Neural network active adjustment method of active surface antenna

Abstract

The invention relates to a neural network active adjustment method for an active surface antenna, and the method aims at a large-scale active surface antenna application data preprocessing method and a convolutional neural network model, and carries out the targeted improvement of the preprocessing method and the convolutional neural network model according to the difference between the preprocessing method and the convolutional neural network model. According to the method, the output quantity of a neural network model is changed into single output, different actuators are respectively trained and predicted, and an emphasis layer is added, so that the model has the capability of self-adapting to node coordinate weights. And finally, analyzing the improved training result, proving the effectiveness of the improvement, and analyzing the effectiveness of the data preprocessing method and the effect of an emphasizing layer in the model according to the result. The method has high practical application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of deep learning, and particularly relates to a neural network active adjustment calculation method based on an active surface antenna, which is used for actively adjusting the position accuracy of the active surface antenna to make the antenna performance optimal. BACKGROUND

[0002] Large reflector antennas are important observation devices in the fields of deep space exploration and radio astronomy. With the continuous improvement of high-frequency observation requirements, large reflector antennas are developing towards large aperture, high frequency and high precision. The surface accuracy of a large reflector antenna directly affects the electrical performance of the antenna. However, the reflector surface of the antenna is usually composed of a large number of panels. Under the action of external loads such as gravity, temperature and wind, the reflector surface inevitably deforms, which directly leads to problems such as gain reduction and pointing deviation of the antenna. Therefore, ensuring the surface accuracy of the antenna reflector is a key link to ensure the performance of the antenna.

[0003] At present, active surface technology is the most effective and direct method to achieve high-frequency reflector antenna surface accuracy of large-scale real panels. Active surface technology mainly adjusts the antenna surface accuracy through actuators supporting the panels, which adjust the pose of the panels by elongation or shortening. Calculating the adjustment amount of the actuator is one of the key contents of the active surface technology to achieve high-precision requirements of the antenna reflector.

[0004] Currently, many scholars have studied the calculation method of the adjustment amount of the actuator. The main idea is to take the normal distance between the nodes of the ideal reflector and the actual reflector as the adjustment amount. Some scholars also use microwave holographic technology to measure the antenna surface error to establish a calculation method for the adjustment amount of the actuator. These methods are simple and effective, and the adjustment amount obtained improves the accuracy of the antenna surface to a certain extent. However, these adjustment methods ignore the influence of the elastic deformation of the panel itself when the actuator adjusts the pose of the panel. Therefore, some scholars have approximately established a linear relationship matrix between the elastic deformation of the panel and the adjustment amount of the actuator, and thus calculated the adjustment amount of the antenna actuator by optimization method. Some scholars have also directly simplified the actuator mechanism and established a structure model of the active reflector antenna containing the simplified actuator, further improving the method of active surface analysis and adjustment amount calculation. These methods calculate the adjustment amount of the actuator around the structure error, and do not consider the influence of the adjustment amount change caused by different complex environments and aging of the control equipment in the actual engineering operation process.

[0005] With the rapid development of computer hardware in recent years, artificial intelligence, especially deep learning technology, has made great progress in application fields. Its intelligent and autonomous learning method through data also opens up new ideas for the calculation method of large reflector antenna active adjustment amount. The present application is aimed at the urgent need for intelligent learning to improve the adjustment amount calculation method of large reflector antenna active surface adjustment technology. The present application innovatively proposes an active surface adjustment amount calculation method based on neural network, which trains and fine-tunes the active adjustment neural network model according to the use data of the active surface technology in the engineering application process, so that the active adjustment amount calculation method has the intelligent attribute of self-learning. This process is a neural network adjustment method for active surface antenna. SUMMARY

[0006] In view of the shortcomings of the previous adjustment method, the present application provides a neural network adjustment method for active surface antenna, which mainly solves the calculation of actuator adjustment amount around the structural error, and the influence of the change of the control amount caused by different complex environments and aging of the control equipment in the engineering actual operation process.

[0007] To achieve the above purpose, the neural network adjustment method provided by the present application comprises the following steps:

[0008] Step (1) generating data set. The 48 actuators are numbered and the actuators are divided into rings. The actuator adjustment amount (including the pre-adjustment coordinate and the post-adjustment coordinate) and the corresponding panel node coordinate are a data set;

[0009] Step (2) using pre-processing strategy for the generated data. The label is subjected to numerical normalization processing, and the adjustment amount is divided by the maximum value. The processing of the features is divided into two parts. The first part is to determine the sorting order, the grouping number and the number of nodes contained in each group after grouping. The second part is to use the sorting order and the grouping result of the first part to process the data set in turn;

[0010] Step (3) using single actuator prediction, changing the multiple output of the model to single output, using the same model for training, training 48 actuators in turn, a total of 48 times, and finally obtaining 48 groups of neural network parameters;

[0011] Step (4) designing a self-learning "emphasis layer", which enables the neural network to autonomously select the importance of each panel node to the connected / unconnected actuator;

[0012] Step (5) based on the single panel active adjustment neural network model, adding an emphasis layer at the bottom layer (close to the input end), and changing the output number to 1 to train a single actuator.

