Phased array online beam calibration method based on deep learning

Abstract

The invention discloses a phased array online beam calibration method based on deep learning. The phased array online beam calibration method comprises three parts of directional diagram data acquisition and processing, initial model construction and model training, and model debugging and application. The method comprises the following steps of: randomly setting amplitude and phase parameters of a phased array by using an amplitude and phase control device, connecting a test antenna to a frequency spectrograph, connecting an upper computer with the frequency spectrograph through a switch, driving the phased array antenna to rotate by using a rotary table, and drawing a directional diagram of a received signal in the upper computer; then constructing a neural network model and training by using a directional diagram; and finally, carrying out model verification by utilizing data of the verification set, and if a verified loss function does not meet the requirements of the method, carrying out model debugging. According to the invention, the amplitude and phase parameters of each channel corresponding to the target beam are directly obtained through the deep learning model, the calculation of the amplitude and phase error value of each channel is avoided, the calibration time and difficulty are greatly reduced, and the calibration efficiency is improved.

Description

Technical Field

[0001] The present invention is a phased array online beam calibration method based on a neural network, which mainly aims at the problem that there is a large difference between the synthetic beam and the target beam caused by manufacturing process errors and array element mutual coupling, and realizes the calibration of the synthetic beam so that the performance meets the design requirements. Background Art

[0002] Digital beamforming technology is a combination of antenna technology and digital signal processing technology. It generates a beam with a specific direction by weighting different elements in the antenna array. Beamforming technology concentrates energy into a certain beam instead of propagating it in all directions. It has the advantages of flexible beam suppression, high signal gain, strong anti-interference ability and high spatial resolution ability. It is widely used in radar signal processing and electronic countermeasure systems.

[0003] Due to errors in the antenna manufacturing process and mutual coupling between antenna elements, especially at high frequencies, the antenna mutual coupling is more serious, resulting in serious amplitude and phase errors between the channels of the phased array antenna, poor channel consistency, and a large difference between the synthesized beam and the target beam, which does not meet the design requirements and requires beam calibration.

[0004] In order to solve the above problems, there are currently methods such as the rotation vector method.

[0005] The rotation vector method is to transmit a calibration signal to the phased array antenna through the detection antenna, control a unit in the phased array antenna to rotate the phase, and keep the other units in the original state. By changing the phase of each unit one by one and measuring the maximum and minimum values ​​of the amplitude of the received signal, the amplitude and phase error values ​​of each array element can be calculated, and finally the calibration and compensation of the entire antenna array can be realized. This method has low requirements on equipment and only needs to measure the amplitude value to complete the calibration, but it has the following disadvantages:

[0006] 1. The rotating vector method can only calibrate the directional pattern of a certain angle at a time, and it takes a long time to complete a calibration; for a new directional pattern, the amplitude and phase error values ​​need to be recalculated for calibration, which is inefficient.

[0007] 2. Insufficient data utilization: When the rotating vector method is used for phased array calibration, only the maximum and minimum amplitude data are used. Other data are also generated in the phased array system but are not used.

[0008] 3. Insensitive to signal changes. For large arrays, changes in the phase of a single array element sometimes make it difficult to cause significant changes in the test signal, which may lead to inaccurate calibration results. Summary of the invention

[0009] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a phased array online beam calibration method based on deep learning. The present invention directly obtains the amplitude and phase parameters of each channel corresponding to the target beam through a deep learning model, avoiding the calculation of the amplitude and phase error values ​​of each channel, thereby completing the calibration of the beam, greatly reducing the time and difficulty of calibration, and improving the efficiency of calibration.

[0010] The objective of the present invention is achieved through the following technical solution: A phased array online beam calibration method based on deep learning, comprising the following steps:

[0011] S1. Set the size of the phased array to be a uniform planar transmitting array of N×N, and the operating frequency to be f0~F0;

[0012] S2, the signal source transmits a total of M+1 frequency signals to form a set frequency set H; the signal emitted by the signal source for the kth time is recorded as f k , k ＝ 0, … ,M; Use the amplitude and phase control device to randomly set K groups of amplitude parameters and phase parameters at each frequency to form an amplitude parameter set Φ A and the phase parameter set Φ P , the amplitude parameter A of the i-th group in the set i and phase parameter P i are all N×N matrices, representing the N×N matrices formed by the amplitude parameters and phase parameters set for each unit in the i-th uniform planar transmitting array, i ＝ 1, … ,K;

[0013] S3, initialize k=0, i=1;

[0014] S4, signal source transmits signal f k , f k Passed to the phased array via connections;

