(X;Y) is used as a statistical measure to calculate the dependency strength between all feature pairs, not just individual variables. The formula for calculating mutual information is: ； Here, X and Y represent any two characteristic variables. The larger the value, the closer the relationship between the two variables. The optimal network topology is determined by a search algorithm so that it conforms to the engineering physical constraint that "temperature changes affect strain, and in turn affect frequency". After determining the network structure, the conditional probability distribution of each node is learned. The core advantage of Bayesian networks lies in decomposing the high-dimensional joint probability distribution into a product of a series of conditional probabilities: ； in, It is a high-dimensional joint probability distribution, representing all variables. arrive The probability of taking a set of preset values ​​at the same time serves to decompose the complex high-dimensional joint probability calculation into a product of a series of simple local conditional probabilities. It is the i-th feature variable; It represents The set of parent nodes is all nodes that directly point to... The variables represented by the nodes; by injecting observation data, the network performs probabilistic inference and calculates the posterior distribution of all variables; The output of probabilistic inference is extracted to form a preliminary fusion feature matrix. This includes continuous features based on posterior expectation, classification features based on probability confidence, and statistical features based on uncertainty measures. Each feature column of the matrix incorporates probabilistic representations of multi-source information. Then, dynamic feature enhancement is performed based on the attention mechanism. Since Bayesian networks represent static dependencies, while actual data interactions are dynamic, an attention mechanism is introduced for enhancement. This mechanism dynamically calculates the mutual influence weights of different feature modalities at a specific time. The output of the attention mechanism is the enhanced feature matrix after dynamic weighting and residual connections. It effectively separates effects from different sources, and the specific process is as follows: The dynamic intermodal attention weights are calculated to adaptively capture complex spatiotemporal dependencies between data, as shown in the following formula: ； in, It is the feature vector of the target mode; is the feature vector of the source mode; d is the feature dimension, which is the optimal dimension determined on the validation set through cross-validation, and is set to 64. These are attention weights; multimodal offset generation is achieved by generating an offset tensor through convolution operations, as shown in the following formula: ； in, It is the output offset tensor, which is used to identify "interference" or "offset" patterns that are mixed in with the target modal features but are caused by other source modes; It is a one-dimensional convolution operation; The kernel size is 3 to capture local temporal patterns; the number of kernels is 32, the stride is 1, and the output offset is concatenated with the target modal feature residual, as shown in the following formula: ； in, These are enhanced features, whose function is to separate the estimated "interference" effect from the original features, thereby enhancing the main features of the target mode. It is the offset tensor; It is the preliminary fusion feature matrix obtained after the initial fusion of Bayesian networks; Finally, the enhanced features are subjected to nonlinear transformation and dimensionality reduction through a fully connected layer to generate the final output of step S2, the building structure data matrix M, which is expressed as follows: ； Where W is the weight matrix, obtained through backpropagation optimization training; b is the bias term, obtained through backpropagation optimization training. ( ) is the ReLU activation function; this matrix M∈ It is a high-level feature representation that has been deeply integrated and refined. It contains complementary and consistent information from multiple sources and will be used as the direct input for step S3.

3. The structural stability assessment and early warning method for multi-source data fusion according to claim 2, characterized in that, Principal component analysis was used to extract core structural factors from the building structure data matrix, resulting in the following core structural factors: The standardized data is projected onto the principal component space to obtain the core factor matrix of the building structure, as shown in the following formula: ； in, It is a standardized building structure data matrix; It is the eigenvector matrix; It is the core factor matrix of building structure. ∈ Each column is a sequence of principal component scores, which are linear combinations of the original multi-source simulated variables, capable of preserving the information that contributes most to the variance in the original data with the lowest possible dimensionality; N is the number of samples; It represents the number of principal components.

