Glass curtain wall delamination loosening terahertz amplitude ratio early warning system

Abstract

The present application relates to a glass curtain wall delamination loose terahertz amplitude ratio early warning system, and in particular to the technical field of building safety warning, which models the glass panel as a graph structure with physical connection weight and multi-dimensional features, uses gated temporal convolution and propagation gated graph convolution to learn the spatio-temporal evolution hidden state of delamination risk and output risk prediction probability, constructs a directed causal graph through Granger causality test and identifies the source and involved panels, and performs causal fusion of the directed causal graph and risk prediction probability to generate a three-level early warning instruction. Thus, the system jumps from single-point threshold alarm to collaborative warning based on spatial topology and causal propagation chain, can perceive the hazard diffusion trend in advance, actively locks the high-risk area and generates differentiated disposal instructions, effectively overcomes the defects that the prior art cannot capture the risk coupling and fault propagation between panels, and significantly improves the accuracy and timeliness of curtain wall safety warning.

Description

Technical Field

[0001] This invention relates to the field of building safety early warning technology, and more specifically, to a terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls. Background Technology

[0002] Large commercial complexes, super high-rise buildings, and landmark public facilities widely use glass curtain walls as their external envelope. The curtain wall area of ​​super high-rise buildings can reach tens of thousands of square meters, consisting of thousands or even tens of thousands of glass panels bonded to an aluminum alloy frame using structural sealant. During long-term service, glass curtain walls withstand multiple stresses: externally, in coastal areas, they bear the high wind speeds of typhoons, while in inland areas, large diurnal temperature variations and ultraviolet radiation cause the sealant's molecular chains to break and age; internally, they experience creep settlement and temperature stress in the main building structure. The expansion and contraction of the keel and the repeated positive and negative pressures caused by wind loads cause the sealant bonding interface to continuously bear the combined stress of shear and peeling. When the sealant at the edge of a glass panel delaminates and loosens due to aging or construction defects, the wind load and temperature stress originally shared by that panel will be redistributed to the adjacent panels. This causes the sealant of the adjacent panels to bear additional shear force exceeding the design value, forming a stress concentration effect. This stress redistribution is common in large curtain wall systems and shows a non-linear accelerating trend as the number of delaminated panels increases, which can easily lead to a chain failure of the curtain wall.

[0003] Currently, detection technologies for delamination and loosening of glass curtain walls are mainly divided into two categories: contact and non-contact. Contact methods, such as manual tapping and listening, and displacement sensor attachment monitoring, are inefficient and difficult to cover high-altitude areas. Non-contact methods include infrared thermal imaging, laser Doppler vibration measurement, and terahertz time-domain spectroscopy. Among these, terahertz detection technology, with its high sensitivity to defects at multi-layer media interfaces, can identify the delamination state of the sealant by analyzing changes in the amplitude ratio of reflected or transmitted signals. However, existing terahertz detection systems, in practical applications, only perform independent threshold judgments on the amplitude ratio data of a single glass pane; once the amplitude ratio of a certain pane exceeds a preset threshold, an alarm is triggered. The single-point isolated alarm mode has fundamental flaws: it cannot capture the spatial topological relationship between panels and ignores the risk coupling effect caused by stress transmission between adjacent panels; it lacks time series analysis capabilities and cannot use historical detection data to predict the trend of fault propagation; therefore, when an early sign of delamination is detected in a panel, maintenance personnel cannot know which adjacent panels will have a significantly increased risk level in the next few months, and can only passively wait for the next inspection to find new panels that exceed the standard; this post-event response early warning mode makes it difficult to fundamentally prevent large-scale chain-reaction detachment accidents, which seriously restricts the engineering application value of terahertz amplitude ratio detection technology in the field of curtain wall safety early warning. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls. The system utilizes a topology graph construction module, a spatiotemporal graph convolution module, a causality verification module, and an early warning decision module to solve the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically including: Topology graph construction module: Each glass panel is treated as a node, and the keel connection between adjacent glass panels is used as an edge. The edge weights are assigned according to the physical connection relationship between the glass panels. At the same time, a multi-dimensional feature vector is extracted for each node. The multi-dimensional feature vector includes the edge-center terahertz amplitude ratio of the current inspection, the trend slope of the historical edge-center amplitude ratio, and the environmental exposure orientation code. The edge-center amplitude ratio values ​​of the past inspections are accumulated to form a historical edge-center amplitude ratio sequence. The feature vectors of the past inspections are stacked to form a node feature tensor, thereby constructing the curtain wall topology graph. Spatiotemporal graph convolution module: Receives the curtain wall topology graph and node feature tensors, and learns the spatiotemporal evolution hidden state of the risk of delamination between glass panels by stacking gated temporal convolution and spatial graph convolution and embedding a propagation gating mechanism with physical constraints. Based on the spatiotemporal evolution hidden state, it outputs the risk prediction probability of each glass panel within a future preset time step. Causality test module: Performs pairwise conditional Granger causality test on the historical edge-center amplitude ratio sequence, constructs a directed causal graph and quantifies the causal strength between each glass panel to identify the source glass panel and the affected glass panel. Early warning decision module: It integrates the directed causal graph with the risk prediction probability, calculates the final early warning probability of each glass panel, and generates a three-level early warning instruction based on the final early warning probability, including high-risk immediate alarm, encrypted monitoring instruction for surrounding glass panels, and source tracing early warning.

[0006] In a preferred embodiment, the specific operation of assigning edge weights based on the physical connection relationship between glass panels in the topology graph construction module is as follows: First, set the initial edge weights according to the type of shared keel between adjacent glass panels. The edge weights of adjacent glass panels sharing vertical keels are set to the first preset value, the edge weights of adjacent glass panels sharing horizontal keels are set to the second preset value, the edge weights of diagonally adjacent glass panels are set to the third preset value, and the edge weights of non-adjacent glass panels are set to zero. Then, normalized mutual information is calculated based on the historical edge-center amplitude ratio sequence of each pair of adjacent glass panels in recent inspections to quantify the nonlinear dependence of the degradation process of the two glass panels. Finally, the normalized mutual information is multiplied by a preset adjustment coefficient and then added to the initial edge weights to obtain the corrected dynamic edge weights.

[0007] In a preferred embodiment, the specific operation for extracting the multidimensional feature vector of each node in the topology graph construction module is as follows: The multidimensional feature vector includes the edge-to-center terahertz amplitude ratio of the current inspection, the trend slope of the historical edge-to-center amplitude ratio, the environmental exposure orientation code, the spectral distortion index, and the glass panel size specification code. Among them, the trend slope of the historical edge-center amplitude ratio is obtained by robust linear regression on the historical edge-center amplitude ratio sequence of multiple past inspections; The environmental exposure orientation coding is based on the angle between the facade where the glass panel is located and the local prevailing wind direction. It uses a cosine function to quantize the segmented structure, with the windward side taking the first coding value, the leeward side taking the second coding value, and the crosswind side taking the middle coding value. The spectral distortion index is calculated based on the power spectral density of the terahertz echo signal obtained when the edge-to-center terahertz amplitude ratio is collected, to capture the changes in scattering characteristics caused by microstructural damage to the sealant; the glass panel size specification code is represented by the normalized logarithmic value of the glass panel area. Meanwhile, the feature vectors of the current inspection and the past multiple inspections are stacked in chronological order to form a node feature tensor. The dimension of the node feature tensor is determined by the number of glass panels, the number of time steps, and the feature dimension.

