Method for forming antenna array radiation pattern using artificial neural network

Abstract

FIELD: antenna technology.SUBSTANCE: invention relates to methods for forming radiation patterns of phased antenna arrays in the event of failure of antenna elements. A method is proposed for generating an antenna array radiation pattern using an artificial neural network with an assessment of the current state of the antenna elements of the antenna array and the identification of failures, the synthesis of a new amplitude distribution for the antenna array based on information about the failures of the antenna elements and the values of the required radiation pattern using a trained artificial neural network based on a variational autoencoder.EFFECT: synthesis of a radiation pattern in the event of failures of antenna elements to optimal parameter values and the implementation of training of an artificial neural network based on a variational autoencoder without a teacher.1 cl, 8 dwg, 1 tbl

Description

[0001] The invention relates to antenna technology, in particular to methods for forming radiation patterns (RP) of antenna arrays.

[0002] A method for selecting the amplitude distribution (AD) and amplitude-phase distribution (APD) for a line of emitters is known, described in [1-Iye T., Susukida Y., Takaya S., Sugiura T., Fujii Y. A Deep Learning Based Antenna Array Calibration Method Using Radiation Power Pattern / / 2022 IEEE 33 rd [Annual International Symposium on Personal, Indoor and Mobile Radio Communications, September 2022], according to which a training sample is formed with random values ​​of amplitude and phase in a given range, the ANN is trained using the training sample, the antenna pattern values ​​are measured, the measured antenna pattern values ​​are fed to the ANN input, and the APF values ​​corresponding to the measured pattern are formed at the ANN output.

[0003] This method has the following disadvantages:

[0004] - the training sample does not cover the RP with AE failures, therefore, there is no possibility of compensating them to reduce the side lobe level (SL).

[0005] - with an increase in the number of AEs of the antenna array and the corresponding scaling of the fully connected ANN, the complexity of training such an ANN when selecting the amplitude-phase distribution increases significantly;

[0006] - the used INS provides for work only with the main plane of the DN.

[0007] A method of matrix amplitude-phase synthesis of a flat phased antenna array is known [2 - Litvinov A.V., Mishchenko S.E., Pomysov A.S., Shchatsky V.V., Atrokhov V.V., Shelkoplyasov S.A. Method of synthesis of an antenna array with a reduced level of side lobes in the presence of channel and module failures / / Journal of Radio Electronics. 2023. No. 9], which is reduced to the formation of a failure vector and an iterative process, at each step of which the weighting function is adjusted to calculate the covariance matrix of the antenna element radiation patterns, taking into account the requirements for the side lobe envelope and the failure vector, and then the problem of determining the amplitude-phase distribution is solved by multiplying the phasing vector of the antenna array of serviceable channels and the inverse covariance matrix.

[0008] This method cannot be applied in real time due to high computational costs.

[0009] The closest in technical essence, selected as a prototype, is the method for adapting the radiation pattern of an active phased antenna array [3-Patent of the Russian Federation No. 2837529, IPC: H01Q 3 / 00], in accordance with which a training sample is formed based on historical data on the operation of the antenna array, including data on various module failures and corresponding adjustments to the APS, a radial basic ANN is trained using the formed training sample, the current states of the AE of the antenna array are determined and AE failures are identified, information on the current APS is collected, a new APS is synthesized that compensates for the failed AEs using the trained radial basic ANN, and the synthesized APS is applied to the antenna array.

[0010] This method has the following disadvantages:

[0011] - restoration of the DN taking into account the failed elements occurs only in the main planes of the DN;

[0012] - ANN training occurs with a teacher, which requires significant computational costs for the formation of a training sample, including data on various module failures and corresponding adjustments to the AFR;

[0013] - the efficiency of the radial basis ANN decreases with the increase in the dimensionality of the input data space, therefore, the efficiency of the ARF correction decreases with the increase in the number of AEs of the antenna array.

[0014] The technical problem that the proposed invention is aimed at solving is the formation of the antenna array pattern taking into account the failures of the AE.

[0015] To solve the specified technical problem, a method is proposed for forming the antenna array directional pattern using an artificial neural network, according to which a training sample is formed, the current states of the antenna array AE are determined, and AE failures are identified.

[0016] According to the invention, a training sample is formed based on various synthesized antenna arrays and their corresponding RPs, an ANN is trained based on a modified variational autoencoder using the formed training sample without a teacher, after determining the current state of the AE of the antenna array and identifying failures, the values ​​of the target RP are formed and stored, a new antenna array is synthesized that corresponds to the target RP and compensates for the AE failures using the trained ANN based on the modified variational autoencoder, and the synthesized antenna array is applied.

