A Diagnostic Method for Phaseless Array Antennas Based on Artificial Neural Networks and Compressed Sensing

Abstract

This invention provides a diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing, relating to the field of antenna array technology. The method includes: calculating the far-field pattern of the array antenna based on the extracted active element radiation pattern; simulating the far-field pattern to obtain a pre-trained new decoder neural network; measuring the amplitude of the damaged radiation pattern of the actual array to obtain the predicted excitation of the damaged elements; using the predicted excitation of the damaged elements as input to the pre-trained new decoder neural network to obtain the excitation of the virtual array; and confirming the position index of the damaged elements and the original excitation of the phaseless array antenna based on the excitation of the virtual array. The advantages of this invention are that it can determine the location of damaged elements and accurately recover the phase of the excitation source from phaseless measurements. Compared with existing technologies, this invention has higher accuracy, efficiency, and robustness.

Description

Technical Field

[0001] This invention relates to the field of antenna array technology, and more specifically, to a diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing. Background Technology

[0002] Active array antennas are widely used in radar and wireless communications due to their advantages such as high gain, low sidelobe levels, and beam scanning and beamforming capabilities. However, due to prolonged operation and exposure to outdoor environments, antenna elements and connected T / R components are prone to failure. Such failures lead to distortion of the array's far-field pattern, resulting in increased sidelobe levels and nulls, reduced directivity, compromised beamforming algorithms, and decreased anti-interference capabilities. Therefore, identifying the location of damaged elements through the array's far-field response is crucial for maintaining the performance of phased array antenna systems.

[0003] Phase-free diagnostic techniques have been effectively integrated into compressed sensing frameworks, enabling the retrieval of amplitude and phase information from pure amplitude samples generated by electromagnetic sources, while reducing the number of sampling points to some extent. Compared to traditional amplitude-phase diagnostic techniques, these methods simplify the measurement process and theoretically enhance anti-interference capabilities. Morabito et al. proposed a convex approximation diagnostic method to address the non-convexity introduced by phase-free measurements, thereby achieving phase-free diagnostics. Furthermore, Fuchs et al. proposed a phase lifting method, which lifts the excitation vector used for recovery into a semi-definite Hermitian matrix and uses trace minimization instead of Hermitian matrix rank minimization, thus transforming the problem into a convex form. However, these approximation methods may significantly reduce diagnostic efficiency and compromise the robustness of the algorithm.

[0004] In recent years, methods based on ANNs and intrinsic mode currents (ANN-EC) have been used to reconstruct sources and accurately identify faulty cells using near-field phase-free sampling data. However, these methods require large amounts of measurement or datasets, thus reducing the efficiency of the training process. Therefore, establishing an artificial neural network framework more suitable for diagnosis and minimizing the number of samples required is of great significance for improving diagnostic accuracy and efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0006] This application provides a diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing, including:

[0007] Based on the extracted active element radiation pattern of the array antenna, the far-field radiation pattern of the array antenna is calculated.

[0008] The far-field radiation pattern of the array antenna is simulated to establish a training set for the decoder neural network. The decoder neural network is then pre-trained using the established training set to obtain a new pre-trained decoder neural network.

[0009] The amplitude of the damaged pattern of the actual array is measured and input into the encoder neural network for training to obtain the predicted excitation of the damaged cells.

[0010] The predicted damaged unit excitation is used as the input of the new decoder neural network after pre-training. The encoder-decoder neural network is cascaded and trained on the cascaded encoder-decoder neural network based on the loss function preset in the compressed sensing framework until training is completed, so as to obtain the excitation of the virtual array.

[0011] Based on the excitation of the virtual array, the location index of the damaged element of the phaseless array antenna and the original excitation are identified.

[0012] Preferably, the relationship between the active element radiation pattern of the array antenna and the far-field radiation pattern of the array antenna is as follows:

[0013]

[0014] In the formula, The far-field radiation pattern of the array antenna, w n and These are the original excitation and active element radiation patterns of the nth element, respectively, where n = 1, 2, ..., N, and N is the total number of antenna elements. Let e ​​be the imaginary unit, β be the base of the natural logarithm, and r be the wavenumber of free space. n The guide vector and the position vector of the nth unit are respectively defined as:

[0015] R n =e x x n +e y y n

[0016] In the formula, e x and e y Let θ be the unit vector in the x and y directions in the Cartesian coordinate system. These represent the pitch angle and azimuth angle, respectively. n ,y n Let r be the coordinate of the nth cell, and let r be the coordinate of the cell. n These are the guide vector and the position vector of the nth unit, respectively.