[0008] In a preferred embodiment, the specific operation of learning the spatiotemporal evolution hidden states in the spatiotemporal graph convolution module is as follows: First, for the feature vector sequence corresponding to each glass panel in the node feature tensor, stacked gated dilated causal convolutional layers are used to extract the time dependency. Each layer contains a hyperbolic tangent convolutional branch and a sigmoid gated branch. The two are multiplied by element-wise to achieve non-linear gating. The dilation rate increases exponentially with the number of layers to expand the time receptive field. Then, based on the hidden state obtained after temporal convolution, spatial message passing is carried out using the curtain wall topology graph to calculate the hazard intensity index of each glass panel. The hazard intensity index maps the deviation of the edge-center terahertz amplitude ratio of the current inspection relative to the preset micro-hazard threshold to between zero and one through the Sigmoid function. By combining dynamic edge weights and normalized mutual information, the propagation gating coefficient is calculated, which significantly amplifies the message transmission weight of the source glass panel. Next, the aggregated messages of neighboring glass panels are aggregated using a weighted average method, with the weights obtained by normalizing the propagation gating coefficients. Finally, the hidden state obtained by convolving the aggregated message with its own time is fused through a linear transformation and activation function to obtain the spatially updated hidden state.

[0009] In a preferred embodiment, the specific operation of outputting the risk prediction probability in the spatiotemporal graph convolution module is as follows: A predetermined number of modules are stacked alternately with temporal convolutional layers and spatial graph convolutional layers. Each module contains one step of temporal convolution and one step of spatial graph convolution. Residual connections are introduced during the stacking process. The input and output of each module are added together and used as the input of the next module to alleviate gradient vanishing. After passing through a preset number of modules, the final spatiotemporal evolution hidden state is obtained. The spatiotemporal evolution hidden state is then input into a two-layer fully connected network. The first layer uses a linear transformation and the ReLU activation function, and the second layer uses a linear transformation and the Sigmoid activation function. The output is a risk prediction probability between zero and one, which represents the possibility that the glass panel will delaminate and loosen within a preset time step in the future.

[0010] In a preferred embodiment, the specific operation of performing the pairwise conditional Granger causality test in the causality testing module is as follows: First, using the risk prediction probability output by the spatiotemporal graph convolution module and the dynamic edge weight output by the topology graph construction module, glass panels with risk prediction probabilities exceeding a preset first threshold and their first-order neighbors, as well as glass panel pairs with normalized mutual information exceeding a preset second threshold output by the topology graph construction module, are selected to form a candidate causal pair set. Then, for each pair of glass panels in the candidate causal pair set, a conditional Granger causality test based on kernel function is adopted. By mapping the historical edge-center amplitude ratio sequence to the regenerated kernel Hilbert space, the kernelized conditional mutual information is calculated as the test statistic, and an approximate F statistic is constructed based on the residual variance ratio of the kernelized ridge regression to determine whether there is a causal relationship from one glass panel to another. Finally, for glass panel pairs that are determined to have a causal relationship, the causal strength is calculated. The causal strength is obtained by fusing the ratio of the F-statistics of the two-way test, the difference of the kernelized conditional mutual information, and the normalized mutual information. The ratio of the F-statistics of the two-way test is the ratio of the approximate F-statistic from the first glass panel to the second glass panel to the approximate F-statistic from the second glass panel to the first glass panel. The difference of the kernelized conditional mutual information is the difference between the kernelized conditional mutual information from the first glass panel to the second glass panel and the difference between the kernelized conditional mutual information from the second glass panel to the first glass panel, thereby constructing a directed causal graph.

[0011] In a preferred embodiment, the specific operation of identifying the source glass panel and the affected glass panel in the causality verification module is as follows: Based on the directed causal graph, calculate the out-degree and in-degree of each glass panel, as well as the weighted out-degree and weighted in-degree; then, define the source index, which is obtained by multiplying the ratio of the difference between the weighted out-degree and the sum of the weighted in-degrees by the ratio of the weighted out-degree to the weighted out-degree plus one. When the source index is greater than the preset third threshold and the output degree is greater than the preset first quantity, the glass panel is marked as the source glass panel; When the source index is less than the preset fourth threshold and the in-degree is greater than the preset second quantity, the glass panel is marked as the affected glass panel.

[0012] In a preferred embodiment, the specific operation of causal fusion of the directed causal graph and the risk prediction probability in the early warning decision module is as follows: First, for each glass panel, all glass panels pointing to the glass panel are obtained as causal parent nodes based on the directed causal graph, and the risk prediction probability and causal strength output by the spatiotemporal graph convolution module of each causal parent node are obtained. Then, each causal parent node is weighted as a proportion of its causal strength to the sum of the causal strengths of all parent nodes. The weight is multiplied by the risk prediction probability of the parent node to obtain the infection probability of a single parent node. Then, the union probability of all parent nodes is calculated to obtain the penetration risk factor. Finally, the risk prediction probability of the glass panel itself is fused with the penetration risk factor through a probability multiplication correction model. The complement of the risk prediction probability is multiplied by the complement of the penetration risk factor after adjustment by the attenuation index, and then the complement is taken to obtain the final warning probability.

[0013] In a preferred embodiment, the specific operation of generating a three-level early warning instruction based on the final early warning probability in the early warning decision module is as follows: When the final warning probability is greater than the preset first alarm threshold and the edge-center terahertz amplitude ratio of the current inspection is greater than the preset second alarm threshold, a high-risk immediate alarm is generated. When the final warning probability is between the preset first alarm threshold and the preset third alarm threshold, an encrypted monitoring command for the surrounding glass panels is generated for the glass panel and all its first-order neighbors. When a glass panel is marked as the source glass panel by the causality verification module, a source tracing warning is generated.

[0014] In a preferred embodiment, the specific operations after generating the early warning instruction in the early warning decision module are as follows: All warning instructions for glass panels are summarized by building floors and facades, and a risk heat map with superimposed directed causal arrows is generated. The color intensity of each glass panel in the risk heat map indicates the magnitude of the final warning probability, and the directed arrows indicate the causal direction and intensity. High-risk real-time alarms are sent to designated maintenance personnel via SMS, encrypted monitoring instructions are written into the preset inspection plan database, and source tracing and early warning are submitted in the form of special reports, thereby realizing the transformation of early warning information into executable maintenance actions.

