Aluminum veneer curtain wall deformation monitoring method based on big data

Abstract

The invention relates to the technical field of curtain wall deformation monitoring, and discloses an aluminum veneer curtain wall deformation monitoring method based on big data. The method comprises the following steps: carrying out environmental factor separation on deformation original data to obtain abnormal deformation features; inputting a deformation mode identification network, and distinguishing connecting piece looseness, aluminum plate deformation and support displacement; analyzing a deformation trend, and predicting a future state; evaluating the safety, judging whether a threshold value is exceeded or not, and obtaining a risk level. According to the method, the deformation data of the curtain wall can be collected in real time, the influence of environmental factors can be effectively separated, the abnormal deformation type can be accurately identified, the deformation development trend can be scientifically predicted, and a complete safety risk assessment system can be established, so that the deformation state of the aluminum veneer curtain wall can be accurately monitored, classified, predicted and pre-warned, and the safety and reliability of a curtain wall system can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of curtain wall deformation monitoring, and in particular to a deformation monitoring method for aluminum veneer curtain walls based on big data. Background Art

[0002] Aluminum veneer curtain walls are a crucial component of the exterior decoration and protection systems of modern high-rise buildings. Their safety and stability have a significant impact on the overall safety of the building. Traditional deformation monitoring of aluminum veneer curtain walls relies primarily on manual inspections and periodic testing. Workers use visual observation and handheld measuring instruments to check the displacement and deformation of the aluminum panels. With technological advancements, some buildings have installed basic monitoring equipment such as strain gauges and displacement sensors. However, the data collection frequency is low and a systematic analysis method is lacking, making it difficult to provide real-time information on the curtain wall's condition. In recent years, technologies such as laser scanning and image recognition have begun to be applied to curtain wall monitoring, comparing before-and-after measurement data to determine deformation. However, these methods typically require specialized equipment and personnel, are costly, and cannot achieve continuous monitoring. Traditional methods for deformation data analysis rely primarily on empirical judgment and simple statistical analysis. They lack the ability to deeply explore and predict deformation characteristics, making it difficult to accurately distinguish between normal and abnormal deformation and providing early warning of potential risks.

[0003] Existing technologies have many shortcomings. First, traditional monitoring methods have poor timeliness and are unable to capture sudden deformation events. The manual inspection cycle is usually quarterly or semi-annual, and potential safety hazards that may arise during this period are difficult to detect in a timely manner. Secondly, single-point monitoring methods lack integrity and cannot fully reflect the deformation state of the curtain wall system, especially changes in key parts such as connection nodes are difficult to monitor. Furthermore, existing deformation data processing methods are simple and crude, lack effective separation of the influence of environmental factors, and it is difficult to identify real abnormal deformations. In addition, existing technologies lack deformation classification and trend prediction capabilities, cannot distinguish between different types of deformation causes, and cannot predict deformation development trends, resulting in a lack of scientific basis for maintenance decisions. Most importantly, existing technologies lack systematic risk assessment methods, cannot quantify deformation risk levels and provide accurate early warnings, which increases the difficulty and potential risks of building safety management. Summary of the Invention

[0004] This application provides a deformation monitoring method for aluminum veneer curtain walls based on big data, which is used to collect curtain wall deformation data in real time, effectively separate the influence of environmental factors, accurately identify abnormal deformation types, scientifically predict deformation development trends, and establish a complete safety risk assessment system, thereby realizing accurate monitoring, classification, prediction and early warning of the deformation status of aluminum veneer curtain walls, and improving the safety and reliability of the curtain wall system.

[0005] In the first aspect, the present application provides a method for monitoring the deformation of aluminum veneer curtain walls based on big data, and the method comprises: performing separation processing on the collected original deformation data of environmental influencing factors to obtain abnormal deformation characteristic data of the aluminum veneer curtain wall; inputting the abnormal deformation characteristic data of the aluminum veneer curtain wall into the curtain wall deformation pattern recognition network to distinguish between loose deformation of connectors, deformation of aluminum plates and displacement of supporting structures to obtain a curtain wall deformation type determination result; analyzing the deformation development trend according to the curtain wall deformation type determination result to obtain a prediction result of the future deformation state of the curtain wall; performing a curtain wall system safety assessment based on the prediction result of the future deformation state of the curtain wall to determine whether the deformation exceeds the safety threshold and obtain a curtain wall safety risk level.

[0006] In the technical solution provided by the present application, by separating the environmental influencing factors from the collected original deformation data, the interference of environmental factors such as temperature change and wind load on the deformation data is effectively eliminated, the purity and reliability of the abnormal deformation feature data are improved, and a high-quality data basis is provided for the subsequent deformation type identification; at the same time, by inputting the abnormal deformation feature data of the aluminum single plate curtain wall into a specially designed curtain wall deformation pattern recognition network, the deep learning technology is used to accurately classify the three main abnormal deformation types of loose connection deformation, aluminum plate deformation and supporting structure displacement, so that the cause of deformation is judged more accurately and the maintenance strategy is formulated more targeted. The recognition network adopts a dual-branch structure and feature fusion mechanism, which can capture The deformation amplitude characteristics and deformation frequency characteristics are captured, which greatly improves the recognition accuracy under complex deformation modes; in addition, based on the deformation type judgment results, the deformation development trend is analyzed, and combined with time series decomposition and multi-model fusion technology, the future deformation state of the curtain wall can be predicted at multiple scales in the short, medium and long term. By introducing the environmental factors to construct an environmentally sensitive deformation prediction model, the prediction accuracy is effectively improved, especially when the environmental conditions change greatly. It shows a significant advantage; finally, by performing a systematic safety assessment on the prediction results of the future deformation state of the curtain wall, the deformation risk is quantified and graded, and a complete closed loop from monitoring to early warning is established, which greatly improves the scientific nature and foresight of the curtain wall system safety management. The artificial intelligence algorithm features applied in the present invention contribute significantly to the solution, especially the deep neural network structure adopted in the curtain wall deformation pattern recognition network, which can learn and extract high-level feature representations from complex deformation data through multi-layer nonlinear transformation and dual-branch feature extraction mechanism, effectively overcoming the limitations of traditional methods in processing high-dimensional nonlinear deformation data; at the same time, the time series prediction model and environmental coupling mechanism integrated in deformation trend prediction can adaptively learn the complex correlation between deformation and environmental factors, greatly improving the accuracy and robustness of the prediction; and the multidimensional risk quantification algorithm and spatial mapping processing technology applied in the safety risk assessment link realize the accurate mapping from single-point risk to overall risk distribution, providing more intuitive and comprehensive decision-making support for building safety management. The in-depth application of these artificial intelligence algorithms has greatly improved the intelligence and precision of aluminum single-plate curtain wall deformation monitoring, and truly realized the technological leap from passive detection to active warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0008] Figure 1This is a schematic diagram of an embodiment of the aluminum veneer curtain wall deformation monitoring method based on big data in the embodiment of the present application. DETAILED DESCRIPTION

[0009] An embodiment of the present application provides a method for monitoring the deformation of an aluminum veneer curtain wall based on big data. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices.

[0010] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the method for monitoring deformation of aluminum single plate curtain wall based on big data includes: Step S101: Separate the environmental factors from the collected deformation raw data to obtain abnormal deformation feature data of the aluminum veneer curtain wall; Step S102: Input the abnormal deformation feature data of the aluminum single plate curtain wall into the curtain wall deformation pattern recognition network to distinguish the loose deformation of the connector, the deformation of the aluminum plate and the displacement of the supporting structure, and obtain the curtain wall deformation type determination result; Step S103: Analyze the deformation development trend based on the curtain wall deformation type determination result to obtain the curtain wall future deformation state prediction result; Step S104: Perform curtain wall system safety assessment based on the prediction result of the curtain wall's future deformation state, determine whether the deformation exceeds a safety threshold, and obtain the curtain wall safety risk level.

