A phased array amplitude and phase calibration method based on machine learning

Abstract

The present invention discloses a phased array amplitude and phase calibration method based on machine learning, which belongs to the field of antenna technology. The method comprises the following steps: S1: obtaining multiple one-to-one corresponding input combinations and output combinations by performing electromagnetic simulation on the phased array antenna array surface to form a data set; S2: establishing a machine learning model; S3: using the machine learning model established in S2 to train and learn the data set described in S1; S4: judging the training result according to the set error value, when the error does not meet the requirement, repeating S3 to continue learning the data set described in S1; when the error meets the requirement, executing S5; S5: inputting the phased array data of the actual test into the machine learning model, and outputting the corresponding amplitude and phase results; S6: using the obtained amplitude and phase data of all units to calibrate the phased array. Machine learning can reduce the number of measurements, and the amplitude and phase distribution of each unit of the actual phased array can be obtained.

Description

Technical Field

[0001] The present invention relates to the field of antenna technology, and in particular to a phased array amplitude and phase calibration method based on machine learning. Background Art

[0002] At present, there are two main amplitude and phase calibration technologies for phased arrays, namely the single-channel detection method and the rotating vector method. Among them, the single-channel detection method is a more commonly used technology. The principle of this method is as follows: the RF port of the phased array antenna and the test probe (usually a standard gain horn or an open rectangular waveguide) are connected to the two ports of the vector network analyzer respectively. Then, by opening the RF channel corresponding to a certain antenna unit and closing other channels at the same time, only this channel works at the same time. For the transmitting phased array, its electromagnetic energy is radiated into space through the antenna unit on the channel, and the test probe receives the electromagnetic waves radiated by the unit and reads the amplitude and phase information through the vector network analyzer. For the receiving phased array, the test probe transmits electromagnetic waves and is received by the antenna unit, and then the amplitude and phase information is read through the vector network analyzer. By opening all channels on the phased array in turn, the amplitude and phase distribution information of the entire phased array can be obtained, and the amplitude and phase compensation information can be further obtained so that the amplitude and phase of all channels are consistent when working, thus completing the calibration work. During the calibration process, the rotating vector method does not need to open a certain channel and close other channels at the same time, but all channels remain open. It is necessary to change the amplitude and phase settings of a channel while keeping the amplitude and phase settings of other channels unchanged, and record the test data of the vector network analyzer twice before and after. In this way, the same operation is performed on all channels in turn, and the amplitude and phase differences between all units are calculated by calculating the data before and after each unit, thus completing the calibration.

[0003] Whether it is the rotating vector method or the single-channel detection method, it is necessary to calibrate by controlling the amplitude and phase changes or the switching state of each unit. Each unit needs to be measured once or even multiple times, and the data needs to be recorded. The larger the phased array, the more channels are tested and the longer it takes. Most existing methods require calibration by testing a large amount of data, which is time-consuming and inefficient.

[0004] The phased array amplitude and phase calibration method in the prior art, such as the amplitude and phase calibration method of a phased array antenna disclosed in CN 108809447B, improves the phase shift resolution of the phase shifter to obtain a higher accuracy of the resultant vector power change curve, so that the relative amplitude and phase of each radiating unit of the phased array antenna can be obtained more accurately, thereby performing amplitude and phase calibration more accurately.

[0005] However, it still requires traversal sampling of all unit sampling points to perform amplitude and phase calibration, which takes a long time and has low test efficiency. Summary of the invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a phased array amplitude and phase calibration method based on machine learning.

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

[0008] A phased array amplitude and phase calibration method based on machine learning, characterized in that it includes the following steps:

[0009] S1: Perform electromagnetic simulation on the phased array antenna surface through electromagnetic simulation software to obtain multiple one-to-one corresponding input combinations and output combinations to form a data set;

[0010] S2: Build a machine learning model;

[0011] S3: Use the machine learning model established in S2 to train and learn the data set described in S1;

[0012] S4: The training result is judged according to the set error value. When the error does not meet the requirement, S3 is repeated to continue learning the data set described in S1. When the error meets the requirement, a trained machine learning model is obtained and S5 is executed.

