Method and device for evaluating health state of building structure

Abstract

The present application relates to a kind of building structure health state evaluation method and device, belong to building reconstruction technical field, method includes the stress part of target building under the action of natural load or artificial excitation, original stress time series data is generated from dynamic strain response, and original stress time series data is marked with building health state category, obtain multi-dimensional marked vector with time stamp;Based on the hierarchical coding architecture of space-time feature decoupling, obtain health state evaluation model, and based on clean stress time series data as marked sample data, and health state evaluation model is trained in combination with unmarked sample data, obtain target health state evaluation model;Based on target health state evaluation model, the stress data of target building collected is analyzed, and the building structure health state of target building is determined, it is favorable to improve the accuracy, continuity and prewarning of building structure health state evaluation.

Description

Technical Field

[0001] This invention relates to the field of building renovation technology, and in particular to a method and apparatus for assessing the health status of building structures. Background Technology

[0002] With the increasing size, complexity, and service life of modern building structures, their safety and performance have become crucial for urban operation and public safety. However, during long-term service, building structures are inevitably affected by a combination of factors, including wind loads, temperature changes, earthquakes, and usage loads, which may lead to problems such as the accumulation of micro-damage, the propagation of local cracks, and fatigue aging at connection points. Traditional structural health monitoring methods mainly rely on periodic manual inspections and limited point-based sensor data, which have significant limitations such as insufficient information coverage, delayed damage detection, and strong subjectivity in judgment, making it difficult to meet the needs for continuous, accurate, and early warning of structural health status. Summary of the Invention

[0003] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method and apparatus for assessing the health status of building structures, which solves the technical problem of assessing whether there is a threat of topology leakage in publicly available data.

[0004] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0005] The first aspect of this invention provides a method for assessing the health status of building structures.

[0006] The method for assessing the health status of building structures proposed in this invention includes:

[0007] The original stress time series data of the stress-bearing parts of the target building under natural load or artificial excitation is obtained by dynamic strain response, and the original stress time series data is labeled with building health status category to obtain a multi-dimensional label vector with timestamp.

[0008] The multidimensional labeled vector with timestamps is denoised to obtain the denoised cleaning stress time series data.

[0009] A hierarchical coding architecture based on spatiotemporal feature decoupling is used to obtain a health status assessment model. The health status assessment model is trained using the cleaning stress time series data as labeled sample data to obtain a target health status assessment model. The total loss function of the health status assessment model is constructed by dynamically weighting the supervised loss and unsupervised loss according to the square root of the proportion of the labeled sample data.

[0010] Based on the target health status assessment model, data analysis is performed on the collected target building stress data to determine the structural health status of the target building.

[0011] In some instances, the step of training the health status assessment model based on the cleaning stress time series data as labeled sample data to obtain the target health status assessment model includes:

[0012] The labeled sample data is input into the input layer of the health status assessment model for model training;

[0013] During model training, the spatiotemporal features of stress are separated by a decoupling convolution mechanism, and the temporal dynamic characteristics and spatial distribution characteristics of the labeled sample data are extracted.

[0014] Dynamic convolution kernels are generated in real time based on the local statistical properties of time feature channels;

[0015] During model training, based on the structural strain energy density as a physical prior, the damage-sensitive feature channel is dynamically enhanced through a gating mechanism, and the feature enhancement region is processed by a hierarchical incremental aggregation method combined with a differential enhancement operator; wherein, the feature change region retains the multi-scale information of the sample data;

[0016] During model training, a spatiotemporal constrained spectral clustering mechanism is used to construct a sample affinity matrix, and the health status assessment model is optimized through a supervised and unsupervised joint loss function to obtain the target health status assessment model; wherein, the unsupervised joint loss function is determined based on the sample affinity matrix.

[0017] In some instances, the separation of the spatiotemporal features of stress through decoupling convolution mechanisms, and the extraction of the temporal dynamic characteristics and spatial distribution characteristics of the labeled sample data, include:

[0018] We apply one-dimensional convolutional operations in the time dimension to cleaning stress time series data, use fixed-length convolutional kernels to capture local time patterns, and enhance nonlinear representation capabilities through activation functions to extract temporal dynamic features.

[0019] Based on the topological relationship of sensor physical location and exponentially decaying location encoding, spatially perceptual convolutional kernel weights are dynamically generated.

[0020] Based on the spatially aware convolutional kernel weights, a one-dimensional convolution operation is performed on the cleaning stress time series data to capture the spatial correlation of the sensor network and output a spatial distribution feature map.

[0021] In some instances, the real-time generation of dynamic convolutional kernels based on local statistical properties of temporal feature channels includes:

[0022] A sliding window is applied to the time feature tensor to calculate the mean, standard deviation, and skewness coefficient of the data within the window;

[0023] The statistical features of mean, standard deviation and skewness coefficient are concatenated into a vector and input into the fully connected layer to generate dynamic convolutional kernel parameters.

[0024] The generated dynamic convolution kernel parameters are applied to the temporal feature channel to extract the adaptively enhanced features, thus obtaining the dynamic convolution output features.

[0025] In some instances, the method of dynamically enhancing damage-sensitive feature channels through gating mechanisms, based on structural strain energy density as a physical prior, includes:

[0026] Based on spatial eigenvalues ​​and material yield strain thresholds, strain energy density estimates are calculated at each time point.

[0027] The dynamic convolution output feature is multiplied by the strain energy density through channels to obtain the first feature operation result;

[0028] The result of the first feature operation is input into a single neuron fully connected layer and activated by the Sigmoid activation function to generate a gated weight matrix.

[0029] The gate weight matrix is ​​applied to the spatial feature tensor, and element-wise multiplication is performed to obtain the second feature operation result.

[0030] The result of the second feature operation is added to the output feature of the dynamic convolution to obtain the enhanced feature tensor.

[0031] In some instances, the hierarchical incremental aggregation method combined with differential enhancement operators is used to process and enhance feature change regions; wherein, the feature change regions retain multi-scale information of the sample data, including:

[0032] The input features enhanced by gated features are set as the initial hierarchical features, serving as the starting point for hierarchical aggregation;

[0033] Based on the starting point of the hierarchical aggregation, and combined with the feature processing of the differential enhancement operator, multiple levels of incremental aggregation iterations are performed to obtain multiple levels of aggregation features.

[0034] Average pooling is performed on the aggregated features at all levels to obtain the global features;

[0035] Multi-scale information fusion is performed based on the global features to enhance the feature variation regions.

[0036] In some instances, the sample affinity matrix is ​​constructed using a spatiotemporally constrained spectral clustering mechanism, and the health status assessment model is optimized through a supervised and unsupervised joint loss function to obtain the target health status assessment model, including:

[0037] Based on the true labels of the labeled samples and the final predicted probabilities, the deviation between the predicted results and the true labels is obtained by calculating the cross-entropy loss.

