Building structure health monitoring method and system based on machine learning

Abstract

The invention relates to a building structure health monitoring method and system based on machine learning, and the method specifically comprises the following steps: firstly, building a target building three-dimensional numerical model through finite element simulation, generating a simulation signal, injecting Gaussian white noise, and adjusting model parameters to form a data set with health category labels; performing data enhancement by combining adaptive wavelet denoising with dynamic normalization, and extracting and enhancing high-resolution time-frequency features through adaptive window short-time Fourier transform and adaptive frequency band enhancement; then, a neural network model fusing structure physical prior guidance and multi-scale space-time interaction is constructed, a feature matrix is modulated, fused and coded to obtain a refined feature vector, and damage state probability distribution output is achieved; and training is carried out by using a feature consistency and prediction smoothness regularization term constraint model, and finally, the trained model is deployed, so that building structure health state evaluation and safety early warning are realized, the monitoring accuracy and reliability are improved, and effective technical support is provided for building safety guarantee.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of building health intelligent monitoring, and in particular to a building structure health monitoring method and system based on machine learning. BACKGROUND

[0002] With the increasingly large and complex of modern building structures, the safety and durability during long-term service have become a major issue of social concern. Structure health monitoring technology aims to realize early identification and safety warning of damage by real-time sensing and analyzing of structural response data, which is a key means to ensure the safety of major infrastructure throughout its life cycle. In recent years, data-driven methods represented by deep learning have brought new light to structure health monitoring. Such methods can automatically learn the complex mapping relationship between damage and response from massive monitoring data, and exhibit strong feature abstraction ability.

[0003] However, the prior art has the following disadvantages: the data preprocessing method is rigid, and the conventional minimum-maximum normalization or fixed threshold wavelet denoising cannot adaptively handle the contradiction between simulation noise and local feature preservation, which easily leads to feature distortion; the time-frequency feature extraction is limited by fixed window function or fixed scale transformation, the time-frequency resolution is insufficient, and there is a lack of focusing mechanism for key damage frequency bands, the feature discrimination is low; the mainstream neural network model is mostly a general architecture, which is difficult to incorporate physical prior knowledge in the field of structure health monitoring, and the modeling ability of multi-scale time evolution characteristics and cross-channel coupling relationship of damage response is limited; the model training is usually only oriented to classification accuracy, and lacks regularization constraints to guide the features to comply with the laws of structural mechanics, and the generalization ability and prediction smoothness are insufficient in the scene of class imbalance and continuous change of damage state.

[0004] Therefore, the present application proposes a building structure health monitoring method and system based on machine learning to solve the above problems. SUMMARY

[0005] The present application is aimed at the deficiencies of the prior art, and develops a building structure health monitoring method and system based on machine learning. The present application can improve monitoring accuracy and reliability, and provides effective technical support for building safety protection.

[0006] The technical solution of the present application to solve the technical problem is a building structure health monitoring method based on machine learning, comprising the following steps: S1, a three-dimensional numerical model of the target building is constructed by using finite element simulation technology, dynamic responses under different load conditions are simulated through the three-dimensional numerical model, simulation signals corresponding to real acquisition are generated, and building health monitoring data is constructed; S2. Inject Gaussian white noise into the simulation signal to simulate simulation noise, and adjust the parameters of the three-dimensional numerical model to generate simulation datasets of different categories. At the same time, label each sample with a health category to form a dataset. S3. Data augmentation is performed on the data in the dataset using a two-stage preprocessing method that combines adaptive wavelet denoising with dynamic normalization to generate augmented data. S4. A two-stage method combining adaptive window short-time Fourier transform and adaptive frequency band enhancement is adopted to extract high-resolution time-frequency features from the enhanced data and enhance the feature components related to structural damage to generate an enhanced time-frequency feature matrix. S5. By integrating a neural network model that combines structural physics prior guidance with multi-scale spatiotemporal interaction, a building structure health classification model is constructed. The enhanced time-frequency feature matrix is ​​subjected to structural physics prior guidance feature modulation. The modulation result is then subjected to multi-source feature fusion and external factor injection to generate a hybrid feature vector. Multi-scale spatiotemporal interactive encoding is then performed to obtain a refined feature vector. Finally, the probability distribution of the damage state is output after damage pattern classification. S6. The learning process of the building structure health classification model is constrained by the regularization term based on feature consistency and prediction smoothness. The total loss function of the model is calculated, and the model is trained based on the loss function. S7. Deploy the trained building structure health classification model into actual building structure health monitoring to assess the health status of building structures and provide safety warnings.

[0007] S1 is as follows: A three-dimensional numerical model is constructed based on the actual geometric dimensions, material properties, component connection methods, and boundary conditions of the target building structure. Different load conditions include environmental excitation, simulated wind load, and simulated traffic load; The simulated signal is the vibration acceleration time history signal corresponding to the actual acquisition situation. The actual acquisition situation is to deploy sensors on the target building structure and collect the actual vibration acceleration time history signal through the sensors.

[0008] S2 is as follows: Methods for adjusting the parameters of the three-dimensional numerical model include locally reducing the material stiffness, introducing microcrack elements, and changing the properties of the connectors. By adjusting the parameters, the structural damage states at different locations and with different degrees of severity are simulated, generating simulation datasets for different health categories, including healthy, minor damage, severe damage, and critical damage. Using data points of the simulated signal within a fixed length as a sample, each sample is labeled with a health category label according to the preset damage state during simulation.

[0009] S3 is as follows: S3.1. Perform stationary wavelet transform on the building structure health monitoring data in the dataset, and then dynamically calculate the adaptive threshold weight of each data point at each scale according to the difference between the local features of the signal and the statistical characteristics at each scale. After weighted reconstruction based on the adaptive threshold weight, the noise-reduced monitoring data is obtained. S3.2 Perform dynamic range normalization on the noise reduction monitoring data to obtain the enhanced data; Specifically, a scaling factor based on local statistical characteristics is used to dynamically adjust the global normalization results to obtain normalized monitoring data.

[0010] S4 is as follows: S4.1 The enhanced data is mapped from the time domain to the time-frequency domain by adaptive window short-time Fourier transform. The adaptive window short-time Fourier transform uses a window function with adaptive bandwidth to perform segmented processing on the time axis to generate a high-resolution initial time-frequency feature matrix. S4.2 Calculate the global average energy of each frequency compartment of the initial time-frequency feature matrix, and generate global frequency band weights to enhance high-energy frequency bands. At the same time, calculate local time-frequency attention weights to focus on time-frequency points with energy abrupt changes. Then, apply the two weights to the initial time-frequency feature matrix respectively, and add them together to obtain the enhanced time-frequency feature matrix.

