JP2023100195A - Learning model generation method for array antenna and learning model generation program - Google Patents

Abstract

To eliminate much labor and time for measurement in estimating/correcting an initial excitation phase error and an initial excitation amplitude error (where only the initial excitation phase error not including the initial excitation amplitude error may be possible) of each element of an array antenna.SOLUTION: A learning model is generated for inputting radiation characteristics of an array antenna and outputting excitation characteristics of elements of the array antenna. In the learning model, a non-corrected radiation pattern of the array antenna is inputted, and the excitation characteristics of the elements of the array antenna are outputted. Namely, by generating the learning model beforehand, the excitation characteristics of the elements of the array antenna can be estimated only by measuring the non-corrected radiation pattern of the array antenna.SELECTED DRAWING: Figure 1

Description

The present disclosure relates to technology for estimating the excitation phase amplitude of each element of an array antenna.

Wireless communication has been required to improve communication capacity, low latency, and high reliability with the times, and the antenna, which is the transmitting end of the radio wave, has been miniaturized by supporting the high frequency band called millimeter wave band. On the other hand, since the spatial propagation loss becomes very large in high frequency bands, array antennas and related technologies have been developed in which a plurality of antenna elements are arranged to compensate for the power loss and radio waves are concentrated and radiated in a specific direction. When an array antenna is used as an antenna of a base station, an error occurs in the excitation phase amplitude of each element of the array antenna due to various hardware factors.

Conventionally, the excitation phase amplitude error of each element of an array antenna has been estimated and corrected by the REV method (Rotating element Electric-field Vector method) or the MEP method (Multi-Element Phase-toggle method) (for example, , Non-Patent Document 1, etc.).

Kiyoshi Mano, Takashi Katagi, "Method for measuring element amplitude and phase of phased array antenna - Element electric field vector rotation method -", Transactions of the Institute of Electronics, Information and Communication Engineers, The Institute of Electronics, Information and Communication Engineers, May 1982, Vol. B65, No. 5 , pp. 555-560.

In the REV method, the phase shifter is rotated once from 0 degrees to 360 degrees for each element of the array antenna, and changes in the combined power of the array are measured. A method for estimating the error and the initial excitation amplitude error. This method only needs to measure the combined power of the array, but the number of times of measuring the combined power of the array increases in proportion to the number of elements of the array antenna.

In the MEP method, the amount of phase shift by a phase shifter is changed for each element of the array antenna, and the array composite electric field (amplitude and phase) is measured when the excitation phase is changed at the same time for multiple elements or all elements at the same time. This method uses the results to estimate the initial excitation phase error and the initial excitation amplitude error of each element of the array antenna. Compared to the REV method, this method can estimate the initial excitation phase error and the initial excitation amplitude error at the same time for multiple elements or all elements at the same time. There is a problem that it is necessary to measure the array synthesized electric field (amplitude and phase), which has a large restriction on .

Furthermore, before performing measurements for estimating and correcting the initial excitation phase error and the initial excitation amplitude error of each element of a normal array antenna, the corrected Not only the radiation pattern but also the uncorrected radiation pattern of the array antenna is measured. As described above, the estimation and correction of the initial excitation phase error and the initial excitation amplitude error of each element of the array antenna require a lot of time and effort for measurement.

Therefore, in order to solve the above problems, the present disclosure estimates and estimates the initial excitation phase error and the initial excitation amplitude error of each element of the array antenna (not including the initial excitation amplitude error, only the initial excitation phase error is also possible). An object of the present invention is to eliminate the need for much labor and time for measurements in correction.

In order to solve the above problem, a learning model is generated that inputs the radiation characteristics of the array antenna and outputs the excitation characteristics of each element of the array antenna. Then, in the learning model, the uncorrected radiation pattern of the array antenna is input, and the excitation characteristic of each element of the array antenna is output. That is, if a learning model is generated in advance, the excitation characteristics of each element of the array antenna can be estimated only by measuring the radiation pattern of the array antenna before correction.

Specifically, the present disclosure provides an excitation phase or excitation phase amplitude of each element of an array antenna, and at least one of a power radiation pattern, a real electric field radiation pattern, and an imaginary electric field radiation pattern of the array antenna. and a power radiation pattern, an electric field real part radiation pattern and an electric field imaginary pattern of the array antenna using the plurality of teacher data acquired in the teacher data acquisition step. and a learning model generation step of inputting at least one of the partial radiation patterns and generating a learning model outputting the excitation phase or the excitation phase amplitude of each element of the array antenna. This is a learning model generation method for an array antenna.

