Method for focusing and regulating radiation near field of MWPT array antenna based on CNN

Abstract

This application discloses a CNN-based near-field focusing control method for MWPT array antenna radiation, belonging to the field of microwave wireless power transfer technology. The method includes: constructing a convolutional neural network; constructing a training dataset with the desired focal position and the corresponding ideal near-field radiation sampling matrix as training samples; training the convolutional neural network using the training dataset, enabling the network to learn the mapping from the focal position to the array's outgoing phase distribution; inputting the actual desired focal position into the trained convolutional neural network, outputting the corresponding array outgoing phase distribution, and configuring the phase of each element of the array antenna according to the outgoing phase distribution to achieve near-field focusing. This method, for any desired focal position, only needs to complete the calculation in a very short time to achieve the desired field distribution control, enabling diversified energy focusing in the target area while reducing microwave radiation in non-target areas, and possessing stronger field distribution constraint capabilities.

Description

Technical Field

[0001] This application relates to the field of microwave wireless power transmission technology, and in particular to a near-field focusing control method for MWPT array antenna radiation based on CNN. Background Technology

[0002] In the near-field power transfer scenario of a microwave wireless power transfer (MWPT) system, the electric field distribution on the focal plane can be controlled by adjusting the phase of the microwaves emitted by each element of the array antenna, thereby achieving the purpose of energy focusing.

[0003] Existing technologies typically rely on two main categories of array antennas: analytical methods based on electromagnetic theory and numerical optimization methods. Analytical methods, based on quasi-optical path rules, construct the spatial geometric relationships between array antenna elements and calculate the required phase for each element based on the target focal position. This allows electromagnetic waves propagating along different paths to coherently superimpose in the target region, achieving energy focusing. However, due to the spherical radiation characteristics of microwaves in the near field, these methods can only achieve energy focusing at the target location, but cannot constrain the electric field distribution at and outside the focal point. This leads to a series of problems such as low microwave energy transmission efficiency, difficulties in beamforming, and microwave irradiation of non-target areas. Numerical optimization-based control methods (such as genetic algorithms and particle swarm optimization) construct an electric field distribution error function and continuously optimize the output excitation configuration of the array elements, thereby minimizing the difference between the target radiation near-field distribution and the actual radiation field distribution of the array antenna. While these methods can overcome the limitations of analytical methods to some extent and achieve more complex electric field control tasks, their computational process typically involves a large number of electromagnetic field iterations, resulting in high computational overhead and slow convergence speed. This is particularly problematic in high-dimensional arrays (large-scale cells) or dynamic real-time control scenarios, where they struggle to meet the demands for rapid response. Furthermore, since optimization problems are often highly non-convex, the algorithms are prone to getting trapped in local optima, thus affecting the final control effect. Therefore, existing electric field control methods based on array antennas still face a significant trade-off between accuracy and efficiency, making it difficult to simultaneously meet the requirements of real-time performance and high-quality electric field reconstruction.

[0004] Existing MWPT methods still have significant shortcomings in controlling the near-field electric field distribution using array antennas. While analytical methods based on quasi-optical path rules are computationally simple and physically clear, they essentially only achieve coherent superposition of target points through path-like phase compensation. They can only constrain the focal position and are unable to finely control the focal shape and the overall electric field distribution of the focal plane, lacking effective constraints on key indicators such as sidelobes. Therefore, their applicability is limited in complex near-field distribution design tasks. On the other hand, numerical optimization methods based on genetic algorithms and particle swarm optimization can globally approximate the target electric field distribution by constructing error functions, improving the degree of freedom of control to some extent. However, they rely on a large number of electromagnetic field forward and inverse iterative calculations, resulting in high computational complexity, slow convergence speed, and a tendency to get trapped in local optima in high-dimensional parameter spaces. This makes it difficult to meet the computational efficiency and stability requirements of large-scale arrays and dynamic real-time control scenarios. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this application provides a CNN-based near-field focusing control method for MWPT array antennas, which solves the problems of low control freedom, inability to constrain the overall electric field distribution of the focal plane, high computational complexity, poor real-time performance, and susceptibility to local optima in existing near-field focusing control methods for array antennas.

[0006] To achieve the aforementioned objectives, the technical solution adopted in this application is as follows: This application provides a CNN-based method for near-field focusing control of MWPT array antenna radiation, including: S1: Construct a convolutional neural network; S2: Construct a training dataset with the desired focal position and the corresponding ideal radiation near-field sampling matrix as training samples; S3: Train a convolutional neural network using the training dataset so that the network learns the mapping from the focal position to the phase distribution of the array output; S4: Input the actual desired focal position into the trained convolutional neural network, output the corresponding array output phase distribution, and configure the phase of each element of the array antenna according to the output phase distribution to achieve near-field focusing.

