Mechanism and data driven cable net efficient adjustment method

Abstract

The invention belongs to the technical field of spaceborne antennas, and particularly relates to a mechanism and data driven cable net efficient adjustment method. Comprising the following steps: S1, collecting initial sample points; s2, calculating a sample response; s3, constructing a sample library; s4, the XGBoost proxy model is used for training a deviation model; s5, solving the optimal adjustment amount of the mesh surface of the corresponding prototype; s6, judging the surface precision of the prototype, carrying out an experiment according to the adjustment amount given in the step S5, and calculating the surface precision after adjustment; and S5, taking the adjusted length of the cable section as an initial value, supplementing the sample points in a small range, and repeating the steps S3 to S5 for continuous adjustment until the profile precision is not obviously reduced any more. According to the method, the cable net model with higher engineering consistency is obtained through a mechanism and data driven modeling technology, so that the mapping relation between the longitudinal cable adjustment amount and the net surface precision is established; on the basis, a cable net profile efficient adjustment algorithm is provided, and rapid and high-precision profile adjustment is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of satellite-borne antennas, and in particular relates to a mechanism and data-driven efficient cable net adjustment method. Background Art

[0002] Mesh antennas are small, light, and have a large aspect ratio, and are widely used in the communications field. The surface accuracy of mesh antennas determines the front-to-back ratio, gain, and other electrical properties of the antenna, and is an important performance indicator of the antenna. Due to the influence of cutting and weaving errors, the surface accuracy of the antenna is difficult to meet the design requirements after the antenna is assembled. It is necessary to manually adjust the length of the longitudinal cable to achieve a high-precision mesh antenna surface.

[0003] At present, the surface adjustment of cable nets in engineering projects is mostly based on simple theoretical models, relying on manual experience and continuous improvement of surface accuracy through a large number of adjustments. Due to the large inconsistency between theoretical modeling and the production of engineering prototypes, the adjustment amount given by the theoretical model often cannot achieve targeted and efficient adjustment, making the adjustment process blind, time-consuming and labor-intensive.

[0004] In order to achieve efficient and high-precision cable net adjustment, it is necessary to establish a cable net model with higher engineering consistency. Based on this model, targeted longitudinal cable adjustment amounts are given through optimization methods to guide the cable net surface adjustment. The higher the engineering consistency of the model used to guide the adjustment, the more engineering guidance value the given adjustment amount has. Therefore, the patent of this invention aims to obtain a cable net model with higher engineering consistency through mechanism and data-driven modeling technology, so as to establish a mapping relationship between the longitudinal cable adjustment amount and the net surface accuracy. On this basis, an efficient cable net surface adjustment algorithm is given to achieve fast and high-precision surface adjustment. Summary of the invention

[0005] The purpose of the present invention is to propose a precise modeling method of a deviation model based on mechanism and data-driven based on the deviation between the theoretical model and the engineering prototype, and to establish a proxy model that accurately characterizes the deviation relationship between the theoretical model and the engineering prototype; the above method is applied to the rapid adjustment of the mesh surface of the engineering prototype, the proxy model is used to accurately predict the adjustment amount, and the shape adjustment of the engineering prototype is guided, thereby achieving the purpose of quickly adjusting the mesh antenna to a higher shape accuracy in a small number of times under precise guidance.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A mechanism and data driven efficient cable net adjustment method comprises the following steps:

[0008] S1, initial sample point collection;

[0009] S2, calculate sample response;

[0010] S3, build sample library;

[0011] S4, XGBoost proxy model training bias model;

[0012] S5, solving the optimal adjustment amount of the corresponding prototype mesh;

[0013] S6. Determine the surface accuracy of the prototype. Perform experiments according to the adjustment amount given in S5 and calculate the surface accuracy after adjustment. Use the length of the cable segment after adjustment as the initial value, add sample points in a small range, and repeat S3 to S5 to make continuous adjustments until the surface accuracy no longer decreases significantly.

