Building curtain wall monitoring data processing visualization method and system

Abstract

The invention discloses a building curtain wall monitoring data processing visualization method and system, and the method carries out the curtain wall health degree monitoring through an extreme learning machine algorithm, improves the significance of a signal in a fragile region through distributing dynamic weights for sensors at different positions, improves the response capability of the signal to local damage, and achieves the real-time monitoring of the health degree of the curtain wall. By simulating the pheromone attenuation effect and dynamically adjusting the weight and bias of each node, the adaptability of the model to complex and high-dimensional data is enhanced, the noise robustness is improved, and the influence of redundant features in the data is reduced. Based on the resonance effect, the adaptive capacity of the hidden layer to nonlinear data is enhanced. And by dynamically adjusting the regularization coefficient, the classification stability is improved. Based on a structural damage propagation consistency constraint term, the sensitivity to early local damage is enhanced, and the fault detection capability of the system is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent buildings, and in particular to a method and system for visualizing the processing of building curtain wall monitoring data. Background Art

[0002] As an important part of modern buildings, building curtain walls are widely used in the facades of high-rise buildings, commercial buildings, etc. They not only have beautiful exterior designs but also play a role in protecting the internal environment of buildings. However, as the service life of building curtain walls increases, affected by factors such as the external environment, temperature changes, wind force, and earthquakes, the curtain walls will gradually show varying degrees of damage. If these damages are not discovered in time, they may lead to potential building safety hazards, affect the service life of the building, and even cause serious safety accidents.

[0003] In the prior art, the sensitivity differences of sensors at different positions in the curtain wall are ignored, and simple weight initialization and processing methods are adopted, resulting in overfitting to noise and a decline in feature extraction ability. Using traditional activation functions cannot fully exploit the deep non-linear structure therein, and using a fixed weight update strategy leads to poor robustness of the model and a decrease in classification accuracy. The physical topology characteristics of the building structure and the damage propagation law are not fully utilized, resulting in weak performance in the detection and prediction of early local damages. Summary of the Invention

[0004] (I) Technical Problems to be Solved

[0005] To solve the above technical problems, the present invention provides a method and system for visualizing the processing of building curtain wall monitoring data.

[0006] (II) Technical Solutions

[0007] To solve the above existing technical problems and achieve the invention purpose, the present invention is realized through the following technical solutions:

[0008] A method for visualizing the processing of building curtain wall monitoring data includes the following steps:

[0009] S1: Data acquisition, real-time acquisition of the health monitoring data of the building curtain wall. Specifically, the physical state information of the curtain wall is obtained through a variety of sensors;

[0010] S2: Data preprocessing, performing data cleaning, denoising, standardization, normalization, and filling of missing data on the collected original data;

[0011] S3: Model the machine learning model for health status classification prediction; the model uses the extreme learning machine algorithm to obtain the health status classification prediction result; specifically, based on the adaptive feature weighting mechanism of the sensor spatial position and combined with the mechanical characteristics of the curtain wall structure, dynamic weights are assigned to sensors at different positions;

[0012] S4: Damage detection and prediction, apply the trained model in real time to detect and predict the damage of the health status of the building curtain wall;

[0013] S5: Visual display, present the monitoring data and the model analysis results to the user in a graphical way;

[0014] S6: System management and monitoring, manage, control and monitor the entire monitoring system.

[0015] Furthermore, the position weights of sensors at different positions in step S3 are:

[0016]

[0017] In the formula, is the spatial weight of the th sensor; is the Euclidean distance from the th sensor to the nearest edge of the curtain wall; is the Euclidean distance from the th sensor to the nearest edge of the curtain wall; is the maximum value of the edge distances of all sensors; is the spatial sensitivity coefficient; is the total number of sensors.

[0018] Furthermore, the extreme learning machine algorithm initializes the weights and bias parameters of the hidden layer nodes of the extreme learning machine based on the biological pheromone decay mechanism, dynamically adjusts the weights and biases of each node according to the feature distribution of the input data, and accepts random noise for correction.

[0019] Furthermore, the extreme learning machine algorithm uses non-linear resonance constraints to the hidden layer activation function, and the activation function is expressed as:

[0020]

[0021] In the formula, is the output of the hidden layer activation function of the extreme learning machine, is the hidden layer activation function of the extreme learning machine; is the weight of the hidden layer node of the extreme learning machine; is the bias of the hidden layer node of the extreme learning machine; is the input of the activation function, representing the feature vector corresponding to the input data; is the intensity of the resonance effect; is the resonance effect frequency.

[0022] Furthermore, the intensity of the resonance effect and the resonance effect frequency are expressed as follows:

[0023]

[0024]

[0025] where is the intensity of the resonance effect in the -th iteration training, controlling the oscillation effect in the non-linear mapping; is the resonance effect frequency in the -th iteration training, controlling the resonance effect frequency; is the first resonance effect intensity adjustment hyperparameter; is the second resonance effect intensity adjustment hyperparameter; is the first resonance effect frequency adjustment hyperparameter; is the second resonance effect frequency adjustment hyperparameter; is the actual label of the -th sample; is the -th sample in the input data.

[0026] Furthermore, the extreme learning machine algorithm is based on a dynamic regularization mechanism of node activity, and dynamically adjusts the regularization coefficient of the hidden layer nodes by calculating the activation variance and covariance matrix of each hidden layer node.