[0015] S5. Get the received signal pattern: Connect the test antenna to the spectrum analyzer, and then connect the host computer to the spectrum analyzer through a switch. Use the turntable to drive the phased array antenna to rotate, and draw the transmission signal f in the host computer. k , the amplitude parameter is A i , the phase parameter is P i When the directional diagram of the received signal is g k,i (f k ,A i ,P i );

[0016] S6, determine whether i is less than K, if so, set i=i+1 and go to step S5; if not, go to the next step;

[0017] S7, determine whether k is less than M, if so, set k = k + 1, and go to step S4; if not, it means that the directional patterns corresponding to all frequencies in the frequency set H have been obtained, and a total of (M + 1) K directional patterns are obtained: g 0,1 (f0,A1,P1),g 0,2 (f0,A2,P2),g M,K (f M ,A K ,P K );

[0018] S8, preprocessing the directional pattern data: denoising and normalizing the directional pattern data, and adjusting the size of the directional pattern according to the size of the subsequent model input layer; then randomly dividing the directional pattern data into two groups: a training set and a validation set, with the training set accounting for a% and the validation set accounting for (1-a)%;

[0019] S9. Build the initial deep learning model: the model includes input layer, hidden layer and output layer in sequence; the input is directional map data; the hidden layer includes convolution layer, pooling layer and fully connected layer in sequence: the number of convolution layers is 1, the convolution kernel size is set to 3×3, and the number of convolution kernels is N; the number of pooling layers is 1, the pooling window is 2×2, the step size is 2, and average pooling is selected; the number of fully connected layers is 1, and the number of neurons is 2N 2 After each hidden layer, a ReLU activation function is connected. The function expression is f(x)=max(0,x). The output is the predicted phased array parameters. The number of neurons in the output layer is 2N. 2 indivual;

[0020] S10, set the initial hyperparameters for model training, including learning rate lr, number of iterations epochs and batch size batch_size; and determine the hyperparameter combination set G;

[0021] S11. Perform model training: Use the training set data to train the model and record the training loss function;

[0022] S12, after the model training is completed, the model is verified; the verification set data is input into the model to determine whether the mean square error between the predicted phased array parameters and the actual results meets the preset requirements. If so, the model training is successful, the model structure and parameters are saved, the model is applied, and step S13 is executed; if not, step S14 is executed;

[0023] S13, apply the model: the target direction map g x After preprocessing, it is input into the model, and the model output is

[0024] Then the phased array parameters output by the model are Input the phased array, obtain the directional pattern corresponding to the model, and observe whether the directional pattern corresponding to the model fits the target directional pattern. If so, the model is successfully applied, the operation is terminated, and the final deep learning model is obtained; if not, increase M and K, and go to step S2 for retraining;

[0025] S14, determine whether the hyperparameters in the hyperparameter combination set G have been explored. If not, change the hyperparameters, continue to explore the hyperparameter combination set, and go to step S11; if yes, compare the training and verification results, find the optimal hyperparameter combination under the current structure and set the optimal hyperparameter combination as the initial hyperparameter for the next model training, and go to the next step;

[0026] S15. Change the model width: Change the number of convolution kernels F in the convolution layer and the number of neurons U in the fully connected layer: F = N + N / 10·m, U = 2N 2 +2N 2 / 10·m, the initial value of m is 0, determine whether m is less than 10, if so, set m=m+1 and go to step S11; if not, go to the next step;

[0027] S16. Change the number of model layers: set n=n+1, m=0, that is, the number of convolutional layers and fully connected layers are increased by one layer, and the number of convolution kernels F and the number of neurons U in the fully connected layer are set to initial values, and then determine whether n is less than or equal to N. If so, go to step S15.

[0028] The beneficial effect of the present invention is that the present invention introduces deep learning in the beam calibration of the phased array. During the calibration process, the amplitude and phase errors of each channel of the phased array are not processed. Instead, the amplitude and phase parameters of each channel corresponding to the target beam are directly obtained through the deep learning model, avoiding the calculation of the amplitude and phase error values ​​of each channel, thereby completing the calibration of the beam, greatly reducing the time and difficulty of calibration, and improving the efficiency of calibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the structural diagram of the calibration system;

[0030] Figure 2 Flowchart of online beam calibration of phased array based on neural network. DETAILED DESCRIPTION

[0031] In view of the problem that there is a large difference between the synthetic beam and the target beam caused by manufacturing process errors and array element mutual coupling, the present invention proposes a phased array online beam calibration method based on deep learning, which realizes the calibration of the synthetic beam, greatly reduces the time and difficulty of calibration, and enables the beam performance to meet the design requirements. The technical solution of the present invention is further explained in conjunction with the accompanying drawings.