[0011] S5 is detailed below: S5.1. Estimate the benchmark mode map that represents the typical distribution of time-frequency energy under the health state from the enhanced time-frequency feature matrix of all health state samples in the dataset, and combine the modulation intensity coefficient to perform physically guided attention modulation and bias enhancement on the enhanced time-frequency feature matrix to generate a physical perception feature matrix that integrates health benchmark prior knowledge. S5.2 Flatten the physical perception feature matrix into a vector, and extract a set of time-domain statistical and envelope spectrum features from the denoised time-domain signal to form a statistical feature vector. At the same time, introduce the feature vector of external environmental factors, and use the feature fusion gating unit to adaptively weight and fuse the three parts of information to obtain a hybrid feature vector. S5.3. Employ a multi-scale spatiotemporal interaction block, extract receptive field features at different times through parallel multi-dilation rate temporal dilated convolution, and interactively modulate the relationship between feature channels with adaptive feature channel interaction modeling. After stacking multiple of this module, a refined feature vector is output. The input to the first multi-scale spatiotemporal interaction block is a mixed feature vector, and the output of the previous multi-scale spatiotemporal interaction block is used as the input to the next multi-scale spatiotemporal interaction block. S5.4 The refined feature vector is mapped to unnormalized scores corresponding to different structural health states through a fully connected layer, and the final classification probability distribution is obtained by using a Softmax function that integrates adaptive temperature scaling and class prior weights.

[0012] The calculation process for the total loss function is as follows: (1) A loss term combining adaptive marginal cross-entropy and class center constraint is adopted, which simultaneously narrows the distance between sample features and their class centers and widens the distance between them and other class centers, and adaptively adjusts the marginal value according to the difficulty of class distinction; (2) Based on the prior knowledge of the continuity of damage evolution and its consistency in time-frequency response, a regularization loss term containing time-frequency feature consistency constraints and prediction distribution smoothness constraints is adopted to encourage similar features of samples of the same type and smooth changes in prediction of continuous or neighboring states, expressed as: (3) The total loss is the weighted sum of the adaptive marginal contrast center loss and the regularization loss.

[0013] The training process for the building structural health classification model is as follows: The dataset is divided into a training dataset and a validation dataset. The data in the training dataset is processed by S3 and S4, and then the model is trained end-to-end. In each training iteration, a batch of sample data is passed through the forward propagation of the model to obtain the predicted probability distribution. Then, the loss value of the current batch is calculated based on the total loss function. An adaptive moment estimation optimizer is used to calculate the gradient of all trainable parameters in the model based on the total loss function, and these parameters are updated through the backpropagation algorithm to minimize the total loss function. During training, model performance is evaluated periodically using a separate validation dataset; The training iterations of the model will continue until the performance metrics on the validation set no longer improve over multiple consecutive training cycles, or until the preset maximum number of iteration cycles is reached. At this point, training stops, the model parameters with the best validation performance are saved, and the training of the building structure health classification model is completed.

[0014] S7 is detailed below: After the building structure health classification model is trained, it is deployed in actual building structure health monitoring. The original vibration signals are collected in real time or periodically by vibration sensors installed on key parts of the building structure to form the vibration data to be monitored. At the same time, ambient temperature and humidity data are collected to form the feature vector of external environmental factors. Then, after the collected data is processed, it is input into the pre-trained building structure health classification model. The model outputs the probability that the building structure belongs to each health category during the monitoring period; the category corresponding to the highest probability is determined as the current health status of the structure, and the monitoring results, probability distribution and key features are visualized; then, different levels of early warning are automatically triggered according to the current health status of the structure.

[0015] This invention also provides a machine learning-based building structure health monitoring system, which implements a machine learning-based building structure health monitoring method. The system structure is as follows: 3D Numerical Model and Simulation Dataset Generation Module: Used to construct a 3D numerical model of the target building using finite element simulation technology, simulate the dynamic response under different load conditions to generate simulation signals, inject Gaussian white noise and adjust model parameters to generate multi-class simulation datasets, and label the health categories of samples. Data preprocessing and enhancement module: A two-stage method combining adaptive wavelet denoising and dynamic normalization is used to process the simulation dataset and generate enhanced data; Time-frequency feature extraction and enhancement module: Through adaptive window short-time Fourier transform combined with adaptive frequency band enhancement, high-resolution time-frequency features are extracted from the enhanced data, and damage-related feature components are enhanced to generate an enhanced time-frequency feature matrix; Health classification model construction and training module: Construct a neural network model that integrates structured physical prior guidance and multi-scale spatiotemporal interaction, perform feature modulation, multi-source fusion and external factor injection on the enhanced time-frequency feature matrix, obtain refined feature vectors through multi-scale spatiotemporal interaction encoding, construct the total loss function based on feature consistency and prediction smoothness regularization term and train the model; Health monitoring and early warning module: Deploy a pre-trained model to assess the health status of building structures, output damage probability distribution, and provide safety early warning.

[0016] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention discloses a machine learning-based building structure health monitoring system and method. It employs a two-stage preprocessing method combining adaptive wavelet denoising and dynamic normalization, dynamically adjusting processing parameters based on local signal features to suppress noise while preserving damage-related local fluctuation features. A two-stage time-frequency feature extraction strategy using adaptive window short-time Fourier transform and adaptive frequency band enhancement generates high-resolution, high-discrimination time-frequency representations through time-varying window functions and a dual-weighting mechanism. A neural network model integrating structural physics prior guidance and multi-scale spatiotemporal interaction is used, achieving refined perception of damage patterns through health benchmark modal spectrum modulation, multi-source feature fusion, and learnable spatiotemporal relationship modeling. A composite loss function combining adaptive marginal contrast center loss and physical regularization guides the model to learn feature representations with high intra-class consistency, inter-class separability, and conformity to the continuity of damage evolution. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0019] Figure 2 This is an example diagram for simulation modeling using refined finite element simulation technology.

[0020] Figure 3 This is an example diagram of a vibration acceleration time history signal.

[0021] Figure 4 This is a schematic diagram of step S4.

[0022] Figure 5 A comparison chart showing the classification accuracy of different classification methods across four injury states. Detailed Implementation

[0023] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0024] Example 1 like Figure 1 As shown, a machine learning-based method for monitoring the health of building structures includes the following steps: S1. A three-dimensional numerical model of the target building is constructed using finite element simulation technology. The dynamic response under different load conditions is simulated through the three-dimensional numerical model, generating simulation signals corresponding to the actual data collected, thus forming building health monitoring data. S2. Inject Gaussian white noise into the simulation signal to simulate simulation noise, and adjust the parameters of the three-dimensional numerical model to generate simulation datasets of different categories. At the same time, label each sample with a health category to form a dataset. S3. Data augmentation is performed on the data in the dataset using a two-stage preprocessing method that combines adaptive wavelet denoising with dynamic normalization to generate augmented data. S4. A two-stage method combining adaptive window short-time Fourier transform and adaptive frequency band enhancement is adopted to extract high-resolution time-frequency features from the enhanced data and enhance the feature components related to structural damage to generate an enhanced time-frequency feature matrix. S5. By integrating a neural network model that combines structural physics prior guidance with multi-scale spatiotemporal interaction, a building structure health classification model is constructed. The enhanced time-frequency feature matrix is ​​subjected to structural physics prior guidance feature modulation. The modulation result is then subjected to multi-source feature fusion and external factor injection to generate a hybrid feature vector. Multi-scale spatiotemporal interactive encoding is then performed to obtain a refined feature vector. Finally, the probability distribution of the damage state is output after damage pattern classification. S6. The learning process of the building structure health classification model is constrained by the regularization term based on feature consistency and prediction smoothness. The total loss function of the model is calculated, and the model is trained based on the loss function. S7. Deploy the trained building structure health classification model into actual building structure health monitoring to assess the health status of building structures and provide safety warnings.