[0011] It is understandable that the execution subject of this application can be an aluminum single plate curtain wall deformation monitoring system based on big data, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0012] Specifically, the process begins with separating environmental factors. In aluminum veneer curtain wall systems, the raw deformation data collected by sensors contains deformation information caused by multiple factors, requiring processing to identify true abnormal deformation signatures. This process first synchronizes the raw deformation data, aligning sensor data distributed across the curtain wall according to timestamps to ensure temporal consistency. Outlier identification is then performed on the time-aligned data, identifying points that significantly deviate from the data mean. These points are typically caused by sensor failures or unexpected events. Waveform decomposition is then performed to decompose the data into high- and low-frequency components, filtering out interference signals caused by environmental vibrations. Temperature compensation is then performed, calculating deformation caused by thermal expansion and contraction based on temperature monitoring data and subtracting it from the data. Wind load separation is then performed, calculating elastic deformation caused by wind loads based on wind pressure monitoring values and eliminating this effect. Finally, time and frequency domain features are extracted, and statistical and spectral parameters are calculated to generate abnormal deformation signature data for the aluminum veneer curtain wall. Inputting this abnormal deformation signature data into the curtain wall deformation pattern recognition network is the second key step. The network structure consists of an input layer, multiple hidden layers, and an output layer. The abnormal deformation feature data first enters the input layer, consisting of 128 neurons, for initialization processing to generate initial feature map data. This data is then processed by the first hidden layer, consisting of 64 neurons, which uses the ReLU activation function for nonlinear transformation. The specific processing process includes weight matrix multiplication, bias vector addition, ReLU activation function processing, batch normalization, random dropout, and position encoding information, ultimately forming the first layer of feature representation data. This data is further processed by the second hidden layer, consisting of 32 neurons, which captures the deep feature correlations of the deformation data and generates the second layer of feature representation data. It is then input into a two-branch structure and processed separately by the deformation amplitude branch and the deformation frequency branch, each consisting of 16 neurons. After parallel processing, a dual-channel feature vector is obtained. These feature vectors are processed by the feature fusion layer, consisting of 24 neurons, which integrates the feature information of the two branches to form the fused feature representation data. Finally, these data are input into the output layer composed of three neurons, corresponding to the three types of deformation: loose connection, aluminum plate deformation, and support structure displacement. The probability value of each type is calculated through the Softmax function to obtain the curtain wall deformation type judgment result.

[0013] After determining the curtain wall deformation type, it is necessary to analyze deformation trends and predict the future deformation state of the curtain wall. This step first performs a time series decomposition on the deformation type determination results, breaking the deformation data into trend terms, seasonal terms, and residual terms to generate deformation decomposition data. Based on this data, time series prediction models corresponding to the deformation types are constructed. Prediction models are established for connector loosening, aluminum plate deformation, and support structure displacement, respectively. Historical deformation data is then input into these models for parameter training and optimization. The trained deformation prediction models are then combined with environmental factor data to construct an environmentally sensitive deformation prediction model that accounts for external factors such as temperature and wind. Based on this model, a rolling forecast operation is performed to perform multi-scale predictions of short-term, medium-term, and long-term deformation, generating deformation trend data for the aluminum veneer curtain wall. Finally, the results are corrected using deformation physical constraints to ensure that the predictions conform to the physical characteristics of the aluminum veneer curtain wall, resulting in a prediction of the curtain wall's future deformation state. The curtain wall system safety assessment is then performed based on the predicted future deformation state. First, the prediction results are compared and analyzed with the safety thresholds of the aluminum veneer curtain wall. The distance between the predicted deformation value and each level of safety threshold is calculated to obtain deformation safety margin data. Safety risk quantification is then performed, converting the deformation rate, deformation acceleration, and deformation amplitude into risk scores to form multidimensional risk scoring data. Next, a weighting is performed to weight and integrate the risks of different dimensions based on the structural characteristics of the aluminum veneer curtain wall to obtain a comprehensive risk index. This index is then compared with a historical deformation database to determine the risk level of the current deformation state within historical cases, forming a relative risk positioning result. Based on this result, a risk classification standard is established, classifying the risk level into four levels: observation, caution, warning, and danger, generating risk classification data. Finally, spatial mapping is performed to generate a safety risk distribution map for the aluminum veneer curtain wall, identifying high-risk areas and key monitoring points, and determining the curtain wall safety risk level.

[0014] In the embodiment of the present application, the deformation monitoring method of aluminum single plate curtain wall based on big data of the present invention separates the environmental influencing factors of the collected original deformation data, effectively eliminates the interference of environmental factors such as temperature change and wind load on the deformation data, improves the purity and reliability of abnormal deformation feature data, and provides a high-quality data basis for subsequent deformation type identification; at the same time, by inputting the abnormal deformation feature data of aluminum single plate curtain wall into a specially designed curtain wall deformation pattern recognition network, the deep learning technology is used to accurately classify the three main abnormal deformation types of loose connection deformation, aluminum plate deformation and supporting structure displacement, so that the cause of deformation is judged more accurately and the maintenance strategy is formulated more targeted. The recognition network adopts a dual-branch structure and feature fusion The combined mechanism can simultaneously capture the deformation amplitude characteristics and deformation frequency characteristics, greatly improving the recognition accuracy under complex deformation modes; in addition, based on the deformation type judgment results, the deformation development trend is analyzed, and combined with time series decomposition and multi-model fusion technology, the future deformation state of the curtain wall can be predicted at multiple scales in the short, medium and long term. By introducing the environmental factors to construct an environmentally sensitive deformation prediction model, the prediction accuracy is effectively improved, especially when the environmental conditions change greatly. It shows a significant advantage; finally, by performing a systematic safety assessment on the prediction results of the future deformation state of the curtain wall, the deformation risk is quantified and graded, and a complete closed loop from monitoring to early warning is established, which greatly improves the scientific nature and foresight of the curtain wall system safety management. The artificial intelligence algorithm features applied in the present invention contribute significantly to the solution, especially the deep neural network structure adopted in the curtain wall deformation pattern recognition network, which can learn and extract high-level feature representations from complex deformation data through multi-layer nonlinear transformation and dual-branch feature extraction mechanism, effectively overcoming the limitations of traditional methods in processing high-dimensional nonlinear deformation data; at the same time, the time series prediction model and environmental coupling mechanism integrated in deformation trend prediction can adaptively learn the complex correlation between deformation and environmental factors, greatly improving the accuracy and robustness of the prediction; and the multidimensional risk quantification algorithm and spatial mapping processing technology applied in the safety risk assessment link realize the accurate mapping from single-point risk to overall risk distribution, providing more intuitive and comprehensive decision-making support for building safety management. The in-depth application of these artificial intelligence algorithms has greatly improved the intelligence and precision of aluminum single-plate curtain wall deformation monitoring, and truly realized the technological leap from passive detection to active warning.

[0015] In a specific embodiment, the process of executing step S101 may specifically include the following steps: Perform time synchronization processing on the original deformation data, and use the timestamp alignment method to time-correct the sensor data to obtain time-aligned data; Perform outlier identification on the time-aligned data, mark the data points that deviate from the mean by more than a threshold, and obtain preliminary cleaned data; Perform waveform decomposition on the preliminary cleaned data to separate high-frequency components from low-frequency components, filter out interference signals caused by environmental vibration, and obtain noise-reduced data; Apply temperature compensation processing to the noise reduction data, calculate the deformation caused by thermal expansion and contraction based on the temperature monitoring data, and deduct it from the data to obtain temperature compensated data; The temperature compensation data is processed for wind load separation, and the elastic deformation caused by wind load is calculated based on the wind pressure monitoring value, and then eliminated from the data to obtain environmental factor exclusion data; The time domain characteristics and frequency domain characteristics of the data are extracted by excluding environmental factors, and the statistical parameters and spectrum parameters are calculated to obtain the abnormal deformation characteristic data of the aluminum single plate curtain wall.