[0014] The initial sample point collection described in step S1 is: using the Latin hypercube sampling method to collect multiple groups of longitudinal cable changes as initial sample points for training the proxy model.

[0015] The calculation of sample response described in step S2 is: bringing the initial sample points obtained in step S1 into the mechanism model to obtain the corresponding ideal model response; inputting the initial sample points on the physical prototype and collecting the response of the measured prototype, and calculating the response deviation.

[0016] The specific steps include:

[0017] T1. According to the selected annular truss antenna cable network, a mechanism model based on the force density method is established, and the layout length of each cable segment is obtained through form-finding design to provide design data for subsequent prototype production.

[0018] In order to amplify the influence of the longitudinal cable change on the position of the mesh node, the longitudinal cable is changed to a cable-spring series form and calculated separately during the form-finding process. The force density model of the cable net structure is:

[0019] Find Q = [q 1 ,q 2 ,…,q n ] T

[0020]

[0021] Sta≤q≤b

[0022] Among them, q is the force density, that is, the ratio of the tension of the cable segment to its length; Q is the diagonal matrix of the force density q.

[0023] The coordinates of the two types of free nodes are P f1 (X f1 ,Y f1 ,Z f1 ),P f2 (X f2 ,Y f2 ,Z f2), represent the coordinates of two types of free nodes, those connected to the spring and those not connected to the spring; P g is the coordinate of the cable segment node fixed on the truss. f ,C g are the topological matrices corresponding to free nodes and boundary nodes respectively; RMS is the root mean square error of the surface; X' f Y' f Z' f is the ideal node coordinate of the cable net; T is the force at both ends of the spring, a and b are the upper and lower limits of the force density value. The unstretched length of the cable segment is calculated by Hooke's law, the change in the longitudinal cable length is obtained by S1, and the unstretched length of the prototype cable segment after the length is changed is obtained.

[0024] T2. Based on the initial sample points, a mechanism model is established to reflect the mapping relationship between the longitudinal cable length and the grid node coordinates. The free node obtained by the mechanism model is denoted as pf 1 .

[0025] pf 1 =[P f1 ;P f2 ;P g ],C=[C f1 ; C f2 ; C g ];

[0026]

[0027] Q=F / L,T=F(121:139,:);

[0028]

[0029] After changing the longitudinal cable adjustment amount, the free node coordinates of the mesh surface are the only variables to be determined. The variables on both sides of the force density equation can be expressed by the free node coordinates. Solving the equation can obtain the free node coordinates pf of each group after the longitudinal cable changes. 1 The number of longitudinal cable segments in this example antenna is 19, and the number of free nodes on the front mesh is 19. Each group of initial sample points can solve the coordinates of 19 free nodes. Each response vector pf 1 The dimension of is 57.

[0030] T3. Make a cable net prototype according to the length of each cable segment obtained by design. According to the longitudinal cable variation of the sample point corresponding to S1, the actual prototype free node coordinates pf are collected through photogrammetry. 2 ;Calculate the output of the physical model pf 1 Output pf with physical prototype 2 The deviation between 2 -pf 1 .

[0031] The construction of the sample library in step S3 is as follows: the input is the initial sample point obtained by S1; the response is the initial sample point obtained by S2, and the response deviation pf is obtained by inputting the mechanism model and the physical prototype respectively. 2 -pf 1 , providing data source for subsequent model training.

[0032] The XGBoost proxy model training deviation model described in step S4 is: the present invention adopts the XGBoost model to construct the deviation model, the input is the longitudinal cable change, and the response is each dimension data of the free node coordinates of the front mesh surface.

[0033] The XGBoost model is a type of regression tree prediction model based on boosting. During training, each iteration generates a regression tree to fit the residual between the previous prediction results and the actual points of the training samples.

[0034]

[0035]

[0036]

[0037]

[0038] The specific steps include:

[0039] U1. Use the XGBoost model to train the sample library and perform hyperparameter tuning on the XGBoost parameters to avoid overfitting.