[0007] Furthermore, the convolutional neural network includes: multiple transposed convolutional layers, multiple convolutional layers, and a batch normalization layer and an activation function sequentially set after each non-output layer; After normalization in each non-output layer, Leaky ReLU is used as the activation function, and hyperbolic tangent is used as the activation function in the output layer.

[0008] Furthermore, the mathematical expression for the batch normalization layer is:

[0009] In the formula, The feature map is after batch normalization. and These are the weights and biases, respectively. The standardized feature map, This represents the feature map obtained after transpose convolution or convolution operations. For the number of floors, For the first Each output channel weight and These are the mean and standard deviation of the current batch, respectively. It is a constant; The Leaky ReLU expression is:

[0010] In the formula, For a modified linear unit with leakage, The value of the input activation function, Leakage coefficient; The expression for the hyperbolic tangent function as the activation function is:

[0011] In the formula, It is the hyperbolic tangent function.

[0012] Furthermore, the construction of the training dataset, which uses the desired focal position and the corresponding ideal radiation near-field sampling matrix as training samples, includes: With the goal of achieving square uniform focusing, an ideal radiation near-field sampling matrix is ​​constructed, the focal center of the ideal radiation near-field sampling matrix is ​​obtained, and different corresponding ideal radiation near-field sampling matrices are obtained by changing the position of the focal center, thus obtaining the training dataset.

[0013] Further, S3 includes: S301: The output matrix is ​​obtained using a convolutional neural network, and the outgoing phase distribution matrix is ​​obtained based on the output matrix; S302: Based on the near-field radiation field calculation formula, the radiation phase distribution matrix is ​​transformed into the corresponding near-field distribution, and the near-field distribution is discretized to obtain the discrete sampling matrix of the radiation field distribution. S303: Based on the discrete sampling matrix of the radiation field distribution, the desired focal position and the corresponding ideal radiation near-field sampling matrix, a mean squared error loss function is constructed to train the convolutional neural network, and the Adam optimizer is used to update the gradient of the parameters of each layer of the convolutional neural network until the mean squared error loss function converges.

[0014] Furthermore, the expressions for the output matrix and the emitted phase distribution matrix are as follows:

[0015]

[0016] in, The output matrix is ​​obtained by using the hyperbolic tangent function as the activation function. For the first element in the output matrix lines and The values ​​of the array antenna elements in the column, This represents the row number of the array antenna elements. This refers to the number of columns of the array antenna elements. The outgoing phase distribution matrix, The first phase distribution matrix is ​​the first... lines and The output phase of the array antenna elements.

[0017] Furthermore, the formula for calculating the near-field radiation field is as follows:

[0018] In the formula, Let be the field distribution function of the near-field radiation. For the first lines and The element radiation pattern of the array antenna elements. The pitch angle, It is the azimuth angle. For the first lines and The emitted amplitude of the array antenna elements in the column, For the first lines and Array antenna elements To the observation point distance, The imaginary unit, The wavelength of the emitted wave. For the first lines and The output phase of the array antenna elements of the column, , and The x, y, and y coordinates of the observation point. and For the first lines and The horizontal and vertical coordinates of the array antenna elements in the column.

[0019] Furthermore, the expression for the mean squared error loss function is:

[0020] In the formula, Let the mean squared error loss function be . For an ideal radiation near-field sampling matrix, The discrete sampling matrix represents the radiation field distribution. The first of the ideal radiation near-field sampling matrix One element, The first discrete sampling matrix of the radiation field distribution One element, The total number of matrix elements. It is the square norm.

[0021] The beneficial effects of this application are: This application provides a CNN-based near-field focusing control method for MWPT array antenna radiation. To quickly obtain the desired focal plane radiation field distribution according to the requirements of MWPT applications, this application specifically designs a convolutional neural network that can rapidly, efficiently, and accurately obtain the corresponding output phase configuration of each element of the array antenna. Furthermore, the trained convolutional neural network can complete the calculation to achieve the desired field distribution control for any desired focal position in a very short time, demonstrating strong real-time advantages. In addition, this method can achieve diverse energy focusing in the target area while reducing microwave radiation in non-target areas, exhibiting stronger field distribution constraint capabilities. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0023] Figure 1 This is a schematic diagram of a quasi-optical path rule provided in an embodiment of this application.

