A Beamforming Computation Method Based on Deep Learning

Abstract

This invention provides a deep learning-based beamforming computation method, comprising: acquiring a labeled beamforming dataset of a real-world system scenario; constructing a neural network architecture based on the actual application scenario or requirements; designing a loss function for the neural network architecture based on the application scenario and beamforming goals and requirements; training the constructed neural network architecture using a mini-batch gradient descent method on a training set; performing a generalization test on the trained neural network architecture using a test set; obtaining a trained neural network architecture when the loss value of the loss function meets the requirements; and calculating beamforming for a new system scenario using the trained neural network architecture. This invention promotes the application of deep learning in beamforming computation and proposes a generalization test for neural networks in beamforming inference, demonstrating good optimal and generalization performance, and achieving millisecond-level response times.

Description

Technical Field

[0001] This invention relates to the field of beamforming technology in wireless communication systems, and more particularly to a beamforming calculation method based on deep learning. Background Technology

[0002] Multi-antenna technology is a core technology in 5G and future mobile communication systems. Beamforming can better leverage the spatial freedom offered by multi-antenna technology to achieve directional transmission, thereby improving spectral efficiency and energy efficiency. However, as mobile communication systems become increasingly complex and dense, achieving high-performance, dynamic, and low-response-time beamforming designs based on traditional solutions faces challenges.

[0003] Deep learning leverages the function approximation capabilities of neural networks to learn the mapping from wireless channels to optimal beamforming, thus solving the aforementioned problems. Therefore, it is necessary to provide a standardized process for beamforming computation based on deep learning. Furthermore, traditional neural networks suffer from poor generalization ability in beamforming design, failing to adapt to dynamically changing wireless environments. Therefore, it is essential to design neural network architectures specifically tailored to application scenarios and requirements to improve algorithm performance. Summary of the Invention

[0004] Embodiments of the present invention provide a deep learning-based beamforming calculation method to achieve efficient calculation of beamforming in wireless communication systems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution.

[0006] A deep learning-based beamforming computation method includes:

[0007] A labeled beamforming dataset of actual wireless communication system scenarios is obtained. The features of the beamforming dataset include real-time channel information and communication requirements. The labels include the optimal beamforming vector. The beamforming dataset is divided into a training set and a test set.

[0008] Build a neural network architecture based on actual application scenarios or needs, and design the loss function of the neural network architecture based on the application scenarios and beamforming goals and needs;

[0009] The constructed neural network architecture is trained using the training set through the mini-batch gradient descent method. The generalization test of the trained neural network architecture is performed using the test set. When the loss value of the loss function of the trained neural network architecture meets the requirements, the trained neural network architecture is obtained.

[0010] The trained neural network architecture is applied to a new wireless communication system scenario to perform beamforming calculations for that scenario.

[0011] Preferably, the acquisition of labeled beamforming datasets for actual wireless communication system scenarios includes: collecting relevant data in actual wireless communication system scenarios as beamforming datasets; or, modeling the actual wireless communication system scenarios as optimization problems and generating beamforming datasets through optimization algorithms.

[0012] Preferably, the step of building a neural network architecture according to actual application scenarios or needs includes:

[0013] A neural network architecture is constructed based on a graph neural network, which includes a graph convolutional neural network and a graph attention network. The model of the neural network architecture is improved into a complex model. Complex channel state information is input into the model of the neural network architecture, and the model of the neural network architecture outputs the corresponding beamforming vector, thereby realizing end-to-end learning.

[0014] Preferably, the loss function for designing the neural network architecture based on the application scenario and beamforming objectives and requirements includes:

[0015] The mean squared error is used as the loss function for designing the neural network architecture, as shown below:

[0016]

[0017] In the formula, U(.) is the utility function designed according to the beamforming target and requirements, {w i} out Beamforming of the neural network output, {w i} target Beamforming of labels in the dataset;

[0018] For unconstrained beamforming problems, the objective function is directly used as the utility function, i.e., U({w i})=OBJ({w i For constrained beamforming problems, a utility function is designed using the penalty function method, i.e. OBJ(.) represents the objective function, J represents the number of constraints, and μ j f represents the penalty coefficient. j (.) indicates a constraint.

