Deep learning and driving combined antenna selection method for optimizing capacity and gain of MIMO (Multiple Input Multiple Output) system

Abstract

The invention provides a deep learning driving joint antenna selection method for optimizing the capacity and gain of an MIMO system, and the method comprises the steps: generating a tag related to the optimal communication performance for each channel matrix, training a Shadow-ResNet classifier based on the generated tag, and extracting features from the channel matrixes through employing a plurality of non-linear activation functions; and classifying an input channel matrix by using the trained classifier, and providing an optimal antenna selection scheme.

Description

Technical Field

[0001] The present invention relates to technical fields such as deep learning, communication engineering and machine learning, and in particular to a deep learning driven joint antenna selection method for optimizing MIMO system capacity and gain. Background Art

[0002] With the development of wireless communication technology, especially the emergence of the fifth generation (5G) and sixth generation (6G) mobile communication systems, improving network performance indicators such as transmission rate, spectrum efficiency and latency has become the key. Multiple-input multiple-output (MIMO) systems have significantly improved channel capacity by utilizing technologies such as spatial diversity, but as the number of antennas increases, problems such as hardware cost, thermal management and system complexity also arise. Smart antenna selection strategies have emerged to reduce the number of active antennas to simplify the system structure and reduce maintenance requirements. At the same time, due to the high dynamics and heterogeneity of the wireless communication environment, traditional model-based methods are difficult to cope with complex environmental changes.

[0003] As a powerful tool, deep learning (DL) has shown advantages in processing large amounts of data and capturing complex patterns. In MIMO systems, deep learning is used to optimize the antenna selection process, that is, to select the best combination from multiple possible antenna configurations to maximize channel gain and minimize interference. This not only improves system performance, but also solves the problem of high computational complexity of traditional algorithms. Specifically, researchers have proposed various antenna selection methods based on machine learning and deep learning, including but not limited to the use of convolutional neural networks (CNN), recurrent neural networks (RNN), support vector machines (SVM), K-nearest neighbors (KNN), genetic algorithms (GA), particle swarm optimization (PSO), etc. On the one hand, the above solutions have the problem of excessively high computing power requirements, and on the other hand, their performance also has a lot of room for improvement and improvement. Summary of the invention

[0004] In view of the defects and shortcomings of the existing technology and the actual needs, the present invention takes advantage of the deep learning (DL) solution and proposes an improved ResNet model, Shadow-ResNet. The model adds a Shadow module to the ResNet architecture to optimize the joint selection process of the transmit and receive antennas.

[0005] In the scheme of the present invention, in order to evaluate the communication performance, analysis is performed from two dimensions: channel capacity and channel gain. First, an exhaustive search (ES) algorithm is used to generate two One-Hot coded antenna combination labels associated with the optimal communication performance for each channel matrix. Then, based on the generated labels, a Shadow-ResNet classifier is trained to extract features from the channel matrix using various nonlinear activation functions. Finally, the trained classifier is used to classify the input channel matrix and provide the optimal antenna selection scheme. Experimental results show that this method can achieve near-optimal communication performance.

[0006] The specific technical solutions are as follows: A deep learning-driven joint antenna selection method for optimizing the capacity and gain of MIMO systems: After generating labels related to the optimal communication performance for each channel matrix, a Shadow-ResNet classifier is trained based on the generated labels to extract features from the channel matrix using multiple nonlinear activation functions; the trained classifier is used to classify the input channel matrix to provide the optimal antenna selection solution.

[0007] Furthermore, generating a label associated with the optimal communication performance for each channel matrix specifically includes: using an exhaustive search algorithm to generate two One-Hot coded antenna combination labels corresponding to the maximized channel capacity and gain for each channel matrix.

[0008] Furthermore, the Shadow-ResNet classifier is constructed by adding a Shadow module to the ResNet architecture.

[0009] Furthermore, the Shadow module first converts the channel matrix H into a grayscale image of size 8×8 with 1 channel, and adjusts the input dimension to , where b is the batch size, h is the height of the image, w is the width of the image, and c is the number of channels; the converted image is then divided into four parts according to the number of channels, and the shape of each part is ; Each part is processed by independent linear mapping and activation function to capture the characteristics of the input signal; the activation functions used in the processing include: ReLU, GELU, SELU and Sigmoid; the processed image is connected with the original input along the channel dimension, and the resulting shape is ; Then, resize the image to , to match the input format of the convolutional layer; finally, a 1×1 convolution is applied to the resized image to map the 2c input channels to c output channels, thus generating the output image, corresponding to the new matrix .