Further, the present disclosure includes an input layer to which at least one of a power radiation pattern, a real electric field radiation pattern, and an imaginary electric field radiation pattern of an array antenna is input, and an excitation phase or excitation of each element of the array antenna. an output layer that outputs a phase amplitude; an excitation phase or excitation phase amplitude of each element of the array antenna; and at least one of a power radiation pattern, a real electric field radiation pattern, and an imaginary electric field radiation pattern of the array antenna. and at least one of a power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern of the array antenna. to the input layer, compute in the intermediate layer, and output the excitation phase or excitation phase amplitude of each element of the array antenna from the output layer. is a model.

According to these configurations, in estimating and correcting the initial excitation phase error and initial excitation amplitude error of each element of the array antenna (not including the initial excitation amplitude error, only the initial excitation phase error is also possible), many measurements Learning models can be generated to save effort and time.

Further, the present disclosure includes a radiation characteristic acquiring step of acquiring at least one of a power radiation pattern, a real part of an electric field radiation pattern, and an imaginary part of an electric field radiation pattern of an array antenna, and an excitation phase of each element of the array antenna. Or in a learning model generated using a plurality of teacher data comprising an excitation phase amplitude and at least one of a power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern of the array antenna, inputting at least one of the power radiation pattern, the real part of the electric field radiation pattern and the imaginary part of the electric field radiation pattern of the array antenna acquired in the radiation characteristics acquisition step, and the excitation phase or and an excitation characteristic estimation step of outputting an excitation phase amplitude.

Further, the present disclosure includes a radiation characteristic acquiring step of acquiring at least one of a power radiation pattern, a real part of an electric field radiation pattern, and an imaginary part of an electric field radiation pattern of an array antenna, and an excitation phase of each element of the array antenna. Or in a learning model generated using a plurality of teacher data comprising an excitation phase amplitude and at least one of a power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern of the array antenna, inputting at least one of the power radiation pattern, the real part of the electric field radiation pattern, and the imaginary part of the electric field radiation pattern of the array antenna acquired in the radiation characteristic acquisition step, and the excitation phase or An excitation characteristic estimation step for outputting an excitation phase amplitude, and an excitation characteristic estimation program for an array antenna for causing a computer to sequentially execute the excitation characteristic estimation step.

According to these configurations, in estimating and correcting the initial excitation phase error and initial excitation amplitude error of each element of the array antenna (not including the initial excitation amplitude error, only the initial excitation phase error is also possible), many measurements Learning models can be used to save effort and time.

Further, in the teacher data acquisition step, the electric power of the array antenna calculated using a simulation for the excitation phase of each element of the array antenna set randomly within a range of 360 degrees In the learning model generation method for an array antenna, at least one of a radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern is acquired as the plurality of teacher data.

According to this configuration, when there is a large excitation phase error in each element of the array antenna due to a hardware factor of each element of the array antenna during manufacture of the array antenna, the excitation phase error of each element of the array antenna is corrected. It can be estimated with some degree of accuracy. Then, simulation can be used to generate a large amount of training data.

Further, according to the present disclosure, in the teacher data acquisition step, the array antenna is randomly set within a predetermined range centered on an excitation phase that linearly changes with respect to the arrangement order number of each element of the array antenna. At least one of a power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern of the array antenna calculated using a simulation for the excitation phase of each element of the antenna; This is a learning model generation method for an array antenna characterized by acquiring as training data for an array antenna.

According to this configuration, during operation of the array antenna, when there is a small excitation phase error in each element of the array antenna due to aging or environmental changes of the array antenna, the excitation phase error of each element of the array antenna can be reduced. It can be estimated with high accuracy. Then, simulation can be used to generate a large amount of teacher data.

Further, in the teaching data acquisition step, the array antenna calculated using a simulation for the excitation amplitude of each element of the array antenna randomly set within a predetermined range not including 0. In the learning model generation method for an array antenna, at least one of an antenna power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern is acquired as the plurality of teacher data.

According to this configuration, the excitation amplitude of all the elements of the array antenna does not become approximately 0 in some of the teacher data, and training and verification are performed appropriately during generation of the learning model. Then, simulation can be used to generate a large amount of teacher data.

Further, in the teaching data acquisition step, the REV method or the MEP method is applied to at least one of a power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern of the array antenna. a learning model generating method for an array antenna, wherein the excitation phase or the excitation phase amplitude of each element of the array antenna measured using an array antenna is acquired as the plurality of teacher data.

