A wind field reconstruction method for large-span spatial structures based on semi-supervised ensemble learning

Abstract

A method for wind field reconstruction of large-span spatial structures based on semi-supervised ensemble learning is proposed. It is proposed to use wind pressure data to obtain the features reflecting wind speed and relative wind direction, and combine the semi-supervised ensemble learning method to achieve high-precision and high-robustness reconstruction of incoming wind fields of large-span spatial structures. The specific implementation process is as follows: A. Preprocess the wind pressure and wind field monitoring data of large-span spatial structures to establish a total data set; B. Use the two-stage K-means clustering algorithm to extract clustering features reflecting wind speed from the average wind pressure data; C. Use the Isomap manifold learning algorithm to extract manifold features reflecting relative wind direction from the average wind pressure coefficient data; D. Add the clustering features and manifold feature data to the total data set to construct a training data set and a test data set; E. Train a semi-supervised ensemble learning model based on the two-stage K-means clustering algorithm and the Isomap manifold learning algorithm feature extraction, input the wind pressure distribution features without corresponding wind fields, and reconstruct the incoming wind field of large-span spatial structures that are missing.

Description

Technical Field

[0001] The invention relates to a large-span space structure wind field reconstruction method, in particular to a large-span space structure incoming flow wind field reconstruction method based on wind pressure distribution feature extraction and semi-supervised ensemble learning, and belongs to the field of structural wind monitoring. Background Art

[0002] Large-span space structures usually have the characteristics of light weight, high flexibility and low damping. They are prone to large deformation and vibration under wind loads and are typical wind-sensitive structures. Wind load is one of the control factors for the service safety of large-span space structures, and its distribution is directly affected by the incoming wind field. Based on the characteristics of the incoming wind field, combined with wind tunnel tests, numerical simulations and field measurements, structural wind-induced effect analysis is an important means to guide the wind-resistant design and service performance evaluation of large-span space structures. Therefore, accurately obtaining the incoming wind field of large-span space structures has theoretical significance and engineering practical value for ensuring the wind-resistant safety of the structure.

[0003] Since the wind field around the large-span space structure is easily disturbed by the roof airflow, and there are significant differences in the wind field measurement results at different spatial locations, there is a certain deviation between the wind field information obtained by direct measurement and the incoming wind field. Therefore, it is difficult to ensure the accuracy of the results by only using wind field data to reconstruct the incoming wind field. Considering that the incoming wind field and structural wind pressure are respectively used as the input and output of the structural system, the wind pressure distribution of the large-span space structure can be used to reconstruct the incoming wind field. By fully mining the existing wind pressure monitoring data, more accurate incoming wind field information can be provided for the analysis of the wind resistance performance of the large-span space structure. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes a large-span spatial structure wind field reconstruction method based on semi-supervised ensemble learning to improve the accuracy and robustness of large-span spatial structure wind field reconstruction. The specific contents include:

[0005] A method for wind field reconstruction of a large-span spatial structure based on semi-supervised ensemble learning comprises the following steps:

[0006] A. Preprocess the wind pressure and wind field monitoring data of large-span spatial structures to establish a total data set;

[0007] B. Use a two-stage K-means clustering algorithm to extract clustering features that reflect wind speed from average wind pressure data;

[0008] C. Using the Isomap manifold learning algorithm to extract manifold features reflecting relative wind direction from the average wind pressure coefficient data;

[0009] D. Add clustering features and manifold feature data to the total data set to construct training data sets and test data sets;

[0010] E. Train a semi-supervised ensemble learning model based on the two-stage K-means clustering algorithm and the Isomap manifold learning algorithm for feature extraction, input the wind pressure distribution characteristics without corresponding wind fields, and reconstruct the incoming wind field with missing large-span spatial structures.

