Method for predicting peak wind pressure of building glass panels based on ultra-precision convolutional neural network

Abstract

The present invention relates to a method for predicting the peak wind pressure of building glass panels based on an ultra-precision convolutional neural network. The ultra-precision convolutional neural network SRCNN is used to reconstruct the ultra-precision pressure spatial distribution, and the spatially sparse pressure measurement values are reconstructed into an ultra-precision pressure distribution, and the peak wind pressure within the glass panel domain is predicted using the ultra-precision pressure distribution. SRCNN simplifies and retains the generator module of the ultra-precision generative adversarial network SRGAN to generate a pressure spatial data distribution close to the true measurement values; based on the high-precision wind tunnel wind pressure spatial measurement data, neural network training is carried out to generate an ultra-precision reconstruction model, and the sparse pressure spatial distribution is input and reconstructed into an ultra-precision pressure spatial distribution similar to the true high-precision wind tunnel measurement values; the area integration method is used to obtain the mean wind pressure coefficient; based on the Cook-Mayne method, the peak pressure coefficient with a 78% non-exceedance probability of the glass panel is calculated to obtain the design value of the surface average pressure coefficient.

Description

Technical Field

[0001] The present invention relates to a method for predicting the peak wind pressure of a glass curtain wall, specifically to a method for predicting the peak wind pressure of glass panels on a building facade based on an ultra-precision convolutional neural network, belonging to the technical field of ultra-precision reconstruction of wind pressure on the facade of high-rise buildings and prediction of peak wind pressure of glass panels. Background Art

[0002] In modern cities, glass curtain walls are widely used in the outer enclosure structures of high-rise buildings. While achieving the lightening of the structure, the potential hidden danger of local strong suction on the glass surface under the action of the wind field cannot be ignored. Accurately predicting the wind pressure of the glass curtain wall is extremely crucial for building wind resistance and disaster prevention. Wind tunnel experiments are widely used to predict the extreme values of the average wind pressure on the outer wall panels of buildings. In principle, this value can be directly calculated by integrating the pressure distribution in the panel domain; however, accurately measuring the pressure distribution of the panel requires a relatively dense arrangement of pressure sensors, which is often difficult to achieve in wind tunnel experiments due to cost and technical limitations. Considering that the experimental measurement points are generally sparse and there is no temporal correlation between the measurement points, Newberry et al. proposed an assumption in their report that the duration of the pressure extreme value is proportional to its spatial range, and determined the relationship between the time-averaging window τ and the ratio of the reference length L of the pressure calculation area to the reference wind speed V, that is, the TVL equation:

[0003]

[0004] Among them, the constant K represents the exponential decay factor of the spatial correlation function of the pressure signal. In its full-scale pressure measurement, K = 4.5 is set, and the time-moving average pressure of a single measurement point within the time-averaging window τ is approximately used to replace the average pressure within the space with a reference length of L; the advantage of the TVL method is that it uses a simple expression to reveal the time filtering effect of the surface average pressure, but its basic assumption has not been fully verified, and its prediction result is extremely dependent on the selection of the K value in the TVL equation and the selection of the filtering measurement point position. Therefore, using the traditional TVL method to predict the surface average pressure is prone to large errors and the operation has uncertainty. Considering the limitations of the current pressure prediction methods, the team of Politecnico di Milano carried out high-precision spatial wind pressure measurements in its wind tunnel laboratory for key blocks on the facade of high-rise buildings, recorded the time series of pressure values at each measurement point in the block within the wind direction angle range from -15° to +30° and from +150° to -135°, sorted out a pressure database, and calculated the relatively accurate and reliable panel average pressure coefficient by integrating based on the high-precision measurement point values. However, up to now, due to the high cost and measurement technical difficulty, high-spatial-precision wind tunnel experiment data is still scarce. Therefore, under the premise that the setting of pressure measurement points is relatively sparse, improving the accuracy of peak pressure assessment has become a major challenge.

[0005] In recent years, machine learning methods have been gradually applied to solve complex problems related to the wind pressure effect of building structures, which has improved the efficiency of wind resistance prediction of structures to a certain extent. As one of the representatives of machine learning algorithms, Convolutional Neural Networks (CNN) has certain advantages in capturing the spatial distribution characteristics of measured point values. Recently, certain achievements have been made in the application of spatial super-resolution research in many fields, such as satellite image reconstruction, downscaling analysis of turbulence, and local precipitation forecasting. In the field of high-precision reconstruction of building facade pressure values, convolutional neural networks also have certain application potential, and are expected to reconstruct an instantaneous high-precision pressure distribution, so as to directly calculate the peak pressure of the surface average and replace the existing approximate methods for peak pressure prediction. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for predicting the peak wind pressure of building glass panels based on a super-resolution convolutional neural network.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A method for predicting the peak wind pressure of building glass panels based on a super-resolution convolutional neural network, characterized in that: based on the super-resolution convolutional neural network SRCNN, a super-resolution pressure spatial distribution is reconstructed, and the spatially sparse pressure measurement values are reconstructed into a super-resolution pressure distribution, and the peak wind pressure within the glass panel domain is predicted using the super-resolution pressure distribution, including the following steps:

[0009] (1) Obtain the training and test sample sets of the pressure distribution, and preprocess the obtained training samples to obtain the training sample pairs of high-precision and low-precision pressure distributions;