4. The structural stability assessment and early warning method for multi-source data fusion according to claim 3, characterized in that, The core factors of the building structure include real-time data and historical data of the core factors of the building structure. A training dataset is constructed based on the historical data of the core factors of the building structure to obtain feature differences, including: To construct the training dataset, a digital twin model that is consistent with the initial state of the real structure must first be established. This model utilizes a highly detailed digital architectural model. The model is obtained by revising it based on the initial baseline monitoring data, thereby effectively incorporating real-world factors such as construction errors and material property deviations, making its simulation data highly reliable. Subsequently, based on this digital twin model Simulations are performed to generate historical data on core structural factors. This historical data undergoes standardization preprocessing, consistent with real-time data, before being used to construct training samples to eliminate dimensional influences and ensure uniform feature scales. To effectively capture the dynamic evolution of structural states, a sliding time window method is used to construct training samples, avoiding the simplistic treatment of time points as independent samples. With a window length of L, a training sample can be represented as... Its corresponding tags This label represents the structural state assessment result at the end of the window; it is obtained through structural mechanics calculations or performance indicators defined by specifications. Finally, the training set... ={ , } Used for model training in step S4; In practical applications, real-time monitoring modules deployed on the building structure continuously collect the latest time-series data. This real-time monitoring data undergoes standardized preprocessing that is completely consistent with historical data. Then, a pre-trained principal component analysis projection matrix is ​​used to project the real-time data onto the same feature space as the historical core factors, generating a core factor sequence representing the current state. This latest sequence is then organized into real-time data of the building structure's core factors according to a predetermined window length L. .

5. The structural stability assessment and early warning method for multi-source data fusion according to claim 4, characterized in that, Based on the Stacking ensemble learning framework, a building structure stability assessment model is constructed. This model is trained using the core building structure factor training dataset to obtain the trained model, which includes: The core of the model building is the Stacking ensemble framework, which consists of two layers: the first layer comprises multiple heterogeneous base models, and the second layer is a meta-model. The base models include CatBoost, XGBoost, GBDT, and ExtraTrees, each independently learning the nonlinear relationship between core structural factors and stability. The CatBoost model excels at handling categorical features and avoiding prediction bias, employing OrderedBoosting. The XGBoost model is based on gradient boosting decision trees and prevents overfitting through regularization. The GBDT model iteratively builds weak learners. The ExtraTrees model increases diversity through random feature selection. The meta-model uses LogisticRegression to fuse the output probabilities of the base models. The mathematical expression for Stacking can be simplified to the following formula: ； Where F represents the core factor vector of the building structure; These are the prediction functions of each base model, outputting the instability probability. LR is the mapping function of the meta-model, which generates the building structure instability probability through a linear combination of these probabilities. ; The training process uses a training dataset, first addressing the data imbalance problem, and then performing hyperparameter optimization and model training.

6. The structural stability assessment and early warning method for multi-source data fusion according to claim 5, characterized in that, The real-time data of the core factors of the building structure are input into the trained building structure stability assessment model, and the building structure stability assessment results are output, including: Input processing: Real-time data is directly input into the pre-trained building structure stability assessment model; Prediction and Decision Making: The building structure stability assessment model outputs the probability of building structure instability. ∈[0,1], the closer the value is to 1, the higher the risk of instability.

7. The structural stability assessment and early warning method for multi-source data fusion according to claim 6, characterized in that, The results of the building structure stability assessment are compared with preset thresholds to obtain graded early warning results, including: This step is based on the structural instability probability obtained in step S4. The results are compared with preset thresholds to generate tiered early warning results; The decision rule is formalized as follows: ； in, The threshold is the probability of structural instability output by the ensemble model, obtained by stacking the results of the fused base models. It is a dynamically adjustable parameter. Setting it low allows for the capture of potential risks while reducing underreporting. Setting it to a higher level ensures the reliability of high-risk warnings; For different warning levels, corresponding tiered response plans are automatically activated: For the Stable level, routine automated monitoring and periodic report generation are maintained without special intervention; for the Low Risk level, a detailed diagnostic report is automatically generated, and engineers are prompted to verify the validity of the data, initiate preliminary analysis procedures, and arrange for on-site manual inspections in the near future; for the Medium Risk level, the depth and frequency of data analysis are automatically increased, while an expert team is notified for remote consultation and assessment, and contingency plans are developed; for the High Risk level, the highest level alarm is immediately triggered, relevant emergency management departments are automatically notified, emergency plans are activated, and an expert team is dispatched to the site for final assessment and disaster relief decision-making.

Images