[0025] In a specific implementation, S1 is as follows: Since the implementation of the system and method of this invention relies on a large amount of vibration monitoring data with clear health status labels, in order to obtain such data, a refined finite element simulation technology is used to construct an accurate three-dimensional numerical model of the target building structure in a computer. This model is established strictly according to the geometric dimensions, material properties, component connection methods and boundary conditions of the actual structure; among which, material properties include the elastic modulus, density, Poisson's ratio, etc. of concrete. like Figure 2 The image shown is an example of simulation modeling using refined finite element simulation technology.

[0026] By simulating the dynamic response of the structure under various load conditions, vibration acceleration time history signals corresponding to those collected by real sensors are generated, which are the building structure health monitoring data.

[0027] like Figure 3 The image shown is an example of the vibration acceleration time history signal of the sample.

[0028] In a specific implementation, S2 is as follows: To closely resemble the actual monitoring environment, Gaussian white noise with specific statistical characteristics is actively injected into the simulation signal to simulate simulation noise. By systematically changing key parameters in the finite element model, structural damage states at different locations and with varying degrees of severity are simulated, thereby generating simulation datasets covering four categories: healthy, minor damage, moderate damage, and severe damage.

[0029] The vibration signal length of each sample is fixed at 8192 data points, and the labels are directly marked according to the damage conditions preset during simulation. The label categories strictly correspond to the four types of structural health states mentioned above.

[0030] Based on this, large-scale, precisely labeled training and validation datasets are generated, providing a foundation for subsequent model construction and training.

[0031] In a specific implementation, S3 is as follows: Building structural health monitoring data is obtained through refined finite element simulation. It is characterized by high dimensionality, simulation noise, and large amplitude scale differences. Conventional preprocessing methods such as minimum-maximum normalization or fixed threshold wavelet denoising are usually unable to adaptively remove simulation noise and ignore local fluctuation features, which can easily lead to distortion in subsequent feature extraction.

[0032] This invention employs a two-stage preprocessing workflow combining adaptive wavelet denoising and dynamic normalization to enhance data robustness and preserve key local features. The specific steps are as follows: S3.1. Perform stationary wavelet transform on the building structure health monitoring data. Then, based on the differences in local signal features and statistical characteristics at each scale, dynamically calculate the adaptive threshold weight for each data point at each scale. After weighted reconstruction based on the adaptive threshold weight, the noise-reduced monitoring data is obtained, as shown below: In the formula, This represents building structural health monitoring data, with the following dimensions: It is a vibration signal obtained from refined finite element simulation, which includes simulation noise; This represents the time point index, identifying the sampling location of the vibration signal in the time domain, with a value range of [value range missing]. ; Representing scale Up to the time point The detail coefficients are obtained through a stationary wavelet transform algorithm, specifically using the Daubechies wavelet basis. The detail coefficient matrix at each scale is obtained through this algorithm, with dimensions of [missing information]. ; Representing scale Up to the time point The approximate coefficient values ​​are obtained through a stationary wavelet transform algorithm, with dimensions of [missing value]. The coarsest scale is used to obtain the low-frequency profile of the signal, characterizing the global trend and low-frequency components of the signal; Indicates building structural health monitoring data In the The vibration amplitude at each time point represents the sampled value of the signal in the time domain; Indicates noise reduction monitoring data In the The vibration amplitude at each time point is a clean signal obtained after adaptive wavelet denoising processing; This represents the total number of wavelet decomposition scales, used to control the depth of multi-scale analysis. Its value is determined by the length of the vibration signal. Calculated as ,in This is a round-down operation; The scale index is a positive integer with a value range of 1 to 2. ; Indicates the first Scale, First An adaptive threshold weight for each data point is used to dynamically adjust the contribution of detail coefficients at each scale to the reconstruction result based on local signal characteristics. This allows for the preservation of important fluctuation features while reducing noise. The calculation method is expressed as follows: ; Indicates the first The sensitivity coefficient of the scale is used to control how sensitive the scale weights are to signal differences, thereby adjusting the noise reduction intensity of that scale. It is set according to the scale noise level, for example, by defining... To make smaller scale (with more high-frequency noise) The larger the noise reduction weight, the smaller the noise reduction weight, thus enhancing the noise reduction effect; Indicates the first The mean absolute value of the scale detail coefficients represents the central tendency of the signal at that scale, and is calculated as follows: ; This represents the natural exponential function.

[0033] It should be noted that, through stationary wavelet transform, building structural health monitoring data can be... Decomposing it into approximation coefficients and detail coefficients allows us to obtain the scale. Up to the time point Detail coefficient value and scale Up to the time point Approximate coefficient values .

[0034] S3.2 To address the issue of large amplitude scale differences, the normalization interval for the noise reduction monitoring data is not fixed at [0,1]. Instead, a scaling factor based on local statistical characteristics is used to dynamically adjust the global normalization result to enhance the retention of local fluctuations, as expressed below: In the formula, Indicates normalized monitoring data In the The vibration amplitude at each time point, with dimension [missing information]. It represents the vibration signal after noise removal and normalization, where the global trend is standardized and local anomalies are enhanced; This represents the function that takes the minimum value. The item is used to calculate noise reduction monitoring data. The minimum value of all elements in the range is used to determine the lower limit of the global magnitude; This represents the function that takes the maximum value. The item is used to calculate noise reduction monitoring data. The maximum value of all elements in the range is used to determine the global amplitude limit; Indicates the first The scaling factor at each time point is used to dynamically adjust the normalization amplitude based on local data fluctuations, in order to preserve local features while addressing large differences in amplitude scale. The calculation method is expressed as follows: In the actual implementation, symmetrical expansion filling is used at the boundary; This represents the local sensitivity parameter, used to control the degree to which local fluctuations affect the scaling factor, thereby adjusting the intensity of local enhancement. An example value is 0.05. This represents the local window size, defining the neighborhood range for calculating local statistics. Examples of possible values ​​are provided. ; This represents the position index within a local window, with a value range of 1. ; This indicates that the noise reduction monitoring data is in the first... The vibration amplitude at each time point.