[0040] Taking the mesh antenna in this example, the input is 19-dimensional longitudinal cable variation, and the output is 57-dimensional free node coordinates; a prediction model for all longitudinal cable variation for each dimension of coordinates is established, totaling 57 models. The final prediction value of the XGBoost model for the i-th sample is The decision tree is f, and the objective function of the model is the sum of the loss function and the model complexity; the objective function obj is parameterized, T represents the number of leaf nodes, ω j Represents the weight of the leaf node, λ and γ are both hyperparameters; the above variables are used as optimization variables to find the optimal decision tree structure.

[0041]

[0042] A greedy algorithm is used to split the regression tree nodes step by step. The larger the Gain value, the more the objective function is reduced after the split, and the better the effect. The process of obtaining a set of Gain values ​​and hyperparameter values ​​corresponding to the optimal objective function is called the tuning process.

[0043] The NRSR algorithm is used to optimize the values ​​of hyperparameters.

[0044]

[0045] r 1 ∈rand(0,1)

[0046] y w =r 1 ×(Mean(Z n+1 +x n )+r 1 ×Δx)

[0047] y b =r 1 ×(Mean(Z n+1 +x n )-r 1 ×Δx)

[0048]

[0049] X w ,X b ∈[x-Δx,x+Δx];λ,γ∈x n

[0050] The NRSR algorithm starts the search for the optimal solution by generating an initial random population within the boundary of the candidate solution, finds the optimal solution of the space within the given hyperparameter range, and obtains the global solution through convergence. The adaptive coefficient of each iteration is represented by δ:

[0051]

[0052] The trap avoidance operator TAO is incorporated to prevent the NRSR algorithm from falling into the local optimum.

[0053]

[0054] U2. Draw the model accuracy of each model, calculate the model accuracy curve of the prediction model, and verify that the accuracy variation range of the deviation model is smaller than that of the pure data model. The prediction effect of the deviation model is more stable.

[0055] The longitudinal index change x of the sample library obtained in S3 i As input, y i As the output, the S4 deviation proxy model is trained. In order to determine the number of samples of the proxy model to meet the model accuracy, after sufficient sampling, the model accuracy is calculated for different numbers of sample points. The model accuracy is expressed in R 2 Make an evaluation. 2 The expression is:

[0056]

[0057] Where N is the number of models, y i is the measured response of the sample, To x i Input S4 to get the predicted response of the proxy model. The model input is 19 dimensions and the output is 57 dimensions. We build proxy models for the mapping relationship between input and output in each dimension, for a total of 57 models. All models need to be verified when verifying the model accuracy.

[0058] Latin hypercube sampling is used to collect the number of samples that make up the model. The number of samples should not be less than (dim+1)(dim+2) / 2, where dim is the input dimension of the sample library (19 dimensions in this example). Therefore, the number of samples in this example should not be less than 210, so a total of 500 samples are drawn as the initial sample library.

[0059] When verifying the proxy model, the present invention adopts the leave-one-out cross-validation method to evaluate the accuracy of the proxy model. The basic idea is: for the proxy model constructed by N sample points, one sample point is reserved as a test point each time, and the proxy model is constructed with the remaining (N-1) sample points. After N times of construction and testing, N prediction errors are obtained to evaluate the model accuracy. In cross-validation, all sample points participate in the construction of the proxy model and also participate in the calculation of model accuracy as test points, making full use of sample point information. There is no need to collect additional test sample points, reducing the call to the original model and improving the overall modeling efficiency. When the model accuracy meets the requirements, collecting and training with a smaller number of sample points can better reduce the time required to build the model.

[0060] The step S5 of solving the optimal adjustment amount of the corresponding prototype mesh is to use the proxy model to solve the corresponding longitudinal cable adjustment amount when the optimal accuracy of the prototype mesh is solved, which belongs to a type of optimization problem.