[0017] The technical result is the synthesis of a DN for AE failures up to optimal parameter values ​​and the implementation of ANN training based on a variational autoencoder without a teacher.

[0018] Thus, the proposed method has the following distinctive features and the sequence of its implementation from the prototype method, which are given in Table 1.

[0019]

[0020] A comparative analysis of the claimed method and the prototype method shows that the claimed method differs in that the set of actions has been changed, namely, 4 new actions have been introduced:

[0021] - train the ANN based on a modified variational autoencoder using the generated training sample without a teacher;

[0022] - generate and store the target DN values;

[0023] - synthesize a new AR that corresponds to the target DN and compensates for AE failures using a trained ANN based on a modified variational autoencoder;

[0024] - apply synthesized AR to the antenna array; and also change the modes of execution of one operation:

[0025] - form a training sample based on various synthesized AR and the corresponding DN.

[0026] The proposed invention is not known from the prior art, and there are no known sources of information containing information on similar technical solutions that have features similar to the features that distinguish the claimed solution from the prototype, as well as properties that coincide with the properties of the claimed solution, therefore it can be considered that it has significant differences, follows from them in a non-obvious way and, therefore, meets the criteria of “novelty” and “inventive step”.

[0027] The essence of the proposed method is disclosed by figures 1-8.

[0028] Fig. 1 shows a structural diagram of a device that implements the proposed method for an antenna array.

[0029] Fig. 2. An example of a target AR.

[0030] Fig. 3 shows an example of the target DN.

[0031] Fig. 4. An example of an AE failure mask.

[0032] Fig. 5 shows an example of a DN taking into account the AE failure mask.

[0033] Fig. 6 shows an example of AR at the output of the INS decoder.

[0034] Fig. 7 shows an example of a synthesized DN using an ANN.

[0035] Fig. 8 shows a comparison of the average UBL in the DN sections for the entire test sample.

[0036] When implementing the method of forming the antenna array radiation pattern using an artificial neural network, the following sequence of actions is performed:

[0037] - form a training sample based on various synthesized AR and the corresponding DN;

[0038] - train the ANN based on a modified variational autoencoder using the generated training sample without a teacher;

[0039] - determine the current states of the antenna array AE and identify AE failures;

[0040] - generate and store the target DN values;

[0041] - synthesize a new AR that corresponds to the target DN and compensates for AE failures using a trained ANN based on a modified variational autoencoder;

[0042] - apply the synthesized AR to the antenna array.

[0043] A simplified structural diagram of a device that allows the implementation of a method for forming an antenna array radiation pattern using an artificial neural network includes N AE 11…1 N , N isolating devices (ID) 21…2 N , N transmitting devices (TU) 31…3 N , TV radio receiving devices (RPU) 41…4 N, digital processing unit (DPU) 5, signal generation unit (SG) 6, electronic computer (EC) 7 and neural network processing unit (NNPU) 8. NNPU 8 consists of encoder 8.1, encoder 8.2, adder 8.3, decoder 8.4, unit for calculating the radiation pattern based on the amplitude distribution (UPA) 8.5, which is one fully connected layer of the ANN with a matrix of constant (non-trainable) weights w={Re(Z) Im(Z)}, of size N×2K, where N is the number of antenna elements, K is the number of angular directions when calculating the RP; in the form of a block matrix composed of the real and imaginary parts of the complex-valued guidance matrix Z, determined by formula 1.

[0044]

[0045] where Z n,k - the phase value of the excitation current of the n-th antenna element corresponding to the k-th angular direction, λ is the wavelength, x n and y n -coordinates of the antenna element, ϕ k and θ k- azimuth and elevation angle of the corresponding angular direction.

[0046] AE outputs 11… 1 N connected to the corresponding first inputs of RU 21…2 N . Second inputs RU 21…2 N connected to the outputs of PU 31…3 N Outputs RU 21…2 N connected to the corresponding inputs of the RPU 41…4 N Outputs of RPU 41…4 N connected to inputs 1-N of the digital control unit 5. The output of the digital control unit 5 is connected to the first input of the computer 7. The first output of the computer 7 is connected to the input of the digital control unit 6. Outputs 1-N of the digital control unit 6 are connected to the corresponding inputs of the control units 31…3 N . The second output of the computer 7 is connected to the encoder 8.2. The third output of the computer 7 is connected to the encoder 8.1. The outputs of the encoder 8.1 and the encoder 8.2 are connected to the corresponding first and second inputs of the adder 8.3. The output of the adder 8.3 is connected to the input of the decoder 8.4. The first output of the decoder 8.4 is connected to the second input of the computer 7. The second output of the decoder 8.4 is connected to the input of the BRAR 8.5. The output of the BRAR 8.5 is connected to the third input of the computer 7.