[0013] The process of generating data set in step (1) is as follows:

[0014] (1a) Generating data sets, giving the model a vertical downward force to simulate the working condition of looking up, randomly adjusting the 48 actuators on the model within the adjustment range of [0, 10] mm, and obtaining the adjusted node displacement of the antenna panel under the action of gravity. The actuator adjustment amount and the corresponding panel node coordinates are a set of data sets;

[0015] (1b) A total of 20480 sets of data sets are generated by the above method, and each set of data is characterized by a 3564x6 array. The first three columns are the ideal panel node coordinates before adjustment, and the last three columns are the adjusted node coordinates. The corresponding label is a one-dimensional array with a length of 48, corresponding to the adjustment amount of the 48 actuators;

[0016] The data preprocessing process in step (2) is as follows:

[0017] (2a) Feature source data change. Each set of changed data contains ideal panel node coordinates before adjustment and adjusted panel node coordinates. At the same time, the data when batch processing is changed to the difference between the ideal panel node coordinates before adjustment and the adjusted panel node coordinates;

[0018] (2b) The ideal panel node coordinates are used to determine the sorting order, the number of groups, and the number of nodes included in each group after grouping;

[0019] (2c) Data normalization operation change. Use the difference between the adjusted panel node coordinates before adjustment and the adjusted panel node coordinates, and directly divide the whole by the maximum adjustment amount;

[0020] (2d) Grouping operation change. Since the reflector antenna is circular, the number of elements in the middle group is greater than or equal to the number of elements in the outer group, so that the middle group has more valid values, and the end group has fewer valid values, and gradually decreases.

[0021] The process of designing a self-learning "emphasis layer" in step (4) is as follows:

[0022] (4a) The emphasis layer uses a single weight to scale the three coordinate values of the node xyz equally. This module is placed at the bottom of the model structure (close to the input end) to redistribute the weights of different panel nodes of the input;

[0023] (4b) The emphasis layer has the same size as the input data, and the calculation formula of this layer is:

[0024] O = I ʘ W

[0025] Where I is the input of the layer, W is the parameter matrix of the layer, and O is the output of the layer, that is

[0026] I = [ X Y Z ] , X = [ x 11 x 12 ⋯ x 1 n x 21 ⋱ ⋮ ⋱ ⋮ x n 1 x nn ] Y = [ y 11 y 12 ⋯ y 1 n y 21 ⋱ ⋮ ⋱ y n 1 y nn ] Z = [ z 11 z 12 ⋯ z 1 n y 21 ⋱ ⋮ ⋱ z n 1 z nn ]

[0027] W = [ w ] = [ [ w 11 w 12 ⋯ w 1 n w 21 ⋱ ⋮ ⋱ w n 1 w nn ] ]

[0028] O = [ X ⊙ w Y ⊙ w Z ⊙ w ] = [ [ x 11 ⋅ w 11 x 12 ⋅ w 12 ⋯ x 1 n ⋅ w 1 n x 21 ⋅ w 21 ⋱ ⋮ ⋱ ⋮ x n 1 ⋅ w n 1 x nn ⋅ w nn ] [ y 11 ⋅ w 11 y 12 ⋅ w 12 ⋯ y 1 n ⋅ w 1 n y 21 ⋅ w 21 ⋱ ⋮ ⋱ ⋮ y n 1 ⋅ w n 1 y nn ⋅ w nn ] [ z 11 ⋅ w 11 z 12 ⋅ w 12 ⋯ z 1 n ⋅ w 1 n z 21 ⋅ w 21 ⋱ ⋮ ⋱ ⋮ z n 1 ⋅ w n 1 z nn ⋅ w nn ] ]

[0029] (4c) By looking at the parameter W of the current emphasis layer, the weight value corresponding to the node coordinate can be viewed to verify other guesses and conclusions.

[0030] The design of the self-learning neural network model in step (5) is as follows:

[0031] (5a) The first logical layer of the model uses four different size convolution kernels, wherein the 1x1 convolution kernel and the 3x3 convolution kernel do not change the feature map size, and the outputs of the two are connected, the 5x5 convolution kernel and the 7x7 convolution kernel reduce the feature map by half, and the output is jump connected to the output of the third layer;

[0032] (5b) The second layer uses a 1x1 convolution kernel and a 3x3 convolution kernel, and the outputs are connected, and the feature map size is not changed;

[0033] (5c) The third layer uses a 3x3 convolution kernel, and the feature map size is halved, and the output of the halved convolution kernel of the first layer is jump connected;

[0034] (5d) The fourth layer uses a 1x1 convolution kernel and a 3x3 convolution kernel, and the outputs are connected, and the feature map size is not changed;

[0035] (5e) The fifth layer only uses a 3x3 convolution kernel, but the feature map is halved, and the number of channels is increased to compress information, so that global average pooling is used for information extraction at the end. After global average pooling, a multilayer perceptron is used to further extract information and output the adjustment amount of the four actuators.

