Building structure dynamic anomaly detection method in combination with physical information

Abstract

The invention provides a building structure dynamic anomaly detection method in combination with physical information, relates to the field of anomaly detection, and provides an SDAD-Net anomaly detection model which is composed of a diffusion module, a multi-scale module and a judgment module for high-dimension and strong-coupling building structure dynamic data. The diffusion module is used for realizing accurate amplification and propagation inhibition of real abnormal disturbance in the dynamic response of the building structure, and the multi-scale module is used for systematically describing evolution characteristics of the dynamic response of the building structure on different time scales by performing time segmentation and cross-scale alignment modeling on a long-term historical sequence; and the judgment module completes building structure dynamic anomaly judgment through time propagation modeling and difference measurement.

Description

Technical Field

[0001] This invention belongs to the field of anomaly detection, specifically relating to a method for detecting dynamic anomalies in building structures that combines physical information. Background Technology

[0002] With the rapid development of high-rise buildings, large-span structures, and complex public buildings in cities, building structures are subjected to multiple dynamic forces such as wind loads, traffic vibrations, changes in ambient temperature, and occasional earthquakes throughout their service life. Their dynamic response exhibits obvious time-varying, non-stationary, and multimodal characteristics. During long-term operation, the structure may gradually develop hidden anomalies such as stiffness degradation, connection loosening, and local damage. These anomalies often do not immediately lead to structural failure, but they will have a continuous cumulative impact on structural safety and durability.

[0003] Existing methods for detecting dynamic anomalies in building structures mostly rely on pure data-driven or statistical feature analysis. They determine changes in structural state by modeling monitoring signals such as acceleration, displacement, or strain in the frequency domain, time-frequency domain, or feature space. However, methods that rely solely on data distribution or empirical features often struggle to distinguish between normal response changes caused by environmental disturbances and genuine anomalies caused by structural performance degradation when faced with complex working conditions. Furthermore, models that ignore structural dynamic mechanism constraints lack physical interpretability during anomaly identification, making them prone to false alarms or missed alarms, and thus failing to meet the reliability and safety requirements of engineering applications.

[0004] Machine learning methods can automatically model large amounts of historical monitoring data to uncover the intrinsic correlation features of structural dynamic responses in the temporal and spatial dimensions, thereby improving the sensitivity and detection accuracy of subtle anomalies. Compared with analysis methods that rely on a single physical index, machine learning models have a stronger adaptability in handling nonlinear, non-stationary dynamic responses and multi-condition changes, providing a new technical path for the intelligent identification and long-term online monitoring of building structural dynamic anomalies. Summary of the Invention

[0005] This invention provides a method for detecting dynamic anomalies in building structures that combines physical information. For high-dimensional and strongly coupled dynamic data of building structures, the proposed SDAD-Net anomaly detection model consists of a diffusion module, a multi-scale module, and a judgment module.

[0006] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps:

[0007] Collect data related to building structure dynamics, construct datasets, and preprocess them;

[0008] Constructing a diffusion module: building short-term time windows and long-term historical sequences for building structural dynamic data, and extracting instantaneous impact directions;

[0009] The global impact intensity is calculated by aggregating the instantaneous impact direction vector, and instantaneous event weights are constructed.

[0010] The instantaneous impact direction vector of the building structure dynamic variables is aggregated, and instantaneous event weights are introduced to construct a correlation weight matrix;

[0011] Construct an initial feature matrix, generate a diffusion transition matrix based on the normalization of the correlation weight matrix, and perform multi-step iterative diffusion update to obtain short-term output features;

[0012] Constructing multi-scale modules: dividing long-term historical sequences, constructing steady-state reference vectors using median statistics, and constructing window direction signature vectors;

[0013] A comprehensive metric is constructed based on the window orientation signature vector, the relationship strength is calculated, and a sparse variable graph is built.

[0014] The abnormal energy parameters are obtained by accumulating the deviation measurement results between the sparse variable diagram and the instantaneous impact direction vector, determining the phase center position, and calculating the weighted arc length scale distance.

[0015] The unnormalized time transition weights are calculated based on the orientation bias parameters and the weighted arc length scale distance, and a time transition matrix is ​​constructed.

[0016] The judgment module is constructed by aggregating short-term output features and introducing them into the time transition matrix to obtain the time propagation representation. A global difference measurement sequence is constructed and compared with the judgment threshold to complete the detection of dynamic anomalies in building structures.

[0017] Preferably, dynamic data related to the building structure are collected, including building structure displacement data, inter-story drift angle data, building component strain data, building external load data, building damping change rate data, building equivalent damping ratio data, and building stiffness degradation index data. The original dataset is constructed and preprocessed, and the dataset is divided into two parts in a ratio of 7:2:1.

[0018] Preferably, input building structure dynamic data Constructing a short-term time window With long-term historical sequence ,in This is the starting index for the short window. For the short window length, For time indexing, To determine the number of time steps in the long-term historical sequence, the building structure dynamics data is truncated multiple times using the short window start index and short window length to form short-term time windows. Each short-term time window is numbered according to the truncating order and used as the short window time segment index. Feature encoding and embedding mapping operations are performed to obtain short-window embeddings. , For the number of dynamic variables of the building structure, For the feature dimensions of building structural dynamic data, index each building structural dynamic variable. In time index Extract the set of feature vectors corresponding to the dynamic variables of the building structure. ,in To retrieve the entire index, The feature vectors for the short-term time window are then used. Median statistics are then performed on the feature vector set to obtain the steady-state reference vector of the variable. The specific mathematical model is as follows:

[0019] ;

[0020] In the formula, The steady-state reference vector, For median statistical operations, an instantaneous change vector is constructed for each time index within a short-term time window using the offset between the feature vector set and the corresponding steady-state reference vector. The instantaneous change vector is then normalized in direction to obtain the index of building structural dynamic variables. In time index The instantaneous impact direction vector, specifically the mathematical model is as follows:

[0021] ;

[0022] In the formula, The instantaneous impact direction vector, It is a vector that changes instantaneously. These are stable parameters.

[0023] Preferably, the instantaneous impact direction vector is collected. Aggregation operations are performed on the vector set, and the modulus of the aggregation result is calculated to obtain the time index. The corresponding global impact intensity scalar has the following specific mathematical model:

[0024] ;

[0025] In the formula, For the number of dynamic variables of the building structure, For indexing building structural dynamic variables, The global impact intensity scalar is used, followed by the introduction of a weight sharpness parameter. An exponential mapping is then performed on the time indices within the short-term time window, and the results are normalized to construct the instantaneous event weights corresponding to each time index within the short-term time window. The specific mathematical model is as follows:

[0026] ;

[0027] In the formula, Weights for instantaneous events It is an exponential mapping function. For the short window length, For the weighted sharpness parameter, The summation variable is the time index. Time index within a short time window The global impact strength scalar.