[0027] Furthermore, the extreme learning machine algorithm updates and calculates the bio-pheromone decay factor based on the relative change of the loss function gradient, expressed as:

[0028]

[0029] where is the bio-pheromone decay factor of the -th iteration; is the bio-pheromone decay factor of the -th iteration; is the decay rate control parameter; is the loss function gradient of the extreme learning machine at the -th iteration; is the loss function gradient of the extreme learning machine at the -th iteration; is the loss function of the extreme learning machine.

[0030] Furthermore, the calculation method of the loss function of the extreme learning machine is:

[0031]

[0032] In the formula, is the loss function of the extreme learning machine; is the number of training samples; represents the output value of the activation function of the hidden layer of the extreme learning machine; is the weight matrix of the hidden layer nodes of the extreme learning machine; is the th sample of the input data; is the bias of the hidden layer nodes of the extreme learning machine; The th actual label of the sample; represents the square of the prediction error of a single sample; is the dynamic regularization term; is the consistency constraint strength; is the consistency constraint term for structural damage propagation.

[0033] The present invention also provides a visualization system for processing building curtain wall monitoring data, which includes: a data acquisition unit, a data preprocessing unit, a machine learning modeling unit, a damage detection and prediction unit, a visualization display unit, and a system management and monitoring unit, wherein:

[0034] The data acquisition unit is used to collect the health monitoring data of the building curtain wall in real time;

[0035] The data preprocessing unit is used to perform denoising, standardization, and normalization processing on the collected original data;

[0036] The machine learning modeling unit is used to learn the health state pattern of the building curtain wall based on the extreme learning machine algorithm through training data, construct a classifier model, and obtain the classification prediction result of the health state;

[0037] The damage detection and prediction unit is used to perform real-time evaluation on the real-time monitoring data based on the machine learning model and output the prediction result of the health state;

[0038] The visualization display unit is used to present the monitoring data and the model analysis result to the user in a graphical manner;

[0039] The system management and monitoring unit is used to manage, control, and monitor the entire monitoring system.

[0040] In addition, to achieve the above object, the present invention also provides a computer-readable storage medium, on which program instructions for a visualization method for processing building curtain wall monitoring data are stored. The program instructions for visualizing the processing of building curtain wall monitoring data can be executed by one or more processors to implement the steps of the visualization method for processing building curtain wall monitoring data as described above.

[0041] (III) Advantageous Effects

[0042] Compared with the prior art, the advantageous effects of the present invention are as follows:

[0043] (1) The present invention avoids the equivalent processing method of sensor data in the conventional method, and significantly improves the sensitivity of early damage warning.

[0044] (2) The present invention simulates the pheromone attenuation effect in nature, enhances the adaptability of the model to complex and high-dimensional data, improves the noise robustness, reduces the influence of redundant features in the data, and can better process noise and heterogeneous data. It enhances the adaptability of the hidden layer to non-linear data, improves the sensitivity to complex patterns and tiny damage signals, and improves the performance of traditional activation functions in complex and dynamic data environments. It avoids the influence of noise and redundant data, and improves the robustness and classification performance of the model.

[0045] (3) The present invention dynamically adjusts the regularization coefficient by calculating the activation variance and covariance matrix of the hidden layer nodes, effectively reducing the activation intensity of high-noise sensitive nodes, enhancing the dependence of the model on stable and effective features, and improving the stability of classification.

[0046] (4) The present invention combines the physical topology structure of the building curtain wall and the damage propagation law, and proposes a structural damage propagation consistency constraint term, which enhances the sensitivity to early local damage and improves the fault detection ability of the system.

[0047] (5) The present invention dynamically adjusts the biological pheromone attenuation factor based on the relative change of the loss function gradient, enabling the model to adapt to the time-varying characteristics of the data, effectively controlling the influence of noise on model training, and enhancing the stability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0049] Figure 1 is a schematic diagram of a visualization method for processing building curtain wall monitoring data according to an embodiment of the present application; Figure 2It is the sensor spatial weight distribution and the edge damage signal enhancement effect diagram according to the embodiments of the present application, Figure 2 where part (a) in Figure 2 represents the sensor spatial weight distribution, Figure 3 and part (b) in Figure 4 represents the edge damage signal enhancement effect; Figure 4 where part (a) in Figure 4 represents the dynamic response surface of the resonant activation function, Figure 5 and part (b) in Figure 5 represents the response surface of the sigmoid activation function; Figure 5 where part (a) in represents the heat map of the original node covariance distribution,

[0050] and part (b) in

[0051] represents the heat map of the covariance distribution after dynamic regularization.

[0052] Specific embodiments

[0053] Referring to Figure 1 , a visualization method for processing building curtain wall monitoring data includes the following steps:

[0054] S1: Data acquisition, which is to collect the health monitoring data of the building curtain wall in real time, and obtain the physical state information of the curtain wall through various sensors (such as temperature sensors, vibration sensors, stress-strain sensors, etc.). The data acquisition process has high-precision and high-real-time requirements.

[0055] It includes the real-time acquisition, preprocessing and transmission of data. The specific functions include: the acquisition of sensor signals, the time sequence synchronization of data, ensuring that the data of various sensors are integrated according to the time sequence and ready for subsequent analysis.

[0056] It also includes detecting the working state of the sensors to ensure the reliability and stability of the data.

[0057] S2: Data preprocessing, which is to perform data cleaning, denoising, standardization, normalization, filling of missing data, etc. on the collected original data to improve the accuracy and efficiency of subsequent analysis and ensure that the data has high-quality features before being input into the machine learning model.