According to this configuration, by using the REV method or the MEP method, the measurement results of the radiation characteristics of the array antenna, which cannot be obtained by simulation, can be included in the training data.

In this way, the present disclosure uses many measurements in estimating and correcting the initial excitation phase error and the initial excitation amplitude error of each element of the array antenna (not including the initial excitation amplitude error, only the initial excitation phase error is also possible). can be made unnecessary.

It is a figure which shows the structure of the learning model generation apparatus of this indication, and an excitation characteristic estimation apparatus. It is a figure which shows the procedure of the learning model generation process of this indication, and an excitation characteristic estimation process. It is a figure which shows the structure of the learning model of 1st Embodiment, an input, and an output. FIG. 4 is a diagram showing teacher data on the input side and the output side of the first embodiment; It is a figure which shows the parameter of the learning model of 1st Embodiment. It is a figure which shows the learning log|history (during training and verification) of 1st Embodiment. FIG. 10 is a diagram showing a modified example of teacher data on the output side of the first embodiment; It is a figure which shows the structure of the learning model of 2nd Embodiment, an input, and an output. FIG. 10 is a diagram showing teacher data on the input side and the output side of the second embodiment; It is a figure which shows the parameter of the learning model of 2nd Embodiment. It is a figure which shows the learning log|history (at the time of training and at the time of verification) of 2nd Embodiment. FIG. 10 is a diagram showing estimated excitation phases of the second embodiment and the REV method;

Embodiments of the present disclosure will be described with reference to the accompanying drawings. The embodiments described below are examples of implementing the present disclosure, and the present disclosure is not limited to the following embodiments.

(Overview of Learning Model for Array Antenna of Present Disclosure)
FIG. 1 shows the configuration of a learning model generating device and an excitation characteristic estimating device according to the present disclosure. FIG. 2 shows procedures of learning model generation processing and excitation characteristic estimation processing of the present disclosure. The learning model generation device M includes a teacher data acquisition unit 1, a neural network 2, and a learning model generation unit 3, and can be realized by installing a learning model generation program shown in steps S1 and S2 in FIG. 2 into a computer. . The excitation characteristic estimation device E includes a radiation characteristic acquisition unit 4, a neural network 2, and an excitation characteristic output unit 5, and can be realized by installing the excitation characteristic estimation program shown in steps S3 and S4 in FIG. .

Here, in the learning model generation stage, a neural network 2 is generated which receives the radiation characteristics of the array antenna and outputs the excitation characteristics of each element of the array antenna. Then, in the excitation characteristic estimation stage, the neural network 2 inputs the uncorrected radiation pattern of the array antenna and outputs the excitation characteristic of each element of the array antenna. That is, if the neural network 2 is generated in advance, the excitation characteristic of each element of the array antenna can be estimated only by measuring the radiation pattern of the array antenna before correction.

First, the learning model generation stage in the learning model generation device M will be described. The teacher data acquisition unit 1 includes the excitation phase or excitation phase amplitude of each element of the array antenna, and at least one of the power radiation pattern, the real part of the electric field radiation pattern, and the imaginary part of the electric field radiation pattern of the array antenna. A plurality of teacher data are obtained (step S1). The learning model generation unit 3 uses the plurality of teacher data acquired by the teacher data acquisition unit 1 to obtain at least one of the array antenna power radiation pattern, electric field real part radiation pattern, and electric field imaginary part radiation pattern. is input to generate a neural network 2 that outputs the excitation phase or excitation phase amplitude of each element of the array antenna (step S2).

The input layer 21 receives at least one of a power radiation pattern, a real electric field radiation pattern, and an imaginary electric field radiation pattern of the array antenna. The output layer 23 outputs the excitation phase or excitation phase amplitude of each element of the array antenna. The intermediate layer 22 has an excitation phase or an excitation phase amplitude of each element of the array antenna, and at least one of a power radiation pattern, a real electric field radiation pattern, and an imaginary electric field radiation pattern of the array antenna. Parameters (coupling coefficients between nodes) are learned using teacher data.

In this way, in estimating and correcting the initial excitation phase error and initial excitation amplitude error of each element of the array antenna (initial excitation phase error alone is also possible, not including the initial excitation amplitude error), a lot of labor and time is required for measurement. A learning model can be generated to eliminate the need for Then, most of the test time of the array antenna can be devoted to the essential characteristic evaluation of the array antenna, not to the estimation and correction of the initial excitation characteristic error of each element of the array antenna.