[0010] Furthermore, the structure of the Shadow-ResNet classifier is based on ResNet18: first, the preprocessed channel matrix is ​​used as the input of the network; after the input layer, the first convolution layer uses 64 convolution kernels of size 3×3 for convolution and performs batch normalization, Shadow module, and ReLU in sequence; the second to fourth layers use the same convolution structure, including convolution, batch normalization, Shadow module, and ReLU; in the fifth layer, 128 convolution kernels of size 1×1 are used, followed by batch normalization, Shadow module, and ReLU; the sixth and seventh layers use 128 convolution kernels of size 3×3, and the output is processed by batch normalization, Shadow module, and ReLU; the eighth layer uses 256 convolution kernels of size 3×3 and applies downsampling with a stride of 2, followed by batch normalization, Shadow module, and ReLU; the ninth layer uses 256 convolution kernels of size 1×1 to increase the number of channels of the feature map, and applies batch normalization. The tenth and eleventh layers use 256 convolution kernels of size 3×3, followed by batch normalization, Shadow module and ReLU; the twelfth layer uses 512 convolution kernels of size 3×3 and applies downsampling with a stride of 2, followed by batch normalization, Shadow module and ReLU; the thirteenth layer uses 512 convolution kernels of size 1×1, applies batch normalization, Shadow module and ReLU; the fourteenth and fifteenth layers use 512 convolution kernels of size 3×3, followed by batch normalization, Shadow module and ReLU; the sixteenth layer uses adaptive average pooling to aggregate the input feature map to 1×1; the seventeenth layer is used as a flattening layer to convert the aggregated feature map into a one-dimensional vector and input it into the fully connected layer of the eighteenth layer to map the feature vector to the corresponding category of the classification task; the fully connected layer contains 784 units; the fully connected layer is activated by the Softmax function, and the loss function uses classification cross entropy.

[0011] Furthermore, the preprocessing first performs data normalization: the complex-valued channel matrix is ​​converted into a real-valued matrix by calculating the modulus; after obtaining the real-valued data, normalization is performed by subtracting the mean and dividing by the standard deviation; and then labels are generated for the normalized data: an exhaustive search method is used to identify all possible antenna subsets for each channel matrix to find the optimal antenna subset that maximizes the channel capacity and gain, and two One-Hot coded antenna combination labels corresponding to the maximized channel capacity and gain are generated for each channel matrix.

[0012] And, an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the steps of the above method when executing the program.

[0013] A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that the computer program implements the steps of the above method when executed by a processor.

[0014] Compared with the prior art, the present invention and its preferred solution optimize the channel capacity and gain by adopting Shadow-ResNet, showing significant advantages in accuracy, computational efficiency and resource utilization. It can provide support for large-scale MIMO applications. The significant innovations are: (1) Based on deep learning (DL), a Shadow module is proposed to solve the problem of joint transmit and receive antenna selection in MIMO systems. The Shadow module can learn powerful features from the input data to make better selection decisions. This innovative method will help optimize the antenna selection strategy and improve the overall efficiency of the communication system.

[0015] (2) A new antenna selection model based on deep learning is proposed, which is an improved ResNet model called Shadow-ResNet, which is used to automatically extract and select sub-channel features. The Shadow-ResNet model is designed with a unique residual connection structure and deep feature extraction function, which enables it to effectively learn important input channel matrix features related to communication performance. It combines the characteristics of the residual network (ResNet) and the Shadow module designed specifically for MIMO systems, and can achieve efficient antenna selection while ensuring communication performance. Simulation results show that the proposed Shadow-ResNet model can achieve performance close to the optimal capacity and gain.

