Metasurface multi-beam rapid generation method based on semi-supervision

Abstract

The invention discloses a metasurface multi-beam rapid generation method based on semi-supervision, and the operation process of the method comprises the following steps: 1, introducing an attention mechanism based on a convolutional neural network, and designing a cross entropy loss function; step 2, training the model by using a large number of samples obtained by caliber field superposition, migrating trained model parameters to a new model, and training based on the samples optimized by a traditional algorithm; step 3, semi-supervised learning is introduced, a model obtained through initial training is utilized to label unlabeled data samples, labeled sample data is added to label data samples, then the network model is retrained, the process is continuously repeated, and the number of the samples is iteratively increased; and 4, inputting required directional diagram data to obtain a corresponding coding matrix. Based on the method, a multi-beam coding problem is converted into a multi-label classification problem, the condition of repeated traversal in a codebook set in a traditional algorithm is avoided, and the time complexity is reduced.

Description

Technical Field

[0001] The invention relates to a semi-supervised hypersurface multi-beam rapid generation method in a multi-user scenario, and belongs to the technical field of wireless communications. Background Art

[0002] Reconfigurable Intelligent Surface (RIS) is a digitally controllable metasurface composed of a large number of reflective or transmissive elements, and its amplitude and phase characteristics can be controlled in real time by electromagnetic components on the control unit. Specifically, RIS can be deployed between the base station and the user to build an additional reflective link to form a virtual LOS link. By properly configuring the RIS units, a highly reflective beam can be formed to improve the link gain, and the coverage blind area with poor signal quality can get better signal quality, thereby improving the capacity of the communication system. In addition, compared with traditional RF systems, RIS saves devices such as phase shifters, and has extremely low hardware costs and energy consumption, which is conducive to large-scale deployment. Therefore, intelligent metasurfaces are widely used in various wireless systems and applications.

[0003] However, most of the existing smart metasurfaces only consider single-link communication scenarios and are still in the stage of theoretical modeling. However, in the actual application of large-scale RIS, there will be problems such as an increase in the number of users and a diversification of user states. The existing single-link system cannot simulate the actual use situation, nor can it meet the actual needs of large-scale RIS. In order to meet the needs of large-scale RIS in actual application scenarios, it is necessary to implement a multi-user channel information feedback mechanism, solve the problem of fast switching of multi-beam codebooks, and further consider the scenario of dynamic user movement.

[0004] In order to solve the problem of fast switching of multi-beam codebooks, we proposed a semi-supervised metasurface multi-beam fast generation method, designed multiple training processes, and used semi-supervised learning to provide a large number of training samples. Finally, the multi-beam codebook can be quickly generated according to the required directional diagram. Compared with traditional optimization algorithms, it reduces the time complexity and is more suitable for the actual deployment of RIS. Summary of the invention

[0005] Aiming at the problem that the existing RIS multi-beam codebook generation method has slow generation speed in multi-user scenarios, the present invention provides a semi-supervised hypersurface multi-beam fast generation method, which reduces the time complexity by using deep learning, avoids the situation of concentrated large-scale traversal in the codebook space, and improves the practicality of RIS.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A semi-supervised metasurface multi-beam rapid generation method introduces deep learning, uses the phase coding matrix obtained by the traditional optimization algorithm as the target value sample to train the neural network, and realizes the mapping relationship between the multi-beam radiation pattern and the phase coding matrix; introduces semi-supervised learning, obtains samples by labeling the unlabeled data by itself, and expands the sample data.

[0008] The model training required for this beamforming method includes the following steps:

[0009] Step 1: Improve the convolutional neural network to make it more suitable for multi-label learning and output. Introduce the attention mechanism, add the SE attention module and residual structure, design the cross entropy loss function, consider the phases represented by different codes, and introduce sample weights to represent the weights of different samples.

[0010] Step 2: Use a large number of samples obtained by aperture field superposition to train the model, transfer the trained model parameters to the new model, and then train it based on the samples optimized by the traditional algorithm.

