A hybrid meta-learning based cross-scenario robust beamforming design method

Abstract

The application discloses a kind of robustness beamforming design methods of cross scene based on hybrid meta-learning.The method selects, mixes multiple scene multiple datasets, and generates action group and stop group with certain sampling means, then initializes trainer, and generates the preset value of beamforming matrix.Training, action group is input into the trainer group based on the training of the Transformer architecture design, first based on the gradient of preset value on action group to optimize it, then back propagation optimization trainer neural network parameters, reach stop condition after saving trainer as pre-training model.In application, pre-training model is deployed according to the channel sample in actual scene, then freeze trainer, only carry out forward propagation, obtain beamforming matrix.Compared with traditional scheme, the method proposed in the application can obtain more channel characteristics during pre-learning and save training resources, adapt faster during deployment, and provide higher communication efficiency when online.

Description

Technical Field

[0001] This invention relates to the field of intelligent wireless communication technology, and in particular to a robust beamforming design method for cross-scenario applications based on hybrid learning. Background Technology

[0002] Beamforming technology has been widely used in the field of wireless communication technology in recent years and plays an important role in 5G and future 6G communications. However, traditional beamforming techniques suffer from high complexity and computational cost, and current beamforming techniques based on traditional neural networks cannot be well applied to various complex scenarios, often experiencing performance degradation due to changes in real-world environments and datasets. Therefore, for beamforming techniques utilizing machine learning, better optimization schemes are needed to improve the technology's performance and scenario adaptability.

[0003] In multi-user communication environments, the complexity of the beamforming problem increases significantly. Beamforming algorithms need to make trade-offs among users to achieve multi-objective optimization, such as adjusting user weights, reducing mutual interference, and increasing throughput. These multi-dimensional optimization objectives often greatly increase the complexity of the optimization function, making the problem a non-convex optimization problem. These factors together make it difficult to directly apply traditional convex optimization methods. Therefore, beamforming optimization in multi-user scenarios is a challenging problem with high research value. Summary of the Invention

[0004] This invention provides a robust beamforming design method based on hybrid learning across scenarios, which can improve the adaptability of beamforming optimization methods to more scenarios and enhance the communication efficiency and user experience of wireless communication in scenarios with a large number of users.

[0005] The systems and methods disclosed in this invention have several innovative aspects, but no single aspect, perspective, or part is fully responsible for the desired properties disclosed in this invention.

[0006] The objective of this invention is achieved through the following technical solution: a robust beamforming design method for cross-scenario applications based on hybrid element learning, comprising the following steps:

[0007] Step 1, Data Acquisition: Select s scenarios from multiple datasets, collect data in each scenario, and collect M data in each scenario. s A collection containing N s Channel data from individual users forms the channel data source.

[0008] Step 2, Constructing the beamforming training dataset: Mix the channel data sources, perform k samplings, and generate an action group containing k channel data sources. and includes k q Stop group of channel data sources

[0009] Step 3, Construct the neural network: This network includes a transmitter trainer. Receiver trainer and weight trainer and the corresponding backpropagation module;

[0010] Step 4, train the neural network: initialize the parameters of the trainer neural network, input the action group into each trainer in batches for forward propagation, use the output as the difference of the preset value and update the preset value, use the updated value as the output Q, J, v of each trainer and combine them into an intermediate result, input it into the loss function, backpropagate on this basis, and repeat p times.

[0011] Step 5, evaluate the neural network: perform forward propagation on the stopping group, input the output of each trainer into the sum rate calculation function, evaluate the current training effect, and compare it with the historical training values ​​of the current dataset. If the training effect no longer improves, end the training and save the model and trainer obtained at the end as the application model.

[0012] Step 6: Online fine-tuning. Collect several channel data in the actual use scenario as a fine-tuning group, train the application model on the fine-tuning group, and freeze the parameters of the obtained trainer neural network as the final model to be used in the scenario.

[0013] Step 7, Practical Application: In actual use scenarios, acquire the user channel, input the acquired channel into the loss function, and input the result of the loss function into the model obtained in Step 6 for forward propagation. Use the output as a beamforming vector.

[0014] Furthermore, in step 1, there are no fewer than two datasets, and each dataset has at least one scenario selected and added to the channel data source. The collected data includes the channel matrix and path loss, and the channel matrix needs to be pre-normalized.

[0015] Furthermore, the number of samplings k in step 2 satisfies the following formula: k = s·M s ·N s The sampling method is sampling with replacement.