[0036] The present application uses a neural network to actively adjust the active surface.

[0037] The present application has the following characteristics:

[0038] (1) In the aspect of data preprocessing, the node coordinates of the input data are changed to node coordinate changes, and the data sorting and reorganization algorithm is improved;

[0039] (2) In the aspect of the convolutional neural network model, the number of model outputs is changed to a single output, different actuators are trained and predicted respectively, and an emphasis layer is added at the bottom of the model to enable the model to have the ability to adapt to the node coordinate weight.

[0040] The active adjustment result shows the effectiveness of the method, and provides an intelligent solution for considering the influence of environment and equipment aging in engineering practice to calculate the adjustment amount. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a convolutional neural network model of antenna active adjustment;

[0042] Figure 2 is an antenna data preprocessing flowchart;

[0043] Figure 3 Comparison of the influence of whether containing an emphasis layer on the prediction error of each ring actuator;

[0044] Figure 4 is a comparison of the average error of each ring actuator with or without an emphasis layer;

[0045] Figure 5 is a finite element model of a reflector antenna and a ring diagram of its actuators. DETAILED DESCRIPTION

[0046] The application will be further described below in combination with the drawings and implementation cases.

[0047] Reference Figure 1 and Figure 2 , a neural network active adjustment method of an active surface antenna, the specific steps are as follows:

[0048] (1) a convolutional neural network model as shown in Figure 1 is established, and the output quantity is 1, that is, only one actuator is trained each time. Similarly, the model parameters are initialized, the loss function is set as L1 loss, the ADAM optimization algorithm is used, and the dynamic learning rate method without using the restart strategy is used for the hyperparameters learning rate and weight decay;

[0049] (2) the model training adopts the small batch gradient descent method, and the batch size is set to 128. The newly added emphasis layer is initialized separately;

[0050] (3) using the preprocessed data set, 150 rounds of training are performed for hyperparameter tuning, and finally the learning rate and weight decay are determined;

[0051] (4) since the model is a symmetrical structure, one actuator is selected for each ring during performance testing. The training conditions with and without the emphasis layer are compared, and the results are shown in Figure 3 , Figure 4 ; and

[0052] (5) the emphasis layer weight is extracted to verify the role of the emphasis layer in the model. In order to make it more representative, four actuators located in different rings are selected, and the directions of the four actuators are all different;

[0053] (6) Train all actuators in turn, and save the trained model parameters of each actuator for use in testing;

[0054] (7) Perform panel node accuracy testing. Generate 128 groups of data and pre-process them as a test set for final accuracy evaluation. In testing, first establish a convolutional neural network model, then load the trained model parameters corresponding to different actuators to predict the adjustment amounts of different actuators. Finally, the average accuracy of the adjustment amounts of the actuators in the test set is obtained. Input the actuator adjustment amounts predicted by the neural network into the antenna model to obtain the corresponding panel node coordinates, and calculate the root mean square error between the panel node coordinates of the test set and the panel node coordinates, to obtain the final surface accuracy.

Claims

1. A neural network active adjustment method of an active surface antenna, characterized in that Comprising the following steps: Step (1): generating a data set, giving the model a downward force to simulate the overhead working condition, within the adjustment range of [0, 10] mm, randomly adjusting 48 actuators on the model, obtaining the adjusted node displacement of the antenna panel under the action of gravity. The actuator adjustment amount and its corresponding panel node coordinates are a set of data sets; Step (2): data preprocessing, determining the order and grouping according to the un-deformed panel node coordinates, and sequentially processing all data sets according to the obtained order and grouping; Step (3): constructing a neural network model structure, the neural network has 5 logical layers, each layer outputs the same size convolution in the depth direction, and each layer after convolution increases the layer normalization layer, and uses the GELU (Gaussian Error Linear Unit) activation function; Step (4): single actuator prediction, change the model's multiple outputs to single output, use the same model for training, train 48 actuators in turn, a total of 48 times, finally obtain 48 groups of neural network parameters; Step (5): design a self-learning "emphasis layer" to enable the neural network to autonomously select the importance of each panel node to the connected / unconnected actuator. At the same time, the emphasis layer uses a single weight to scale the node xyz three coordinate values to the same size. This module is placed in the first layer after the antenna panel data input, and is used to redistribute the weights of different panel nodes. The emphasis layer is the same size as the input data. The calculation formula of this layer is: O = I ʘ W Where I is the input of this layer, W is the parameter matrix of this layer, and O is the output of this layer.

2. The neural network active adjustment method of an active surface antenna according to claim 1, characterized in that, In step (2), the data preprocessing is realized by the following method: The label is normalized by dividing the maximum adjustment amount. The features are processed in two parts. The first part determines the order, the number of groups, and the number of nodes included in each group after grouping. The second part uses the order and grouping results of the first part to sequentially process the data sets.

3. The neural network active adjustment method of an active surface antenna according to claim 1, characterized in that: Change the model output quantity to single output to train and predict different actuators respectively, and add an emphasis layer to enable the model to have the ability to adapt to node coordinate weights.

Images