[0011] Step three: introduce semi-supervised learning, use the model obtained from the initial training to label the unlabeled data samples, add the labeled sample data to the labeled data samples, retrain the network model, and repeat the process continuously.

[0012] Step 4: After training, input the required directional pattern data into the model to obtain the corresponding encoding matrix.

[0013] Among them, the residual structure is introduced in step 1, which is constructed by connecting three convolutional layers in series: 1×1, 3×3, and 1×1, with the SE attention module connected in parallel in the middle, and finally added to the output. At the same time, considering the mapping relationship of the actual phase, the network uses 3-bit quantized phase encoding to represent 8 phase values ​​from 0° to 315°, with a difference of 45° between each phase value.

[0014] Among them, in step 2, the transfer learning method is used to train the initial model, which specifically includes:

[0015] A preliminary sample set is obtained based on the aperture field superposition method, and a preliminary model is obtained by training with the preliminary sample set.

[0016] The prepared sample set is optimized using traditional optimization algorithms to obtain the initial sample set, and the convolutional layer parameters of the prepared model are extracted and transferred to the initial model. Finally, the initial model is trained using the initial sample set until the loss function and accuracy function converge.

[0017] The general process of step three is to use the trained initial model to label the unlabeled data samples, and then add the labeled sample data back to the labeled data samples, and repeat the process until convergence or the expected effect is achieved. Specifically, the process uses the three-training method, that is, three models with different network structures are obtained from step two, and the same unlabeled sample is labeled respectively. If two models have the same label, they will enter the confidence screening. After passing the confidence screening, they can enter the labeled data set and enter the self-training iteration process.

[0018] Beneficial effects: Compared with the prior art, the present invention adopts the above technical solution and has the following advantages:

[0019] The present invention provides a method for fast generation of multi-beams based on a semi-supervised metasurface, introduces deep learning, uses the phase coding matrix obtained by the traditional optimization algorithm as a target value sample for training a neural network, and transforms the multi-beam prediction problem into a multi-label classification problem. First, this method mainly solves the problem of slow generation and poor accuracy of multi-beam codebooks in traditional optimization algorithms, reduces time complexity, and is more suitable for practical application scenarios; secondly, compared with the prior art, this method transforms the beam generation problem of multi-user multi-beams under the RIS-assisted system into a multi-label classification problem, transforming it from the traditional antenna optimization field into a neural network optimization problem, and opening up new ideas for technical optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0021] Figure 1 It is the main step to realize the fast multi-beam generation method based on semi-supervision.

[0022] Figure 2 This is a training flowchart of the semi-supervised multi-beam rapid generation method.

[0023] Figure 3 It is a confidence screening flow chart. DETAILED DESCRIPTION

[0024] In order to more clearly and in detail illustrate the embodiments of the present invention or the technical solutions in the prior art, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and embodiments.

[0025] It should be noted that the described embodiments are only some embodiments of the present invention, not all embodiments, and are only used to explain the present application, not to limit the present application.

[0026] The purpose of the present invention is to provide a semi-supervised metasurface multi-beam fast generation method, which uses the phase coding matrix obtained by the traditional optimization algorithm as a target value sample for training the neural network and converts the multi-beam coding problem into a multi-label classification problem.

[0027] In order to make the above-mentioned objects, features and advantages of the present invention more understandable, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0028] The main steps to implement the metasurface multi-beam rapid generation method are as follows: Figure 1 As shown, it includes the following steps:

[0029] Step 1: Improve the convolutional neural network to make it more suitable for multi-label learning and output. Introduce the attention mechanism, add the SE attention module and residual structure, design the cross entropy loss function, consider the phases represented by different codes, and introduce sample weights to represent the weights of different samples.

[0030] Step 2: Use a large number of samples obtained by aperture field superposition to train the model, transfer the trained model parameters to the new model, and then train it based on the samples optimized by the traditional algorithm.

[0031] Step three: introduce semi-supervised learning, use the model obtained from the initial training to label the unlabeled data samples, add the labeled sample data to the labeled data samples, retrain the network model, and repeat the process continuously.

[0032] Step 4: After training, input the required directional pattern data into the model to obtain the corresponding encoding matrix.

