Digital Monitoring and Protection System for Historical Buildings Based on Multi-Scale Features

Abstract

This invention discloses a digital monitoring and protection system for historical buildings based on multi-scale features. The system includes a building data acquisition module, a raw data optimization module, a feature extraction model construction module, a state recognition model construction module, and a building digital monitoring module. This invention relates to the field of digital monitoring technology for historical buildings, specifically a digital monitoring and protection system for historical buildings based on multi-scale features. This invention obtains raw data through building data acquisition; employs data optimization methods such as data alignment, data cleaning, data standardization, and dataset segmentation; uses a deep learning model as the feature extraction model, decoupling features driven by temporal modeling and physical laws to learn the deep correlation features between time-varying building behavior and structural damage; and uses an improved clustering model as the state recognition model, achieving robust identification of building states by introducing physical threshold constraint spectral clustering, dynamic time kernel functions, and multi-scale probability fusion mechanisms.

Description

Technical Field

[0001] This invention relates to the field of digital monitoring technology for historical buildings, specifically to a digital monitoring and protection system for historical buildings based on multi-scale features. Background Technology

[0002] The digital monitoring and protection system for historical buildings is a system that uses modern information technology to monitor, collect, and analyze historical buildings in real time. It accurately obtains changes in structural, environmental, temperature, and humidity parameters of buildings through digital means, promptly detects potential damage or dangers, effectively extends the service life of historical buildings, avoids damage caused by natural aging or external factors, and provides strong technical support for the protection of historical culture.

[0003] However, traditional digital monitoring and protection systems for historical buildings suffer from several technical problems. They often rely on fixed signal processing algorithms or shallow machine learning models to extract features, which are insufficiently adaptable to the nonlinear degradation process unique to buildings. They also struggle to distinguish between environmental disturbances and actual structural damage, lack a mechanism for embedding physical laws, and are susceptible to sensor noise and changes in operating conditions. Furthermore, traditional digital monitoring and protection systems for historical buildings cannot capture the time-varying characteristics of building damage evolution in terms of state identification, ignore material constitutive relationships and environment-structure interaction effects, tend to forcibly classify data with different physical mechanisms, and their single-scale feature analysis cannot simultaneously take into account the global macroscopic response and local microscopic damage of the building. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a digital monitoring and protection system for historical buildings based on multi-scale features. Traditional digital monitoring and protection systems for historical buildings often rely on fixed signal processing algorithms or shallow machine learning models for feature extraction, resulting in insufficient adaptability to the unique nonlinear degradation processes of buildings, difficulty in distinguishing between environmental disturbances and actual structural damage, and a lack of physical law embedding mechanisms, making them susceptible to sensor noise and changes in operating conditions. This solution creatively employs a deep learning model as the feature extraction model. Through adaptive temporal modeling and physical law-driven feature decoupling, it automatically learns the deep correlation features between the time-varying behavior of buildings and structural damage, reflecting the complex environment-structure coupling effect of buildings. Furthermore, orthogonal constraints are used to ensure the independence between features, providing more discriminative input for subsequent state identification. Addressing the technical challenges of traditional digital monitoring and protection systems for historical buildings—namely, their inability to capture the time-varying characteristics of building damage evolution, their neglect of material constitutive relationships and environment-structure interaction effects, their tendency to forcibly categorize data with different physical mechanisms, and their difficulty in simultaneously considering both the global macroscopic response and local microscopic damage through single-scale feature analysis—this solution creatively employs an improved clustering model as the state identification model. By introducing physical threshold constraint spectral clustering and adapting to the progressive nature of building damage through a dynamic time kernel function, combined with a multi-scale probabilistic fusion mechanism, it achieves fine-grained perception and robust identification of changes in building state.

[0005] The technical solution adopted by the present invention is as follows: The digital monitoring and protection system for historical buildings based on multi-scale features provided by the present invention includes a building data acquisition module, a raw data optimization module, a feature extraction model construction module, a status recognition model construction module, and a building digital monitoring module;

[0006] The building data acquisition module obtains the original dataset of historical buildings by collecting data.

[0007] The original data optimization module uses data alignment, data cleaning, data standardization and dataset splitting to obtain the monitoring dataset, monitoring training set and monitoring test set.

[0008] The feature extraction model construction module constructs a deep learning model as a feature extraction model to provide building attribute features for subsequent digital building monitoring.

[0009] The state recognition model construction module constructs an improved clustering model as a state recognition model for subsequent digital building monitoring.

[0010] The building digital monitoring module specifically performs building digital monitoring by using the feature extraction model in combination with the state recognition model to obtain a reference result of the building digital monitoring state.

[0011] Furthermore, in the building data acquisition module, the historical building original dataset specifically includes the building's past original dataset and the building's current original dataset. Both the building's past original dataset and the building's current original dataset include building structure data, building environment data, and building material performance data. The building's past original dataset also includes building damage level labeling data.

[0012] Furthermore, in the original data optimization module, the data alignment is used to solve the problem of sensor spatiotemporal asynchrony in historical building monitoring. Specifically, it is achieved by synchronizing the clocks of each device and constructing a timestamp alignment matrix to obtain data with strictly synchronized timelines.

