Curtain wall system connecting node stress state recognition method based on deep learning

Abstract

The application discloses a curtain wall system connecting node stress state recognition method based on deep learning, and the method comprises the following steps: collecting original stress time series data and synchronous environment temperature data of a curtain wall connecting node to form a sample set; based on dynamic time clustering, the original stress time series data of each sample is segmented and divided, and an adaptive feature mapping function combining segmented information and a wavelet base function is used for feature mapping, and then dimension reduction is performed through principal component analysis to obtain a dimension reduction feature vector; a deep learning recognition network is constructed, and a double supervision loss function containing a weighted time series focal loss and an attention consistency regularization term is used to train the deep learning recognition network; and the obtained enhanced feature sequence and external physical features are input into the trained deep learning recognition network to output a stress state category of a node to be recognized. The application realizes automatic and engineering deployable transformation from original multi-source monitoring data to a clear state grade.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for identifying the stress state of connection nodes in a curtain wall system based on deep learning. Background Technology

[0002] With the increasing emphasis on building structural safety, curtain walls, as a crucial external envelope system for modern high-rise buildings, directly impact the overall structural safety and service life. Curtain wall systems connect glass panels, metal frames, and the main structure through numerous connection nodes. These nodes endure alternating stresses under complex environmental excitations such as wind loads, temperature variations, and seismic forces, making them highly susceptible to stress concentration, fatigue damage, and even localized failure. This can trigger a chain reaction, threatening structural safety. Therefore, real-time and accurate monitoring and identification of the stress state at curtain wall connection nodes has become a vital and highly challenging task in the field of structural health monitoring.

[0003] The existing technologies suffer from the following shortcomings: First, feature extraction methods are rigid. Traditional fixed-scale wavelet transforms or global normalization struggle to adapt to the non-stationarity and local abrupt changes in stress signals, resulting in limited feature discrimination. Second, models neglect physical background and crucial temporal information. Conventional deep learning models, such as fully connected networks or standard recurrent neural networks, fail to effectively integrate environmental stimulus data like temperature and treat all time steps equally, failing to focus on truly important segments. Third, feature representation capabilities are insufficient. Conventional methods typically use raw or simply normalized features directly, lacking in-depth exploration and enhancement of the complex distribution structure of the feature space, thus limiting the performance ceiling of the classifier. Fourth, training objectives are singular and lack interpretability constraints. Traditional cross-entropy loss functions aim only at correct classification, failing to consider the consistency of temporal patterns in curtain wall stress data and failing to guide the model to learn decision-making criteria with clear physical meaning, leading to doubts about the model's generalization ability and reliability.

[0004] Therefore, proposing a deep learning-based method for identifying the stress state of connection nodes in curtain wall systems to address the difficulties in existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method for identifying the stress state of connection nodes in a curtain wall system, forming an end-to-end intelligent identification process that enables automated and engineerable deployment from raw multi-source monitoring data to a clearly defined state level.

[0006] To achieve the above objectives, the present invention provides the following solution: A deep learning-based method for identifying the stress state of connection nodes in a curtain wall system includes the following steps: S1. Collect the original stress time series data and synchronous ambient temperature data of the curtain wall connection nodes to form a sample set, and label the corresponding stress state category of each sample. S2. For the original stress time series data of each sample, segmentation is performed based on dynamic time clustering, and feature mapping is performed using an adaptive feature mapping function that combines segmentation information with wavelet basis functions to obtain the mapped feature vector. Then, dimensionality reduction is performed through principal component analysis to obtain the dimensionality-reduced feature vector. S3. Fit a Gaussian mixture model using the dimensionality-reduced feature vectors, calculate the posterior probability of each sample's dimensionality-reduced feature vector belonging to each Gaussian component, form a posterior probability feature vector, and concatenate the dimensionality-reduced feature vector with the posterior probability feature vector to generate an enhanced feature vector. S4. Construct a deep learning recognition network, including a gated recurrent unit for temporal encoding, an adaptive temporal attention mechanism that fuses external physical information, and a classification layer; the adaptive temporal attention mechanism calculates attention weights by fusing the hidden state of the gated recurrent unit, the posterior probability feature vector, and the external physical features of the current time step. S5. The deep learning recognition network is trained using a dual-supervised loss function that includes weighted temporal focus loss and attention consistency regularization term. S6. Process the original stress time series data and ambient temperature data of the node to be identified sequentially through steps S2-S4, and input the obtained enhanced feature sequence and external physical features into the trained deep learning recognition network to output the stress state category of the node to be identified.

[0007] Preferably, in S2, feature mapping is performed using an adaptive feature mapping function that combines piecewise information with wavelet basis functions to obtain mapped feature vectors. Then, dimensionality reduction is performed through principal component analysis to obtain dimensionality-reduced feature vectors. Specifically, this includes: K-Means clustering was performed on the time index set of the original stress time series data of the samples, dividing it into K time segments, and obtaining the center point index and width parameter of each segment. The segment information was used to construct a feature mapping function that integrates adaptive weights and wavelet basis functions. This feature mapping function extracts frequency domain features within a local time window through wavelet basis functions and dynamically adjusts the feature contribution of different segments in combination with adaptive weight coefficients to generate a mapped feature vector. Then, dimensionality reduction was performed through principal component analysis to obtain the dimensionality-reduced feature vector.