[0032] The calibration system used in the present invention is as follows Figure 1 As shown, it includes an amplitude and phase control device, an N×N uniform planar transmitting array (TX), a turntable, a signal source, a host computer, a switch, a spectrum analyzer and a test antenna (RX). Then, the amplitude and phase control device is used to randomly set the amplitude parameters and phase parameters of the phased array, the test antenna is connected to the spectrum analyzer, and then the host computer is connected to the spectrum analyzer through the switch. The turntable is used to drive the phased array antenna to rotate, and the directional diagram of the received signal is drawn in the host computer.

[0033] Before calibration, the scale and parameters of the phased array need to be given, and then according to the design ideas of the present invention, the directional pattern data and deep learning model are obtained, and then the model is trained, debugged and applied to achieve the purpose of phased array beam calibration.

[0034] like Figure 2 As shown, the present invention is a phased array online beam calibration method based on deep learning, which is generally divided into three parts: 1. Collection and processing of directional pattern data; 2. Initial model construction and model training; 3. Model debugging and application.

[0035] The first part is the collection and processing of the directional pattern data, and the specific process is shown in steps S1 to S8.

[0036] S1. Set the size of the phased array to be a 10×10 uniform planar transmitting array with an operating frequency of 10 GHz to 30 GHz;

[0037] S2, the signal source transmits a total of 201 frequency signals to form a set frequency set H; the signal transmitted by the signal source for the kth time is recorded as f k , f k To set the frequency in frequency set H, k=0,…,200; use the amplitude and phase control device to randomly set 500 sets of amplitude parameters and phase parameters at each frequency to form an amplitude parameter set Φ A and the phase parameter set Φ P , the amplitude parameter A of the i-th group in the set i and phase parameter P i are all 10×10 matrices, representing the 10×10 matrices formed by the amplitude parameters and phase parameters set for each unit in the i-th uniform planar transmitting array, i ＝ 1, … ,500;

[0038] S3, initialize k=0, i=1;

[0039] S4, signal source transmits signal f k , f k Delivered to the phased array via a coaxial cable connection;

[0040] S5. Get the received signal pattern: Connect the test antenna to the spectrum analyzer, and then connect the host computer to the spectrum analyzer through a switch. Use the turntable to drive the phased array antenna to rotate, and draw the transmission signal f in the host computer. k , the amplitude parameter is A i , the phase parameter is P i When the directional diagram of the received signal is g k,i (f k ,A i ,P i );

[0041] S6, determine whether i is less than 500, if so, set i=i+1 and go to step S5; if not, go to the next step;

[0042] S7, determine whether k is less than 200, if so, set k = k + 1, and go to step S4; if not, it means that the directional patterns corresponding to all frequencies in the frequency set H have been obtained, and a total of 201 500 = 100500 directional patterns are obtained: g 0,1 (f0,A1,P1),g 0,2 (f0,A2,P2),g 200,500 (f 200 ,A 500 ,P 500 );

[0043] S8. Preprocess the directional map data: denoise and normalize the directional map data, and adjust the size of the directional map according to the size of the subsequent model input layer; the specific methods of preprocessing are: 1. Denoising: use filtering technology to smooth the image and remove noise. 2. Normalization: normalize the image pixel values ​​to [0,1] to facilitate comparison and processing, while improving computational efficiency. 3. Size adjustment: according to the size of the subsequent model input layer [227,227], adjust the size of the directional map appropriately. Then randomly divide the directional map data into 2 groups: training set and validation set, with the training set accounting for 70% and the validation set accounting for 30%;

[0044] The second part is initial model construction and model training, as shown in steps S9 to S11.

[0045] S9. According to the set phased array scale, construct the initial deep learning model: the model includes input layer, hidden layer and output layer in sequence; the input is directional pattern data, and the input layer size is [227,227]; the hidden layer includes convolution layer, pooling layer and fully connected layer in sequence: the initial number of convolution layers is 1, the convolution kernel size is set to 3×3, the number of convolution kernels is 10, and the feature information of the image is extracted through the convolution kernel to obtain 10 output feature maps; the number of pooling layers is 1, the pooling window is 2×2, the step size is 2, and average pooling is selected to reduce the dimension of the output feature map by half; the number of initial fully connected layers is 1, and the number of neurons is 2·10 2 = 200, the features extracted from the previous layer can be integrated and reorganized; since the model needs to introduce nonlinearity rather than stacking linear networks, a ReLU activation function is connected after each layer in the hidden layer, and the function expression is f(x) = max(0,x); the output is the predicted phased array parameters, and the number of neurons in the output layer is 200;

[0046] S10. After the initial model is constructed, the initial hyperparameters for model training are set, including the learning rate lr, the number of iterations epochs, and the batch size batch_size; and the hyperparameter combination set G is determined, and the hyperparameter combination set G includes multiple sets of hyperparameter combinations; in this embodiment, the initial learning rate is set to 0.01, the number of iterations is 100, and the batch size is 250.