[0013] The data cleaning is used to ensure the accuracy and integrity of the data. Specifically, it involves eliminating impulse noise through median filtering and removing baseline drift through empirical mode decomposition to obtain cleaned data.

[0014] The data standardization is used to eliminate dimensional distortion caused by environmental factors. Specifically, it corrects displacement deformation data and power-standardized vibration energy through temperature compensation, and uses the Z-Score standardization method to obtain standardized data.

[0015] The dataset segmentation is used to divide the dataset, specifically by segmenting the original historical dataset of the building to obtain a monitoring training set and a monitoring test set.

[0016] The existing original dataset of the building is optimized through data alignment, data cleaning, and data standardization to obtain a monitoring dataset. The past original dataset of the building is optimized through data alignment, data cleaning, data standardization, and dataset segmentation to obtain a monitoring training set and a monitoring test set.

[0017] Furthermore, in the feature extraction model construction module, a model is used to construct the model required for extracting building attribute features, specifically a deep learning model as the feature extraction model. The deep learning model is specifically a temporal-structural decoupling encoder model, including a feature extraction module, a temporal feature branch, a structural feature branch, and a decoupling fusion module.

[0018] The feature extraction model construction module specifically includes constructing a feature extraction module, constructing a temporal feature branch, constructing a structural feature branch, constructing a decoupling and fusion module, and constructing and training the model.

[0019] The constructed feature extraction module is used to transform the raw data into a feature matrix that the model can process. Specifically, it constructs temporal, frequency, and structural features of the building as input to the model, including:

[0020] Temporal feature extraction is used to capture the instantaneous energy changes of historical building structures. Specifically, dynamic energy changes are calculated using the Teager energy operator to obtain the temporal features of the building.

[0021] Frequency domain feature extraction is used to identify frequency response changes caused by the aging of building materials. Specifically, it extracts spectral features through Fourier transform to obtain the building's frequency domain features.

[0022] Structural feature construction is used to quantify the degree of building damage. Specifically, it involves selecting extreme values, gradients, and acceleration statistics from the input data to construct features, thereby obtaining structural features of the building that reflect structural stability.

[0023] The constructed temporal feature branch is used to monitor the environmental impact of historical buildings. Specifically, it obtains the temporal features of the buildings by combining convolutional networks and temporal attention mechanisms. The content includes:

[0024] Adaptive temporal convolution is used to extract local temporal patterns. Specifically, it captures the correlation between adjacent time points by sliding a one-dimensional convolutional kernel to obtain primary temporal features.

[0025] An activation function is designed to enhance sensitivity to abnormal fluctuations. Specifically, an adaptive threshold-switching bimodal nonlinear transformation is used as the activation function for building attributes.

[0026] Multi-scale temporal dilation is used to perceive multi-scale changes. Specifically, it captures long-term and short-term dependencies in parallel through convolutional layers with different dilation rates to obtain multi-scale temporal features.

[0027] The design time attention is used to focus on key time nodes. Specifically, it processes multi-scale temporal features through self-attention weights, dynamically allocates feature importance, and obtains architectural temporal features.

[0028] The constructed structural feature branch is used to assess the building damage state. Specifically, it obtains the building structural state features through feature transformation guided by physical constraints, including:

[0029] Design physical constraints to embed material mechanical properties, specifically by constraining the feature space distribution through constitutive equations to obtain physical structural features that conform to physical laws;

[0030] The design of structural attention is used to locate high-risk building components. Specifically, the physical structural features are reduced in dimensionality by using a convolutional layer with a kernel size of 1×1, and then the feature weights of key structural points are enhanced by attention gating to obtain the structural state features of the building.

[0031] The decoupled fusion module is used to realize temporal-structural feature separation and joint expression. Specifically, it forces feature independence by decoupling orthogonal constraints and feature pooling, and constructs a joint expression to obtain building attribute features.

[0032] The construction and training of the model specifically involves constructing a deep learning model through the construction of a feature extraction module, a construction of a temporal feature branch, a construction of a structural feature branch, and a construction of a decoupling and fusion module; training the model based on the monitoring training set; and verifying the model performance based on the monitoring test set to obtain a deep learning model, which serves as the feature extraction model.

[0033] Furthermore, in the state recognition model construction module, the model used to construct the model required to identify the level of building damage state is specifically an improved clustering model, which serves as the state recognition model.

[0034] The state recognition model construction module specifically includes constructing a feature pyramid module, dynamic time kernel K-means clustering, physical constraint spectrum clustering, feature pyramid co-clustering, and building and training the model.

[0035] The constructed feature pyramid module is used to generate multi-scale feature representations to adapt to different types of historical building structures. Specifically, it obtains a three-level feature pyramid through linear projection transformation.