[0008] Preferred adaptive weighting coefficients The calculation is based on the proximity of the current stress value to the statistical mean of each segment. The weight value decreases as the difference increases, and the sum is ensured to be 1 through normalization.

[0009] Preferably, in S3, a Gaussian mixture model is fitted using dimensionality-reduced feature vectors to ultimately generate enhanced feature vectors, specifically including: Using the expectation-maximization algorithm, based on the dimensionality-reduced features of all training samples... If a Gaussian mixture model containing C multivariate Gaussian distribution components is fitted, then the model parameter set is: For each sample's dimensionality-reduced features, its rank (belonging to the 1st) is calculated using the fitted Gaussian mixture model parameters. The posterior probability of each Gaussian component; classifying a sample as belonging to all... The posterior probabilities of the Gaussian components are arranged in order to form The original dimensionality-reduced feature vector is concatenated vertically with the posterior probability feature vector to form the final enhanced feature vector. .

[0010] Preferably, in S4, the adaptive temporal attention mechanism calculates attention weights by fusing the hidden state of the gated recurrent unit, the posterior probability feature vector, temperature, and external physical features of the current time step, specifically including: Constructing external physical features The feature is based on the mutual information weighting of the enhanced features and temperature data, and is scaled in combination with the temperature change rate; the attention score is calculated by fusing the latent state, posterior probability feature vector and external physical features, and entropy regularization is introduced to encourage attention to time steps with high uncertainty; finally, the attention score is normalized to obtain the attention weight.

[0011] Preferably, in S5, the dual-supervised loss function includes a weighted temporal focus loss. Attention Consistency Regularization The total loss function is a weighted sum of the two, where the weighted temporal focus loss is used to handle class imbalance and focus-difficult samples, and the attention consistency regularization term is used to improve the physical interpretability of attention.

[0012] Preferred, weighted timing focus loss The calculation is based on class balance weights and temporal attention weights, and the focus is enhanced on samples that are difficult to classify by using focus adjustment parameters.

[0013] Preferred attention consistency regularization term The robustness of network model decisions can be enhanced by encouraging the clarity of the posterior probability distribution and the compactness of features in high-attention time steps.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a deep learning-based method for identifying the stress state of connection nodes in a curtain wall system as described above.

[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention proposes an innovative deep learning-based method for identifying the stress state of curtain wall connection nodes. Its innovations are mainly as follows: First, it employs a feature mapping method combining dynamic temporal clustering and adaptive wavelet weighting, which can self-adjust based on local signal characteristics, overcoming the shortcomings of fixed-scale analysis in capturing multi-scale fluctuations and local stress concentrations. Second, it introduces a Gaussian mixture model for probabilistic modeling of the feature space, using the posterior probability of a sample belonging to each component as a soft-assignment feature for enhancement, improving the discriminative power of features against complex latent patterns. Third, it constructs a recognition network that integrates external physical information and adaptive attention. By explicitly introducing temperature data and designing mutual information weighted features, the attention mechanism can dynamically focus on key time steps that are significantly affected by environmental stimuli and exhibit drastic changes in internal features. Fourth, it designs a dual-supervised loss function combining weighted temporal focus loss and attention consistency regularization, forcing the model to ensure that its attention focuses on feature segments with clear physical meaning and typical patterns while maintaining classification accuracy, thus enhancing the interpretability and robustness of the decision. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the deep learning-based method for identifying the stress state of connection nodes in a curtain wall system provided by this invention. Figure 2 This is a comparison diagram of the stress state identification results of the method of the present invention and other existing methods in the embodiments; Figure 3 This is a diagram showing the analysis results of the adaptive temporal attention mechanism of the present invention in the embodiment; Among them, (a) is the original stress monitoring time series signal, (b) is the synchronous ambient temperature monitoring data, and (c) is the weight distribution of the adaptive attention mechanism. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, the present invention provides a method for identifying the stress state of connection nodes in a curtain wall system based on deep learning, comprising the following steps: S1. Collect the original stress time series data and synchronous ambient temperature data of the curtain wall connection nodes to form a sample set, and label the corresponding stress state category of each sample. S2. For the original stress time series data of each sample, segmentation is performed based on dynamic time clustering, and feature mapping is performed using an adaptive feature mapping function that combines segmentation information with wavelet basis functions to obtain the mapped feature vector. Then, dimensionality reduction is performed through principal component analysis to obtain the dimensionality-reduced feature vector. S3. Fit a Gaussian mixture model using the dimensionality-reduced feature vectors, calculate the posterior probability of each sample's dimensionality-reduced feature vector belonging to each Gaussian component, form a posterior probability feature vector, and concatenate the dimensionality-reduced feature vector with the posterior probability feature vector to generate an enhanced feature vector. S4. Construct a deep learning recognition network, including a gated recurrent unit for temporal encoding, an adaptive temporal attention mechanism that integrates external physical information, and a classification layer; the adaptive temporal attention mechanism calculates attention weights by integrating the hidden state of the gated recurrent unit, the posterior probability feature vector, and the external physical features of the temperature and the current time step. S5. The deep learning recognition network is trained using a dual-supervised loss function that includes weighted temporal focus loss and attention consistency regularization term. S6. Process the original stress time series data and ambient temperature data of the node to be identified sequentially through steps S2-S4, and input the obtained enhanced feature sequence and external physical features into the trained deep learning recognition network to output the stress state category of the node to be identified.