[0047] Let us consider the implementation of the method for forming the directional pattern of an antenna array using an artificial neural network using the device shown in Fig. 1.

[0048] Before starting work, it is necessary to form a training sample and train the ANN. To form a training sample, a set of amplitude distributions was selected in the form of vectors corresponding to the distribution in the vertical (W y ) and horizontal (W x ) planes. In the numerical experiments conducted, this set contained nonparametric window functions (Bartlett, Blackman, Hamming, Hann, etc.) and parametric ones (Tukey, Taylor, Kaiser, Gauss, Chebyshev windows) with different parameters. The amplitude distribution in the entire antenna array A was then calculated as the outer product of each W x with each W yOne example of such an antenna array is shown in Figure 2. In the second step, the pattern was calculated for each antenna array using formula 2 on a grid of angles [-90°, 90°] with a step of 1°. The step of the angular grid was chosen such that it did not exceed half the width of the main lobe of the pattern. An example of the calculated pattern is shown in Figure 3.

[0049]

[0050] where Z n,k - the phase value of the excitation current of the n-th AE corresponding to the k-th angular direction; M n AE failure mask (1 - active AE, 0 - failed AE); A n - the amplitude of the excitation current of the n-th AE.

[0051] The AE failure mask was generated not at the training set creation stage, but directly during the loading of input data during ANN training. Each time the input data readout function was called during training, a matrix with a size identical to the amplitude distribution matrix was randomly generated, filled with the values ​​0 (failed) and 1 (active). The probability of an AE failure in the matrix was defined by a random variable with a uniform distribution ranging from 0% to 15% of the total number of channels. An example of an AE and DN failure mask with failed AEs is shown in Figures 4 and 5, respectively.

[0052] The modification of the ANN based on the variational autoencoder means the addition of Encoder 8.2 followed by the summation operation at the outputs of Encoder 8.1 and Encoder 8.2. Encoder 8.1 encodes the target DN into the vector X (1)latent space of dimension L, Encoder 8.2 encodes information about AE failures, presented in the form of a binary matrix (0 - failed and 1 - active) in the vector X (2) the same space. The vector encoding the information belongs to the so-called latent space. The latent space is a manifold in which similar objects have close coordinates. Vectors X (1) and X (2) are summed up and the sum is fed to the input of the decoder, which generates an amplitude distribution based on complete information about the requirements for the DN and about the AE failures.

[0053] During ANN training, the decoder output data is fed to the BRAR to calculate the error function, and during ANN operation, the data is directly used as the AR. When training an ANN based on a variational autoencoder, two different error functions are used. To check the correspondence between the target pattern and the pattern synthesized by the ANN, the standard deviation is used:

[0054]

[0055] where T k - target DN;

[0056] Y k - DN obtained at the output of the INS;

[0057] K - the number of angular directions when calculating the DN.

[0058] An additional error function is also used in the form of the Kullback-Leibler divergence (KL divergence), which is a measure of the distance between two probability distributions P and Q. Calculating the KL divergence is used during ANN training to ensure that the distribution of each vector coordinate in the latent space converges to a priori distribution, which is normally chosen. The KL divergence is calculated using the following formula:

[0059]

[0060] where P i - the probability of event i in distribution P;

[0061] Qi - the probability of event i in distribution Q;

[0062] K - the number of angular directions when calculating the DN.

[0063] Training is carried out iteratively until the error function values ​​reach a specified minimum level corresponding to the required accuracy of the model.

[0064] After training the ANN based on the variational autoencoder, using computer 7, the current states of the AE of the antenna array are determined from the output of the digital encoding unit 5 and information on AE failures is read in the form of a binary failure matrix (failure mask) and pre-set requirements for the shape of the pattern in the reception and transmission modes. The values ​​of the target pattern are fed to the input of encoder 8.1, and the values ​​of the AE failure mask are fed to the input of encoder 8.2, where these values ​​are encoded into latent space vectors. The obtained vectors are transmitted to the first and second inputs of adder 8.3 and the resulting vector of values ​​is transmitted to decoder 8.4, where the decoding of the received vector of values ​​into the antenna array takes into account the AE failures. The new synthesized antenna array corresponding to the target pattern and compensating AE failures are transmitted from the first output of decoder 8.4 to the second input of computer 7. The values ​​of the AE amplitudes l1…1 N are transmitted to the input of the UFS 6 from the computer 7 together with commands for controlling the frequency and phase of the emitted signals. Using PU 31…3 Nperform frequency upshifting, filtering and amplification of emitted signals with amplitudes corresponding to the values ​​at outputs 1-N of UFS 6. Signals from outputs of PU 31…3 N are fed to the second inputs of RU 21…2 N , designed to switch the receiving and transmitting paths with a single antenna system and to isolate the transmitted and received signals. In parallel with the emission of signals via AE 11…1 N receive signals directed at AE 11… 1 N , using RPU 41…4 N Using RPU 41…4 N, carry out filtering, amplification, frequency down-conversion and separation of quadrature components of the signals received from each antenna element. An example of the calculated AR at the output of the INS decoder and the corresponding RP are shown in Figures 6 and 7, respectively. Next, the received signals go to the corresponding inputs 1-N of the digital audio system 5, where analog-to-digital conversion is performed. In digital form, the received signals are fed to the input of the computer 7, where digital formation of the RP is performed by means of weighted summation with the amplitude values ​​obtained at the first output of the decoder 8.4 at the start of operation. In the event that a signal from one of the AEs 11… 1 N stops arriving at the input of the UDC 5, the corresponding AE is considered to have failed, the binary matrix of the AE is updated and the procedure for selecting the AR corresponding to the DN in the reception and transmission modes is repeated.