[0038] Based on the deviation between the predicted results and the true labels, and combining the global feature vector similarity of the samples and the proximity of the collection time, the similarity weight between samples is determined.

[0039] Based on the affinity matrix and the similarity weights between samples, the unsupervised loss term is obtained by calculating the sum of the product of the squared Euclidean distance between the sample prediction probability vectors and the similarity weights.

[0040] The supervised loss and unsupervised loss are dynamically weighted and combined according to the square root of the proportion of labeled data to obtain the total loss function;

[0041] The health status assessment model is optimized based on the total loss function to obtain the target health status assessment model.

[0042] In some instances, the method includes: optimizing a health status assessment model using a joint supervised and unsupervised loss function.

[0043] Based on the quasi-balance factor and parameter offset penalty term, gradient oscillations are suppressed by improving the gradient calculation and parameter update mechanism of the optimizer.

[0044] In some instances, the step of training the health status assessment model based on the cleaning stress time series data as labeled sample data to obtain the target health status assessment model includes:

[0045] Batch gradient descent is used for iterative training of the model. In each iteration, a batch of samples is randomly sampled from the labeled sample data. Forward propagation is performed to calculate the prediction results, back propagation is performed to update the parameters, and an adaptive optimizer is used to adjust the gradient.

[0046] During training, the loss function and classification accuracy on the validation set are monitored in real time. When the total loss function decreases by less than the first threshold for several consecutive times and the fluctuation range of the validation set accuracy is lower than the second threshold, the model is considered to have converged, and the target health status assessment model is obtained.

[0047] A second aspect of the present invention provides a building structure health status assessment device, comprising:

[0048] The data acquisition unit is used to acquire the original stress time series data generated by the dynamic strain response of the stress-bearing parts of the target building under natural load or artificial excitation, and to label the original stress time series data with the building health status category to obtain a multi-dimensional label vector with timestamps.

[0049] The data denoising unit is used to denoise the multidimensional labeled vector with timestamps to obtain denoised cleaning stress time series data.

[0050] The model training unit is used to obtain a health status assessment model based on a hierarchical coding architecture with decoupled spatiotemporal features, and to train the health status assessment model using the cleaning stress time series data as labeled sample data to obtain a target health status assessment model; wherein, the total loss function of the health status assessment model is constructed by dynamically weighting the supervised loss and unsupervised loss according to the square root of the proportion of the labeled sample data.

[0051] The status assessment unit is used to perform data analysis on the collected stress data of the target building based on the target health status assessment model, and to determine the structural health status of the target building.

[0052] This invention discloses a method for assessing the structural health status of a building, comprising: acquiring raw stress time-series data of the stress-bearing parts of a target building under natural loads or artificial excitation, generated by dynamic strain response, and labeling the raw stress time-series data with building health status categories to obtain a multi-dimensional label vector with timestamps; denoising the multi-dimensional label vector with timestamps to obtain denoised clean stress time-series data; acquiring a health status assessment model based on a hierarchical coding architecture with spatiotemporal feature decoupling, and training the health status assessment model using clean stress time-series data as labeled sample data and combined with unlabeled sample data to obtain a target health status assessment model; wherein, the total loss function of the health status assessment model is constructed by dynamically weighting the supervised loss and unsupervised loss according to the square root of the proportion of labeled sample data; and performing data analysis on the collected stress data of the target building based on the target health status assessment model to determine the structural health status of the target building. In this application, by constructing a health status assessment model and automatically assessing the structural health status of the target building using this model, the accuracy, continuity, and predictability of the structural health status assessment are improved. Attached Figure Description

[0053] Figure 1 A flowchart of a method for assessing the health status of building structures provided in an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram comparing the noise suppression effects of different filtering methods provided in the embodiments of the present invention;

[0055] Figure 3 This is a schematic diagram comparing the accuracy of health status classification using different methods provided in embodiments of the present invention.

[0056] Figure 4 This is a schematic diagram comparing recall rates under different degrees of damage, provided by an embodiment of the present invention.

[0057] Figure 5 This is a schematic diagram of the building structure health status assessment device provided in an embodiment of the present invention. Detailed Implementation

[0058] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] The structural health status assessment method proposed in this invention is used to address the problem of data topology leakage threat assessment. By processing power data, performing correlation path and inference relationship analysis, and constructing a link prediction model, the method assesses the topology leakage threat generated by publicly available power grid data, which helps to improve the accuracy of data topology leakage threat assessment.

[0060] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0061] Figure 1 A flowchart illustrating a method for assessing the health status of building structures, provided as an embodiment of the present invention. Figure 1 As shown, the method for assessing the health status of building structures proposed in this embodiment of the invention includes:

[0062] Step 100: Obtain the original stress time series data of the stress-bearing parts of the target building under natural load or artificial excitation, generated by dynamic strain response, and label the original stress time series data with building health status category to obtain a multi-dimensional label vector with timestamp.

[0063] Step 110: Denoise the multidimensional label vector with timestamps to obtain the denoised cleaning stress time series data;

[0064] Step 120: Based on the hierarchical coding architecture of spatiotemporal feature decoupling, obtain the health status assessment model, and use the cleaning stress time series data as labeled sample data to train the health status assessment model to obtain the target health status assessment model; wherein, the total loss function of the health status assessment model is constructed by dynamically weighting the supervised loss and unsupervised loss according to the square root of the proportion of the labeled sample data.

[0065] Step 130: Based on the target health status assessment model, perform data analysis on the collected target building stress data to determine the structural health status of the target building.

[0066] In this exemplary embodiment, the hierarchical coding architecture includes: a spatiotemporal decoupling module, a dynamic kernel generation module, a gating enhancement module, an incremental aggregation module, and a dual-stream classification module;

[0067] The hierarchical coding architecture based on spatiotemporal feature decoupling obtains a health status assessment model, including:

[0068] The health status assessment model is constructed based on a spatiotemporal decoupling module, a dynamic kernel generation module, a gating enhancement module, an incremental aggregation module, and a dual-stream classification module, combined with a spectral clustering optimization layer; wherein,

[0069] The spatiotemporal decoupling module is used to separate features through parallel temporal convolutional paths and spatial sensing convolutional paths.

[0070] The dynamic kernel generation module is used to adjust the convolution parameters in real time based on time feature statistics to adapt to non-stationary signals;

[0071] The gated enhancement module is used to utilize the prior weighted fusion of spatiotemporal features based on strain energy density;

[0072] The incremental aggregation module is used to extract multi-scale damage-sensitive features through four-level hierarchical differential pooling;

[0073] The dual-stream classification module is used to process global features and damage critical point features respectively;

[0074] The spectral clustering optimization layer is used to fuse supervised and unsupervised losses.

[0075] In this exemplary embodiment, the training data is mainly acquired through a distributed fiber optic strain sensor network deployed at key stress-bearing parts of the building structure (such as beam-column joints, cantilever ends, shear wall connections, etc.).