[0021] Specifically, the method of the present invention includes: S1. Data acquisition for stress monitoring at curtain wall connection nodes. By deploying high-precision fiber optic gratings or resistance strain gauge stress sensors at key connection nodes (such as support connection points, rod connectors, etc.) of the actual curtain wall structure, the dynamic stress response of the nodes is continuously monitored over a long period of time at a fixed sampling frequency, and the ambient temperature data is recorded simultaneously, thereby obtaining the original stress-temperature time series monitoring data.

[0022] The acquisition time for each sample is fixed to ensure data comparability. The continuous signal is eventually discretized into a time-series vector of 8192 time points, forming an original stress monitoring data sample.

[0023] To obtain the necessary labels for supervised training, this study employs a combination of finite element simulation analysis, laboratory loading experiments, and on-site expert evaluation for data annotation. Specifically, based on known curtain wall design parameters and material properties, finite element software is used to simulate the mechanical response of nodes under different load conditions (such as wind load, temperature load, gravity load, and their combinations), generating simulation data with clearly defined stress state labels. Simultaneously, typical connection node specimens are subjected to progressive loading in the laboratory, and acoustic emission or digital image correlation techniques are used to assist in determining their stress states from elasticity and yielding to local damage, recording the corresponding sensor data. In addition, for some long-term on-site monitoring data, structural engineers interpret and label the status based on historical inspection records, maintenance reports, and structural health assessment results.

[0024] Ultimately, all samples were labeled with M discrete stress state categories. These categories comprehensively cover key states of nodes, from "normal working state" and "slight stress concentration" to "potential damage warning" and "dangerous state," forming a complete dataset for subsequent model training and validation. Each sample contains an 8192-dimensional original stress time-series vector, a synchronized temperature time-series sequence, and a category label representing its true stress state.

[0025] S2, Adaptive multi-scale feature mapping of stress monitoring data The stress monitoring data of the curtain wall connection nodes exhibits multi-scale fluctuation characteristics, with obvious local stress concentration and interference from environmental noise. Conventional methods use global normalization or fixed-scale wavelet transform for feature processing, which is difficult to adapt to the local feature changes between different samples and within samples, resulting in insufficient feature discrimination of the extracted features and inability to effectively characterize complex stress states.

[0026] This invention employs a method combining dynamic temporal clustering and adaptive wavelet packet transform to extract multi-scale features from the original stress data, and performs dimensionality reduction through principal component analysis to enhance the model's sensitivity to local stress changes and latent patterns. The specific steps are as follows: 1) Segmentation based on dynamic temporal clustering Dynamic clustering of the time index divides the entire time series into multiple intervals with different statistical characteristics to adaptively capture the multi-scale time-series structure of stress data. Specifically, for the i-th stress monitoring data sample Its time index set The K-Means clustering algorithm was used to divide the data into groups. The center point index and width parameter of each time segment are obtained through clustering. This segmentation information is used to guide subsequent adaptive feature mapping to obtain the center point index of the k-th time segment. Width parameter of the k-th time segment .

[0027] in, The i-th stress monitoring data sample is an 8192-dimensional time-series vector, directly acquired by the sensor. This represents the total number of cluster segments. It is a preset hyperparameter used to determine the number of multi-scale feature basis functions generated. An example value is 10. This represents a segmented index, with a value range of 100. ; This represents the center point index of the k-th time segment. It is an integer value that represents the center position of the segment on the time axis. The width parameter of the k-th time segment is a positive real number used to describe the span of the segment on the time axis, and will be used as the scaling parameter of the wavelet basis function later.

[0028] 2) Construct an adaptive multi-scale feature mapping function By utilizing segmented information, a feature mapping function integrating adaptive weights and wavelet basis functions is constructed to map the original time-series data to a new feature space that can simultaneously reflect multi-scale fluctuations and local features, expressed as:

[0029] In the formula, The feature mapping function is used to dynamically fuse features at different time scales in the original stress time series data and map them to a new feature space that can better capture local fluctuations and patterns. This represents the original stress value collected by the i-th sample at time t; Denotes the wavelet basis function, defined as ,in Used in Centered on Extracting frequency domain features within a local time window of a given scale can effectively capture local fluctuation patterns of stress signals; This represents the adaptive weight coefficient corresponding to the k-th segment at time t, which is used to dynamically adjust the weights of the features contributed by different segments based on the data characteristics at the current time.