[0061]

[0062] The optimization variable is the longitudinal cable adjustment ΔL, and the constraint condition depends on the longitudinal cable adjustable range (ΔL min ,ΔL max ), the objective function is the mesh accuracy, using RMS evaluation, where pf 0 is the ideal node coordinate on the surface, p pre is the predicted value of the mesh surface coordinates. The longitudinal cable adjustment amount corresponding to the optimal value of the mesh surface accuracy obtained by the optimization is used as the adjustment amount to guide the assembly of the prototype.

[0063] Step S6 uses the longitudinal cable adjustment amount obtained in step S5 to adjust the prototype, and calculates the surface accuracy of the front mesh of the prototype after adjustment. Determine whether the RMS surface accuracy can be further reduced: take the cable length after each adjustment as the initial value, sample data in a small range, repeat steps S3 to S5, train the prediction model, and make multiple adjustments until the surface accuracy of the front mesh after adjustment is no longer significantly reduced.

[0064] The beneficial effects of the present invention are:

[0065] 1. The present invention establishes a precise model of the cable net driven by mechanism and data, improves the engineering consistency of the theoretical model, accurately reveals the mapping relationship between the longitudinal cable length and the coordinates of the front mesh node, and provides an accurate calculation model for subsequent mesh adjustment.

[0066] 2. Based on the mechanism and data-driven precise cable net model, an efficient cable net adjustment optimization model was established with the longitudinal cable adjustment amount as the optimization variable and the net surface accuracy as the optimization target. Through optimization design, a rapid adjustment strategy for the net surface accuracy suitable for the current cable net node position was given, achieving the purpose of rapid adjustment to higher surface accuracy in a small number of times. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The accompanying drawings are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0068] Figure 1 It is a schematic diagram of the process of the present invention;

[0069] Figure 2 This is a schematic diagram of the cable net antenna structure;

[0070] Figure 3 This is a schematic diagram of the surface structure and node numbering on the cable net;

[0071] Figure 4 This is a schematic diagram comparing the evaluation results of two types of surrogate models;

[0072] Figure 5 It is a curve chart showing the change of mesh surface accuracy during the adjustment process;

[0073] Figure 6 This is the root mean square error diagram of each node on the network before adjustment;

[0074] Figure 7 This is the root mean square error diagram of each node on the network after adjustment;

[0075] Figure 8 To adjust the amount of adjustment of the longitudinal cables corresponding to each node on the front and rear net surfaces;

[0076] Among them: 1-front cable net; 2-longitudinal cable; 3-spring; 4-rear cable net, 5-annular truss. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and specific examples, but this does not limit the present invention.

[0078] like Figure 1 As shown, a method for checking the engineering consistency and rapid adjustment of a mesh antenna cable network model is provided. A proxy model of an engineering prototype is established, and the model is used to predict the corresponding longitudinal cable adjustment amount under the optimal mesh surface accuracy to achieve efficient adjustment. The cable network antenna structure of this embodiment is as follows Figure 2 As shown, it consists of 139 cable segments, and its design structural parameters are shown in Table 1.

[0079] Table 1 Cable net structure parameters

[0080]

[0081] The specific implementation of this example is as follows:

[0082] Step S1, using Latin hypercube sampling to collect multiple groups of longitudinal cable changes as initial sample points.

[0083] The adjustable cable segment of the ring mesh antenna is the longitudinal cable connecting the upper and lower meshes. The position of the free node of the front mesh is changed by adjusting the length of the longitudinal cable. The adjustment amount of the longitudinal cable is obtained by the Latin hypercube sampling method. The range of the adjustment amount is determined to be (-26.5mm, 26.5mm). 500 sets of changes are randomly obtained by applying random errors. The front mesh node is shown in the figure below. Figure 3 As shown, each node is connected to one longitudinal cable, the number of longitudinal cable segments is 19, and the node numbers correspond to the longitudinal cable numbers one by one.

[0084] Step S2, calculating the sample response.