[0058] S3: Machine learning model building, which is used for health state classification prediction. The model adopts the extreme learning machine algorithm, learns the health state pattern of the building curtain wall through the training data, constructs a classifier model, and obtains the health state classification prediction result, which specifically includes the following:

[0059] a. Adaptive feature weighting of the sensor spatial position;

[0060] In the health monitoring of the building curtain wall, sensors are usually deployed at key structural positions (such as the four corners, joints, middle, etc.). Sensors at different positions have significant differences in sensitivity to the structural health state. For example, sensors at the edges and joints are more likely to capture local deformations and stress concentrations, while middle sensors are more sensitive to the overall vibration mode.

[0061] Conventional methods treat the sensor data at different positions equally, ignoring the association between their spatial positions and structural vulnerability, resulting in key signals being submerged by noise.

[0062] The present invention adopts an adaptive feature weighting mechanism based on the sensor spatial position, combines the mechanical properties of the curtain wall structure, and assigns dynamic weights to sensors at different positions to enhance the significance of signals in vulnerable areas. Specifically, the sensor position weight is defined as:

[0063]

[0064] In the formula, is the spatial weight of the th sensor, reflecting the contribution of the spatial positions of different sensors to the health state monitoring; is the The Euclidean distance from the i-th sensor to the nearest edge of the curtain wall, which characterizes the distance of the sensor from the structurally vulnerable areas (such as edges, joints, etc.). The closer the distance, the higher the structural vulnerability, and the greater the weight of the sensor should be; is the Euclidean distance from the i-th sensor to the nearest edge of the curtain wall (the closer the edge distance, the higher the structural vulnerability); is the maximum value of the edge distances of all sensors, which is used for normalization; is the spatial sensitivity coefficient, which controls the rate of attenuation of the weights of sensors near the edge. A larger spatial sensitivity coefficient will make the weights of sensors closer to the edge higher; is the total number of sensors; is the exponential function with the base of the natural constant.

[0065] Furthermore, the input data is processed by weighting and is expressed as , where is the -th sample in the input data, which characterizes the sample weighted by the spatial weights of the sensors; is the original input feature vector; represents element-wise multiplication; is the spatial weight vector, is the spatial weight of the i-th sensor.

[0066] In one embodiment, to verify the enhancement effect of the spatial weighting mechanism on the damage signals in the edge area of the curtain wall, analyze the spatial weight distribution of the sensors and its signal enhancement effect. The experiment associates the deployment positions of the sensors with the distances from the curtain wall edge, and compares with the equivalent processing mode of the traditional method for sensor data to obtain Figure 2 results. The experimental results show that the present invention constructs a spatial weight function related to the structural vulnerability, enables the sensors closer to the curtain wall edge to obtain higher feature weights, effectively amplifies the intensity of the damage signals at the edge joints, while the conventional method is difficult to extract key features from complex background noise due to ignoring the spatial position information, indicating that the adaptive weighting mechanism based on structural mechanics significantly improves the sensitivity of early damage warning.

[0067] b. Initialize the weights and bias parameters of the hidden layer nodes of the extreme learning machine based on the bioinformatics pheromone decay mechanism

[0068] In the classification of the health status of building curtain walls, the data comes from various sensors installed on the curtain walls, such as temperature sensors, vibration sensors, stress-strain sensors, displacement sensors, and humidity sensors, etc. These sensors collect the physical state data of the curtain walls in real time. Due to the diversity of sensors and the complexity brought by different types of physical quantities, the data is usually high-dimensional and contains noise. The high-dimensionality and noise of the data make the training and learning of the model more difficult.

[0069] In the conventional extreme learning machine, the weights and biases of the hidden layer nodes are usually randomly set during parameter initialization. This method lacks sensitivity to the characteristics of the data itself and is prone to overfitting to noise during the training process, thus reducing the generalization ability and classification performance of the model. When facing high-dimensional, complex, and noisy sensor data, it cannot adapt to the diversity and internal structure of the data.

[0070] In the present invention, a bio-pheromone decay mechanism is adopted to initialize the weights and biases of the hidden layer nodes. The bio-pheromone decay mechanism simulates the decay effect of pheromones in nature, dynamically adjusts the weights and biases of each node according to the feature distribution of the input data, and accepts random noise for correction, enabling the extreme learning machine model to more flexibly adapt to the changes in the data. During the training process, random noise terms are considered for perturbation to improve the robustness of the extreme learning machine model in complex data. Therefore, when there is noise in the input data, the model can still adapt to the noise situation, expressed as

[0071]

[0072]

[0073] In the formula, is the weight of the th hidden layer node of the extreme learning machine; is the bias of the th hidden layer node of the extreme learning machine; is the th sample in the input data; is the bio-pheromone decay factor, controlling the decay degree; is the value of the th sample in the input data after passing through the Sigmoid function, representing the non-linear transformation of the data, is the Sigmoid activation function; is the random noise term of the weight, following a normal distribution with a mean of 0 and a variance of the identity matrix; is the random noise term of the bias, following a normal distribution with a mean of 0 and a variance of the identity matrix; is the number of training samples.