[0030] 3) Calculate the adaptive weighting coefficients The weighting coefficients are dynamically calculated based on the proximity of the current stress value to the statistical mean of its respective segments, enabling the feature mapping to flexibly respond to local changes in the data, as expressed below:

[0031] In the formula, This represents the feature contribution weight corresponding to the k-th time segment at time t. Its value is between 0 and 1, and the summation over all segments is 1. This represents the weight attenuation coefficient, a preset positive hyperparameter used to adjust the sensitivity of the current stress value to the difference between the current stress value and the piecewise mean in the weight calculation. The larger the value, the more the weight will be biased towards the segment mean that is closest to the current value; the example value is 0.5. This represents the statistical mean of all stress data within the k-th time segment, used to represent the typical stress level of that segment; Represented by natural constant An exponential function with base 0; Indicates difference from Segmented index.

[0032] 4) Forming the mapping feature vector By performing feature mapping on all time points of the sample, a new feature vector with the same dimension as the original data is obtained, represented as:

[0033] In the formula, The mapped feature vector of the i-th sample is a new feature vector obtained after adaptive multi-scale feature mapping, with dimension 1. It contains enhanced multi-scale information; This indicates the transpose operation.

[0034] 5) Principal Component Analysis Dimensionality Reduction Principal component analysis is performed on the mapped feature vectors of all training samples to retain the main variance components, thereby achieving feature dimensionality reduction, reducing computational complexity, and filtering noise. This is represented as:

[0035] In the formula, Let represent the dimensionality-reduced feature vector of the i-th sample, with dimension . ; This represents the projection matrix obtained from principal component analysis, with dimensions of... Its column vectors are the frontmost values ​​of the covariance matrix of the training sample set. A principal feature vector is used to project high-dimensional features onto a low-dimensional principal component space; The feature dimension after dimensionality reduction is expressed as follows: The preset hyperparameter determines the number of principal components to retain; an example value is 50.

[0036] S3. Feature Enhancement Based on Gaussian Mixture Model Even after dimensionality reduction, the feature vector may still be affected by intra-class feature differences and residual noise. Conventional feature standardization methods only linearly scale the features, ignoring the distribution structure of the data in the feature space, resulting in insufficient feature discriminative power.

[0037] This invention employs a Gaussian mixture model to probabilistically model the spatial distribution of dimensionality-reduced features. The posterior probability of each sample belonging to each Gaussian component is used as an additional distribution feature and concatenated with the original dimensionality-reduced features to construct a more discriminative enhanced feature vector. The specific steps are as follows: 1) Fitting Gaussian mixture model parameters Assume the reduced feature space is composed of The stress is composed of a mixture of multiple Gaussian distribution components, each corresponding to a potential stress state mode. An expectation-maximization algorithm is used, based on the dimensionality-reduced features of all training samples. If the parameters of the fitted Gaussian mixture model are obtained, then the model parameter set is: ; in, This represents the total number of mixture components in the Gaussian mixture model. It is a preset hyperparameter that corresponds to the number of potential different stress state modes in the feature space. Its value can be determined by the Bayesian information criterion, and the example value is 8. This represents the index of the c-th component in the mixture, with a value ranging from 1 to 2. ; Let the mixing weight of the c-th Gaussian component satisfy the following condition: and This is used to characterize the prior importance of the component in the entire mixture model; Let c be the mean vector of the c-th Gaussian component, with dimension 1. Used to represent the component in The center position in the 3D feature space; Let c be the covariance matrix of the c-th Gaussian component, with dimension 1. is a symmetric positive definite matrix used to describe the distribution of eigenvalues ​​within the component and the correlation between its dimensions.

[0038] 2) Calculate the posterior probability that the sample belongs to each component. For each sample's dimensionality-reduced features, its rank (belonging to the 1st) is calculated using the fitted Gaussian mixture model parameters. The posterior probability of each Gaussian component, reflecting the correlation strength between sample features and each potential stress state mode, is expressed as:

[0039] In the formula, Represents the dimensionality reduction features of the i-th sample. Given the given conditions, the posterior probability that the sample belongs to the c-th Gaussian component satisfies the following condition: ; The mean is The covariance matrix is The probability density function of the multivariate Gaussian distribution is expressed as follows: ; Represents the covariance matrix The determinant; Represents the covariance matrix The inverse matrix.

[0040] 3) Construct the posterior probability feature vector The sample belongs to all The posterior probabilities of the Gaussian components are arranged in order to form A probability distribution vector of dimension is represented as:

[0041] In the formula, Let represent the posterior probability feature vector of the i-th sample. The column vector, whose The elements are This vector encodes the soft assignment information of the sample on the potential stress state mode.

[0042] 4) Generate enhanced feature vectors The original dimensionality-reduced features are vertically concatenated with the posterior probability distribution vector to form the final enhanced feature vector. This vector contains both the original feature information of the sample and its soft assignment information in the fitted distribution, expressed as:

[0043] In the formula, This represents the enhanced feature vector of the i-th sample, with dimension . .