[0076] The sensor continuously monitors the dynamic strain response of the structure under natural loads (wind load, temperature change) and artificial excitations (vibration test, load test) at a fixed sampling frequency (e.g., 10Hz), forming the original stress time series data stream.

[0077] The data attributes include timestamps, sensor spatial location codes, original micro-strain values ​​(unit με), and synchronously collected environmental temperature and humidity parameters. The annotation work is jointly completed by structural engineers using periodic manual inspection reports (recording crack width and location), load test results (deformation data under graded loading), and finite element simulation results (damage condition simulation). The annotation categories are divided into four health states: normal (no damage characteristics), micro-damage (abnormal local strain but no visible cracks), moderate damage (cracks ≤0.2mm appear in critical areas), and severe damage (cracks ≥0.3mm appear in main load-bearing components or multi-point damage coupling).

[0078] The annotation process employs a time alignment mechanism to precisely match manually detected damage events with sensor data for the corresponding time periods, forming a multi-dimensional annotation vector with timestamps.

[0079] In this exemplary embodiment, due to environmental noise interference and local abnormal fluctuations in the stress sensor monitoring data, conventional moving average filtering will blur the minute characteristics of structural damage.

[0080] This application uses a dual-channel adaptive filtering mechanism to process the original stress time series data to eliminate environmental noise interference and local abnormal fluctuations, while preserving the minute characteristics of structural damage. The specific steps are as follows:

[0081] 1) Calculate the adaptive mixed weights

[0082] The gradient values ​​of the original stress time series data are calculated using discrete difference to characterize the rate of stress change. An adaptive threshold, calculated based on the standard deviation of the gradient value sequence, is then used to dynamically generate hybrid weights for subsequent fusion of high- and low-frequency components. This ensures a sensitive response to structural damage characteristics, expressed as follows:

[0083] ;

[0084] In the formula, This represents the mixing weight at time point i, used to control the fusion ratio of high-frequency and low-frequency components, with a value range of... ;

[0085] Parameters that control the steepness of the weighting function transition, such as... ;

[0086] Let be the gradient value at the i-th time point, calculated as follows: , characterizing the discrete difference of stress time series;

[0087] This represents the absolute value operator;

[0088] The adaptive threshold is calculated as follows: ;

[0089] It is the standard deviation of the gradient value sequence, used to identify abnormal fluctuations;

[0090] This represents the natural exponential function.

[0091] 2) Apply dual-channel filtering to generate a clean signal

[0092] The outputs of low-pass and high-pass filters are dynamically fused using hybrid weights. The hybrid weights enhance low-frequency components to suppress high-frequency noise, while the complementary weights enhance high-frequency components to preserve damage characteristics, generating denoised clean stress time-series data, represented as follows:

[0093] ;

[0094] In the formula, This represents the time-series vector of filtered cleaning stress;

[0095] This is the original stress time series vector;

[0096] Here is the low-pass filter function, and the low-pass filter cutoff frequency is... ;

[0097] The Nyquist frequency is calculated as follows: ;

[0098] The sampling frequency is used to preserve low-frequency trends;

[0099] Here is the high-pass filter function, and the high-pass filter cutoff frequency is... Used to extract high-frequency details;

[0100] This represents element-wise multiplication;

[0101] Let be a mixed weight vector, where the i-th element is the mixed weight;

[0102] As complementary weight vectors, calculated as minus Element-wise values.

[0103] In one embodiment, the noise suppression effects of different filtering methods are compared to verify the superiority of the dual-channel adaptive filtering technology of the present invention in terms of noise suppression and preservation of damage features. Figure 2This is a schematic diagram comparing the noise suppression effects of different filtering methods provided in the embodiments of the present invention. Figure 2 The paper compares the processing effects of the original signal, conventional moving average filtering, wavelet denoising, and the method of this invention. The time axis is in seconds, and the micro-strain value is in με. Experimental data show that the original signal has significant environmental noise interference and local abnormal fluctuations. Although moving average filtering smooths the noise, it over-smooths the signal in the damage feature regions (red shading) of 3.4-3.7 seconds and 7.0-7.4 seconds, resulting in a decrease in the amplitude of key damage features. Wavelet denoising is better than moving average in preserving high-frequency features, but feature distortion still exists in the damage region. In contrast, the method of this invention achieves noise suppression effects comparable to wavelet denoising in the non-damage region, while completely preserving feature mutations in the damage region, and the feature amplitude is about 30% higher than that of conventional methods. This proves that the dual-channel adaptive filtering, through a dynamic mixed weight mechanism, effectively preserves the micro-amplitude features of structural damage while suppressing environmental noise.

[0104] In some instances, the step of training the health status assessment model based on the cleaning stress time series data as labeled sample data to obtain the target health status assessment model includes:

[0105] The labeled sample data is input into the input layer of the health status assessment model for model training;

[0106] During model training, the spatiotemporal features of stress are separated by a decoupling convolution mechanism, and the temporal dynamic characteristics and spatial distribution characteristics of the labeled sample data are extracted.

[0107] Dynamic convolution kernels are generated in real time based on the local statistical properties of time feature channels;

[0108] During model training, based on the structural strain energy density as a physical prior, the damage-sensitive feature channel is dynamically enhanced through a gating mechanism, and the feature enhancement region is processed by a hierarchical incremental aggregation method combined with a differential enhancement operator; wherein, the feature change region retains the multi-scale information of the sample data;

[0109] A dual-stream classification architecture is constructed, and decision optimization is achieved through a weight fusion mechanism that adapts to damage development.

[0110] During model training, a spatiotemporal constrained spectral clustering mechanism is used to construct a sample affinity matrix, and the health status assessment model is optimized through a supervised and unsupervised joint loss function to obtain the target health status assessment model; wherein, the unsupervised joint loss function is determined based on the sample affinity matrix.

[0111] In this exemplary embodiment, the step of separating the spatiotemporal features of stress and extracting the temporal dynamic characteristics and spatial distribution characteristics of the labeled sample data through a decoupling convolution mechanism during model training includes:

[0112] We apply one-dimensional convolutional operations in the time dimension to cleaning stress time series data, use fixed-length convolutional kernels to capture local time patterns, and enhance nonlinear representation capabilities through the ReLU activation function to extract temporal dynamic features.

[0113] Based on the topological relationship of sensor physical location and exponentially decaying location encoding, spatially perceptual convolutional kernel weights are dynamically generated.

[0114] Based on the spatially aware convolutional kernel weights, a one-dimensional convolution operation is performed on the cleaning stress time series data to capture the spatial correlation of the sensor network and output a spatial distribution feature map.