[0011] Further, step A specifically includes:

[0012] A1. Obtain wind pressure and wind field monitoring data for large-span spatial structures;

[0013] A2. According to the wind pressure time history data p(t i ) Calculate the average wind pressure p mean , pulsating wind pressure root mean square p rms , wind pressure coefficient C p (t i ), the calculation formula is:

[0014]

[0015] Where: n is the total number of wind pressure time series data samples, p(t i ) is t i Roof wind pressure value at the moment; p0 and p ∞ are the average total pressure and average static pressure at the reference point respectively; ρ is the air density; v ∞ is the incoming wind speed at the reference point.

[0016] A3. According to the wind pressure coefficient C p (t i ) Calculate the average wind pressure coefficient C p_mean , Pulsating wind pressure coefficient C p_rms , wind pressure deflection S k and wind pressure peak K u , the calculation formula is:

[0017]

[0018] A4. Use wind pressure distribution characteristics and incoming wind field characteristics to establish a total data set; wind pressure distribution characteristics are used as model input, including the average wind pressure p mean , pulsating wind pressure root mean square p rms , average wind pressure coefficient C p_mean , Pulsating wind pressure coefficient C p_rms , wind pressure deflection S k and wind pressure peak K u ; The incoming wind field characteristics are used as model output, including wind speed U and wind direction θ.

[0019] Further, step B specifically includes:

[0020] B1. Assume that there are M samples of average wind pressure distribution data to be analyzed, and each sample contains the average wind pressure of N measuring points of a large-span spatial structure under a certain wind field. Then the average wind pressure data set used for cluster analysis is:

[0021] P={P i |P i =(p i1 ,p i2 ,…,p iN ),i=1,2,…,M} (3)

[0022] B2. Use the K-means clustering algorithm to perform the first clustering on the input large-span spatial structure average wind pressure data, select k1 clusters, calculate the Euclidean distance from each wind pressure distribution sample point to its cluster center, and obtain the distance data set after clustering.

[0023] B3. Distance dataset D 1 As input, the K-means clustering algorithm is used again for the second clustering, k2 clusters are selected, and the secondary cluster center corresponding to each distance sample point is calculated as the distance data set after the secondary clustering.

[0024] B4. Transform each sample point in the distance dataset after secondary clustering according to the following formula to obtain the clustering feature dataset λ c :

[0025]

[0026] Where: ρ is the incoming air density.

[0027] Further, step C specifically includes:

[0028] C1. Assume that there are M samples of the average wind pressure coefficient distribution data to be analyzed, and each sample contains the average wind pressure coefficient of N measuring points of a large-span spatial structure under a certain wind field. Then the average wind pressure coefficient data set used for cluster analysis is:

[0029] C p ={C pi |C pi =(cp i1 ,cp i2 ,…,cpi N ),i=1,2,…,M} (5)

[0030] C2. Calculation of data set C p The Euclidean distance between each data point in is used to construct the adjacency graph G using the k-nearest neighbor method;

[0031] C3. Use Dijkstra or Floyd-Warshall algorithm to calculate the shortest path distance between any two points in the adjacency graph G and generate the geodesic distance matrix D;

[0032] C4. Use the multidimensional scaling (MDS) algorithm to reduce the geodesic matrix D to a low-dimensional space with dimension d = 2 to obtain the manifold feature data set λ m As shown below, each data point corresponds to 2 low-dimensional manifold features:

[0033] λ m ={(λ m1i ,λ m2i ),(λ m1i ,λ m2i ),…,(λ m1M ,λ m2M )} (6)

[0034] Further, step D specifically includes:

[0035] D1. Use clustering features and manifold feature data as model input features and add them to the total data set according to the corresponding incoming wind field;

[0036] D2. Divide the total data set with clustering features and manifold feature data into a training data set and a test data set. The data with wind field labels is the training data set, and the data without wind field labels is the test data set.

[0037] Further, step E specifically includes:

[0038] E1. A simple average ensemble method is used to construct a semi-supervised ensemble learning model based on the two-stage K-means clustering algorithm and the Isomap manifold learning algorithm feature extraction. The ensemble learning model uses the random forest model and the extreme random tree model for integration. The calculation formula is:

[0039]

[0040] Where: Y(x) is the simple average ensemble prediction value, y i (x) is the predicted value of the ith model, and T is the number of models used in the simple average ensemble method.