[0035] It should be noted that, Item calculation with the first After a certain point in time The local mean calculated from each data point represents the average behavior of the signal within a local window. It is used to quantify the deviation between the current data point and the local average behavior, thereby dynamically adjusting the scaling factor during the normalization process to enhance the preservation of local fluctuation characteristics.

[0036] In specific implementation methods, such as Figure 4 As shown, S4 is as follows: Vibration data contains frequency and time domain features of structural health status. Conventional short-time Fourier transform or fixed-scale wavelet transform has inherent limitations in terms of insufficient time and frequency resolution and cannot adaptively focus on key frequency bands related to structural damage, resulting in low feature discrimination and affecting the performance of subsequent classification models.

[0037] This invention employs a two-stage method combining adaptive window short-time Fourier transform and adaptive frequency band enhancement to extract high-resolution time-frequency features from preprocessed vibration data and enhance feature components related to structural damage. The specific steps are as follows: S4.1. The normalized monitoring data is mapped from the time domain to the time-frequency domain using the adaptive window short-time Fourier transform. The adaptive window short-time Fourier transform uses a window function with adaptive bandwidth to perform piecewise processing on the time axis to generate a high-resolution initial time-frequency feature matrix, expressed as: In the formula, Represents the initial time-frequency characteristic matrix The Middle The first time frame, the first The value of each frequency bin, with dimension . It characterizes the energy distribution of the signal in the time-frequency domain, reflecting the frequency band and time-varying modes of damage-related vibrations in structures; Represents the time frame index, with a value range of 100. ; The total number of time frames is represented by the window length. and step length Calculation, expressed as For example, if , ,but ; This represents the frequency bin index, with a value range of [value range missing]. ; This indicates the total number of frequency bins, used to control the frequency resolution of the time-frequency representation; an example value is 128. Indicates the first The first time frame, the first Each frequency compartment corresponds to an adaptive Gaussian window function, whose center position and bandwidth vary with time and frequency to achieve high-resolution time-frequency analysis, denoted as: ; Represents the imaginary unit, satisfying ; Indicates the first The normalized frequency value of each frequency compartment is used to identify the frequency domain location, ranging from... Between, the calculation method is expressed as ; Indicates the first The position of the center point of each time frame in the original time series is used to locate the center of the window function, and is determined by the time segmentation parameter. Indicates the first The window function bandwidth parameter corresponding to each frequency band is used to control the bandwidth of the window function. The higher the frequency, the narrower the bandwidth, to adapt to the rapid changes in the signal in the high-frequency region. The calculation method is expressed as follows: ; This represents a constant to prevent the denominator from being zero; it is a minimal constant, and examples of its values ​​are given below. .

[0038] In the specific implementation process, Determined by the method of time segmentation, assuming the time series is divided into... Segments, each segment has a window length of [length missing]. ,but It can be defined as the first The time index of the segment center point, for example, if the segments are non-overlapping and uniform, then define... , where window length If segments overlap, adjustments need to be made based on the overlap rate, such as the window length. overlap rate Then define .

[0039] It should be noted that, The term is a Fourier kernel function used to transform a windowed time-domain signal into the frequency domain. In order to capture both time and frequency information in time-frequency analysis, it allows the analysis of the frequency domain characteristics of the signal within each time frame, thereby generating a high-resolution time-frequency representation.

[0040] S4.2 Calculate the global average energy of each frequency compartment in the initial time-frequency feature matrix, and generate global frequency band weights to enhance high-energy frequency bands. Simultaneously, calculate local time-frequency attention weights to focus on time-frequency points with energy abrupt changes. Then, apply both weights to the initial time-frequency feature matrix and sum them to obtain the enhanced time-frequency feature matrix, expressed as: In the formula, Represents the enhanced time-frequency feature matrix The Middle The first time frame, the first The value of each frequency bin, with dimension [missing information]. This characterizes the adaptively enhanced time-frequency energy distribution; Indicates the first The global frequency band weight of each frequency compartment is used to adaptively enhance key frequency bands that may be related to structural damage based on the average energy of that frequency over the entire time period. The calculation method is expressed as follows: ; Indicates the first The first time frame, the first The local time-frequency attention weights for each frequency bin are used to highlight higher-energy frequency components within the same time frame, improving the local discriminative power of features. The calculation method is expressed as follows: . Indicates the first The global average energy of each frequency compartment is derived from the initial time-frequency characteristic matrix. In the The average of the absolute values ​​of all time frames across a frequency cell is used to represent the average intensity of that frequency component over the entire time range. The calculation method is expressed as follows: ; The slope parameter of the Sigmoid function is used to control the steepness of the weight changes, thereby adjusting the smoothness of the weight transition. The example value is 5. This represents the energy threshold parameter, used to determine whether a frequency band is significant, thereby filtering important frequency bands. An example value is 0.1. This represents the attention temperature parameter, which controls the concentration of the weight distribution, thereby adjusting the intensity of attention. An example value is 0.5.

[0041] In a specific implementation, S5 is as follows: The enhanced time-frequency feature matrix contains key time-frequency information related to structural damage. The damage patterns of building structures are closely related to their physical properties, and damage evolution exhibits spatiotemporal correlation. Conventional convolutional neural networks or recurrent neural networks struggle to incorporate prior knowledge of structural physics and have limited ability to model cross-scale spatiotemporal feature interactions. This invention constructs a neural network model that integrates prior knowledge of structural physics with multi-scale spatiotemporal interactions to achieve high-precision and robust classification of the health status of building structures. The specific steps are as follows: S5.1. Using the enhanced time-frequency feature matrix of all healthy state samples in the training set, estimate the baseline mode map representing the typical distribution of time-frequency energy under healthy states. Combine this with the modulation intensity coefficient to perform physically guided attention modulation and bias enhancement on the enhanced time-frequency feature matrix, generating a physical perception feature matrix that integrates prior knowledge of the healthy baseline, expressed as: In the formula, Represents the physical perception feature matrix The Middle The first time frame, the first The value of each frequency bin, with dimension [missing information]. It integrates original time-frequency energy with prior physical knowledge to enhance the damage-related frequency band while suppressing the health baseline signal; The modulation intensity coefficient is a trainable scalar parameter used to control the strength of the physical prior's guidance on the attention weights. Indicates the first The first time frame, the first The modulation coefficients of each frequency cell are used to quantify the deviation of the current feature point from the health baseline, and are calculated as follows: ; The bias amplification factor is a trainable scalar parameter used to amplify significant deviations from a healthy baseline. This represents the hyperbolic tangent activation function, which maps the input to the interval [0, 1]. ; In the reference modal spectrum, the first The first time frame, the first The values ​​of each frequency bin are obtained by representing the initial time-frequency feature matrix of all healthy samples in the training set. The average along the upper sample dimension is used to characterize the typical distribution of time-frequency energy in a healthy state; The reference modal spectrum is shown in the first... The standard deviation of the frequency dimension of each time frame is used to normalize the bias, and is calculated as follows: ; express The mean; The reference modal spectrum is shown in the first... On each frequency module, all time frames ( from arrive A vector composed of the values ​​of ). The reference modal spectrum is shown in the first... All frequency modules in each time frame ( from arrive A vector composed of the values ​​of ). This indicates the operation of calculating variance; This represents a very small constant to prevent the denominator from being zero. Examples of its values ​​are: .