[0074] The bio - pheromone decay factor is dynamically adjusted by calculating the similarity between input samples, enabling the initialization process to be adaptively adjusted according to the input data. For high - dimensional heterogeneous data from multiple types of sensors (such as temperature, vibration, etc.), through the dynamic adjustment of the similarity - driven bio - pheromone decay factor, the initialization of noise weights is reduced, and the adaptability of the model to the multi - physical - field coupling characteristics of the curtain wall is improved. The calculation method is expressed as:

[0075]

[0076] In the formula, is the similarity between sample and sample , which is specifically calculated by cosine similarity and represents the similarity degree between sample and sample . is the th sample in the input data, which is used to measure the spatial correlation of sensor data. Due to physical connections (such as the transfer of vibration by keels) of curtain - wall sensors, the data of adjacent sensors have a high similarity. High similarity will reduce the bio - pheromone decay factor and inhibit the initialization of redundant features; and are the Euclidean norms of sample and respectively, which represent the size and direction of the sample.

[0077] In one embodiment, a systematic verification of the noise robustness of the bio - pheromone decay initialization mechanism is carried out. By constructing a test environment with different noise levels and comparing it with the traditional random initialization method, the experiment simulates common noise scenarios such as temperature drift and electromagnetic interference in sensor data. The results are shown in Figure 3 . The results show that in a strong - noise environment, the initialization method based on bio - pheromone decay effectively suppresses the weight proportion of noise features by dynamically adjusting node parameters and maintains a stable feature - extraction ability. In contrast, the traditional random initialization method shows feature confusion under noise interference, resulting in a sharp decline in classification performance. This technology realizes the directional enhancement of the essential features of sensor data by simulating the natural pheromone decay law.

[0078] c. Using a non - linear resonance constraint for the hidden - layer activation function

[0079] In the task of classifying the health status of building curtain walls, since the collected data has non - linear characteristics, for example, different types of sensors may measure changes in different dimensions at the same time, and the relationships between these changes are often complex and non - linear.

[0080] The hidden layer activation functions of traditional extreme learning machines, such as the Sigmoid activation function or the ReLU activation function, although they can capture the basic nonlinear relationships in the data, often fail to fully represent the deep nonlinear structures in the data when faced with complex and dynamic data changes.

[0081] The hidden layer activation functions of conventional extreme learning machines, such as the Sigmoid activation function or the ReLU activation function, do not consider how to further enhance the nonlinear mapping ability of the hidden layer through the resonance effect in physics, resulting in poor classification performance of the extreme learning machine model's nonlinear expression ability when dealing with complex data.

[0082] In the present invention, by applying a nonlinear resonance constraint to the hidden layer activation function and mimicking the resonance effect in physics to enhance the adaptability of the hidden layer to nonlinear data, the activation function can not only capture the basic nonlinear mapping of the data but also enhance the complexity in the mapping process through the resonance effect, thereby effectively improving the expression ability of the model when dealing with complex nonlinear relationships. The activation function of the extreme learning machine is expressed as:

[0083]

[0084] where, is the output of the hidden layer activation function of the extreme learning machine, representing the mapping of the input data after passing through the hidden layer, is the hidden layer activation function of the extreme learning machine; is the weight of the hidden layer nodes of the extreme learning machine; is the bias of the hidden layer nodes of the extreme learning machine; is the input of the activation function, representing the feature vector corresponding to the input data; is the resonance effect intensity, controlling the degree of influence of resonance. The damage propagation of the curtain wall has nonlinear oscillation characteristics (such as crack propagation accompanied by stress waves). When the error is high, the resonance effect intensity is enhanced to capture the mutation signal; is the resonance effect frequency, controlling the speed of resonance change, simulating the natural frequency of the curtain wall material (such as the high-frequency response of the glass curtain wall and the low-frequency response of the concrete), enhancing the ability to capture abnormalities in specific frequency bands; is the sine function.

[0085] The resonance effect intensity and the resonance effect frequency are dynamically adjusted according to the error in each iteration training to ensure that the model can adjust the effect of the nonlinear mapping according to the change of the error in the training process. For the nonlinearity of the curtain wall data (such as the temperature-stress coupling effect), the resonance term is used to enhance the expression ability of the activation function for complex patterns and improve the classification accuracy of minor damages, which is expressed as:

[0086]

[0087]

[0088] In the formula, is the resonance effect intensity in the -th iteration training, which controls the oscillation effect in the non-linear mapping; is the resonance effect frequency in the -th iteration training, which controls the resonance effect frequency; is the first resonance effect intensity adjustment hyperparameter; is the second resonance effect intensity adjustment hyperparameter; is the first resonance effect frequency adjustment hyperparameter; is the second resonance effect frequency adjustment hyperparameter; is the actual label (target output) of the -th sample; is the -th sample in the input data. Preferably, is set to 2, is set to 0.3, is set to 3, is set to 0.5.

[0089] In one embodiment, the non-linear expression capabilities of the resonance activation function and the traditional activation function are compared. The experiment analyzes the enhancement effect of the resonance mechanism on non-linear mapping by constructing a joint parameter space of the input features and the training process. See Figure 4 The results show that while maintaining the basic non-linear characteristics, the resonance activation function proposed in the present invention generates multi-scale response patterns by dynamically adjusting the resonance parameters, and can effectively characterize the stress wave oscillation characteristics in the damage propagation of the curtain wall. The traditional activation function presents a single response characteristic and is difficult to capture the high-frequency mutation signals in complex mechanical behaviors. The activation mechanism based on the physical resonance principle significantly improves the model's analytical ability for multi-physical field coupling characteristics.

[0090] d. Weight adjustment of the extreme learning machine based on the biological pheromone decay mechanism and the adaptive adjustment mechanism;

[0091] The health state data of the building curtain wall often contains a large amount of sensor data, and these data may be affected by various factors such as environmental factors and sensor noise.