[0016] (3) A one-hot encoding multi-category label generation algorithm is also proposed. A label generation method based on two criteria, channel capacity maximization and channel gain maximization, is proposed. The residual neural network classifier is trained using large-scale channel matrix data, and the output of the residual neural network is converted into an antenna combination scheme with optimal communication performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: Figure 1 A partial channel matrix diagram is simulated for an embodiment of the present invention; Figure 2 This is a structural diagram of Shadow according to an embodiment of the present invention: Figure 3 The following is a diagram of the Shadow-ResNet network architecture according to an embodiment of the present invention: Figure 4 This is a relationship diagram between the channel capacity loss variance and the signal-to-noise ratio according to an embodiment of the present invention: Figure 5This is a relationship diagram between accuracy and signal-to-noise ratio under different algorithms based on capacity in an embodiment of the present invention: Figure 6 A graph showing the relationship between channel capacity and signal-to-noise ratio under different capacity-based algorithms according to an embodiment of the present invention; Figure 7 A channel gain diagram of different algorithms in gain-based antenna selection according to an embodiment of the present invention; Figure 8 Comparison of channel capacity of MIMO system according to an embodiment of the present invention Figure 1 ; Among them, a: Comparison of channel capacity of MIMO system under the optimal method configuration: (8,8;2,2), (9,9;2,2), and (10,10;2,2); b: Comparison of channel capacity of MIMO system under the suboptimal Shadow-ResNet method configuration: (8,8;2,2), (9,9;2,2), and (10,10;2,2); c: Comparison of channel capacity of MIMO system under the decoupling method configuration: (8,8;2,2), (9,9;2,2), and (10,10;2,2); d: Comparison of channel capacity of MIMO system under the Shadow-ResNet method configuration: (8,8;2,2), (9,9;2,2), and (10,10;2,2): Fig. 9 Comparison of channel capacity of MIMO system according to an embodiment of the present invention Figure 2 ; Among them, a: Comparison of channel capacity of MIMO system under the optimal method configuration: (8,8;2,2), (8,8;3,3), and (8,8;4,4); b: Comparison of channel capacity of MIMO system under the suboptimal Shadow-ResNet method configuration: (8,8;2,2), (8,8;3,3), and (8,8;4,4); c: Comparison of channel capacity of MIMO system under the decoupling method configuration: (8,8;2,2), (8,8;3,3), and (8,8;4,4); d: Comparison of channel capacity of MIMO system under the Shadow-ResNet method configuration: (8,8;2,2), (8,8;3,3), and (8,8;4,4): Fig.10 The following is a comparison diagram of channel gains of MIMO systems configured as (8,8;2,2), (9,9;2,2), and (10,10;2,2) under different algorithms in an embodiment of the present invention: Fig.11 This is a comparison diagram of channel gains of MIMO systems configured as (8,8;2,2), (8,8;3,3), and (8,8;4,4) under different algorithms in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the features and advantages of this patent more obvious and easy to understand, the following embodiments are specifically described in detail as follows: It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0019] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0020] The embodiment of the present invention proposes a joint transmitter-receiver antenna selection model Shadow-ResNet based on deep learning. The Shadow module is used to enhance feature extraction by utilizing various nonlinear activation functions and channel transformations, effectively improving the model's ability to capture complex relationships within the data. In the implementation process of the scheme, labels are first generated based on maximizing two criteria, namely channel capacity and gain, and these labels are implemented using one-hot encoding. Then, the Shadow-ResNet model is trained using large-scale simulated data. Finally, the trained model is used to assign category labels and identify the optimal antenna subset.

[0021] 1.MIMO system First, consider the downlink of a single-user MIMO system. The transmitter sends RF signal to the receiving end, where These signals are received by an RF receive path.

[0022] The transmitter of a single-user MIMO system is equipped with antennas, and the receiver has antennas. The channel matrix of the system is expressed as ,in represents the channel coefficient between the i-th transmitting antenna and the j-th receiving antenna. The received signal can be expressed as follows: (1) in, represents the average signal-to-noise ratio (SNR), Indicates the transmission signal. Indicates the received signal. represents Gaussian white noise. The channel capacity of the system can be expressed as follows: (2) in, The dimension is The identity matrix of Representation Matrix The conjugate transpose of , det() is the determinant calculation symbol.

[0023] Further assume that the transmitting end of the single-user MIMO system is equipped with antennas, and the receiver has antennas, assuming that space-time trellis codes (STTC) are used for space-time coding, the MIMO channel can be regarded as a combination of multiple independent single-input single-output (SISO) sub-channels. Therefore, the equivalent channel gain can be expressed as follows: (3) in, is the channel matrix elements of which represents the channel coefficient between the i-th transmit antenna and the j-th receive antenna.