[0015] The beneficial effects of this invention are as follows: By modeling the glass panel as a graph structure with physical connection weights and multidimensional features, the spatiotemporal evolution of the hidden state of delamination risk is learned using gated temporal convolution and propagation-gated graph convolution, and the risk prediction probability is output. A directed causal graph is constructed through Granger causality test to identify the source and affected panels. The directed causal graph and the risk prediction probability are causally fused to generate three-level early warning instructions. As a result, the system has leaped from single-point threshold alarm to collaborative early warning based on spatial topology and causal propagation chain. It can perceive the trend of hidden danger spread in advance, actively lock high-risk areas and generate differentiated handling instructions, effectively overcome the defects of existing technology that cannot capture risk coupling and fault propagation between panels, and significantly improve the accuracy and timeliness of curtain wall safety early warning. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, 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. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] Example 1: This example provides the following... Figure 1-2 The terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls, as shown, specifically includes: Topology graph construction module: Each glass panel is treated as a node, and the keel connection between adjacent glass panels is used as an edge. The edge weights are assigned according to the physical connection relationship between the glass panels. At the same time, a multi-dimensional feature vector is extracted for each node. The multi-dimensional feature vector contains at least the edge-to-center terahertz amplitude ratio of the current inspection, the trend slope of the historical edge-to-center amplitude ratio, and the environmental exposure orientation code. The edge-to-center amplitude ratio values ​​of the past multiple inspections are accumulated to form a historical edge-to-center amplitude ratio sequence. The feature vectors of the past multiple inspections are stacked to form a node feature tensor, thereby constructing the curtain wall topology graph. Spatiotemporal graph convolution module: Receives the curtain wall topology graph and node feature tensors, and learns the spatiotemporal evolution hidden state of the risk of delamination between glass panels by stacking gated temporal convolution and spatial graph convolution and embedding a propagation gating mechanism with physical constraints. Based on the spatiotemporal evolution hidden state, it outputs the risk prediction probability of each glass panel within a future preset time step. Causality test module: Performs pairwise conditional Granger causality test on the historical edge-center amplitude ratio sequence, constructs a directed causal graph and quantifies the causal strength between each glass panel to identify the source glass panel and the affected glass panel. Early warning decision module: It integrates the directed causal graph with the risk prediction probability, calculates the final early warning probability of each glass panel, and generates a three-level early warning instruction based on the final early warning probability, including high-risk immediate alarm, encrypted monitoring instruction for surrounding glass panels, and source tracing early warning.

[0021] In this embodiment, it is specifically necessary to explain the operation of assigning edge weights based on the physical connection relationship between glass panels in the topology graph construction module as follows: First, initial edge weights are set according to the type of shared keel between adjacent glass panels. The edge weight of adjacent glass panels sharing a vertical keel is set to the first preset value, the edge weight of adjacent glass panels sharing a horizontal keel is set to the second preset value, the edge weight of diagonally adjacent glass panels is set to the third preset value, and the edge weight of non-adjacent glass panels is set to zero. Among them, the first preset value is set to 1.0, the second preset value is set to 0.8, and the third preset value is set to 0.5. These values ​​are determined according to the ratio of the section moment of inertia of vertical keels and horizontal keels in the curtain wall design code. Vertical keels usually bear a larger bending moment, so they have the highest weight. Then, normalized mutual information is calculated based on the historical edge-center amplitude ratio sequences of each pair of adjacent glass panels in recent inspections to quantify the nonlinear dependence of the degradation process of the two glass panels. The specific calculation process of normalized mutual information is as follows: First, for a pair of adjacent glass panels, the edge-center amplitude ratio values ​​of each of their last 12 inspections are obtained, forming two time series of length 12. Next, the values ​​of each series are discretized into 10 equal-width intervals, and the number of data points in each interval is counted to obtain their respective edge probability distributions. Then, the joint probability distribution of the two sequences falling into each interval combination is counted. Mutual information is equal to the joint probability distribution multiplied by the common logarithm of the ratio of the joint probability to the edge probability product, and then summed over all interval combinations. Information entropy is equal to the edge probability multiplied by the common logarithm of the edge probability, then negative and summed. Finally, normalized mutual information is equal to mutual information divided by the sum of the two information entropies minus mutual information, and its value ranges from 0 to 1. The larger the value, the tighter the coupling of the degradation processes of the two panels. Finally, the normalized mutual information is multiplied by a preset adjustment coefficient and added to the initial edge weights to obtain the corrected dynamic edge weights. This allows the edge weights to simultaneously integrate the physical connection strength and the statistical correlation of actual monitoring data. The preset adjustment coefficient is set to 0.4, which was determined through cross-validation experiments. This value can moderately introduce data-driven statistical correlations while maintaining the dominance of physical connections. The formula for calculating the dynamic edge weights is: dynamic edge weights equal to the initial edge weights plus the adjustment coefficient multiplied by the normalized mutual information. These dynamic edge weights will replace the initial edge weights as the edge attributes of the graph structure in the subsequent spatiotemporal graph convolution module, so that the modeling of stress transfer paths not only depends on the fixed keel connection type but can also be adaptively adjusted according to the actual monitoring data. In the topology graph construction module, the specific operation for extracting the multidimensional feature vector of each node is as follows: The multidimensional feature vector includes at least the edge-to-center terahertz amplitude ratio of the current inspection, the trend slope of the historical edge-to-center amplitude ratio, the environmental exposure orientation code, the spectral distortion index, and the glass panel size specification code. The historical edge-center amplitude ratio trend slope is obtained by robust linear regression of the historical edge-center amplitude ratio sequence from multiple past inspections. The specific steps of robust linear regression are as follows: First, obtain the edge-center amplitude ratio values ​​of the glass panel in the past 8 inspections, and construct a linear regression model with the inspection number as the independent variable and the amplitude ratio value as the dependent variable. Then, replace the squared loss function of ordinary least squares with the Huber loss function. The Huber loss function uses squared loss when the absolute value of the residual is less than a preset threshold (set to 0.1), and uses linear loss when the absolute value of the residual is greater than or equal to the threshold, thereby reducing the impact of outliers. Next, iterative reweighted least squares is used to solve the regression coefficients. Each iteration updates the weights according to the current residual until the coefficients converge. The final slope is the historical edge-center amplitude ratio trend slope, with positive values ​​indicating worsening degradation and negative values ​​indicating improvement. The environmental exposure orientation code is based on the angle between the facade where the glass panel is located and the local prevailing wind direction. It is quantified piecewise using a cosine function, with the windward side receiving the first code value, the leeward side receiving the second code value, and the crosswind side receiving the middle code value. The specific quantification process is as follows: First, obtain the prevailing wind direction angle provided by the local meteorological department for the entire year. Then, measure the angle between the normal direction of the facade where the glass panel is located and the due north direction, and calculate the difference between the two to obtain the angle value. The environmental exposure orientation code is equal to 0.5 multiplied by 1 plus the cosine value of the angle. When the angle is 0 degrees (windward side), the code value is 1.0; when the angle is 180 degrees (leeward side), the code value is 0.0; and when the angle is 90 degrees (crosswind side), the code value is 0.5. This code value reflects the direct intensity of the wind load on the panel. The spectral distortion index is calculated based on the power spectral density of the terahertz echo signal obtained when the edge-to-center terahertz amplitude ratio is acquired, to capture changes in scattering characteristics caused by microstructural damage in the sealant. The specific calculation process of the spectral distortion index is as follows: First, a fast Fourier transform is performed on the terahertz echo signal to obtain the power spectral density function in the frequency range from 0.1 terahertz to 1.0 terahertz. Then, the power spectral density values ​​at all frequency points are summed to obtain the total power, and the power spectral density value at each frequency point is divided by the total power to obtain the normalized power distribution. Next, the product of the normalized power and the common logarithm of the normalized power at each frequency point is calculated, and the summation over all frequency points is taken as a negative value to obtain the spectral entropy complexity. When microcracks or delamination occur in the sealant, the echo signal will experience main frequency splitting and bandwidth broadening, resulting in a more uniform power spectral distribution and a significantly increased spectral entropy complexity, thus sensitively reflecting microstructural damage. The size specification code of the glass panel is represented by the normalized logarithm of the glass panel area. The specific calculation is as follows: First, measure the actual area of ​​the glass panel in square meters; then take the natural logarithm of the area and divide it by the natural logarithm of the largest glass panel area in the entire building to obtain a code value between 0 and 1. Larger panels bear greater bending moments under wind loads and have more obvious stress concentration, so their code values ​​are higher. Simultaneously, the feature vectors of the current inspection and the past multiple inspections are stacked in chronological order to form a node feature tensor. The dimension of this node feature tensor is determined by the number of glass panels, the number of time steps, and the feature dimension. Specifically, the number of time steps is set to 12, that is, the data of the most recent 12 inspections are retained; the feature dimension is set to 5, corresponding to the five feature components mentioned above. Therefore, the shape of the node feature tensor is the total number of glass panels multiplied by 12 and then multiplied by 5. This tensor will serve as one of the inputs to the spatiotemporal graph convolution module, providing complete spatiotemporal feature information for subsequent learning.