[0047] S11. Perform model training: Use the training set data to train the model and record the training loss function.

[0048] The specific steps of model training are: 1. Input the directional pattern data and calculate the predicted phased array parameters A* and P* through forward propagation; 2. Calculate the loss function (Loss), where the mean square error is selected as the loss function, let Y * =(A * ,P * ), similarly, the actual phased array parameters Y = (A, P),

[0049] Y ab Represents the element in the ath row and bth column of the matrix Y. The loss function reflects the error between the true value and the predicted value. 3. Based on the calculated loss function, the back propagation algorithm is used to calculate the gradient of the parameters of each layer from the output layer to the input layer layer by layer, also known as the error term. 4. The gradient descent algorithm is used to update the parameters and weights of each layer to reduce the loss function. 5. Continue training until the loss function converges to a certain threshold, then stop training.

[0050] The third part is model debugging and application: after the model training is completed, the model is verified using the data of the verification set. If the verified loss function does not meet the requirements of the present invention, the model is debugged, and the iterative debugging method is adopted. First, by exploring the hyperparameter combination set G, by comparing the training and verification results, the optimal hyperparameter combination under the current model structure is found. If the loss function still does not meet the requirements, the model hyperparameters are set to the optimal hyperparameter combination, and then the model structure is adjusted: first keep the number of model layers unchanged, adjust the width of the model, if the verification loss function still does not meet the requirements, adjust the depth of the model, that is, the number of layers of the model, until the model structure and parameters that meet the verification loss function requirements are found. The specific process is shown in steps S12 to S16.

[0051] S12, after the model training is completed, the model is verified; the verification set data is input into the model to determine whether the mean square error between the predicted phased array parameters and the actual results meets the preset requirements, that is, whether the mean square error is less than the preset threshold value of 0.05; if so, the model training is successful, the model structure and parameters are saved, the model is applied, and step S13 is executed; if not, step S14 is executed;

[0052] S13, apply the model: the target direction map g x After preprocessing, it is input into the model, and the model output is Then the phased array parameters output by the model are Input the phased array, get the directional pattern corresponding to the model, and observe whether the directional pattern corresponding to the model fits the target directional pattern. If so, the model is successfully applied, the operation is ended, and the final deep learning model is obtained; if not, it is considered that the target directional pattern g x If there is a big difference with the data set, increase M and K, set M = 300, K = 600, that is, increase the amount of data, and go to step S2 for retraining;

[0053] S14, using an iterative debugging method to determine whether the hyperparameters in the hyperparameter combination set G have been explored. If not, change the hyperparameters, continue to explore the hyperparameter combination set, and go to step S11; if yes, compare the training and verification results, find the optimal hyperparameter combination under the current structure, set the optimal hyperparameter combination as the initial hyperparameter for the next model training, and go to the next step;

[0054] S15. Keep the number of model layers unchanged and change the model width: change the number of convolution kernels F in the convolution layer and the number of neurons U in the fully connected layer: F = 10 + 10 / 10·m, U = 200 + 200 / 10·m, the initial value of m is 0, determine whether m is less than 10, if so, set m = m + 1 and go to step S11; if not, go to the next step; that is, the number of convolution kernels F and the number of neurons U in the fully connected layer are at most doubled, and after each adjustment, the model is retrained and debugged. When the number of convolution kernels F and the number of neurons U in the fully connected layer reach the upper limit of the adjustment, if the verification loss function still does not meet the requirements, adjust the depth of the model, that is, the number of layers of the model, and execute S16; in the process of retraining the model, first use the optimal hyperparameter combination of the previous model structure for training. If the training and verification are successful, the model can be directly applied. If the training or verification is unsuccessful, continue to explore the remaining hyperparameters in G, which can save debugging time and improve debugging efficiency.

[0055] S16. Change the number of model layers: Select and adjust the number of convolutional layers and fully connected layers n, with the initial value of n being 1. Set n=n+1, and add one layer each time, with a maximum of 10 layers; after each change in the number of layers, set m=0 (that is, the number of convolutional layers and fully connected layers are increased by one layer, and the number of convolution kernels F and the number of neurons U in the fully connected layer are set to the initial values, and the number of convolution kernels F and the number of neurons U in the fully connected layer are readjusted based on the change in the number of model layers). Then determine whether n is less than or equal to 10, and if so, go to step S15. After each adjustment, retrain and debug the model until the model structure and parameters that meet the requirements of the verification loss function are found.