[0013] S5: Input the actual tested phased array data into the trained machine learning model and output the corresponding amplitude and phase results;

[0014] S6: Use the obtained amplitude and phase data of all units to perform phased array calibration.

[0015] Preferably, the output combination is the excitation amplitude and excitation phase information of the phased array antenna unit. For an M×N phased array antenna, in the electromagnetic simulation software, a power Ai and a phase Pi are set for each unit, and M×N amplitude and phase information are obtained to obtain an output combination.

[0016] Preferably, the power Ai ranges from 0 to 1 W, and the phase Pi ranges from 0 to 360°.

[0017] Preferably, the input combination is to obtain the amplitude and phase information above the phased array antenna plane through electromagnetic simulation software. After setting an output combination for the phased array antenna unit, an input combination is obtained by selecting m×n points at d above the phased array antenna plane in the electromagnetic simulation software and simulating the m×n points to obtain the amplitude and phase information at the positions of the m×n points.

[0018] Preferably, the m×n points are evenly distributed above the phased array, where m<M, n<N. The closer m×n is to M×N, the higher the accuracy of the machine learning model, and accordingly, more points need to be tested in actual testing.

[0019] Preferably, the training learning can select error calculation input combination data from the simulation data outside the data set and input it into the machine learning model. The machine learning model correspondingly outputs the same number of error calculation output combinations. The error data can be obtained by comparing the output combination consisting of the amplitude and phase of each unit set in the electromagnetic simulation software.

[0020] Preferably, the S5 includes measuring the M×N phased array antenna actually tested, taking m×n points at a position d above the M×N phased array actually tested, with a distribution distance consistent with the electromagnetic simulation model, and testing the amplitude and phase at these points to form an input combination, and then inputting the input combination into the machine learning model to obtain an output combination, i.e., the amplitude and phase excitation data of each unit of the M×N phased array actually tested.

[0021] Preferably, after the calibration data is calculated based on the amplitude and phase excitation data of each unit of the M×N phased array actually tested, the phased array is subjected to amplitude and phase calibration.

[0022] Preferably, the S2 uses the Pytorch machine learning library to build a machine learning model, including the following steps:

[0023] S21: Set the number of neurons in the neural network input layer to m×n;

[0024] S22: Set the number of neurons in the output layer of the neural network to M×N;

[0025] S23: Add K hidden layers, the number of neurons in each layer is set to L, and K is greater than or equal to 3;

[0026] S24: Select the ReLU function as the activation function of the neural network input layer, the first hidden layer, and the second hidden layer, and do not process the last hidden layer;

[0027] S25: Set all parameters in the machine learning model to 0;

[0028] The amplitude and phase information X collected at the m×n points above the phased array is used as the input of the neural network, the amplitude and phase excitation Y set for each antenna unit in the simulation software is used as the true value of the neural network output, and {X, Y} constitutes a training set; data of a first preset value is selected from the training set as training data, data of a second preset value is selected as evaluation data, and data of a third preset value is selected as test data; when the error of the evaluation data and the error of the training data are greater than the fourth preset value, the values ​​of K and L are adjusted until the error of the evaluation data and the error of the training data are less than the fifth preset value.

[0029] Preferably, the S3 uses a small batch stochastic gradient descent method to train the data set, and each parameter update includes the following steps:

[0030] The data of the sixth preset value in the data set is randomly selected for forward propagation to calculate the neural network prediction value, and then the MSE error between the neural network prediction value and the true value of the neural network output is calculated, and finally the MSE error between the neural network prediction value and the true value of the neural network output is back-propagated to transfer the gradient from the neural network output layer to the neural network input layer layer by layer, and update the network parameters of each layer; when the average phase deviation of all units is within the first preset range and the average amplitude deviation is within the second preset range, the training is terminated; wherein MES is mean square error, and the mean square error is used to measure the gap between the neural network prediction value and the true value of the neural network output.