[0010] (2) Determine the super-resolution convolutional neural network architecture. The convolutional neural network is simplified from the super-resolution generative adversarial network SRGAN, including an input layer, a generator, and an output layer. The generator is used to generate a pressure spatial data distribution close to the real measured values, and adopts a deep fully convolutional architecture, including m consecutive residual blocks with skip connections. Each residual block includes a transposed convolutional layer, followed by a rectified linear unit ReLU layer as the activation function; the transposed convolutional layer, including a convolutional kernel with a size of K×K, is used for upsampling the low-dimensional spatial features; the residual block structure is beneficial to alleviating the disappearance of gradient descent during the training stage;

[0011] (3) For each wind direction angle, train the super-resolution convolutional neural network in step (2), initialize the weights and biases of each deconvolution layer, and iteratively optimize the network model by minimizing the loss function. After multiple iterations of training, obtain the optimal solutions for each weight and bias in the network model to get the optimal SRCNN model. Use the trained SRCNN model to reconstruct the pressure measurement values with rough spatial accuracy on the key stress-bearing blocks of the building facade into a super-accurate pressure distribution.

[0012] (4) Based on the super-accurate pressure reconstruction results, use the area integral method to obtain the mean wind pressure within the glass panel domain. Within a sufficient pressure time history, calculate the peak pressure coefficient with a 78% non-exceedance probability for the glass panel based on the Cook-Mayne method to obtain the design value of the surface mean pressure coefficient.

[0013] Furthermore, in the above method for predicting the peak wind pressure of building glass panels based on a super-resolution convolutional neural network, for training the SRCNN network model to achieve super-accurate pressure reconstruction of the key stress-bearing blocks on the building facade under different wind direction conditions, for step (1), use the high-spatial-accuracy pressure database measured on the surface of the building model in the atmospheric boundary layer wind tunnel experiment of the Polytechnic University of Milan as the training and test set for the SRCNN. The database is a public database that records the spatial and temporal pressure changes in the key stress-bearing areas of the building facade at different wind direction angles, namely the blocks near the top corner of the facade and the blocks near the edge of the middle height of the building.

[0014] Furthermore, in the above method for predicting the peak wind pressure of building glass panels based on a super-resolution convolutional neural network, in step (1), the dimension of the real high-precision pressure measurement data is N×M×L, where N represents the number of training samples, that is, the training set contains the pressure measurement values at N consecutive time steps, and M×L represents the dimension of the pressure measurement points with real spatial accuracy. Divide the M×L real measurement data at each time step into M / S×L / S groups of measurement blocks with a dimension of S×S. Extract the pressure data of the measurement point farthest from the building sidewall within each measurement block to form relatively sparse pressure measurement data with a dimension of N×M / S×L / S as the low-precision input for the SRCNN, where S is the downsampling ratio coefficient from high-precision measurement points to low-precision measurement points. At the corresponding time step, use the real pressure data as the high-precision label for the low-precision pressure input.

[0015] Furthermore, in the above method for predicting the peak wind pressure of building glass panels based on a super-resolution convolutional neural network, in step (3), the loss function L(x, y) is the mean square error function. where y i represents the real high-precision measurement value, and x i is the low-precision input, G(x i) is the super-resolution pressure distribution reconstructed by the SRCNN model at the corresponding time step, and the Adam optimization is used to minimize the training loss function.

[0016] Furthermore, in the above method for predicting the peak wind pressure of building glass panels based on a super-resolution convolutional neural network, in step (3), a S-fold super-resolution reconstruction model S×SRCNN is generated through training.

[0017] Furthermore, in the above method for predicting the peak wind pressure of building glass panels based on a super-resolution convolutional neural network, in step (4), based on the pressure distribution reconstructed by super-resolution, the area integral method is adopted to obtain the surface-averaged pressure coefficients of different glass panel domains within the key stress-bearing blocks on the facade at each instant.

[0018] where n is the total number of pressure measurement points within the panel domain at the true accuracy, C p,i represents the pressure coefficient predicted by SRCNN at the i-th measurement point, A i is the influence area of the i-th measurement point, and A is the total area of the glass panel domain.

[0019] Furthermore, in the above method for predicting the peak wind pressure of building glass panels based on a super-resolution convolutional neural network, in step (4), based on the Cook-Mayne method, the design values of the surface-averaged pressure coefficients are obtained, and the total duration of pressure prediction is divided into n′ groups. Each group at full scale contains 10-minute time history data of the surface-averaged pressure coefficients (C p，AA ), and the design pressure coefficient value with a 78% non-exceedance probability is expressed as: where m and σ represent the mean value and the root mean square value of C p，AA in n′ sample groups respectively.

[0020] Furthermore, the above method for predicting the peak wind pressure of building glass panels based on a super-resolution convolutional neural network specifically includes the following steps:

[0021] (1) Based on the wind pressure measurement value database of a high-spatial-accuracy wind tunnel experiment, obtain the training and test sample sets of the pressure distribution, and preprocess the obtained training samples to obtain the training sample pairs of high-precision and low-precision pressure distributions.

[0022] 1a) To train the SRCNN network model and achieve ultra-precise reconstruction of the pressure in the key stress areas of the building facade under different wind directions, a high-spatial-precision pressure database measured on the surface of a high-rise building model in the PoliMi wind tunnel experiment of the Milan Polytechnic University's atmospheric boundary layer is used as the training and test set for the SRCNN; the database records the spatial and temporal pressure changes in the key stress areas of the building facade at -15° to +30° and +150° to -135° under different wind direction angles, namely block A near the top corner of the facade and block B near the edge at the middle height of the building.

[0023] 1b) For each wind direction angle, the database collects the high-precision pressure distribution at 150,000 consecutive time steps. The wind pressure data of the first 60,000 time steps are selected to generate the training set, and the data of the subsequent 90,000 time steps are used as the test set.