[0016] Furthermore, in step 2, the channel data source stops group. It is generated in the following ways:

[0017]

[0018] In this case, the elements of the stopping group are unique, and the expected number of elements is close to the value calculated by the following formula:

[0019] Where e≈2.718 is the natural logarithm;

[0020] Count the elements in the obtained stopping group. If the number of elements obtained is equal to the number of elements in the stopping group, then... If the deviation exceeds the preset deviation threshold, the action group and stop group will be generated again.

[0021] Furthermore, in step 4, the training of each trainer uses the inverse of the sum rate as a reference, and the specific formula for the sum rate is:

[0022]

[0023] L = -S E

[0024] Where L represents the loss function, S E The values ​​represent the sum and rate, n represents the noise amplitude, Q, J, and v represent the outputs of the weighted trainer, receiver trainer, and transmitter trainer, respectively, and h... T This represents the transpose of the channel matrix from the base station to a specific user, with the subscript indicating the user number corresponding to each output element. The noise used is derived based on the following formula:

[0025] n 2 =σ 2 ·L path

[0026] Where L path For path loss, σ 2 Let V be the variance of the circular complex Gaussian noise.

[0027] Furthermore, the trainer is designed based on the Transformer architecture; during training, the trainer... and First, initialize the weights, the transmitter matrix, and the receiver matrix to initial weights Q0, the transmitter matrix v0, and the receiver matrix J0, respectively, as preset values ​​D. Then, input the preset values ​​into the gradient layer to obtain its gradient g with respect to the loss function. r Then g r The input to the training layer is used to extract features and transform them into a dense vector K, with unique positional information added. The training layer then inputs K into an association layer containing an array of association elements. The association elements in the association layer decompose and analyze the features of different aspects of K, and input the analysis results into a deassociation layer. The deassociation layer calculates the output. The output of the deassociation layer is used as the difference of a preset value, which is added to the preset value, and the result is used as the output of the corresponding training layer.

[0028] Further, in step 4, the specific training process is as follows: initialize the parameters of each trainer neural network, and randomly generate initial weights Q0, initial transmitting matrix v0, and initial receiving matrix J0 as preset values; input the preset values ​​into the loss function, calculate the value of the loss function, and find the gradient of the value with respect to Q0, v0, and J0; input the gradient into each trainer for forward propagation to obtain the outputs Q, v, and J of each trainer; input Q, v, and J into the loss function, calculate the gradient of the value of the loss function with respect to Q, v, and J, and perform backpropagation to update the parameters of the trainer neural network.

[0029] Further, in step 4, during the first backpropagation, the parameter changes of the trainer's correlation layer and discorrelation layer (hereinafter collectively referred to as the hidden layer) are r times the gradient of the sum and rate; during the i-th backpropagation, the parameter changes of the trainer's hidden layer are r times the gradient of the sum and rate. times;

[0030] Here, r is a hyperparameter, specified by the trainer, and its specific value is not exactly the same for the three trainers.

[0031] Furthermore, in step 5, the evaluation criteria for training effectiveness are as follows: input the channel portion of all data from the stopping group sequentially, calculate the sum rate when the signal-to-noise ratio is fixed; take the average sum rate obtained from the stopping group samples as the final evaluation result; compare the final evaluation result with the historical results, and if the sum rate no longer shows an upward trend, then training stops.

[0032] Further, in step 6, the online fine-tuning specifically involves: collecting several channel data points as a fine-tuning group in the actual usage scenario, and training the neural network trained using the aforementioned method on the fine-tuning group a small number of times; in step 7, the actual application specifically involves: fixing the parameters of each trainer's neural network, and randomly generating initial weights Q0, initial transmitter matrix v0, and initial receiver matrix J0 as preset values; inputting the preset values ​​into the loss function, calculating the value of the loss function, and obtaining the gradient of the value with respect to Q0, v0, and J0; inputting the gradients into each trainer for forward propagation to obtain the outputs Q, v, and J of each trainer, which are then used as the design results of the weights, transmitter, and receiver, respectively, for application.

[0033] The beneficial effects of this invention are:

[0034] This invention generates channel data sources containing more scene features through dataset generation and mixing methods, enabling models to learn richer beamforming features better. This is beneficial for improving the model's ability to adapt to diverse beamforming requirements, reducing the computational resources used by the trained model when adapting to new application scenarios, quickly grasping the characteristics of new scenarios, and reducing the need for samples.