[0031] The beneficial effects of the present invention are:

[0032] 1) Amplitude and phase calibration can be performed without traversal testing of all phased array antenna units, which shortens the calibration time and improves the test efficiency.

[0033] 2) Only m×n tests are required to quickly obtain calibration data for all units. Since m×n is less than M×N, the reduction in the number of measurements can effectively improve the measurement efficiency, which is M×N / (m×n) times that of the traditional method. This is of great significance for the calibration process of ultra-large-scale phased arrays or mass-produced phased array products, especially because mass-produced phased array products can use the same training set, which greatly saves time. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the phased array amplitude and phase calibration method based on machine learning;

[0035] Figure 2 This is a top view of an M×N scale phased array and m×n selected measurement points;

[0036] Figure 3 This is a side view of an M×N phased array and m×n selected measurement points. DETAILED DESCRIPTION

[0037] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0038] The following will take machine learning as an example to introduce an embodiment of the present invention, but the learning and training process is not limited to machine learning. Other computer software algorithms with training and learning functions such as artificial intelligence, neural networks, deep learning, etc. can be applied to the present invention. Other algorithm solutions that use basically the same design ideas and achieve basically the same effects as the present invention should all be within the scope of protection of the present invention.

[0039] See also Figure 1-Figure 3 , the present invention provides a technical solution:

[0040] A phased array amplitude and phase calibration method based on machine learning, characterized in that it includes the following steps:

[0041] S1: Perform electromagnetic simulation on the phased array antenna surface through electromagnetic simulation software to obtain multiple one-to-one corresponding input combinations and output combinations to form a data set;

[0042] S2: Build a machine learning model;

[0043] S3: Use the machine learning model established in S2 to train and learn the data set described in S1;

[0044] S4: The training result is judged according to the set error value. When the error does not meet the requirement, S3 is repeated to continue learning the data set described in S1. When the error meets the requirement, a trained machine learning model is obtained and S5 is executed.

[0045] S5: Input the actual tested phased array data into the trained machine learning model and output the corresponding amplitude and phase results;

[0046] S6: Use the obtained amplitude and phase data of all units to perform phased array calibration.

[0047] Compared with traditional phased array calibration technology, this solution can greatly reduce the number of measurements, and can obtain the amplitude and phase distribution of each unit of the actual phased array. Further, the calibration data can be calculated based on the amplitude and phase distribution to calibrate the phased array. Since the number of measurement points (i.e., the number of times) is reduced, the test efficiency is greatly improved, which can save a lot of calibration time.

[0048] In some embodiments, the output combination is the excitation amplitude and excitation phase information of the phased array antenna unit. For an M×N scale phased array antenna, in the electromagnetic simulation software, a power Ai and a phase Pi are set for each unit, and M×N amplitude and phase information are obtained to obtain an output combination. In order to facilitate the understanding of the use process of the present invention, the present invention takes an M×N scale rectangularly arranged phased array as an example to explain the technical principle of the present invention. In actual operation, the scope of application of the present invention is not limited to an M×N scale rectangular distributed phased array, and can be applied to any arranged or distributed phased array.

[0049] In some embodiments, the power Ai ranges from 0 to 1 W, and the phase Pi ranges from 0 to 360°. Phase is the position of a wave at a specific moment in its cycle: a measure of whether it is at a peak, a trough, or a point in between. Phase describes the measure of the change in the signal waveform, usually in degrees (angles), also known as the phase angle. When the signal waveform changes in a periodic manner, one cycle of the waveform is 360°.