[0024] In this embodiment, by maximizing the channel capacity ( ) and maximize channel gain ( ) to optimize the communication performance, the objective function can be expressed as: (4) (5) in, and Respectively represent the number of transmitting antennas selected from the transmitting antennas and from The number of receiving antennas selected from the receiving antennas, The dimension is The identity matrix of represents the average signal-to-noise ratio (SNR). B is the sub-channel matrix selected from the original channel matrix, expressed as .

[0025] 2. Generate tags As a preprocessing step, data normalization is first performed: before neural network processing, the complex-valued channel matrix is ​​converted to a real-valued matrix by calculating the modulus. After obtaining the real-valued data, it is normalized by subtracting the mean and dividing by the standard deviation.

[0026] Then, labels are generated for the normalized data: an exhaustive search method (i.e., the multi-class label generation algorithm based on channel capacity and gain described below) is used to identify all possible antenna subsets for each channel matrix to find the optimal antenna subset that maximizes channel capacity and gain. To improve the performance of label generation, this embodiment further uses one-hot encoding, which provides a method for processing discrete features and expands the attribute dimensions of these features to a certain extent.

[0027] To improve the performance of label generation, this embodiment uses a one-hot encoding, which provides a method for processing discrete features and expands the attribute dimensions of these features to a certain extent. In this case, set and In this case, the total number of tags is , and s=784. For a one-hot encoding scheme with s=784 categories, each instance will be represented by a vector containing 784 elements, of which 783 elements are 0 and a single element is 1. The position of 1 in the vector indicates the category to which the instance belongs. The correspondence between antenna subsets and one-hot encodings is shown in Table 1, where each subset uniquely matches a one-hot encoding.

[0028] Table 1 Antenna subset corresponding one-hot encoding

[0029] Algorithm 1 is a multi-class label generation algorithm based on channel capacity and gain designed based on the exhaustive method in this embodiment, where J is the total number of channel matrices. and They represent the channel capacity and channel gain corresponding to different antenna subsets respectively. A represents the label based on the maximization channel capacity standard, and B represents the label based on the maximization channel gain standard. is a specific set of row and column indices in the channel matrix H, representing different transmit and receive antenna combinations.

[0030] Algorithm 1 Multi-category label generation algorithm: enter:

[0031] Output: J one-hot encodings of A and B.

[0032] 3. Shadow module The advantage of ResNet is that it can effectively alleviate the gradient vanishing problem in deep neural networks through residual connections, allowing information to flow more freely between layers. This is especially important in communication signal analysis. In wireless communication environments, multipath propagation and co-channel interference are common problems that degrade signal quality and require efficient allocation of limited spectrum and power resources.

[0033] To meet these challenges, the present invention first proposes the design of Shadow module, which enhances the model's understanding of complex signal environment. It optimizes signal separation and interference suppression algorithms, thereby improving the clarity and reliability of signal reception. Figure 2 Demonstrated the proposed Shadow module.

[0034] First, the channel matrix H is converted into a grayscale image of size 8×8 with 1 channel, and the input dimension is adjusted to , where b is the batch size, h represents the height of the image, w represents the width of the image, and c represents the number of channels. Then, the converted image is divided into four parts according to the number of channels, and the shape of each part is . Each part is processed through independent linear mapping and activation functions to capture the characteristics of the input signal. This approach enhances the model's ability to capture complex nonlinear relationships. The processed results are then concatenated to improve the feature representation, thereby improving the performance of the antenna selection task. During the processing, a variety of activation functions are used. The ReLU activation function helps alleviate the gradient vanishing problem and accelerates convergence during training. The GELU activation function effectively performs random regularization by scaling the input according to its cumulative distribution function (CDF), further enhancing the generalization ability of the model. The SELU activation function has a self-normalization property, which is particularly important in deep networks because it helps maintain the mean and variance of the activation values ​​and promotes faster convergence. At the same time, the Sigmoid function regulates the feature flow, enabling the model to emphasize or weaken the influence of certain features based on the learned representation. Next, the processed image is concatenated with the original input along the channel dimension, and the resulting shape is Then, resize the image to , to match the input format of the convolutional layer. Finally, a 1×1 convolution is applied to the resized image to map the 2c input channels to c output channels, producing the output image, which is then converted into the new matrix .