[0035] This invention, through sampling methods, ensures that the action group and the stop group have the same distribution to a greater extent compared with traditional methods. At the same time, it ensures that the model can be evaluated using data not used in training, which is beneficial to improving the reliability and validity of model evaluation.

[0036] This invention enhances the model's ability to grasp scene features and break through local optima by rationally designing the trainer and learning mechanism. In this embodiment, since the fine-tuned model does not need to backpropagate in the application scenario, it is advantageous to achieve better communication results with significantly less computational resources compared to traditional beamforming methods. This invention, through the rational design of a Transformer-based trainer and a corresponding training mechanism, facilitates the model's ability to simultaneously focus on different information in different subspace representations, improving its ability to learn input information, promoting gradient propagation, and standardizing the states of layers in the trainer. Furthermore, it also enables efficient computation on hardware, accelerating the training process. Specifically, based on common Transformer implementations, this invention adds a gradient layer before the traditional mapping layer and an adder after the traditional disassociation layer. This gives the invention a stronger ability to break through local optima in complex multi-user problems, reduces the interference of local optimizations with zero first-level gradients, and makes training easier to obtain the global optimum. In addition, the training of this invention is based on the calculation of the loss function by formula, which does not require the prior preparation of beamforming vectors, which helps to reduce development costs and also helps to adapt to application scenarios that are difficult to predict during the training phase. Attached Figure Description

[0037] Details of one or more implementations of the subject matter described in this invention are set forth in the accompanying drawings and the following description. Other aspects, features, and advantages will be obtained or understood from this description, the drawings, and the claims. It should be noted that the following drawings are schematic diagrams, and their relative dimensions may not be drawn to scale.

[0038] Figure 1 The above is a flowchart of the method provided in the embodiments of the present invention.

[0039] Figure 2 The structure and training flowchart of the neural network and trainer provided in the embodiments of the present invention are shown.

[0040] Figure 3 The architecture of the trainer provided in the embodiments of the present invention.

[0041] Figure 4 The figures show the test results under different signal-to-noise ratio environments provided in the embodiments of the present invention.

[0042] Figure 5The figure shows the online fine-tuning results of the dataset group used in training under a new scenario provided by an embodiment of the present invention.

[0043] Figure 6 The figure shows the online fine-tuning results for a new scenario using a dataset group that was not used during training, as provided in an embodiment of the present invention.

[0044] Figure 7 The image shows the online beamforming results after freezing the neural network, which is the result of online fine-tuning of a new scenario based on an unused dataset group provided in this embodiment of the invention. Detailed Implementation

[0045] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0046] This invention provides a robust beamforming design method for cross-scenario applications based on hybrid element learning.

[0047] The present invention provides a detailed description of its key aspects in the following sections. It should be understood that numerous specific details are listed in the following description, intended to provide a thorough understanding of the invention. However, the invention can also be practiced in ways different from those described herein. Those skilled in the art will be able to make appropriate adjustments and variations without departing from the core spirit of the invention. Therefore, the scope of the invention is not limited to the specific embodiments described below.

[0048] 1. Regarding the generation of action group and stop group data, one specific implementation step of this invention is as follows:

[0049] Step 1: Select s typical scenarios from the g datasets and collect data in each scenario, collecting M data per scenario. s A collection containing N s Channel data from individual users forms the channel data source. The number of datasets is generally no less than two. The typical scenarios require diverse channel characteristics, and each dataset must have at least one scenario included. In each scenario, there is one dataset loaded with N... tThe base station has one antenna. Each user has only one receiving antenna. The collected channel data should be evenly distributed in the selected scenario and include information such as the channel matrix and path loss. The channel matrix in the collected channel data needs to be normalized. Normalization means that, while ensuring the phase information of the channel matrix remains unchanged, its magnitude is reduced to 1 by dividing it. Normalization also includes discarding data when the channel matrix is ​​empty, assuming that the base station and user cannot communicate effectively in this case. Channels from the channel data source are selected without repetition. Normalization also includes preprocessing the shape of the channel matrix: for each element of the data source, it should be flattened into a 2D tensor, and the real and imaginary parts should be separated and concatenated; the final generated shape is (n1, n2), where n1 = N. S *N t *2, n2=1.