[0050] In some embodiments, the input combination is to obtain the amplitude and phase information above the phased array antenna plane through electromagnetic simulation software. After setting an output combination for the phased array antenna unit, an input combination is obtained by selecting m×n points at d above the phased array antenna plane in the electromagnetic simulation software, and simulating the m×n points to obtain the amplitude and phase information of the positions of the m×n points. Different from the rotating vector method and the single-channel switch method, the amplitude and phase data of all channels need to be measured, that is, for an array of M×N scale, at least M×N measurements are required. However, using the calibration technology proposed in the present invention, only m×n tests are required to quickly obtain the calibration data of all units. Since m×n is less than M×N, the reduction in the number of measurements can effectively improve the measurement efficiency, and its test efficiency is M×N / (m×n) times that of the traditional method. This is of great significance for the calibration process of ultra-large-scale phased arrays or mass-produced phased array products, especially mass-produced phased array products can use the same training set, which greatly saves time. In this embodiment, the input combination and output combination obtained by electromagnetic simulation are one-to-one corresponding, and a large number of such corresponding relationships form a data set. Since a set of amplitude and phase values, i.e., one output combination of data, is set for all phased array units in electromagnetic simulation, the amplitude and phase data measured at m×n points above the phased array are unique, so the relationship between the input combination and the output combination is one-to-one.

[0051] In some embodiments, the m×n points are evenly distributed above the phased array, where m<M, n<N. The closer m×n is to M×N, the higher the accuracy of the machine learning model, and accordingly, more points need to be tested in actual testing. The number of points selected cannot be too small, otherwise it will lead to large errors in subsequent output results. In order to facilitate the understanding of the use of the present invention, the test points above the phased array are distributed in an m×n rectangular shape as an example to illustrate the technical principles of the present invention. In actual operation, the scope of application of the present invention is not limited to the m×n rectangular measurement point distribution, and can be applied to any suitable distribution of measurement points.

[0052] In some embodiments, the training learning can select error calculation input combination data from the simulation data outside the data set and input it into the machine learning model. The machine learning model correspondingly outputs the same number of error calculation output combinations. The error data can be obtained by comparing the output combination composed of the amplitude and phase of each unit set in the electromagnetic simulation software.

[0053] In some embodiments, the S5 includes measuring the M×N phased array antenna actually tested, taking m×n points at a distance d above the M×N phased array actually tested, with a distribution distance consistent with the electromagnetic simulation model, and testing the amplitude and phase at these points to form an input combination, and then inputting the input combination into the machine learning model to obtain an output combination, namely, the amplitude and phase excitation data of each unit of the M×N phased array actually tested.

[0054] In some embodiments, after the calibration data is calculated based on the amplitude and phase excitation data of each unit of the M×N phased array actually tested, the phased array is calibrated for amplitude and phase.

[0055] In some embodiments, the S2 uses the Pytorch machine learning library to build a machine learning model, including the following steps:

[0056] S21: Set the number of neurons in the neural network input layer to m×n;

[0057] S22: Set the number of neurons in the output layer of the neural network to M×N;

[0058] S23: Add K hidden layers, the number of neurons in each layer is set to L, and K is greater than or equal to 3;

[0059] S24: Select the ReLU function as the activation function of the neural network input layer, the first hidden layer, and the second hidden layer, and do not process the last hidden layer;

[0060] S25: Set all parameters in the machine learning model to 0;

[0061] The amplitude and phase information X collected at the m×n points above the phased array is used as the input of the neural network, the amplitude and phase excitation Y set for each antenna unit in the simulation software is used as the true value of the neural network output, and {X, Y} constitutes a training set; data of a first preset value is selected from the training set as training data, data of a second preset value is selected as evaluation data, and data of a third preset value is selected as test data; when the error of the evaluation data and the error of the training data are greater than the fourth preset value, the values ​​of K and L are adjusted until the error of the evaluation data and the error of the training data are less than the fifth preset value.