[0016] Specifically, the raw deformation data is time-synchronized, using timestamp alignment to align data collected by different sensors. Multiple sensors on an aluminum veneer curtain wall are distributed at different locations, resulting in slight differences in acquisition times. This asynchrony can lead to bias in data analysis. Time synchronization compares the timestamps of each sensor's data to identify a baseline time point. Then, all sensor data are linearly interpolated to a uniform time interval, ensuring that all data points correspond to the same time series. For example, if sensor A collects data at 10:00:01 and sensor B at 10:00:02, timestamp alignment can align both data points to the same time point of 10:00:00, forming a complete time series dataset. The second step is outlier identification on the time-aligned data. Outliers are data points that significantly deviate from the normal range and are typically caused by sensor failure, external interference, or unexpected events. Outlier identification uses statistical methods to calculate the mean and standard deviation of the data, set a threshold range, and flag data points that fall outside this range. The specific operation first calculates the mean and standard deviation of the deformation data within a time window. Then, a threshold multiple is set (usually 3 times the standard deviation). Any data point that deviates from the mean by more than the threshold is marked as an outlier. These outliers can be directly deleted or replaced with the average of the adjacent points, thus obtaining preliminary clean data. In aluminum veneer curtain wall monitoring, outliers often indicate sensor failure or sudden external force on the curtain wall.

[0017] The third key step is to perform waveform decomposition on the initially cleaned data. Environmental vibrations such as traffic noise and wind vibrations can create high-frequency interference signals in the raw data, while actual curtain wall deformations often manifest as low-frequency changes. Waveform decomposition uses spectral analysis to convert time-domain signals into frequency-domain signals, separating high-frequency and low-frequency components. Specifically, a wavelet transform is used to perform multi-scale decomposition of the data. Appropriate wavelet basis functions (such as the db4 wavelet basis) are selected, and the number of decomposition layers (e.g., 4 layers) is set to decompose the original signal into multiple frequency bands. The low-frequency portion representing the actual deformation is then retained, while the high-frequency portion representing the environmental vibration is filtered out to achieve noise reduction. Actual deformation of aluminum veneer curtain walls typically manifests as signals with frequencies below 1 Hz, while environmental vibrations are often high-frequency signals above 1 Hz.

[0018] Applying temperature compensation processing to the noise reduction data is the fourth step. Aluminum has a large thermal expansion coefficient. Temperature changes will cause obvious thermal expansion and contraction deformation of the aluminum veneer curtain wall. This deformation is not a structural abnormality. Temperature compensation processing calculates and eliminates the normal deformation caused by temperature changes by establishing a temperature-deformation relationship model. In the specific implementation, first use historical data to establish a temperature sensitivity model of the aluminum veneer curtain wall to determine the linear relationship coefficient between temperature change and deformation. Then, based on the real-time temperature data during the monitoring period, calculate the expected deformation caused by temperature changes at each time point, deduct this part of the deformation from the total deformation, and obtain the temperature-compensated data. The linear expansion coefficient of the aluminum veneer is approximately 23× / ℃, when the ambient temperature changes by 20℃, a 5-meter-long aluminum panel can produce thermal deformation of about 2.3mm. This deformation needs to be eliminated through temperature compensation processing.

[0019] The fifth step is to separate the temperature-compensated data for wind loads. The wind loads faced by the exterior walls of high-rise buildings can cause elastic deformation of the aluminum veneer curtain wall. This deformation will automatically recover after the wind subsides and is not an abnormal deformation. The wind load separation process is based on wind pressure monitoring data and a wind load-deformation relationship model to calculate and eliminate normal elastic deformation caused by wind. In specific implementation, the wind pressure distribution on the curtain wall surface is monitored by a pressure sensor, and a mechanical relationship model between wind pressure and deformation is established in combination with the elastic modulus and structural characteristics of the aluminum veneer. This model is used to calculate the expected elastic deformation under a specific wind pressure, and this part of the deformation is further eliminated from the temperature-compensated data to obtain deformation data after eliminating environmental factors. In actual applications, when the wind pressure is 1kPa, a typical aluminum veneer can produce an elastic deformation of approximately 1.5mm. This deformation is identified and eliminated through wind load separation.

[0020] After excluding environmental factors, the data is used to extract time domain features and frequency domain features. Time domain features reflect the direct statistical characteristics of deformation, including parameters such as mean, standard deviation, peak value, peak-to-valley difference, skewness, and kurtosis; frequency domain features are obtained through Fourier transform, reflecting the periodicity and frequency distribution characteristics of deformation, including main frequency, spectral energy distribution, etc. These features together constitute the characteristic index system of abnormal deformation of aluminum veneer curtain wall, providing a data basis for subsequent deformation pattern recognition. The feature extraction process first divides the deformation data into time windows of fixed length, calculates statistical parameters within each window, and then applies fast Fourier transform to obtain spectral features, ultimately forming a data set containing multi-dimensional features, namely the abnormal deformation feature data of aluminum veneer curtain wall.

[0021] In a specific embodiment, the process of executing step S102 may specifically include the following steps: The abnormal deformation feature data of the aluminum veneer curtain wall is input into the input layer of the curtain wall deformation pattern recognition network. The input layer consists of 128 neurons, and the deformation features are initialized to obtain the initial feature mapping data. The initial feature map data is processed through the first hidden layer of the curtain wall deformation pattern recognition network. The first hidden layer consists of 64 neurons and uses the ReLU activation function to perform nonlinear transformation on the data to obtain the first layer of feature representation data. The first layer of feature representation data is processed through the second hidden layer of the curtain wall deformation pattern recognition network. The second hidden layer consists of 32 neurons, which captures the deep feature associations of the deformation data to obtain the second layer of feature representation data. The second layer of feature representation data is input into the dual-branch structure of the curtain wall deformation pattern recognition network, which is processed in parallel by the deformation amplitude branch and the deformation frequency branch composed of 16 neurons respectively to obtain a dual-channel feature vector; The dual-channel feature vectors are processed by the feature fusion layer of the curtain wall deformation pattern recognition network. The feature fusion layer consists of 24 neurons, which integrates the feature information of the two branches to obtain fused feature representation data. The fused feature representation data is input into the output layer of the curtain wall deformation pattern recognition network. The output layer consists of three neurons, corresponding to the three types of deformation: loose connection parts, aluminum plate deformation, and support structure displacement. The probability value of each type is calculated by the Softmax function to obtain the curtain wall deformation type judgment result.

[0022] Specifically, the abnormal deformation feature data of aluminum veneer curtain walls, obtained after separation of environmental factors, is input into the network. The input layer consists of 128 neurons, each corresponding to a feature dimension. These features include time-domain features (such as mean, standard deviation, and peak value) and frequency-domain features (such as dominant frequency and band energy distribution). The input layer normalizes the raw feature data, mapping all features of different dimensions to the same numerical range, typically the interval [-1, 1]. This prevents features with large values from dominating the network learning process. After normalization, the data undergoes a linear transformation and is perturbed by random noise to enhance the model's generalization and prevent overfitting. This generates initial feature map data. The initial feature map data is then processed by the first hidden layer, which consists of 64 neurons, half the number of neurons in the input layer, completing a dimensionality reduction process. The data is first multiplied by the weight matrix of the first hidden layer. Each input feature is multiplied by its corresponding weight and the sum is calculated to form the linear transformation result. A bias vector is then added to complete the affine transformation, followed by a nonlinear transformation using the ReLU activation function (rectified linear unit). The characteristic of the ReLU function is that it sets all negative values to zero and leaves positive values unchanged, expressed as f(x) = max(0,x). This nonlinear transformation enables the network to learn complex feature patterns, not just linear relationships. After the activation function, batch normalization is applied to reduce internal covariate shift and accelerate network training. A random dropout technique randomly shuts down some neurons to improve network robustness. Finally, combined with positional encoding information, the spatial location information of the sensors in the curtain wall structure is incorporated into the feature representation, resulting in the first layer of feature representation data.