[0036] The dynamic time kernel K-means clustering is used for clustering temporal features. Specifically, it obtains category labels based on temporal features by combining adaptive wavelet kernel optimization with kernel K-means clustering. The content includes:

[0037] The optimal scale is selected to capture the time-varying characteristic period. Specifically, the optimal analysis scale for feature alignment is obtained by optimizing the scale parameters through wavelet transform.

[0038] Distance matrix calculation is used to measure temporal feature similarity. Specifically, it is performed by calculating the temporal distance matrix through wavelet space Euclidean distance.

[0039] Kernel spatial clustering is used to obtain category labels based on temporal features. Specifically, it iteratively optimizes the cluster centers using the kernel K-means clustering method to obtain category labels based on temporal features.

[0040] The physically constrained spectral clustering is used to cluster structural features. Specifically, it constrains sample correlation through physical thresholds and performs spectral clustering to obtain category labels based on structural features, including:

[0041] Similarity matrix construction is used to avoid spurious correlation structural features. Specifically, it controls the relationship between samples by setting a physical threshold to obtain a physically constrained similarity matrix.

[0042] Laplace matrix construction is used to model the damage propagation path in building structures;

[0043] The eigenvalues ​​and eigenvectors are calculated by performing eigenvector decomposition on the Laplacian matrix to obtain the eigenvalues ​​and eigenvectors of the Laplacian matrix.

[0044] K-means clustering is performed to obtain category labels based on structural features. Specifically, the eigenvalues ​​of the Laplacian matrix are sorted in ascending order, the eigenvectors corresponding to the top n eigenvalues ​​are selected, and K-means clustering is used to obtain category labels based on structural features.

[0045] The feature pyramid collaborative clustering is used to fuse multi-scale clustering results to eliminate single-view bias. Specifically, it obtains the final identification result through probabilistic fusion and consensus optimization, including:

[0046] Obtain multi-scale clustering probabilities to weight the credibility of each scale. Specifically, obtain the clustering probabilities of each scale through temperature scaling probability transformation.

[0047] The initial consensus allocation matrix is ​​used to balance the contributions of different scales. Specifically, it is obtained by dynamically weighting the multi-scale clustering probabilities using the silhouette coefficient.

[0048] Consensus clustering optimization specifically involves minimizing divergence through Laplace regularization to obtain the final identification result label;

[0049] The construction and training of the model specifically involves constructing an improved clustering model through the construction of the feature pyramid module, the dynamic time kernel K-means clustering, the physical constraint spectrum clustering, and the feature pyramid co-clustering. The model is then trained based on the monitoring training set, and its performance is verified based on the monitoring test set to obtain the improved clustering model, which serves as the state recognition model.

[0050] Furthermore, in the building digital monitoring module, the monitoring dataset is first used as the input to the feature extraction model to obtain a building attribute feature set based on the monitoring dataset. The building attribute feature set based on the monitoring dataset is then used as the input to the state recognition model to obtain a building digital monitoring state reference result. The building digital monitoring state reference result is specifically the final recognition result label obtained by the state recognition model.

[0051] The beneficial effects achieved by the present invention using the above solution are as follows:

[0052] (1) In view of the technical problems of traditional digital monitoring and protection systems for historical buildings, which mostly use fixed signal processing algorithms or shallow machine learning models to extract features, are not adaptable to the nonlinear degradation process unique to buildings, have difficulty distinguishing between environmental interference and real structural damage, lack physical law embedding mechanism, and are easily affected by sensor noise and changes in operating conditions, this solution creatively adopts a deep learning model as the feature extraction model. Through adaptive temporal modeling and physical law-driven feature decoupling, it automatically learns the deep correlation features between the time-varying behavior of buildings and structural damage, reflects the complex environment-structure coupling effect of buildings, and ensures the independence between features through orthogonal constraints, providing more discriminative input for subsequent state recognition.

[0053] (2) In view of the technical problems of traditional digital monitoring and protection systems for historical buildings, such as the inability to capture the time-varying characteristics of building damage evolution in terms of state identification, ignoring material constitutive relations and environment-structure interaction effects, forcibly classifying data with different physical mechanisms, and the difficulty of taking into account the global macroscopic response and local microscopic damage of buildings by single-scale feature analysis, this solution creatively adopts an improved clustering model as the state identification model. By introducing physical threshold constraint spectrum clustering, and adapting to the gradual characteristics of building damage through dynamic time kernel function, combined with multi-scale probability fusion mechanism, fine-grained perception and robust identification of building state changes are realized. Attached Figure Description

[0054] Figure 1 A schematic diagram of the modules of the digital monitoring and protection system for historical buildings based on multi-scale features provided by the present invention;

[0055] Figure 2 A flowchart illustrating the process of optimizing the raw data module;

[0056] Figure 3 A flowchart illustrating the process of building a module for the feature extraction model;

[0057] Figure 4 A flowchart illustrating the process of building a module for the state recognition model;

[0058] Figure 5 This is a flowchart illustrating the process of constructing temporal feature branches in the feature extraction model building module.