[0042] S5.2 To comprehensively characterize the structural state, the physical sensing feature matrix is ​​flattened into a vector, and a set of time-domain statistical and envelope spectrum features are extracted from the denoised time-domain signal to form a statistical feature vector. Simultaneously, an external environmental factor feature vector is introduced. Through a feature fusion gating unit, the three parts of information are adaptively weighted and fused to obtain a hybrid feature vector, expressed as: In the formula, Represents a mixed feature vector with dimension . It adaptively integrates time-frequency, time-domain statistics, and multi-source information from the external environment; This represents the fusion weight matrix, with dimension 1. , is a trainable parameter matrix that achieves information compression and fusion by projecting it onto a dimension; The dimension of the mixed feature vector; The dimension of the statistical feature vector; The dimension representing the feature vector of external environmental factors; This represents a vector concatenation operation; This represents the operation of flattening a matrix into a vector. The implementation will Flattened to a length of ; Represents a statistical feature vector with dimension . From noise reduction monitoring data Extracted time-domain indicators and monitoring data based on normalization The envelope spectrum features are composed of time-domain indices including mean, standard deviation, skewness, kurtosis, and root mean square. This represents the feature vector of external environmental factors, with dimension 1. Example of a value: Including, for example, ambient temperature and ambient humidity Information is obtained by simulating environmental factors through sensor data collection. This represents the element-wise multiplication operation; This represents the Sigmoid activation function, with an output range of 1 / 2. , used to generate adaptive gating weights; This represents the gate weight matrix, with dimension 1. , is a trainable parameter matrix used to construct a trainable gating mechanism; Represents the gated bias vector, with dimension . , is a trainable parameter vector.

[0043] In the specific implementation process, the characteristic vector of external environmental factors Including ambient temperature and ambient humidity Two elemental properties: temperature changes will cause changes in the elastic modulus of a material, leading to changes in the overall stiffness of the structure, which in turn affects its natural frequency; humidity may affect the quality of concrete structures (water absorption) or the stiffness of wood structures, thereby changing their dynamic response.

[0044] From noise reduction monitoring data The extracted time-domain indicators include mean (the average amplitude of the signal, reflecting the static offset or DC component of the vibration), standard deviation (the intensity of signal fluctuations around the mean, reflecting vibration energy), skewness (the asymmetry of the signal amplitude distribution; positive skewness indicates a large number of large-amplitude pulses, possibly caused by impact damage), kurtosis (the sharpness of the signal amplitude distribution; high kurtosis indicates a significant impact component in the signal, strongly correlated with bearing failure, crack development, etc.), and root mean square (the effective amplitude of the signal); envelope spectrum characteristics are obtained by normalizing monitoring data. The envelope signal is obtained by performing Hilbert transform, and then spectrum analysis is performed on it. The envelope spectrum can highlight the periodic impact characteristics of the signal (such as the characteristic frequency of bearing failure, the sideband of gear meshing frequency, etc.).

[0045] It should be noted that the physical perception feature matrix It primarily captures the frequency and time-varying patterns of signals, but may lose or weaken some intuitive time-domain statistical characteristics, while statistical feature vectors... From noise reduction monitoring data and normalized monitoring data Classical statistical features and envelope spectrum features are extracted to effectively supplement time-frequency features, providing the model with information from different perspectives and more physically interpretable information, thereby improving the model's robustness.

[0046] It should also be noted that the dynamic characteristics of building structures (such as frequency and damping) are significantly affected by environmental factors (temperature and humidity). Ignoring these factors may cause the model to misjudge normal performance fluctuations caused by environmental changes as damage. Therefore, we introduce feature vectors of external environmental factors. The aim is to enable the model to sense and “eliminate” environmental interference, focusing on changes caused by real damage, thereby improving the reliability of the monitoring system.

[0047] S5.3 To capture the complex multi-scale dependencies that may exist between damage features and feature channels over time, a multi-scale spatiotemporal interaction block is adopted. This block extracts features from different temporal receptive fields through parallel multi-dilation rate temporal dilated convolutions, and interactively modulates the relationships between feature channels modeled by adaptive feature channel interaction. Stacking multiple such modules outputs a refined feature vector. The interaction process of a single module is represented as follows: In the formula, Indicates input to the first The feature vectors of a multi-scale spatiotemporal interaction block are defined. For the first multi-scale spatiotemporal interaction block, the following definition is provided. ; Indicates input to the first The feature vector of the multi-scale spatiotemporal interaction block, that is, the first... The feature vectors output by the multi-scale spatiotemporal interaction blocks are deeply integrated with the multi-scale temporal information and the correlation information between feature channels to obtain higher-level abstract features that better reflect the spatiotemporal evolution of damage. Presentation layer normalization operation; This represents the dilation rate of dilated convolution, belonging to a predefined set of dilation rates. ,For example This is used to control the size of the receptive field of the convolutional kernel in the time dimension; This represents the modified linear unit activation function, used to introduce nonlinearity, enabling the model to learn and represent more complex functional relationships; This indicates the corresponding expansion rate. The temporal convolution kernel parameters have a dimension of . , are trainable parameters; This represents the kernel size, with examples of possible values. ; Indicates the expansion rate One-dimensional dilated convolution operations allow networks to efficiently capture long-range temporal dependencies; This represents the element-wise multiplication operation; This represents the activation function used for spatial attention weights, specifically the Softplus function; This represents the graph adjacency matrix with dimension . , used to model the spatial relationship between feature channels, are trainable parameters; It represents the interaction operation of feature channels. It models the dependencies between feature channels through a learnable adjacency matrix, so that the feature integrates information from other channels on each channel, thereby simulating the propagation and coupling effect of damage among multiple physical quantities.

[0048] In the specific implementation process, the definition The item is simplified as The number of time steps is For dimension tensor First, it is reshaped into dimensions. A two-dimensional matrix, then with a trainable graph adjacency matrix. Performing matrix multiplication yields the dimension. The result is that it is eventually reshaped back to the original dimension. , represented as ,in, Reshaping into dimensions is tensor, Reshaping into dimensions is The matrix.

[0049] Stacking After multiple multi-scale spatiotemporal interaction blocks, a refined feature vector is obtained. ,in This represents the total number of layers in the multi-scale spatiotemporal interaction block stack; an example value is 4.