[0017] Preferably, the amplitude of the damaged pattern of the measured actual array is calculated using the following formula:

[0018]

[0019] In the formula, The amplitude of the damaged pattern. Let |·| be the actual excitation of the nth element of the damaged array, |·| be the modulus operation of the complex number, and N be the total number of antenna elements. Let be the radiation pattern of the active element in the nth element, where e represents the base of the natural logarithm and β is the wavenumber in free space. Let r be the imaginary unit and R be the number of elements in the array. n These are the guide vector and the position vector of the nth unit, respectively.

[0020] Preferably, the simulation processing of the far-field radiation pattern of the array antenna to establish a training set for the decoder neural network includes:

[0021] Based on the far-field radiation pattern of the array antenna, the training set of the decoder neural network is established by simulating the excitation of the virtual array corresponding to different damage conditions of different damaged units and the amplitude information of the corresponding damaged radiation pattern. The simulation of different damage conditions of different damaged units involves randomly selecting S array antenna units from N array antenna units as damaged units, randomly setting the magnitude of the corresponding virtual excitation, and calculating the amplitude of the damaged radiation pattern.

[0022] Preferably, the decoder neural network is pre-trained using the established training set to obtain a new pre-trained decoder neural network. The decoder neural network has four layers, with the first to last layers being fully connected layers. Except for the third layer, which uses the ReLU activation function, the activation functions of the remaining layers are all Sigmoid functions. After the decoder neural network completes pre-training, the real and imaginary parts of the virtual excitation are used as inputs to the decoder neural network, allowing it to output the predicted amplitude of the damaged radiation pattern. The number of nodes in the decoder neural network is 2N, 1024, 512, and M, where M is the number of sampling points for the far-field radiation pattern amplitude sampling. During pre-training, the training parameters include a learning rate of 0.001, a batch size of 32, and an epoch count of 1000.

[0023] Preferably, the input is fed into an encoder neural network for training. The encoder neural network has six layers, with the first to last layers being fully connected layers. Except for the fifth layer, which uses the ReLU activation function, the activation functions of the remaining layers are all Sigmoid functions. The number of nodes in the encoder neural network is M, 256, 512, 1024, 512, and 2N. During pre-training, the training parameters include a learning rate of 0.0001 and an epoch count of 1000.

[0024] Preferably, the virtual array is obtained by first subtracting the original array pattern from the pattern of the damaged array. The pattern of the array is called a virtual array. This is the virtual excitation for the nth unit corresponding to the virtual orientation pattern.

[0025] Preferably, the cascaded encoder-decoder neural network is trained based on a loss function preset in the compressed sensing framework until training is complete, wherein the calculation formula of the loss function preset in the compressed sensing framework is as follows:

[0026] Loss encoder =||P P -P A ||2+τ||Δw|| SL0

[0027] In the formula, Loss encoder Let P be the loss function, τ be the regularization parameter, and P be the regularization parameter. P and P A These represent the predicted power pattern and the actual power pattern, respectively, where the actual power pattern P... A Amplitude of the damaged unit The relationship is: ||·||2 and||·|| SL0 These represent the L2 norm and smooth L0 norm operations, respectively, where ||Δw|| SL0 Represents the smooth L0 norm operation on the virtual excitation Δw: N is the total number of antenna elements, |Δw n | represents the magnitude of the nth virtual stimulus, g σ (|Δw n |) represents a family of Gaussian functions. σ is the smooth L0 norm parameter, and e represents the base of the natural logarithm, which is dynamically adjusted according to different training cycles. This represents the limit where σ approaches 0.