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0060] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0061] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0062] Example 1, see Figure 1 The present invention provides a digital monitoring and protection system for historical buildings based on multi-scale features, comprising a building data acquisition module, a raw data optimization module, a feature extraction model construction module, a status recognition model construction module, and a building digital monitoring module;

[0063] The building data acquisition module obtains the original dataset of historical buildings by collecting data.

[0064] The original data optimization module uses data alignment, data cleaning, data standardization and dataset splitting to obtain the monitoring dataset, monitoring training set and monitoring test set.

[0065] The feature extraction model construction module constructs a deep learning model as a feature extraction model to provide building attribute features for subsequent digital building monitoring.

[0066] The state recognition model construction module constructs an improved clustering model as a state recognition model for subsequent digital building monitoring.

[0067] The building digital monitoring module specifically performs building digital monitoring by using the feature extraction model in combination with the state recognition model to obtain a reference result of the building digital monitoring state.

[0068] In Example 2, in the building data acquisition module, the historical building original dataset specifically includes the building's past original dataset and the building's current original dataset. Both the building's past original dataset and the building's current original dataset include building structure data, building environment data, and building material performance data. The building's past original dataset also includes building damage level labeling data.

[0069] The building structure data specifically includes vibration acceleration data, displacement deformation data, and building tilt angle data. The vibration acceleration data is specifically the vibration response collected by acceleration sensors installed at the beam-column joints of the building. The displacement deformation data is specifically the displacement data collected by displacement gauges installed at key joints of the building. The building tilt angle data is specifically the building tilt angle collected by inclinometers deployed at the load-bearing columns and base of the building.

[0070] The building environment data specifically includes ambient temperature and humidity data covering the interior and exterior of the building as well as different heights, and wind speed and direction data around the building.

[0071] The building material performance data specifically includes building material acoustic emission data and building material electrical resistance change data;

[0072] The building damage level labeling data specifically refers to building damage status level labels, including intact, minor damage, moderate damage, and severe damage.

[0073] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the original data optimization module, the data alignment is used to solve the problem of sensor spatiotemporal asynchrony in historical building monitoring. Specifically, it obtains data with strict time axis synchronization by synchronizing the clocks of each device and constructing a timestamp alignment matrix.

[0074] The data cleaning is used to ensure the accuracy and integrity of the data. Specifically, it involves eliminating impulse noise through median filtering and removing baseline drift through empirical mode decomposition to obtain cleaned data.

[0075] The data standardization is used to eliminate dimensional distortions caused by environmental factors. Specifically, it involves correcting displacement deformation data and power-standardized vibration energy through temperature compensation, and then using the Z-Score standardization method to obtain standardized data. The formulas used for temperature compensation correction of displacement deformation data and power-standardized vibration energy are as follows:

[0076] ;

[0077] In the formula, This represents the displacement and deformation data after compensation. This represents the original displacement and deformation data, where Te represents the material expansion coefficient. This indicates the change in building ambient temperature. This represents the standardized vibration acceleration data. This represents the raw vibration acceleration data, where T represents the length of the time window. This represents the instantaneous raw vibration acceleration data at time t;

[0078] The dataset segmentation is used to divide the dataset, specifically by segmenting the original historical dataset of the building to obtain a monitoring training set and a monitoring test set.

[0079] The existing original dataset of the building is optimized through data alignment, data cleaning, and data standardization to obtain a monitoring dataset. The past original dataset of the building is optimized through data alignment, data cleaning, data standardization, and dataset segmentation to obtain a monitoring training set and a monitoring test set.

[0080] Example 4, see Figure 1 , Figure 3 and Figure 5 This embodiment is based on the above embodiment. In the feature extraction model construction module, a model is used to construct the model required to extract building attribute features. Specifically, a deep learning model is constructed as the feature extraction model. The deep learning model is specifically a temporal-structural decoupling encoder model, which includes a feature extraction module, a temporal feature branch, a structural feature branch, and a decoupling fusion module.

[0081] The feature extraction model construction module specifically includes constructing a feature extraction module, constructing a temporal feature branch, constructing a structural feature branch, constructing a decoupling and fusion module, and constructing and training the model.

[0082] The constructed feature extraction module is used to transform the raw data into a feature matrix that the model can process. Specifically, it constructs temporal, frequency, and structural features of the building as input to the model, including:

[0083] Temporal feature extraction is used to capture the instantaneous energy changes of historical building structures. Specifically, dynamic energy changes are calculated using the Teager energy operator to obtain the temporal features of the building. The formula used is as follows:

[0084] ;

[0085] In the formula, Indicates the temporal characteristics of architecture. This represents the Teager energy operator function. This represents the input data for the deep learning model;

[0086] Frequency domain feature extraction is used to identify frequency response changes caused by the aging of building materials. Specifically, it extracts spectral features through Fourier transform to obtain the building's frequency domain features.

[0087] Structural feature construction is used to quantify the degree of building damage. Specifically, it involves selecting extreme values, gradients, and acceleration statistics from the input data to construct features, thereby obtaining structural features of the building that reflect structural stability.