[0024] 1c) On the two blocks of the facade, 200 pressure measurement points are densely arranged respectively; during the preprocessing process, the pressure data of 224 and 166 original resolution pressure measurement points closest to the building edge in block A and block B are used as the output labels of the SRCNN model. The distance between adjacent pressure measurement points in the PoliMi wind tunnel experiment is very small, and the minimum distance from the measurement point to the building edge is 3 mm. When moving away from the edge, the measurement point spacing gradually increases; considering the feasibility of convolution operations during the model training stage, the two-dimensional spline interpolation method is used to complement the missing pressure measurement values, and target output matrices with dimensions of 15×15 and 15×12 are formed for block A and block B respectively.

[0025] 1d) For block A and block B, the 15×15 and 15×12 target output data of each time step sample are respectively divided into groups of 5×5 and 5×4 matrices with dimensions of 3×3. The pressure data of the measurement points farthest from the building side wall are extracted within each block to form relatively sparse pressure measurement point data, with dimensions of 5×5 and 5×4 respectively, as the low-precision input of the SRCNN; the downsampling ratio coefficient of high-precision measurement points to low-precision measurement points is S = 3.

[0026] (2) Determine the ultra-precise convolutional neural network architecture.

[0027] 2a) Determine the number of residual block layers m and the convolution kernel size K×K of the convolutional layer. The number of residual blocks and the size of the convolution kernel determine the scale of the SRCNN. The larger the number of residual blocks and the convolution kernel size, the stronger the non-linear learning ability of the SRCNN, which will lead to too many network parameters and the model is prone to overfitting; the smaller the number of residual blocks and the convolution kernel size, the weaker the non-linear learning ability and the model is underfitted. The number of residual block layers of the SRCNN is determined by experimental adjustment to be m = 16, and the convolution kernel size is determined to be K×K = 3×3.

[0028] 2b) The trained SRCNN model reconstructs a 15×15 super-resolution pressure matrix based on a 5×5 low-precision pressure matrix; this SRCNN model is called a triple super-precision model, denoted as 3×SRCNN;

[0029] (3) For each wind direction angle, train the SRCNN model in step (2);

[0030] 3a) Initialize the weights and biases of SRCNN;

[0031] 3b) Determine the training parameters and optimization method of SRCNN. Given a learning rate of 0.0001, a batch size of 16, and the training loss function L(x, y) as the mean squared error function,

[0032]

[0033] where, y i represents the true high-precision measured point value, x i is the low-precision input, and G(x i ) is the super-precision pressure distribution reconstructed by the SRCNN model at the corresponding time step; minimize the training loss function using Adam optimization; use early stopping technology during training, with a tolerance of 10;

[0034] 3c) Using the trained SRCNN model, reconstruct the pressure measurements with rough spatial accuracy at 90,000 time steps in the test set into a super-precision pressure distribution;

[0035] (4) Based on the super-precision reconstructed pressure distribution, predict the peak value of the mean pressure coefficient on the inner surface of the glass panel domain and the design value of the pressure coefficient;

[0036] 4a) Using the area integration method, obtain the mean pressure coefficient of different glass panel domains within the key force-bearing blocks on the facade at each instant,

[0037] where, n is the total number of pressure measurement points within the panel domain at the true accuracy, C p,i represents the pressure coefficient predicted by SRCNN at the i-th measurement point, A i is the influence area of the i-th measurement point, and A is the total area of the glass panel domain;

[0038] 4b) Within a sufficient pressure time history, based on the Cook-Mayne method, calculate the peak pressure coefficient with a 78% non-exceedance probability for the glass panel and obtain the design value of the mean pressure coefficient; divide the total duration of pressure prediction into 6 groups. At full scale, each group contains 10 minutes of mean pressure coefficient (C p，AA ) time history data, and the design pressure coefficient value with a 78% non-exceedance probability is expressed as: Among them, m and σ respectively represent the average value and the root mean square value of C in 6 sample groups p，AA of the 6 sample groups

[0039] Compared with the prior art, the present invention has remarkable advantages and beneficial effects, which are specifically embodied in the following aspects:

[0040] ① The present invention uses a super-precision convolutional neural network to reconstruct spatially sparse pressure measurement values into a super-precision pressure distribution, and predicts the peak wind pressure design value within the glass panel area of high-rise buildings based on the super-precision pressure distribution data;

[0041] ② The present invention makes up for the deficiencies such as high working cost and great technical difficulty of high-precision pressure measurement, and improves the efficiency of obtaining the super-precision pressure distribution on the facade of high-rise buildings; compared with the direct integration method of sparse pressure data, the peak value of the pressure coefficient calculated by the present invention based on the super-precision reconstructed pressure can more accurately reflect the true peak value of the surface average pressure coefficient;

[0042] ③ Compared with the traditional TVL method, since the present invention fully considers the spatial distribution characteristics of the pressure on the building facade, the prediction accuracy of the designed value of the surface average pressure coefficient obtained is improved; and because it is not restricted by the selection of time-filtered measurement points and the filtering duration, the robustness of the prediction of the peak value of the surface average pressure coefficient inside the glass panel is stronger.