[0023] The first-layer feature representation data is further processed by the second hidden layer, where the number of neurons is reduced to 32, allowing for the extraction of more abstract, high-level features. The processing also includes weight matrix multiplication, bias vector addition, ReLU activation, and batch normalization. The second hidden layer is designed to capture deep feature correlations within the deformation data, particularly the relationships between sensor data at different locations and the temporal evolution of the deformation pattern. Unlike the first hidden layer, which primarily extracts local features, the second hidden layer integrates broader information to form an understanding of the overall deformation pattern, generating a second-layer feature representation data at a higher level of abstraction. The second-layer feature representation data is then fed into the network's innovative dual-branch structure, specifically tailored to the deformation characteristics of aluminum veneer curtain walls. This dual-branch structure comprises a deformation amplitude branch and a deformation frequency branch, each consisting of 16 neurons. The amplitude branch focuses on time-domain features such as deformation magnitude and rate of change, while the frequency branch focuses on frequency-domain features such as deformation periodicity and frequency characteristics. This branching design is based on the observation that different types of deformation in aluminum veneer curtain walls exhibit distinct amplitude and frequency characteristics: loosening of connectors typically manifests as small-amplitude, high-frequency localized vibrations; deformation of the aluminum sheet manifests as large, slowly changing permanent deformations; and displacement of the supporting structure manifests as a medium-amplitude, gradually increasing overall displacement. By processing these two parallel branches, the network can extract both the amplitude and frequency characteristics of the deformation, forming a dual-channel feature vector.

[0024] The dual-channel feature vectors are processed by the feature fusion layer, which consists of 24 neurons, more than in any single branch, ensuring sufficient capacity to integrate information from both branches. The feature fusion layer first concatenates the feature vectors from the two branches and then performs a weighted combination of all features using a fully connected approach. During the fusion process, an attention mechanism assigns different weights to the information from the two branches, adaptively adjusting the importance of amplitude and frequency features based on the characteristics of the current deformation data. The fusion layer also considers the spatial correlation of deformation patterns, correlating information from adjacent sensors. This results in a fused feature representation that incorporates both time-frequency characteristics and spatial correlations.

[0025] The fused feature representation data is input to the output layer, which consists of only three neurons, one for each of the three deformation types: loose connector deformation, aluminum plate deformation, and support structure displacement. The output layer first performs a linear transformation on the fused features to obtain raw scores for the three deformation types. The Softmax function is then applied to these scores to convert them into probability distributions. The Softmax function ensures that the sum of all output values is 1, consistent with the properties of probability. This ultimately forms a probability distribution for the three deformation types, indicating the likelihood that the current aluminum veneer curtain wall deformation belongs to each type, thereby determining the curtain wall deformation type.

[0026] In a specific embodiment, the process of executing the step of processing the initial feature map data through the first hidden layer of the curtain wall deformation pattern recognition network may specifically include the following steps: Perform weight matrix multiplication on the initial feature map data, perform matrix multiplication on the input data and the weight matrix of the first hidden layer to obtain the linear transformation result data; Add the bias vector to the linear transformation result data, add the linear transformation result to the neuron bias vector, complete the affine transformation, and obtain the bias adjustment data; The bias-adjusted data is processed through the ReLU activation function, which sets values less than zero to zero and values greater than zero to remain unchanged, introducing nonlinear characteristics and obtaining activated feature data. Perform batch normalization on the activated feature data, calculate the mean and variance of the batch data, standardize the data, and obtain normalized feature data; A random dropout operation with a dropout rate of 0.2 is applied to the normalized feature data to randomly set some neuron outputs to zero to prevent the network from overfitting and obtain regularized feature data; The regularized feature data is combined with the position coding information, and the coding vector representing the spatial position of the deformation feature in the curtain wall structure is added to enhance the spatial correlation of the features and obtain the first layer of feature representation data.

[0027] Specifically, the initial feature map data is a multidimensional vector containing information about the various deformation features of the aluminum veneer curtain wall. The weight matrix is a parameter matrix derived from network training, reflecting the importance and correlation between different features. A matrix multiplication operation multiplies the initial feature map data by the weight matrix, achieving a linear transformation from the input space to the hidden layer space. In the specific calculation process, assuming the initial feature map data is a 128-dimensional vector and the first hidden layer has 64 neurons, the dimensions of the weight matrix are 128×64. Matrix multiplication multiplies each input feature by its corresponding weight and sums the results, generating 64 weighted combinations that constitute the linear transformation result data. This linear combination enables the network to selectively emphasize certain features and suppress others, thereby extracting key characteristic patterns of aluminum veneer curtain wall deformation. After the linear transformation, a bias vector is added to the result data. The bias vector is a set of constants that corresponds to each neuron and adds a fixed offset to the output of each neuron. This step adds the linear transformation result to the bias vector of the first hidden layer to complete the affine transformation. In a hidden layer of 64 neurons, the bias vector also contains 64 elements, one for each linear transformation result. The inclusion of the bias vector increases the model's flexibility, enabling the neural network to learn a baseline level of features and better fit the data distribution. For aluminum veneer curtain wall deformation monitoring, different deformation types may have different baseline levels. The introduction of the bias term allows the network to adjust for these differences, resulting in more accurate bias-adjusted data.

[0028] The bias-adjusted data needs to be processed by the ReLU activation function to introduce nonlinear characteristics. The ReLU activation function is a simple but effective nonlinear function that sets all negative values to zero and keeps all positive values unchanged. This operation may seem simple, but it can effectively enhance the network's expressive power and enable the network to learn complex nonlinear relationships. In the deformation monitoring of aluminum veneer curtain walls, there are often complex nonlinear relationships between the features of different deformation types. For example, the loose deformation of connectors may manifest as a specific combination pattern of certain features rather than a simple linear relationship. The application of the ReLU function enables the network to capture these nonlinear patterns, and it also has the advantages of simple calculation and good gradient transfer. After ReLU processing, activated feature data is formed, in which some of the originally negative neuron outputs are set to zero, showing sparse activation characteristics.

[0029] Batch normalization of activated feature data is an important step to improve the stability of network training. Batch normalization first calculates the mean and variance of the current batch data, and then standardizes the data so that its mean is 0 and its variance is 1. In the specific operation, for each feature dimension in a batch, the mean and variance of all samples on that dimension are calculated, and then the mean is subtracted from the current data and divided by the square root of the variance to achieve standardization. This processing makes the data distribution more stable, reduces the problem of internal covariate shift, and accelerates the network training process. For the deformation data of aluminum veneer curtain walls, there may be distribution differences in data from different batches. Batch normalization helps the network adapt to this change, improves the ability to recognize deformation patterns under different working conditions, and obtains normalized feature data.

[0030] Applying random dropout to normalized feature data is an effective technique for preventing network overfitting. This operation randomly sets the outputs of some neurons to zero according to a preset dropout rate (here, 0.2), simulating the effect of integrating multiple different network structures. During each forward propagation, approximately 20% of the neurons are temporarily shut down, their outputs set to zero, and they do not participate in the calculation of the current batch. This randomness prevents the network from over-relying on specific features, forcing it to learn more robust feature combinations. For deformation monitoring of aluminum veneer curtain walls, since the collected data may contain noise or redundancy, random dropout helps the network filter out these interferences and focus on the truly important deformation feature patterns, thereby obtaining regularized feature data with better generalization capabilities. Position encoding information is an encoding vector that represents the spatial position of the deformation feature within the curtain wall structure, reflecting the physical position relationship of the sensor on the aluminum veneer curtain wall. The spatial distribution of deformation data is crucial for deformation monitoring of aluminum veneer curtain walls, as different types of deformation often exhibit distinct spatial patterns: loose connections often manifest as localized abnormal deformation; aluminum plate deformation manifests as overall deformation of a single panel area; and support structure displacement manifests as the coordinated deformation of multiple adjacent panels. By combining regularized feature data with positional encoding information, the network can learn the spatial correlation of deformations and understand the mutual influence of deformations in adjacent regions, ultimately generating a first-layer feature representation containing rich spatial information.