[0044] S4. Construct a deep learning recognition network The stress state evolution of curtain wall connection nodes is a complex temporal process. Its state depends not only on the current stress characteristics, but also on the coupling influence of historical trends and external environmental excitations such as temperature and wind pressure. Conventional deep learning models, such as fully connected networks, only classify static features and ignore the temporal dependence of stress data and its physical background, resulting in insufficient modeling ability of the correlation between the state evolution process and external excitations.

[0045] This invention constructs a deep recognition network that integrates external physical features and adaptive temporal attention. This network takes enhanced feature sequences and constructed external physical features as input, captures temporal dependencies through gated recurrent units, and introduces an adaptive attention mechanism that integrates external physical information. It dynamically focuses on key time steps that are highly correlated with external excitations and exhibit dramatic changes in internal features, thereby achieving accurate identification of stress states. The specific steps are as follows: 1) Construction and fusion of external physical features Considering that the stress at curtain wall nodes is significantly affected by the thermal expansion and contraction effect caused by changes in ambient temperature, synchronously monitored temperature time-series data is introduced, and a mutual information weighted feature between the temperature and stress enhancement characteristics is constructed to quantify the time-varying influence of environmental excitation on the stress state. Specifically, for the first For each sample, its enhanced feature vector sequence is represented as: ,in This represents the total number of time steps. Furthermore, the synchronously acquired temperature time-series data is ; Furthermore, at each time step, the mutual information between the temperature and each dimension of the corresponding enhanced feature vector is calculated, and this information is used as weights to generate temperature-stress coupled features, expressed as:

[0046] In the formula, Indicates the first The sample at the th The external physical features constructed by the time step are scalars; Indicates the first The sample at the th Enhanced feature vectors at time steps The One portion, The range of values ​​is ; Indicates the first Time step, enhancing feature number Dimensional components With temperature value The normalized mutual information weights between them are calculated as follows: ; Represents the mutual information estimation function; Indicates the first The sample at the th The ambient temperature value monitored at the time step; This represents the rate of temperature change between adjacent time steps, calculated as follows: It is used to capture the thermal stress impact caused by drastic temperature changes; The temperature change sensitivity coefficient is a preset positive hyperparameter used to adjust the influence of the rate of temperature change on the feature. An example value is shown below. ; This represents the hyperbolic tangent activation function, used to map the rate of temperature change to... Use a range to avoid values ​​that are too large.

[0047] Furthermore, the external physical features and the enhancement features are concatenated along the feature dimension to form a fused feature vector. Its dimensions are .

[0048] 2) Adaptive temporal attention mechanism To dynamically evaluate the importance of features at different time steps for the final stress state classification, an adaptive attention mechanism is adopted to fuse the hidden state of the gated recurrent unit, the posterior probability feature vector obtained from the previous step, and the external physical features of the current time step to jointly calculate the attention weight.

[0049] Assuming that after processing through a gated loop unit, a temporal hidden state sequence is obtained. ,in The dimension is , Let be the dimension of the hidden layer of the gated loop unit. Then the ... Attention score at each time step The calculation is as follows:

[0050] In the formula, Indicates the first The first sample Attention score at each time step; This represents a trainable weight vector with dimension . ; The weight matrix, representing the forces acting on the hidden state, is a trainable parameter with dimension O(n). ; The weight matrix, representing the weights applied to the posterior probability features, is a trainable parameter with dimension O(n). ; The hidden state bias vector is a trainable parameter with dimension . ; Indicates the first The sample at the th The posterior probability feature vector corresponding to the time step; The external feature amplification factor is a preset positive hyperparameter used to emphasize the importance of time steps where external physical features are significant. An example value is shown below. ; Indicates the first External physical feature scalar of a sample The absolute value; This represents the entropy regularization coefficient, a preset positive hyperparameter used to encourage the model to focus on time steps with high uncertainty in the probability distribution. An example value is shown below. ; Represents the posterior probability feature vector Information entropy, calculated as follows .

[0051] Furthermore, all weight scores are normalized to obtain the final attention weights. .

[0052] 3) Attention-weighted aggregation and state recognition The calculated attention weights are used to sum the hidden states of the gated recurrent units at all time steps to obtain a context vector summarizing the key information of the entire time series. Finally, stress state recognition is achieved through a classification layer, as shown below:

[0053] Furthermore, the context vector is input into the fully connected classification layer to obtain the stress state category. The predicted probability distribution is expressed as:

[0054] In the formula, Indicates the first The context vector of each sample, with dimensions of... ; The model predicts the first The stress state category probability vector for each sample, with dimension . , This represents the total number of stress state categories. This represents the trainable weight matrix of the classification layer, with dimension . ; This represents the trainable bias vector of the classification layer, with dimension . ; This represents the normalized exponential function.

[0055] S5. Model Training and Loss Function Calculation To effectively train the constructed deep learning model, a loss function that can simultaneously optimize feature representation, temporal dependency modeling, and final classification performance is required. Conventional cross-entropy loss only focuses on the classification boundary and does not consider the consistency of intra-class temporal patterns in curtain wall stress data or the physical meaning of the model's focus on key time steps, resulting in insufficient robustness and interpretability of the model.