[0050] Step 2: Mix the channel data sources described in Step 1, perform k samplings, and generate an action group containing k channel data sources. and includes k q Stop group of channel data sources The number of samplings k satisfies the following formula:

[0051] k = s·M s ·N s

[0052] Sampling uses sampling with replacement. The sampled data is placed into the action group without removing duplicate values. The stop group is generated by collecting all unsampled channel data from the channel data source. For any channel number in the channel data source, its probability P entering the stop group satisfies the following formula:

[0053]

[0054] For a not-too-small k, this probability satisfies Let n be the base of the natural logarithm. It can be proven that the number of data points in the stopping group, n... T Follows a Gaussian distribution:

[0055]

[0056] It can be proven that n T There is a 95% probability that it falls within the interval. Based on this, it can be considered that for k not less than 100, a 25% deviation is a fairly random event, and therefore resampling should be performed.

[0057] 2. In terms of neural networks and their training, a specific implementation step of this invention is as follows:

[0058] Step 1: Construct a neural network and design a trainer that includes a transmitter. Receiver trainer Weight trainer The neural network consists of the required components, and the parameters of the three trainer neural networks are initialized. Specifically, if Θ is used respectively... F Θ J and Θ Q Indicates the trainer and Given the initial parameters of a neural network, and assuming these initial values ​​are arbitrarily chosen within the linear space of their values, this process can be represented by the following formula:

[0059]

[0060]

[0061]

[0062] in, The trainers are represented in order. and The linear space formed by the initial parameter values ​​of the neural network in the diagram. Representative at The parameters are arbitrarily selected. This step also includes the initialization of parameters such as the initial transmitter matrix, the initial receiver matrix, and the initial weights (hereinafter referred to as "preset values"):

[0063]

[0064]

[0065]

[0066] st||F||=||J||=||Q||=1

[0067] Here, "∈" means "belongs to", that is, the initial parameters are arbitrarily chosen from their respective linear spaces. These represent the linear space containing the sending matrix, receiving matrix, and weights in the current scenario, respectively. Representative at Let st||F||=||J||=||Q||=1 represent constraining the magnitude of each element in the aforementioned linear space to be 1. It can be proven that this initialization method, in the multi-user environment addressed in this invention, facilitates the model's better exploration of the parameter space. This initialization method provides a more comprehensive starting point for subsequent gradient descent or other optimization algorithms, helps avoid the model getting trapped in local optima, and improves training performance on non-convex optimization problems.

[0068] Figure 2 The specific process of the trainer is illustrated. In this embodiment, the trainer is designed based on the Transformer architecture. Because beamforming optimization in a multi-user environment suffers from the problem of zero first-order gradients at some points (i.e., flat regions or local optima), relying solely on first-order gradient information may lead the optimization algorithm into poor local optima. Based on this, this invention utilizes higher-order gradient information to approach the global optimum. Specifically, for each trainer, the pre-initialized parameter I is first input into the gradient layer to obtain its gradient with respect to the loss function. Then The mapping layer of the trainer extracts features and transforms them into a dense vector K, adding unique positional information. The trainer inputs K into an association layer containing an array of associationors. The associationors decompose and analyze different aspects of K's features, and the analysis results are input into the deassociation layer to obtain the output I*. Here, the output I* is used as the difference from the input I to update the input, thus obtaining the output O = I + I*. By combining first-order and second-order gradient information, this update strategy allows the model to escape local optima or flat regions when the first-order gradient is zero, using the second-order gradient to increase the chance of finding the global optimum or close to the global optimum.

[0069] Step 2, train the neural network. Use the generated action sets as training cases sequentially for training. During training, first randomly initialize the parameters and preset values ​​of the three trainer neural networks: initial weights Q0, initial receiver matrix J0, and initial transmitter matrix v0. Input the preset values ​​into the loss function and calculate the gradient of the loss function with respect to the preset values. Then, input the training cases and the gradient of the loss function with respect to the preset values ​​into the trainer for forward propagation to obtain the outputs Q, J, and v of each trainer. Input Q, J, and v into the loss function and backpropagate based on this. The formula for the loss function is:

[0070]

[0071] In the formula, L represents the loss function, and n represents the noise amplitude. For cases where preset values ​​Q0, J0, and v0 are input to the loss function, Q, J, and v in the above formula need to be replaced with initial weights, initial receiver matrix trainer, and initial transmitter matrix. However, for cases where the outputs Q, J, and v of the trainer are input to the loss function, Q, J, and v in the above formula directly represent Q, J, and v themselves. In the formula, h... TThis represents the transpose of the channel matrix from the base station to a specific user. In the formula, the subscript represents the user number corresponding to the output element. The internal parameters of the trainer are updated based on the gradient of the loss function and the learning rate specified by the trainer. After repeating this process p times, proceed to step 3, where p is a hyperparameter specified by the trainer. It should be noted that before inputting the loss function, the preset values ​​need to be normalized. This ensures that the phase of each element in the preset value vector remains unchanged, the magnitude of each element in V is 1, and the sum of the magnitudes of each element in Q and J is 1. During training, the aforementioned forward and backward propagation processes are repeated multiple times. This method can further include, during the first backward propagation, the parameters of the trainer's hidden layer change by r times the gradient of the sum of the rates, while during the i-th backward propagation, the parameters of the trainer's hidden layer change by r times the gradient of the sum of the rates. The method further includes the fact that r is pre-specified by the trainer, and the specific values ​​may not be exactly the same for the three trainers described in this method.