[0085] The change of each group of longitudinal cables obtained in S1 is used as the sample initial point, and the free node coordinates pf after the longitudinal cable adjustment are solved through the relevant theory of the force density model. 1 ,

[0086] Find Q = [q 1 ,q 2 ,…,q n ] T

[0087]

[0088]

[0089]

[0090] Sta≤q≤b

[0091] Among them, q is the force density, that is, the ratio of the tension of the cable segment to its length; Q is the diagonal matrix of the force density q.

[0092] The coordinates of the two types of free nodes are P f1 (X f1 ,Y f1 ,Z f1 ),P f2 (X f2 ,Y f2 ,Z f2 ), represents the coordinates of free nodes on the two types of cable nets, those connected to springs and those not connected to springs; P g is the coordinate of the cable segment node fixed on the truss. f ,C g are the topological matrices corresponding to free nodes and boundary nodes respectively; RMS is the root mean square error of the surface; X' f Y' f Z' f are the coordinates of the ideal nodes of the cable net; T is the force at both ends of the spring, and a and b are the upper and lower limits of the force density value respectively.

[0093] T=k@Δl

[0094]

[0095]

[0096] The unstretched length of each cable segment is calculated by back-calculating the Hooke's law, and the longitudinal cable change is obtained from S1, and the original longitudinal cable length of the current prototype is obtained. Where E is the elastic modulus, A is the cross-sectional area of ​​the cable segment, L is the stretched length of the cable segment, and L 0 is the unstretched length of the cable segment. After applying the adjustment amount to the original length of the longitudinal cable and changing the original length of the longitudinal cable, the adjusted free node coordinates are obtained.

[0097] pf 1 =[P f1 ;P f2 ;P g ],C=[C f1 ; C f2 ; C g ];

[0098]

[0099] Q=F / L,T=F(121:139,:);

[0100]

[0101] After the adjustment is applied, the free node coordinates are the only variables to be determined. The variables on both sides of the above force density equation can be expressed by the free node coordinates. In this example, the number of longitudinal cable segments of the antenna is 19, and the number of free nodes on the front mesh surface is 19. Each group of initial sample points can solve the coordinates of 19 free nodes. Each response vector pf 1 The dimension of is 57.

[0102] The original prototype is built under the guidance of the mechanism model, and the coordinates of each set of free nodes after adjustment are recorded, which is denoted as pf 2 . The output of the computational model is pf 1 Output pf with physical prototype 2 The deviation between 2 -pf 1 .

[0103] Step S3, building a sample library.

[0104] According to the initial sample points obtained by S1 and S2, the response of the mechanism model and the actual prototype response, the input is the change of the longitudinal cable, and the output is the difference between the two groups of responses pf 2 -pf 1 The sample library provides data for subsequent training.

[0105] Step S4, the XGBoost proxy model trains the deviation model to construct a proxy model that reflects the mapping relationship between the longitudinal cable change and the front mesh free node coordinate difference.

[0106] The present invention uses the XGBoost model to construct a bias proxy model, and constructs a proxy model with each dimensional coordinate of the 19 front mesh free nodes. The XGBoost model is a type of regression tree prediction model based on boosting. During training, each iteration generates a regression tree to fit the residual between the previous prediction result and the actual point of the training sample.

[0107]

[0108]

[0109]

[0110]

[0111] The final prediction value of the XGBoost model for the i-th sample is The decision tree is f, and the objective function of the model is the sum of the loss function and the model complexity; the objective function obj is parameterized, T represents the number of leaf nodes, ω jRepresents the weight of the leaf node, λ and γ are both hyperparameters; the above variables are used as optimization variables to find the optimal decision tree structure.

[0112]

[0113] The greedy algorithm is used to split the regression tree nodes step by step. The larger the Gain value, the more the objective function is reduced after the split, and the better the effect. The process of obtaining a set of Gain values ​​and hyperparameter values ​​corresponding to the optimal objective function is called the tuning process. The NRSR algorithm is used to calculate the value of the hyperparameter.