[0024] Figure 2 This is a flowchart illustrating a CNN-based near-field focusing control method for MWPT array antenna radiation, provided in an embodiment of this application.

[0025] Figure 3 A CNN structure diagram is provided for an embodiment of this application.

[0026] Figure 4 This is a schematic diagram of a visualization of an ideal radiation near-field discrete sampling matrix provided in an embodiment of this application.

[0027] Figure 5A near-field radiation field calculation diagram is provided for an embodiment of this application.

[0028] Figure 6 This is a schematic diagram of a training process loss function curve provided in an embodiment of this application.

[0029] Figure 7 This application provides a near-field square uniform focusing calculation result for an embodiment of the present application.

[0030] Figure 8 This is a schematic diagram of a simulation structure provided for an embodiment of this application.

[0031] Figure 9 An electric field value for a focal plane cross section is provided in an embodiment of this application.

[0032] Figure 10 Another near-field square uniform focusing calculation result provided for the embodiments of this application.

[0033] Figure 11 Another focal plane cross-sectional electric field value provided for embodiments of this application. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.

[0035] Example 1: like Figure 1 As shown, in the energy focusing process using the analytical method, by M × N A transmitting antenna composed of array antenna elements needs to be positioned at the target focal point. For energy focusing, the output phase expression for each array antenna element is as follows:

[0036] In the formula, For the first lines and The output phase of the array antenna elements of the column, For the first lines and Array antenna elements To the target focus position Euclidean distance, , and The coordinates of the target focus on the horizontal, vertical, and y-axis. As a reference phase, The wavelength of the radiated electromagnetic wave. For the first One target focus, and This represents the number of rows and columns of the array antenna elements.

[0037] Based on this, this application provides a CNN-based method for controlling the near-field focusing of MWPT array antenna radiation. This method can be found in [reference needed]. Figure 2 , Figure 2 The diagram shown is a flowchart illustrating a CNN-based near-field focusing control method for MWPT array antenna radiation, as provided in an embodiment of this application, including: S1: Construct a convolutional neural network.

[0038] Furthermore, the convolutional neural network includes: multiple transposed convolutional layers, multiple convolutional layers, and batch normalization layers and activation functions sequentially set after each non-output layer; after normalization of each non-output layer, Leaky ReLU is used as the activation function, and in order to make the output layer have a clear physical meaning (the output phase of each unit), a smooth and bounded hyperbolic tangent function is used as the activation function in the output layer.

[0039] The mathematical expression for the batch normalization layer is:

[0040] In the formula, The feature map is after batch normalization. and These are the weights and biases, respectively. The standardized feature map, This represents the feature map obtained after transpose convolution or convolution operations. For the number of floors, For the first Each output channel weight and These are the mean and standard deviation of the current batch, respectively. To prevent small constants from taking the denominator as 0.

[0041] The Leaky ReLU expression is:

[0042] In the formula, For a modified linear unit with leakage, The value of the input activation function, The leakage coefficient is a small positive number. Compared to the traditional ReLU activation function, its difference lies in... xThe region <0 introduces a non-zero gradient, which effectively alleviates the problem of neuron death.

[0043] The expression for the smooth and bounded hyperbolic tangent function as the activation function is:

[0044] In the formula, The function is a hyperbolic tangent. At this point, the data from the output layer is mapped to the interval [-1, 1]. This symmetrical and continuously bounded data distribution will help to further map it to the radiation phase of the antenna array elements, thereby calculating the radiation field.

[0045] In one embodiment of this application, the CNN structure designed to calculate the output phase of a 19×19 array antenna is as follows: Figure 3 As shown, where Tconv h Indicates the first h Layer transposed convolution, Conv g Indicates the first g Convolutional layers. During CNN network training, parameters are constantly updated, and the data received by each layer from the previous layer is constantly changing. This non-stationary situation, where the distribution of input data changes between intermediate layers during training, is called internal covariate shift. This can lead to problems such as decreased model convergence speed, training instability (e.g., gradient explosion or vanishing gradient problems), and high sensitivity to hyperparameters. Therefore, to stabilize the training process, batch normalization is added after each non-output layer. By standardizing the results of each layer to a distribution with zero mean and unit variance, internal covariate shift can be significantly mitigated.

[0046] S2: Construct a training dataset with the desired focal position and the corresponding ideal radiation near-field sampling matrix as training samples.