[0022] In this embodiment, it is specifically necessary to explain the specific operation of learning the spatiotemporal evolution hidden states in the spatiotemporal graph convolution module as follows: First, for the feature vector sequence corresponding to each glass panel in the node feature tensor, a stacked gated dilated causal convolutional layer is used to extract the temporal dependency. Each layer contains a hyperbolic tangent convolutional branch and a sigmoid gated branch. The two are non-linearly gated through element-wise multiplication. The dilation rate increases exponentially with the number of layers to expand the temporal receptive field. Specifically, three gated dilated causal convolutional layers are set. The dilation rate of the first layer is set to 1, the dilation rate of the second layer is set to 2, and the dilation rate of the third layer is set to 4. The calculation process of each layer is as follows: First, the input feature vector sequence is fed into two parallel dilated causal convolutional kernels. The output of one kernel is activated by the hyperbolic tangent function to obtain the feature branch, and the output of the other kernel is activated by the sigmoid function to obtain the gated branch. Then, the feature branch and the gated branch are multiplied element-wise to obtain the output of the layer. Through this gating mechanism, the model can selectively retain the temporal features related to the delamination evolution and suppress irrelevant noise. After three stacked layers, each glass panel obtains a compressed temporal representation with a dimension of 64. Then, based on the hidden states obtained after temporal convolution, spatial message passing is performed using the curtain wall topology map to calculate the hazard intensity index for each glass panel. The hazard intensity index maps the deviation of the edge-center terahertz amplitude ratio of the current inspection relative to the preset micro-hazard threshold to a range of zero to one using the Sigmoid function; the preset micro-hazard threshold is set to 1.2. The specific calculation process of the hazard intensity index is as follows: First, subtract the micro-hazard threshold of 1.2 from the edge-center terahertz amplitude ratio of the current inspection to obtain the difference; then, multiply the difference by the kurtosis parameter of 5.0 to obtain the scaled difference; next, take a negative exponential function with the natural constant e as the base of the scaled difference to obtain the exponent term; finally, add 1 to the exponent term and take the reciprocal to obtain the hazard intensity index; when the current amplitude ratio is less than 1.2, the hazard intensity index is close to 0; when the current amplitude ratio is greater than or equal to 1.2, the hazard intensity index rapidly approaches 1, thus quantifying the severity of the hazard source. By combining dynamic edge weights and normalized mutual information, a propagation gating coefficient is calculated. This coefficient significantly amplifies the message transmission weight of the source glass panel. The specific calculation process for the propagation gating coefficient is as follows: First, the hazard intensity index is multiplied by an amplification factor of 2.0 to obtain the amplification term. Then, the amplification term is incremented by 1 to obtain the amplification factor. Next, the dynamic edge weights are multiplied by the amplification factor and then by the normalized mutual information to obtain the propagation gating coefficient. This coefficient increases significantly when the hazard intensity index of the source glass panel is high, thereby giving the source panel a greater influence in subsequent message aggregation and simulating the physical process of stress transmission to neighbors. Next, a weighted average is used to aggregate the aggregated messages of neighboring glass panels, with the weights obtained by normalizing the propagation gating coefficients. Specifically, for each glass panel, its first-order neighbor glass panel set is first obtained; then, the propagation gating coefficients of all neighbors are summed to obtain the normalized denominator; next, for each neighbor, its propagation gating coefficient is divided by the normalized denominator to obtain the normalized weight; finally, the hidden state obtained after temporal convolution of each neighbor is multiplied by the corresponding normalized weight and summed to obtain the aggregated message vector, which has a dimension of 64; finally, the hidden state obtained after temporal convolution of the aggregated message with itself is linearly... The transformation and activation function are fused to obtain the spatially updated hidden state. The specific fusion process is as follows: First, the hidden state obtained after temporal convolution is multiplied by a learnable weight matrix to obtain the self-transformation vector; the aggregated message is multiplied by another learnable weight matrix to obtain the message transformation vector; then the self-transformation vector and the message transformation vector are added together, and a learnable bias vector is added to obtain the fusion vector; finally, the ReLU activation function is applied to the fusion vector to obtain the spatially updated hidden state, which has a dimension of 64. This hidden state contains both the node's own temporal evolution information and the spatial coupling information passed from its neighbors. In the spatiotemporal graph convolution module, the specific operation for outputting the risk prediction probability is as follows: Temporal convolutional layers and spatial graph convolutional layers are stacked alternately in a predetermined number of modules. Each module contains one step of temporal convolution and one step of spatial graph convolution. Residual connections are introduced during the stacking process. The input and output of each module are added together and used as the input of the next module to alleviate gradient vanishing. The predetermined number of modules is set to 3. The internal order of each module is as follows: first, a gated dilated causal convolution (temporal convolution) is performed, then a propagation gated graph convolution (spatial graph convolution) is performed, and then the input of the module (i.e., the output of the previous module) is added to the output of the current module element-wise to obtain the final output of the module. Residual connections allow gradients to be directly backpropagated to shallow layers, avoiding the gradient vanishing problem in deep networks. After 3 modules, the hidden state dimension of each glass panel is expanded to 128. This hidden state is called the spatiotemporal evolution hidden state, which encodes the spatiotemporal dependencies of the past 12 inspections. After passing through a preset number of modules, the final spatiotemporal evolution hidden state is obtained. This hidden state is then input into a two-layer fully connected network. The first layer uses a linear transformation and the ReLU activation function, while the second layer uses a linear transformation and the Sigmoid activation function. The output is a risk prediction probability between zero and one, representing the likelihood of the glass panel becoming delaminated or loose within a preset future time step. The specific structure of the two-layer fully connected network is as follows: the first fully connected layer linearly transforms the 128-dimensional spatiotemporal evolution hidden state into a 64-dimensional intermediate vector, and then applies the ReLU activation function; the second fully connected layer linearly transforms the 64-dimensional intermediate vector into a scalar, and then applies the Sigmoid activation function, outputting a probability value between 0 and 1. The preset future time step corresponds to an inspection interval of 1 to 3 months. The closer this probability value is to 1, the higher the risk of the glass panel becoming delaminated or loose within the future time period. The training process of the spatiotemporal graph convolutional module is as follows: First, historical inspection data is collected, including the edge-to-center terahertz amplitude ratio sequence of each glass panel, and the delamination status labels (0 indicates intact, 1 indicates delamination and loosening) verified by manual opening after each inspection. The data is divided into training set, validation set, and test set. The loss function is binary cross-entropy loss, the optimizer is Adam optimizer, the initial learning rate is set to 0.001, and it decays to 0.5 times the original rate every 10 epochs. The batch size is set to 32, and the total number of training epochs is 100. During the training process, early stopping is used simultaneously. Training is stopped when the validation set loss no longer decreases for 10 consecutive epochs to prevent overfitting. After training, the network parameters are frozen for risk prediction in actual deployment.