Next, the excitation characteristic estimation stage in the excitation characteristic estimation device E will be described. The radiation characteristic acquisition unit 4 acquires at least one of the power radiation pattern, the real part of the electric field radiation pattern, and the imaginary part of the electric field radiation pattern of the array antenna (step S3). The excitation characteristic output unit 5 outputs the power radiation pattern of the array antenna, the electric field real part radiation pattern, and the electric field acquired by the radiation characteristic acquisition unit 4 in the neural network 2 generated by the learning model generation unit 3 of the learning model generation device M. At least one of the imaginary part radiation patterns is input, and the excitation phase or excitation phase amplitude of each element of the array antenna is output (step S4).

The neural network 2 inputs at least one of the array antenna power radiation pattern, electric field real part radiation pattern, and electric field imaginary part radiation pattern to the input layer 21, and in the intermediate layer 22 parameters (couplings between nodes coefficient), and outputs the excitation phase or excitation phase amplitude of each element of the array antenna from the output layer 23 .

In this way, in estimating and correcting the initial excitation phase error and initial excitation amplitude error of each element of the array antenna (initial excitation phase error alone is also possible, not including the initial excitation amplitude error), a lot of labor and time is required for measurement. A learning model can be used to eliminate the need for Then, most of the test time of the array antenna can be devoted to the essential characteristic evaluation of the array antenna, not to the estimation and correction of the initial excitation characteristic error of each element of the array antenna.

(Specific example of learning model for array antenna of the first embodiment)
FIG. 3 shows the configuration, inputs and outputs of the learning model of the first embodiment. In the first embodiment, the neural network 2 inputs the electric field real part radiation pattern and the electric field imaginary part radiation pattern of the array antenna to the input layer 21, and performs calculations based on parameters (coupling coefficients between nodes) in the intermediate layer 22. Then, the excitation phase amplitude of each element of the array antenna is output from the output layer 23 .

FIG. 4 shows teacher data on the input side and the output side of the first embodiment. The teacher data acquisition unit 1 determines the excitation phase of each element of the array antenna randomly set within a range of 360 degrees, and the excitation phase of each element of the array antenna randomly set within a predetermined range not including 0. The electric field real part radiation pattern and the electric field imaginary part radiation pattern of the array antenna, which are calculated using simulation for the excitation amplitude of the element, are acquired as a plurality of teacher data.

ここで、θはアレーアンテナの放射方向であり（ブロードサイド方向は０°）、ｋはアレーアンテナの各素子の配列順序番号であり（第１実施形態では全８素子）、λはアレーアンテナの放射波長であり（第１実施形態では搬送波周波数は２８ＧＨｚ）、ｄはアレーアンテナの素子間隔である（第１実施形態では６ｍｍ）。そして、ｇ_ｋ（θ）はアレーアンテナの各素子の単体放射パターンであり（すべて既知の指向性関数）、Ａ_ｋはアレーアンテナの各素子の励振振幅であり、δ_ｋはアレーアンテナの各素子の励振位相である。 The electric field radiation pattern E(θ) of the array antenna is expressed by Equation (1).

Here, θ is the radiation direction of the array antenna (the broadside direction is 0°), k is the array order number of each element of the array antenna (8 elements in total in the first embodiment), and λ is the direction of the array antenna. is the radiation wavelength (the carrier frequency is 28 GHz in the first embodiment), and d is the element interval of the array antenna (6 mm in the first embodiment). Then, g _k (θ) is the single radiation pattern of each element of the array antenna (all known directivity functions), A _k is the excitation amplitude of each element of the array antenna, and δ _k is each element of the array antenna. is the excitation phase of

In the left column of FIG. 4, three types of teacher data on the output side are excitation phase δ _k of each element of the array antenna, excitation amplitude A _k (0.75≦A _k ≦1) of each element of the array antenna, and , indicates. The right column of FIG. 4 shows, as three kinds of input-side teacher data, the electric field real part radiation pattern E _r (θ) of the array antenna and the electric field imaginary part radiation pattern E _i (θ) of the array antenna.

FIG. 5 shows the parameters of the learning model of the first embodiment. The number M of teacher data was fixed at 750,000. Since the training data ratio p _train is fixed at 0.96, the number of training data is 720,000 and the number of verification data is 30,000.