[0047] In one embodiment of this application, to enable the CNN network to map from the focal point to the radiation field distribution, training data needs to be constructed and used for training. The following section describes the construction of an ideal radiation near-field sampling matrix, aiming to achieve "square uniform focusing" as the radiation near-field control objective. Visualization, such as Figure 4 As shown in the figure, the red part of the ideal near-field discrete sampling matrix visualization is the focal region, which is square in shape and oriented with... The center of the focus has an electric field value of 1, while the blue area represents the region outside the focus, where the electric field value is 0. At this point, the matrix... With the center of focus This forms training data pairs. By changing the position of the focal center, different ideal electric field sampling matrices can be obtained. Experimental results show that this method can obtain approximately 1.0 × 10⁻⁶ units of light per minute.4 One training data pair.

[0048] S3: Train a convolutional neural network using the training dataset so that the network learns the mapping from the focal position to the phase distribution of the array output.

[0049] Further, S3 includes: S301: The output matrix is ​​obtained using a convolutional neural network, and the output phase matrix is ​​obtained based on the output matrix. The expression is as follows:

[0050]

[0051] in, The output matrix is ​​obtained by using the hyperbolic tangent function as the activation function. For the first element in the output matrix lines and The values ​​of the array antenna elements in the column, The outgoing phase distribution matrix, The first phase distribution matrix is ​​the first... lines and The output phase of the array antenna elements.

[0052] In one embodiment of this application, it is set that , Then the model output is represented as a matrix. as follows:

[0053] Then the matrix The matrix is ​​obtained by taking the absolute value of each element and then multiplying by 360. :

[0054] At this time, the matrix The position of each element in the array corresponds strictly to the actual physical arrangement of the antenna elements, and their values ​​are assigned to the output phase of the corresponding element. The physical meaning of: such as This indicates that, based on the model calculations, the output phase value of the antenna element located at position (5,13) should be... Next, we will... As input to the array radiation field distribution calculation, after processing according to the corresponding calculation rules, the radiation field distribution sampling matrix obtained by the array antenna under the current phase excitation will be obtained. .

[0055] S302: Electromagnetic waves propagating in the near-field region exhibit significant spherical wave characteristics. According to the near-field radiation field calculation formula, ... The near-field radiation field is calculated by converting it into a discrete sampling matrix of radiation field distribution:

[0056] In the formula, Let be the field distribution function of the near-field radiation. For the first lines and The element radiation pattern of the array antenna elements. The pitch angle, It is the azimuth angle. For the first lines and The emitted amplitude of the array antenna elements in the column, For the first lines and Array antenna elements To the observation point distance, The imaginary unit, The wavelength of the emitted wave. For the first lines and The output phase of the array antenna elements of the column, , and The x, y, and y coordinates of the observation point. and For the first lines and The horizontal and vertical coordinates of the array antenna elements in the column.

[0057] In one embodiment of this application, such as Figure 5 As shown, For the first array antenna elements To the target focus position The Euclidean distance, and the phase distribution calculated by CNN. The formula for calculating the near-field radiation field can be transformed into its corresponding near-field distribution. At this point, using a step size of 1mm, [the following is a sequence of steps]... Discrete sampling will yield the results calculated by the CNN network. Discrete sampling matrix of the excited field distribution .

[0058] S303: Based on the discrete sampling matrix of the radiation field distribution, the desired focal position, and the corresponding ideal near-field sampling matrix, a mean squared error loss function is constructed to train the convolutional neural network. The Adam optimizer is then used to update the gradients of the parameters of each layer of the convolutional neural network until the mean squared error loss function converges. This ensures that the calculated field distribution approximates the ideal field distribution as closely as possible, thereby indirectly enabling it to capture the desired focal position. To the required array output phase distribution The mean squared error loss function is expressed as follows:

[0059] In the formula, Let the mean squared error loss function be . For an ideal radiation near-field sampling matrix, The discrete sampling matrix represents the radiation field distribution. The first of the ideal radiation near-field sampling matrix One element, The first discrete sampling matrix of the radiation field distribution One element, The total number of matrix elements. It is the square norm.

[0060] In one embodiment of this application, the number of CNN training epochs is set to 200, and the training loss function is obtained as follows: Figure 6 As shown, in the initial stage of training, since the optimization and updates of model parameters after initialization are based on the training set, and the model's generalization ability on unseen samples is evaluated by the validation set, the loss function value of the validation set is usually higher than that of the training set. As training progresses, the model gradually learns an effective mapping between input and output, so the values ​​of both gradually decrease and eventually tend to converge smoothly. The loss curves of the training set and the validation set show the same trend, and there is no obvious separation between them. No obvious overfitting or underfitting was observed, indicating that the model training process is stable.