[0088] The constructed temporal feature branch is used to monitor the environmental impact of historical buildings. Specifically, it obtains the temporal features of the buildings by combining convolutional networks and temporal attention mechanisms. The content includes:

[0089] Adaptive temporal convolution is used to extract local temporal patterns. Specifically, it captures the correlation between adjacent time points by sliding a one-dimensional convolutional kernel to obtain primary temporal features. The formula used is as follows:

[0090] ;

[0091] In the formula, Indicates primary tense characteristics, Represents a one-dimensional convolution function. Represents the frequency domain characteristics of buildings;

[0092] An activation function is designed to enhance sensitivity to abnormal fluctuations. Specifically, an adaptive threshold-switching bimodal nonlinear transform is used as the building attribute activation function, and the formula is as follows:

[0093] ;

[0094] In the formula, This represents the building attribute activation function, where x represents the input independent variable. Represents the hyperbolic tangent function. This represents the sigmoid activation function. This represents the SiLU activation function. , and These are mutually exclusive learnable parameters;

[0095] Multi-scale temporal dilation is used to perceive multi-scale changes. Specifically, it captures long-short-term dependencies in parallel through convolutional layers with different dilation rates to obtain multi-scale temporal features. The formula used is as follows:

[0096] ;

[0097] In the formula, This represents the temporal characteristics at the first scale. This represents the temporal characteristics at the second scale. Indicates the temporal characteristics at the third scale. This represents a dilated convolution function with a dilation rate of 1. This represents a dilated convolution function with a dilation rate of 3. This represents a dilated convolution function with a dilation rate of 5. Represents multi-scale temporal features. This represents a learnable multi-scale weight matrix;

[0098] The design time attention is used to focus on key time nodes. Specifically, it processes multi-scale temporal features through self-attention weights, dynamically allocates feature importance, and obtains architectural temporal features.

[0099] The constructed structural feature branch is used to assess the building damage state. Specifically, it obtains the building structural state features through feature transformation guided by physical constraints, including:

[0100] The design incorporates physical constraints to embed the mechanical properties of materials. Specifically, it constrains the feature space distribution through constitutive equations to obtain physical structural features that conform to physical laws. The formulas used are as follows:

[0101] ;

[0102] In the formula, Represents the physical constraint function. Indicates the yield strength of building materials. Indicates the current strain of building materials. Indicates the strain threshold of building materials. Indicates the ultimate strain of building materials. Indicates the structural features of a building. The constitutive matrix of building materials is a diagonal matrix composed of material parameters, and Pe represents the physical structural characteristics. This represents element-wise multiplication;

[0103] The design incorporates structural attention to locate high-risk building components. Specifically, it reduces the dimensionality of physical structural features using a 1×1 convolutional layer, and then strengthens the feature weights of key structural points through attention gating to obtain the building structural state features. The formula used is as follows:

[0104] ;

[0105] In the formula, Indicates activation of structural features, Represents the ReLU activation function. This represents a convolution function with a kernel size of 1×1. Indicates the structural characteristics of a building. Represents learnable structural weights. Represents the tensor product;

[0106] The decoupled fusion module is used to achieve temporal-structural feature separation and joint representation. Specifically, it forces feature independence by decoupling orthogonal constraints and feature pooling, and constructs a joint representation to obtain building attribute features. The formula used is as follows:

[0107] ;

[0108] In the formula, Indicates the temporal attributes of a building. Indicates the structural state and attribute characteristics of a building. Indicates orthogonal loss. This indicates the calculation of the Frobenius norm. This represents the global average pooling function. Indicates the temporal characteristics of architecture. represents the max pooling function, Tr represents the transpose operation, and Ar represents the building attribute feature;

[0109] The construction and training of the model specifically involves constructing a deep learning model through the construction of a feature extraction module, a construction of a temporal feature branch, a construction of a structural feature branch, and a construction of a decoupling and fusion module; training the model based on the monitoring training set; and verifying the model performance based on the monitoring test set to obtain a deep learning model, which serves as the feature extraction model.

[0110] By performing the above operations, this solution addresses the technical problems of traditional digital monitoring and protection systems for historical buildings, which often rely on fixed signal processing algorithms or shallow machine learning models to extract features. These systems are insufficiently adaptable to the unique nonlinear degradation processes of buildings, struggle to distinguish between environmental interference and actual structural damage, lack a physical law embedding mechanism, and are susceptible to sensor noise and changes in operating conditions. This solution creatively adopts a deep learning model as the feature extraction model. Through adaptive temporal modeling and physical law-driven feature decoupling, it automatically learns the deep correlation features between the time-varying behavior of buildings and structural damage, reflecting the complex environment-structure coupling effect of buildings. Orthogonal constraints ensure the independence between features, providing more discriminative input for subsequent state identification.

[0111] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In the state recognition model construction module, a model is used to construct the model required to identify the level of building damage state. Specifically, an improved clustering model is constructed as the state recognition model.

[0112] The state recognition model construction module specifically includes constructing a feature pyramid module, dynamic time kernel K-means clustering, physical constraint spectrum clustering, feature pyramid co-clustering, and building and training the model.