[0062] Since a large amount of simulation data (e.g., 2560 sets of one-to-one corresponding input and output combination data) is obtained in steps S21-S24 (the electromagnetic simulation model is exactly the same as the phased array antenna array surface of the actual test, which can improve the accuracy of the actual test calibration), the amplitude and phase information X collected at the m×n points above the phased array is used as the input of the neural network, and the amplitude and phase excitation Y set for each antenna unit in the simulation software is used as the true value of the neural network output, then {X, Y} constitutes a training set. 70% is randomly selected as training data, 15% as evaluation data, and 15% as test data. The evaluation is to observe the fitting of the network during neural network training. If the error of the evaluation data is much larger than the error of the training data, then serious overfitting has occurred, and the scale of the neural network, that is, K and L, needs to be adjusted. If the difference is not large, then the neural network is considered to be well fitted.

[0063] In some embodiments, the S3 uses a small batch stochastic gradient descent method to train the data set, and each parameter update includes the following steps:

[0064] The data of the sixth preset value in the data set is randomly selected for forward propagation to calculate the neural network prediction value, and then the MSE error between the neural network prediction value and the true value of the neural network output is calculated, and finally the MSE error between the neural network prediction value and the true value of the neural network output is back-propagated to transfer the gradient from the neural network output layer to the neural network input layer layer by layer, and update the network parameters of each layer; when the average phase deviation of all units is within the first preset range and the average amplitude deviation is within the second preset range, the training is terminated; wherein MES is mean square error, and the mean square error is used to measure the gap between the neural network prediction value and the true value of the neural network output.

[0065] For example, if there are 2560 groups of {X, Y} in total and the batch size is 256, then 256 groups of {X, Y} are randomly selected for forward propagation to calculate the predicted value of the neural network each time the parameters are updated. Then, the MSE error between the predicted value and the true value is calculated. Finally, the error back propagation transfers the gradient from the output layer to the input layer layer by layer to update the network parameters of each layer. The training end condition can be set to the average phase deviation of all units ±5° and the average amplitude deviation ±0.5dB, and the training ends.

[0066] When all units of the phased array antenna are working, a continuous synthetic field can be generated above the aperture of the phased array antenna. Each unit has independent amplitude and phase. When the amplitude and phase of individual units or all units change, the synthetic field above the aperture is also different. At this time, a finite number of points are selected above the aperture of the phased array, and the amplitude and phase data collected at these points each time will be different. A large amount of unit amplitude and phase data is given by the simulation software to obtain the amplitude and phase data of the corresponding measurement points above the aperture of the antenna array surface. The amplitude and phase data of the measurement points above the aperture of the antenna array surface are used as the input combination, and a large amount of unit amplitude and phase data given by the simulation software is used as the output combination. Then, the machine learning model is used for learning and training, and when a certain error accuracy is achieved, the measured data above the aperture of the phased array can be input into the machine learning model to obtain the amplitude and phase data of each unit, and then calibration is performed. Since the finite number of points selected above the phased array antenna is less than the number of units, this is more efficient than the traditional calibration method.

[0067] Similarly, within the near field range of the phased array radiation aperture, measurement points sparser than the original phased array scale are used (the number of measurement points is less than the number of antenna units), and a data set is obtained by learning simulation through a computer software algorithm, and the actual test point data is input into the software model to obtain the amplitude and phase information of each unit of the actual phased array. All of these methods fall within the scope of protection of the present invention.