[0035] 4. Shadow-ResNet Network Architecture Based on the Shadow module, the Shadow-ResNet proposed in this invention is based on the specific architecture of ResNet18, such as Figure 3As shown. First, the preprocessed channel matrix is ​​used as the input of the network. After the input layer, the first convolution layer uses 64 convolution kernels of size 3×3 for convolution to extract basic features. Next, batch normalization is applied to accelerate training and improve the stability of the network. The Shadow module is then used to further enhance the feature extraction capability, and the ReLU activation function is used to introduce nonlinearity, enabling the network to learn more complex mapping relationships. The second to fourth layers use the same convolution structure, including convolution, batch normalization, Shadow module, and ReLU, to further extract deep features. In the fifth layer, 128 convolution kernels of size 1×1 are used, followed by batch normalization, Shadow module, and ReLU. The sixth and seventh layers use 128 convolution kernels of size 3×3, and the output is processed by batch normalization, Shadow module, and ReLU. The eighth layer uses 256 convolution kernels of size 3×3 and applies a stride of 2 to achieve downsampling, followed by batch normalization, Shadow module, and ReLU. The ninth layer uses 256 1×1 convolution kernels to increase the number of channels of the feature map, and applies batch normalization, Shadow module, and ReLU. The tenth and eleventh layers use 256 3×3 convolution kernels, followed by batch normalization, Shadow module, and ReLU. The twelfth layer uses 512 3×3 convolution kernels and applies downsampling with a stride of 2, followed by batch normalization, Shadow module, and ReLU. The thirteenth layer uses 512 1×1 convolution kernels, applies batch normalization, Shadow module, and ReLU. Similarly, the fourteenth and fifteenth layers use 512 3×3 convolution kernels, followed by batch normalization, Shadow module, and ReLU. Finally, the sixteenth layer uses adaptive average pooling to aggregate the input feature map to 1×1 to achieve global feature aggregation. The 17th layer is a flattening layer, which converts the pooled feature map into a one-dimensional vector, which is then input into the 18th fully connected layer to map the feature vector to the corresponding category of the classification task. The fully connected layer contains 784 units. The Softmax function activates the fully connected layer, and the loss function uses the classification cross entropy.

[0036] The solution of the present invention is tested and verified by specific test examples as follows: Since real antenna data is very expensive, simulated data is used, where each element of the channel matrix is ​​generated by a complex Gaussian random variable with zero mean and unit variance. The size of the channel matrix sample set is 200,000. The signal-to-noise ratio (SNR) is set to 10. The number of transmit and receive antennas is 8. Two specific transmit and receive antennas are selected in the simulation. The generated simulated partial channel matrix example is as follows: Figure 1In the experiments of this embodiment, the focus is on evaluating the effectiveness of the proposed method in achieving two goals: maximizing channel capacity and maximizing channel gain.

[0037] To verify the performance of Shadow-ResNet, this embodiment selects six classic deep learning (DL) models: ResNet, LeNet, RNN, VGG, AlexNet and LSTM. Traditional machine learning (ML) models KNN and SVM are also added for comparison. For KNN, Euclidean distance is used, and 5-fold cross validation is used to select the best value of parameter K. SVM uses radial basis function (RBF) kernel, and the regularization parameter is set to 1. The parameters of other DL models are shown in Table 2.

[0038] Table 2 Network setting parameters corresponding to different methods

[0039] Antenna selection based on capacity This embodiment compares the differences between the subsets selected by various algorithms and the best subset selected by exhaustive search (ES), and calculates the accuracy. Table 3 shows the antenna subset selection accuracy of different capacity-based algorithms. Specifically, the accuracies of Shadow-ResNet, ResNet, LeNet, RNN, VGG, AlexNet, LSTM, KNN, and SVM are 66.50%, 61.73%, 34.10%, 34.14%, 21.59%, 4.69%, 33.38%, 0.98%, and 4.45%, respectively. Shadow-ResNet has the highest accuracy of 66.50%, indicating that it effectively extracts and utilizes the high-dimensional features of antenna signals. This highlights its great potential in optimizing antenna selection. Figure 4 The relationship between the variance of channel capacity loss and the signal-to-noise ratio (SNR) is shown. Figure 5 The relationship between the accuracy and signal-to-noise ratio of different methods is shown. It can be observed that Shadow-ResNet is more robust to changes in signal-to-noise ratio. Figure 6 In , the relationship between channel capacity and signal-to-noise ratio (SNR) is shown.