[0056] This invention employs a dual-supervised loss function, comprising a weighted temporal focus loss and an attention-consistency-based regularization term. This loss function forces the model to make correct classifications while ensuring that the features corresponding to its high attention weight time steps have higher intra-class compactness and pattern attribution clarity in the latent feature space, thereby improving the reliability and physical interpretability of the recognition results. The specific steps are as follows: 1) Weighted time series focus loss To address the potential imbalance in the number of samples and the differences in classification difficulty among samples in stress state categories, an adaptive weight is introduced on the basis of the standard cross-entropy loss. This weight consists of two parts: a class balancing weight and a temporal focus weight. The latter is dynamically adjusted based on the attention weight of each time step and the classification difficulty of the samples at that time step. Specifically, for the first For each sample, the true label is: The model predicts the probability vector as follows: The weighted time series focus loss is defined as:

[0057] In the formula, Indicates the weighted time-series focus loss; This represents the total number of samples in the training batch; Indicate category The balancing weights are usually calculated based on the inverse frequency of the class in the training set, and are used to alleviate the class imbalance problem. Indicates the first The sample at the th Attention weights at time steps; Represents the prediction probability vector Corresponding to the real category The amount; The attention weight sharpening index is greater than 100%. The preset positive hyperparameters are used to amplify the contribution of high-attention time steps, forcing the model to concentrate classification confidence on a few key steps. Examples of values ​​are provided. ; This represents the focus adjustment parameter, which is a preset non-negative hyperparameter. The model will pay more attention to difficult samples with low classification probabilities, with the following example values: .

[0058] 2) Attention consistency regularization To ensure that the model's attention mechanism focuses on feature patterns with clear physical meaning and strong discriminative power, a regularization term based on Gaussian mixture model components is adopted. This regularization term encourages a more explicit posterior probability distribution corresponding to high attention time steps, and its dimensionality-reduced features are closer to the latent pattern centers to which they belong, expressed as:

[0059] In the formula, This represents a regularization term for attention consistency. Indicates the first Time step posterior probability feature vector The information entropy, by minimizing this, encourages a more certain attribution of patterns in high-attention time steps; Indicates the first The Gaussian component index to which the time step feature most likely belongs is calculated as follows: ; In the Gaussian mixture model, the first... The mean vector of each component ; Indicates the first The sample at the th Dimensionally reduced feature vectors at each time step; The feature compactness penalty coefficient is a preset positive hyperparameter used to balance the contributions of the entropy term and the feature distance term. An example value is given below. ; This represents the L2 norm.

[0060] Furthermore, the total loss function for model training is a weighted sum of the above two terms, expressed as:

[0061] In the formula, Represents the total loss function; This represents the regularization strength coefficient, a preset positive hyperparameter used to control the impact of attention consistency constraints on the overall optimization objective. An example value is shown below. .

[0062] S6, Deep Learning Recognition Network Training After completing data collection, feature extraction and enhancement, and network structure construction, the training phase of the deep learning recognition network begins.

[0063] This stage utilizes the training dataset collected and labeled in step S1, and iteratively optimizes all trainable parameters in the network according to the data processing and network forward propagation process defined in S2 to S4. The training process employs a gradient-based optimization algorithm. The fused features and prediction results obtained from forward propagation are combined with the total loss function defined in S5, and the gradient of the loss function with respect to the parameters of each layer is calculated using the backpropagation algorithm. The parameters are then updated to minimize the total loss.

[0064] During training, data is input in batches. Each batch consists of a random number of samples drawn from the training set. The batch then sequentially performs adaptive multi-scale feature mapping, dimensionality reduction, Gaussian mixture model feature enhancement, external physical feature fusion, gated recurrent unit temporal encoding, adaptive attention weighting, and final classification prediction. The sum of the weighted temporal focus loss and the attention consistency regularization term for that batch is calculated as the total loss, and parameter updates are then performed. This iterative process continues until a preset stopping condition is met.

[0065] There are usually two stopping conditions: one is to reach the pre-set maximum number of training epochs to ensure sufficient training; the other is to monitor the performance on an independent validation set, and when the loss function value on the validation set no longer decreases or starts to increase in several consecutive training epochs, i.e., when an overfitting trend occurs, training is stopped early to preserve the model parameters with the best generalization ability.

[0066] Based on this, the training process outputs a deep learning recognition network model with optimized parameters that can effectively identify nodal stress states from raw stress and temperature data.

[0067] S7. Stress State Identification of Curtain Wall System Connection Nodes Once the deep learning recognition network is trained, it can be applied to the online or offline stress state identification of connection nodes in actual curtain wall systems.

[0068] For monitoring data of new nodes or time periods to be identified, the data processing is first performed according to the exact same procedure as in the training phase: raw stress time-series data and synchronous temperature data of the node within a fixed duration are collected to form an 8192-dimensional raw sample. Then, strictly following step S2, the sample is subjected to adaptive multi-scale feature mapping and dimensionality reduction using the dynamic time clustering segmentation parameters and principal component analysis projection matrix determined in the training phase, resulting in its dimensionality-reduced feature vector. Then, based on the Gaussian mixture model parameters fitted to the training set in step S3, the posterior probability feature vector of the sample is calculated and concatenated with the dimensionality-reduced feature vector to generate an enhanced feature vector. Finally, following step S4, the corresponding external physical features of the sample are constructed and input together with the enhanced feature sequence into the pre-trained deep learning recognition network.