[0068] The above is only a preferred embodiment 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 embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A phased array amplitude and phase calibration method based on machine learning, characterized in that: The following steps are involved: S1: Perform electromagnetic simulation on the phased array antenna surface through electromagnetic simulation software to obtain multiple one-to-one corresponding input combinations and output combinations to form a data set; S2: Build a machine learning model; S3: Use the machine learning model established in S2 to train and learn the data set described in S1; S4: The training result is judged according to the set error value. When the error does not meet the requirement, S3 is repeated to continue learning the data set described in S1. When the error meets the requirement, a trained machine learning model is obtained and S5 is executed. S5: Input the actual tested phased array data into the trained machine learning model and output the corresponding amplitude and phase results; S6: calibrate the phased array using the obtained amplitude and phase data of all units; The output combination is the excitation amplitude and excitation phase information of the phased array antenna unit. For an M×N scale phased array antenna, in the electromagnetic simulation software, a power Ai and a phase Pi are set for each unit, and M×N amplitude and phase information are obtained to obtain an output combination; The input combination is to obtain the amplitude and phase information above the phased array antenna plane through electromagnetic simulation software. After setting an output combination for the phased array antenna unit, m×n points are selected at d above the phased array antenna plane in the electromagnetic simulation software, and the m×n points are simulated to obtain the amplitude and phase information of the positions of the m×n points to obtain an input combination; The S2 uses the Pytorch machine learning library to build a machine learning model, including the following steps: S21: Set the number of neurons in the neural network input layer to m×n; S22: Set the number of neurons in the output layer of the neural network to M×N; S23: Add K hidden layers, the number of neurons in each layer is set to L, and K is greater than or equal to 3; S24: Select the ReLU function as the activation function of the neural network input layer, the first hidden layer, and the second hidden layer, and do not process the last hidden layer; S25: Set all parameters in the machine learning model to 0; The amplitude and phase information X collected at the m×n points above the phased array is used as the input of the neural network, the amplitude and phase excitation Y set for each antenna unit in the simulation software is used as the true value of the neural network output, and {X, Y} constitutes a training set; data of a first preset value is selected from the training set as training data, data of a second preset value is selected as evaluation data, and data of a third preset value is selected as test data; when the error of the evaluation data and the error of the training data are greater than the fourth preset value, the values ​​of K and L are adjusted until the error of the evaluation data and the error of the training data are less than the fifth preset value; The S3 described above uses a small batch stochastic gradient descent method to train the data set, and each parameter update includes the following steps: The data of the sixth preset value in the data set is randomly selected for forward propagation to calculate the neural network prediction value, and then the MSE error between the neural network prediction value and the true value of the neural network output is calculated, and finally the MSE error between the neural network prediction value and the true value of the neural network output is back-propagated to transfer the gradient from the neural network output layer to the neural network input layer layer by layer, and update the network parameters of each layer; when the average phase deviation of all units is within the first preset range and the average amplitude deviation is within the second preset range, the training is terminated; wherein MES is mean square error, and the mean square error is used to measure the gap between the neural network prediction value and the true value of the neural network output.

2. The phased array amplitude and phase calibration method based on machine learning according to claim 1, characterized in that: The power Ai ranges from 0 to 1 W, and the phase Pi ranges from 0 to 360°.

3. The phased array amplitude and phase calibration method based on machine learning according to claim 1, characterized in that: The m×n points are evenly distributed above the phased array, wherein m<M, n<N.

4. The phased array amplitude and phase calibration method based on machine learning according to claim 1, characterized in that: The training learning can select error calculation input combination data from the simulation data outside the data set and input it into the machine learning model. The machine learning model correspondingly outputs the same number of error calculation output combinations. The error data can be obtained by comparing the output combination composed of the amplitude and phase of each unit set in the electromagnetic simulation software.

5. The phased array amplitude and phase calibration method based on machine learning according to claim 1, characterized in that: The S5 includes measuring the M×N phased array antenna actually tested, taking m×n points at a position d above the M×N phased array actually tested, with a distribution distance consistent with the electromagnetic simulation model, and testing the amplitude and phase at these points to form an input combination, and then inputting the input combination into the machine learning model to obtain an output combination, i.e., the amplitude and phase excitation data of each unit of the M×N phased array actually tested.

6. The phased array amplitude and phase calibration method based on machine learning according to claim 5, characterized in that: After the calibration data is calculated based on the amplitude and phase excitation data of each unit of the M×N phased array actually tested, the phased array is calibrated for amplitude and phase.

Images