[0072] Step 3, evaluate the neural network. A set of replicas of the trainer is created using a deep replication method. While keeping the parameters of each trainer's neural network unchanged, the preset values ​​of each replica are reset. Using samples from the stopping group, the preset values ​​are updated d times in the manner of Step 2, where d is specified by the trainer and generally does not exceed 30. During evaluation, this embodiment fixes the noise of the loss function to n. 2 The last updated preset value and the aforementioned samples are input into the loss function. The outputs of the three trainers are input into the sum rate calculation function. With a fixed signal-to-noise ratio, the sum rate value is recorded. This process is repeated for the samples in the stopping group, and the average value is taken as the final result of this evaluation. The final result of this evaluation is compared with historical results. If the sum rate no longer shows an upward trend, training is considered to be stopped; otherwise, step 2 is repeated. The criterion for the sum rate no longer showing an upward trend is that the increase in sum rate over h / p consecutive times is less than c. s Or there was no increase, where h and c s The hyperparameter, h, is specified by the trainer and is typically set to a large value. When dealing with complex non-convex optimization problems such as multi-user scenarios and reflective smart surfaces, the approach proposed in this step not only helps prevent overfitting on the training set, thus improving the model's generalization ability to unforeseen data, but also significantly improves computational efficiency and saves computational resources. Furthermore, the stopping method proposed in this step provides an adaptive regularization mechanism, allowing the model to automatically stop training based on its performance on the validation set without requiring manual adjustment of relevant parameters by the trainer.

[0073] Step 4: Package and save the parameters of the three trainer neural networks to any easily accessible location.

[0074] 3. Regarding the online deployment and application of this invention, a specific implementation step is as follows:

[0075] Step 1, online fine-tuning: Collect several channel data sets in the actual usage scenario as fine-tuning groups. Train the neural network trained using the aforementioned method on these fine-tuning groups a limited number of times. Taking the scenario selected in this embodiment as an example, 40 training iterations can be performed. The training method can refer to the aforementioned steps, but testing and evaluation on the stopping group is not required. This step is based on reinforcement learning and transfer learning design, which helps to solve the problem of data distribution drift between the training scenario and the actual scenario (i.e., the distribution of training data and actual usage data differs). It can be demonstrated that by training on the fine-tuning groups collected in the actual scenario, the model can more accurately capture the statistical characteristics of the target environment, thereby improving its performance in the actual production environment. Specifically, if P train (x,y) represents the distribution of the training data, while P actual (x,y) represents the distribution of data in a real-world application environment. The differences in this distribution can be represented by the distance D. KL To quantify, see below:

[0076]

[0077] The goal of online fine-tuning is to reduce this inconsistency by adjusting the model parameters Θ through a small number of iterations on real-world data. The parameter update can be represented as:

[0078]

[0079] Where L is the loss function, α is the learning rate, and ∂L is the gradient of the aforementioned loss function with respect to the parameter Θ. Training based on this formula can be performed online to fine-tune the model from its pre-trained state Θ. pre-trained Towards a P that is more adapted to the actual use environment actual State Θ fine-tuned The transformation is conducive to achieving the aforementioned D KL This reduces the cost, thereby improving the model's performance in real-world applications. It should be understood that online fine-tuning can take many forms, not limited to those mentioned above, with the primary goal of adapting the trainer to the specific scenarios in which it will be used.

[0080] Step 2, practical application: Freeze the neural network parameters obtained from the aforementioned fine-tuning as the actual model used in this scenario, and randomly generate initial weights Q0, initial transmitter matrix v0, and initial receiver matrix J0 as preset values; input the preset values ​​into the loss function, calculate the value of the loss function, and find the gradient of the value with respect to Q0, v0, and J0; input the gradient into each trainer for forward propagation to obtain the outputs Q, v, and J of each trainer; use the obtained Q, v, and J as the design results of the weights, transmitter, and receiver, respectively, as outputs for application.