[0092] The conventional extreme learning machine adjusts the weights of the hidden layer nodes through a fixed weight update method, without dynamically adjusting according to the contribution degree of each node, so that the model cannot effectively focus on meaningful features in the face of data complexity and noise, resulting in a decline in classification performance.

[0093] By adopting a bio - pheromone attenuation mechanism and an adaptive adjustment mechanism, in each iterative training process, the weights are dynamically adjusted according to the contribution degree of each hidden - layer node, ensuring that the extreme learning machine model always focuses on the most informative features during the training process, while avoiding over - reliance on noise and redundant data, thereby improving the classification performance and robustness, which is expressed as:

[0094]

[0095]

[0096] In the formula, is the updated weight of the -th hidden - layer node of the extreme learning machine in the -th iterative training; is the current weight of the -th hidden - layer node of the extreme learning machine in the -th iterative training; is the bio - pheromone attenuation factor, controlling the attenuation degree; is the weight change of the -th hidden - layer node of the extreme learning machine in the -th iterative training; is the -th sample in the input data; is the adjustment factor of the -th sample, controlling the rate of weight update.

[0097] Furthermore, the weight change can be weighted and adjusted according to the importance of each sample, so as to more accurately optimize the weights of the model. The calculation method of the adjustment factor of the sample is expressed as:

[0098]

[0099] In the formula, is the -th sample in the input data.

[0100] The numerator part ( ) is the derivative of the Sigmoid function, used to represent the sensitivity of the activation function.

[0101] Adjustment factor Essentially, it adjusts the importance update speed and influence of each sample (influence refers to the impact of sample diversity on model performance. For example, sample diversity can be measured by the eigenvalue variance of the sample. A large eigenvalue variance of the sample indicates a large sample diversity). The importance of a sample is jointly determined by its activation value (calculated by the Sigmoid function) and sensitivity (represented by the magnitude of the derivative of the Sigmoid). The more active the sample and the larger its activation amplitude (Sigmoid approaching 0.5), the greater its impact on importance update; while when the sample shows a "negative" state (Sigmoid approaching 0 or 1) during training, its influence on the model is relatively small.

[0102] In addition, the derivatives of all samples are normalized in the denominator so that the adjustment factors of all samples are within a unified scale range, avoiding the adjustment factors of some samples being too large or too small, thus affecting the balance of model training.

[0103] Therefore, by weighted adjustment of the importance of samples, this formula enables the model to pay more attention to those samples that make greater contributions to classification during training, avoiding excessive interference from noise samples or samples with weak information to the training process, thereby improving the training efficiency and robustness of the model. e. Hidden layer output dynamic regularization and node activity optimization

[0104] In the classification of the health status of building curtain walls, the high-dimensional characteristics of sensor data will lead to redundancy or over-activation of the output of hidden layer nodes. Some nodes may be sensitive to noise and generate abnormal activation values, affecting the classification accuracy.

[0105] Conventional extreme learning machines directly perform linear regression to solve after the output of the hidden layer, lacking dynamic regularization processing of the hidden layer output matrix and unable to suppress the negative impact of redundant nodes.

[0106] The present invention adopts a dynamic regularization mechanism based on node activity. By calculating the activation variance and covariance matrix of each hidden layer node, the regularization coefficient is dynamically adjusted to suppress the contribution of noise-sensitive nodes and enhance robustness. The regularization term of the hidden layer output matrix of the extreme learning machine is defined as:

[0107]

[0108] In the formula, is the output vector of the th hidden layer node of the extreme learning machine; is the output vector of the th hidden layer node of the extreme learning machine; is the regularization coefficient of the th hidden layer node of the extreme learning machine: is the dynamic regularization term, reflecting the redundancy and covariance relationship of the output of hidden layer nodes; is the number of hidden layer nodes; is the covariance function; is the L2 norm; is the -th and -th covariance penalty coefficient of hidden layer nodes of the extreme learning machine.

[0109] Furthermore, the regularization coefficient of hidden layer nodes is calculated based on the variance proportion of node output, ensuring that the model depends on nodes with higher stability, thereby improving the robustness of classification. The calculation method is expressed as:

[0110]

[0111] In the formula, is the variance of the output of the -th hidden layer node of the extreme learning machine, reflecting the sensitivity of the -th hidden layer node of the extreme learning machine to input changes. Nodes with large variance are sensitive to input changes (such as being affected by temperature drift). By reducing their weights through the regularization coefficient, environmental noise is suppressed; is the variance of the output of the -th hidden layer node of the extreme learning machine.

[0112] Furthermore, the covariance penalty coefficient adjusts the regularization intensity dynamically. The curtain wall sensor network has spatial correlation (such as adjacent sensors vibrating synchronously). High covariance nodes impose penalties on redundant features to avoid over-activation, forcing the extreme learning machine model to preferentially depend on nodes with stable activation and low correlation with noise, thereby improving classification stability. The calculation method is expressed as:

[0113]

[0114] In the formula, represents the variance function, which is used to measure the dispersion degree of node output and optimize the regularization intensity between nodes.