[0019] Preferably, the step of training the constructed neural network architecture using the training set via mini-batch gradient descent, and then performing generalization testing on the trained neural network architecture using the test set, to obtain the trained neural network architecture after verifying that the loss value of the loss function of the trained neural network architecture meets the requirements, includes:

[0020] The parameters in the neural network architecture are initialized, and the samples in the training set are input into the neural network architecture in batches. The neural network architecture infers the optimal beamforming mapping of the wireless channel through deep learning, and obtains the corresponding beamforming for each sample. The trained neural network is generalized using the test set, and the loss value of the loss function of the sample is calculated. The coefficients of the neural network architecture are updated according to the loss value using backpropagation and gradient descent. The above process is repeated until the loss value tends to converge, and the trained neural network architecture is obtained.

[0021] Preferably, applying the trained neural network architecture to a new wireless communication system scenario to perform beamforming calculations for the new wireless communication system scenario includes:

[0022] Acquire labeled beamforming datasets for new wireless communication system scenarios. The features of the beamforming datasets include real-time channel information and communication requirements, and the labels include the optimal beamforming vector.

[0023] The samples of different parameters in the beamforming dataset of the new system scenario are input into the trained neural network architecture, and the trained neural network architecture outputs the beamforming vector of the new wireless communication system scenario.

[0024] As can be seen from the technical solutions provided by the embodiments of the present invention above, the beamforming calculation scheme proposed in the embodiments of the present invention has good optimal performance and generalization performance, and can achieve millisecond-level response.

[0025] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart illustrating a deep learning-based beamforming calculation method provided in an embodiment of the present invention;

[0028] Figure 2 This is a structural diagram of a beamforming computing platform based on deep learning, provided for an embodiment of the present invention. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0030] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0031] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0032] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0033] This invention relates to a deep learning-based beamforming computation scheme and a platform supporting its deployment. The method constructs a dataset by optimizing algorithms and using real data, then builds a neural network architecture tailored to application scenarios and requirements based on deep learning. Next, the neural network is trained using supervised learning, and finally, the trained neural network is deployed to a new scenario for beamforming computation. To facilitate the deployment of the proposed computation scheme, this invention also proposes a supporting platform for dataset management, neural network training, and commercial services.

[0034] The processing flow of a deep learning-based beamforming calculation method provided in this embodiment of the invention is as follows: Figure 1 As shown, the processing steps include the following:

[0035] Step S10: Obtain labeled beamforming dataset.

[0036] For some common mobile communication scenarios, such as typical cellular systems, IoT systems, and industrial internet systems, and combined with the personalized needs of application scenarios, including throughput, latency, and energy consumption, we model machine learning tasks. On this basis, we obtain labeled beamforming datasets through actual system measurement or simulation optimization algorithms. The characteristics of the beamforming datasets include real-time channel information, communication requirements, etc., and the labels are the optimal beamforming vectors, etc.

[0037] There are two methods to obtain the required beamforming dataset: collecting relevant data in actual communication system scenarios as beamforming datasets; or modeling the actual system scenario as an optimization problem and generating beamforming datasets through optimization algorithms.

[0038] Step S20: Construct the architecture of the neural network and design the loss function of the neural network.

[0039] To enhance the learning and generalization capabilities of the constructed neural network, this invention builds a neural network architecture based on some common graph neural networks, including Graph Convolutional Network (GCN) and Graph Attention Network (GAT). The neural network architecture model is improved to a complex model, which allows complex channel state information to be directly input into the model and output corresponding beamforming vectors, thus achieving end-to-end learning.

[0040] The loss function of the neural network architecture is constructed based on the design goals and requirements of beamforming to ensure the usability of the obtained beamforming results. For example, mean square error can be used as the loss function, as shown below.

[0041]

[0042] In the formula, U(.) is the utility function designed according to the beamforming target and requirements, {w i} out Beamforming of the neural network output, {w i} target Beamforming of labels in the dataset.