[0028] Preferably, the step of determining the location index of the damaged element of the phaseless array antenna and the original excitation based on the excitation of the virtual array, wherein the excitation of the virtual array is Δwn The relationship between the number of encoders (n = 1, 2, ..., N) and the actual output of the encoder is as follows:

[0029] v=[Re(Δw n ),Im(Δw n ); n = 1, 2, ..., N]

[0030] In the formula, Re(Δw) n ) and Im(Δw n ) represent the virtual excitation Δw respectively n The operations of extracting the real and imaginary parts are performed, where v is the 2N-dimensional output vector of the encoder, and the index of the damaged cell is Δw. n The index n corresponding to ≠0 is the original excitation of the damaged unit, which is the virtual excitation Δw. n N is the total number of antenna elements.

[0031] The beneficial effects of this invention are as follows:

[0032] This invention proposes a phase-free method for diagnosing damaged array antennas by using sparse recovery technology based on compressed sensing in artificial neural networks, which has high accuracy and efficiency. The cascaded encoder-decoder neural network can adapt to various damage scenarios and is supplemented by a smooth L0 norm method to enhance the recovery of excitation sparsity. This process eliminates the need for complex mathematical modeling and also eliminates mutual coupling effects in the array through training data.

[0033] This invention can not only determine the location of damaged units, but also accurately recover the phase of the excitation source from phase-free measurements. Compared with existing diagnostic algorithms based on compressed sensing and artificial neural networks, this invention has higher accuracy, efficiency, and robustness.

[0034] This invention addresses the problem of restoring the excitation distribution of damaged elements in an array antenna by measuring amplitude alone, thereby determining the location of the damaged elements. This invention is applied to both ideal omnidirectional linear antenna arrays and real antenna arrays that have been damaged, evaluating the ability to diagnose damaged elements and reconstruct the excitation magnitude of damaged elements.

[0035] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing as described in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram showing the encoder loss and mean square error of the recovery excitation as a function of Epoch in the implementation of ideal array antenna diagnosis in this embodiment of the invention.

[0039] Figure 3 This is a schematic diagram comparing the recovery excitation of ideal array diagnosis with the actual virtual excitation and the phase-free L1 norm method in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the damage pattern and sampling point distribution for implementing real array antenna diagnosis in this embodiment of the invention;

[0041] Figure 5 This is a schematic diagram comparing the recovery excitation amplitude and phase of the actual array antenna diagnosis implemented in the embodiments of the present invention with the actual situation. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0043] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0044] Example 1:

[0045] This embodiment provides a diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing.

[0046] See Figure 1 The figure shows that the method includes steps S100, S200, S300, S400 and S500.

[0047] S100. Based on the extracted active element radiation pattern of the array antenna, calculate the far-field radiation pattern of the array antenna.

[0048] It is understood that the relationship between the active element radiation pattern of the array antenna and the far-field radiation pattern of the array antenna in this step S100 is as follows:

[0049]

[0050] In the formula, The far-field radiation pattern of the array antenna, w n and These are the original excitation and active element radiation patterns of the nth element, respectively, where n = 1, 2, ..., N, and N is the total number of antenna elements. Let e ​​be the imaginary unit, β be the base of the natural logarithm, and r be the wavenumber of free space. n The guide vector and the position vector of the nth unit are respectively defined as:

[0051] R n =e x x n +e y y n

[0052] In the formula, e x and e y Let θ be the unit vector in the x and y directions in the Cartesian coordinate system. These represent the pitch angle and azimuth angle, respectively. n ,y n Let r be the coordinate of the nth cell, and let r be the coordinate of the cell. n These are the guide vector and the position vector of the nth unit, respectively.

[0053] S200. Simulate the far-field radiation pattern of the array antenna to establish a training set for the decoder neural network. Use the established training set to pre-train the decoder neural network to obtain a new pre-trained decoder neural network.

[0054] It is understood that the simulation processing of the far-field radiation pattern of the array antenna described in step S200 to establish the training set of the decoder neural network includes:

[0055] Based on the far-field radiation pattern of the array antenna, the training set of the decoder neural network is established by simulating the excitation of the virtual array corresponding to different damage conditions of different damaged units and the amplitude information of the corresponding damaged radiation pattern. The simulation of different damage conditions of different damaged units involves randomly selecting S array antenna units from N array antenna units as damaged units, randomly setting the magnitude of the corresponding virtual excitation, and calculating the amplitude of the damaged radiation pattern.