[0028] Preferably, the dynamic variables of the building structure and In the short-term time window length parameter Within a given timeframe, the instantaneous impact direction vectors of the two building structure dynamic variables at each time index are aggregated across time, and instantaneous event weights are introduced. The dynamic variables of the building structure are obtained. point to The instantaneous shock consistency evidence scalar, specifically the mathematical model is as follows:

[0029] ;

[0030] In the formula, For instantaneous impact consistency evidence scalar, These are the building structure dynamic variables. In time index Instantaneous impact direction vector and building structure dynamic variables In time index The instantaneous impact direction vector at the point is used to introduce an adaptive threshold parameter, and combined with the boundary sharpening parameter, threshold projection processing is performed on the instantaneous impact consistency evidence scalar to obtain the association weight. The specific mathematical model is as follows:

[0031] ;

[0032] In the formula, For Sigmoid mapping function, For boundary sharpening parameters, For adaptive threshold parameters, As a correlation weight, during the threshold projection process, the dynamic variables of the building structure are... By introducing a fixed outgoing edge mass budget parameter and applying mass conservation constraints during the calculation of the adaptive threshold parameter, the dynamic variables of the building structure are controlled. The sum of the association weights pointing to other variables is always equal to the out-edge quality budget parameter. The specific mathematical model is as follows:

[0033] ;

[0034] In the formula, To obtain the outgoing quality budget parameters, the correlation weights obtained between the dynamic variables of each building structure are aggregated to construct a correlation weight matrix. .

[0035] Preferably, for each time index Extracting feature vectors from short-term time windows Using the encoded features corresponding to the time index as the initial state, an initial feature matrix is ​​constructed. The specific mathematical model is as follows:

[0036] ;

[0037] In the formula, Time index within a short time window The initial feature matrix, Let be the feature vector of the short-term time window, where To retrieve all indexes, based on the association weight matrix Normalization is performed according to the dimension of building structure dynamic variables. The correlation weights of each building structure dynamic variable pointing to other building structure dynamic variables are proportionally adjusted so that the sum of the outward correlation weights of each building structure dynamic variable satisfies the constraint conditions. A diffusion transfer matrix of mass conservation constraints is constructed. The specific mathematical model is as follows:

[0038] ;

[0039] In the formula, The diffusion transfer matrix, It is a vector of all 1s. To diagonalize the matrix, preset diffusion step number and diffusion step size parameters are introduced. Starting from the initial feature matrix, multi-step iterative diffusion updates are performed on the feature matrices at each time index. During each diffusion iteration, the neighborhood information of the current feature matrix is ​​redistributed through the diffusion transfer matrix, and the update magnitude is adjusted in conjunction with the corresponding diffusion step size parameter. The specific mathematical model is as follows:

[0040] ;

[0041] In the formula, For the first After the diffusion iteration, the time index eigenmatrix For the first After the diffusion iteration, the time index eigenmatrix For diffusion step size parameters, The range of values ​​for the diffusion step index. The feature matrices obtained at each diffusion stage are aggregated along the diffusion step dimension to obtain short-term output features, using the diffusion step number parameter. The specific mathematical model is as follows:

[0042] ;

[0043] In the formula, For the first The feature matrix obtained after diffusion iterations This is a short-term output characteristic.

[0044] Preferably, for long-term historical sequences The timeline is divided into multiple consecutive time segments, and indexed by time segment. The segments are distinguished, and the dynamic variables of each building structure and their corresponding first-order components are defined. Each time segment is divided into time segments, and the time index set for each time segment is determined. Extracting feature embedding representations of building structure dynamic variables at each time index within time segments. ,in For indexing building structural dynamic variables, For time indexing, To retrieve all indexes, median statistical processing is then performed on the feature embedding representations of each time index within the time index set to obtain the index of building structure dynamic variables. and time segment index The corresponding steady-state reference vector, specifically the mathematical model, is as follows:

[0045] ;

[0046] In the formula, The steady-state reference vector, For median statistical processing, for each time index in the time index set, the offset of the feature embedding representation relative to the steady-state reference vector is calculated, and the offset is normalized in direction. This is then used to define the offset for each time index within the time index set. By introducing a time weight parameter and performing weighted aggregation, the window direction signature vector is obtained. The specific mathematical model is as follows:

[0047] ;

[0048] In the formula, For window orientation signature vector, These are stable parameters.

[0049] Preferably, based on the window orientation signature vector Dynamic variables of building structure In the Time segments and building structural dynamic variables Alignment is performed between adjacent time segments, and a maximum lag window parameter is introduced. Multiple candidate time alignment paths are constructed within a range not exceeding the maximum lag window number before and after the current time segment. Under each candidate time alignment path, the dynamic variables of the building structure are calculated. The window direction signature vector and building structure dynamic variables corresponding to the current time segment To measure the directional similarity between window directional signature vectors under corresponding lag time segments, a distance penalty parameter is introduced, and the maximum aligned path is selected to obtain a comprehensive metric value. The specific mathematical model is as follows:

[0050] ;

[0051] This is a comprehensive metric. Index for time lag steps, For building structural dynamic variables In the A window orientation signature vector for each time segment. For building structural dynamic variables In relation to the first Time segment lag Window direction signature vector at each segment, For distance penalty parameters, To select the alignment path with the highest comprehensive metric value within the lag range, a window weight parameter is introduced, and the comprehensive metric values ​​of all time segments are weighted and aggregated to obtain the relationship strength. The specific mathematical model is as follows:

[0052] ;

[0053] In the formula, For the strength of the relationship, For window weight parameters, The total number of time segments is used to map the relationship strength through a monotonically nonlinear mapping, and then the relationship strength is determined by an adaptive threshold parameter and an outgoing edge quality budget parameter. A sparse variable graph is constructed using common constraints. The mapping process is adjusted by the boundary sharpening coefficient, and each building structure dynamic variable is subjected to projection processing with a fixed total outgoing edge. The specific mathematical model is as follows:

[0054] ;

[0055] In the formula, For sparse variable graphs, It is a monotonic nonlinear mapping function. This is the boundary sharpening factor. This is an adaptive threshold parameter.