[0113] The constructed feature pyramid module is used to generate multi-scale feature representations to adapt to different types of historical building structures. Specifically, it obtains a three-level feature pyramid through linear projection transformation, using the following formula:

[0114] ;

[0115] In the formula, This represents the architectural attributes and characteristics at the first scale. This represents the architectural attributes at the second scale. This represents the architectural attributes at the third scale. This represents the learnable weight matrix for the first scale. This represents the learnable weight matrix for the second scale. Represents the learnable weight matrix for the third scale;

[0116] The dynamic time kernel K-means clustering is used for clustering temporal features. Specifically, it obtains category labels based on temporal features by combining adaptive wavelet kernel optimization with kernel K-means clustering. The content includes:

[0117] The optimal scale is selected to capture the time-varying periodicity. Specifically, the optimal analysis scale for feature alignment is obtained by optimizing the scale parameters through wavelet transform. The formula used is as follows:

[0118] ;

[0119] In the formula, This represents the optimal analytical scale. The scale parameter of the wavelet transform is represented, and num represents the total number of samples. Represents the continuous wavelet transform function. This represents the temporal attribute features of the building in the i-th sample. This represents the temporal attribute features of the building in the j-th sample. This indicates the calculation of the L2 norm;

[0120] Distance matrix calculation, used to measure temporal feature similarity, is specifically performed by calculating the temporal distance matrix using wavelet spatial Euclidean distance. The formula used is as follows:

[0121] ;

[0122] In the formula, The elements of the time-domain distance matrix represent the wavelet space Euclidean distance between the i-th sample and the j-th sample.

[0123] Kernel spatial clustering is used to obtain category labels based on temporal features. Specifically, it iteratively optimizes the cluster centers using the kernel K-means clustering method to obtain category labels based on temporal features.

[0124] The physically constrained spectral clustering is used to cluster structural features. Specifically, it constrains sample correlation through physical thresholds and performs spectral clustering to obtain category labels based on structural features, including:

[0125] Similarity matrix construction, used to avoid spurious correlation structural features, specifically involves controlling the relationship between samples by setting a physical threshold to obtain a physically constrained similarity matrix. The formula used is as follows:

[0126] ;

[0127] In the formula, This represents the elements of the physical constraint similarity matrix, where pa represents the proportional parameter controlling the rate of similarity decay. This represents the structural state attribute features of the i-th sample. This represents the structural state attribute features of the j-th sample. This represents the k-th dimension of the building structure state attribute features of the i-th sample. This represents the k-th dimension of the building structure state attribute features of the j-th sample. The preset physical threshold representing the k-th dimension of the building structure's state attribute characteristics;

[0128] The Laplace matrix is ​​constructed to model the damage propagation path in building structures, using the following formula:

[0129] ;

[0130] In the formula, Lm represents the Laplacian matrix, and D represents the degree matrix of the physical constraint similarity matrix. Represents the physical constraint similarity matrix;

[0131] The eigenvalues ​​and eigenvectors are calculated by performing eigenvector decomposition on the Laplacian matrix to obtain the eigenvalues ​​and eigenvectors of the Laplacian matrix.

[0132] K-means clustering is performed to obtain category labels based on structural features. Specifically, the eigenvalues ​​of the Laplacian matrix are sorted in ascending order, the eigenvectors corresponding to the top n eigenvalues ​​are selected, and K-means clustering is used to obtain category labels based on structural features.

[0133] The feature pyramid collaborative clustering is used to fuse multi-scale clustering results to eliminate single-view bias. Specifically, it obtains the final identification result through probabilistic fusion and consensus optimization, including:

[0134] The multi-scale clustering probabilities are obtained and used to weight the confidence at each scale. Specifically, the clustering probabilities at each scale are obtained through temperature scaling probability transformation, using the following formula:

[0135] ;

[0136] In the formula, Let m represent the clustering probability at the m-th scale. This represents the Mahalanobis distance calculation function. This represents the building attribute characteristics at the m-th scale. This represents the p-th fusion cluster center, and tem represents the temperature parameter. Let q be the fusion cluster center. This represents the weighting of the dynamic time kernel K-means cluster center transformation. Represents the weights for the transformation of the physical constraint spectrum cluster centers. This represents the p-th dynamic time kernel K-means cluster center. This represents the p-th physical constraint spectrum cluster center;

[0137] The initial consensus assignment matrix, used to balance contributions at different scales, is obtained by dynamically weighting multi-scale clustering probabilities using silhouette coefficients. The formula used is as follows:

[0138] ;

[0139] In the formula, Represents the initial consensus allocation matrix. The consensus assignment matrix, representing the weighted weights of the contour coefficients at the m-th scale, has the following dimensions: ,in Indicates the total number of tags;

[0140] Consensus clustering optimization specifically involves minimizing divergence through Laplacian regularization to obtain the final identification result label, using the following formula:

[0141] ;

[0142] In the formula, Fun represents the objective function to be optimized, and U represents the consensus allocation matrix, which is iteratively updated using the gradient descent method. Represents the regularization parameter. The trace function of the matrix is ​​used to calculate the trace, and La represents the label of the final recognition result. Represents the elements of the consensus allocation matrix;

[0143] The construction and training of the model specifically involves constructing an improved clustering model through the construction of the feature pyramid module, the dynamic time kernel K-means clustering, the physical constraint spectrum clustering, and the feature pyramid co-clustering. The model is then trained based on the monitoring training set, and its performance is verified based on the monitoring test set to obtain the improved clustering model, which serves as the state recognition model.