[0061] S4: Input the actual desired focal position into the trained convolutional neural network, output the corresponding array output phase distribution, and configure the phase of each element of the array antenna according to the output phase distribution to achieve near-field focusing.

[0062] In one embodiment of this application, after training is complete, the desired focal position is input into the network, and the network outputs the corresponding array emission phase distribution within 0.2 seconds. (Using a focal position...) A square uniform focusing test was performed at the focal center, and the network output results are as follows: Figure 7As shown, (a) is the phase distribution result of the unit output; (b) is the electric field distribution of the focal plane calculated based on this phase distribution. It can be seen that... Figure 7 Figure (a) shows the calculated phase distribution of the unit output. Due to the randomness of the model calculation process, the results presented in the figure do not exhibit the regular distribution characteristics described by the quasi-optical path rule. At this time, the electric field distribution at the focal plane located at z=500mm excited by such an excitation distribution is as follows... Figure 7 As shown in (b), it is not difficult to see that a distinct square bright spot appears with (0 mm, 0 mm) as the geometric center in the figure.

[0063] To further verify, such as Figure 8 As shown, (a) is a schematic diagram of the simulation unit structure; (b) is the unit return loss; and (c) is a schematic diagram of the simulation array. Specifically, an array was constructed in the electromagnetic simulation software CST Studio Suite (CST for short) as shown. Figure 8 The array antenna element shown in (a) (dimensions in mm, PEC is the material name) is used to simulate and verify the near-field extrapolation of the model. The element return loss is as follows: Figure 8 As shown in (b), the material parameters of Rogers RT5880 used are: relative permittivity. Loss tangent Arrange them as follows: Figure 8 The 19×19 square antenna array shown in (c) is simulated in full-wave mode according to the cell phase distribution results calculated by CNN. The simulation results are as follows. Figure 9 As shown, (a) x =0mm cross-section electric field value; (b) y =0 mm cross-section electric field value. As shown in the figure, the model calculation results agree well with the CST simulation results, and the focusing position is as expected. Although the electric field value at (0 mm, 0 mm) in the figure is approximately E =0.92, which is not the expected maximum electric field, but considering the coverage area of ​​the focused bright spot, such an error is still within an acceptable range. Furthermore, the electric field value in the region outside the focal spot is generally suppressed to... E = Below 0.1 indicates that in addition to achieving focusing at the expected location, the model also has a good effect on suppressing the electric field in the region outside the focus, which reduces electromagnetic energy loss and improves energy utilization.

[0064] In one embodiment of this application, using A square uniform focusing calculation test was performed at another focal position, and the calculation results are as follows: Figure 10 As shown, (a) is the result of the unit phase calculation; (b) is the electric field distribution of the focal plane.

[0065] It's easy to see that the model's calculations still demonstrate the ability to achieve uniform square focusing at the preset position. Next, we will perform simulation verification in CST and extract... x =50 mm and y The electric field curve obtained at the focal plane at 50 mm is shown below. Figure 11 As shown, (a) is x =Electric field value of 50 mm cross section; (b) is y =Electric field value of 50 mm cross section. The calculated results in the figure agree well with the simulation, and it can be clearly distinguished that the focus is located at the preset position. The electric field in the area outside the focus is also well suppressed.

[0066] This application provides a CNN-based near-field focusing control method for MWPT array antenna radiation. To quickly obtain the desired focal plane radiation field distribution according to the requirements of MWPT applications, this application specifically designs a convolutional neural network that can rapidly, efficiently, and accurately obtain the corresponding output phase configuration of each element of the array antenna. Furthermore, the trained convolutional neural network can complete the calculation to achieve the desired field distribution control for any desired focal position in a very short time, demonstrating strong real-time advantages. In addition, this method can achieve diverse energy focusing in the target area while reducing microwave radiation in non-target areas, exhibiting stronger field distribution constraint capabilities.

[0067] It should be noted that those skilled in the art will recognize that the embodiments described herein are for the purpose of helping readers understand the principles of this application, and should be understood as not limiting the scope of protection of this application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this application without departing from the essence of this application, and these modifications and combinations are still within the scope of protection of this application.