[0056] Preferably, time-segmented indexing Determine the time segment length parameter contained in this time segment. And according to the relative time position within the time segment. Each time point within the time segment is numbered, and the instantaneous impact direction vectors corresponding to all building structural dynamic variables are obtained, based on a sparse variable graph. Dynamic variables of building structure The instantaneous impact direction vectors are weighted and converged to obtain the dynamic variables of the building structure. The corresponding structural alignment reference vector, for the building structural dynamic variables The deviation between the instantaneous impact direction vector and its corresponding structural alignment reference vector is measured, and the dynamic variables of the building structure are analyzed. The deviation measurement results are accumulated to obtain the abnormal energy parameters. The specific mathematical model is as follows:

[0057] ;

[0058] In the formula, These are abnormal energy parameters. These are the building structure dynamic variables. and In the original time index The instantaneous impact direction vector at the point, For the number of dynamic variables of the building structure, As a structural alignment reference vector, a phase center parameter is introduced, and the relative time position within the time segment is used as the weight benchmark. The scalar anomaly energy parameters at each time position are weighted and converged. A stabilizing parameter is introduced during the convergence process to obtain the phase center position. The specific mathematical model is as follows:

[0059] ;

[0060] In the formula, The position of the phase center. The time segment length parameter, To stabilize the parameters, the phase center position is then linearly normalized to obtain the normalized phase center parameters. By accumulating and processing the anomalous energy parameters at all relative time positions within the time segment, the total energy parameter is obtained. Then, following the order of the time segment index, starting from the initial time segment, each time segment is accumulated sequentially. During the accumulation process, a context-sensitive parameter is introduced to adjust the weight of the total energy parameter of each time segment, resulting in a weighted arc length scale. The specific mathematical model is as follows:

[0061] ;

[0062] In the formula, For weighted arc length scale, Context-sensitive parameter This is a cumulative index, representing the time segment from the start time segment to the current time segment. Intermediate time segment index, For the first The total energy parameter for each time segment, for any two and Calculate the energy-weighted arc length scale for each time segment. and The difference yields the weighted arc-length scaled distance. .

[0063] Preferably, for any and Two time segments are used. Based on the difference between the time segments and by introducing a stable parameter for scaling normalization, a directional bias parameter is constructed. The specific mathematical model is as follows:

[0064] ;

[0065] In the formula, For direction offset parameters, To stabilize the parameters, an arc length attenuation parameter is introduced, and combined with a weighted arc length scaling distance, the weighted arc length scaling distance is exponentially modulated. Simultaneously, a bias intensity parameter is introduced, and the normalized phase center parameter is used as a modulation factor to construct an unnormalized time-transfer weight. The specific mathematical model is as follows:

[0066] ;

[0067] In the formula, For unnormalized time transition weights, For weighted arc length scaled distance, For arc length attenuation parameters, Normalized phase center parameters, For bias intensity parameters, For the exponential mapping operation, the unnormalized time transition weights of each time segment are then subjected to row-level normalization to obtain the time transition matrix. The specific mathematical model is as follows:

[0068] ;

[0069] In the formula, This is the time transition matrix. For stable parameters, The total number of time segments. This serves as an intermediate index for traversing all possible target time segments. This refers to the unnormalized time transition weights corresponding to different target time segments during the normalization process.

[0070] Preferably, the short-term output characteristics The resulting three-dimensional tensor representation, formed by aggregation along the diffusion step dimension, is indexed by building structural dynamic variables. Short-window time segment index and the characteristic dimensions of building structural dynamic data Organize to obtain a short-term structure representation. Using the short-window time segment index as the alignment reference, a time transition matrix is ​​introduced. After weighted combination processing, the time propagation representation is obtained. The specific mathematical model is as follows:

[0071] ;

[0072] In the formula, For the purpose of time propagation, Given the total number of time segments, the short-term structure representation and the time propagation representation within the same time segment are aligned node-by-node, and a difference metric is constructed based on the differences after alignment. Then, the difference measures are aggregated into a global difference measure sequence. The global difference measurement sequence is compared with the set judgment threshold. Based on the comparison result, a building structure dynamic anomaly judgment label is generated to obtain the building structure dynamic anomaly detection result.

[0073] The advantages of the technical effects provided by the present invention in the above technical solution are as follows:

[0074] This invention's diffusion module, through a step-by-step construction of "steady-state reference, directional impact, consistency constraint, and mass conservation diffusion," achieves precise amplification and propagation suppression of real abnormal disturbances in the dynamic response of building structures. Utilizing the median steady-state reference vector within a short time window, it effectively weakens the interference of environmental noise and occasional fluctuations on features, allowing subsequent analysis to focus on the true trend of change relative to the steady state. Secondly, by jointly modeling the instantaneous impact direction vector and the global impact intensity, it uniformly maps multivariate and multidimensional dynamic response changes into physically meaningful impact representations, improving the temporal localization capability of abnormal events. Furthermore, it introduces cross-variable impact consistency evidence and an adaptive threshold projection mechanism, constructing a sparse and stable correlation weight matrix under mass conservation constraints to avoid excessive diffusion without physical basis and enhance the interpretability of the relationships between variables. Finally, based on the diffusion transfer matrix, it performs multi-step feature diffusion updates, enabling local abnormal information to propagate along reasonable structural correlation paths, thereby strengthening abnormal patterns and suppressing isolated noise. This provides stable and physically consistent feature inputs for subsequent multi-scale analysis and judgment modules, effectively improving the reliability and robustness of building structure dynamic anomaly detection under complex conditions.

[0075] This invention's multi-scale module systematically characterizes the evolution of building structure dynamic response across different time scales by performing time segmentation and cross-scale alignment modeling on long-term historical sequences. The module introduces a median steady-state reference vector within each time segment and performs directional normalization and weighted convergence on feature offsets to form a window direction signature vector. This weakens the influence of amplitude and environmental differences while highlighting the dominant trend of structural dynamic response, thereby improving the ability to identify slowly accumulating anomalies. By constructing multiple candidate time alignment paths and introducing the maximum hysteresis window number, adaptive matching of response patterns between different time segments is achieved, effectively capturing the hysteresis propagation and scale migration characteristics of anomalies in the time dimension. Furthermore, by combining distance penalty and window weights, the strength of multi-scale relationships is weighted and integrated. A sparse variable graph is constructed under adaptive threshold and out-of-edge quality budget constraints to ensure the stability and physical consistency of multi-scale correlation results. Finally, the cumulative effect and evolution path of anomalies are quantitatively characterized by indicators such as anomaly energy, phase center, and weighted arc length scaling, providing reliable and interpretable multi-scale structural information support for subsequent anomaly determination.