[0031] Taking a real-world aluminum curtain wall monitoring case as an example, a sensor network installed on the exterior wall of a high-rise building collected deformation data across 128 feature dimensions. This data was first multiplied by a trained weight matrix consisting of 64 columns of weight vectors, each corresponding to a hidden layer neuron. The resulting 64-dimensional linear transformation revealed that certain feature combinations related to deformation of connection nodes had high values. A bias vector was then added to adjust the activation threshold of each neuron, making it more likely to activate the corresponding neuron when data indicating loose connections was present. After processing with the ReLU activation function, the outputs of approximately 30% of the neurons were set to zero, while the remaining neurons retained their original positive information. Batch normalization calculated the statistical distribution of all samples in the batch, normalizing the data and making it easier for the network to capture differences between different deformation types. A random dropout operation disabled approximately 13 neurons, forcing the network to use the remaining features for judgment. Finally, by adding coded information representing the sensor's position within the curtain wall structure, it was discovered that multiple adjacent sensing points in the northeast corner formed a clear deformation correlation pattern. This spatial correlation feature was significantly enhanced after combining it with the position encoding, allowing the displacement deformation of the supporting structure in this area to be successfully identified in subsequent classification.

[0032] In a specific embodiment, the process of inputting the fused feature representation data into the output layer of the curtain wall deformation pattern recognition network may specifically include the following steps: The fused feature representation data is transformed by the output layer weight matrix multiplication. The fused feature is multiplied with the weight matrices of the connector loosening deformation neurons, aluminum plate deformation neurons, and support structure displacement neurons in the output layer to obtain the original output score data of the three deformation types. The loose connection deformation bias, aluminum plate deformation bias, and support structure displacement bias are added to the original output score data of the three deformation types, completing the linear transformation of the output layer to obtain the adjusted output score data of the three deformation types. A Softmax function is performed on the adjusted output score data of the three deformation types to convert the loose connection deformation score, aluminum plate deformation score, and support structure displacement score into a probability distribution form, thereby obtaining the probability vectors of the three deformation types. Confidence thresholds were applied to the probability vectors of the three deformation types, marking the probability values of loose connection deformation, aluminum plate deformation, and support structure displacement below 0.15 as uncertain categories, thus obtaining the confidence-filtered probability data of the three deformation types. Based on the confidence-filtered probability data of the three deformation types, majority voting was performed to vote on the classification results of the three types of loose connection deformation, aluminum plate deformation, and support structure displacement in five consecutive time windows, and the time-series smoothed classification results of the three deformation types were obtained. The time-series smoothing classification results of the three deformation types are constrained and checked against the physical characteristics of the aluminum veneer curtain wall deformation to determine whether the three classification results of loose connection deformation, aluminum plate deformation, and supporting structure displacement conform to the physical constraint relationship, thereby obtaining the curtain wall deformation type determination result.

[0033] Specifically, the fused feature representation data undergoes an output layer weight matrix multiplication transformation. The fused feature representation data is a 24-dimensional feature vector obtained through processing by the previous layers of the network, containing high-level feature information for various deformation patterns. The output layer weight matrix contains three sets of weight vectors, one for each deformation type: connector loosening, aluminum plate deformation, and support structure displacement. The matrix dimensions are 24×3. During the weight matrix multiplication transformation, a dot product operation is performed on the 24-dimensional fused features with the weight vector corresponding to each deformation type, resulting in three scalar values. These three values represent the degree of match between the current deformation data and the three deformation types, with higher values indicating a higher degree of match. For example, when the input fused features contain significant periodic vibration features, the weight vector corresponding to connector loosening will assign higher weights to these features, resulting in a higher raw output score for connector loosening. Through this weight multiplication transformation, the network maps the multidimensional features into three raw output scores, representing the preliminary judgment results for the three deformation types.

[0034] The second step in completing the linear transformation of the output layer is to add corresponding bias values to the original output score data for the three deformation types. Each deformation type corresponds to a bias value, which is used to adjust the judgment threshold. In this step, the original output scores for the three deformation types of loose connection deformation, aluminum plate deformation, and support structure displacement are added to the corresponding bias values to obtain the adjusted output scores. The bias value serves to provide a baseline adjustment for the determination of deformation type and to compensate for any uneven distribution of different deformation types. For example, in actual aluminum veneer curtain wall monitoring, aluminum plate deformation is typically less frequent than loose connection deformation. By assigning a higher positive bias value to aluminum plate deformation, this uneven distribution can be balanced, ensuring that any signs of aluminum plate deformation are correctly identified. The three scores after bias adjustment constitute the adjusted output score data, but the ranges of these scores are inconsistent, making direct comparison difficult and requiring further processing.

[0035] Applying the Softmax function to the adjusted output scores for the three deformation types is a key step in converting the scores into probabilistic form. The Softmax function maps any three score values in the arbitrary range to probability values between 0 and 1, with the sum of the three probabilities being 1, consistent with the properties of a probability distribution. Specifically, each score value is exponentially calculated and then divided by the sum of the three exponential values. This process causes scores with larger values to occupy a higher proportion in the probability distribution, while the differences between smaller scores are magnified in the probability space, helping to provide a clearer judgment. For example, if the adjusted scores for the three deformation types are 2.5 (loose connector), 1.3 (aluminum plate deformation), and 0.8 (support structure displacement), respectively, after the Softmax function calculation, they are converted into probability distributions of approximately 0.65, 0.25, and 0.10, indicating that there is a 65% probability of loose connector, a 25% probability of aluminum plate deformation, and a 10% probability of support structure displacement. This probabilistic representation makes the deformation type judgment more intuitive and easier to understand.

[0036] Applying a confidence threshold to the probability vectors of the three deformation types is a crucial step in ensuring the reliability of the judgment results. A confidence threshold of 0.15 is set, and any probability value below this threshold for the three deformation types is labeled as uncertain. This step aims to filter out judgment results with insufficient confidence and avoid erroneous decisions under uncertainty. When the probabilities of all deformation types are below the threshold, it indicates that the current deformation signature does not match known deformation patterns and may represent a new deformation type or sensor noise. In this case, it should be labeled as "unknown" to prompt further human inspection. Conversely, if the probability value of a deformation type is significantly above the threshold, the judgment result is highly reliable. For example, in the deformation data detected in the northeast corner of an aluminum single-panel curtain wall, after network processing, the probabilities of the three deformation types are 0.08 (connector loosening), 0.87 (aluminum plate deformation), and 0.05 (support structure displacement). After applying the confidence threshold, connector loosening and support structure displacement are labeled as uncertain, while aluminum plate deformation is confirmed as the primary deformation type in this area.

[0037] Performing majority voting based on confidence-filtered probability data for the three deformation types is a key step in enhancing the temporal stability of the determination. Deformation type determination should not be based solely on data from a single time point but should consider the continuity of the time series. Majority voting calculates the results of each deformation type determination within five consecutive time windows and determines the final type using the principle of "the minority obeys the majority." Specifically, for each deformation type, the number of times that type is determined within the five time windows is counted, and the type with the highest frequency is determined as the dominant deformation type for that area. For example, if the deformation type determination results for a region in five consecutive time windows are: aluminum plate deformation, aluminum plate deformation, loose connector deformation, aluminum plate deformation, and aluminum plate deformation, majority voting will determine the deformation type for that region as aluminum plate deformation. This time series smoothing effectively reduces the interference of transient noise and outliers on the determination results, improving the stability and reliability of deformation type determination.