[0056] It should be noted that a virtual array refers to obtaining the array by first subtracting the original array pattern from the pattern of the damaged array. The pattern of the array is called a virtual array. This is the virtual excitation for the nth unit corresponding to the virtual orientation pattern.

[0057] Understandably, step S200 also includes pre-training the decoder neural network using the established training set to obtain a new pre-trained decoder neural network. The decoder neural network has four layers, with the first to last layers being fully connected layers. Except for the third layer, which uses the ReLU activation function, the activation functions of the remaining layers are all Sigmoid functions. The number of nodes in the decoder neural network are 2N, 1024, 512, and M, respectively, where M is the number of sampling points for the amplitude sampling of the far-field radiation pattern. During the pre-training process, the training parameters include a learning rate of 0.001, a batch size of 32, and an epoch count of 1000.

[0058] Once the decoder neural network has completed its pre-training, the real and imaginary parts of the virtual excitation are used as inputs to the decoder neural network, which can then output the amplitude of the predicted damaged pattern.

[0059] The input to the decoder can be represented as a vector:

[0060] v=[Re(Δw n ),Im(Δw n ); n = 1, 2, ..., N]

[0061] In the formula, v is the 2N-dimensional input vector of the decoder, and Re(Δw) n ) and Im(Δw n ) represent the virtual excitation Δw respectively n After extracting the real and imaginary parts, the decoder output can be represented as a vector. Where u is the M-dimensional output vector of the decoder. The m-th pitch and azimuth sampling positions in the far field.

[0062] S300 measures the amplitude of the damaged pattern of the actual array and inputs it into the encoder neural network for training to obtain the predicted excitation of the damaged units.

[0063] It is understood that the amplitude of the damaged pattern of the actual array measured in step S300 is calculated using the following formula:

[0064]

[0065] In the formula, The amplitude of the damaged pattern. Let |·| be the actual excitation of the nth element of the damaged array, |·| be the modulus operation of the complex number, and N be the total number of antenna elements. Let be the radiation pattern of the active element in the nth element, where e represents the base of the natural logarithm and β is the wavenumber in free space. Let r be the imaginary unit and R be the number of elements in the array. n These are the guide vector and the position vector of the nth unit, respectively.

[0066] The input is fed into an encoder neural network for training. The encoder neural network has six layers, with the first to last layers being fully connected layers. Except for the fifth layer, which uses the ReLU activation function, the activation functions of the remaining layers are all Sigmoid functions. The number of nodes in the encoder neural network are M, 256, 512, 1024, 512, and 2N, respectively. During pre-training, the training parameters include a learning rate of 0.0001 and an epoch count of 1000.

[0067] Preferably, the virtual array is obtained by first subtracting the original array pattern from the pattern of the damaged array.

[0068]

[0069] In the formula, The pattern of the array is called a virtual array. The amplitude of the damaged radiation pattern is N, where N is the total number of antenna elements. This is the virtual excitation for the nth unit corresponding to the virtual radiation pattern. Let be the radiation pattern of the active element in the nth element, where e represents the base of the natural logarithm and β is the wavenumber in free space. Let r be the imaginary unit and R be the number of elements in the array. n These are the guide vector and the position vector of the nth unit, respectively.

[0070] S400. The predicted excitation of the damaged unit is used as the input of the new decoder neural network after pre-training. The encoder-decoder neural network is cascaded and trained on the cascaded encoder-decoder neural network based on the loss function preset in the compressed sensing framework until training is completed, so as to obtain the excitation of the virtual array.

[0071] It is understandable that the calculation formula for the loss function in the compressed sensing framework is preset in this S400 step as follows:

[0072] Loss encoder =||P P -P A ||2+τ||Δw|| SL0

[0073] In the formula, Loss encoder Let P be the loss function, τ be the regularization parameter, and P be the regularization parameter. P and P A These represent the predicted power pattern and the actual power pattern, respectively, where the actual power pattern P... A Amplitude of the damaged unit The relationship is: ||·||2 and||·|| SL0 These represent the L2 norm and smooth L0 norm operations, respectively, where ||Δw|| SL0 Represents the smooth L0 norm operation on the virtual excitation Δw: N is the total number of antenna elements, |Δw n | represents the magnitude of the nth virtual stimulus, g σ (|Δw n |) represents a family of Gaussian functions. σ is the smooth L0 norm parameter, and e represents the base of the natural logarithm, which is dynamically adjusted according to different training cycles. This represents the limit where σ approaches 0.