[0076] Based on the outputs of the diffusion module and the multi-scale module, the judgment module of this invention reliably determines the dynamic anomalies of building structures through time propagation modeling and difference measurement. This module uses a time transition matrix to weightedly combine structural features of different time segments, explicitly characterizing the propagation and continuation of anomaly information in the time dimension, avoiding the random interference caused by single time slice judgment. Subsequently, the short-term structural representation and time propagation representation within the same time segment are aligned node by node. Based on the alignment difference, a difference measurement is constructed, making anomaly identification focus on the degree of deviation of the structural response before and after time propagation, rather than absolute amplitude changes, thereby enhancing the sensitivity to hidden and gradual anomalies. Finally, the difference measurement is aggregated into a global difference measurement sequence and compared with the judgment threshold to output a clear anomaly label result. This judgment module can comprehensively utilize time propagation information and structural difference features, effectively suppress false alarms caused by noise and environmental disturbances, and improve the stability, robustness and engineering interpretability of anomaly detection results. It is suitable for the identification of dynamic anomalies of building structures under complex working conditions. Attached Figure Description

[0077] Figure 1 This is a step-by-step diagram of a method for detecting dynamic anomalies in building structures that incorporates physical information.

[0078] Figure 2 This is a diagram of the diffusion module.

[0079] Figure 3 This is a multi-scale module diagram.

[0080] Figure 4 This is an error distribution diagram of the SDAD-Net anomaly detection model.

[0081] Figure 5 This is a screenshot of the SDAD-Net anomaly detection model.

[0082] Figure 6 This is a histogram showing the distribution of comprehensive judgment indicators for the SDAD-Net anomaly detection model. Detailed Implementation

[0083] This invention proposes a method for detecting dynamic anomalies in building structures by incorporating physical information, the steps of which are as follows: Figure 1 As shown, for high-dimensional and strongly coupled building structure dynamic data, the SDAD-Net anomaly detection model is proposed, consisting of a diffusion module, a multi-scale module, and a judgment module. The diffusion module accurately amplifies and suppresses the propagation of real anomalous disturbances in the building structure dynamic response. The multi-scale module systematically characterizes the evolution characteristics of the building structure dynamic response at different time scales by performing time segmentation and cross-scale alignment modeling on long-term historical sequences. The judgment module completes the determination of building structure dynamic anomalies through time propagation modeling and difference measurement.

[0084] Collect data related to the dynamics of building structures, construct datasets, and preprocess them.

[0085] Furthermore, dynamic data related to the building structure were collected. Data acquisition involved a unified deployment and synchronous acquisition of data from the actual building structure test object: high-precision displacement sensors with a range of ±100mm, resolution ≤0.01mm, and sampling frequency of 50–100Hz were deployed on key floors to acquire structural displacement data; inter-story drift gauges were deployed at the edges of each floor slab or at the core tube location to calculate the inter-story drift angle based on the drift difference between adjacent floors, with the drift angle accuracy controlled within 10⁻⁻⁴. 4 The data is measured in rads. Resistance strain gauges or fiber optic strain sensors with a range of ±3000 με are attached to the beam ends, column bases, and surfaces of key load-bearing components, with a sampling frequency of 100 Hz, to collect strain data of building components. Anemometers, accelerometers, and ground motion recording devices are deployed on the top of the structure or in the external environment to acquire external load data of the building, with wind speed sampling frequency of 1–5 Hz and ground motion and acceleration signal sampling frequency of not less than 200 Hz. The building damping change rate is obtained by sliding time window mode identification of the structure's free vibration or environmental excitation response, with the time window length set to 30–60 s and step size of 5–10 s. The building equivalent damping ratio is calculated based on frequency domain or time domain modal parameter inversion and updated within the same time window. The building stiffness degradation index is characterized by identifying the relative change rate of the structural principal vibration mode frequency over time. All data must be synchronized with a unified clock and aligned with timestamps to form a multi-source, multi-parameter original dataset that can be directly used for training and verification of physical information constraint anomaly detection models. The dataset is then preprocessed and divided into datasets in a 7:2:1 ratio.

[0086] Build a diffusion module, such as Figure 2 As shown, the module is: to construct a short-term time window and a long-term historical sequence for the building structure dynamic data, and to extract the instantaneous impact direction.

[0087] Furthermore, input the building structure dynamic data. Constructing a short-term time window With long-term historical sequence ,in This is the starting index for the short window. For the short window length, For time indexing, To determine the number of time steps in the long-term historical sequence, the building structure dynamics data is truncated multiple times using the short window start index and short window length to form short-term time windows. Each short-term time window is numbered according to the truncating order and used as the short window time segment index. Feature encoding and embedding mapping operations are performed to obtain short-window embeddings. , For the number of dynamic variables of the building structure, For the feature dimensions of building structural dynamic data, index each building structural dynamic variable. In time index Extract the set of feature vectors corresponding to the dynamic variables of the building structure. ,in To retrieve the entire index, The feature vectors for the short-term time window are then used. Median statistics are then performed on the feature vector set to obtain the steady-state reference vector of the variable. The specific mathematical model is as follows:

[0088] ;

[0089] In the formula, The steady-state reference vector, For median statistical operations, an instantaneous change vector is constructed for each time index within a short-term time window using the offset between the feature vector set and the corresponding steady-state reference vector. The instantaneous change vector is then normalized in direction to obtain the index of building structural dynamic variables. In time index The instantaneous impact direction vector, specifically the mathematical model is as follows:

[0090] ;

[0091] In the formula, The instantaneous impact direction vector, It is a vector that changes instantaneously. These are stable parameters.

[0092] In this embodiment, the short window length is preferably set within the range of 20 to 60 time steps. In a specific implementation, 30 time steps are selected as the short window length. This setting ensures that the window contains the complete transient impact process while avoiding interference from low-frequency environmental changes on steady-state estimation due to an excessively long window, thereby improving the response sensitivity to sudden anomalies. The stability parameter is preferably set at 10⁻ 6 Within the range of ~10⁻³, 10⁻³ was selected in the specific implementation. 4 This parameter can effectively prevent numerical divergence or instability during the direction normalization process when the instantaneous change amplitude is small or close to zero, while not significantly biasing the true impact direction.

[0093] The global impact intensity is calculated by aggregating the instantaneous impact direction vector, and instantaneous event weights are constructed.

[0094] Furthermore, the instantaneous impact direction vector is collected. Aggregation operations are performed on the vector set, and the modulus of the aggregation result is calculated to obtain the time index. The corresponding global impact intensity scalar has the following specific mathematical model:

[0095] ;

[0096] In the formula, For the number of dynamic variables of the building structure, For indexing building structural dynamic variables, The global impact intensity scalar is used, followed by the introduction of a weight sharpness parameter. An exponential mapping is then performed on the time indices within the short-term time window, and the results are normalized to construct the instantaneous event weights corresponding to each time index within the short-term time window. The specific mathematical model is as follows:

[0097] ;

[0098] In the formula, Weights for instantaneous events It is an exponential mapping function. For the short window length, For the weighted sharpness parameter, The summation variable is the time index. Time index within a short time window The global impact strength scalar.