[0043] For unconstrained beamforming problems, the objective function can be directly used as the utility function, i.e., U({w i})=OBJ({w i For constrained beamforming problems, a utility function can be designed using the penalty function method, i.e. In the formula, OBJ(.) represents the objective function, J represents the number of constraints, and μ j f represents the penalty coefficient. j (.) indicates a constraint.

[0044] Step S30: Train the neural network using the beamforming dataset and perform a generalization test on the neural network.

[0045] The neural network is trained using the constructed beamforming dataset. The neural network learns the mapping from the wireless channel to the optimal beamforming through deep learning until the neural network has good performance on the training set.

[0046] Specifically, the training set is input into the neural network in batches for inference and the corresponding beamforming is output. The loss value is calculated based on the constructed loss function. The neural network coefficients are updated using backpropagation and gradient descent. The above process is repeated until the loss value tends to converge.

[0047] Step S40: Apply the trained neural network to the new scene and perform beamforming calculations for the new scene.

[0048] Acquire labeled beamforming datasets for new wireless communication system scenarios. The features of the beamforming datasets include real-time channel information and communication requirements, and the labels include the optimal beamforming vector.

[0049] The samples of different parameters in the beamforming dataset of the new wireless communication system scenario are input into the trained neural network architecture, and the trained neural network architecture outputs the beamforming vector of the new wireless communication system scenario.

[0050] A test set, generated with different antenna numbers, user numbers, base station numbers, and channel information than the training set, is used as the input to the neural network, and the output is a corresponding beamforming vector. This evaluates the generalization ability of the neural network in different scenarios. The constructed neural network architecture is theoretically analyzed, and methods for constructing neural networks for different learning tasks are proposed. Then, neural network architecture, hyperparameter search, and model compression are performed sequentially to optimize the neural network architecture.

[0051] This invention fully considers the scene and the goals and requirements of beamforming, helps the neural network learn feasible results, and improves the learning efficiency and effect of the neural network.

[0052] After training, the model is tested on a test set with different antenna numbers, user numbers, base station numbers, and channel information compared to the training set. The designed model is validated and evaluated based on various metrics on different test sets, and the generalization error is analyzed, providing conditions for practical deployment. This paper integrates standardized steps for solving beamforming computation problems based on deep learning, simplifies the solution process for similar problems, and provides benchmark algorithm results and a performance evaluation system, contributing to the exchange and development of related research directions.

[0053] Table 1 shows the performance of the proposed computational scheme in terms of computational optimality, generalization ability, and computation time. The number of antennas and users in the training set are (4,2) and (8,3), respectively; the number of antennas and users in the test set are (4,2), (8,2), (8,3), and (8,4), respectively. It can be seen that when the number of antennas and users in the training and test sets are the same (i.e., the test generalization ability to the channel), the proposed graph attention network's performance loss is within 10%, which is better than multilayer perceptrons and graph convolutional neural networks. When the number of users in the training and test sets is different (i.e., the test generalization ability to the channel and the number of users), the proposed graph attention network's performance is within approximately 20%, which is better than graph convolutional neural networks, while multilayer perceptrons do not possess this generalization ability. Furthermore, all proposed neural networks can achieve millisecond-level response times, making them suitable for dynamic wireless environments. This demonstrates the superiority of the proposed method.

[0054] Table 1 Performance Demonstration of the Calculation Scheme

[0055]

[0056] This invention provides a deep learning-based beamforming computing platform, such as... Figure 2 As shown, the cloud deployment method is adopted, and the functions of each main module are as follows.

[0057] Dataset Module

[0058] The dataset module is mainly responsible for generating or selecting datasets for supervised learning. If a suitable public dataset exists, it can be used directly. If not, an optimization algorithm or field measurement can be used to generate the corresponding dataset, which will then be added to the public dataset.

[0059] Dataset generation module

[0060] Real data generation

[0061] This platform integrates some common communication hardware devices and common scenarios (MISO, etc.). Through a visual operation interface, relevant parameters (such as the number of antennas and the number of users) can be set, and relevant dataset generation jobs can be submitted. The platform controls the relevant devices to generate datasets through remote commands and collects relevant data into the database.