[0040] Table 3 Comparison of various indicators of different methods based on capacity antenna:

[0041] Gain-based Antenna Selection,Table 4 shows the comparative analysis of various performance indicators when,different antenna selection methods based on channel gain are applied.,In terms of accuracy, Shadow-ResNet, ResNet, LeNet, RNN, VGG, AlexNet, LSTM,,KNN, and SVM achieved 63.50%, 61.70%, 33.81%, 34.14%, 34.12%, 3.70%, 33.98%, 1.96%, and 51.99%, respectively.,Compared to other algorithms, Shadow-ResNet shows significantly higher,accuracy on this best subset identification problem, highlighting its effectiveness,and strong feature extraction capabilities for this specific task.,Table 5 shows the comparative analysis of various performance indicators when,different antenna selection methods based on channel gain are applied. In terms of accuracy, Shadow-ResNet, ResNet, LeNet, RNN, VGG, AlexNet, LSTM, KNN, and SVM achieved 63.50%, 61.70%, 33.81%, 34.14%, 34.12%, 3.70%, 33.98%, 1.96%, and 51.99%, respectively. Compared with other algorithms, Shadow-ResNet showed significantly higher accuracy on this best subset identification problem, highlighting its effectiveness and strong feature extraction capabilities for this specific task. Figure 7 As shown in Figure 2, the channel gain comparison results between different algorithms show that the channel gain of Shadow-ResNet is close to that of the exhaustive search (ES) algorithm. Shadow-ResNet maintains a high predicted subchannel gain while minimizing the channel gain loss, which shows its potential in the design of efficient communication systems.

[0042] Table 4 compares various indicators of different methods based on gain antenna:

[0043] In practical applications, the model is trained offline and the trained model is used for real-time antenna selection. One of the main challenges facing MIMO systems is the long time required for real-time selection, which consumes a lot of computing resources and introduces additional delays. This embodiment compares the training and testing time of Shadow-ResNet with other algorithms. The data set includes 160,000 samples for training and 40,000 samples for testing. Table 5 shows the training and testing time of each method based on channel capacity. As shown in Table 6, the results of the training and testing time of each method based on channel gain show that Shadow-ResNet has the smallest testing time.

[0044] Table 5 compares the training and testing time of different channel capacity algorithms: method Training time Testing time Shadow-ResNet 347.7s 0.0802s ResNet 276.8s 0.0830s LetNet 800.1s 0.0794s RNN 395.4s 0.0793s VGG 1024.4s 0.0787s AlexNet 410.9s 0.0798s LSTM 87.0s 0.0801s SVM 1247.8s 3956.0s KNN 0.021s 100.8s Table 6 compares the training and testing time of different channel gain algorithms: method Training time Testing time Shadow-ResNet 349.4s 0.0752s ResNet 277.5s 0.0788s LetNet 779.4s 0.0827s RNN 392.3s 0.0804s VGG 389.4s 0.0805s AlexNet 412.3s 0.0796s LSTM 86.4s 0.0808s SVM 655.8s 3952.0s KNN 0.020s 101.7s This embodiment also increases the number of transmitting and receiving antennas to evaluate the robustness of Shadow-ResNet, such as Figure 8 On the other hand, Fig.10 As mentioned above, for antenna selection based on maximizing channel gain, the channel gain of all algorithms decreases as the number of antennas increases. To further explore the performance of Shadow-ResNet, the number of RF chains is increased by 3 and 4 respectively. Fig. 9 As shown in Figure 2, in the antenna selection based on channel capacity, the channel capacity of all algorithms shows an increasing trend as the number of RF chains increases. Similarly, regarding channel gain, Fig.11 It also shows an upward trend. This is because the added RF chains provide more spatial diversity and spatial multiplexing. To further explore the performance of Shadow-ResNet, the number of RF chains is increased by 3 and 4 respectively. Fig. 9 As shown in Figure 2, in the antenna selection based on channel capacity, the channel capacity of all algorithms shows an increasing trend as the number of RF chains increases. Similarly, regarding channel gain, Fig.11 It also shows an upward trend. This is because the added RF chains provide more spatial diversity and spatial multiplexing.