[0041] E2. Input the wind pressure distribution characteristic data without corresponding wind field and reconstruct the missing incoming wind field of the large-span spatial structure.

[0042] The advantages of the present invention are:

[0043] (1) The present invention is based on a two-stage K-means clustering algorithm and an Isomap manifold learning algorithm, and uses average wind pressure distribution and average wind pressure coefficient distribution to extract clustering features and manifold features, which can effectively characterize the wind speed and relative wind direction of the incoming wind field, respectively.

[0044] (2) When extracting the characteristics of the incoming wind field, the present invention utilizes all wind pressure data including wind pressure distribution without wind field marking and wind pressure distribution with wind field marking, which can fully explore the potential correlation between the wind pressure distribution corresponding to the wind field to be reconstructed and the wind pressure distribution corresponding to the known wind field, thereby avoiding the wind pressure distribution data of a large number of unknown wind fields from being unable to be utilized.

[0045] (3) The wind field of the large-span spatial structure is reconstructed under different numbers of wind pressure measurement points. The reconstruction accuracy of wind speed and wind direction of the present invention is better than that of the traditional supervised learning method, showing good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the present invention;

[0047] Figure 2 It is a diagram of the measurement point arrangement of the wind tunnel test in the gymnasium of Zhejiang University in the embodiment of the present invention;

[0048] Figure 3 It is the clustering characteristics and wind speed diagram of the present invention;

[0049] Figure 4 It is the flow shape characteristic diagram of different wind fields of the present invention;

[0050] Figure 5 is a comparison diagram of the wind speed reconstruction value and the true value of the present invention;

[0051] Figure 6 is a comparison diagram of the wind direction reconstruction value and the true value of the present invention;

[0052] Figure 7 It is a determination coefficient diagram of wind speed reconstruction results corresponding to different numbers of wind pressure measurement points of the present invention;

[0053] Figure 8 It is a determination coefficient diagram of wind direction reconstruction results corresponding to different numbers of wind pressure measurement points of the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further described in detail below in conjunction with the accompanying drawings and a wind tunnel test example.

[0055] like Figure 1 This embodiment relates to a large-span spatial structure wind field reconstruction method based on semi-supervised ensemble learning, comprising the following steps:

[0056] A. Preprocess the wind pressure and wind field monitoring data of large-span space structures to establish a total data set, including:

[0057] A1. The wind tunnel test data of Zhejiang University Gymnasium were selected, including wind pressure and wind field data collected at 36 wind direction angles (0° to 360° with 10° intervals) and 3 wind speeds (5.4m / s, 9.7m / s, 16.0m / s), totaling 108 wind field conditions; the wind pressure measurement points were arranged as follows: Figure 2 As shown, five groups of working conditions with wind pressure measuring points N=5, N=10, N=20, N=30 and N=40 were set for verification;

[0058] A2. According to the wind pressure time history data p(t i ) Calculate the average wind pressure p mean , pulsating wind pressure root mean square p rms , wind pressure coefficient C p (t i );

[0059] A3. According to the wind pressure coefficient C p (t i ) Calculate the average wind pressure coefficient C p_mean , Pulsating wind pressure coefficient C p_rms , wind pressure deflection S k and wind pressure peak K u ;

[0060] A4. Use wind pressure distribution characteristics and incoming wind field characteristics to establish a total data set; wind pressure distribution characteristics are used as model input, including the average wind pressure p mean , pulsating wind pressure root mean square p rms , average wind pressure coefficient C p_mean , Pulsating wind pressure coefficient C p_rms , wind pressure deflection S k and wind pressure peak K u ; The incoming wind field characteristics are used as model output, including wind speed U and wind direction θ.