[0115] In this exemplary embodiment, during model training, the spatiotemporal features of stress are separated by a decoupling convolution mechanism to extract the temporal dynamic characteristics and spatial distribution characteristics of the labeled sample data. The specific steps are as follows:

[0116] 1) Extracting temporal dynamic features

[0117] A one-dimensional convolution operation along the time dimension is applied to the cleaning stress time series data. A fixed-length convolution kernel is used to capture local temporal patterns, and the nonlinear representation capability is enhanced by the ReLU activation function to extract temporal dynamic features, as shown below:

[0118] ;

[0119] In the formula, Represents the time feature matrix;

[0120] To correct the activation function of the linear unit, a nonlinearity is introduced;

[0121] This represents a one-dimensional temporal convolution operation;

[0122] The temporal convolution kernel weight matrix is ​​shown in the superscript. Indicates the convolution kernel parameters: One output channel, This represents the kernel time length.

[0123] 2) Generate spatially aware convolutional kernels

[0124] Based on the topological relationship of the sensor's physical location and exponentially decaying location encoding, spatially aware convolutional kernel weights are dynamically generated. These weights consider the sensor's spatial neighborhood set and location distance, ensuring that the convolution operation is strongly correlated with spatial distribution, as expressed below:

[0125] ;

[0126] In the formula, The target sensor index is Reference sensor index is Spatial-aware convolution kernel weights;

[0127] Indicates the first A set of spatial neighborhood indexes for each sensor;

[0128] For neighborhood set The cardinality;

[0129] Indicates the first Time-series data of cleaning stress from each sensor;

[0130] The position encoding function is calculated as follows:

[0131] .

[0132] 3) Extracting spatial distribution features

[0133] By using spatially aware convolutional kernels to perform one-dimensional convolution operations on cleaning stress time-series data in the spatial dimension, the spatial correlation of the sensor network is captured, and the spatial distribution feature map is output, represented as follows:

[0134] ;

[0135] In the formula, Represents the spatial characteristic matrix;

[0136] This represents a one-dimensional convolution operation.

[0137] The spatial convolution kernel weight matrix is ​​represented by the superscript. This indicates 8 output channels. is the spatial length of the convolution kernel.

[0138] In some instances, the real-time generation of dynamic convolutional kernels based on local statistical properties of temporal feature channels includes:

[0139] A sliding window is applied to the time feature tensor to calculate the mean, standard deviation, and skewness coefficient of the data within the window;

[0140] The statistical features of mean, standard deviation and skewness coefficient are concatenated into a vector and input into the fully connected layer to generate dynamic convolutional kernel parameters.

[0141] The generated dynamic convolution kernel parameters are applied to the temporal feature channel to extract the adaptively enhanced features, thus obtaining the dynamic convolution output features.

[0142] In this exemplary embodiment, dynamic convolution kernels are generated in real time using local statistical characteristics based on time feature channels. The ability to capture non-stationary damage features is enhanced by adaptively adjusting the kernel weights. The specific steps are as follows:

[0143] 1) Calculate local statistical characteristics

[0144] A sliding window is applied to the time characteristic tensor to calculate the mean, standard deviation, and skewness coefficient of the data within the window. The mean represents the reference stress level, the standard deviation represents the stress fluctuation intensity, and the skewness coefficient represents the asymmetry of the stress distribution, expressed as:

[0145] ;

[0146] ;

[0147] ;

[0148] In the formula, express The mean value of the strain characteristics within the window is used to characterize the reference stress level;

[0149] express The standard deviation of the strain characteristics within the window is used to characterize the intensity of stress fluctuations;

[0150] express The skewness coefficient of the strain characteristics within the window is used to characterize the asymmetry of stress distribution;

[0151] The length of the sliding window is calculated as follows: ,in For example, half the window width. Control the scope of statistics;

[0152] Let i be the time feature value at the i-th time point, with superscript. Indicates a point-in-time index.

[0153] 2) Generate dynamic convolution kernel parameters

[0154] The statistical features of mean, standard deviation, and skewness coefficient are concatenated into a vector, which is then input into the fully connected layer to generate dynamic convolutional kernel parameters, as follows:

[0155] ;

[0156] In the formula, for The generated dynamic convolution kernel;

[0157] This means concatenating the three statistics into an input vector;

[0158] For fully connected layer operations, the input dimension is... The output dimension is .

[0159] 3) Perform dynamic convolution operation

[0160] The generated convolutional kernel is applied to the temporal feature channel to extract the adaptively enhanced features:

[0161] ;

[0162] In the formula, for No. Dynamic convolutional output features of channels;

[0163] for Dynamic kernel in position The weight value;

[0164] For time features in the first No. The value of the channel;

[0165] The channel index variable represents the channel index during convolution operations;

[0166] This indicates the time point and channel index of the output feature;

[0167] This represents the time offset and channel index of the input feature.

[0168] S304, Gating Feature Enhancement Mechanism

[0169] Damage-sensitive features of building structures are easily overwhelmed by principal stress components, and conventional attention mechanisms lack mechanical prior constraints, resulting in insufficient enhancement of key features with unclear physical meaning.

[0170] In some instances, during model training, based on the structural strain energy density as a physical prior, the damage-sensitive feature channels are dynamically enhanced through a gating mechanism, including:

[0171] Based on spatial eigenvalues ​​and material yield strain thresholds, strain energy density estimates are calculated at each time point.

[0172] The dynamic convolution output feature is multiplied by the strain energy density through channels to obtain the first feature operation result;

[0173] The result of the first feature operation is input into a single neuron fully connected layer and activated by the Sigmoid activation function to generate a gated weight matrix.

[0174] The gate weight matrix is ​​applied to the spatial feature tensor, and element-wise multiplication is performed to obtain the second feature operation result.

[0175] The result of the second feature operation is added to the output feature of the dynamic convolution to obtain the enhanced feature tensor.

[0176] In this exemplary embodiment, during model training, based on the structural strain energy density as a physical prior, the damage-sensitive feature channel is dynamically enhanced through a gating mechanism. The specific steps are as follows:

[0177] 1) Calculate the strain energy density estimate

[0178] Based on spatial eigenvalues ​​and the material yield strain threshold, the strain energy density estimate at each time point is calculated. This estimate is normalized to the total number of cells and characterizes the strain energy, expressed as follows:

[0179] ;

[0180] In the formula, for No. Strain energy density estimation of the channel;

[0181] This represents the total number of building structure units, used for normalization.

[0182] for Spatial characteristic value of the e-th unit;

[0183] is the yield strain threshold of the e-th element, characterizing the mechanical properties of the material;

[0184] Indicates the time point and structural unit index.

[0185] 2) Generate strain energy gating weights

[0186] The dynamic features are multiplied by the strain energy density using a channel multiplication. The result is input into a single-neuron fully connected layer and activated by a Sigmoid function to generate a gated weight matrix, as shown below:

[0187] ;

[0188] In the formula, This is the gated weight matrix;

[0189] This represents the Sigmoid activation function, which compresses the output to the interval [0, 1].

[0190] This represents a fully connected layer with a single neuron, enabling feature compression.