[0099] In this embodiment, the weight sharpness parameter is preferably set in the range of 0.5 to 5. In a specific implementation, 2 is selected as the weight sharpness parameter. This setting can moderately enhance the weight ratio corresponding to the high impact intensity time index while maintaining the continuous and smooth change of time weight.

[0100] The instantaneous impact direction vector of the building structure dynamic variables is aggregated, and instantaneous event weights are introduced to construct a correlation weight matrix.

[0101] Furthermore, regarding the dynamic variables of building structures and In the short-term time window length parameter Within a given timeframe, the instantaneous impact direction vectors of the two building structure dynamic variables at each time index are aggregated across time, and instantaneous event weights are introduced. The dynamic variables of the building structure are obtained. point to The instantaneous shock consistency evidence scalar, specifically the mathematical model is as follows:

[0102] ;

[0103] In the formula, For instantaneous impact consistency evidence scalar, These are the building structure dynamic variables. In time index Instantaneous impact direction vector and building structure dynamic variables In time index The instantaneous impact direction vector at the point is used to introduce an adaptive threshold parameter, and combined with the boundary sharpening parameter, threshold projection processing is performed on the instantaneous impact consistency evidence scalar to obtain the association weight. The specific mathematical model is as follows:

[0104] ;

[0105] In the formula, For Sigmoid mapping function, For boundary sharpening parameters, For adaptive threshold parameters, As a correlation weight, during the threshold projection process, the dynamic variables of the building structure are... By introducing a fixed outgoing edge mass budget parameter and applying mass conservation constraints during the calculation of the adaptive threshold parameter, the dynamic variables of the building structure are controlled. The sum of the association weights pointing to other variables is always equal to the out-edge quality budget parameter. The specific mathematical model is as follows:

[0106] ;

[0107] In the formula, To obtain the outgoing quality budget parameters, the correlation weights obtained between the dynamic variables of each building structure are aggregated to construct a correlation weight matrix. .

[0108] In this embodiment, the boundary sharpening parameter is preferably set in the range of 1 to 10, and 5 is selected in the specific implementation. This setting can ensure that the association weight remains sensitive to changes in consistency while avoiding the weight mapping being too smooth or over-binarized, thereby improving the clarity of the boundary discrimination between variables. The adaptive threshold parameter is preferably initially set in the range of 0.3 to 0.7, and 0.5 is selected in the specific implementation. This setting can adaptively distinguish between significant consistency and random correlation under different operating conditions and noise levels, reducing the problems of false connections and missing connections. The outgoing edge quality budget parameter is preferably initially set in the range of 0.8 to 1.2, and 1.0 is selected in the specific implementation. By fixing the total outgoing edge, the conservation and comparability of the association weight in the variable dimension are ensured, and the uncontrolled diffusion and amplification effect of individual variables due to abnormal fluctuations is prevented.

[0109] An initial feature matrix is ​​constructed, and a diffusion transition matrix is ​​generated based on the normalization of the correlation weight matrix. Multi-step iterative diffusion update is performed to obtain short-term output features.

[0110] Furthermore, for each time index Extracting feature vectors from short-term time windows Using the encoded features corresponding to the time index as the initial state, an initial feature matrix is ​​constructed. The specific mathematical model is as follows:

[0111] ;

[0112] In the formula, Time index within a short time window The initial feature matrix, Let be the feature vector of the short-term time window, where To retrieve all indexes, based on the association weight matrix Normalization is performed according to the dimension of building structure dynamic variables. The correlation weights of each building structure dynamic variable pointing to other building structure dynamic variables are proportionally adjusted so that the sum of the outward correlation weights of each building structure dynamic variable satisfies the constraint conditions. A diffusion transfer matrix of mass conservation constraints is constructed. The specific mathematical model is as follows:

[0113] ;

[0114] In the formula, The diffusion transfer matrix, It is a vector of all 1s. To diagonalize the matrix, preset diffusion step number and diffusion step size parameters are introduced. Starting from the initial feature matrix, multi-step iterative diffusion updates are performed on the feature matrices at each time index. During each diffusion iteration, the neighborhood information of the current feature matrix is ​​redistributed through the diffusion transfer matrix, and the update magnitude is adjusted in conjunction with the corresponding diffusion step size parameter. The specific mathematical model is as follows:

[0115] ;

[0116] In the formula, For the first After the diffusion iteration, the time index eigenmatrix For the first After the diffusion iteration, the time index eigenmatrix For diffusion step size parameters, The range of values ​​for the diffusion step index. The feature matrices obtained at each diffusion stage are aggregated along the diffusion step dimension to obtain short-term output features, using the diffusion step number parameter. The specific mathematical model is as follows:

[0117] ;

[0118] In the formula, For the first The feature matrix obtained after diffusion iterations This is a short-term output characteristic.

[0119] In this embodiment, the diffusion step size parameter is preferably set in the range of 0.05 to 0.3, and 0.1 is selected in the specific implementation. This setting can ensure the smooth convergence of the diffusion process while avoiding feature oscillation or excessive smoothing caused by an excessively large step size, so that the abnormal information can be propagated gradually and controllably along the associated structure. The diffusion number parameter is preferably set in the range of 2 to 6 steps, and 4 steps can be selected in the specific implementation. This setting can fully integrate the multi-hop structure association information while preventing the introduction of irrelevant variable information due to excessively deep diffusion layers, which would weaken the locality of the anomaly.

[0120] Constructing multi-scale modules, such as Figure 3 As shown, the module consists of: dividing the long-term historical sequence, constructing a steady-state reference vector using median statistics, and constructing a window direction signature vector.

[0121] Furthermore, regarding the long-term historical sequence The timeline is divided into multiple consecutive time segments, and indexed by time segment. The segments are distinguished, and the dynamic variables of each building structure and their corresponding first-order components are defined. Each time segment is divided into time segments, and the time index set for each time segment is determined. Extracting feature embedding representations of building structure dynamic variables at each time index within time segments. ,in For indexing building structural dynamic variables, For time indexing, To retrieve all indexes, median statistical processing is then performed on the feature embedding representations of each time index within the time index set to obtain the index of building structure dynamic variables. and time segment index The corresponding steady-state reference vector, specifically the mathematical model, is as follows:

[0122] ;

[0123] In the formula, The steady-state reference vector, For median statistical processing, for each time index in the time index set, the offset of the feature embedding representation relative to the steady-state reference vector is calculated, and the offset is normalized in direction. This is then used to define the offset for each time index within the time index set. By introducing a time weight parameter and performing weighted aggregation, the window direction signature vector is obtained. The specific mathematical model is as follows:

[0124] ;

[0125] In the formula, For window orientation signature vector, These are stable parameters.