[0081] Example 2

[0082] A test is provided to evaluate the sum rate under different signal-to-noise ratios and different scenarios. The aforementioned test focuses on the results of several datasets in DeepMIMO O1 and WAIR-D scenario group 1 of the aforementioned embodiments.

[0083] It should be understood that the tests provided in this disclosure are used only as a means of facilitating the understanding of this disclosure.

[0084] For the simulation results provided in this disclosure, the set parameters satisfy: s = 1000, M s *N s =135; for Figure 4 n x =500,s t =SNR, where SNR is distributed between -20 dB and 20 dB (inclusive); for Figure 5 , Figure 6 n x =500,s t =20. The simulation data provided in this disclosure uses scenarios that satisfy the following assumptions: each scenario has 4 users; each scenario has 1 base station transmitting electromagnetic waves at a baseband frequency of 6.0 GHz, and the base station has 4 antennas arranged in a 2×2 configuration, with the spacing between the two antennas in each row and column set to half the wavelength of the transmitted electromagnetic wave's baseband frequency in vacuum; the bandwidth used for communication is set to 48.6 MHz; and all users in each scenario have 1 antenna arranged in a 1×1 configuration. The upper reference limit in the figure is obtained iteratively using the weighted minimum mean square error (WMMSE) method.

[0085] This disclosure is based on this setting, and the results of all samples are processed and recorded using an arithmetic mean. In the attached image, Figure 4 The summation rate of this method under different signal-to-noise ratio environments is demonstrated; Figure 5The results demonstrate the changes in sum and rate during fine-tuning in different scenarios within the dataset used in the aforementioned training, simulating the effects during deployment in different scenarios using the dataset used in the aforementioned training. Figure 6 The results demonstrate the sum and rate changes during fine-tuning in the scenario of the unused dataset set during the aforementioned training, simulating the effect during deployment in the scenario of the unused dataset set during the aforementioned training. Figure 7 The simulation demonstrates the sum and rate changes in a real-world application scenario within a dataset not used in the aforementioned training, simulating the beamforming process after the deployment of the real-world application scenario.

[0086] It should be understood that this disclosure also provides sums and rates of other algorithms in the specification and accompanying figures. The sums and rates of these other algorithms provided are only for the convenience of reference and understanding of the present invention and the simulation effects provided in this disclosure of the embodiments.

[0087] This disclosure provides an analysis of the aforementioned tests based on the tests described above. It should be understood that the analysis of the tests provided in this disclosure is solely for facilitating reference and understanding of the invention and the simulation effects of the embodiments described herein.

[0088] Analysis of the aforementioned tests of this invention indicates that: Figure 4 This demonstrates that the present embodiment can adapt to different signal-to-noise ratio environments quite well; Figure 5 and Figure 6 This demonstrates that this embodiment can adapt relatively well to communication environments with different characteristics than the channel data used in the aforementioned training, across different scenarios and datasets. Analysis of the aforementioned tests of this invention indicates that: based on... Figure 5 and Figure 6 It can be observed that the algorithm proposed in this invention has a high starting point of adaptation and a fast ascent speed, demonstrating that the method used in this embodiment is beneficial for reducing the consumption of computing resources, accelerating its deployment in the environment, and achieving more ideal communication quality; based on Figure 7 It can be observed that, in dynamically changing scenarios, the method used in this embodiment is beneficial for obtaining more ideal communication results and for adapting to the changing channel conditions in real-world scenarios.

[0089] Corresponding to the aforementioned embodiment of a robust beamforming design method for cross-scenario based on hybrid learning, the present invention also provides an embodiment of a robust beamforming design system for cross-scenario based on hybrid learning.

[0090] This invention provides a robust beamforming design system for cross-scenario applications based on hybrid element learning, comprising one or more processors for implementing a robust beamforming design decoding method for cross-scenario applications based on hybrid element learning as described in the above embodiments.

[0091] An embodiment of the robust beamforming design system based on hybrid learning for cross-scenario applications, as described in this invention, is applicable to any device with data processing capabilities. The device may include various hardware and software components, including but not limited to circuits, application-specific integrated circuits (ASICs), and processors. In the operational description shown in the accompanying drawings, the operations can be implemented by corresponding matching devices and functional components.