[0023] In this embodiment, it is specifically necessary to explain the specific operation of performing the pairwise conditional Granger causality test in the causality test module as follows: First, using the risk prediction probability output by the spatiotemporal graph convolution module and the dynamic edge weights output by the topology graph construction module, glass panels with risk prediction probabilities exceeding a preset first threshold and their first-order neighbors, as well as glass panel pairs with normalized mutual information output by the topology graph construction module exceeding a preset second threshold, are selected to form a candidate causal pair set. The preset first threshold is set to 0.5, and the preset second threshold is set to 0.6. First-order neighbors refer to glass panels directly connected to the current glass panel through the keel, i.e., neighbors with dynamic edge weights greater than zero. Through this screening, the number of glass panel pairs to be tested is reduced from the full square level to the linear level, focusing on high-risk areas and strongly coupled areas, avoiding meaningless testing of a large number of intact panels. Then, for each pair of glass panels in the candidate causal pair set, a conditional Granger causality test based on kernel functions is used. By mapping the historical edge-center amplitude ratio sequence to the regenerating kernel Hilbert space, the kernelized conditional mutual information is calculated as the test statistic, and an approximate F-statistic is constructed based on the residual variance ratio of the kernelized ridge regression to determine whether there is a causal relationship from one glass panel to another. The specific calculation process of the kernelized conditional Granger causality test is as follows: First, for a pair of glass panels, their historical edge-center amplitude ratio sequences with a lag order of 3 are obtained to form lag vectors; the lag order of 3 is determined according to the Akaike information criterion; then, a Gaussian radial basis kernel function is selected, and the kernel bandwidth parameter is set to the reciprocal of the lag vector dimension, i.e., one-sixth; next, a restricted kernel ridge regression model is constructed, which predicts the current value of the target glass panel using only the lag vector of the target glass panel to obtain the restricted residuals; An unrestricted kernel ridge regression model is then constructed, and the current value of the target is predicted using the lag vectors of the target glass panel and the lag vectors of the candidate cause glass panels to obtain the unrestricted residuals. The regularization parameter of the kernel ridge regression is selected through five-fold cross-validation to minimize the prediction error. Then, the difference between the variance of the restricted residuals and the variance of the unrestricted residuals is calculated, divided by the variance of the unrestricted residuals, and multiplied by the degree of freedom adjustment factor to obtain the approximate F-statistic. Simultaneously, the kernelized conditional mutual information is calculated based on kernel density estimation. Specifically, in the high-dimensional feature space, the log-expected value of the ratio of the conditional probability density of the current value of the target glass panel to that of the candidate cause lag vector is estimated. This statistic approximately follows an F-distribution under the null hypothesis. The null hypothesis is rejected by comparing the approximate F-statistic with the critical value (significance level of 0.05). If rejected, it is determined that there is a causal relationship from the candidate cause glass panel to the target glass panel. Finally, for glass panel pairs where a causal relationship is determined, the causal strength is calculated. The causal strength is obtained by fusing the F-statistic ratio of the two-way test, the difference in kernelized conditional mutual information, and the normalized mutual information. The F-statistic ratio of the two-way test is the ratio of the approximate F-statistic from the first glass panel to the second glass panel to the approximate F-statistic from the second glass panel to the first glass panel. The difference in kernelized conditional mutual information is the difference between the kernelized conditional mutual information from the first glass panel to the second glass panel and vice versa, thus constructing a directed causal graph. The specific fusion calculation process for the causal strength is as follows: First, the F-statistic ratio of the two-way test is used as a directional weight. A ratio greater than 1 indicates a causal relationship between the first glass panel and the second glass panel. If the influence is stronger, the opposite is true if it is less than 1; then, take the maximum value between the difference of the kernelized conditional mutual information and zero to ensure non-negativity. This difference reflects the directional strength of the nonlinear dependency; next, multiply the directional weight by the difference of the kernelized conditional mutual information, and then multiply by the normalized mutual information to obtain the causal strength; the normalized mutual information has been calculated in the topology graph construction module, and its value ranges from 0 to 1, reflecting the statistical coupling tightness of the degradation process of the two panels; finally, the causal strength ranges from 0 to 1, and the larger the value, the stronger the causal drive from the first glass panel to the second glass panel; record all glass panel pairs with significant causal relationships and their causal strengths to form a directed causal graph. Each node in the graph represents a glass panel, and each directed edge represents a causal relationship. The thickness or color of the edge can represent the magnitude of the causal strength; In the causality test module, the specific steps for identifying the source glass panel and the affected glass panel are as follows: Based on a directed causal graph, calculate the out-degree and in-degree of each glass panel, as well as the weighted out-degree and weighted in-degree. Out-degree refers to the number of directed edges originating from that glass panel and pointing to other glass panels; in-degree refers to the number of directed edges originating from other glass panels and pointing to that glass panel. Weighted out-degree is the sum of the causal strengths of all outgoing edges, and weighted in-degree is the sum of the causal strengths of all incoming edges. For example, if glass panel A has three outgoing edges with causal strengths of 0.6, 0.3, and 0.1 respectively, then its weighted out-degree is 1.0; if it has one incoming edge with a causal strength of 0.8, then its weighted in-degree is 0.8. Then, the source index is defined. The source index is obtained by multiplying the ratio of the difference between the weighted out-degree and the weighted in-degree to the sum of the weighted out-degree and the weighted out-degree plus one by the ratio of the weighted out-degree and the weighted out-degree plus one. The specific calculation process of the source index is as follows: First, calculate the difference between the weighted out-degree and the weighted in-degree, then divide it by the sum of the weighted out-degree plus the weighted in-degree plus a small constant (ten to the power of negative eight) to obtain the first factor, with a value range of negative one to positive one; then calculate the weighted out-degree divided by the weighted out-degree plus one to obtain the second factor, with a value range of 0 to 1; finally, multiply the first factor by the second factor to obtain the source index. This index comprehensively considers the net outflow direction and the absolute outflow intensity. A positive value indicates that the glass panel exerts a greater causal influence on other glass panels, and a negative value indicates that it is more affected by other glass panels. When the source index is greater than the preset third threshold and the output degree is greater than or equal to the preset first quantity, the glass panel is marked as the source glass panel; the preset third threshold is set to 0.5 and the preset first quantity is set to 2; that is, when the source index is greater than 0.5 and the output degree is greater than or equal to 2, the glass panel is identified as the source glass panel, which means that its delamination and degradation is one of the main reasons for the degradation of the surrounding glass panels. When the source index is less than the preset fourth threshold and the in-degree is greater than or equal to the preset second quantity, the glass panel is marked as an affected glass panel. The preset fourth threshold is set to -0.5 and the preset second quantity is set to 2. That is, when the source index is less than -0.5 and the in-degree is greater than or equal to 2, the glass panel is identified as an affected glass panel, which means that its degradation is mainly caused by the causal influence of multiple upstream glass panels. The identification results are used in the early warning decision module to generate source tracing early warning and surrounding glass panel encrypted monitoring instructions. Specifically, for nodes marked as source glass panels, the early warning decision module generates source tracing early warning, prompting priority to deal with the glass panel to block the propagation chain. For nodes marked as affected glass panels, even if their current risk prediction probability is not high, the surrounding glass panel encrypted monitoring instructions will be triggered, shortening the next inspection cycle of the glass panel and its first-order neighbors from the usual 3 months to 2 to 4 weeks, realizing active tracking and monitoring of potential hazard spread areas.