[0074] S500: Based on the excitation of the virtual array, confirm the location index of the damaged element of the phaseless array antenna and the original excitation.

[0075] It is understandable that the excitation of the virtual array in this S500 step is Δw n The relationship between the number of encoders (n = 1, 2, ..., N) and the actual output of the encoder is as follows:

[0076] v=[Re(Δw n ),Im(Δw n ); n = 1, 2, ..., N]

[0077] In the formula, Re(Δw) n ) and Im(Δw n) represent the virtual excitation Δw respectively n The operations of extracting the real and imaginary parts are performed, where v is the 2N-dimensional output vector of the encoder, and the index of the damaged cell is Δw. n The index n corresponding to ≠0 is the original excitation of the damaged unit, which is the virtual excitation Δw. n N is the total number of antenna elements.

[0078] This invention proposes a phaseless array antenna diagnostic method based on artificial neural networks and compressed sensing, utilizing sparse recovery techniques based on compressed sensing within artificial neural networks. This method boasts high accuracy and efficiency, and its cascaded encoder-decoder neural network adapts to various damage scenarios. Furthermore, it incorporates a smooth L0 norm method to enhance excitation sparsity recovery. Compared to existing diagnostic algorithms, this invention offers higher accuracy, efficiency, and robustness.

[0079] Example 2:

[0080] like Figure 2 and Figure 3 As shown, this embodiment provides a phaseless array antenna diagnostic method based on artificial neural networks and compressed sensing, which is applied to actual damaged ideal omnidirectional antenna linear arrays and real antenna arrays to evaluate the ability to diagnose damaged elements of the array and the ability to reconstruct the excitation magnitude of damaged elements. In this embodiment, a 100-element half-wave pitch ideal antenna linear array is used for illustration.

[0081] Specifically, the array consists of N = 100 ideal omnidirectional antenna elements with an antenna spacing of λ / 2, where λ is the free-space wavelength. The antenna elements are excited using Chebyshev synthesis weighting with a sidelobe level of -20dB. The number of damaged elements is selected as S = 8, and elements in the array are randomly assigned to be damaged, with the excitation of the damaged elements set to 0. The sampling quantity is set to M = 80, the smoothing L0 norm parameter σ = [1, 0.5, 0.25, 0.125, 0.062, 0.031, 0.0162, 0.0078, 0.0039], and the regularization parameter τ = 10. Finally, during the pattern sampling process, Gaussian white noise with a signal-to-noise ratio (SNR) of 30dB is applied to the sampled signal. The specific steps of this invention are used to diagnose the damaged elements of the aforementioned 100-element half-wave spacing ideal antenna array.

[0082] It needs to be specifically explained that, such as Figure 2 The results shown in the figure illustrate the variation of encoder loss and mean square error of the recovered excitation with Epoch. From... Figure 2As can be seen, unlike the loss function of conventional artificial neural networks, the loss function undergoes periodic increases and decreases under different smoothing L0 norm parameters σ. This is attributed to the fact that the sparsity penalty in the loss function increases with the smoothing L0 norm. Furthermore, the diagnostic mean square error of the array gradually decreases with each epoch, and the final mean square error result demonstrates that this invention possesses good diagnostic accuracy.

[0083] Specifically, it should be noted that, such as Figure 3 The results shown correspond to the amplitude and phase comparisons of the actual virtual excitation of the array, the excitation diagnosed by this method, and the excitation diagnosed by the phase-free L1 norm method, respectively. From Figure 3 As can be seen from this, the diagnostic method proposed in this invention has higher diagnostic accuracy than the existing phase-free L1 norm method. It can successfully recover the amplitude and phase of the virtual excitation of the damaged unit in a noisy environment, thereby more accurately determining the location and damage condition of the damaged unit.