[0092] The apparatus embodiments generally correspond structurally to the method embodiments. For similar content, please refer to the relevant description of the method embodiments. The above apparatus embodiments are illustrative, wherein units defined as separate components may or may not be physically independent; components identified as units may or may not be physical units. These components may be concentrated in one location or distributed across multiple network units. Depending on actual needs, some or all modules may be selected to achieve the objectives of the present invention. For those skilled in the art, this invention can be understood and practiced without excessive innovative effort. Various modifications to the implementations described in this disclosure may be apparent to those skilled in the art, and the universal principles defined herein may be applied to other implementations without departing from the spirit or scope of this disclosure. Therefore, the claims are not intended to be limited to the implementations shown herein, but should be granted a scope consistent with this disclosure, the principles disclosed herein, and the novel features.

[0093] This invention further provides a computer-readable storage medium that stores a computer program. When this program is executed on a processor, it can implement the robust beamforming design method for cross-scenario applications based on hybrid learning described earlier. This computer-readable storage medium can be an internal storage element of any data processing device as described in the foregoing embodiments, such as a hard disk or memory. Alternatively, the computer-readable storage medium can also refer to external storage devices, such as plug-in hard disks, smart memory cards (SMCs), SD cards, flash memory cards, CDs, laser discs, optical discs, DVDs, floppy disks, and Blu-ray discs. This computer-readable storage medium may also encompass both the internal storage portion of a device and external storage devices. In addition to storing the aforementioned computer program, this storage medium can also store other required programs and data, or temporarily store data that has been output or is yet to be output.

[0094] Those skilled in the art will understand that the functions described in this application can be implemented by hardware, software, firmware, or a combination of these components. When implemented in software, these functions can be stored on a computer-readable medium or transmitted as one or more instructions or codes on that medium. The computer-readable medium includes both computer-readable storage media and communication media, the latter being media that facilitates the transfer of a computer program from one location to another. This storage medium can be any general-purpose or special-purpose, computer-accessible medium.

[0095] Obviously, the above description and embodiments are merely illustrative of specific examples of the present invention and are not intended to limit it. Those skilled in the art can make various modifications or adjustments based on the above description. Clearly, it is impossible to list all possible implementations in detail, but all reasonable variations made within the scope of this description should be considered within the scope of the present invention. It is worth noting that although the accompanying drawings may depict certain operations in a specific order, this does not mean that these operations must be performed in this specific order, nor does it mean that all described operations must be performed to obtain the desired effect. The drawings may merely illustrate some operational flows in a schematic manner, but may include other steps not explicitly described. For example, some steps may be performed before, after, or in the middle of the described operations. In some situations, multitasking or parallel processing may be advantageous. At the same time, although the independence of some components may be described herein, this does not mean that such independence is required in all cases. In fact, the various parts or systems described can often be integrated into one software or divided into multiple different software programs. In general, other implementations and variations should be considered within the scope of protection of the present invention. In some instances, the operations described in the claims may not be performed in a fixed order, but the expected results can still be obtained.

Claims

1. A robust beamforming design method for cross-scenario applications based on hybrid element learning, characterized in that, The method includes the following steps: Step 1, Data Acquisition: Select data from multiple datasets. Data will be collected in the above scenarios, with data collected in each scenario. Includes Channel data from individual users forms the channel data source. ; Step 2, Constructing the Beamforming Training Dataset: The channel data sources are mixed and then... Sub-sampling, generating a sample containing Action group for each channel data source and includes Stop group of channel data sources ; Step 3, Construct the neural network: This network includes a transmitter trainer. Receiver trainer and weight trainer and the corresponding backpropagation module; the trainer is designed based on the Transformer architecture; during training, the trainer... , and First, initialize the weights, the transmitting matrix, and the receiving matrix to the initial weights Q0, the initial transmitting matrix v0, and the initial receiving matrix J0, respectively, as preset values. Input the preset values ​​into the gradient layer to obtain its gradient with respect to the loss function. Then The input to the training layer's mapping layer extracts features and transforms them into dense vectors. And add unique location information; the trainer will then Input is an association layer containing multiple association arrays; the associations in the association layer are decomposed and analyzed. The different aspects of the characteristics are analyzed and the results are input into the decorrelation layer; the decorrelation layer calculates the output; the output of the decorrelation layer is used as the difference of the preset value, which is added to the preset value, and the result is used as the output of the corresponding trainer; Step 4, Training the Neural Network: Initialize the parameters of the trainer neural network, input the action components into each trainer in batches for forward propagation, use the output as the difference of the preset value and update the preset value, use the updated value as the output of each trainer, input it into the loss function, backpropagate on this basis, and repeat. The specific training process is as follows: Initialize the parameters of each trainer's neural network, and randomly generate initial weights Q0, initial matrix v0 at the sending end, and initial matrix J0 at the receiving end as preset values; input the preset values ​​into the loss function, calculate the value of the loss function, and find the gradient of the value with respect to Q0, v0, and J0; input the gradient into each trainer for forward propagation to obtain the outputs Q, v, and J of each trainer; input Q, v, and J into the loss function, calculate the gradient of the value of the loss function with respect to Q, v, and J, and perform backpropagation to update the parameters of the trainer's neural network; Step 5, evaluate the neural network: perform forward propagation on the stopping group, input the output of each trainer into the sum rate calculation function, evaluate the current training effect, and compare it with the historical training values ​​of the current dataset. If the training effect no longer improves, end the training and save the model and trainer obtained at the end as the application model. Step 6: Online fine-tuning. Collect several channel data in the actual use scenario as a fine-tuning group, train the application model on the fine-tuning group, and freeze the parameters of the obtained trainer neural network as the final model to be used in the scenario. Step 7, Practical Application: In actual use scenarios, acquire the user channel, input the acquired channel into the loss function, and input the result of the loss function into the model obtained in Step 6 for forward propagation. Use the output as a beamforming vector.