[0115] In one embodiment, the regulation effect of the dynamic regularization mechanism on the correlation of hidden layer nodes is analyzed through a covariance heat map, as shown in Figure 5, The experiment compared the distribution characteristics of the activation patterns of hidden layer nodes before and after regularization processing. The results show that through the adjustment of the regularization coefficient based on node variance adaption, the present invention effectively suppresses the activation intensity of high-noise sensitive nodes and reduces the covariance correlation between spatially redundant nodes. Compared with traditional static regularization methods, this technology can dynamically optimize the node contribution degree according to real-time data characteristics, significantly improving the model's robustness to interference factors such as sensor drift while ensuring the feature expression ability. The dynamic optimization strategy based on node activity provides stability guarantee for long-term health monitoring.

[0116] f. Update of the bio-pheromone decay factor driven by error feedback

[0117] The sensor data of building curtain walls has time-varying characteristics. The noise level and data distribution may change in different time periods, and it is necessary to dynamically adjust the bio-pheromone decay factor in the initialization parameters to adapt to the time-varying characteristics of the building curtain wall sensor data.

[0118] The present invention calculates the update of the bio-pheromone decay factor based on the relative change of the loss function gradient, and then realizes dynamic adaptation to the time-varying characteristics of the data, expressed as:

[0119]

[0120] In the formula, is the bio-pheromone decay factor of the th iteration; is the bio-pheromone decay factor of the th iteration; is the exponential function with the base of the natural constant; is the decay rate control parameter; is the loss function gradient of the extreme learning machine at the th iteration; is the loss function gradient of the extreme learning machine at the th iteration; is the loss function of the extreme learning machine.

[0121] Based on this, when the loss gradient increases ( ), it indicates that the model is in an unstable state, and it is necessary to enhance the decay effect (reduce ) to reduce the influence of noise; when the loss gradient decreases ( ), it indicates that the model tends to converge, and it is necessary to weaken the decay effect (increase ) to retain effective features.

[0122] g. Calculate the loss function of the extreme learning machine;

[0123] In the classification of the health status of building curtain walls, structural damage often has the characteristic of spatial propagation (for example, cracks extend from the edge to the center, and stress concentration areas cause abnormalities in adjacent sensors).

[0124] Conventional loss functions only consider the single-point prediction error and do not model the damage propagation law among sensor networks, resulting in insufficient sensitivity to early local damage.

[0125] In the present invention, a structural damage propagation consistency constraint term is adopted in the loss function. By integrating the physical connection topology of the curtain wall and the prior knowledge of damage propagation, the prediction result of the model is constrained to conform to the engineering mechanics law, and it is jointly constituted by combining with the regularization term. The calculation method is as follows:

[0126]

[0127] In the formula, is the loss function of the extreme learning machine, which represents the overall error between the predicted value of the model and the actual label; is the number of training samples, that is, the number of samples of the input data; represents the output value of the activation function of the hidden layer of the extreme learning machine, which represents the predicted value of the th sample after passing through the hidden layer; is the weight matrix of the hidden layer nodes of the extreme learning machine, which is initialized and dynamically adjusted through the biological pheromone attenuation mechanism; is the th sample of the input data; is the bias of the hidden layer nodes of the extreme learning machine, which is initialized and dynamically adjusted through the biological pheromone attenuation mechanism; The th sample's actual label (target output value), that is, the true classification result of the health status of the building curtain wall; represents the square of the prediction error of a single sample, which is used to quantify the deviation degree between the predicted value and the true value; is the dynamic regularization term, which is used to suppress the negative impacts of redundant nodes and noise-sensitive nodes; is the consistency constraint strength; is the structural damage propagation consistency constraint term. Preferably, is set to 0.2.

[0128] The structural damage propagation consistency constraint term encodes the mechanical properties of the curtain wall into the loss function through topological connection, forcing the model to learn the prediction result that conforms to the actual damage propagation mode. The constraint on the correlation of the prediction changes of adjacent sensors can amplify the chain effect of local tiny abnormalities and improve the early fault detection ability. The calculation method is expressed as:

[0129]

[0130] In the formula, is the set of topological connection edges of the curtain wall sensor network, representing the physical connection relationship; denotes the th sample of the th sensor's feature, denotes the th sample of the th sensor's feature; is the temporal gradient of the loss function of the extreme learning machine's previous iteration with respect to the sensor data, approximated by the difference of predicted values at adjacent time steps; is the damage propagation attenuation factor, calculated according to the mechanical transfer coefficient from the th node to the th node, expressed as , where is the L2 norm, is the three-dimensional coordinate of the th sensor, is the three-dimensional coordinate of the th sensor, is the angle between the connection edge and the direction of the principal stress of the curtain wall, is the first material property parameter, is the second material property parameter. For example, for a concrete curtain wall, take , .

[0131] In one embodiment, the set of topological connection edges of the curtain wall sensor network is specifically , and each edge represents that in the physical structure of the curtain wall, there is a direct mechanical transfer path between the th sensor and the building component (such as glass panel, keel, connecting piece, etc.) where the th sensor is located. The construction method can be that the two sensors are directly connected through the curtain wall load-bearing structure (such as keel, frame), or it can be digitally generated based on the node connection relationship in the curtain wall CAD drawing.

[0132] h. Model convergence condition;

[0133] In view of the characteristic that the health monitoring of the building curtain wall needs to run continuously for a long time, a double-threshold convergence condition is defined, expressed as:

[0134]

[0135] In the formula, is the first error threshold, preset by humans; is the actual label (target output) of the th sample; The second error threshold, which is preset manually. Preferably, it is set to 0.01, it is set to 0.02.

[0136] When the double - threshold convergence condition holds, the iteration stops, indicating that the model training is completed.