[0144] By performing the above operations, this solution addresses the technical problems of traditional digital monitoring and protection systems for historical buildings, such as the inability to capture the time-varying characteristics of building damage evolution in terms of state identification, the neglect of material constitutive relations and environment-structure interaction effects, the tendency to forcibly classify data with different physical mechanisms, and the difficulty of taking into account both the global macroscopic response and local microscopic damage in single-scale feature analysis. This solution creatively adopts an improved clustering model as the state identification model. By introducing physical threshold constraint spectral clustering and adapting to the gradual characteristics of building damage through a dynamic time kernel function, combined with a multi-scale probabilistic fusion mechanism, it achieves fine-grained perception and robust identification of changes in building state.

[0145] Example 6, see Figure 1 This embodiment is based on the above embodiment. In the building digital monitoring module, specifically, the monitoring dataset is first used as the input of the feature extraction model to obtain a building attribute feature set based on the monitoring dataset. The building attribute feature set based on the monitoring dataset is then used as the input of the state recognition model to obtain a building digital monitoring state reference result. The building digital monitoring state reference result is specifically the final recognition result label obtained by the state recognition model.

[0146] 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 process, method, article, or apparatus.

[0147] 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.

[0148] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A digital monitoring and protection system for historic buildings based on multi-scale features, characterized by: The system comprises a building data acquisition module, an original data optimization module, a feature extraction model construction module, a state recognition model construction module and a building digital monitoring module. The building data acquisition module acquires historical building original data sets by data acquisition, and the historical building original data sets specifically include building past original data sets and building existing original data sets. The original data optimization module adopts data optimization methods of data alignment, data cleaning, data standardization and data set segmentation to obtain a monitoring data set, a monitoring training set and a monitoring test set. The feature extraction model construction module is used for constructing a model required for extracting building property features, specifically a deep learning model as a feature extraction model, and the deep learning model is specifically a time-state decoupling encoder model comprising a feature extraction module, a time-state feature branch, a structure feature branch and a decoupling fusion module. The state recognition model construction module is used for constructing a model required for recognizing building damage state grades, specifically an improved clustering model as a state recognition model, and the specific content includes constructing a feature pyramid module, dynamic time kernel K-means clustering, physical constraint spectral clustering, feature pyramid collaborative clustering and model construction and training. The building digital monitoring module specifically comprises the following steps: taking the monitoring data set as the input of the feature extraction model to obtain a building property feature set based on the monitoring data set, and taking the building property feature set based on the monitoring data set as the input of the state recognition model to obtain a building digital monitoring state reference result, wherein the building digital monitoring state reference result is specifically a final recognition result label obtained by the state recognition model.

2. The multi-scale feature based historic building digitalization monitoring protection system according to claim 1, wherein: The feature extraction model construction module specifically includes constructing a feature extraction module, constructing a time-state feature branch, constructing a structure feature branch, constructing a decoupling fusion module and constructing and training a model.

3. The multi-scale feature based historic building digitalization monitoring protection system according to claim 2, wherein: The feature extraction module is used for converting original data into a feature matrix that can be processed by a model, and specifically constructs building time domain, frequency domain and structured features as the input of the model, and the content includes: time domain feature extraction, which is used for capturing the instantaneous energy change of a historical building structure, and specifically calculates the dynamic energy change by a Teager energy operator to obtain building time domain features; frequency domain feature extraction, which is used for identifying the frequency response change caused by the aging of building materials, and specifically extracts frequency spectrum features by Fourier transform to obtain building frequency domain features; structure feature construction, which is used for quantifying the building damage degree, and specifically selects extreme values, gradients and acceleration statistics of input data to construct features and obtain building structured features reflecting the stability of the structure; the time-state feature branch construction, which is used for monitoring the environmental impact on a historical building, and specifically obtains building time-state features by combining a convolutional network and a time attention mechanism, and the content includes: adaptive time domain convolution, which is used for extracting local time sequence patterns, and specifically captures the correlation between adjacent time points by a one-dimensional convolution kernel sliding to obtain primary time-state features; An activation function is designed to enhance the sensitivity to abnormal fluctuations, specifically a dual-mode nonlinear transformation with adaptive threshold switching is adopted as the building attribute activation function; A multi-scale time-domain expansion is used to perceive multi-scale changes, specifically, long and short-term dependencies are captured in parallel through convolution layers with different expansion rates to obtain multi-scale time features; A time attention mechanism is designed to focus on key time nodes, specifically, the multi-scale time features are processed through self-attention weights to dynamically allocate feature importance, and building time features are obtained; The structural feature branch is constructed to evaluate the building damage state, specifically, the building structure state features are obtained through a feature transformation guided by physical constraints, including: A physical constraint is designed to embed material mechanical properties, specifically, the feature space distribution is constrained by the constitutive equation to obtain physical structure features that conform to physical laws; A structure attention mechanism is designed to locate high-risk building components, specifically, a 1x1 convolution layer is used to reduce the dimensionality of the physical structure features, and then the feature weights of key structure points are enhanced through attention gating to obtain building structure state features; The decoupling fusion module is constructed to realize time-structure feature separation and joint expression, specifically, the feature independence is enforced through decoupling orthogonal constraints and feature pooling, and joint expression is constructed to obtain building attribute features; The model is constructed and trained, specifically, the feature extraction model is constructed through the feature extraction module, the time feature branch, the structure feature branch, and the decoupling fusion module, the model is trained based on the monitoring training set, the model performance is verified based on the monitoring test set, and the deep learning model is obtained as the feature extraction model.