[0191] This indicates channel multiplication, which involves multiplying the eigenvalues ​​at the same channel position.

[0192] This is for dynamic convolution output features;

[0193] This is for estimating strain energy density.

[0194] 3) Enhanced Fusion of Execution Features

[0195] The gating weights are applied to the spatial feature tensor, multiplied element-wise, and then added to the dynamic features to obtain the enhanced feature tensor, represented as:

[0196] ;

[0197] In the formula, The enhanced feature tensor is used as the output of this step;

[0198] This represents element-wise multiplication;

[0199] It is a spatial feature tensor.

[0200] In some instances, the hierarchical incremental aggregation method combined with differential enhancement operators is used to process and enhance feature change regions; wherein, the feature change regions retain multi-scale information of the sample data, including:

[0201] The input features enhanced by gated features are set as the initial hierarchical features, serving as the starting point for hierarchical aggregation;

[0202] Based on the starting point of the hierarchical aggregation, and combined with the feature processing of the differential enhancement operator, multiple levels of incremental aggregation iterations are performed to obtain multiple levels of aggregation features.

[0203] Average pooling is performed on the aggregated features at all levels to obtain the global features;

[0204] Multi-scale information fusion is performed based on the global features to enhance the feature variation regions.

[0205] In this exemplary embodiment, a hierarchical incremental aggregation method combined with a differential enhancement operator is used to process and enhance the feature variation region; wherein, the feature variation region retains the multi-scale information of the sample data to avoid gradient vanishing, while enhancing the local details of the damage features, the specific steps are as follows:

[0206] 1) Initialize feature levels

[0207] The input features enhanced by gated features are set as the initial hierarchical features, serving as the starting point for hierarchical aggregation, as follows:

[0208] ;

[0209] In the formula, This represents the initial hierarchical features, serving as the starting point for the aggregation process;

[0210] The input features are enhanced by gating, with the same dimensions as the original features, representing the enhanced stress temporal features.

[0211] 2) Execution-level incremental aggregation

[0212] Incremental aggregation is performed iteratively across multiple levels, pairing feature blocks from the previous layer, applying difference enhancement operators to strengthen regions of feature difference, and then performing max pooling to compress feature dimensions, thus achieving multi-scale feature preservation, as shown below:

[0213] ;

[0214] ;

[0215] In the formula, Indicates the first Level 1 Block output characteristics;

[0216] This is a hierarchical index, with a value range of [value range missing]. This indicates four aggregation levels;

[0217] This represents the max pooling operation, used for feature dimensionality reduction while preserving salient features;

[0218] The difference enhancement operator is defined as the second formula;

[0219] For the first Level 1 The feature tensor of the block;

[0220] For the first Level 1 The feature tensor of the block;

[0221] The first input feature of the differential enhancement operator;

[0222] This is the second input feature of the differential enhancement operator;

[0223] This represents the output of the differential enhancement operator;

[0224] This represents a one-dimensional convolution operation with a kernel size of . , used for feature transformation.

[0225] 3) Calculate global features

[0226] Average pooling is performed on the aggregated features across all levels, and then the average is taken to obtain the global features. Multi-scale information is fused to solve the gradient vanishing problem, as shown below:

[0227] ;

[0228] In the formula, This represents the global features of the final output, which are used for subsequent classification tasks.

[0229] This indicates an average pooling operation, which compresses the feature dimension.

[0230] This application can construct a two-stream classification architecture to process global features and local mutation features separately, and achieve decision optimization through a weight fusion mechanism that adapts to damage development. The specific steps are as follows:

[0231] 1) Global Classification Stream Processing

[0232] The global features are transformed by a fully connected layer and normalized using Softmax to output the global health state probability distribution, which is represented as follows:

[0233] ;

[0234] In the formula, This represents the output of the global classification stream, which indicates the predicted probability distribution of various health states.

[0235] This represents the normalized exponential function, which converts the output into a probability distribution;

[0236] This represents a fully connected layer operation with an input dimension of 128 and an output dimension of... ;

[0237] This represents the total number of health status categories, such as: normal, minor injury, moderate injury, and severe injury.

[0238] 2) Local peak point detection

[0239] Calculate the L2 norm of the second derivative of the enhancement feature with respect to time, locate the moment when the feature mutation is most drastic, and take it as the critical point of structural damage, expressed as:

[0240] ;

[0241] In the formula, This indicates the moment when the detected characteristic peak occurs, corresponding to the critical point of structural damage.

[0242] This indicates the time parameter corresponding to the maximum value. ;

[0243] Indicates L2 norm calculation;

[0244] Indicates enhanced features Regarding time The second derivative is calculated by discretization using the central difference method;

[0245] The output features from the gated feature enhancement module contain spatiotemporal dimension information;

[0246] 3) Local Classification Flow Processing

[0247] At the detected peak moment, an enhanced feature fragment is extracted, subjected to fully connected layer transformation and Softmax normalization, and the output is the health state probability based on the damage threshold, represented as:

[0248] ;

[0249] In the formula, The output of the local classification flow represents the probability of a healthy state based on the damage threshold.

[0250] Indicates the peak time The extracted enhanced feature fragment;

[0251] This represents a fully connected layer operation with an input dimension of 64 and an output dimension of 64. ;

[0252] The index represents the characteristic peak time, indicating the critical time point where structural damage mutations are most severe.

[0253] 4) Adaptive decision fusion

[0254] The adaptive weights of the global flow are dynamically calculated based on the cumulative number of damage warnings. The global classification results are then weighted and fused with the local classification results to obtain the final fusion decision output, expressed as:

[0255] ;

[0256] ;

[0257] In the formula, Indicates the fusion decision output;

[0258] This represents the adaptive weight of the global flow, with a value range of (0, 1).

[0259] Represents the adaptive weights of the local flow;

[0260] Indicates the cumulative number of damage warnings;

[0261] This is the weighting adjustment coefficient, with a fixed value. ;

[0262] This represents the natural exponential function.

[0263] In some instances, during model training, a spatiotemporally constrained spectral clustering mechanism is used to construct a sample affinity matrix, and a health status assessment model is optimized using a supervised and unsupervised joint loss function to obtain the target health status assessment model, including:

[0264] Based on the true labels of the labeled samples and the final predicted probabilities, the deviation between the predicted results and the true labels is obtained by calculating the cross-entropy loss.

[0265] Based on the deviation between the predicted results and the true labels, and combining the global feature vector similarity of the samples and the proximity of the collection time, the similarity weight between samples is determined.

[0266] Based on the affinity matrix and the similarity weights between samples, the unsupervised loss term is obtained by calculating the sum of the product of the squared Euclidean distance between the sample prediction probability vectors and the similarity weights.

[0267] The supervised loss and unsupervised loss are dynamically weighted and combined according to the square root of the proportion of labeled data to obtain the total loss function;

[0268] The health status assessment model is optimized based on the total loss function to obtain the target health status assessment model.