[0024] In this embodiment, it is specifically necessary to explain the specific operation of causal fusion of the directed causal graph and the risk prediction probability in the early warning decision module as follows: First, for each glass panel, all glass panels pointing to the glass panel are obtained as causal parent nodes based on the directed causal graph, and the risk prediction probability and causal strength output by the spatiotemporal graph convolution module of each causal parent node are obtained. The set of causal parent nodes consists of all nodes pointing to the current glass panel in the directed causal graph output by the causal verification module. The risk prediction probability of each causal parent node is output by the spatiotemporal graph convolution module, and the causal strength is output by the causal verification module, with a value range of 0 to 1. Then, the weight of each causal parent node is calculated as the proportion of its causal strength to the sum of the causal strengths of all parent nodes. This weight is multiplied by the risk prediction probability of the parent node to obtain the infection probability of a single parent node. The union probability of all parent nodes is then calculated to obtain the penetration risk factor. The specific calculation process for the penetration risk factor is as follows: First, the causal strengths of all causal parent nodes are summed to obtain the total causal strength. Then, for each causal parent node, its causal strength is divided by the total causal strength to obtain the weight of that parent node. Next, the risk prediction probability of that parent node is multiplied by this weight to obtain the infection probability of that parent node on the current glass panel. Finally, the complements of the infection probabilities of all parent nodes (i.e., one minus the infection probability) are multiplied together, and then the result of the multiplication is subtracted from one to obtain the penetration risk factor. This factor ranges from 0 to 1, representing the comprehensive penetration impact of all upstream parent nodes on the current glass panel through the causal path. If the current glass panel has no causal parent nodes, the penetration risk factor is zero. Finally, the risk prediction probability of the glass panel itself and the penetration risk factor are fused using a probability multiplication correction model. The model calculates the final warning probability by multiplying the complement of the self-risk prediction probability by the complement of the penetration risk factor adjusted by a decay index. The specific calculation process of the probability multiplication correction model is as follows: First, calculate the complement of the self-risk prediction probability (one minus the self-risk prediction probability); then, calculate the complement of the penetration risk factor (one minus the penetration risk factor) and perform a decay exponent operation on this complement, with the decay exponent set to 0.8; next, multiply the complement of the self-risk prediction probability with the decayed complement of the penetration risk factor to obtain a joint complement; finally, subtract the joint complement from one to obtain the final warning probability. The characteristics of this fusion model are: when the self-risk prediction probability is high, the final warning probability rapidly approaches one; when the self-risk prediction probability is low but the penetration risk factor is high, the final warning probability can be raised to a moderate level without completely obscuring its own safety, thus achieving causal propagation perception while maintaining physical rationality. In the early warning decision module, the specific operation for generating a three-level early warning instruction based on the final early warning probability is as follows: A high-risk immediate alarm is generated when the final warning probability is greater than the preset first alarm threshold and the edge-center terahertz amplitude ratio of the current inspection is greater than the preset second alarm threshold. The preset first alarm threshold is set to 0.8 and the preset second alarm threshold is set to 1.6. That is, a high-risk immediate alarm is triggered only when the final warning probability exceeds 0.8 and the current measured amplitude ratio exceeds 1.6, so as to ensure the reliability of the alarm and avoid false alarms based solely on predicted probability or single measured anomalies. When the final warning probability is between the preset first alarm threshold and the preset third alarm threshold, or when there is at least one causal parent node of the glass panel whose final warning probability is greater than the preset fourth alarm threshold and whose causal strength is greater than the preset fifth threshold, an encrypted monitoring command for surrounding glass panels is generated for the glass panel and all its first-order neighbors. The preset first alarm threshold is 0.8, and the preset third alarm threshold is 0.5, meaning that encrypted monitoring is triggered when the final warning probability is between 0.5 and 0.8. The preset fourth alarm threshold is set to 0.6, and the preset fifth threshold is set to 0.3, meaning that when there is at least one causal parent node whose final warning probability is greater than 0.6 and whose causal strength is greater than 0.3, encrypted monitoring is triggered even if the final warning probability of the current glass panel is not between 0.5 and 0.8. First-order neighbors refer to glass panels that are directly connected to the current glass panel through the keel, i.e., neighbors with dynamic edge weights greater than zero. The encrypted monitoring command shortens the next inspection cycle of the current glass panel and all its first-order neighbors from the usual 3 months to 2 to 4 weeks, realizing proactive tracking and monitoring of potential hazard diffusion areas. When a glass panel is marked as a source glass panel by the causality test module, a source tracing warning is generated. The marking conditions for a source glass panel are: a source index greater than 0.5 and an out-degree greater than or equal to 2. This warning is independent of the final warning probability. Even if the current risk of the source panel is not high, it still needs to be dealt with first to block the transmission chain because it has an amplifying effect. The source tracing warning is submitted in the form of a special report, which lists the location of the source panel, the causal transmission path, and the recommended disposal measures. In the early warning decision module, the specific operations after generating an early warning command are as follows: All warning instructions for glass panels are summarized by building floor and facade, generating a risk heatmap with overlaid directed causal arrows. The color intensity of each glass panel in the risk heatmap represents the final warning probability, and the directed arrows represent the causal direction and intensity. The generation process of the risk heatmap is as follows: First, the final warning probability of each glass panel is mapped to a color range: green for a warning probability of 0, yellow for 0.5, orange for 0.8, and red for 1.0. Intermediate values ​​are obtained through linear interpolation. Then, the building facade outline is drawn at the bottom of the heatmap, and the corresponding color is filled at each glass panel location. Next, directed causal arrows are overlaid on the heatmap, with arrows pointing from the source glass panel to the affected glass panels. The arrow thickness is proportional to the causal intensity: a causal intensity of 0.1 corresponds to a thin line, and a causal intensity of 1.0 corresponds to a thick line. Finally, legends and floor labels are added to form a complete visualization report. High-risk real-time alarms are sent to designated maintenance personnel via SMS or application push notifications. Encrypted monitoring instructions are written into the pre-set inspection plan database. Source tracing and early warning are submitted as special reports, thereby transforming early warning information into executable maintenance actions. The high-risk real-time alarm sending process is as follows: the system automatically generates an SMS message containing the building name, floor, facade, glass panel number, final warning probability, current amplitude ratio, and suggested measures, and sends it to the pre-registered maintenance personnel's mobile phones within 10 seconds via SMS gateway or application push service. The encrypted monitoring instruction writing process is as follows: the system automatically creates a new inspection task in the inspection plan database. The task content includes the target glass panel number, first-order neighbor list, planned inspection date (current date plus 2 to 4 weeks), and detection items (edge-center amplitude ratio, spectral distortion index, etc.), and sets the task status to pending execution. The format of the special report for source tracing and early warning is as follows: the cover page includes the building name and the date the report was generated; the first part of the main text lists the number, location, source index, output degree, and average causal intensity of all source glass panels; the second part displays the causal propagation path of each source panel in tabular form (i.e., a list of downstream panels directly affected by it and the causal intensity); the third part provides disposal suggestions (such as prioritizing the replacement of sealant, reinforcing the pressure strips, etc.); all early warning information is archived in the early warning log database for subsequent auditing and analysis.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls, characterized in that, Specifically, it includes: Topology graph construction module: Each glass panel is treated as a node, and the keel connection between adjacent glass panels is used as an edge. The edge weights are assigned according to the physical connection relationship between the glass panels. At the same time, a multi-dimensional feature vector is extracted for each node. The multi-dimensional feature vector includes the edge-center terahertz amplitude ratio of the current inspection, the trend slope of the historical edge-center amplitude ratio, and the environmental exposure orientation code. The edge-center amplitude ratio values ​​of the past inspections are accumulated to form a historical edge-center amplitude ratio sequence. The feature vectors of the past inspections are stacked to form a node feature tensor, thereby constructing the curtain wall topology graph. Spatiotemporal graph convolution module: Receives the curtain wall topology graph and node feature tensors, and learns the spatiotemporal evolution hidden state of the risk of delamination between glass panels by stacking gated temporal convolution and spatial graph convolution and embedding a propagation gating mechanism with physical constraints. Based on the spatiotemporal evolution hidden state, it outputs the risk prediction probability of each glass panel within a future preset time step. Causality test module: Performs pairwise conditional Granger causality test on the historical edge-center amplitude ratio sequence, constructs a directed causal graph and quantifies the causal strength between each glass panel to identify the source glass panel and the affected glass panel. Early warning decision module: It integrates the directed causal graph with the risk prediction probability, calculates the final early warning probability of each glass panel, and generates a three-level early warning instruction based on the final early warning probability, including high-risk immediate alarm, encrypted monitoring instruction for surrounding glass panels, and source tracing early warning.