[0043] Other features and advantages of the present invention will be described in the subsequent specification, and part of them will become obvious from the specification, or be understood by implementing the specific embodiments of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the structures specifically pointed out in the written specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0045] Figure 1 : Schematic flow diagram of the present invention;

[0046] Figure 2a : Schematic diagram of the building model of the high-precision wind tunnel experiment pressure measurement value dataset;

[0047] Figure 2b : Schematic diagram of the wind direction angle conditions involved in the high-precision wind tunnel experiment pressure measurement value dataset;

[0048] Figure 2c: Schematic diagram of high-precision measured points and low-precision measured point inputs of the neural network within dataset block A of the pressure measurement values in the high-precision wind tunnel experiment;

[0049] Figure 2d : Schematic diagram of high-precision measured points and low-precision measured point inputs of the neural network within dataset block B of the pressure measurement values in the high-precision wind tunnel experiment;

[0050] Figure 3 : Schematic diagram of the ultra-precision convolutional neural network structure of the present invention;

[0051] Figure 4a : Comparison chart of the prediction errors of the existing TVL method and the method of the present invention for the peak pressure coefficient within the glass panel area in block A;

[0052] Figure 4b : Comparison chart of the prediction errors of the existing TVL method and the method of the present invention for the peak pressure coefficient within the glass panel area in block B;

[0053] Figure 5a : Comparison chart of the prediction errors of the existing TVL method and the method of the present invention for the designed pressure coefficient value within the glass panel area in block A;

[0054] Figure 5b : Comparison chart of the prediction errors of the existing TVL method and the method of the present invention for the designed pressure coefficient value within the glass panel area in block B. Detailed implementation manners

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Usually, the components of the embodiments of the present invention described and shown here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed present invention, but only represents the selected embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0056] It should be noted that: similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present invention, orientation terms and order terms are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0057] When the layout of wind pressure measurement points is relatively sparse, aiming at the limitations of the traditional TVL method in predicting the surface average pressure in the glass panel area, the present invention uses the super-resolution convolutional neural network SRCNN to predict the design value of the peak wind pressure on the glass curtain wall of high-rise buildings based on the super-accurate pressure distribution data, so as to improve the accuracy of peak pressure coefficient prediction and enhance the robustness of prediction to meet the requirements of actual engineering applications.

[0058] First, on the key stress-bearing blocks of the building facade, use the convolutional neural network to perform super-accurate reconstruction of the pressure distribution; then, based on the super-accurate pressure distribution on the blocks, predict the peak pressure within the glass panel area; during the prediction process, it is the peak pressure of the glass panel within the range of the key stress-bearing blocks of the building facade.

[0059] As Figure 1 shown, the method for predicting the peak wind pressure of the glass panel on the building facade based on the super-resolution convolutional neural network reconstructs the super-accurate pressure spatial distribution based on the super-resolution convolutional neural network SRCNN, reconstructs the spatially sparse pressure measurement values into a super-accurate pressure distribution, and uses the super-accurate pressure distribution to predict the peak wind pressure within the glass panel area. The steps are as follows:

[0060] (1) Obtain the training and test sample sets of the pressure distribution, preprocess the obtained training samples, and obtain the training sample pairs of high-precision and low-precision pressure distributions;

[0061] (2) Determine the super-resolution convolutional neural network architecture. The convolutional neural network is simplified from the super-resolution generative adversarial network SRGAN, including an input layer, a generator, and an output layer. The generator is used to generate a pressure spatial data distribution close to the real measurement value, and adopts a deep fully convolutional architecture, including m consecutive residual blocks with skip connections. Each residual block includes a transposed convolutional layer, followed by a rectified linear unit ReLU layer as the activation function; the transposed convolutional layer, including a convolutional kernel of size K×K, is used to upsample the low-dimensional spatial features; the residual block structure is beneficial to alleviating the disappearance of gradient descent during the training stage;

[0062] (3) For each wind direction angle, train the super-resolution convolutional neural network in step (2), initialize the weights and biases of each transposed convolutional layer, and use the method of minimizing the loss function to iteratively optimize the network model. After multiple iterative trainings, obtain the optimal solutions of each weight and bias in the network model, and obtain the optimal SRCNN model; use the trained SRCNN model to reconstruct the pressure measurement values with rough spatial accuracy on the key stress-bearing blocks of the building facade into a super-accurate pressure distribution;

[0063] (4) Based on the ultra-precise pressure reconstruction results, the area integration method is adopted to obtain the mean wind pressure within the glass panel domain. During a sufficient pressure time history, based on the Cook-Mayne method, the peak pressure coefficient with a 78% non-exceedance probability of the glass panel is calculated to obtain the design value of the surface mean pressure coefficient.

[0064] For training the SRCNN network model to achieve ultra-precise pressure reconstruction of the key stressed blocks on the building facade under different wind directions, for step (1), the high-spatial-precision pressure database measured on the surface of the building model in the atmospheric boundary layer wind tunnel experiment of the Politecnico di Milano is used as the training and test set of the SRCNN. The database records the spatial and temporal pressure changes in the key stressed areas on the building facade at different wind angles, namely the blocks near the top corner of the facade and the blocks near the edge at the middle height of the building.

[0065] In step (1), the dimension of the real high-precision pressure measurement data is N×M×L. N represents the number of training samples, that is, the training set contains the pressure measurement point values at N consecutive time steps, and M×L represents the pressure measurement point dimension of the real spatial precision. The real measurement data of M×L at each time step is divided into M / S×L / S groups of measurement point blocks with a dimension of S×S. The pressure data of the measurement point farthest from the building sidewall is extracted within each measurement point block to form relatively sparse pressure measurement data with a dimension of N×M / S×L / S as the low-precision input of the SRCNN, where S is the downsampling ratio coefficient from high-precision measurement points to low-precision measurement points. At the corresponding time step, the real pressure data is used as the high-precision label for the low-precision pressure input.

[0066] In step (2), the number of layers m of the residual block is determined by experimental adjustment. In step (2), the convolution kernel size K of the deconvolution layer is determined by experimental adjustment using the step-by-step reduction method.