[0065] In accordance with the proposed method, a numerical analysis was carried out implementing the method for generating the antenna array radiation pattern using an artificial neural network in the MATLAB programming environment. As an example, an antenna array measuring 40 by 40 AEs was selected, located at the nodes of a rectangular grid with a step of 0.5 λ. The training sample was formed in the form of target RP values ​​corresponding to the antenna array in the form of distribution vectors in the vertical (W y ) and horizontal (W x) planes. The distributions contained nonparametric window functions (Bartlett, Blackman, Hamming, Hann, etc. windows) and parametric ones (Tukey, Taylor, Kaiser, Gauss, Chebyshev windows) with different parameters. A variety of window functions was required to ensure a variety of patterns in terms of sidelobe level, gain, and main lobe width. For simplicity of calculations, we assume the pattern of a single active element in the antenna array to be isotropic. The antenna array pattern in the entire antenna array A was then calculated as the outer product of each W x with each W y In this case, the DN will be a complex vector F of length K on a grid of angles [-90°, 90°] with a step of 1°, calculated using the following formula 2.

[0066] The angular grid pitch was chosen so that it did not exceed half the width of the main lobe of the pattern. As a result, the vector size D was 91 2=8281. A total of 603 distribution vectors of length 40 were obtained for each plane. The sizes of the training and test samples were 358609 and 5000 DN in complex form, respectively.

[0067] To feed the ANN input, the 8281-length D vector modules were resized to 91×91, converted to dB, and uniformly scaled to ensure maximum contrast in the input images. To calculate the error function at the ANN output, the complex D vectors were not resized, but the real and imaginary parts were written separately to match the T value. l×16562 in the formula for the error value L according to formula 3.

[0068] Now let us consider the synthesis results averaged over the entire test set. The comparison results are presented in Figure 8. The solid blue, dashed red, and dotted black lines indicate the SLL values ​​as a function of the pattern cross-section orientation angle y, averaged over 5000 samples from the test set for the target pattern, the pattern with AE failures, and the pattern synthesized using a trained ANN based on a variational autoencoder, respectively. Figure 8 shows that the use of the ANN reduces the average SLL across all pattern cross-sections, and therefore across the entire angular sector. In the principal planes, which correspond to cross-section orientation angles of γ=0° and γ=90°, a 0.5 dB reduction in the average sidelobe level is observed.

[0069] The greatest reduction in UBL is observed for section orientation angles γ ⊂ [40°, 50°] and is approximately 2 dB.

[0070] Thus, the patented method for forming the antenna array radiation pattern using an artificial neural network is practically feasible and ensures the stated technical result - the synthesis of the antenna pattern during AE failures to optimal parameter values ​​and the implementation of ANN training based on a variational autoencoder without a teacher.

Claims

A method for generating an antenna array radiation pattern using an artificial neural network, which involves generating a training sample, determining the current states of the antenna elements of the antenna array, and identifying antenna element failures, characterized in that the training sample is generated based on various synthesized amplitude distributions and the radiation patterns corresponding thereto, modifying the artificial neural network based on a variational autoencoder by adding a second encoder to the architecture of the artificial neural network, an operation of summing the vectors at the outputs of the first and second encoders, and a block for calculating the amplitude distribution at the decoder output, training the artificial neural network based on a variational autoencoder using the generated training sample without a teacher, and, after determining the current state of the antenna elements of the antenna array and identifying failures, generating and storing the values ​​of the target radiation pattern,synthesize a new amplitude distribution corresponding to the target radiation pattern and compensating for failures of antenna elements using a trained artificial neural network based on a variational autoencoder, apply the synthesized amplitude distribution to the antenna array.

Images