[0269] In this exemplary embodiment, during model training, a spatiotemporally constrained spectral clustering mechanism is used to construct a sample affinity matrix. The model is optimized through a joint supervised and unsupervised loss function to enhance the prediction consistency of similar samples. The specific steps are as follows:

[0270] 1) Supervised classification loss calculation

[0271] Based on the true labels of the labeled samples and the final predicted probabilities, the deviation between the predicted results and the true labels is obtained by calculating the cross-entropy loss, as follows:

[0272] ;

[0273] In the formula, This is the supervised loss term, which measures the deviation between the predicted results and the true labels;

[0274] The one-hot encoded value of the sample's true label in the c-th category;

[0275] The probability value of the final predicted probability vector output in step S6 is the probability value of the c-th category.

[0276] It is the natural logarithm function.

[0277] 2) Construction of spatiotemporal constraint affinity matrix

[0278] Combining the global feature vector similarity of samples with the proximity of collection time, a similarity weight between samples is defined. The similarity is calculated based on the feature space distance and the time window threshold, and is expressed as:

[0279] ;

[0280] In the formula, Let be the similarity weight between the i-th sample and the j-th sample. ;

[0281] Let be the global feature vector of the i-th sample;

[0282] Let be the global feature vector of the j-th sample;

[0283] For vectors The squared norm represents the spatial distance of features;

[0284] The Gaussian kernel width parameter controls the similarity decay rate.

[0285] Let be the timestamp of the i-th sample.

[0286] Let j be the timestamp of the collection of the j-th sample;

[0287] For indicator functions, when The value is 1 when the time is right, and 0 otherwise. For example, the time window threshold, ;

[0288] It is a natural exponential function.

[0289] 3) Calculation of unsupervised consistency loss

[0290] Based on the affinity matrix, the sum of the product of the squared Euclidean distance between the predicted probability vectors of samples and the similarity weights is calculated as an unsupervised loss term, forcing similar samples to have similar prediction distributions, expressed as:

[0291] ;

[0292] In the formula, The unsupervised loss term forces similar samples to have similar prediction distributions;

[0293] Let be the final predicted probability vector for the i-th sample;

[0294] Let be the final predicted probability vector for the j-th sample;

[0295] The squared Euclidean distance between the predicted vectors.

[0296] 4) Adaptive weighted total loss combination

[0297] The supervised and unsupervised losses are dynamically weighted and combined according to the square root of the proportion of labeled data to obtain the total loss function, which is expressed as:

[0298] ;

[0299] In the formula, This is the total loss function;

[0300] For example, the unsupervised loss weight coefficients, ;

[0301] To label the number of samples;

[0302] This represents the total number of samples;

[0303] The weight of unsupervised terms is dynamically adjusted based on the square root of the percentage of labeled data.

[0304] In some instances, this application can improve the gradient calculation and parameter update mechanism of the optimizer by using a class balance factor and a parameter offset penalty term to suppress gradient oscillations and enhance the generalization ability for minority class samples. The specific steps are as follows:

[0305] 1) Calculate the normalization factor for the equilibrium gradient.

[0306] Based on the sample size of each class, a class-balanced gradient normalization factor is calculated. This factor is used to apply class weights in gradient estimation, reducing the influence of the class with the larger sample size on the optimization direction. It is expressed as:

[0307] ;

[0308] In the formula, This represents the balancing factor for the c-th category, used to adjust the gradient magnitude;

[0309] Let c be the number of samples in the c-th category;

[0310] The number of samples in the k-th category;

[0311] This represents the largest number of samples across all categories;

[0312] Indexed by category;

[0313] Use the index for category traversal.

[0314] 2) Construct parameter update rules with stability constraints

[0315] The step of training the health status assessment model based on the cleaning stress time series data as labeled sample data to obtain the target health status assessment model includes:

[0316] During model training, gradient first-moment and second-moment estimates are used, combined with a class balance factor, to add a parameter offset penalty term to the parameter update rule. This constrains the parameter update magnitude and avoids drastic parameter shifts caused by overfitting. The constrained parameter update is expressed as:

[0317] ;

[0318] ;

[0319] ;

[0320] ;

[0321] In the formula, express loss function Regarding model parameters The gradient;

[0322] Gradient operator;

[0323] for Gradient first moment estimation;

[0324] for Gradient first moment estimation;

[0325] This is the total loss function;

[0326] The exponential decay rate is estimated by the first moment, such as... ;

[0327] for Gradient second moment estimation;

[0328] for Second moment estimation of gradient at time;

[0329] The exponential decay rate estimated by the second moment, such as, ;

[0330] This represents element-wise multiplication;

[0331] Represents the L2 norm;

[0332] For class-balanced gradient normalization factor vector;

[0333] for Updated model parameters;

[0334] for Current model parameters;

[0335] These are the initial values ​​for the model parameters;

[0336] For example, the parameter offset penalty coefficient, Control the strength of historical parameter constraints;

[0337] Based on the learning rate, such as, ;

[0338] For example, the numerical stability constant, .

[0339] In this exemplary embodiment, the model can be trained iteratively using mini-batch gradient descent. In each iteration, batches of samples are randomly sampled from the labeled dataset, and forward propagation is performed sequentially to calculate the prediction results, backpropagation is performed to update the parameters, and the gradient is adjusted by the adaptive optimizer in S308.

[0340] During training, the loss function and classification accuracy on the validation set are monitored in real time: when the total loss decreases by less than a threshold (e.g., 0.1%) over 20 consecutive epochs and the accuracy fluctuation range of the validation set (reserved from 10% of the training set) is less than 1%, the model is considered to have converged.

[0341] The storage strategy adopts a two-track system, saving the current optimal model parameters each time the performance of the validation set improves;

[0342] The final iteration version is saved at the end of training.

[0343] All model parameters, optimizer states, and training metadata (such as loss curves and hyperparameters) are persistently stored, supporting training interruption recovery and model version rollback.

[0344] In one embodiment, Figure 3 This diagram illustrates a comparison of the accuracy of health status classification using different methods provided in embodiments of the present invention. Figure 3 As shown, the accuracy of different methods for classifying health status is compared, and the experiment evaluates the performance difference between the model of this invention and traditional methods in the task of classifying building health status. The bar chart clearly shows the classification accuracy (in %) of Support Vector Machine, Random Forest, One-Dimensional Convolutional Neural Network, Long Short-Term Memory Network, and the method of this invention. The experimental results show that traditional machine learning methods (SVM approximately 68%, Random Forest approximately 73%) are limited by their feature representation capabilities and have difficulty capturing the spatiotemporal correlation of stress data. Although deep learning methods (One-Dimensional Convolutional Neural Network approximately 79%, Long Short-Term Memory Network approximately 81%) have improved, their sensitivity to micro-damage features is still insufficient. The method of this invention significantly leads with an accuracy of approximately 93%, which is more than 12 percentage points higher than the best baseline method. The experimental results show that in the spatiotemporal feature decoupling mechanism and the dual-stream decision fusion architecture, the spatiotemporal decoupling encoding separates the dynamic fluctuation pattern and spatial distribution characteristics of stress, while the dual-stream classifier simultaneously captures the overall trend and local mutations, thereby achieving more accurate discrimination of the four levels of health status (normal / micro-damage / moderate damage / severe damage).