[0126] A comprehensive metric is constructed based on the window orientation signature vector, the relationship strength is calculated, and a sparse variable graph is built.

[0127] Furthermore, based on the window orientation signature vector Dynamic variables of building structure In the Time segments and building structural dynamic variables Alignment is performed between adjacent time segments, and a maximum lag window parameter is introduced. Multiple candidate time alignment paths are constructed within a range not exceeding the maximum lag window number before and after the current time segment. Under each candidate time alignment path, the dynamic variables of the building structure are calculated. The window direction signature vector and building structure dynamic variables corresponding to the current time segment To measure the directional similarity between window directional signature vectors under corresponding lag time segments, a distance penalty parameter is introduced, and the maximum aligned path is selected to obtain a comprehensive metric value. The specific mathematical model is as follows:

[0128] ;

[0129] This is a comprehensive metric. Index for time lag steps, For building structural dynamic variables In the A window orientation signature vector for each time segment. For building structural dynamic variables In relation to the first Time segment lag Window direction signature vector at each segment, For distance penalty parameters, To select the alignment path with the highest comprehensive metric value within the lag range, a window weight parameter is introduced, and the comprehensive metric values ​​of all time segments are weighted and aggregated to obtain the relationship strength. The specific mathematical model is as follows:

[0130] ;

[0131] In the formula, For the strength of the relationship, For window weight parameters, The total number of time segments is used to map the relationship strength through a monotonically nonlinear mapping, and then the relationship strength is determined by an adaptive threshold parameter and an outgoing edge quality budget parameter. A sparse variable graph is constructed using common constraints. The mapping process is adjusted by the boundary sharpening coefficient, and each building structure dynamic variable is subjected to projection processing with a fixed total outgoing edge. The specific mathematical model is as follows:

[0132] ;

[0133] In the formula, For sparse variable graphs, It is a monotonic nonlinear mapping function. This is the boundary sharpening factor. This is an adaptive threshold parameter.

[0134] In this embodiment, the distance penalty parameter is preferably set in the range of 0.1 to 1.0, and 0.5 is selected in the specific implementation. This setting can apply a moderate penalty to long-distance time segments while allowing reasonable lag alignment, thereby highlighting the main time path of abnormal evolution and avoiding interference from accidental similarity. The window weight parameter is preferably set in the range of 0.5 to 1.5, and 1.0 is selected in the specific implementation. This setting ensures that each time segment has a balanced influence by default, while retaining the flexibility to emphasize key time periods according to the working conditions. The boundary sharpening coefficient is preferably set in the range of 1 to 10, and 4 is selected in the specific implementation. This setting can ensure that weak correlations are effectively suppressed while clearly retaining strong correlation boundaries, avoiding the relationship graph from being too dense or completely discrete.

[0135] The abnormal energy parameters are obtained by accumulating the deviation measurement results of the sparse variable diagram and the instantaneous impact direction vector, determining the phase center position, and calculating the weighted arc length scale distance.

[0136] Furthermore, time-segmented indexing Determine the time segment length parameter contained in this time segment. And according to the relative time position within the time segment. Each time point within the time segment is numbered, and the instantaneous impact direction vectors corresponding to all building structural dynamic variables are obtained, based on a sparse variable graph. Dynamic variables of building structure The instantaneous impact direction vectors are weighted and converged to obtain the dynamic variables of the building structure. The corresponding structural alignment reference vector, for the building structural dynamic variables The deviation between the instantaneous impact direction vector and its corresponding structural alignment reference vector is measured, and the dynamic variables of the building structure are analyzed. The deviation measurement results are accumulated to obtain the abnormal energy parameters. The specific mathematical model is as follows:

[0137] ;

[0138] In the formula, These are abnormal energy parameters. These are the building structure dynamic variables. and In the original time index The instantaneous impact direction vector at the point, For the number of dynamic variables of the building structure, As a structural alignment reference vector, a phase center parameter is introduced, and the relative time position within the time segment is used as the weight benchmark. The scalar anomaly energy parameters at each time position are weighted and converged. A stabilizing parameter is introduced during the convergence process to obtain the phase center position. The specific mathematical model is as follows:

[0139] ;

[0140] In the formula, The position of the phase center. The time segment length parameter, To stabilize the parameters, the phase center position is then linearly normalized to obtain the normalized phase center parameters. By accumulating and processing the anomalous energy parameters at all relative time positions within the time segment, the total energy parameter is obtained. Then, following the order of the time segment index, starting from the initial time segment, each time segment is accumulated sequentially. During the accumulation process, a context-sensitive parameter is introduced to adjust the weight of the total energy parameter of each time segment, resulting in a weighted arc length scale. The specific mathematical model is as follows:

[0141] ;

[0142] In the formula, For weighted arc length scale, Context-sensitive parameter This is a cumulative index, representing the time segment from the start time segment to the current time segment. Intermediate time segment index, For the first The total energy parameter for each time segment, for any two and Calculate the energy-weighted arc length scale for each time segment. and The difference yields the weighted arc-length scaled distance. .

[0143] In this embodiment, the context-sensitive parameter is preferably set in the range of 0.6 to 0.95. In a specific implementation, 0.8 is selected as the context-sensitive parameter. This setting can ensure the temporal continuity of the abnormal energy while avoiding the excessive dominance of early segmented energy on subsequent judgments.

[0144] The unnormalized time transition weights are calculated based on the orientation bias parameter and the weighted arc length scale distance, and a time transition matrix is ​​constructed.

[0145] Furthermore, for any and Two time segments are used. Based on the difference between the time segments and by introducing a stable parameter for scaling normalization, a directional bias parameter is constructed. The specific mathematical model is as follows:

[0146] ;

[0147] In the formula, For direction offset parameters, To stabilize the parameters, an arc length attenuation parameter is introduced, and combined with a weighted arc length scaling distance, the weighted arc length scaling distance is exponentially modulated. Simultaneously, a bias intensity parameter is introduced, and the normalized phase center parameter is used as a modulation factor to construct an unnormalized time-transfer weight. The specific mathematical model is as follows:

[0148] ;

[0149] In the formula, For unnormalized time transition weights, For weighted arc length scaled distance, For arc length attenuation parameters, Normalized phase center parameters, For bias intensity parameters, For the exponential mapping operation, the unnormalized time transition weights of each time segment are then subjected to row-level normalization to obtain the time transition matrix. The specific mathematical model is as follows:

[0150] ;

[0151] In the formula, This is the time transition matrix. For stable parameters, The total number of time segments. This serves as an intermediate index for traversing all possible target time segments. This refers to the unnormalized time transition weights corresponding to different target time segments during the normalization process.