Claims

1. A method for controlling the near-field focusing of MWPT array antenna radiation based on CNN, characterized in that, include: S1: Construct a convolutional neural network; S2: Construct a training dataset with the desired focal position and the corresponding ideal radiation near-field sampling matrix as training samples; S3: Train a convolutional neural network using the training dataset so that the network learns the mapping from the focal position to the phase distribution of the array output; S4: Input the actual desired focal position into the trained convolutional neural network, output the corresponding array output phase distribution, and configure the phase of each element of the array antenna according to the output phase distribution to achieve near-field focusing.

2. The CNN-based MWPT array antenna radiation near-field focusing control method according to claim 1, characterized in that, The convolutional neural network includes: multiple transposed convolutional layers, multiple convolutional layers, and a batch normalization layer and an activation function sequentially set after each non-output convolutional layer; After normalization in each non-output convolutional layer, Leaky ReLU is used as the activation function, and hyperbolic tangent is used as the activation function in the output convolutional layer.

3. The CNN-based MWPT array antenna radiation near-field focusing control method according to claim 2, characterized in that, The mathematical expression for the batch normalization layer is: In the formula, The feature map is after batch normalization. and These are the weights and biases, respectively. The standardized feature map, This represents the feature map obtained after transpose convolution or convolution operations. For the number of floors, For the first Each output channel weight and These are the mean and standard deviation of the current batch, respectively. It is a constant; The Leaky ReLU expression is: In the formula, For a modified linear unit with leakage, The value of the input activation function, Leakage coefficient; The expression for the hyperbolic tangent function as the activation function is: In the formula, It is the hyperbolic tangent function.

4. The CNN-based MWPT array antenna radiation near-field focusing control method according to claim 1, characterized in that, The construction of the training dataset, which uses the desired focal position and the corresponding ideal radiation near-field sampling matrix as training samples, includes: With the goal of achieving square uniform focusing, an ideal radiation near-field sampling matrix is ​​constructed, the focal center of the ideal radiation near-field sampling matrix is ​​obtained, and different corresponding ideal radiation near-field sampling matrices are obtained by changing the position of the focal center, thus obtaining the training dataset.

5. The CNN-based MWPT array antenna radiation near-field focusing control method according to claim 1, characterized in that, The S3 includes: S301: The output matrix is ​​obtained using a convolutional neural network, and the outgoing phase distribution matrix is ​​obtained based on the output matrix; S302: Based on the near-field radiation field calculation formula, the radiation phase distribution matrix is ​​transformed into the corresponding near-field distribution, and the near-field distribution is discretized to obtain the discrete sampling matrix of the radiation field distribution. S303: Based on the discrete sampling matrix of the radiation field distribution, the desired focal position and the corresponding ideal radiation near-field sampling matrix, a mean squared error loss function is constructed to train the convolutional neural network, and the Adam optimizer is used to update the gradient of the parameters of each layer of the convolutional neural network until the mean squared error loss function converges.

6. The CNN-based MWPT array antenna radiation near-field focusing control method according to claim 5, characterized in that, The expressions for the output matrix and the output phase distribution matrix are as follows: in, The output matrix is ​​obtained by using the hyperbolic tangent function as the activation function. For the first element in the output matrix lines and The values ​​of the array antenna elements in the column, This represents the row number of the array antenna elements. This refers to the number of columns of the array antenna elements. The outgoing phase distribution matrix, The first phase distribution matrix is ​​the first... lines and The output phase of the array antenna elements.

7. The CNN-based MWPT array antenna radiation near-field focusing control method according to claim 6, characterized in that, The formula for calculating the near-field radiation field is as follows: In the formula, Let be the field distribution function of the near-field radiation. For the first lines and The element radiation pattern of the array antenna elements. The pitch angle, It is the azimuth angle. For the first lines and The emitted amplitude of the array antenna elements in the column, For the first lines and Array antenna elements To the observation point distance, The imaginary unit, The wavelength of the emitted wave. For the first lines and The output phase of the array antenna elements of the column, , and The x, y, and y coordinates of the observation point. and For the first lines and The horizontal and vertical coordinates of the array antenna elements in the column.

8. The CNN-based MWPT array antenna radiation near-field focusing control method according to claim 7, characterized in that, The expression for the mean squared error loss function is: In the formula, Let the mean squared error loss function be . For an ideal radiation near-field sampling matrix, The discrete sampling matrix represents the radiation field distribution. The first of the ideal radiation near-field sampling matrix One element, The first discrete sampling matrix of the radiation field distribution One element, The total number of matrix elements. It is the square norm.

Images