[0345] In one embodiment, recall rates are compared under different degrees of damage to focus on the detection sensitivity of various methods for early damage. Figure 4 This is a schematic diagram comparing recall rates under different degrees of damage, provided as an embodiment of the present invention. Figure 4 As shown in the line graph, the recall rates (in %) of the conventional method and the present invention are compared in four healthy states. Data shows that under normal conditions, the two methods perform similarly, but as the degree of damage increases, the advantages of the present invention gradually become apparent. In micro-damage detection, the recall rate of the present invention is approximately 89%, an improvement of over 23 percentage points compared to the conventional method; in moderate damage detection, it improves by approximately 20 percentage points to 92%; and in severe damage detection, it maintains a high sensitivity of approximately 96%. This step-by-step advantage verifies the effectiveness of the incremental aggregation module and the gating enhancement mechanism. The four-level hierarchical differential pooling enhances multi-scale damage characteristics, while the strain energy density-weighted gating mechanism significantly improves the response intensity of the damage-sensitive channel.

[0346] After the model training is completed, the evaluation of the newly acquired building stress data involves four steps:

[0347] First, real-time noise suppression is performed by inputting the raw stress data into a dual-channel adaptive filter to generate a clean time-series signal. Then, the feature extraction engine inputs the clean data into the loaded evaluation model and sequentially performs spatiotemporal decoupling coding, dynamic convolution enhancement, gated feature fusion, and hierarchical aggregation to output a global feature vector and damage critical point features.

[0348] Then, the two-stream classifier simultaneously calculates the global health probability distribution (reflecting long-term degradation trends) and the local mutation probability distribution (capturing sudden damage), and generates the final four-dimensional state probability vector through adaptive weight fusion.

[0349] The final decision-making module determines the health level based on the maximum probability (e.g., a red alert is triggered if the probability of serious injury is >0.7).

[0350] This invention provides a device for assessing the health status of building structures. Figure 5 This is a schematic diagram of the building structure health status assessment device provided in an embodiment of the present invention. Figure 5 As shown, it includes:

[0351] The data acquisition unit 50 is used to acquire the original stress time series data generated by the dynamic strain response of the stress-bearing parts of the target building under natural load or artificial excitation, and to label the original stress time series data with the building health status category to obtain a multi-dimensional label vector with timestamps.

[0352] The data denoising unit 51 is used to denoise the multidimensional label vector with timestamps to obtain denoised cleaning stress time series data.

[0353] The model training unit 52 is used to obtain a health status assessment model based on a hierarchical coding architecture with decoupled spatiotemporal features, and to train the health status assessment model based on the cleaning stress time series data as labeled sample data to obtain a target health status assessment model; wherein, the total loss function of the health status assessment model is constructed by dynamically weighting the supervised loss and unsupervised loss according to the square root of the proportion of the labeled sample data.

[0354] The status assessment unit 53 is used to perform data analysis on the collected target building stress data based on the target health status assessment model to determine the structural health status of the target building.

[0355] In this exemplary embodiment, the hierarchical coding architecture includes: a spatiotemporal decoupling module, a dynamic kernel generation module, a gating enhancement module, an incremental aggregation module, and a dual-stream classification module;

[0356] The hierarchical coding architecture based on spatiotemporal feature decoupling obtains a health status assessment model, including:

[0357] The health status assessment model is constructed based on a spatiotemporal decoupling module, a dynamic kernel generation module, a gating enhancement module, an incremental aggregation module, and a dual-stream classification module, combined with a spectral clustering optimization layer; wherein,

[0358] The spatiotemporal decoupling module is used to separate features through parallel temporal convolutional paths and spatial sensing convolutional paths.

[0359] The dynamic kernel generation module is used to adjust the convolution parameters in real time based on time feature statistics to adapt to non-stationary signals;

[0360] The gated enhancement module is used to utilize the prior weighted fusion of spatiotemporal features based on strain energy density;

[0361] The incremental aggregation module is used to extract multi-scale damage-sensitive features through four-level hierarchical differential pooling;

[0362] The dual-stream classification module is used to process global features and damage critical point features respectively;

[0363] The spectral clustering optimization layer is used to fuse supervised and unsupervised losses.

[0364] In this application, by constructing a health status assessment model and using the health status assessment model to automatically assess the structural health status of the target building, it is beneficial to improve the accuracy, continuity and predictability of the structural health status assessment.

[0365] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the systems / devices based on the methods described in the above embodiments of the present invention, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0366] This invention provides a computer-readable storage medium storing a building structure health status assessment program thereon. When the building structure health status assessment program is executed by a processor, it implements the building structure health status assessment method described in the above embodiments.

[0367] This invention provides an electronic device, including a memory, a processor, and a building structure health status assessment program stored in the memory and executable on the processor. When the processor executes the building structure health status assessment program, it implements the building structure health status assessment method described in the above embodiments.

[0368] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0369] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0370] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "over," or "on top" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," or "beneath" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0371] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0372] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method of assessing the health state of a building structure, characterized by, The method comprises the following steps: acquiring original stress time series data generated by dynamic strain response of a stress part of a target building under the action of natural load or artificial excitation, and labeling the original stress time series data with a building health state category to obtain a multi-dimensional labeled vector with a time stamp; performing denoising processing on the multi-dimensional labeled vector with a time stamp to obtain clean stress time series data after denoising; training a health state evaluation model based on the clean stress time series data as labeled sample data to obtain a target health state evaluation model; wherein a total loss function of the health state evaluation model is constructed by dynamically weighting a supervised loss and an unsupervised loss according to the square root of the proportion of sample labeled data; the health state evaluation model is obtained based on a pre-decoupled spatiotemporal feature hierarchical encoding architecture; performing data analysis on collected target building stress data based on the target health state evaluation model to determine the building structure health state of the target building; wherein the training of the health state evaluation model based on the clean stress time series data as labeled sample data to obtain the target health state evaluation model comprises: inputting the labeled sample data into an input layer of the health state evaluation model for model training; during the model training, separating the spatiotemporal features of the stress by a decoupling convolution mechanism, extracting the time dynamic characteristics and spatial distribution characteristics of the labeled sample data, and generating a dynamic convolution kernel based on the local statistical characteristics of the time feature channel in real time; during the model training, dynamically enhancing the damage sensitive feature channel based on the structure strain energy density as a physical prior through a gating mechanism, and processing the feature change area by combining a difference enhancement operator through a hierarchical incremental aggregation method; wherein the feature change area retains the multi-scale information of the sample data; during the model training, constructing a sample affinity matrix by using a spatiotemporal constrained spectral clustering mechanism, and optimizing the health state evaluation model by a supervised and unsupervised joint loss function to obtain the target health state evaluation model; wherein the unsupervised joint loss function is determined based on the sample affinity matrix.