Since the electric field real part radiation pattern E _r (θ) and the electric field imaginary part radiation pattern E _i (θ) of the array antenna were calculated in increments of 1° at -90° ≤ θ ≤ 90°, the number of input layer nodes N _i is , 181×2=362. The number of intermediate layers L _h was searched for 1≦L _h ≦4 and was optimized to 2. The number of intermediate layer nodes N _h was searched for 100≦N _h ≦1500, and was optimized to 700 and set. The hidden layer activation function φ _h was set fixed at 'ReLU'. Since the excitation phase _.delta.k and excitation amplitude _A.sub.k of each element of the array antenna are set for eight elements, the number of output layer nodes _N.sub.o is fixed at 8.times.2=16. The output layer activation function φ _о was set fixed at 'sigmоid'.

The batch size b was optimized and set to 8192 after being explored at 16, 64, 256, . The dropout rate p was set fixed at zero. Among 'SGD', 'RMSprop' and 'Adam', 'Adam' was adopted as the optimizer. Training, validation and estimation were then performed.

FIG. 6 shows the learning history (at the time of training and at the time of verification) of the first embodiment. The normalization error Loss is the root mean square error calculated for the normalized amplitude and normalized phase that take values from 0 to 1 and are calculated as estimated values at the 16 output nodes. The normalized error Loss during training decreased to 10 ^-4 at the learning times Epoch ~1000. The normalization error Loss at the time of verification decreases to 10 ^-3 at the learning times Epoch ~ 100, and approaches the normalization error Loss at the time of training asymptotically, indicating no overfitting.

ここで、δ_ｋ（ハット‘＾’記号なし）は、検証データにおける励振位相設定値であり、δ_ｋ（ハット‘＾’記号あり）は、検証段階における励振位相推定値である。そして、Ａ_ｋ（ハット‘＾’記号なし）は、検証データにおける励振振幅設定値であり、Ａ_ｋ（ハット‘＾’記号あり）は、検証段階における励振振幅推定値である。なお、励振位相推定精度δ_ＲＭＳＥにおいて、励振位相（０°～３６０°）の循環性を考慮している。 The excitation phase estimation accuracy δ _RMSE and the excitation amplitude estimation accuracy A _RMSE are expressed by Equation (2).

where δ _k (without the hat '^' symbol) is the excitation phase setting in the verification data and δ _k (with the hat '^' symbol) is the excitation phase estimate in the verification stage. Then A _k (without the hat '^' symbol) is the excitation amplitude setting in the verification data and A _k (with the hat '^' symbol) is the excitation amplitude estimate in the verification stage. Note that the excitation phase estimation accuracy δ _RMSE considers the cyclicity of the excitation phase (0° to 360°).

The excitation phase estimation accuracy δ _RMSE was 1.19°. On the other hand, in the analog beamformer, the number of bits of the phase shifter is at most 6 bits, and the resolution of the phase shifter is at most 5.625°. That is, the excitation phase estimation accuracy δ _RMSE exceeded the resolution of the phase shifter. The excitation amplitude estimation accuracy A _RMSE was 0.27%. On the other hand, in analog beamformers, the resolution of the gain controller is at most 0.5 dB, and the finest adjustment of the gain controller is at most 0.8% within its dynamic range. Therefore, the excitation amplitude estimation accuracy A _RMSE exceeded the finest adjustment amount of the gain controller.

In this way, when there is a large excitation phase error in each element of the array antenna due to the hardware factors of each element of the array antenna during manufacture of the array antenna, the excitation phase error of each element of the array antenna can be reduced to some extent. It can be estimated with accuracy (see the method of setting the excitation phase δ _k in the upper left column of FIG. 4). Then, in some teacher data, the excitation amplitude of all the elements of the array antenna does not become almost 0, and training and verification are appropriately executed when generating the learning model (the excitation amplitude in the lower left column of FIG. 4) See how to set the amplitude A _k ). Furthermore, simulations can be used to generate large amounts of training data.

FIG. 7 shows a modified example of teacher data on the output side of the first embodiment. The teacher data acquisition unit 1 randomly sets the excitation phase of each element of the array antenna within a predetermined range centering on the excitation phase that varies linearly with respect to the arrangement order number of each element of the array antenna. It is also possible to acquire the electric field real part radiation pattern and the electric field imaginary part radiation pattern of the array antenna, which are calculated using simulation, as a plurality of teacher data.