2. The robust beamforming design method for cross-scenario applications based on hybrid element learning as described in claim 1, characterized in that, In step 1, there are no fewer than two datasets, and each dataset has at least one scenario selected and added to the channel data source. The collected data includes the channel matrix and path loss, and the channel matrix needs to be pre-normalized.

3. The robust beamforming design method for cross-scene applications based on hybrid element learning according to claim 1, characterized in that, The number of samplings in step 2 Satisfy the following formula: The sampling method is sampling with replacement.

4. The robust beamforming design method for cross-scene applications based on hybrid element learning according to claim 1, characterized in that, In step 2, the channel data source is stopped. It is generated in the following ways: In this case, the elements of the stopping group are unique, and the expected number of elements is close to the value calculated by the following formula: ,in It is the natural logarithm; Count the elements in the obtained stopping group. If the number of elements obtained is equal to the number of elements in the stopping group, then... If the deviation exceeds the preset deviation threshold, the action group and stop group will be generated again.

5. The robust beamforming design method for cross-scene applications based on hybrid element learning according to claim 1, characterized in that, In step 4, the training of each trainer uses the inverse of the sum rate as a reference. The specific formula for the sum rate is: in, Represents the loss function. Representative and rate, The amplitude represents the noise level, Q, J, and v represent the outputs of the weighted trainer, receiver trainer, and transmitter trainer, respectively, and h T This represents the transpose of the channel matrix from the base station to a specific user, with the subscript indicating the user number corresponding to each output element. The noise used is derived based on the following formula: in For path loss, Let V be the variance of the circular complex Gaussian noise.

6. The robust beamforming design method for cross-scene applications based on hybrid element learning according to claim 1, characterized in that, In step 4, during the first backpropagation, the parameters of the trainer's association layer and disassociation layer (hereinafter collectively referred to as the hidden layer) change as the gradient of the rate. times; in the During the second backpropagation, the parameters of the hidden layer of the trainer change as a gradient of the rate. times; in These are hyperparameters, specified by the trainer. For the three trainers, their specific values ​​are not exactly the same.

7. The robust beamforming design method for cross-scene applications based on hybrid element learning according to claim 1, characterized in that, In step 5, the evaluation criteria for training effectiveness are as follows: input the channel portion of all data in the stopping group in sequence, calculate the sum rate when the signal-to-noise ratio is fixed; take the average sum rate obtained from the stopping group samples as the final evaluation result; compare the final evaluation result with the historical results, and if the sum rate no longer shows an upward trend, then training stops.

8. The robust beamforming design method for cross-scene applications based on hybrid element learning according to claim 1, characterized in that, In step 6, online fine-tuning specifically involves: collecting several channel data points as a fine-tuning group in the actual usage scenario, and training the neural network trained using the aforementioned method on the fine-tuning group a small number of times; in step 7, practical application specifically involves: fixing the parameters of each trainer's neural network, and randomly generating initial weights Q0, initial transmitter matrix v0, and initial receiver matrix J0 as preset values; inputting the preset values ​​into the loss function, calculating the value of the loss function, and obtaining the gradient of the value with respect to Q0, v0, and J0; inputting the gradient into each trainer for forward propagation to obtain the outputs Q, v, and J of each trainer, which are then used as the design results of the weights, transmitter, and receiver, respectively, for application.

Images