[0038] The final step in ensuring that the results of the time-smoothed classification of the three deformation types are consistent with the physical characteristics of aluminum veneer curtain wall deformation is to perform constraint checking on them. Deformation of aluminum veneer curtain walls is subject to certain physical constraints, such as: deformation types in adjacent areas typically exhibit spatial continuity; deformation types in the same area should not change frequently within a short period of time; and specific deformation types have characteristic spatial distribution patterns. Constraint checking compares the time-smoothed classification results with these physical characteristics to verify compliance with these constraints. If a result violates a physical rule, appropriate adjustments are made or the result is flagged for further verification. For example, if a single aluminum panel is identified as supporting structure displacement, but all surrounding panels are identified as loose connectors, this isolated determination violates the principle of spatial continuity and warrants reassessment. Through physical constraint checking, erroneous determinations that violate physical laws are eliminated, ultimately resulting in a curtain wall deformation type determination that both conforms to the data characteristics and satisfies the physical constraints.

[0039] Taking the monitoring of a high-rise aluminum veneer curtain wall as an example, the fused feature representation data, after multiplication of the output layer weight matrix, yields three raw scores: 3.2 (connector loosening), 1.8 (aluminum plate deformation), and 0.5 (support structure displacement). After adding corresponding bias values (-0.5, 0.2, and 0.3, respectively), the adjusted scores become 2.7, 2.0, and 0.8. Using a softmax function, these scores are converted to probability distributions: 0.62 (connector loosening), 0.31 (aluminum plate deformation), and 0.07 (support structure displacement). Applying a confidence threshold of 0.15, the support structure displacement (0.07) is labeled as uncertain, while the connector loosening and aluminum plate deformation remain as valid judgments. Majority voting revealed that in five consecutive time windows, this region was judged as connector loosening four times and aluminum plate deformation once. Therefore, the time series smoothing result confirms the connection loosening. Finally, a physical constraint check was performed, and it was found that the judgment result was consistent with the deformation pattern of the surrounding area and met the typical spatial distribution characteristics of loose connection deformation - manifested as local high-frequency small-amplitude vibration, confirming that the curtain wall deformation type in this area was loose connection deformation.

[0040] In a specific embodiment, the process of executing step S103 may specifically include the following steps: Perform time series decomposition processing on the curtain wall deformation type determination results, decompose the deformation data into trend terms, seasonal terms and residual terms, and obtain deformation decomposition data; Based on the deformation decomposition data, a time series prediction model for the corresponding deformation type is constructed. Prediction models are established for loose deformation of connectors, deformation of aluminum plates, and displacement of supporting structures to obtain a deformation type prediction model. Inputting historical deformation data into the deformation type prediction model, executing the parameter training process, optimizing the model parameters, and obtaining a trained deformation prediction model; The trained deformation prediction model is combined with environmental factor data to build a deformation-environment coupling prediction framework to obtain an environment-sensitive deformation prediction model. Based on the environment-sensitive deformation prediction model, a rolling prediction operation is performed to perform multi-scale predictions on short-term, medium-term and long-term deformations, and the deformation trend data of the aluminum veneer curtain wall is obtained; The deformation trend data of the aluminum veneer curtain wall is combined with the physical constraints of the deformation to correct the results, ensuring that the prediction results are consistent with the physical characteristics of the aluminum veneer curtain wall deformation and obtaining the prediction results of the future deformation state of the curtain wall.

[0041] Specifically, the deformation data is decomposed into trend terms, seasonal terms, and residual terms to obtain deformation decomposition data. The time series decomposition process uses the seasonal trend decomposition method to decompose the aluminum single-panel curtain wall deformation time series data into three independent components. The trend term reflects the long-term change trend of the deformation data, which is manifested as an upward or downward trend over time; the seasonal term represents the periodic deformation pattern of the deformation data due to periodic changes in environmental factors such as temperature and humidity; the residual term contains random fluctuations that cannot be explained by the trend term and seasonal term, and usually contains abnormal deformation information. Based on the deformation decomposition data, corresponding time series prediction models are then constructed for different deformation types. For loose deformation of connectors, the autoregressive integrated moving average model is used. This model is particularly suitable for capturing deformation data with random fluctuation characteristics. For aluminum plate deformation, the long short-term memory network model is used. This model can effectively capture the nonlinear trend characteristics of aluminum plate deformation through memory units. For the displacement of supporting structures, a type of deformation that has obvious periodicity and is greatly affected by environmental factors, an autoregressive integrated moving average model with seasonal decomposition is used, which can handle both trend and seasonal characteristics.

[0042] After inputting historical deformation data into the deformation type prediction model, a parameter training process is performed. During the training process, the parameters of the loose connection deformation model are adjusted, including the autoregressive order, the differential order, and the moving average order. The optimal parameter combination is determined through a grid search method. The long-short-term memory network model of aluminum plate deformation requires optimization of parameters such as the number of neurons, the number of hidden layers, the learning rate, and the batch size, and uses an early stopping method to prevent overfitting. The seasonal autoregressive integrated moving average model of the support structure displacement requires additional determination of seasonal period parameters, which are set according to the temperature change period.

[0043] After parameter training is complete, the optimized deformation prediction model is combined with environmental factor data to construct a coupled deformation-environment prediction framework. This framework incorporates environmental parameters such as temperature, humidity, wind speed, and sunshine data as external input variables. The dynamic correlation between deformation and environmental factors is captured through a gated recurrent unit network, which is capable of processing long-term dependencies and selectively memorizing important information. This coupled deformation-environment prediction framework employs an attention mechanism to assign weights based on the impact of different environmental factors on deformation, thereby improving prediction accuracy. Based on the environmentally sensitive deformation prediction model, a rolling forecast is performed to generate multi-scale predictions for short-term, medium-term, and long-term deformations. The short-term forecast covers deformation trends within 24 hours, using a rolling forecast with a step size of 1 hour. Each forecast uses the previous 72 hours of actual observations. The medium-term forecast covers deformation developments within a week, using a rolling forecast with a step size of 6 hours. It integrates short-term prediction results with actual observations. The long-term forecast covers deformation trends within a month, using a rolling forecast with a step size of 24 hours. It combines seasonal patterns in historical data with long-term forecasts of environmental factors.

[0044] Finally, the deformation trend data for the aluminum veneer curtain wall is corrected based on the physical constraints of the deformation. These constraints include the elastic deformation limit of the aluminum veneer, the maximum load-bearing capacity of the connectors, and the displacement safety threshold of the supporting structure. The correction process utilizes a physical constraint optimization algorithm to constrain the predicted results to a physically reasonable range, avoiding unreasonable results where the predicted values exceed the physical properties of the material.

[0045] Taking the monitoring of the aluminum veneer curtain wall of a high-rise building as an example, in analyzing connector loosening and deformation, the deformation data was first separated into trend terms through time series decomposition, revealing a slowly increasing trend in connector deformation. The seasonal term showed that deformation increased with rising daytime temperatures and decreased with falling nighttime temperatures. The residual term captured irregular, sudden deformations. After constructing an autoregressive integrated moving average model, parameter training determined the optimal parameter combination to be (2,1,1), i.e., second-order autoregression, first-order difference, and first-order moving average. After incorporating environmental factor data, a rolling forecast was performed using a deformation-environment coupling prediction framework. The forecast results indicated that connector deformation would exceed 75% of the design value within the next seven days. After combining physical constraint corrections, the system generated early warning information indicating connector loosening and deformation, indicating that connectors in specific areas require maintenance and reinforcement, effectively mitigating potential safety hazards.

[0046] In a specific embodiment, the process of executing step S104 may specifically include the following steps: Compare and analyze the predicted results of the curtain wall's future deformation state with the safety threshold of the aluminum veneer curtain wall, calculate the distance between the deformation prediction value and each level of safety threshold, and obtain the deformation safety margin data; Based on the deformation safety margin data, safety risk quantification calculation is performed, and the deformation rate, deformation acceleration and deformation amplitude are converted into risk scores to obtain multi-dimensional risk scoring data; Weights are assigned to multi-dimensional risk scoring data, and risks in different dimensions are weighted and integrated according to the structural characteristics of the aluminum veneer curtain wall to obtain a comprehensive risk index; Compare the comprehensive risk index with the historical deformation database to determine the risk level of the current deformation state in the historical cases and obtain the risk positioning result; Based on the risk positioning results, a risk level classification standard is established, and the risks are divided into four levels: observation level, attention level, warning level and danger level, and the risk level classification data is obtained; Spatial mapping is performed on the risk level classification data to generate a safety risk distribution map of the aluminum veneer curtain wall, identify high-risk areas and key monitoring points, and obtain the curtain wall safety risk level.