[0152] In this embodiment, the arc length attenuation parameter is preferably set in the range of 0.3 to 1.5, and 0.8 is selected in the specific implementation. This setting can effectively attenuate the time segments that are far apart while allowing the abnormal energy to spread reasonably along the time segments, so as to avoid the unconstrained spread of the abnormal influence in the time dimension and thus highlight the main evolution path.

[0153] The judgment module is constructed by aggregating short-term output features and introducing them into the time transition matrix to obtain the time propagation representation. A global difference measurement sequence is constructed and compared with the judgment threshold to complete the detection of dynamic anomalies in building structures.

[0154] Furthermore, short-term output characteristics The resulting three-dimensional tensor representation, formed by aggregation along the diffusion step dimension, is indexed by building structural dynamic variables. Short-window time segment index and the characteristic dimensions of building structural dynamic data Organize to obtain a short-term structure representation. Using the short-window time segment index as the alignment reference, a time transition matrix is ​​introduced. After weighted combination processing, the time propagation representation is obtained. The specific mathematical model is as follows:

[0155] ;

[0156] In the formula, For the purpose of time propagation, Given the total number of time segments, the short-term structure representation and the time propagation representation within the same time segment are aligned node-by-node, and a difference metric is constructed based on the differences after alignment. Then, the difference measures are aggregated into a global difference measure sequence. The global difference measurement sequence is compared with the set judgment threshold. Based on the comparison result, a building structure dynamic anomaly judgment label is generated to obtain the building structure dynamic anomaly detection result.

[0157] Furthermore, the SDAD-Net anomaly detection model was written in Python, and the experiments were run on a Windows operating system. PyTorch was used as the framework in the CUDA 11.27 environment, and training was performed on a GeForce RTX 3090. The optimizer used was Adam, the initial learning rate was set to 0.001, the training batch size was set to 64, the training period was set to 100, and the dataset consisted of 60 days of building structure dynamics-related data, which were preprocessed and then input into the SDAD-Net anomaly detection model.

[0158] Furthermore, the error distribution diagram of the SDAD-Net anomaly detection model is shown below. Figure 4As shown, the overall error distribution approximates a normal distribution, mainly concentrated near zero, indicating that the model has high prediction accuracy and stability under normal operating conditions, with no significant systematic shift. A small number of tail error samples exist on both sides of the distribution, reflecting significant deviations in the structural response at certain moments; these samples correspond to potential abnormal states. The SDAD-Net anomaly detection model performance is shown in the figure below. Figure 5 As shown, the horizontal axis represents time, the vertical axis represents the anomaly indicator, and the dashed line represents the anomaly judgment threshold. During the initial normal operation phase, the anomaly indicator is generally at a low level with minimal fluctuations, indicating a stable system state. The anomaly indicator gradually increases and significantly approaches and exceeds the threshold in the middle to later stages, indicating a sustained shift in the structural response characteristics and the gradual accumulation of potential anomalies. Later, the indicator shows significant fluctuations in the high-level range, reflecting a more pronounced and somewhat unstable anomaly state. The histogram of the comprehensive judgment index distribution of the SDAD-Net anomaly detection model is shown below. Figure 6 As shown, the indicators exhibit a clear multi-peak distribution, with the low-value range having the highest sample proportion, mainly concentrated in a small range, corresponding to the system being in a normal or slightly disturbed state. The medium- and high-value ranges have relatively fewer samples, but the distribution is more dispersed, reflecting that the system has significant anomalies or state deterioration in some time periods. There are clear distribution intervals between different ranges, indicating that the comprehensive judgment indicators can effectively distinguish between normal and abnormal states, providing an intuitive basis for setting anomaly thresholds and classifying states.

Claims

1. A method for detecting dynamic anomalies in building structures by combining physical information, characterized in that, Includes the following steps: Collect data related to building structure dynamics, construct datasets, and preprocess them; Constructing a diffusion module: building short-term time windows and long-term historical sequences for building structural dynamic data, and extracting instantaneous impact directions; The global impact intensity is calculated by aggregating the instantaneous impact direction vector, and instantaneous event weights are constructed. The instantaneous impact direction vector of the building structure dynamic variables is aggregated, and instantaneous event weights are introduced to construct a correlation weight matrix; Construct an initial feature matrix, generate a diffusion transition matrix based on the normalization of the correlation weight matrix, and perform multi-step iterative diffusion update to obtain short-term output features; Constructing multi-scale modules: dividing long-term historical sequences, constructing steady-state reference vectors using median statistics, and constructing window direction signature vectors; A comprehensive metric is constructed based on the window orientation signature vector, the relationship strength is calculated, and a sparse variable graph is built. The abnormal energy parameters are obtained by accumulating the deviation measurement results between the sparse variable diagram and the instantaneous impact direction vector, determining the phase center position, and calculating the weighted arc length scale distance. The unnormalized time transition weights are calculated based on the orientation bias parameters and the weighted arc length scale distance, and a time transition matrix is ​​constructed. The judgment module is constructed by aggregating short-term output features and introducing them into the time transition matrix to obtain the time propagation representation. A global difference measurement sequence is constructed and compared with the judgment threshold to complete the detection of dynamic anomalies in building structures.

2. The method for detecting dynamic anomalies in building structures by combining physical information according to claim 1, characterized in that, Inputting building structure dynamic data, short-term time windows and long-term historical sequences are constructed. The building structure dynamic data is truncated multiple times using the short-window start index and short-window length to form short-term time windows. Each short-term time window is numbered according to the truncating order and used as a short-window time segment index. Feature encoding and embedding mapping operations are performed to obtain the short-window embedding. For each building structure dynamic variable index, the corresponding feature vector set is extracted within the time index. Then, median statistics are performed on the feature vector set to obtain the steady-state reference vector of the variable. For each time index within the short-term time window, an instantaneous change vector is constructed using the offset between the feature vector set and the corresponding steady-state reference vector. The instantaneous change vector is then normalized in direction to obtain the instantaneous impact direction vector of the building structure dynamic variable index at the time index.

3. The method for detecting dynamic anomalies in building structures by combining physical information according to claim 2, characterized in that, The instantaneous impact direction vectors are aggregated into a vector set for aggregation operations. The magnitude of the aggregation result is calculated to obtain the global impact intensity scalar corresponding to the time index. Then, a weight sharpness parameter is introduced, and the time indexes within the short-term time window are subjected to exponential mapping. The exponential mapping result is normalized to construct the instantaneous event weights corresponding to each time index within the short-term time window.