[0137] S4: Damage detection and assessment. The trained model is applied in real - time to evaluate the health status of the building curtain wall, and then to judge whether there is damage.

[0138] The damage detection and assessment are based on the machine - learning model to evaluate the real - time monitoring data in real - time and output the evaluation results of the health status.

[0139] In one embodiment, the prediction results are different health - status levels, which are divided into 5 levels in total. The higher the level, the better the health status.

[0140] It should be noted that the categories of the prediction results correspond to the classification categories of the extreme learning machine, depending on the annotation categories of the training samples constructed for training the extreme - learning - machine model. For example, if the annotation categories are different health - status levels, which are divided into 5 levels in total, and the higher the level, the better the health status, then the types of the prediction results are the same.

[0141] S5: Visualization display. The monitoring data and the model analysis results are presented to the user in a graphical way.

[0142] The visualization - display unit is responsible for presenting the real - time health - monitoring data, prediction results, and the running status of the model to the user in the form of easy - to - understand charts, heat maps, trend charts, etc., to help the user intuitively understand the health status of the building curtain wall.

[0143] The functions of the visualization - display unit include graphical display of real - time data streams, dynamic display of prediction results, statistical analysis and trend display of historical data, etc. Through the visualization - display unit, the user can conveniently view and analyze the health status of the building curtain wall and make maintenance decisions.

[0144] S6: System management and monitoring. Manage, control, and monitor the entire monitoring system.

[0145] It includes managing the hardware and software components of the system, monitoring the running status of the monitoring system, and ensuring the stability and efficiency of the system. Specifically, it includes device - status monitoring, data - stream management, system - log recording, error alarm and repair, etc.

[0146] In this embodiment, by assigning dynamic weights to sensors at different positions, the significance of signals in vulnerable areas is enhanced. In particular, higher weights are given to sensors near the edges and joints of the building curtain wall, thereby improving the signal's response ability to local damage, avoiding the equivalent processing method of sensor data in conventional methods, and significantly enhancing the sensitivity of early damage warning. By simulating the pheromone attenuation effect in nature, the weights and biases of each node are dynamically adjusted, enhancing the model's adaptability to complex and high-dimensional data, improving the noise robustness, and reducing the influence of redundant features in the data. Compared with traditional random initialization methods, it can better handle noise and heterogeneous data. In addition, by imitating the resonance effect in physics and introducing it into the hidden layer activation function of the extreme learning machine, the adaptability of the hidden layer to non-linear data is enhanced. It not only captures the basic non-linear mapping of the data but also improves the sensitivity to complex patterns and tiny damage signals through the resonance effect, improving the performance of traditional activation functions in complex and dynamic data environments. By combining the biological pheromone attenuation mechanism and the adaptive adjustment mechanism, the weights are dynamically adjusted according to the contribution degree of hidden layer nodes during each iteration, focusing on meaningful features, avoiding the influence of noise and redundant data, and enhancing the robustness and classification performance of the model. By calculating the activation variance and covariance matrix of hidden layer nodes and dynamically adjusting the regularization coefficient, the negative impacts of redundant nodes and noise-sensitive nodes are suppressed, effectively reducing the activation intensity of high-noise-sensitive nodes, enhancing the model's dependence on stable and effective features, and improving the stability of classification. Combining the physical topology structure of the building curtain wall and the damage propagation law, a structural damage propagation consistency constraint term is proposed, which can ensure that the model takes into account the spatial propagation characteristics between sensors during prediction, enhancing the sensitivity to early local damage and improving the fault detection ability of the system. Dynamically adjusting the biological pheromone attenuation factor based on the relative change of the loss function gradient enables the model to adapt to the time-varying characteristics of the data, effectively controlling the influence of noise on model training and enhancing the stability of the model.

[0147] An embodiment of the present invention also proposes a visualization system for processing building curtain wall monitoring data, including a data acquisition unit, a data preprocessing unit, a machine learning modeling unit, a damage detection and prediction unit, a visualization display unit, and a system management and monitoring unit, where:

[0148] The data acquisition unit is used to collect the health monitoring data of the building curtain wall in real time and support various data upload methods such as databases, files, and HTTP.

[0149] The data preprocessing unit is used to denoise, standardize, and normalize the collected raw data, and also has the functions of cache and multi-dimensional analysis.

[0150] The machine learning unit uses machine learning algorithms to build a prediction model to evaluate the health status of the building and detect damages.

[0151] The machine learning modeling unit is used to build a prediction model based on the extreme learning machine algorithm to evaluate the health status of the building and detect damages, learn the health status pattern of the building curtain wall through training data, build a classifier model, and obtain the classification prediction result of the health status.

[0152] The damage detection and prediction unit is used to perform real-time evaluation on the real-time monitoring data based on the machine learning model and output the prediction result of the health status.

[0153] The visualization display unit is used to present the monitoring data and the model analysis results to the user in a graphical way. The visualization elements include an interactive dashboard, reports, and a WEB interface to help the user clearly understand the health status of the building.

[0154] The system management and monitoring unit is responsible for ensuring the normal operation of the entire system by monitoring the health status of the hardware and software, managing the data flow, and performing daily maintenance.

[0155] In addition, an embodiment of the present invention further provides a computer-readable storage medium, on which program instructions for the visualization method of building curtain wall monitoring data processing are stored. The program instructions for the visualization of building curtain wall monitoring data processing can be executed by one or more processors to implement the steps of the visualization method of building curtain wall monitoring data processing as described above.