[0047] Specifically, the predicted results of the curtain wall's future deformation state are compared and analyzed with the safety threshold of the aluminum veneer curtain wall. The safety threshold refers to the safety limit of the aluminum veneer curtain wall under different usage conditions, including the safety threshold of loose connection deformation, the safety threshold of aluminum plate deformation, and the safety threshold of supporting structure displacement. During the comparative analysis, the deformation prediction value at each future time point is subtracted from the safety threshold of the corresponding position, and the difference between the two is calculated. This difference is the deformation safety margin data. The deformation safety margin data indicates the distance between the current deformation state and the safety limit. The larger the value, the larger the safety margin, the smaller the value, the closer to the dangerous state, and a negative value indicates that the safety threshold has been exceeded. When performing safety risk quantification calculations based on deformation safety margin data, it is necessary to consider the three key indicators of deformation rate, deformation acceleration, and deformation amplitude at the same time. Deformation rate refers to the rate of change of deformation per unit time, calculated by dividing the difference in deformation between adjacent time points by the time interval. Deformation acceleration refers to the rate of change of deformation rate, calculated by dividing the difference in deformation rate between adjacent time points by the time interval. Deformation amplitude refers to the absolute magnitude of the deformation. During the safety risk quantification process, these three indicators are first standardized to a range between 0 and 10. Then, based on the standardized indicator values, a risk score comparison table is used to convert each indicator into a corresponding risk score, resulting in a multidimensional risk score consisting of deformation rate risk score, deformation acceleration risk score, and deformation amplitude risk score.

[0048] Assigning weights to multidimensional risk score data is a core step in risk assessment. Due to the structural characteristics of aluminum veneer curtain walls, the importance of various indicators varies for different deformation types. For loose connector deformation, the deformation acceleration indicator is given a higher weight, as loose connectors often manifest as sudden changes in acceleration. For aluminum plate deformation, the deformation amplitude indicator is given a higher weight, as aluminum plate deformation is primarily limited by the maximum deformation. For support structure displacement, the deformation rate indicator is given a higher weight, as the rate of displacement has a significant impact on the overall safety of the curtain wall. The weight coefficients for each indicator are determined using the Analytic Hierarchy Process (AHP). The weight coefficients are then multiplied by the corresponding risk scores and summed to produce a comprehensive risk index, which comprehensively reflects the overall risk level of the curtain wall deformation. Comparing the comprehensive risk index with a historical deformation database is a key step in risk identification. The historical deformation database contains a large amount of historical monitoring data and corresponding risk assessment results. By calculating the similarity between the current comprehensive risk index and the risk indexes of historical cases, the most similar historical cases are identified and the current deformation state's position within the historical risk spectrum is determined. Similarity calculation uses the Euclidean distance method, using the distance between the current risk index and the historical case risk index as the similarity metric. A smaller distance indicates a higher similarity. Based on the similarity ranking results, the top N most similar historical cases are selected and their risk level distribution is statistically analyzed to obtain the risk location results.

[0049] Based on the risk location results, a risk classification standard was established, categorizing the risks into four levels: observation, caution, warning, and danger. The observation level indicates normal deformation and requires routine observation; the caution level indicates minor deformation anomalies, requiring increased monitoring frequency; the warning level indicates significant deformation anomalies, requiring specialized inspections; and the danger level indicates deformation is approaching or exceeding safety thresholds and requires immediate action. Risk classification utilizes a fuzzy comprehensive evaluation method, mapping the comprehensive risk index to four risk levels. Corresponding risk intervals and levels are then established to generate risk classification data. Spatial mapping of the risk classification data is a crucial step in risk visualization. This spatial mapping process links risk level information with the physical location of the aluminum veneer curtain wall. Based on the spatial coordinates of each monitoring point and the corresponding risk level data, an interpolation algorithm is used to calculate the risk distribution across the entire curtain wall surface. The interpolation algorithm uses an inverse distance weighted interpolation method. Based on the risk level and spatial location of known monitoring points, the risk level of unmonitored areas is estimated, generating a continuous risk distribution field. Finally, a safety risk distribution map of the aluminum veneer curtain wall is generated based on the risk distribution field data. Different colors are used to identify areas with different risk levels. Blue represents the observation level, yellow represents the attention level, orange represents the warning level, and red represents the danger level. High-risk areas and key monitoring points are intuitively displayed to obtain the curtain wall safety risk level.

[0050] Taking the monitoring of the aluminum veneer curtain wall of a high-rise building as an example, when conducting a safety assessment of the loosening and deformation of the connectors of aluminum veneer No. 3, the predicted deformation value for the next seven days was first compared with the safety threshold for loosening and deformation of the connectors. The calculated minimum safety margin was 1.2mm, occurring on the fifth day. Risk quantification was then performed based on this safety margin data. With a deformation rate of 0.3mm / day, a deformation acceleration of 0.05mm / day², and a deformation amplitude reaching 78% of the design value, these values translated into risk scores of 6, 7, and 8, respectively. Based on the characteristics of loosening and deformation of the connectors, weighting coefficients were assigned: 0.3 for the deformation rate, 0.5 for the deformation acceleration, and 0.2 for the deformation amplitude, resulting in a calculated comprehensive risk index of 6.9. Comparing this risk index with a historical database revealed a high degree of similarity with multiple cases of loose connectors, most of which were assessed as warning-level risks. Based on the risk identification results, the risk level of loosening and deformation of the No. 3 aluminum veneer connector was determined to be a warning level, and the area was marked as an orange warning zone in the curtain wall safety risk distribution map. At the same time, an early warning message containing maintenance recommendations was generated, suggesting that the connectors in this area be reinforced within the next three days, effectively preventing potential safety risks.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for monitoring deformation of aluminum veneer curtain wall based on big data, characterized in that: The method comprises: The collected original deformation data are processed to separate the environmental factors and obtain the abnormal deformation characteristic data of the aluminum single plate curtain wall; Inputting the abnormal deformation feature data of the aluminum single plate curtain wall into the curtain wall deformation pattern recognition network to distinguish the loose deformation of the connector, the deformation of the aluminum plate and the displacement of the supporting structure, and obtaining the curtain wall deformation type determination result; Analyze the deformation development trend according to the curtain wall deformation type determination result to obtain the prediction result of the curtain wall future deformation state; A curtain wall system safety assessment is performed based on the predicted result of the curtain wall's future deformation state to determine whether the deformation exceeds a safety threshold and obtain a curtain wall safety risk level.

2. The method for monitoring deformation of aluminum veneer curtain wall based on big data according to claim 1 is characterized in that: The collected deformation original data is subjected to environmental influence factor separation processing to obtain abnormal deformation characteristic data of the aluminum single plate curtain wall, including: Performing time synchronization processing on the deformation raw data, and performing time correction on the sensor data using a timestamp alignment method to obtain time-aligned data; Performing outlier identification on the time-aligned data, marking data points whose deviation from the mean exceeds a threshold, and obtaining preliminary cleaned data; Performing waveform decomposition processing on the preliminary cleaned data to separate high-frequency components from low-frequency components, filtering out interference signals caused by environmental vibrations, and obtaining noise-reduced data; Applying temperature compensation processing to the noise reduction data, calculating the deformation caused by thermal expansion and contraction based on the temperature monitoring data, and deducting it from the data to obtain temperature compensated data; Performing wind load separation processing on the temperature compensation data, calculating the elastic deformation caused by the wind load based on the wind pressure monitoring value, and eliminating it from the data to obtain environmental factor exclusion data; The time domain characteristics and frequency domain characteristics of the environmental factor exclusion data are extracted, and the statistical parameters and spectrum parameters are calculated to obtain the abnormal deformation characteristic data of the aluminum single plate curtain wall.