[0067] In step (3), the loss function L(x, y) is the mean square error function. where y i represents the real high-precision measurement point value, x i is the low-precision input, G(x i ) is the ultra-precise pressure distribution reconstructed by the SRCNN model at the corresponding time step, and the Adam optimization is used to minimize the training loss function.

[0068] In step (3), a S-fold ultra-precise reconstruction model S×SRCNN is generated through training.

[0069] In step (4), based on the pressure distribution reconstructed with ultra-precision, the area integration method is adopted to obtain the surface mean pressure coefficients of different glass panel domains within the key stressed blocks on the facade at each instant.

[0070] where n is the total number of pressure measurement points in the lower panel area of the true accuracy, C p,i represents the pressure coefficient predicted by SRCNN at the i-th measurement point, A i is the influence area of the i-th measurement point, and A is the total area of the glass panel area.

[0071] In step (4), based on the Cook-Mayne method, the design value of the surface-averaged pressure coefficient is obtained. The total duration of pressure prediction is divided into n′ groups. At full scale, each group contains 10-minute time history data of the surface-averaged pressure coefficient (C p，AA ), and the design pressure coefficient value with a 78% non-exceedance probability is expressed as: where m and σ represent the mean and root mean square values of C p，AA in n′ sample groups respectively.

[0072] The present invention utilizes the super-resolution convolutional neural network SRCNN to predict the peak wind pressure on the glass curtain wall of high-rise buildings based on the super-resolution pressure distribution data. SRCNN simplifies and retains the generator module Generator of the super-resolution generative adversarial network SRGAN to generate a pressure spatial data distribution close to the true measurement value. The generator adopts a deep fully convolutional architecture, including 16 consecutive residual blocks with skip connections. Each residual block includes a transposed convolutional layer followed by a rectified linear unit ReLU layer as the activation function; the transposed convolutional layer is essentially a convolutional operation for upsampling the low-dimensional space features, and the residual block structure helps to alleviate the problem of vanishing gradient descent during the training stage.

[0073] Based on the high-precision wind tunnel wind pressure spatial measurement data, the neural network is trained to generate a super-resolution reconstruction model. The model can, under the corresponding wind direction angle conditions, reconstruct the sparse pressure spatial distribution input into a super-resolution pressure spatial distribution similar to the true high-precision wind tunnel measurement value. The present invention adopts the area integral method based on the super-resolution wind pressure data at each instant to obtain the mean wind pressure coefficient within the glass panel area. Within a sufficient pressure time history, based on the Cook-Mayne method, the peak pressure coefficient with a 78% non-exceedance probability of the glass panel is calculated to obtain the design value of the surface-averaged pressure coefficient. Compared with directly integrating the sparse pressure data, the peak pressure coefficient calculated based on the super-resolution reconstructed pressure can better reflect the true surface-averaged peak. Compared with the traditional time filtering method (time-length-velocity, TVL), the prediction accuracy of the peak surface-averaged pressure coefficient is higher, and due to being not restricted by the time filtering measurement points and filtering duration, the prediction robustness is stronger.

[0074] In specific applications, the specific process is as follows:

[0075] (1) Based on the wind pressure measurement value database of high-spatial-precision wind tunnel experiments, obtain the training and test sample sets of the pressure distribution, and preprocess the obtained training samples to obtain the training sample pairs of high-precision and low-precision pressure distributions.

[0076] 1a) To train the SRCNN network model and achieve the super-precision reconstruction of the pressure on the key stressed blocks of the building facade under different wind direction conditions, use the high-spatial-precision pressure database on the surface of the high-rise building model measured in the PoliMi wind tunnel experiment of the Milan Polytechnic University's atmospheric boundary layer as the training and test sets of the SRCNN. This database is an open database (DOI: 10.5281 / zenodo.3906588); refer to Figure 2a 、 2b , the database records the spatial and temporal pressure changes in the key stressed areas of the building facade at different wind direction angles from -15° to +30° and +150° to -135°, that is, block A near the top corner of the facade and block B near the edge at the middle height of the building.

[0077] 1b) For each wind direction angle, the database collects the high-precision pressure distributions of 150,000 consecutive time steps. Select the wind pressure data of the first 60,000 time steps (true scale 2550s) to generate the training set, and the data of the subsequent 90,000 time steps (true scale 3825s) as the test set.

[0078] 1c) On the two blocks of the facade, 200 pressure measurement points are densely arranged respectively; during the preprocessing process, use the data of 224 and 166 original resolution pressure measurement points closest to the building edge in block A and block B as the output labels of the SRCNN model, Figure 2c 、 2d as indicated by the asterisks in the model-scale block areas shown in Figure 2c 、 2d ; compared with most previous wind tunnel experiments, the distance between adjacent pressure measurement points in the PoliMi wind tunnel experiment is very small, and the minimum distance from the measurement point to the building edge is 3mm (equivalent to 0.15m in the original size). When far from the edge, the measurement point spacing gradually increases; considering the feasibility of the convolution operation in the model training stage, use the two-dimensional spline interpolation method to

[0079] fill in the missing pressure measurement values represented by the squares inFigure 2c , 2d as shown by the circular symbol in; the proportionality coefficient for downsampling high-precision measurement points to low-precision measurement points is S = 3;

[0080] (2) Determine the architecture of the super-resolution convolutional neural network;