[0050] It should be noted that conventional temporal networks either handle time dependencies or feature channel relationships separately. This invention modulates multi-scale temporal convolutions with adaptive feature channel interactions, using dilated convolutions with different dilation rates in parallel to capture different scale patterns of the damage response that may exist in time (such as instantaneous impact and slowly varying drift), and utilizes a learnable graph adjacency matrix. This study models the correlation strength between different feature channels (potentially representing different frequency bands or physical quantities), simulates the propagation and coupling effects of damage among multiple physical quantities, and utilizes the "spatial relationship" information after feature channel interaction processing to weight multi-scale temporal features. This guides the extraction process of temporal features with the semantic relationships between features, simulating the cognitive process of "analyzing time-varying patterns with emphasis based on the correlation between features." (Graph adjacency matrix) As a trainable matrix, its first... Line number Column elements It can be understood as the first The first feature channel and the first The model automatically learns the correlation strength between feature channels through training data. For example, it may learn that there is a strong correlation between "energy anomaly in a certain high-frequency band" and "phase change in a certain low-frequency band", which may correspond to a certain damage pattern, thereby strengthening this cross-channel collaborative perception during feature extraction.

[0051] S5.4. The refined feature vector is mapped to unnormalized scores corresponding to different structural health states through a fully connected layer, and the final classification probability distribution is obtained by using a Softmax function that integrates adaptive temperature scaling and class prior weights, as follows: In the formula, This indicates that the model predicts the sample belongs to the first... The probability of a damage state, with a range of The sum of the probabilities of all categories is ; This represents the category index, which is a positive integer with a value range of 1 to 2. ; Indicates corresponding to the first The class's output weight vector has dimensions of , are trainable parameters used to linearly map refined feature vectors to scalar scores; Indicates corresponding to the first The output bias scalar of the class is a trainable parameter; Indicates the first The adaptive temperature parameter of a class is a trainable positive scalar parameter used to adjust the smoothness of the output probability distribution of that class. Indicates the first Prior weights of classes are used to mitigate class imbalance problems, and are calculated as follows: ; This represents a smoothing factor used to smooth prior weights. The calculation, when hour, The penalty is directly proportional to the inverse of the category frequency, and it dominates when... At that time, all No category prior, take It can strike a balance between mitigating class imbalance and preventing excessive distortion of the prediction distribution, with an example value of 0.8; The total number of categories representing the structural health status, specifically including: healthy, minor injury, moderate injury, and severe injury, for a total of four categories; This represents the total number of samples in the training set. Indicates the first training set The number of class samples.

[0052] In a specific implementation, S6 is as follows: S6.1 The total loss function is calculated as follows: To effectively train neural network models and guide them to learn robust feature representations that conform to the principles of structural mechanics, a regularization term based on feature consistency and prediction smoothness is used to constrain the model learning process during loss function calculation. This encourages accurate classification while achieving consistency in feature representations and smoothness in prediction. The specific steps are as follows: (1) To enhance the intra-class compactness and inter-class discriminativeness of the learned features, a loss term combining adaptive marginal cross-entropy and class center constraints is adopted. This term simultaneously narrows the distance between sample features and their respective class centers while widening the distance between them and other class centers. The marginal value is adaptively adjusted according to the difficulty of class discrimination, as shown below: , In the formula, This represents the adaptive margin contrast center loss, which combines the angular margin classification loss with the center loss and introduces a learnable adaptive margin. This represents the sample index, which is a positive integer. Indicates the training batch size; This represents a logarithmic function, with the default base being the natural constant. Indicates the first Index of the true class of each sample; Represent the natural logarithm function; Represents the natural exponential function; Represents the cosine function; This represents the feature scaling factor, a fixed scalar parameter used to amplify differences in the cosine values ​​of the angle. Examples of possible values ​​are provided. ; Indicates the first Refined feature vectors of each sample With category Corresponding class center vector The angle between them; Indicates category The adaptive margin is a trainable non-negative scalar parameter. As a trainable parameter, it means that the model can automatically learn a larger margin for difficult-to-distinguish categories and a smaller margin for easily distinguishable categories, making the optimization process more refined. Indices representing non-real categories, used for traversing categories. All categories other than; Indicates the first Refined feature vectors of each sample With category Corresponding class center vector The angle between them; This represents the weighting coefficient of the center loss term, used to balance the cross-entropy loss and the center loss. An example of its value is shown below. ; Indicates the first The refined feature vector of each sample; Indicates category The class center vector is a trainable parameter vector with dimension equal to the first... Refined feature vectors of each sample same; This represents the L2 norm.

[0053] It should be noted that, through The requirement is that not only is the cosine value of the true category greater than that of other categories, but it is also required to be significantly larger by a margin. This forces clearer boundaries between classes, and, through Weighted central loss term This brings the characteristics of similar samples closer together, making the intra-class distribution more compact.

[0054] (2) Based on the prior knowledge of the continuity of damage evolution and its consistency in time-frequency response, a regularization loss term containing time-frequency feature consistency constraints and prediction distribution smoothness constraints is adopted to encourage similar features of samples of the same type and smooth changes in prediction of continuous or neighboring states, expressed as: In the formula, This represents the regularization loss, which serves as the overall objective function to be minimized during model training, taking into account the requirements of classification accuracy and the physical rationality and smoothness of model predictions / features. The weight coefficients for feature consistency regularization are shown in the example below. This is used to control the relative weight of the feature consistency regularization term in the regularization loss; The weight coefficients represent the prediction smoothness regularization, with examples of possible values. This is used to control the relative weight of the prediction smoothness regularization term in the regularization loss; This represents the set of all true class labels in a batch of training. For example, if a batch has two classes, "healthy" and "minor injury", then... ; Represents a set The number of unique categories appearing in the data; Indicates the category label index; Indicates that the current batch belongs to the category The number of samples; Indicates the relationship with the first All samples belong to the same category. The index of another sample; Indicates belonging to a category The The physical perception feature matrix of each sample; Indicates belonging to a category The The physical perception feature matrix of each sample; This represents the total number of steps / points in a continuous time series or a gradual change sequence of operating conditions. Indicates the time step or sequence point index in the sequence; Indicates the model in the sequence number 1 The probability distribution vector of the class predicted at each step (or point); Indicates the model in the sequence number 1 The probability distribution vector of the class predicted at each step (or point); This represents the Kullback-Leibler divergence, used to measure the difference between two probability distributions; This represents the Frobenius norm.

[0055] (3) The total loss of model training is the weighted sum of the adaptive marginal contrast center loss and the regularization loss. By optimizing this total loss, the model is guided to learn robust features with high discriminative power and in accordance with physical laws, which is expressed as: In the formula, The total loss function for model training represents the overall optimization objective of model training. This represents the weighting coefficient of the regularization loss, used to control the importance of the regularization term in the total loss. An example value is 0.1.

[0056] S6.2 The model training process is as follows: After completing the construction of the building structure health classification model, the model is trained end-to-end using the training dataset generated by the refined finite element simulation in step S1 and preprocessed and feature extracted in steps S2 and S3.