[0084] Example 3:

[0085] like Figure 4 and Figure 5 As shown, this embodiment provides a phaseless array antenna diagnostic method based on artificial neural networks and compressed sensing, which is applied to actual damaged ideal omnidirectional antenna linear arrays and real antenna arrays to evaluate the ability to diagnose damaged elements of the array and the ability to reconstruct the excitation magnitude of damaged elements. In this embodiment, a 25-element half-wave pitch circular ring antenna array is used for illustration.

[0086] Specifically, the array consists of N = 5 × 5 real loop antennas with an antenna spacing of λ / 2, where λ is the free-space wavelength. The antenna elements are excited using uniform weighting; the number of damaged elements is selected as S = 4, and elements in the array are randomly assigned to be damaged, with the excitation of the damaged elements set to 0; the sampling quantity is set to M = 32, the smoothing L0 norm parameter is consistent with implementation method 1, and the regularization parameter τ = 10; finally, during the pattern sampling process, Gaussian white noise with a signal-to-noise ratio (SNR) of 30 dB is applied to the sampled signal. The specific steps of this invention are used to diagnose the damaged elements of the aforementioned 25-element half-wave spacing real antenna array.

[0087] It needs to be specifically explained that, such as Figure 4 The results show the normalized radiation pattern distribution of the actual loop antenna in the upper half-space, and the positions of 32 random sampling points are marked. The radiation pattern amplitude is sampled at these locations, and after considering noise, the data is input into an encoder-decoder artificial neural network for diagnostics.

[0088] Specifically, it should be noted that, such as Figure 5The results shown correspond to the diagnostic results of two different damage scenarios—amplitude damage and phase damage. Amplitude damage refers to a deviation in the amplitude of the damaged element compared to the original amplitude, while phase damage refers to a deviation in the phase of the damaged element compared to the original phase. It can be seen that, regardless of whether it is amplitude damage or phase damage, the actual virtual excitation of the array and the diagnostic excitation of this method fit well, indicating that this invention can effectively diagnose the array based on the amplitude and phase of the recovered virtual array.

[0089] Example 4:

[0090] Corresponding to the above method embodiments, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the phaseless array antenna diagnostic method based on artificial neural networks and compressed sensing described above.

[0091] A computer program is stored on a readable storage medium, and when the computer program is executed by a processor, it implements the steps of the phaseless array antenna diagnostic method based on artificial neural networks and compressed sensing in the above method embodiments.

[0092] Specifically, the readable storage medium can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other readable storage medium capable of storing program code.

[0093] In summary, the phaseless array antenna diagnostic method proposed in this invention, based on artificial neural networks and compressed sensing, not only eliminates the need for complex mathematical modeling but also eliminates mutual coupling effects in the array through training data. Furthermore, this invention can not only determine the location of damaged elements but also accurately recover the phase of the excitation source from phaseless measurements, exhibiting higher accuracy, efficiency, and robustness. The location of damaged elements can be determined simply by recovering the excitation distribution of the damaged elements through amplitude measurement.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0095] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing, characterized in that, include: Based on the extracted active element radiation pattern of the array antenna, the far-field radiation pattern of the array antenna is calculated. The far-field radiation pattern of the array antenna is simulated to establish a training set for the decoder neural network. The decoder neural network is then pre-trained using the established training set to obtain a new pre-trained decoder neural network. The amplitude of the damaged pattern of the actual array is measured and input into the encoder neural network for training to obtain the predicted excitation of the damaged cells. The predicted damaged unit excitation is used as the input of the new decoder neural network after pre-training. The encoder-decoder neural network is cascaded and trained on the cascaded encoder-decoder neural network based on the loss function preset in the compressed sensing framework until training is completed, so as to obtain the excitation of the virtual array. Based on the excitation of the virtual array, the location index of the damaged element of the phaseless array antenna and the original excitation are identified; The relationship between the active element radiation pattern and the far-field radiation pattern of the array antenna is as follows: In the formula, This represents the far-field radiation pattern of the array antenna. and The first The original excitation and active element radiation pattern of each unit. , This represents the total number of antenna elements. For imaginary units, The base of the natural logarithm. For the wavenumber in free space, and respectively the guide vector and the first The position vector of each unit is defined as: ， In the formula, and In Cartesian coordinate system and The unit vector of direction, and These represent the pitch angle and azimuth angle, respectively. For the first The coordinates of each unit, and respectively the guide vector and the first The position vector of each unit.