[0062] Simulation data generation

[0063] Users can choose from existing simulation code of optimization algorithms on the platform or upload their local optimization algorithm code. After setting the relevant dataset parameters, they submit the simulation dataset generation job. The platform will generate a simulation dataset based on the job task. The platform has a built-in optimization algorithm code verification tool to verify the usability of the optimization algorithm code.

[0064] Public datasets

[0065] This module collects publicly available datasets from communication-related papers. Users can directly copy these datasets for model training or upload their local datasets for public learning and exchange. The platform includes a built-in dataset verification tool to validate the usability of user-submitted datasets.

[0066] Deep learning experiment module

[0067] The deep learning experiment module is mainly used for model training and includes functions such as cloud server rental, model architecture and hyperparameter search, and model compression.

[0068] Cloud server rental

[0069] Users can choose the appropriate server model according to their training needs, and configure the appropriate image to generate server instances according to different experimental environment requirements. The instances contain all publicly available datasets for model training. The data of each instance will be maintained for one month when the server is off. If the instance is not used for more than one month, it will be released.

[0070] Model Architecture and Hyperparameter Search

[0071] The model architecture search module allows users to select the most suitable model architecture, such as the number of neural network layers and the number of hidden units, based on the performance of models with different settings on small datasets.

[0072] The model architecture search task mainly consists of three parts: search space design, search strategy selection, and performance estimation. The exploration strategy selects an architecture from a predefined search space. This architecture is then passed to a performance evaluator to obtain a score, which represents the performance of this network structure on a specific task. This process is repeated until the search finds the optimal network structure.

[0073] Hyperparameter search is similar to model architecture search, requiring the setting of three parts: the search space for hyperparameters, the selection of search strategies, and performance estimation. Unlike model architecture search, which yields an optimal model architecture, hyperparameter search yields a set of optimal hyperparameters.

[0074] The performance metrics generated during the model architecture and hyperparameter search process will be visualized on the platform, allowing users to control the search process in real time based on the data.

[0075] Model compression

[0076] Neural networks in wireless communication systems typically need to be deployed on devices with limited computing resources or strict latency requirements. However, the neural networks obtained directly from training are usually computationally and energy-intensive. Therefore, model compression is needed to reduce model size and accelerate model training and inference without significantly degrading model performance. Model compression techniques can be divided into two categories: pruning and quantization. Pruning methods explore redundancy in model weights and attempt to remove or trim redundant and non-critical weights. Quantization refers to compressing the model by reducing the number of bits required to represent or activate weights.

[0077] Users can select relevant model compression algorithms from MMRazor (model compression algorithm library) on the platform to compress models, and the performance of the compressed model will be displayed in the visualization interface.

[0078] Performance evaluation

[0079] The performance evaluation function compares the trained model with various mainstream methods (such as optimization algorithms, reinforcement learning, etc.) on relevant specified metrics, such as method performance, generalization ability, computation time, training cost, etc. The comparison results of various methods will be visualized.

[0080] Simulation test

[0081] The trained model can be remotely deployed in a real environment controlled by the platform for testing, and test data can be collected and displayed on the relevant interface. This allows for a more realistic verification of the model's actual performance and deployment effect, providing guidance for further optimization of the model structure and related hyperparameters.

[0082] Experimental Management

[0083] Experimental data from several modules, including model architecture and hyperparameter search, model compression, and simulation testing, will be stored on a remote server. Users can view relevant training logs and reproduce experiments in the experiment management module.

[0084] Academic Forum Module

[0085] open source projects

[0086] Users can publish relevant open-source projects for others to reference and learn from. These projects include introductions to relevant datasets, related code, runtime environment requirements, and even training details. Other users can directly clone these open-source projects to their own server instances for running and reproduction.

[0087] Paper Reproduction

[0088] This module contains open-source code related to deep learning in communication, along with introductions to the papers. Users can also reproduce the relevant papers themselves and then share their projects publicly for others to exchange and learn from.

[0089] Academic exchange

[0090] Users can post questions and learn in the technical forum, whether it's about platform usage or paper-related issues. Through everyone's posts and answers, not only can a good learning atmosphere be fostered and the development of related research directions be promoted, but it can also help some users with weaker foundations to get started quickly.