[0033] Among them, step 1 introduces the SE attention module and residual structure based on the convolutional neural network, where the convolutional layer can be expressed by the formula;

[0034]

[0035] The attention module is used to distinguish the weights of different channels. Specifically, the unit phases at different positions in the encoding matrix may have different influencing factors.

[0036] The entire module can be divided into two calculation parts. The first part is to perform a pooling process on the output of the convolution layer. The output data of the convolution layer can be expressed as (H, W, C). After pooling, a variable (1, 1, C) is obtained. This process uses global average pooling, which can distribute different weights to different channels. Formula F sq (·) is expressed as:

[0037]

[0038] The second part is implemented using two fully connected layers. First, the first fully connected layer is used for dimensionality reduction, then the second fully connected layer is used for dimensionality increase after the ReLu activation function is passed, and finally the activation function Si gmoid is linked. The formula is expressed as:

[0039] S=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0040] Finally, the obtained S vector is multiplied by the original feature map U, which is equivalent to establishing the connection between the attention module and the original channel. By giving different weights to different channels, the function of obtaining more important information is achieved. The expression is as follows:

[0041] x′=F scale (u c ,s c )=s c u c

[0042] The residual structure is constructed by connecting three convolutional layers in series: 1×1, 3×3, and 1×1, with the SE attention module connected in parallel in the middle, and then adding the output. Considering the mapping relationship of the actual phase, the network uses 3-bit quantized phase encoding to represent 8 phase values ​​from 0° to 315°, with a difference of 45° between each phase value.

[0043] The loss function of the model adopts the cross entropy loss function, and its formula is as follows:

[0044]

[0045] Among them, o[i] represents the model prediction value, t[i] represents the label value, and w represents the weight of different samples. Through w, a weight can be added to the model output.

[0046] In step 2, the transfer learning method is used to train the initial model, including:

[0047] Based on the aperture field superposition method, a preliminary sample set Data is obtained. s The sample set is divided into a training set and a test set in a ratio of 9:1 and preprocessed. The training set is used to train the preparatory model Model s .

[0048] Use traditional optimization algorithms to prepare sample set Data s Optimize to get the initial sample set Data t , and the convolutional layer parameters W of the prepared model s Extracted as the initial model Model t The initial parameters of .

[0049] Finally, use the initial sample set Data t The initial model is trained until the loss function converges.

[0050] Step 3 introduces semi-supervised learning to solve the problem of insufficient sample quantity and insufficient precision. It mainly uses the trained model to label the unlabeled samples, adds the labeled sample data to the sample set and retrains the network model. Its overall architecture is as follows: Figure 2 shown.

[0051] The three-training method is used in this process. Based on the training in step 2, three models with different network parameters are obtained. The three models are randomly divided into training sets and test sets during the initial training to avoid model consistency. At the same time, the residual structures also have certain differences.

[0052] When training unlabeled samples, the three models have different labeling strategies due to the difference in parameters. For the same sample data, the voting method is used to label the unlabeled samples. If two models have the same labeling for the same sample, the next step of confidence screening is entered.

[0053] The label samples of the three network models are:

[0054]

[0055] Among them, T1, T2, and T3 are the labeled sample sets corresponding to different models Model1, Model2, and Model3 respectively, and then unlabeled samples are used for annotation training:

[0056] U=(Ux1,Ux2,…,Ux m )

[0057] After a series of confidence screening, the annotated pseudo-label data is obtained:

[0058] F0={(Ux2,F0y2),(Ux3Fx6)5,…(Ux f F 0f )}

[0059] Where f≤m, the pseudo-label sample data and the existing label data T1 are used as training data sets to self-train Model1, and the unlabeled samples are continuously predicted and labeled to obtain labeled samples, for example:

[0060] F i ={(Ux7,F i y7),(Ux 15 ,F i y 15 ),…,(Ux fi ,Fi y fi )}

[0061] By continuously adding such pseudo-label data to the training samples and iterating the self-training until convergence, the output model Model1 can be obtained.