[0156] The above-described embodiments are only used to describe the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for visualizing building curtain wall monitoring data processing, characterized in that: The steps include: S1: Data collection: real-time collection of health monitoring data of building curtain walls. Specifically, the physical status information of the curtain walls is obtained through a variety of sensors. S2: Data preprocessing, including data cleaning, denoising, standardization, normalization, and missing data filling of the collected raw data; S3: Machine learning model building for health status classification prediction; the model uses the extreme learning machine algorithm to obtain health status classification prediction results; specifically, based on the adaptive feature weighting mechanism of the sensor's spatial position and combined with the mechanical characteristics of the curtain wall structure, dynamic weights are assigned to sensors at different positions; S4: Damage detection and prediction, real-time application of the trained model for damage detection and prediction of the health status of the building curtain wall; S5: Visual display, presenting monitoring data and model analysis results to users in a graphical manner; S6: System management and monitoring, manage, control and monitor the entire monitoring system.

2. The method for visualizing building curtain wall monitoring data processing according to claim 1, characterized in that: The position weights of different position sensors in step S3 are: , where For the The spatial weight of each sensor; For the The Euclidean distance from each sensor to the nearest edge of the curtain wall; For the The Euclidean distance from each sensor to the nearest edge of the curtain wall; is the maximum value of all sensor edge distances; is the spatial sensitivity coefficient; is the total number of sensors.

3. The method for visualizing building curtain wall monitoring data processing according to claim 1, characterized in that: The extreme learning machine algorithm initializes the weight and bias parameters of the extreme learning machine hidden layer nodes based on the biopheromone decay mechanism, dynamically adjusts the weight and bias of each node according to the characteristic distribution of input data, and accepts random noise for correction.

4. The method for visualizing building curtain wall monitoring data processing according to claim 1, characterized in that: The extreme learning machine algorithm adopts nonlinear resonance constraint to the hidden layer activation function, and the activation function is expressed as: , where is the output of the hidden layer activation function of the extreme learning machine, is the hidden layer activation function of the extreme learning machine; is the weight of the hidden layer node of the extreme learning machine; is the bias of the hidden layer nodes of the extreme learning machine; It is the input of the activation function, representing the feature vector corresponding to the input data; is the resonance effect intensity; is the resonance effect frequency.

5. The method for visualizing building curtain wall monitoring data processing according to claim 4, characterized in that: The resonance effect intensity and the resonance effect frequency are expressed as follows: , , where For the The intensity of the resonance effect in the iterative training controls the oscillation effect in the nonlinear mapping; For the The resonance effect frequency in the iterative training controls the resonance effect frequency; Tune hyperparameters for first resonance effect strength; Tuning hyperparameters for the second resonance effect strength; Adjust hyperparameters for the first resonant effect frequency; Adjust hyperparameters for the second resonance effect frequency; For the The actual labels of the samples; The first samples.

6. The method for visualizing building curtain wall monitoring data processing according to claim 1, characterized in that: The extreme learning machine algorithm is based on a dynamic regularization mechanism of node activity, and dynamically adjusts the regularization coefficient of the hidden layer nodes by calculating the activation variance and covariance matrix of each hidden layer node.

7. The method for visualizing building curtain wall monitoring data processing according to claim 1, characterized in that: The extreme learning machine algorithm performs an update calculation of the biopheromone attenuation factor based on the relative change of the loss function gradient, which is expressed as: , where For the The biopheromone decay factor of the iteration; For the The biopheromone decay factor of the iteration; is the decay rate control parameter; For extreme learning machine The gradient of the loss function at iterations; For extreme learning machine The gradient of the loss function at iterations; is the loss function of the extreme learning machine.

8. The method for visualizing building curtain wall monitoring data processing according to claim 1, characterized in that: The loss function of the extreme learning machine is calculated as follows: , where is the loss function of the extreme learning machine; is the number of training samples; Represents the output value of the hidden layer activation function of the extreme learning machine; is the weight matrix of the hidden layer nodes of the extreme learning machine; The first samples; is the bias of the hidden layer nodes of the extreme learning machine; No. The actual labels of the samples; Characterizes the square of the prediction error of a single sample; is the dynamic regularization term; is the consistency constraint strength; is the consistency constraint term for structural damage propagation.

9. A building curtain wall monitoring data processing visualization system, based on the building curtain wall monitoring data processing visualization method according to any one of claims 1 to 8, comprising a data acquisition unit, a data preprocessing unit, a machine learning modeling unit, a damage detection and prediction unit, a visualization display unit, and a system management and monitoring unit, wherein: The data acquisition unit is used to collect health monitoring data of building curtain walls in real time; The data preprocessing unit is used to perform denoising, standardization and normalization on the collected raw data; The machine learning modeling unit is used to learn the health status pattern of the building curtain wall through training data based on the extreme learning machine algorithm, build a classifier model, and obtain the health status classification prediction result; The damage detection and prediction unit is used to evaluate the real-time monitoring data based on the machine learning model and output the prediction results of the health status; The visualization display unit is used to present the monitoring data and model analysis results to the user in a graphical manner; The system management and monitoring unit is used to manage, control and monitor the entire monitoring system.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions for a method for visualizing building curtain wall monitoring data processing, and the program instructions for visualizing building curtain wall monitoring data processing can be executed by one or more processors to implement the steps of the method for visualizing building curtain wall monitoring data processing as described in any one of claims 1-8.

Images