[0081] 2a) Determine the number of residual block layers m and the convolution kernel size K×K of the convolutional layer. The number of residual blocks and the size of the convolution kernel determine the scale of SRCNN. The larger the number of residual blocks and the convolution kernel size, the stronger the non-linear learning ability of SRCNN, which will lead to too many network parameters and the model is prone to overfitting; the smaller the number of residual blocks and the convolution kernel size, the weaker the non-linear learning ability and the model is underfitting. The number of residual block layers of SRCNN is determined by experimental adjustment to be m = 16, and the convolution kernel size is determined to be K×K = 3×3;

[0082] The trained SRCNN model reconstructs a 15×15 super-resolution pressure matrix based on a 5×5 low-precision pressure matrix; this SRCNN model is called a triple super-precision model, denoted as 3×SRCNN;

[0083] The constructed SRCNN structure is as Figure 3 shown;

[0084] (3) For each wind direction angle, train the SRCNN model in step (2);

[0085] 3a) Initialize the weights and biases of SRCNN;

[0086] 3b) Determine the training parameters and optimization method of SRCNN. Given a learning rate of 0.0001, a batch size of 16, and the training loss function L(x, y) is the mean square error function,

[0087]

[0088] where y i represents the true high-precision measurement point value, x i is the low-precision input, and G(x i ) is the super-precision pressure distribution reconstructed by the SRCNN model at the corresponding time step; minimize the training loss function using Adam optimization; use the early stopping technique during training with a tolerance of 10;

[0089] 3c) Use the trained SRCNN model to reconstruct the pressure measurements with rough spatial accuracy at 90,000 time steps (true scale 3825 s) in the test set into a super-precision pressure distribution;

[0090] (4) Based on the super-precision reconstructed pressure distribution, predict the peak value of the mean pressure coefficient on the inner surface of the glass panel domain and the design value of the pressure coefficient;

[0091] 4a) By using the area integral method, the surface-averaged pressure coefficients of different glass panel regions within the key stressed blocks of the lower facade at each instant are obtained.

[0092] Among them, n is the total number of pressure measurement points within the panel region under the true accuracy, and C p,i represents the pressure coefficient predicted by SRCNN at the i-th measurement point, and A i is the influence area of the i-th measurement point, and A is the total area of the glass panel region.

[0093] 4b) During a sufficient pressure time history, based on the Cook-Mayne method, the peak pressure coefficient with a 78% non-exceedance probability of the glass panel is calculated to obtain the design value of the surface-averaged pressure coefficient; the total duration of pressure prediction is divided into 6 groups, and each group at full scale contains 10-minute surface-averaged pressure coefficient (C p，AA ) time history data, and the design pressure coefficient value with a 78% non-exceedance probability is expressed as: Among them, m and σ respectively represent the mean value and root mean square value of C p，AA in the 6 sample groups.

[0094] The training environment of the SRCNN model is Tensorflow-1.14.0-CPU.

[0095] During the test, the wind direction angles that are likely to cause extreme negative pressure in each block of the building facade, namely block A at the +10° wind direction angle and block B at the 180° wind direction angle, are used to evaluate the prediction performance of the proposed solution of the present invention for the peak value of the surface-averaged pressure coefficient and the design value of the pressure coefficient within the glass panel region. The wind pressure prediction is respectively carried out by using the method of the present invention, the solution directly based on sparse pressure measurement points, and the traditional TVL method, and compared with the results of the true surface-averaged wind pressure peak value and the design value.

[0096] The test focuses on analyzing the panel regions near the top corner in block A of the facade and near the middle height edge in block B. Two sizes of glass panel regions are considered within each block, namely square panels with full sizes of 1.5 m × 1.5 m and 3 m × 3 m. The above panel regions are likely to experience extreme positive or negative pressure events. Therefore, PoliMi wind tunnel experiments installed pressure measurement points with high spatial accuracy at the corresponding positions to accurately estimate the "true" surface-averaged pressure coefficient during the wind tunnel test. Based on the super-accuracy reconstruction data (C p,LR ), sparse pressure input data (C p,LR ), and the spatial-averaged pressure coefficients of the true high-precision measurement data (C p,AA ) are all defined by .

[0097] To statistically analyze the accuracy of peak pressure coefficient prediction, a threshold is set to identify extreme negative pressure.Figure 4a , 4b The box plots of Figure 4a and 4b illustrate the average error of each predicted peak relative to the true peak under different time periods. The error of the TVL method in predicting the panel peak pressure coefficient is much larger than that of the SRCNN method. For example, for Tile A, the relative error of the TVL method is around 20 - 30%, while that of the SRCNN is less than 10%. For Tile B, the relative error of the TVL method is around 6 - 10%, while that of the SRCNN is less than 2%. In addition, the prediction error of the TVL method has a more dispersed distribution due to different selected filtering measurement points or filtering time lengths, while the error of the SRCNN method is more robust.

[0098] As can be shown from Table 1 below, when the SRCNN model is trained with 10,000 sample numbers, compared with the prediction error of the peak pressure coefficient calculated directly based on low-precision measurement points (LR), the relative error of the prediction error of the present invention can be reduced by about 50%, indicating that the SRCNN method can more accurately predict the peak surface average pressure coefficient than directly integrating from the sparse pressure measurement points.

[0099] Table 1:

[0100]

[0101]

[0102] From Figure 5a , 5b it can be seen that for panels of different sizes in two key stress-bearing blocks, the relative error of the designed value of the pressure coefficient predicted based on the SRCNN model is lower than the relative error calculated by the TVL method.

[0103] In summary, the present invention uses a super-precision convolutional neural network to reconstruct the spatially sparse pressure measurement values into a super-precision pressure distribution, and predicts the designed value of the peak wind pressure in the glass panel area of high-rise buildings based on the super-precision pressure distribution data.