[0061] B. A two-stage K-means clustering algorithm is used to extract clustering features that reflect wind speed from the average wind pressure data, including:

[0062] B1. Use the K-means clustering algorithm to perform the first clustering on the input average wind pressure data, select k1 clusters, calculate the Euclidean distance from each wind pressure distribution sample point to its cluster center, and obtain the distance data set D after clustering. 1 ;

[0063] B2. Distance dataset D 1As input, the K-means clustering algorithm is used again for the second clustering, k2 clusters are selected, and the secondary cluster center corresponding to each distance sample point is calculated as the distance data set D after the secondary clustering. 2 :

[0064] B3. Distance dataset D after secondary clustering 2 Each sample point in is transformed according to the following formula to obtain the clustering feature data set λ c ;

[0065] C. The Isomap manifold learning algorithm is used to extract the manifold features reflecting the relative wind direction from the average wind pressure coefficient data, including:

[0066] C1. Calculate the average wind pressure coefficient data set C p The Euclidean distance between each data point in is used to construct the adjacency graph G using the k-nearest neighbor method;

[0067] C2. Use Dijkstra or Floyd-Warshall algorithm to calculate the shortest path distance between any two points in the adjacency graph G and generate the geodesic distance matrix D;

[0068] C3. Use the multidimensional scaling (MDS) algorithm to reduce the geodesic matrix D to a low-dimensional space with dimension d = 2 to obtain the manifold feature data set λ m As shown below, each data point corresponds to two low-dimensional manifold features λ m1 and λ m2 ;

[0069] D. Add clustering features and manifold feature data to the total data set to construct training data sets and test data sets, including:

[0070] D1. Use clustering features and manifold feature data as model input features and add them to the total data set according to the corresponding incoming wind field;

[0071] D2. Divide the total data set with clustering features and manifold feature data into a training data set and a test data set. The data with wind field labels is the training data set, and the data without wind field labels is the test data set.

[0072] E. Train a semi-supervised ensemble learning model based on the two-stage K-means clustering algorithm and the Isomap manifold learning algorithm feature extraction, input the wind pressure distribution characteristics without corresponding wind field, and reconstruct the incoming wind field with missing large-span spatial structure. In the working condition where the number of wind pressure measurement points N = 40, the clustering feature λ c The relationship with wind speed U is as follows Figure 3 As shown, the two manifold characteristics λ of different incoming wind fields m1 and λ m2 Relationship Figure 4As shown, the wind speed reconstruction value U p With the true value U t For example Figure 5 As shown, the wind direction reconstruction value θ p and the true value θ t For example Figure 6 As shown in the figure, for different wind pressure measurement points, the determination coefficient R of the wind speed reconstruction result is 2 like Figure 7 As shown in the figure, the determination coefficient R of the wind direction reconstruction result is 2 like Figure 8 shown.

[0073] The contents described in the implementation cases of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be deemed to be limited to the specific forms described in the implementation cases. The protection scope of the present invention also extends to the equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for wind field reconstruction of a large-span spatial structure based on semi-supervised ensemble learning, comprising the following steps: A. Preprocess the wind pressure and wind field monitoring data of large-span spatial structures to establish a total data set; B. Use a two-stage K-means clustering algorithm to extract clustering features that reflect wind speed from average wind pressure data; C. Using the Isomap manifold learning algorithm to extract manifold features reflecting relative wind direction from the average wind pressure coefficient data; D. Add clustering features and manifold feature data to the total data set to construct training data sets and test data sets; E. Train a semi-supervised ensemble learning model based on the two-stage K-means clustering algorithm and the Isomap manifold learning algorithm feature extraction, input the wind pressure distribution characteristics without corresponding wind field, and reconstruct the incoming wind field with missing large-span spatial structure; specifically include: E1. A simple average ensemble method is used to construct a semi-supervised ensemble learning model based on the two-stage K-means clustering algorithm and the Isomap manifold learning algorithm feature extraction. The ensemble learning model uses the random forest model and the extreme random tree model for integration. The calculation formula is: Where: Y(x) is the simple average ensemble prediction value, y i (x) is the predicted value of the i-th model, and T is the number of models used in the simple average ensemble method; E2. Input the wind pressure distribution characteristic data without corresponding wind field and reconstruct the missing incoming wind field of the large-span spatial structure.