[0114]

[0115] r 1 ∈rand(0,1)

[0116] y w =r 1 ×(Mean(Z n+1 +x n )+r 1 ×Δx)

[0117] y b =r 1 ×(Mean(Z n+1 +x n )-r 1 ×Δx)

[0118]

[0119] X w ,X b ∈[x-Δx,x+Δx];λ,γ∈x n

[0120] The NRSR algorithm starts the search for the optimal solution by generating an initial random population within the boundary of the candidate solution, finds the optimal solution of the space within the given hyperparameter range, and obtains the global solution through convergence. The adaptive coefficient of each iteration is represented by δ:

[0121]

[0122] The trap avoidance operator TAO is incorporated to prevent the NRSR algorithm from falling into the local optimum.

[0123]

[0124] The model accuracy of the XGBoost model is available in R 2 The higher the model accuracy, the closer the predicted value is to the actual prototype value. 2 To indicate. 2The closer it is to 1, the better the accuracy.

[0125] R 2 The expression is:

[0126]

[0127] Where N is the number of models, y i is the measured response of the sample, For the sample, x i Input S4 to get the predicted response after the proxy model. The model input is 19 dimensions and the output is 57 dimensions. The proxy models of all input-to-output mapping relationships are established separately. There are 57 models in total. All models need to be verified when verifying the model accuracy.

[0128] Latin hypercube sampling is used to collect the number of samples that compose the model. The number of samples should not be less than (dim+1)(dim+2) / 2, where dim is the input dimension of the sample library, which is 19 dimensions. Therefore, the number of samples should not be less than 210. In this example, 500 samples are extracted as the initial sample library.

[0129] The present invention adopts the leave-one-out cross-validation method to evaluate the accuracy of the proxy model. The basic idea is: for the proxy model constructed by N sample points, one sample point is reserved as a test point each time, and the remaining (N-1) sample points are used to construct the proxy model. After N times of construction and testing, N prediction errors are obtained to evaluate the model accuracy. Cross-validation makes full use of sample point information and does not require additional collection of test sample points, which reduces the call to the original model and improves the overall modeling efficiency. The model accuracy of each prediction model composed of the data model and the mechanism-data deviation model is calculated separately under the same set of inputs. The range of model accuracy is as follows: Figure 4 As shown, the accuracy of the data model ranges from 0.76 to 0.95, and the accuracy of the deviation model ranges from 0.88 to 0.99, which proves that under the same set of inputs, the model accuracy of the deviation model is more stable than that of the data model.

[0130] Step S5, solving the optimal adjustment amount of the prototype.

[0131] Solving the optimal adjustment means predicting and searching for a set of adjustment values ​​within the adjustable range of the longitudinal cable, so that the mesh surface accuracy of the prototype after adjustment is the best. The mesh surface accuracy is characterized by RMS.

[0132]

[0133] (X f ,Y f ,Z f ) is the free node coordinate of the front mesh surface, (X' f ,Y' f ,Z'f ) is the free node seat of the front mesh predicted by the model. The whole optimization process can be expressed as:

[0134]

[0135] Step S6, judging the surface accuracy of the prototype.

[0136] Apply the adjustment predicted by S5 to the prototype, obtain the coordinates of the free nodes of the front mesh surface at this time by photogrammetry, and calculate the RMS value of the mesh accuracy at this time. To ensure that the mesh accuracy predicted by the model is optimal, after each prediction and adjustment, use the adjusted cable segment length as the initial value, and add 210 groups in a small range as a new sample library. Repeat steps S3 to S5, and collect the mesh accuracy after each adjustment until it no longer decreases significantly.