[0069] The network's forward propagation process automatically completes temporal encoding, adaptive attention calculation to integrate external information, and aggregation of key information. Finally, the network output layer (Softmax classification layer) provides a probability distribution vector for M stress state categories, with the category with the highest probability being identified as the stress state of the curtain wall connection node during the current monitoring period. This identification result can be used to trigger early warnings, guide maintenance decisions, or assess the overall structural safety of the curtain wall system, achieving end-to-end intelligent identification from multi-source monitoring data to a clear stress state level.

[0070] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a deep learning-based method for identifying the stress state of connection nodes in a curtain wall system as described above.

[0071] Example 1 In specific embodiments, the identification performance of different methods on five typical stress state categories is evaluated. For example... Figure 2As shown in the figure, the experiment compares four methods: fixed-scale wavelet transform combined with traditional classification methods, global normalization combined with gated recurrent unit networks, traditional convolutional neural network methods, and the complete method proposed in this invention. Fixed-scale wavelet transform represents traditional signal processing methods, using fixed-scale wavelet basis functions to extract frequency domain features before inputting them into a support vector machine for classification; global normalization combined with gated recurrent unit networks represents traditional deep learning temporal methods, only globally standardizing the input data before using a gated recurrent unit network for modeling; traditional convolutional neural network methods use standard convolutional layers and pooling layers to automatically extract features. The method of this invention includes complete adaptive multi-scale feature mapping, Gaussian mixture model feature enhancement, and an adaptive attention mechanism that integrates external physical features. In the experimental configuration, all methods used the same training dataset, containing five categories of stress state samples with a balanced number of samples in each category, and the test set was independently partitioned. The horizontal axis in the figure represents the five stress state categories, from normal working state to severely damaged state, with the recognition difficulty gradually increasing; the vertical axis represents the recognition accuracy, expressed as a percentage but labeled as a decimal in the figure. As can be clearly seen from the bar chart, the recognition accuracy of the method of the present invention is significantly higher than that of the other three methods in all categories. This reflects that the method of the present invention can better capture local stress concentration features through adaptive multi-scale feature mapping, improve feature discriminativeness through Gaussian mixture model feature enhancement, and enhance the modeling ability of complex stress states through attention mechanism that integrates external physical features.

[0072] Example 2 In specific embodiments, the working process of the adaptive temporal attention mechanism in this invention is visualized and analyzed, such as... Figure 3As shown in (a)-(c), this demonstrates how the model automatically focuses on key segments in the stress monitoring time series signal. The experiment used a complete stress monitoring sample containing multiple local fluctuations and noise, and simultaneously provided ambient temperature data. The trained model calculated and output the attention weights at each time point. The horizontal axis represents continuous "time" in "seconds". The results are presented in three vertically aligned subplots, showing the "original stress monitoring time series signal", "synchronized ambient temperature monitoring data", and "adaptive attention mechanism weight distribution". The vertical axis for the stress signal is in "megapascals", the vertical axis for the temperature data is in "degrees Celsius", and the vertical axis for the attention weights represents the relative importance score without a specific unit. In the upper stress signal subplot, the signal exhibits an overall fluctuating pattern with several high-amplitude local pulse peaks. Two main pulse regions are highlighted with semi-transparent blocks. The middle temperature subplot shows that the temperature fluctuates periodically over time. The bottom attention weight subplot shows that the weights are not uniformly distributed but form several significant spikes. The locations of these spikes highly coincide with the highlighted pulse regions in the stress signal and are also correlated with the inflection point regions of the temperature change curve. For example, at the pulse peaks around 3 seconds and 7 seconds on the horizontal axis, the attention weights reach local maxima, demonstrating that the attention mechanism of this invention can effectively "ignore" stationary and noisy periods, automatically allocating computational resources and high-importance scores to critical time steps that may significantly affect stress concentration, potential damage, or environmental stimuli. This capability enhances the interpretability of the model and makes its decisions more focused on physically meaningful signal features.

[0073] This invention also achieves the following technical effects: First, it achieves higher accuracy in various stress state identification tasks, especially excelling in identifying challenging conditions such as minor stress concentrations and potential damage warnings. Second, the model possesses good physical interpretability; its adaptive attention mechanism automatically and appropriately highlights key time periods such as stress pulses and temperature inflection points, making the decision-making basis visible. Third, the model exhibits stronger robustness to noise and intra-class variations, thanks to multi-scale adaptive feature extraction and probability distribution-based feature enhancement strategies. Finally, it establishes an end-to-end intelligent identification process, enabling automated and engineerable deployment from raw multi-source monitoring data to clearly defined state levels.