2. A method for wind field reconstruction of a large-span spatial structure based on semi-supervised ensemble learning as claimed in claim 1, characterized in that: Step A specifically includes: A1. Obtain wind pressure and wind field monitoring data for large-span spatial structures; A2. According to the wind pressure time history data p(t i ) Calculate the average wind pressure p mean , pulsating wind pressure root mean square p rms , wind pressure coefficient C p (t i ), the calculation formula is: Where: n is the total number of wind pressure time series data samples, p(t i ) is t i Roof wind pressure value at the moment; p0 and p ∞ are the average total pressure and average static pressure at the reference point respectively; ρ is the air density; v ∞ is the incoming wind speed at the reference point; A3. According to the wind pressure coefficient C p (t i ) Calculate the average wind pressure coefficient C p_mean , Pulsating wind pressure coefficient C p_rms , wind pressure deflection S k and wind pressure peak K u , the calculation formula is: A4. Use wind pressure distribution characteristics and incoming wind field characteristics to establish a total data set; wind pressure distribution characteristics are used as model input, including the average wind pressure p mean , pulsating wind pressure root mean square p rms , average wind pressure coefficient C p_mean , Pulsating wind pressure coefficient C p_rms , wind pressure deflection S k and wind pressure peak K u ; The incoming wind field characteristics are used as model output, including wind speed U and wind direction θ.

3. A method for wind field reconstruction of a large-span spatial structure based on semi-supervised ensemble learning as claimed in claim 1, characterized in that: Step B specifically includes: B1. Assume that there are M samples of average wind pressure distribution data to be analyzed, and each sample contains the average wind pressure of N measuring points of a large-span spatial structure under a certain wind field. Then the average wind pressure data set used for cluster analysis is: P={P i |P i =(p i1 ,p i2 ,…,p iN ),i=1,2,…,M} (3) B2. Use the K-means clustering algorithm to perform the first clustering on the input large-span spatial structure average wind pressure data, select k1 clusters, calculate the Euclidean distance from each wind pressure distribution sample point to its cluster center, and obtain the distance data set after clustering. B3. Distance dataset D 1 As input, the K-means clustering algorithm is used again for the second clustering, k2 clusters are selected, and the secondary cluster center corresponding to each distance sample point is calculated as the distance data set after the secondary clustering. B4. Transform each sample point in the distance dataset after secondary clustering according to the following formula to obtain the clustering feature dataset λ c : Where: ρ is the incoming air density.

4. The method for wind field reconstruction of a large-span spatial structure based on semi-supervised ensemble learning as claimed in claim 1, wherein step C specifically comprises: C1. Assume that there are M samples of the average wind pressure coefficient distribution data to be analyzed, and each sample contains the average wind pressure coefficient of N measuring points of a large-span spatial structure under a certain wind field. Then the average wind pressure coefficient data set used for cluster analysis is: Cp={Cpi|Cpi=(cpi1,cpi2,…,cpi N ),i=1,2,…,M} (5) C2. Calculation of data set C p The Euclidean distance between each data point in is used to construct the adjacency graph G using the k-nearest neighbor method; C3. Use Dijkstra or Floyd-Warshall algorithm to calculate the shortest path distance between any two points in the adjacency graph G and generate the geodesic distance matrix D; C4. Use the multidimensional scaling (MDS) algorithm to reduce the geodesic matrix D to a low-dimensional space with dimension d = 2 to obtain the manifold feature data set λ m As shown below, each data point corresponds to 2 low-dimensional manifold features: l m ={(λ m1i ,l m2i ),(l m1i ,l m2i ),…,(λ m1M ,l m2M )} (6).

5. The method for wind field reconstruction of a large-span spatial structure based on semi-supervised ensemble learning according to claim 1, characterized in that: Step D specifically includes: D1. Use clustering features and manifold feature data as model input features and add them to the total data set according to the corresponding incoming wind field; D2. Divide the total data set with clustering features and manifold feature data into a training data set and a test data set. The data with wind field labels is the training data set, and the data without wind field labels is the test data set.

Images