2. The terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls according to claim 1, characterized in that: In the topology graph construction module, the specific operation of assigning edge weights based on the physical connection relationships between glass panels is as follows: First, set the initial edge weights according to the type of shared keel between adjacent glass panels. The edge weights of adjacent glass panels sharing vertical keels are set to the first preset value, the edge weights of adjacent glass panels sharing horizontal keels are set to the second preset value, the edge weights of diagonally adjacent glass panels are set to the third preset value, and the edge weights of non-adjacent glass panels are set to zero. Then, normalized mutual information is calculated based on the historical edge-center amplitude ratio sequence of each pair of adjacent glass panels in recent inspections to quantify the nonlinear dependence of the degradation process of the two glass panels. Finally, the normalized mutual information is multiplied by a preset adjustment coefficient and then added to the initial edge weights to obtain the corrected dynamic edge weights.

3. The terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls according to claim 2, characterized in that: In the topology graph construction module, the specific operation for extracting the multidimensional feature vector of each node is as follows: The multidimensional feature vector includes the edge-to-center terahertz amplitude ratio of the current inspection, the trend slope of the historical edge-to-center amplitude ratio, the environmental exposure orientation code, the spectral distortion index, and the glass panel size specification code. Among them, the trend slope of the historical edge-center amplitude ratio is obtained by robust linear regression on the historical edge-center amplitude ratio sequence of multiple past inspections; The environmental exposure orientation coding is based on the angle between the facade where the glass panel is located and the local prevailing wind direction. It uses a cosine function to quantize the segmented structure, with the windward side taking the first coding value, the leeward side taking the second coding value, and the crosswind side taking the middle coding value. The spectral distortion index is calculated based on the power spectral density of the terahertz echo signal obtained when the edge-to-center terahertz amplitude ratio is collected, to capture the changes in scattering characteristics caused by microstructural damage to the sealant; the glass panel size specification code is represented by the normalized logarithmic value of the glass panel area. Meanwhile, the feature vectors of the current inspection and the past multiple inspections are stacked in chronological order to form a node feature tensor. The dimension of the node feature tensor is determined by the number of glass panels, the number of time steps, and the feature dimension.