[0104] The present invention makes up for the deficiencies such as high working cost and great technical difficulty of high-precision pressure measurement, and improves the efficiency of obtaining the super-precision pressure distribution on the facade of high-rise buildings; compared with the method of directly integrating sparse pressure data, the peak of the pressure coefficient calculated based on the super-precision reconstructed pressure can more accurately reflect the true peak of the surface average pressure coefficient.

[0105] Compared with the traditional TVL method, the present invention improves the prediction accuracy of the designed value of the surface average pressure coefficient because it fully considers the spatial distribution characteristics of the pressure on the building facade; and because it is not restricted by the selection of time filtering measurement points and filtering time lengths, the prediction of the peak of the surface average pressure coefficient in the glass panel is more robust.

[0106] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention. It should be noted that like reference numerals and letters denote like items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0107] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present invention, and all of them should be covered within the protection scope of the present invention.

[0108] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising a..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

Claims

1. A method for predicting the peak wind pressure of building glass panels based on a super-precision convolutional neural network, characterized in that: Based on the super-resolution convolutional neural network SRCNN to reconstruct the super-resolution pressure spatial distribution, the spatially sparse pressure measurement values are reconstructed into a super-resolution pressure distribution, and the peak wind pressure within the glass panel domain is predicted using the super-resolution pressure distribution, including the following steps: (1) Obtain the training and test sample sets of the pressure distribution, and preprocess the obtained training samples to obtain the training sample pairs of high-precision and low-precision pressure distributions; (2) Determine the super-resolution convolutional neural network architecture. The convolutional neural network is simplified from the super-resolution generative adversarial network SRGAN, including an input layer, a generator, and an output layer. The generator is used to generate a pressure spatial data distribution close to the real measurement values. It adopts a deep fully convolutional architecture, including m consecutive residual blocks with skip connections. Each residual block includes a transposed convolutional layer, followed by a rectified linear unit ReLU layer as the activation function; the transposed convolutional layer, which contains a convolutional kernel of size K×K, is used to upsample the low-dimensional spatial features; the residual block structure helps to alleviate the vanishing gradient during the training stage; (3) For each wind direction angle, train the super-resolution convolutional neural network in step (2). Initialize the weights and biases of each transposed convolutional layer, and use the method of minimizing the loss function to iteratively optimize the network model. After multiple iterations of training, obtain the optimal solutions of each weight and bias in the network model to get the optimal SRCNN model; use the trained SRCNN model to reconstruct the pressure measurement values with rough spatial accuracy on the key stress-bearing blocks of the building facade into a super-resolution pressure distribution; (4) Based on the super-resolution pressure reconstruction result, use the area integration method to obtain the average wind pressure within the glass panel domain. Within a sufficient pressure time history, based on the Cook-Mayne method, calculate the peak pressure coefficient with a 78% non-exceedance probability for the glass panel, and obtain the design value of the surface average pressure coefficient.

2. The method for predicting the peak wind pressure of building glass panels based on a super-precision convolutional neural network according to claim 1, wherein: To train the SRCNN network model and achieve the super-resolution reconstruction of the pressure on the key stress-bearing blocks of the building facade under different wind direction conditions, for step (1), use the high-spatial-accuracy pressure database measured in the atmospheric boundary layer wind tunnel experiment of the Politecnico di Milano as the training and test set for SRCNN. The database is a public database, which records the spatial and temporal pressure changes in the key stress-bearing areas of the building facade at different wind direction angles, that is, the blocks near the top corner of the facade and the blocks near the edge of the middle height of the building.

3. The method for predicting the peak wind pressure of building glass panels based on a super-precision convolutional neural network according to claim 1, wherein: In step (1), the dimension of the real high-precision pressure measurement data is N×M×L, where N represents the number of training samples, that is, the training set contains the pressure measurement values of N consecutive time steps, and M×L represents the pressure measurement dimension of the real spatial accuracy; Divide the M×L real measurement data at each time step into M / S×L / S groups of measurement blocks with a dimension of S×S. Extract the pressure data of the measurement point farthest from the building side wall within each measurement block to form relatively sparse pressure measurement data with a dimension of N×M / S×L / S, which is used as the low-precision input of SRCNN, where S is the downsampling ratio coefficient from high-precision measurement points to low-precision measurement points; at the corresponding time step, use the real pressure data as the high-precision label of the low-precision pressure input.

4. The method for predicting the peak wind pressure of building glass panels based on a super-precision convolutional neural network according to claim 1, wherein: In step (3), the loss function \(L(x, y)\) is the mean square error function, where \(y\) i represents the measured values of true high-precision points, and \(x\) i is the low-precision input. \(G(x\) i ) is the super-resolution pressure distribution reconstructed by the SRCNN model at the corresponding time step. The Adam optimization is used to minimize the training loss function.

5. The method for predicting the peak wind pressure of building glass panels based on a super-precision convolutional neural network according to claim 1 or 4, characterized in that: In step (3), an S-fold super-resolution reconstruction model S×SRCNN is generated through training.

6. The method for predicting the peak wind pressure of building glass panels based on a super-precision convolutional neural network according to claim 1, wherein: In step (4), based on the pressure distribution reconstructed with ultra-high precision, the area integral method is adopted to obtain the area-averaged pressure coefficients of different glass panel domains within the key stress-bearing blocks on the lower facade at each instant. Among them, n is the total number of pressure measurement points in the lower panel area of the true accuracy, C p,i represents the pressure coefficient predicted by SRCNN at the i-th measurement point, A i is the influence area of the i-th measurement point, and A is the total area of the glass panel area.