[0074] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0075] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying the stress state of connection nodes in a curtain wall system based on deep learning, characterized in that, Includes the following steps: S1. Collect the original stress time series data and synchronous ambient temperature data of the curtain wall connection nodes to form a sample set, and label the corresponding stress state category of each sample. S2. For the original stress time series data of each sample, segmentation is performed based on dynamic time clustering, and feature mapping is performed using an adaptive feature mapping function that combines segmentation information with wavelet basis functions to obtain the mapped feature vector. Then, dimensionality reduction is performed through principal component analysis to obtain the dimensionality-reduced feature vector. S3. Fit a Gaussian mixture model using the dimensionality-reduced feature vector, calculate the posterior probability of each sample's dimensionality-reduced feature vector belonging to each Gaussian component, form a posterior probability feature vector, and concatenate the dimensionality-reduced feature vector with the posterior probability feature vector to generate an enhanced feature vector. S4. Construct a deep learning recognition network, including a gated recurrent unit for temporal encoding, an adaptive temporal attention mechanism that fuses external physical information, and a classification layer; the adaptive temporal attention mechanism calculates attention weights by fusing the hidden state of the gated recurrent unit, the posterior probability feature vector, and the external physical features of the current time step. S5. The deep learning recognition network is trained using a dual-supervised loss function that includes weighted temporal focus loss and attention consistency regularization term. S6. Process the original stress time series data and ambient temperature data of the node to be identified sequentially through steps S2-S4, and input the obtained enhanced feature sequence and external physical features into the trained deep learning recognition network to output the stress state category of the node to be identified.

2. The method for identifying the stress state of connection nodes in a curtain wall system based on deep learning according to claim 1, characterized in that, In step S2, an adaptive feature mapping function combining piecewise information and wavelet basis functions is used to perform feature mapping, resulting in a mapped feature vector. Then, principal component analysis is used for dimensionality reduction to obtain a dimensionality-reduced feature vector. Specifically, this includes: K-Means clustering was performed on the time index set of the original stress time series data of the samples, dividing it into K time segments, and obtaining the center point index and width parameter of each segment. The segment information was used to construct a feature mapping function that integrates adaptive weights and wavelet basis functions. This feature mapping function extracts frequency domain features within a local time window through wavelet basis functions and dynamically adjusts the feature contribution of different segments in combination with adaptive weight coefficients to generate a mapped feature vector. Then, dimensionality reduction was performed through principal component analysis to obtain the dimensionality-reduced feature vector.

3. The method for identifying the stress state of connection nodes in a curtain wall system based on deep learning according to claim 2, characterized in that, The adaptive weighting coefficient The calculation is based on the proximity of the current stress value to the statistical mean of each segment. The weight value decreases as the difference increases, and the sum is ensured to be 1 through normalization.

4. The method for identifying the stress state of connection nodes in a curtain wall system based on deep learning according to claim 1, characterized in that, In step S3, a Gaussian mixture model is fitted using dimensionality-reduced feature vectors to ultimately generate enhanced feature vectors, specifically including: Using the expectation-maximization algorithm, based on the dimensionality-reduced features of all training samples... If a Gaussian mixture model containing C multivariate Gaussian distribution components is fitted, then the model parameter set is: For each sample's dimensionality-reduced features, its rank (belonging to the 1st) is calculated using the fitted Gaussian mixture model parameters. The posterior probability of each Gaussian component; classifying a sample as belonging to all... The posterior probabilities of the Gaussian components are arranged in order to form... The original dimensionality-reduced feature vector is concatenated vertically with the posterior probability feature vector to form the final enhanced feature vector. .

5. The method for identifying the stress state of connection nodes in a curtain wall system based on deep learning according to claim 1, characterized in that, In S4, the adaptive temporal attention mechanism calculates attention weights by fusing the hidden state of the gated recurrent unit, the posterior probability feature vector, temperature, and external physical features of the current time step. Specifically, it includes: Constructing external physical features The feature is based on the mutual information weighting of the enhanced features and temperature data, and is scaled in combination with the temperature change rate; the attention score is calculated by fusing the latent state, posterior probability feature vector and external physical features, and entropy regularization is introduced to encourage attention to time steps with high uncertainty; finally, the attention score is normalized to obtain the attention weight.

6. The method for identifying the stress state of connection nodes in a curtain wall system based on deep learning according to claim 1, characterized in that, In S5, the dual-supervised loss function includes a weighted temporal focus loss. Attention Consistency Regularization The total loss function is a weighted sum of the two, where the weighted temporal focus loss is used to handle class imbalance and focus-difficult samples, and the attention consistency regularization term is used to improve the physical interpretability of attention.

7. The method for identifying the stress state of connection nodes in a curtain wall system based on deep learning according to claim 6, characterized in that, The weighted time-series focus loss The calculation is based on class balance weights and temporal attention weights, and the focus is enhanced on samples that are difficult to classify by using focus adjustment parameters.

8. The method for identifying the stress state of connection nodes in a curtain wall system based on deep learning according to claim 7, characterized in that, The attention consistency regularization term The robustness of network model decisions can be enhanced by encouraging the clarity of the posterior probability distribution and the compactness of features in high-attention time steps.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based method for identifying the stress state of connection nodes in a curtain wall system as described in claims 1-8.

Images