4. The method for detecting dynamic anomalies in building structures by combining physical information according to claim 3, characterized in that, For pairs of building structure dynamic variables, within the time range of the short-term time window length parameter, the instantaneous impact direction vectors corresponding to the two building structure dynamic variables at each time index are aggregated across time, and instantaneous event weights are introduced to obtain an instantaneous impact consistency evidence scalar pointing from one building structure dynamic variable to another. An adaptive threshold parameter is introduced, and threshold projection processing is performed on the instantaneous impact consistency evidence scalar in combination with boundary sharpening parameters to obtain association weights. During the threshold projection process, a fixed outbound mass budget parameter is introduced for the building structure dynamic variables, and a mass conservation constraint is applied during the calculation of the adaptive threshold parameter, so that the sum of the association weights of the building structure dynamic variables pointing to other building structure dynamic variables is always equal to the outbound mass budget parameter. The association weights obtained between each building structure dynamic variable are aggregated to construct an association weight matrix.

5. The method for detecting dynamic anomalies in building structures by combining physical information according to claim 4, characterized in that, For each time index, the encoded features of the corresponding time index in the feature vector of the short-term time window are extracted as the initial state to construct an initial feature matrix. Based on the correlation weight matrix, the matrix is ​​normalized according to the dimension of building structure dynamic variables. The correlation weights of each building structure dynamic variable pointing to other building structure dynamic variables are proportionally adjusted so that the sum of the outward correlation weights of each building structure dynamic variable satisfies the constraint conditions. A diffusion transfer matrix with mass conservation constraints is constructed. Preset diffusion step number parameters and diffusion step size parameters are introduced. Starting from the initial feature matrix, the feature matrix of each time index is iteratively updated through multiple steps. In the k-th diffusion iteration, the neighborhood information of the current feature matrix is ​​redistributed through the diffusion transfer matrix, and the update magnitude is adjusted in combination with the corresponding diffusion step size parameters. The feature matrices obtained in each diffusion stage are aggregated along the diffusion step dimension to obtain short-term output features.

6. The method for detecting dynamic anomalies in building structures by combining physical information according to claim 1, characterized in that, For long-term historical sequences, the timeline is divided into multiple consecutive time segments, and each segment is distinguished by a time segment index. For each building structure dynamic variable and its corresponding time segment, a time index set for that time segment is determined. The feature embedding representations of the building structure dynamic variables at each time index within the time segment are extracted. Then, median statistical processing is performed on the feature embedding representations of each time index within the time index set to obtain a steady-state reference vector corresponding to the building structure dynamic variable index and the time segment index. For each time index within the time index set, the offset of the feature embedding representation relative to the steady-state reference vector is calculated, and the offset is normalized in direction. A time weight parameter is introduced for each time index within the time index set for weighted aggregation to obtain the window direction signature vector.

7. The method for detecting dynamic anomalies in building structures by combining physical information according to claim 6, characterized in that, The relationship between building structure dynamic variables in adjacent time segments is aligned based on the window orientation signature vector. A maximum lag window number parameter is introduced. Multiple candidate time alignment paths are constructed within the range of the maximum lag window number before and after the current time segment. Under each candidate time alignment path, the directional similarity metric between the window orientation signature vector of the building structure dynamic variable in the current time segment and the window orientation signature vector of another building structure dynamic variable in the corresponding lag time segment is calculated. A distance penalty parameter is introduced, and the maximum alignment path is selected to obtain a comprehensive metric value. Then, a window weight parameter is introduced, and the comprehensive metric values ​​of all time segments are weighted and converged to obtain the relationship strength. After the relationship strength is monotonically nonlinearly mapped, a sparse variable graph is constructed by constraining the adaptive threshold parameter and the out-edge quality budget parameter. The mapping process is adjusted by the boundary sharpening coefficient, and a projection processing with a fixed total out-edge amount is performed on each building structure dynamic variable.

8. The method for detecting dynamic anomalies in building structures by combining physical information according to claim 7, characterized in that, The time segment index determines the length parameter of the time segment contained in the time segment, and the time points within the time segment are numbered according to their relative time positions. The instantaneous impact direction vectors corresponding to all building structure dynamic variables are obtained. Based on a sparse variable graph, the instantaneous impact direction vector of one variable in a pair of building structure dynamic variables is weighted and converged to obtain a structural alignment reference vector corresponding to the other building structure dynamic variable. The deviation between the instantaneous impact direction vector of the other building structure dynamic variable and its corresponding structural alignment reference vector is measured, and the deviation measurement results of the other building structure dynamic variable are accumulated to obtain anomaly energy parameters. A phase center parameter is introduced and the relative time within the time segment is used to determine the abnormal energy parameters. Using location as the weighting benchmark, the scalar anomalous energy parameters at each time location are weighted and converged. A stabilizing parameter is introduced during the convergence process to obtain the phase center location. Then, the phase center location is linearly normalized to obtain the normalized phase center parameter. The anomalous energy parameters of all relative time locations within the time segment are accumulated to obtain the total energy parameter. Then, according to the order of the time segment index, starting from the initial time segment, each time segment is accumulated sequentially. During the accumulation process, a context-sensitive parameter is introduced to adjust the weight of the total energy parameter of each time segment to obtain the weighted arc length scale. For any two time segments, the difference in their weighted arc length scale is calculated to obtain the weighted arc length scale distance.

9. The method for detecting dynamic anomalies in building structures by combining physical information according to claim 8, characterized in that, For any two time segments, scale normalization is performed based on the difference between the time segments and by introducing stable parameters. A directional bias parameter is constructed, an arc length attenuation parameter is introduced, and the weighted arc length scale distance is exponentially modulated by combining the weighted arc length scale distance. At the same time, a bias intensity parameter is introduced, and the normalized phase center parameter is used as a modulation factor to construct unnormalized time transition weights. Then, row-level normalization is performed on the unnormalized time transition weights of each time segment to obtain the time transition matrix.

10. The method for detecting dynamic anomalies in building structures by combining physical information according to claim 1, characterized in that, The three-dimensional tensor representation formed by aggregating short-term output features along the diffusion step dimension is organized using the building structure dynamic variable index, short-window time segment index, and feature dimensions of building structure dynamic data to obtain a short-term structural representation. Using the short-window time segment index as the alignment benchmark, a time transition matrix is ​​introduced for weighted combination processing to obtain a time propagation representation. The short-term structural representation and the time propagation representation under the same time segment are aligned node by node, and a difference metric is constructed based on the difference after alignment. Then, the difference metrics are aggregated into a global difference metric sequence, and the global difference metric sequence is compared with a set judgment threshold. Based on the comparison result, a building structure dynamic anomaly judgment label is generated to obtain the building structure dynamic anomaly detection result.

Images