2. The diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing according to claim 1, characterized in that, The amplitude of the damaged pattern of the actual array is calculated using the following formula: In the formula, The amplitude of the damaged pattern. For the damaged array The actual incentive for each unit For the modulo operation of complex numbers, This represents the total number of antenna elements. For the first The radiation pattern of the active element in each cell. The base of the natural logarithm. For the wavenumber in free space, For imaginary units, and respectively the guide vector and the first The position vector of each unit.

3. The diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing according to claim 1, characterized in that, The simulation processing of the far-field radiation pattern of the array antenna to establish the training set of the decoder neural network includes: Based on the far-field radiation pattern of the array antenna, the training set for the decoder neural network is established by simulating the excitation of the virtual array corresponding to different damage conditions of different damaged elements, and the amplitude information of the corresponding damaged radiation patterns. The simulation of different damage conditions of different damaged elements is used to... Randomly select from the array antenna elements Each array antenna element is a damaged element, and the magnitude of the corresponding virtual excitation is randomly set, and the amplitude of the damaged pattern is calculated.

4. The diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing according to claim 1, characterized in that, The decoder neural network is pre-trained using the established training set to obtain a new pre-trained decoder neural network. The decoder neural network has four layers, with the first to the last layer being fully connected layers. Except for the third layer, which uses the ReLU activation function, the activation functions of the remaining layers are all Sigmoid functions. After the decoder neural network completes pre-training, the real and imaginary parts of the virtual excitation are used as inputs to the decoder neural network, which can then output the amplitude of the predicted damaged radiation pattern; the number of nodes in the decoder neural network are respectively 1024, 512 and ,in The number of sampling points for far-field radiation pattern amplitude sampling. During pre-training, the training parameters include a learning rate of 0.001, a batch size of 32, and an epoch count of 1000.

5. The diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing according to claim 1, characterized in that, The input is then fed into an encoder neural network for training. The encoder neural network has six layers, with the first to last layers being fully connected layers. Except for the fifth layer, which uses the ReLU activation function, the remaining layers all use the Sigmoid activation function. The number of nodes in the encoder neural network is as follows: 256, 512, 1024, 512 and During the pre-training process, the training parameters included a learning rate of 0.0001 and an epoch count of 1000.

6. The diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing according to claim 1, characterized in that, The virtual array is obtained by first subtracting the original array pattern from the pattern of the damaged array: In the formula, The pattern of the array is called a virtual array. The amplitude of the damaged pattern. This represents the total number of antenna elements. The first corresponding to the virtual orientation pattern Virtual stimulus for each unit For the first The radiation pattern of the active element in each cell. The base of the natural logarithm. For the wavenumber in free space, For imaginary units, and respectively the guide vector and the first The position vector of each unit.

7. The diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing according to claim 1, characterized in that, The concatenated encoder-decoder neural network is trained based on a loss function preset in the compressed sensing framework until training is complete. The calculation formula for the loss function preset in the compressed sensing framework is as follows: In the formula, For loss function, For regularization parameters, and These represent the predicted power pattern and the actual power pattern, respectively, where the actual power pattern... Amplitude of the damaged unit The relationship is: , and These represent L2 norm and smooth L0 norm operations, respectively. Indicates virtual incentives Smoothing L0 norm operations: ,in, This represents the total number of antenna elements. Indicates the first The magnitude of a virtual incentive, It is a family of Gaussian functions. , For smooth L0 norm parameters, The base of the natural logarithm is dynamically adjusted according to different training cycles. express The limit that approaches 0.

8. The diagnostic method for phaseless array antennas based on artificial neural networks and compressed sensing according to claim 1, characterized in that, The location index of the damaged element of the phaseless array antenna and the original excitation are determined based on the excitation of the virtual array, wherein the excitation of the virtual array is... The relationship between the encoder's output and the actual output is as follows: In the formula, and These represent virtual incentives. Real part and imaginary part operations For encoder The output vector of dimension, the index of the damaged unit is... Corresponding index The original stimulus of the damaged unit is the virtual stimulus. , This represents the total number of antenna elements.

Images