[0045] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0046] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by the processor to execute the above method. The storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0047] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

[0049] The present invention is not limited to the above-mentioned optimal implementation mode. Anyone can derive other various forms of deep learning driven joint antenna selection methods for optimizing MIMO system capacity and gain under the inspiration of the present invention. All equal changes and modifications made according to the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A deep learning driven joint antenna selection method for optimizing MIMO system capacity and gain, characterized by: After generating labels associated with the optimal communication performance for each channel matrix, a Shadow-ResNet classifier is trained based on the generated labels to extract features from the channel matrix using a variety of nonlinear activation functions; Use the trained classifier to classify the input channel matrix and provide the best antenna selection solution; The generating of a label related to the optimal communication performance for each channel matrix specifically includes: using an exhaustive search algorithm to generate two One-Hot coded antenna combination labels corresponding to the maximized channel capacity and gain for each channel matrix.

2. The deep learning driven joint antenna selection method for optimizing MIMO system capacity and gain according to claim 1, characterized in that: The Shadow-ResNet classifier is constructed by adding a Shadow module to the ResNet architecture.

3. The deep learning driven joint antenna selection method for optimizing MIMO system capacity and gain according to claim 2, characterized in that: The Shadow module first converts the channel matrix H into a grayscale image of size 8×8 with 1 channel, and adjusts the input dimension to , where b is the batch size, h is the height of the image, w is the width of the image, and c is the number of channels; the converted image is then divided into four parts according to the number of channels, and the shape of each part is ; Each part is processed by independent linear mapping and activation function to capture the input signal characteristics; The activation functions used in the processing include: ReLU, GELU, SELU and Sigmoid; the processed image is concatenated with the original input along the channel dimension, and the resulting shape is ; Then, resize the image to , to match the input format of the convolutional layer; finally, a 1×1 convolution is applied to the resized image to map the 2c input channels to c output channels, thus generating the output image, corresponding to the new matrix .

4. The deep learning driven joint antenna selection method for optimizing MIMO system capacity and gain according to claim 2, characterized in that: The structure of the Shadow-ResNet classifier is based on ResNet18: first, the preprocessed channel matrix is ​​used as the input of the network; after the input layer, the first convolution layer uses 64 convolution kernels of size 3×3 for convolution and performs batch normalization, Shadow module, and ReLU in sequence; the second to fourth layers use the same convolution structure, including convolution, batch normalization, Shadow module, and ReLU; in the fifth layer, 128 convolution kernels of size 1×1 are used, followed by batch normalization, Shadow module, and ReLU; the sixth and seventh layers use 128 convolution kernels of size 3×3, and the output is processed by batch normalization, Shadow module, and ReLU; the eighth layer uses 256 convolution kernels of size 3×3 and applies downsampling with a stride of 2, followed by batch normalization, Shadow module, and ReLU; the ninth layer uses 256 convolution kernels of size 1×1 to increase the number of channels of the feature map, applies batch normalization, Shadow module and ReLU; the tenth and eleventh layers use 256 convolution kernels of size 3×3, followed by batch normalization, Shadow module and ReLU; the twelfth layer uses 512 convolution kernels of size 3×3 and applies downsampling with a stride of 2, followed by batch normalization, Shadow module and ReLU; the thirteenth layer uses 512 convolution kernels of size 1×1, applies batch normalization, Shadow module and ReLU; the fourteenth and fifteenth layers use 512 convolution kernels of size 3×3, followed by batch normalization, Shadow module and ReLU; the sixteenth layer uses adaptive average pooling to aggregate the input feature map to 1×1; the seventeenth layer is used as a flattening layer to convert the aggregated feature map into a one-dimensional vector and input it into the fully connected layer of the eighteenth layer to map the feature vector to the corresponding category of the classification task; the fully connected layer contains 784 units; the fully connected layer is activated by the Softmax function, and the loss function uses classification cross entropy.

5. The deep learning driven joint antenna selection method for optimizing MIMO system capacity and gain according to claim 4, characterized in that: The preprocessing first performs data normalization: the complex-valued channel matrix is ​​converted into a real-valued matrix by calculating the modulus; after obtaining the real-valued data, normalization is performed by subtracting the mean and dividing by the standard deviation; and then labels are generated for the normalized data: an exhaustive search method is used to identify all possible antenna subsets of each channel matrix to find the optimal antenna subset that maximizes the channel capacity and gain, and two One-Hot coded antenna combination labels corresponding to the maximized channel capacity and gain are generated for each channel matrix.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the deep learning driven joint antenna selection method for optimizing MIMO system capacity and gain are implemented as described in any one of claims 1-5.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the deep learning driven joint antenna selection method for optimizing MIMO system capacity and gain are implemented as described in any one of claims 1-5.

Images