4. The terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls according to claim 3, characterized in that: In the spatiotemporal graph convolution module, the specific operation for learning the hidden states of spatiotemporal evolution is as follows: First, for the feature vector sequence corresponding to each glass panel in the node feature tensor, stacked gated dilated causal convolutional layers are used to extract the time dependency. Each layer contains a hyperbolic tangent convolutional branch and a sigmoid gated branch. The two are multiplied by element-wise to achieve non-linear gating. The dilation rate increases exponentially with the number of layers to expand the time receptive field. Then, based on the hidden state obtained after temporal convolution, spatial message passing is carried out using the curtain wall topology graph to calculate the hazard intensity index of each glass panel. The hazard intensity index maps the deviation of the edge-center terahertz amplitude ratio of the current inspection relative to the preset micro-hazard threshold to between zero and one through the Sigmoid function. By combining dynamic edge weights and normalized mutual information, the propagation gating coefficient is calculated, which significantly amplifies the message transmission weight of the source glass panel. Next, the aggregated messages of neighboring glass panels are aggregated using a weighted average method, with the weights obtained by normalizing the propagation gating coefficients. Finally, the hidden state obtained by convolving the aggregated message with its own time is fused through a linear transformation and activation function to obtain the spatially updated hidden state.

5. The terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls according to claim 4, characterized in that: In the spatiotemporal graph convolution module, the specific operation for outputting the risk prediction probability is as follows: A predetermined number of modules are stacked alternately with temporal convolutional layers and spatial graph convolutional layers. Each module contains one step of temporal convolution and one step of spatial graph convolution. Residual connections are introduced during the stacking process. The input and output of each module are added together and used as the input of the next module to alleviate gradient vanishing. After passing through a preset number of modules, the final spatiotemporal evolution hidden state is obtained. The spatiotemporal evolution hidden state is then input into a two-layer fully connected network. The first layer uses a linear transformation and the ReLU activation function, and the second layer uses a linear transformation and the Sigmoid activation function. The output is a risk prediction probability between zero and one, which represents the possibility that the glass panel will delaminate and loosen within a preset time step in the future.

6. The terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls according to claim 5, characterized in that: In the causality testing module, the specific steps for performing a pairwise conditional Granger causality test are as follows: First, using the risk prediction probability output by the spatiotemporal graph convolution module and the dynamic edge weight output by the topology graph construction module, glass panels with risk prediction probabilities exceeding a preset first threshold and their first-order neighbors, as well as glass panel pairs with normalized mutual information exceeding a preset second threshold output by the topology graph construction module, are selected to form a candidate causal pair set. Then, for each pair of glass panels in the candidate causal pair set, a conditional Granger causality test based on kernel function is adopted. By mapping the historical edge-center amplitude ratio sequence to the regenerated kernel Hilbert space, the kernelized conditional mutual information is calculated as the test statistic, and an approximate F statistic is constructed based on the residual variance ratio of the kernelized ridge regression to determine whether there is a causal relationship from one glass panel to another. Finally, for glass panel pairs that are determined to have a causal relationship, the causal strength is calculated. The causal strength is obtained by fusing the ratio of the F-statistics of the two-way test, the difference of the kernelized conditional mutual information, and the normalized mutual information. The ratio of the F-statistics of the two-way test is the ratio of the approximate F-statistic from the first glass panel to the second glass panel to the approximate F-statistic from the second glass panel to the first glass panel. The difference of the kernelized conditional mutual information is the difference between the kernelized conditional mutual information from the first glass panel to the second glass panel and the difference between the kernelized conditional mutual information from the second glass panel to the first glass panel, thereby constructing a directed causal graph.

7. The terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls according to claim 6, characterized in that: In the causality test module, the specific steps for identifying the source glass panel and the affected glass panel are as follows: Based on the directed causal graph, calculate the out-degree and in-degree of each glass panel, as well as the weighted out-degree and weighted in-degree; then, define the source index, which is obtained by multiplying the ratio of the difference between the weighted out-degree and the sum of the weighted in-degrees by the ratio of the weighted out-degree to the weighted out-degree plus one. When the source index is greater than the preset third threshold and the output degree is greater than the preset first quantity, the glass panel is marked as the source glass panel; When the source index is less than the preset fourth threshold and the in-degree is greater than the preset second quantity, the glass panel is marked as the affected glass panel.

8. The terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls according to claim 7, characterized in that: In the early warning decision-making module, the specific operation of causal fusion between the directed causal graph and the risk prediction probability is as follows: First, for each glass panel, all glass panels pointing to the glass panel are obtained as causal parent nodes based on the directed causal graph, and the risk prediction probability and causal strength output by the spatiotemporal graph convolution module of each causal parent node are obtained. Then, each causal parent node is weighted as a proportion of its causal strength to the sum of the causal strengths of all parent nodes. The weight is multiplied by the risk prediction probability of the parent node to obtain the infection probability of a single parent node. Then, the union probability of all parent nodes is calculated to obtain the penetration risk factor. Finally, the risk prediction probability of the glass panel itself is fused with the penetration risk factor through a probability multiplication correction model. The complement of the risk prediction probability is multiplied by the complement of the penetration risk factor after adjustment by the attenuation index, and then the complement is taken to obtain the final warning probability.

9. A terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls according to claim 8, characterized in that: In the early warning decision module, the specific operation for generating a three-level early warning instruction based on the final early warning probability is as follows: When the final warning probability is greater than the preset first alarm threshold and the edge-center terahertz amplitude ratio of the current inspection is greater than the preset second alarm threshold, a high-risk immediate alarm is generated. When the final warning probability is between the preset first alarm threshold and the preset third alarm threshold, an encrypted monitoring command for the surrounding glass panels is generated for the glass panel and all its first-order neighbors. When a glass panel is marked as the source glass panel by the causality verification module, a source tracing warning is generated.

10. A terahertz amplitude ratio early warning system for delamination and loosening of glass curtain walls according to claim 9, characterized in that: In the early warning decision module, the specific operations after generating an early warning command are as follows: All warning instructions for glass panels are summarized by building floors and facades, and a risk heat map with superimposed directed causal arrows is generated. The color intensity of each glass panel in the risk heat map indicates the magnitude of the final warning probability, and the directed arrows indicate the causal direction and intensity. High-risk real-time alarms are sent to designated maintenance personnel via SMS, encrypted monitoring instructions are written into the preset inspection plan database, and source tracing and early warning are submitted in the form of special reports, thereby realizing the transformation of early warning information into executable maintenance actions.

Images