4. The multi-scale feature based historic building digitalization monitoring protection system according to claim 1, wherein: In the state recognition model construction module, the feature pyramid module is constructed to generate multi-scale feature expression to adapt to different types of historical building structures, specifically, a three-level feature pyramid is obtained through linear projection transformation; The dynamic time kernel K-means clustering is used to cluster time features, specifically, the adaptive wavelet kernel optimization is combined with kernel K-means clustering to obtain class labels based on time features, including: Optimal scale selection is used to capture the period of time-varying characteristics, specifically, the optimal analysis scale of feature alignment is obtained through wavelet transform scale parameter optimization; Distance matrix calculation is used to measure the similarity of time features, specifically, the time domain distance matrix is obtained through wavelet space Euclidean distance calculation; Kernel space clustering is used to obtain class labels based on time features, specifically, the clustering center is iteratively optimized through the kernel K-means clustering method to obtain class labels based on time features; The physical constraint spectral clustering is used to cluster structure features, specifically, the sample correlation is constrained by the physical threshold and spectral clustering is performed to obtain class labels based on structure features, including: Similarity matrix construction is used to avoid pseudo-related structure features, specifically, the physical threshold is set to control the relationship between samples to obtain a physical constraint similarity matrix; The Laplace matrix is constructed to model the building structure damage conduction path. The characteristic value and the characteristic vector are calculated, specifically, the characteristic value and the characteristic vector of the Laplacian matrix are obtained by performing the characteristic vector decomposition on the Laplacian matrix; The K-means clustering is performed to obtain the class label based on the structural feature, specifically, the characteristic values of the Laplacian matrix are sorted in ascending order, the characteristic vectors corresponding to the first n characteristic values are selected, the K-means clustering method is used for clustering to obtain the class label based on the structural feature; The feature pyramid collaborative clustering is used to fuse the multi-scale clustering results to eliminate the single-view bias, specifically, the final recognition result is obtained by probability fusion and consensus optimization, including: The multi-scale clustering probability is obtained to weight the credibility of each scale, specifically, the clustering probability of each scale is obtained by temperature scaling probability conversion; The initial consensus allocation matrix is used to balance the contribution of different scales, specifically, the initial consensus allocation matrix is obtained by dynamically weighting the multi-scale clustering probability through the contour coefficient; The consensus clustering optimization is performed, specifically, the final recognition result label is obtained by minimizing the divergence through Laplace regularization; The model is constructed and trained, specifically, the improved clustering model is constructed through the feature pyramid module, the dynamic time kernel K-means clustering, the physical constraint spectral clustering and the feature pyramid collaborative clustering, the model is trained based on the monitoring training set, the model performance is verified based on the monitoring test set, and the improved clustering model is obtained as the state recognition model.

5. The multi-scale feature based historic building digitalization monitoring protection system according to claim 1, wherein: In the building data acquisition module, the building past original data set and the building existing original data set both include building structure data, building environment data and building material performance data, and the building past original data set further includes building damage level annotation data.

6. The multi-scale feature based historic building digitalization monitoring protection system according to claim 1, wherein: In the original data optimization module, the data alignment is used to solve the problem of time and space asynchronization of sensors in historical building monitoring, specifically, the data on the time axis are strictly synchronized by synchronizing the clocks of each device and constructing a timestamp alignment matrix; The data cleaning is used to ensure the accuracy and integrity of the data, specifically, the cleaned data is obtained by removing impulse noise through median filtering and removing baseline drift through empirical mode decomposition; The data standardization is used to eliminate the dimension distortion caused by environmental factors, specifically, the standardized data is obtained by correcting displacement deformation data through temperature compensation and standardizing vibration energy through power, and adopting the Z-Score standardization method; The data set segmentation is used to segment the data set, specifically, the monitoring training set and the monitoring test set are obtained by segmenting the building past original data set; The building existing original data set is optimized through the data alignment, the data cleaning and the data standardization to obtain the monitoring data set, and the building past original data set is optimized through the data alignment, the data cleaning, the data standardization and the data set segmentation to obtain the monitoring training set and the monitoring test set.

Images