[0057] The training process begins by calculating a baseline modality map using the enhanced time-frequency feature matrix of all healthy state samples in the training set. This map is solidified as physical prior knowledge and used for feature modulation in step S401 of all subsequent training batches. In each training iteration, the sample data of a batch are sequentially passed through the forward propagation of the model to obtain the predicted probability distribution. Then, the loss value of the current batch is calculated according to the total loss function defined in step S402, which includes adaptive marginal contrast center loss and regularization terms.

[0058] An adaptive moment estimation optimizer is used to calculate the gradient of all trainable parameters in the model based on the loss function, and these parameters are updated through the backpropagation algorithm to minimize the total loss function.

[0059] During training, the model performance is periodically evaluated using an independent validation set, monitoring metrics such as classification accuracy on unseen data. Training iterations continue until performance metrics on the validation set no longer improve over multiple consecutive training epochs, or until the preset maximum number of iterations is reached. At this point, training stops, the model parameters with the best validation performance are saved, and training of the building structure health classification model is complete.

[0060] In a specific implementation, S7 is as follows: Once the building structure health classification model is trained, it can be deployed for actual building structure health monitoring. During monitoring, the system collects raw vibration signals in real time or periodically through vibration sensors installed on key parts of the building structure, forming the vibration data to be monitored. This data first flows into a preprocessing and feature extraction pipeline that is identical to that used in the training phase. It then undergoes adaptive wavelet denoising and dynamic range normalization in step S2, and adaptive windowed short-time Fourier transform and adaptive time-frequency feature enhancement in step S3, transforming it into an enhanced time-frequency feature matrix. Simultaneously, the system collects ambient temperature and humidity data to form an external environmental factor feature vector, and extracts statistical feature vectors from the preprocessed time-domain signal.

[0061] Then, the processed features are input into the pre-trained building structure health classification model. The model performs forward inference, and after a series of internal calculations such as feature modulation guided by structural physics priors, multi-source feature fusion, and multi-scale spatiotemporal interactive coding, it finally outputs the probability that the building structure belongs to each of the following categories: "healthy", "minor damage", "moderate damage" and "severe damage" during the monitoring period.

[0062] The system determines the current structural health status based on the category corresponding to the highest probability and visualizes the monitoring results, probability distribution, and key characteristics to the user. Once the system determines that a damage state has occurred (minor, moderate, or severe), it will automatically trigger different levels of warnings, thereby achieving automated and intelligent assessment of building structural health status and safety early warning.

[0063] Example 2 A machine learning-based building structure health monitoring system implements a machine learning-based building structure health monitoring method. The system structure is as follows: 3D Numerical Model and Simulation Dataset Generation Module: Used to construct a 3D numerical model of the target building using finite element simulation technology, simulate the dynamic response under different load conditions to generate simulation signals, inject Gaussian white noise and adjust model parameters to generate multi-class simulation datasets, and label the health categories of samples. Data preprocessing and enhancement module: A two-stage method combining adaptive wavelet denoising and dynamic normalization is used to process the simulation dataset and generate enhanced data; Time-frequency feature extraction and enhancement module: Through adaptive window short-time Fourier transform combined with adaptive frequency band enhancement, high-resolution time-frequency features are extracted from the enhanced data, and damage-related feature components are enhanced to generate an enhanced time-frequency feature matrix; Health classification model construction and training module: Construct a neural network model that integrates structured physical prior guidance and multi-scale spatiotemporal interaction, perform feature modulation, multi-source fusion and external factor injection on the enhanced time-frequency feature matrix, obtain refined feature vectors through multi-scale spatiotemporal interaction encoding, construct the total loss function based on feature consistency and prediction smoothness regularization term and train the model; Health monitoring and early warning module: Deploy a pre-trained model to assess the health status of building structures, output damage probability distribution, and provide safety early warning.

[0064] Example 3 like Figure 5 As shown, the classification accuracy of different classification methods on four damage states is compared to evaluate the classification performance of the neural network model constructed in this invention, which integrates physical prior guidance and multi-scale spatiotemporal interaction. Conventional techniques compared include traditional support vector machines based on handcrafted features, convolutional neural networks using only time-frequency images, and long short-term memory networks processing sequential data. The technology of this invention fully implements the entire process from physical prior guidance modulation to multi-scale spatiotemporal interactive coding. Figure 5 The horizontal axis represents the four structural health states, and the vertical axis represents the classification accuracy, displayed as a percentage. Experimental results show that the accuracy of the method in this invention is significantly higher than the other three comparative methods for each health state. Particularly noteworthy is the significant improvement in accuracy for minor and moderate injuries, two states that are difficult to distinguish and easily misclassified. Traditional support vector machines show relatively low accuracy in these two categories, demonstrating that the model in this invention, by incorporating prior knowledge of health baseline modalities, multi-source feature fusion, and modeling the relationships between feature channels, can more precisely capture the subtle differences between different injury states, thereby achieving high-precision health state identification.

[0065] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for monitoring the health of a building structure based on machine learning, characterized by, The method comprises the following steps: S1, a three-dimensional numerical model of the target building is constructed by using finite element simulation technology, the dynamic response under different load conditions is simulated through the three-dimensional numerical model, the simulation signal corresponding to the real acquisition situation is generated, and the building health monitoring data is constructed; S2, Gaussian white noise is injected into the simulation signal to simulate simulation noise, and the parameters of the three-dimensional numerical model are adjusted to generate different types of simulation data sets, and each sample is labeled with a health category to form a data set; S3, the data in the data set is preprocessed by using a two-stage preprocessing method of adaptive wavelet denoising combined with dynamic normalization, and enhanced data is generated; S4, a two-stage method of adaptive window short-time Fourier transform combined with adaptive frequency band enhancement is used to extract high-resolution time-frequency features from the enhanced data, and the feature components related to structural damage are strengthened to generate an enhanced time-frequency feature matrix; S5, a neural network model is constructed by fusing structure physical prior guidance and multi-scale space-time interaction to build a building structure health classification model, the enhanced time-frequency feature matrix is subjected to structure physical prior guided feature modulation, and then the modulation result is subjected to multi-source feature fusion and external factor injection to generate a mixed feature vector, and then multi-scale space-time interaction coding is performed to obtain a refined feature vector, and then the refined feature vector is subjected to damage mode classification to output the probability distribution of the damage state; S6, the building structure health classification model learning process is constrained based on the regularization term of feature consistency and prediction smoothness, the total loss function of the model is calculated, and the model is trained based on the loss function; S7, the trained building structure health classification model is deployed to the actual building structure health monitoring to evaluate and warn the building structure health state.

2. The method of claim 1, wherein the machine learning-based building structure health monitoring method is characterized by, S1 is as follows: A three-dimensional numerical model is constructed according to the actual geometric size, material properties, component connection mode and boundary conditions of the target building structure; Different load conditions include environmental excitation, simulated wind load and simulated traffic load; The simulation signal is a vibration acceleration time history signal corresponding to the real acquisition situation, and the real acquisition situation is that sensors are deployed on the target building structure to collect real vibration acceleration time history signals.