[0137] This case conducted an experiment of accurate modeling and rapid adjustment of 18 units with a scale of 6 units, such as Figure 5 As shown in the figure, after three adjustments, the mesh surface accuracy dropped from 1.9mm to 0.08mm. Figure 6 This is the initial state of the mesh before the prototype is adjusted, and the mesh accuracy is 1.9mm; Figure 7 This is the final state of the mesh after three adjustments, with a mesh accuracy of 0.08mm; Figure 8 The adjustment amount of each longitudinal cable segment from the initial state to the final state of the mesh surface is shown in Table 2. The prediction accuracy of the final state mesh surface and the measured accuracy of the adjusted prototype are shown in Table 2.

[0138] Table 2 Comparison of prediction accuracy after mesh adjustment and actual measurement accuracy of prototype

[0139]

[0140] The present invention is simple in theory and easy to implement. The prediction results of the present invention are highly consistent with the actual prototype engineering, providing effective guidance for quickly calculating the optimal adjustment amount of the prototype. The above description shows and describes the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other examples. It can be applied to various other antenna structures and can be modified within the scope of the inventive concept described herein through the above teachings or related technologies or knowledge. Changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention, and should all be within the scope of protection of the claims attached to the present invention.

Claims

1. A mechanism and data driven efficient cable net adjustment method, characterized in that: The steps include: S1, initial sample point collection; S2, calculate sample response; S3, build sample library; S4, XGBoost proxy model training bias model; S5, solving the optimal adjustment amount of the corresponding prototype mesh; S6, judging the surface accuracy of the prototype, conducting experiments according to the adjustment amount given in S5 and calculating the surface accuracy after adjustment; Taking the adjusted cable segment length as the initial value, add sample points in a small range and repeat S3 to S5 to make continuous adjustments until the surface accuracy no longer decreases significantly.

2. The mechanism and data driven cable net efficient adjustment method according to claim 1, characterized in that: In step S1, a Latin hypercube sampling method is used to collect multiple groups of longitudinal cable changes as initial sample points for training the proxy model.

3. The method according to claims 1 and 2, characterized in that In step S2, the following steps are included: T1. According to the selected annular truss antenna cable network, a mechanism model based on the force density method is established, and the layout length of each cable segment is obtained through form-finding design to provide design data for subsequent prototype production. T2. Based on the initial sample points, a mechanism model is established to reflect the mapping relationship between the longitudinal cable length and the mesh node coordinates. The free node obtained by the mechanism model is recorded as pf1. T3. Make a cable net prototype according to the designed length of each cable segment, and collect the actual prototype free node coordinates pf2 through photogrammetry according to the longitudinal cable variation of the sample point corresponding to S1. The deviation between the output pf1 of the computer model and the output pf2 of the physical prototype is pf2-pf1.

4. The method according to claims 2 and 3, characterized in that: In step S3, the initial sample points and response deviations are combined to establish a sample library to provide a data source for subsequent model training.

5. The mechanism and data driven cable net efficient adjustment method according to claim 1, characterized in that: In step S4, the following steps are included: U1. Use the XGBoost model to train the sample library and perform hyperparameter tuning on the XGBoost parameters to avoid overfitting. U2. Draw the model accuracy of each model, calculate the model accuracy curve of the prediction model, and verify that the accuracy variation range of the deviation model is smaller than that of the pure data model. The prediction effect of the deviation model is more stable.

6. The mechanism and data driven cable net efficient adjustment method according to claim 1, characterized in that: In step S5, the optimal mesh adjustment amount of the prototype is solved, and the proxy model trained in S4 is used to solve the corresponding longitudinal cable adjustment amount when the mesh accuracy is optimal.

7. The mechanism and data driven cable net efficient adjustment method according to claim 1, characterized in that: In step S6, the prototype is adjusted using the longitudinal cable adjustment amount obtained in step S5, and the mesh surface accuracy of the front mesh surface of the adjusted prototype is calculated to determine whether the mesh surface accuracy RMS can be further reduced. Taking the cable segment length after each adjustment as the initial value, sample data in a small range, repeat steps S3 to S5, repeatedly train the prediction model, and make multiple adjustments until the mesh accuracy after the previous mesh adjustment no longer decreases significantly.

Images