[0056] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A phased array online beam calibration method based on deep learning, characterized in that: The following steps are involved: S1. Set the size of the phased array to be a uniform planar transmitting array of N×N, and the operating frequency to be f0~F0; S2, the signal source transmits a total of M+1 frequency signals to form a set frequency set H; the signal emitted by the signal source for the kth time is recorded as f k , The amplitude and phase control device is used to randomly set K groups of amplitude parameters and phase parameters at each frequency to form an amplitude parameter set Φ A and the phase parameter set Φ P , the amplitude parameter A of the i-th group in the set i and phase parameter P i are all N×N matrices, representing the N×N matrices formed by the amplitude parameters and phase parameters set for each unit in the i-th uniform planar transmitting array, i=1,…,K; S3, initialize k=0, i=1; S4, signal source transmits signal f k , f k Passed to the phased array via connections; S5. Get the received signal pattern: Connect the test antenna to the spectrum analyzer, and then connect the host computer to the spectrum analyzer through a switch. Use the turntable to drive the phased array antenna to rotate, and draw the transmission signal f in the host computer. k , the amplitude parameter is A i , the phase parameter is P i When the directional diagram of the received signal is g k,i (f k ,A i ,P i ); S6, determine whether i is less than K, if so, set i=i+1 and go to step S5; if not, go to the next step; S7, determine whether k is less than M, if so, set k = k + 1, and go to step S4; if not, it means that the directional patterns corresponding to all frequencies in the frequency set H have been obtained, and a total of (M + 1) K directional patterns are obtained: g 0,1 (f0,A1,P1),g 0,2 (f0,A2,P2),g M,K (f M ,A K ,P K ); S8, preprocessing the directional pattern data: denoising and normalizing the directional pattern data, and adjusting the size of the directional pattern according to the size of the subsequent model input layer; then randomly dividing the directional pattern data into two groups: a training set and a validation set, with the training set accounting for a% and the validation set accounting for (1-a)%; S9. Build the initial deep learning model: the model includes input layer, hidden layer and output layer in sequence; the input is directional map data; the hidden layer includes convolution layer, pooling layer and fully connected layer in sequence: the number of convolution layers is 1, the convolution kernel size is set to 3×3, and the number of convolution kernels is N; the number of pooling layers is 1, the pooling window is 2×2, the step size is 2, and average pooling is selected; the number of fully connected layers is 1, and the number of neurons is 2N 2 After each hidden layer, a ReLU activation function is connected. The function expression is f(x)=max(0,x). The output is the predicted phased array parameters. The number of neurons in the output layer is 2N. 2 indivual; S10, set the initial hyperparameters for model training, including learning rate lr, number of iterations epochs and batch size batch_size; and determine the hyperparameter combination set G; S11. Perform model training: Use the training set data to train the model and record the training loss function; S12, after the model training is completed, the model is verified; the verification set data is input into the model to determine whether the mean square error between the predicted phased array parameters and the actual results meets the preset requirements. If so, the model training is successful, the model structure and parameters are saved, the model is applied, and step S13 is executed; if not, step S14 is executed; S13, apply the model: the target direction map g x After preprocessing, it is input into the model, and the model output is Then the phased array parameters output by the model are Input the phased array, obtain the directional pattern corresponding to the model, and observe whether the directional pattern corresponding to the model fits the target directional pattern. If so, the model is successfully applied, the operation is terminated, and the final deep learning model is obtained; if not, increase M and K, and go to step S2 for retraining; S14, determine whether the hyperparameters in the hyperparameter combination set G have been explored. If not, change the hyperparameters, continue to explore the hyperparameter combination set, and go to step S11; if yes, compare the training and verification results, find the optimal hyperparameter combination under the current structure and set the optimal hyperparameter combination as the initial hyperparameter for the next model training, and go to the next step; S15. Change the model width: Change the number of convolution kernels F in the convolution layer and the number of neurons U in the fully connected layer: F = N + N / 10·m, U = 2N 2 +2N 2 / 10·m, the initial value of m is 0, determine whether m is less than 10, if so, set m=m+1 and go to step S11; If not, proceed to the next step; S16. Change the number of model layers: set n=n+1, m=0, that is, the number of convolutional layers and fully connected layers are increased by one layer, and the number of convolution kernels F and the number of neurons U in the fully connected layer are set to initial values, and then determine whether n is less than or equal to N. If so, go to step S15.

Images