In the left column of FIG. 7, the linearly increasing excitation phase δ _k is shown by a solid line with respect to the arrangement order number k of each element of the array antenna, and within a predetermined range centered on this solid line (for example, aging or The randomly set excitation phase δ _k is shown by dashed and dashed lines. In the middle column of FIG. 7 , the linearly decreasing excitation phase δ _k is indicated by a solid line with respect to the arrangement order number k of each element of the array antenna, and within a predetermined range centered on this solid line (for example, aging or The randomly set excitation phase δ _k is shown by dashed and dashed lines. In the right column of FIG. 7, the constant excitation phase δ _k is shown by a solid line with respect to the arrangement order number k of each element of the array antenna, and within a predetermined range centered on this solid line (for example, aging or environmental change ) and randomly set excitation phases δ _k are shown by dashed and dashed lines.

Then, during operation of the array antenna, when there is a small excitation phase error in each element of the array antenna due to aging or environmental changes of the array antenna, the excitation phase error of each element of the array antenna can be estimated with higher accuracy. can do. Then, simulation can be used to generate a large amount of training data. Note that the teacher data shown in FIG. 7 can be applied not only to the first embodiment but also to the second embodiment.

(Specific example of learning model for array antenna of the second embodiment)
FIG. 8 shows the configuration, inputs and outputs of the learning model of the second embodiment. In the second embodiment, the neural network 2 inputs the power radiation pattern (the square of the electric field radiation pattern) of the array antenna to the input layer 21, and performs computation based on the parameters (coupling coefficients between nodes) in the intermediate layer 22. Then, the excitation phase of each element of the array antenna is output from the output layer 23 .

FIG. 9 shows teacher data on the input side and the output side of the second embodiment. The teacher data acquisition unit 1 calculates the power radiation of the array antenna calculated using simulation for the excitation phase (relative excitation phase) of each element of the array antenna set randomly within a range of 360 degrees. A pattern (the square of the electric field radiation pattern) is acquired as a plurality of teacher data.

ここで、θはアレーアンテナの放射方向であり（ブロードサイド方向は０°）、ｋはアレーアンテナの各素子の配列順序番号であり（第２実施形態では全８素子）、λはアレーアンテナの放射波長であり（第２実施形態では搬送波周波数は２８ＧＨｚ）、ｄはアレーアンテナの素子間隔である（第２実施形態では６ｍｍ）。そして、ｇ_ｋ（θ）はアレーアンテナの各素子の単体放射パターンであり（すべて既知の指向性関数）、Ａはアレーアンテナの各素子の一定励振振幅であり、δ_ｋはアレーアンテナの各素子の励振位相である。 A power radiation pattern P(θ) of the array antenna is expressed by Equation (3).

Here, θ is the radiation direction of the array antenna (broadside direction is 0°), k is the arrangement order number of each element of the array antenna (8 elements in total in the second embodiment), and λ is the direction of the array antenna. is the radiation wavelength (the carrier frequency is 28 GHz in the second embodiment), and d is the element spacing of the array antenna (6 mm in the second embodiment). Then g _k (θ) is the single radiation pattern of each element of the array antenna (all known directivity functions), A is the constant excitation amplitude of each element of the array antenna, and δ _k is each element of the array antenna. is the excitation phase of

The left column of FIG. 9 shows the excitation phase δ _k (δ ₁ is fixed at 0°, the others are relative values to δ ₁ ) of each element of the array antenna as three types of teacher data on the output side. The right column of FIG. 9 shows the power radiation pattern P(θ) of the array antenna as the three types of training data on the input side.

FIG. 10 shows the parameters of the learning model of the second embodiment. The number M of teacher data was fixed at 750,000. Since the training data ratio p _train is fixed at 0.96, the number of training data is 720,000 and the number of verification data is 30,000.

The power radiation pattern P(θ) of the array antenna was calculated in increments of 1° for −90°≦θ≦90°, so the number of input layer nodes _Ni was fixed at 181. The number of hidden layers L _h was searched for 2≦L _h ≦8 and was optimized to 6. The number of intermediate layer nodes N _h was searched for 200≦N _h ≦1500, and was optimized to 900 and set. The hidden layer activation function φ _h was set fixed at 'ReLU'. Since the relative value of the excitation phase _δk of each element of the array antenna is set for seven elements, the number of output layer nodes N _о is fixed at seven. The output layer activation function φ _о was set fixed at 'sigmоid'.

The batch size b was optimized and set to 16384 after being explored at 16, 64, 256, ..., 8192, 16384. The dropout rate p was set fixed at zero. Among 'SGD', 'RMSprop' and 'Adam', 'Adam' was adopted as the optimizer. Training, validation and estimation were then performed.