3. The method for monitoring deformation of aluminum veneer curtain wall based on big data according to claim 1, characterized in that: The abnormal deformation feature data of the aluminum single plate curtain wall is input into the curtain wall deformation pattern recognition network to distinguish the loose deformation of the connector, the deformation of the aluminum plate and the displacement of the supporting structure, and obtain the curtain wall deformation type determination result, including: The abnormal deformation feature data of the aluminum single plate curtain wall is input into the input layer of the curtain wall deformation pattern recognition network, the input layer is composed of 128 neurons, the deformation feature is initialized, and initial feature mapping data is obtained; The initial feature map data is processed by the first hidden layer of the curtain wall deformation pattern recognition network, the first hidden layer consists of 64 neurons, and the data is nonlinearly transformed using the ReLU activation function to obtain the first layer feature representation data; The first layer of feature representation data is processed through the second hidden layer of the curtain wall deformation pattern recognition network. The second hidden layer consists of 32 neurons and captures the deep feature associations of the deformation data to obtain the second layer of feature representation data. The second layer of feature representation data is input into the dual-branch structure of the curtain wall deformation pattern recognition network, which is processed in parallel by the deformation amplitude branch and the deformation frequency branch composed of 16 neurons respectively to obtain a dual-channel feature vector; The dual-channel feature vector is processed by the feature fusion layer of the curtain wall deformation pattern recognition network. The feature fusion layer is composed of 24 neurons, which integrates the feature information of the two branches to obtain fused feature representation data; The fused feature representation data is input into the output layer of the curtain wall deformation pattern recognition network. The output layer consists of three neurons, corresponding to the three types of deformation: loose connection parts, aluminum plate deformation, and support structure displacement. The probability value of each type is calculated by the Softmax function to obtain the curtain wall deformation type judgment result.

4. The method for monitoring deformation of aluminum veneer curtain wall based on big data according to claim 3 is characterized in that: The initial feature map data is processed through the first hidden layer of the curtain wall deformation pattern recognition network. The first hidden layer consists of 64 neurons and uses the ReLU activation function to perform nonlinear transformation on the data to obtain the first layer of feature representation data, including: Performing a weight matrix multiplication operation on the initial feature map data, performing a matrix multiplication operation on the input data and the weight matrix of the first hidden layer to obtain linear transformation result data; Performing bias vector addition on the linear transformation result data, adding the linear transformation result to the neuron bias vector to complete affine transformation, and obtaining bias adjustment data; The bias adjustment data is processed through the ReLU activation function, values less than zero are set to zero, values greater than zero remain unchanged, nonlinear characteristics are introduced, and activated feature data is obtained; Performing batch normalization on the activated feature data, calculating the mean and variance of the batch data, and standardizing the data to obtain normalized feature data; Applying a random dropout operation with a dropout rate of 0.2 to the normalized feature data, randomly setting some neuron outputs to zero to prevent network overfitting, and obtaining regularized feature data; The regularized feature data is combined with the position coding information, and a coding vector representing the spatial position of the deformation feature in the curtain wall structure is added to enhance the spatial correlation of the feature, thereby obtaining the first layer of feature representation data.

5. The method for monitoring deformation of aluminum veneer curtain wall based on big data according to claim 4 is characterized in that: The fused feature representation data is input into the output layer of the curtain wall deformation pattern recognition network. The output layer consists of three neurons, corresponding to the three types of loose connection deformation, aluminum plate deformation, and support structure displacement. The probability value of each type is calculated by the Softmax function to obtain the curtain wall deformation type judgment result, including: Performing an output layer weight matrix multiplication transformation on the fused feature representation data, multiplying the fused feature with the weight matrices of the connector loosening deformation neurons, aluminum plate deformation neurons, and support structure displacement neurons in the output layer to obtain original output score data of the three deformation types; Add the loose connection deformation bias value, aluminum plate deformation bias value, and support structure displacement bias value to the original output score data of the three deformation types, complete the linear transformation of the output layer, and obtain the adjusted output score data of the three deformation types; Performing a Softmax function operation on the adjusted output score data of the three deformation types, converting the loose deformation score of the connection, the deformation score of the aluminum plate, and the displacement score of the support structure into a probability distribution form, and obtaining probability vectors of the three deformation types; Apply confidence threshold screening to the probability vectors of the three deformation types, marking the probability values of loose connection deformation, aluminum plate deformation, and support structure displacement below 0.15 as uncertain categories, and obtain the confidence-filtered probability data of the three deformation types; Based on the confidence-filtered probability data of the three deformation types, majority voting is performed to vote on the classification results of the three types of loose connection deformation, aluminum plate deformation, and support structure displacement in five consecutive time windows to obtain time-series smoothed classification results of the three deformation types; The time-series smoothing classification results of the three deformation types are constrained and checked against the physical characteristics rules of the aluminum single-panel curtain wall deformation to determine whether the three classification results of loose connection deformation, aluminum plate deformation, and supporting structure displacement conform to the physical constraint relationship, thereby obtaining the curtain wall deformation type determination result.

6. The method for monitoring deformation of aluminum veneer curtain wall based on big data according to claim 1, characterized in that: The step of analyzing the deformation development trend based on the curtain wall deformation type determination result to obtain the curtain wall future deformation state prediction result includes: Performing time series decomposition processing on the curtain wall deformation type determination result, decomposing the deformation data into trend terms, seasonal terms, and residual terms to obtain deformation decomposition data; Constructing a time series prediction model corresponding to the deformation type based on the deformation decomposition data, and establishing prediction models for loose deformation of connectors, deformation of aluminum plates, and displacement of supporting structures, respectively, to obtain a deformation type prediction model; Inputting historical deformation data into the deformation type prediction model, executing a parameter training process, optimizing model parameters, and obtaining a trained deformation prediction model; Combining the trained deformation prediction model with environmental factor data to construct a deformation-environment coupling prediction framework to obtain an environment-sensitive deformation prediction model; Based on the environment-sensitive deformation prediction model, a rolling prediction operation is performed to perform multi-scale predictions on short-term, medium-term and long-term deformations to obtain deformation trend data of the aluminum veneer curtain wall; The deformation trend data of the aluminum veneer curtain wall is combined with the physical constraint conditions of deformation to perform result correction to ensure that the prediction result conforms to the physical characteristics of the aluminum veneer curtain wall deformation, and obtain the prediction result of the future deformation state of the curtain wall.

7. The method for monitoring deformation of aluminum veneer curtain wall based on big data according to claim 1, characterized in that: The curtain wall system safety assessment is performed based on the prediction result of the future deformation state of the curtain wall to determine whether the deformation exceeds the safety threshold and obtain the curtain wall safety risk level, including: Comparing and analyzing the predicted results of the future deformation state of the curtain wall with the safety threshold of the aluminum single-plate curtain wall, calculating the distance between the deformation prediction value and the safety thresholds at each level, and obtaining deformation safety margin data; Performing safety risk quantification calculation based on the deformation safety margin data, converting the deformation rate, deformation acceleration and deformation amplitude into risk scores to obtain multi-dimensional risk scoring data; Weighting the multi-dimensional risk scoring data, and weighting and integrating the risks of different dimensions according to the structural characteristics of the aluminum veneer curtain wall to obtain a comprehensive risk index; Comparing the comprehensive risk index with the historical deformation database to determine the risk level position of the current deformation state in the historical cases and obtain a risk positioning result; Establishing a risk level classification standard based on the risk positioning results, dividing the risk into four levels: observation level, attention level, warning level, and danger level, and obtaining risk level classification data; A spatial mapping process is performed on the risk level classification data to generate a safety risk distribution map of the aluminum single plate curtain wall, identify high-risk areas and key monitoring points, and obtain the curtain wall safety risk level.

Images