3. The method of claim 1, wherein the method further comprises: S2 is as follows: The method for adjusting the parameters of the three-dimensional numerical model includes locally reducing the material stiffness, introducing micro-crack elements and changing the properties of the connecting parts, and the simulation data set of different health categories is generated by adjusting the parameters to simulate the structural damage state at different positions and different severity, and the health categories include health, slight damage, severe damage and serious damage; The data points of the simulation signal within a fixed length are taken as a sample, and the health category label is labeled for each sample according to the preset damage state during simulation.

4. The method of claim 1, wherein the method further comprises: S3 is as follows: S3.1, stationary wavelet transform is performed on the building structure health monitoring data in the data set, and then the adaptive threshold weight of each data point at each scale is dynamically calculated based on the difference between the local features of the signal and the statistical characteristics of each scale, and the denoised monitoring data is obtained after weighting and reconstruction based on the adaptive threshold weight; S3.2, the denoised monitoring data is subjected to dynamic range normalization processing to obtain enhanced data; The global normalization result is dynamically adjusted by using a scaling factor based on local statistical characteristics to obtain normalized monitoring data.

5. The method of claim 1, wherein the method further comprises: S4 is specifically as follows: S4.1, the enhanced data is mapped from the time domain to the time-frequency domain by adaptive window short-time Fourier transform, the adaptive window short-time Fourier transform uses a window function with an adaptive bandwidth to segment the time axis, and generates a high-resolution initial time-frequency feature matrix; S4.2, the global average energy of each frequency bin of the initial time-frequency feature matrix is calculated, and a global frequency band weight is generated to strengthen the high-energy frequency band, and a local time-frequency attention weight is calculated to focus on the energy mutation time-frequency point, and then the two weights are respectively applied to the initial time-frequency feature matrix, and the enhanced time-frequency feature matrix is obtained after addition.

6. The method of claim 1, wherein the method further comprises: S5 is specifically as follows: S5.1, the enhanced time-frequency feature matrix of all health state samples in the data set is estimated to represent the baseline mode atlas of the typical distribution of time-frequency energy under the health state, and the enhanced time-frequency feature matrix is subjected to physical guided attention modulation and deviation enhancement combined with the modulation intensity coefficient, to generate a physical perception feature matrix fused with health baseline prior knowledge; S5.2, the physical perception feature matrix is flattened into a vector, and a group of time domain statistics and envelope spectrum features are extracted from the noise reduction time domain signal to form a statistical feature vector, and an external environmental factor feature vector is introduced, and the three parts of information are adaptively weighted and fused by a feature fusion gate unit to obtain a mixed feature vector; S5.3, a multi-scale space-time interaction block is used to extract different time receptive field features through parallel multi-dilation rate time-space hole convolution, and the features are interactively modulated with the inter-feature channel relationship modeled by adaptive feature channel interaction, and a refined feature vector is output after stacking multiple modules; Wherein, the input of the first multi-scale space-time interaction block is the mixed feature vector, and the output of the last multi-scale space-time interaction block is the input of the next multi-scale space-time interaction block; S5.4, the refined feature vector is mapped to the unnormalized score corresponding to different structural health states through a fully connected layer, and a Softmax function fused with adaptive temperature scaling and class prior weight is used to obtain the final classification probability distribution.

7. The method of claim 1, wherein the method further comprises: The calculation process of the total loss function is specifically as follows: (1) the loss term combining adaptive margin cross-entropy and class center constraint is used, which simultaneously narrows the distance between sample features and their own class center and pushes away the distance from other class centers, and the margin value is adaptively adjusted according to the difficulty of class distinction; (2) based on the prior knowledge that damage evolution is continuous and should have consistency in time-frequency response, a regularization loss term containing time-frequency feature consistency constraint and prediction distribution smoothness constraint is used to encourage similar features of samples of the same class and smooth changes in the prediction of continuous or adjacent states, which is expressed as: (3) the total loss is the weighted sum of the adaptive margin contrast center loss and the regularization loss.

8. The method of claim 1, wherein the building is a building. The training process of the structural health classification model is specifically as follows: The data set is divided into a training data set and a validation data set, and the data in the training data set is processed by S3 and S4, and the model is trained end to end; In each training iteration, a batch of sample data is sequentially passed through the forward propagation of the model to obtain the predicted probability distribution, and then the loss value of the current batch is calculated according to the total loss function; An adaptive matrix estimator optimizer is used to calculate the gradient of all trainable parameters in the model according to the total loss function, and these parameters are updated through the back propagation algorithm to minimize the total loss function; During the training process, the performance of the model is periodically evaluated using an independent validation dataset; The training iteration of the model will continue until the performance indicator on the validation set no longer improves for consecutive multiple training cycles, or the maximum number of preset iteration cycles is reached, at which point the training stops and the model parameters with the best validation performance are saved, completing the training of the building structure health classification model.

9. The method of claim 4, wherein the method further comprises: S7 is as follows: After the building structure health classification model is trained, it is deployed in actual building structure health monitoring. Real-time or periodic collection of original vibration signals by vibration sensors installed at key positions of the building structure forms the vibration data to be monitored; at the same time, the environmental temperature and humidity data are synchronously collected to form the external environmental factor feature vector; Then, the collected data is processed and input into the trained building structure health classification model, and the model outputs the probability of the building structure belonging to each health category during the monitoring period; Determine the class corresponding to the highest probability as the current structure's health status, and visualize the monitoring results, probability distribution, and key features; Then, according to the current structure's health status, automatically trigger different levels of early warning.

10. A machine learning based building structure health monitoring system, performing a machine learning based building structure health monitoring method according to any one of claims 1-9, characterized in that, The system structure is as follows: Three-dimensional numerical model and simulation dataset generation module: used to construct a three-dimensional numerical model of the target building using finite element simulation technology, simulate the dynamic response under different load conditions to generate simulation signals, inject Gaussian white noise and adjust model parameters to generate a multi-class simulation dataset, and label the sample health category; Data preprocessing and enhancement module: uses a two-stage method of adaptive wavelet denoising combined with dynamic normalization to process the simulation dataset to generate enhanced data; Time-frequency feature extraction and enhancement module: extract high-resolution time-frequency features from enhanced data by adaptive window short-time Fourier transform combined with adaptive frequency band enhancement to strengthen damage-related feature components and generate enhanced time-frequency feature matrix; Health classification model construction and training module: construct a neural network model that integrates structure physical prior guidance and multi-scale time-space interaction, perform feature modulation, multi-source fusion, and external factor injection on the enhanced time-frequency feature matrix, obtain refined feature vectors through multi-scale time-space interaction coding, construct a total loss function based on feature consistency and prediction smoothness regularization terms, and train the model; Health monitoring and early warning module: deploy the trained model to evaluate the building structure health status, output damage probability distribution, and realize safety warning.

Images