FIG. 11 shows the learning history (during training and verification) of the second embodiment. The normalization error Loss is the root mean square error calculated for the normalization phase that takes a value from 0 to 1 and is calculated as an estimated value at the seven output nodes. The normalized error Loss during training decreased to 10 ^-3 at the learning times Epoch-100. The normalization error Loss at the time of verification decreases to 10 ⁻² at the learning times Epoch˜100, and approaches the normalization error Loss at the time of training asymptotically, indicating no overfitting. However, since the optimizer 'Adam' was adopted, the normalization error Loss during training and verification sometimes increased in a spike-like manner at a certain number of times of learning Epoch.

ここで、δ_ｋ（ハット‘＾’記号なし）は、検証データにおける励振位相設定値であり、δ_ｋ（ハット‘＾’記号あり）は、検証段階における励振位相推定値である。なお、励振位相推定精度δ_ＲＭＳＥにおいて、励振位相（０°～３６０°）の循環性を考慮している。また、励振位相δ_１は０°に固定され、他の励振位相δ_ｋはδ_１に対する相対値である。 The excitation phase estimation accuracy δ _RMSE is expressed by Equation (4).

where δ _k (without the hat '^' symbol) is the excitation phase setting in the verification data and δ _k (with the hat '^' symbol) is the excitation phase estimate in the verification stage. Note that the excitation phase estimation accuracy δ _RMSE considers the cyclicity of the excitation phase (0° to 360°). Also, the excitation phase _.delta.1 is fixed at 0.degree., and the other excitation phases _.delta.k are relative values to _.delta.1 .

The excitation phase estimation accuracy δ _RMSE was 7.45°. On the other hand, in the analog beamformer, the number of bits of the phase shifter is at most 6 bits, and the resolution of the phase shifter is at most 5.625°. That is, the excitation phase estimation accuracy δ _RMSE is a measure of the resolution of the phase shifter.

Thus, the excitation phase error of each element of the array antenna can be estimated by inputting the power radiation pattern P(θ) of the array antenna, which can be easily measured. When manufacturing the array antenna, when there is a large excitation phase error of each element of the array antenna due to hardware factors of each element of the array antenna, the excitation phase error of each element of the array antenna can be corrected with a certain degree of accuracy. can be estimated (see the method of setting the excitation phase δ _k in the left column of FIG. 9). Furthermore, simulations can be used to generate large amounts of training data.

FIG. 12 shows the estimated excitation phases of the second embodiment and the REV method. In the left column of FIG. 12, the estimated excitation phase δ _k ( δ ₁ is fixed at 0°) is indicated by a solid line, and the estimated excitation phase δ _k (δ ₁ is fixed at 0°) by the REV method is indicated by a broken line. In the right column of FIG. 12, the calculation result of the power radiation pattern P(θ) of the array antenna of the second embodiment (based on the estimated excitation phase δ _k and Equation 3) is shown by a solid line, and the power of the REV method array antenna A dashed line indicates the calculation result of the radiation pattern P(θ) (based on the estimated excitation phase δ _k and Equation 3).

Then, compared with the calculation result of the power radiation pattern P(θ) of the array antenna of the second embodiment, the calculation result of the power radiation pattern P(θ) of the array antenna by the REV method shows the uncorrected power of the array antenna. It was close to the measurement result of the radiation pattern P(θ). And the excitation phase estimation error δ _RMSE in the second embodiment compared to the REV method was 11.9°. However, even with the calculation result of the power radiation pattern P(θ) of the array antenna of the second embodiment, the measurement result of the pre-correction power radiation pattern P(θ) of the array antenna was sufficiently reproduced.

Considering the degree of agreement between the measurement results of FIG. 12 and the second embodiment, the teacher data acquiring unit 1 calculates the power radiation pattern of the array antenna by using the REV method or the MEP method. The excitation phase of the element can also be acquired as a plurality of teaching data.

Then, by using the REV method or the MEP method, the measurement results of the radiation characteristics of the array antenna, which cannot be obtained by simulation, can be included in the training data. The teacher data shown here can be applied not only to the second embodiment but also to the first embodiment.

The array antenna learning model generation method, learning model, excitation characteristic estimation method, and excitation characteristic estimation program of the present disclosure are analog beamformers (the initial correction of the excitation characteristic error is not reset, and the electric field radiation pattern can also be measured. can be applied) as well as digital beamformers (where the initial correction of the excitation characteristic error may be reset and only the power radiation pattern can be measured).