7. The method for predicting the peak wind pressure of building glass panels based on a super-precision convolutional neural network according to claim 1 or 6, characterized in that: In step (4), based on the Cook-Mayne method, the design value of the surface-averaged pressure coefficient is obtained. The total duration of pressure prediction is divided into n′ groups, and each group at full scale contains 10-minute time history data of the surface-averaged pressure coefficient (C p，AA ). The design pressure coefficient value with a non-exceedance probability of 78% is expressed as: where m and σ represent the mean value and the root mean square value of C p，AA in n′ sample groups, respectively.

8. The method for predicting the peak wind pressure of building glass panels based on a super-precision convolutional neural network according to claim 1, wherein: Specifically, it includes the following steps: (1) Based on the database of wind pressure measurement values from high-spatial-accuracy wind tunnel experiments, obtain the training and test sample sets of the pressure distribution, and preprocess the obtained training samples to obtain the training sample pairs of high-precision and low-precision pressure distributions. 1a) To train the SRCNN network model and achieve the super-resolution reconstruction of the pressure on the key stress blocks of the building facade under different wind direction conditions, use the high-spatial-accuracy pressure database on the surface of the high-rise building model measured in the PoliMi wind tunnel experiment of the Milan Polytechnic University's atmospheric boundary layer as the training and test set of SRCNN; the database records the spatial and temporal pressure changes of the key stress areas on the building facade at -15° to +30° and +150° to -135° under different wind direction angles, that is, block A near the top corner of the facade and block B near the edge of the middle height of the building. 1b) For each wind direction angle, the database collects 150,000 consecutive time steps of high-precision pressure distributions, selects the wind pressure data of the first 60,000 time steps to generate the training set, and the data of the subsequent 90,000 time steps as the test set. 1c) On the two blocks of the facade, 200 pressure measurement points are densely arranged respectively; during the preprocessing process, use the pressure data of 224 and 166 original resolution pressure measurement points closest to the building edge in block A and block B as the output labels of the SRCNN model. The spacing between adjacent pressure measurement points in the PoliMi wind tunnel experiment is very small, and the minimum distance from the measurement point to the building edge is 3 mm. When far from the edge, the measurement point spacing gradually increases; considering the feasibility of convolution operations during the model training stage, use the two-dimensional spline interpolation method to fill in the missing pressure measurement values, and form target output matrices with dimensions of 15×15 and 15×12 for block A and block B respectively. 1d) For block A and block B, divide the 15×15 and 15×12 target output data of each time step sample into groups of 5×5 and 5×4 matrix blocks with a dimension of 3×3 respectively, and extract the pressure data of the measurement point farthest from the building side wall in each block to form relatively sparse pressure measurement point data, with dimensions of 5×5 and 5×4 respectively, as the low-precision input of SRCNN; the proportionality coefficient for downsampling high-precision measurement points to low-precision measurement points is S = 3. (2) Determine the super-resolution convolutional neural network architecture. 2a) Determine the number of residual blocks m and the convolution kernel size K×K of the convolutional layer. The number of residual blocks and the size of the convolution kernel determine the scale of SRCNN. The larger the number of residual blocks and the convolution kernel size, the stronger the non-linear learning ability of SRCNN, which will lead to too many network parameters and the model is prone to overfitting; the smaller the number of residual blocks and the convolution kernel size, the weaker the non-linear learning ability and the model is underfitting; the number of residual blocks of SRCNN is determined to be m = 16 through experimental adjustment, and the convolution kernel size is determined to be K×K = 3×3. 2b) The trained SRCNN model reconstructs a 15×15 super-resolution pressure matrix based on a 5×5 low-precision pressure matrix; this SRCNN model is called a triple super-precision model, denoted as 3×SRCNN; (3) For each wind direction angle, train the SRCNN model in step (2); 3a) Initialize the weights and biases of SRCNN; 3b) Determine the training parameters and optimization method of SRCNN, with a given learning rate of 0.0001, a batch size of 16, and the training loss function L(x,y) being the mean squared error function, Among them, y i represents the measured point value of true high precision, and x i is the input of low precision. G(x i ) is the super-precision pressure distribution reconstructed by the SRCNN model at the corresponding time step. The Adam optimization is used to minimize the training loss function. The early stopping technique is used during training, and the tolerance is 10; 3c) Use the trained SRCNN model to reconstruct the pressure measurements with rough spatial accuracy at 90,000 time steps in the test set into a super-precision pressure distribution; (4) Based on the super-precision reconstructed pressure distribution, predict the peak value of the mean pressure coefficient on the inner surface and the design value of the pressure coefficient within the glass panel domain; 4a) By using the area integral method, the surface average pressure coefficients of different glass panel regions within the key force-bearing blocks on the lower facade at each instant are obtained. Among them, n is the total number of pressure measurement points in the lower panel area of the true accuracy, C p,i represents the pressure coefficient predicted by SRCNN at the i-th measurement point, A i is the influence area of the i-th measurement point, and A is the total area of the glass panel area; 4b) Based on the Cook-Mayne method, calculate the peak pressure coefficient with a 78% non-exceedance probability for the glass panel within a sufficient pressure duration to obtain the design value of the surface-averaged pressure coefficient; divide the total duration of pressure prediction into 6 groups, and each group at full scale contains 10-minute surface-averaged pressure coefficient (C p，AA ) time history data, and the design pressure coefficient value with a 78% non-exceedance probability is expressed as: where m and σ represent the mean value and the root mean square value of C p，AA in the 6 sample groups respectively.

Images