[0091] In summary, the beamforming calculation scheme proposed in this embodiment of the invention has good optimal performance and generalization performance, and can achieve millisecond-level response.

[0092] This invention promotes the further application of deep learning in beamforming computation and proposes a test for the generalization ability of neural networks in beamforming inference. Through this method, those skilled in the art can more systematically and efficiently customize, train, and test deep neural networks. Finally, the open-source module and academic forum module designed in this method will also contribute to the development of beamforming technology in mobile communication systems.

[0093] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0094] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0095] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0096] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A beamforming calculation method based on deep learning, characterized in that, include: A labeled beamforming dataset of actual wireless communication system scenarios is obtained. The features of the beamforming dataset include real-time channel information and communication requirements. The labels include the optimal beamforming vector. The beamforming dataset is divided into a training set and a test set. Build a neural network architecture based on actual application scenarios or needs, and design the loss function of the neural network architecture based on the application scenarios and beamforming goals and needs; The constructed neural network architecture is trained using the training set through the mini-batch gradient descent method. The generalization test of the trained neural network architecture is performed using the test set. When the loss value of the loss function of the trained neural network architecture meets the requirements, the trained neural network architecture is obtained. The trained neural network architecture is applied to the new wireless communication system scenario to perform beamforming calculations for the new wireless communication system scenario; The loss function for constructing the neural network architecture is based on the design goals and requirements of beamforming. The mean squared error is used as the loss function for designing the neural network architecture, as shown below: In the formula, U(.) is the utility function designed according to the beamforming target and requirements, {w i } out Beamforming of the neural network output, {w i } target Beamforming of labels in the dataset; For unconstrained beamforming problems, the objective function is directly used as the utility function, i.e., U({w i })=OBJ({w i For constrained beamforming problems, a utility function is designed using the penalty function method, i.e. OBJ(.) represents the objective function, J represents the number of constraints, and μ j f represents the penalty coefficient. j (.) indicates a constraint; The process of training the constructed neural network architecture using the training set via mini-batch gradient descent, and then performing generalization testing on the trained neural network architecture using the test set, yields a trained neural network architecture once the loss value of the loss function of the trained neural network architecture meets the requirements. This process includes: The parameters in the neural network architecture are initialized, and the samples in the training set are input into the neural network architecture in batches. The neural network architecture infers the optimal beamforming mapping of the wireless channel through deep learning, and obtains the corresponding beamforming for each sample. The trained neural network is generalized using the test set, and the loss value of the loss function of the sample is calculated. The coefficients of the neural network architecture are updated according to the loss value using backpropagation and gradient descent. The above process is repeated until the loss value tends to converge, and the trained neural network architecture is obtained.

2. The beamforming calculation method based on deep learning according to claim 1, characterized in that, The acquisition of labeled beamforming datasets for actual wireless communication system scenarios includes: collecting relevant data in actual wireless communication system scenarios as beamforming datasets; or, modeling the actual wireless communication system scenarios as optimization problems and generating beamforming datasets through optimization algorithms.

3. The beamforming calculation method based on deep learning according to claim 1, characterized in that, The aforementioned construction of neural network architectures based on actual application scenarios or needs includes: A neural network architecture is constructed based on a graph neural network, which includes a graph convolutional neural network and a graph attention network. The model of the neural network architecture is improved into a complex model. Complex channel state information is input into the model of the neural network architecture, and the model of the neural network architecture outputs the corresponding beamforming vector, thereby realizing end-to-end learning.

4. The beamforming calculation method based on deep learning according to claim 1, characterized in that, The process of applying the trained neural network architecture to a new wireless communication system scenario and performing beamforming calculations for that scenario includes: Acquire a labeled beamforming dataset for a new wireless communication system scenario. The features of the beamforming dataset include real-time channel information and communication requirements. The labels include the optimal beamforming vector. The samples of different parameters in the beamforming dataset of the new system scenario are input into the trained neural network architecture, and the trained neural network architecture outputs the beamforming vector of the new wireless communication system scenario.

Images