M: learning model generation device E: excitation characteristic estimation device 1: teacher data acquisition unit 2: neural network 3: learning model generation unit 4: radiation characteristic acquisition unit 5: excitation characteristic output unit 21: input layer 22: intermediate layer 23: output layer

Claims (8)

Acquiring a plurality of teacher data including an excitation phase or an excitation phase amplitude of each element of the array antenna and at least one of a power radiation pattern, an electric field real part radiation pattern and an electric field imaginary part radiation pattern of the array antenna a training data acquisition step for
inputting at least one of a power radiation pattern, a real electric field radiation pattern, and an imaginary electric field radiation pattern of the array antenna using the plurality of training data obtained in the training data obtaining step; a learning model generation step of generating a learning model that outputs the excitation phase or excitation phase amplitude of each element of the array antenna;
A method for generating a learning model for an array antenna, comprising: In the training data acquisition step, the power radiation pattern and the electric field real part of the array antenna calculated using a simulation for the excitation phase of each element of the array antenna set randomly within a range of 360 degrees. 2. The learning model generation method for an array antenna according to claim 1, wherein at least one of a radiation pattern and an electric field imaginary part radiation pattern is acquired as said plurality of teacher data. In the training data acquisition step, excitation of each element of the array antenna is randomly set within a predetermined range centered on an excitation phase that varies linearly with respect to the arrangement order number of each element of the array antenna. At least one of a power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern of the array antenna, which are calculated for the phase using a simulation, is obtained as the plurality of teacher data. 2. The method of generating a learning model for an array antenna according to claim 1, characterized by: In the training data acquisition step, the power radiation pattern of the array antenna calculated using a simulation with respect to the excitation amplitude of each element of the array antenna set randomly within a predetermined range not including 0; 4. The learning of the array antenna according to claim 1, wherein at least one of an electric field real part radiation pattern and an electric field imaginary part radiation pattern is acquired as the plurality of teacher data. Model generation method. In the training data acquisition step, for at least one of a power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern of the array antenna, REV method (Rotating element Electric-field Vector method) or 2. An excitation phase or an excitation phase amplitude of each element of the array antenna measured using an MEP method (Multi-Element Phase-toggle method) is acquired as the plurality of teacher data. 5. A learning model generation method for an array antenna according to any one of 4 to 4. an input layer to which at least one of a power radiation pattern, a real part of the electric field radiation pattern and an imaginary part of the electric field radiation pattern of the array antenna is input;
an output layer that outputs the excitation phase or excitation phase amplitude of each element of the array antenna;
a plurality of teacher data comprising an excitation phase or an excitation phase amplitude of each element of the array antenna and at least one of a power radiation pattern, an electric field real part radiation pattern and an electric field imaginary part radiation pattern of the array antenna; an intermediate layer whose parameters are learned using
and inputting at least one of a power radiation pattern, a real electric field radiation pattern and an imaginary electric field radiation pattern of the array antenna into the input layer, calculating in the intermediate layer, and A learning model of an array antenna for causing a computer to output the excitation phase or excitation phase amplitude of each element from the output layer. a radiation characteristic acquiring step of acquiring at least one of a power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern of the array antenna;
a plurality of teacher data comprising an excitation phase or an excitation phase amplitude of each element of the array antenna and at least one of a power radiation pattern, an electric field real part radiation pattern and an electric field imaginary part radiation pattern of the array antenna; inputting at least one of the power radiation pattern, the electric field real part radiation pattern, and the electric field imaginary part radiation pattern of the array antenna acquired in the radiation characteristic acquiring step into the learning model generated using the an excitation characteristic estimation step of outputting an excitation phase or an excitation phase amplitude of each element of the array antenna;
A method for estimating excitation characteristics of an array antenna, comprising: a radiation characteristic acquiring step of acquiring at least one of a power radiation pattern, an electric field real part radiation pattern, and an electric field imaginary part radiation pattern of the array antenna;
a plurality of teacher data comprising an excitation phase or an excitation phase amplitude of each element of the array antenna and at least one of a power radiation pattern, an electric field real part radiation pattern and an electric field imaginary part radiation pattern of the array antenna; inputting at least one of the power radiation pattern, the electric field real part radiation pattern, and the electric field imaginary part radiation pattern of the array antenna acquired in the radiation characteristic acquiring step into the learning model generated using the an excitation characteristic estimation step of outputting an excitation phase or an excitation phase amplitude of each element of the array antenna;
A program for estimating the excitation characteristics of an array antenna for causing a computer to execute in order.

Images