2. The building structure health state evaluation method according to claim 1, wherein the separation of the spatiotemporal features of the stress by the decoupling convolution mechanism, and the extraction of the time dynamic characteristics and spatial distribution characteristics of the labeled sample data, comprise: applying a one-dimensional convolution operation in the time dimension to the clean stress time series data, using a fixed length convolution kernel to capture local time patterns, and using an activation function to enhance the nonlinear representation ability to extract time dynamic characteristics; dynamically generating a spatial perception convolution kernel weight based on the sensor physical position topology relationship and the exponential decay position encoding; performing a one-dimensional convolution operation in the spatial dimension on the clean stress time series data based on the spatial perception convolution kernel weight, capturing the spatial correlation of the sensor network, and outputting a spatial distribution feature map.

3. The building structure health state evaluation method according to claim 2, wherein the real-time generation of a dynamic convolution kernel based on the local statistical characteristics of the time feature channel, comprises: applying a sliding window to the time feature tensor to calculate the mean, standard deviation and skewness coefficient of the data in the window; The mean, standard deviation and skewness coefficient statistical features are spliced into a vector and input into a full connection layer to generate dynamic convolution kernel parameters; The generated dynamic convolution kernel parameters are applied to the time feature channel to extract self-adaptive enhanced features and obtain dynamic convolution output features.

4. The building structure health state evaluation method according to claim 3, wherein The structure strain energy density is used as a physical prior, and a damage-sensitive feature channel is dynamically enhanced through a gating mechanism, including: Based on the spatial feature value and the material yield strain threshold, the strain energy density estimation at each time point is calculated; The dynamic convolution output features are multiplied by the strain energy density to obtain a first feature operation result; The first feature operation result is input into a single neuron full connection layer and activated by a Sigmoid function to generate a gating weight matrix; The gating weight matrix is applied to the spatial feature tensor for element-wise multiplication to obtain a second feature operation result; The second feature operation result is added to the dynamic convolution output features to obtain an enhanced feature tensor.

5. The building structural health state evaluation method according to claim 4, wherein The feature change region is processed by a hierarchical incremental aggregation method combined with a difference enhancement operator; wherein the feature change region retains multi-scale information of the sample data, including: The input features after gated feature enhancement are set as initial hierarchical features as the starting point of hierarchical aggregation; Based on the starting point of hierarchical aggregation, the features are processed by the difference enhancement operator for incremental aggregation iteration at multiple levels to obtain aggregated features at multiple levels; The aggregated features at all levels are averaged and pooled to obtain global features; The global features are used for multi-scale information fusion to strengthen the feature change region.

6. The building structural health state evaluation method according to claim 5, wherein The spectral clustering mechanism with spatiotemporal constraints is used to construct a sample affinity matrix, and a supervised and unsupervised joint loss function is used to optimize a health state evaluation model to obtain the target health state evaluation model, including: Based on the true labels and final prediction probabilities of the labeled samples, the deviation between the prediction results and the true labels is obtained by calculating the cross-entropy loss; Based on the deviation between the prediction results and the true labels, the similarity weight between samples is determined based on the global feature vector similarity and the proximity of the collection time of the samples; Based on the affinity matrix and the similarity weight between samples, the sum of the product of the Euclidean distance square between sample prediction probability vectors and the similarity weight is obtained as an unsupervised loss term; The supervised loss and the unsupervised loss are dynamically weighted and combined according to the square root of the proportion of labeled data to obtain a total loss function; The health state evaluation model is optimized based on the total loss function to obtain the target health state evaluation model.

7. The building structural health state evaluation method according to Claim 1, wherein During the process of optimizing the health state evaluation model by a supervised and unsupervised joint loss function, Based on the class balance factor and the parameter offset penalty term, the gradient calculation and parameter update mechanism of the improved optimizer is used to suppress gradient oscillation.

8. The building structural health state evaluation method according to claim 6, wherein The cleaning stress time series data is used as labeled sample data to train the health state evaluation model to obtain a target health state evaluation model, including: The model is iteratively trained in a batch gradient descent manner; wherein, in each iteration, a batch of samples is randomly sampled from the labeled sample data, the forward propagation is performed to calculate the prediction result, the back propagation is performed to update the parameters, and the gradient is adjusted by the adaptive optimizer; During the training process, the loss function and the classification accuracy on the validation set are monitored in real time; wherein, when the descending amplitude of the total loss function for a plurality of consecutive times is less than a first threshold and the fluctuation range of the validation set accuracy is lower than a second threshold, it is determined that the model converges, and the target health state evaluation model is obtained.

9. A building structure health state evaluation device characterized by comprising: It comprises: A data acquisition unit is configured to obtain original stress time series data generated by dynamic strain response of a force-bearing part of a target building under the action of natural load or artificial excitation, and to label the original stress time series data with a building health state category to obtain a multi-dimensional labeled vector with a timestamp; A data denoising unit is configured to perform denoising processing on the multi-dimensional labeled vector with a timestamp to obtain clean stress time series data after denoising; A model training unit is configured to obtain a health state evaluation model based on a hierarchical encoding architecture for decoupling spatio-temporal features, and to train the health state evaluation model based on the clean stress time series data as labeled sample data to obtain a target health state evaluation model; wherein, the total loss function of the health state evaluation model is constructed by dynamically weighting the supervised loss and the unsupervised loss according to the square root of the proportion of sample labeled data; A state evaluation unit is configured to analyze the collected stress data of the target building based on the target health state evaluation model to determine the building structure health state of the target building; wherein, the health state evaluation model is trained based on the clean stress time series data as labeled sample data to obtain a target health state evaluation model, which comprises: Inputting the labeled sample data into the input layer of the health state evaluation model for model training; During the model training process, the spatio-temporal features of the stress are separated by a decoupling convolution mechanism to extract the time dynamic characteristics and spatial distribution characteristics of the labeled sample data, and a dynamic convolution kernel is generated in real time based on the local statistical characteristics of the time feature channel; During the model training process, the damage sensitive feature channel is dynamically enhanced by a gating mechanism based on the structural strain energy density as a physical prior, and the feature change region is processed by combining a difference enhancement operator through a hierarchical incremental aggregation method; wherein, the feature change region retains the multi-scale information of the sample data; During the model training process, a sample affinity matrix is constructed by a spatio-temporal constrained spectral clustering mechanism, and the health state evaluation model is optimized by a supervised and unsupervised joint loss function to obtain the target health state evaluation model; wherein, the unsupervised joint loss function is determined based on the sample affinity matrix.

Images