[0062] The confidence screening process is very important and is the most critical part of the entire semi-supervised process. The screening process is shown in Figure 3 .

[0063] Confidence screening ensures the quality of training samples and directly affects the accuracy of the training model. The quality and quantity of samples can be adjusted by adjusting the confidence screening criteria. The entire process is divided into two parts.

[0064] The first part uses numerical calculation to analyze the coding matrix output by the model to determine whether the main lobe of the directional pattern corresponding to the coding is accurate. If the phase difference is within the error range, it can pass.

[0065] The second part uses numerical calculation to analyze the sidelobe value of the directional pattern corresponding to the code. If it is below the set threshold, it will pass.

[0066] The numerical calculation uses two-dimensional inverse Fourier transform to quickly calculate the direction map, and the calculation formula is:

[0067]

[0068] In the calculation, the wave number is:

[0069]

[0070] The calculation formulas for the path and phase of the incident wave are expressed as:

[0071]

[0072] Similarly, the phase calculation formula of the reflected wave is:

[0073]

[0074] After obtaining the directional pattern matrix E(u,v), we know that the maximum coordinate of the regional matrix represents the pointing angle of the beam, so we can first calculate the expected beam pointing direction of the regional matrix:

[0075]

[0076] Since E(u,v) is a multi-beam pattern, it can be Search within a certain range of the area, and find the maximum value coordinates in the area:

[0077] (u i,v i )=find(E i (u,v)=max(E i (u,v)))

[0078] If the expected beam pointing direction is consistent with the actual beam pointing direction, it proves that the pointing directions of each beam are correct and the coding matrix can be adopted.

[0079] The implementation of the semi-supervised metasurface multi-beam rapid generation method proposed above can quickly generate the phase coding matrix of RIS according to the required pattern features. Compared with the traditional optimization algorithm, it reduces the time complexity, avoids repeated traversal in the sample set space, and improves the applicability in practical scenarios.

[0080] The above description in conjunction with the accompanying drawings describes the implementation modes of the present invention in detail, but the present application is not limited to the above specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present application, ordinary technicians in this field can make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A semi-supervised hypersurface multi-beam rapid generation method, characterized in that: Introducing deep learning, the phase encoding matrix obtained by the traditional optimization algorithm is used as the target value sample to train the neural network; After the initial training, a large number of unlabeled samples are used for self-training; a confidence function is introduced to screen the unlabeled samples.

2. The method according to claim 1, characterized in that The multi-beam pattern is taken as input to realize the mapping relationship between the multi-beam pattern and the phase coding matrix, and the multi-beam coding problem is converted into a multi-label classification problem.

3. According to the semi-supervised hypersurface multi-beam rapid generation method of claim 1, it is characterized in that: The following steps are involved: Step 1: Improve the model based on convolutional neural network, introduce attention mechanism, and design cross entropy loss function; Step 2: Use a large number of samples obtained by aperture field superposition to train the model, transfer the trained model parameters to the new model, and then train it based on the samples optimized by the traditional algorithm; Step 3: Introduce semi-supervised learning, use the model obtained from the initial training to label the unlabeled data samples, add the labeled sample data to the labeled data samples and retrain the network model, and repeat this process to iteratively expand the number of samples; Step 4: Input the required directional pattern data into the model to obtain the corresponding coding matrix.

4. The method for rapid generation of multi-beams based on semi-supervised metasurface according to claim 3, characterized in that: The neural network model undergoes two different training processes. The first process is initial training, which uses a large number of labeled samples for initial training; the second process is iterative self-training, which uses the voting method to label unlabeled samples and uses the expanded labeled samples to self-train the model.

5. The method according to claim 4, characterized in that Transfer learning is used to generate initial training samples. The phase coding matrix obtained based on the traditional optimization algorithm can be used to obtain a large number of initial labeled samples for training using the aperture superposition method.

6. The method for rapid generation of multi-beams based on semi-supervised metasurface according to claim 4, characterized in that: A confidence function is designed to screen pseudo-label samples. The screened samples are added to the labeled data samples and trained iteratively.

Images