Deep learning-based beamforming device and method scalable to base station antennas and the number of users

Abstract

The present disclosure relates to a deep learning-based beamforming device and method scalable to the number of base station antennas and users. The device includes: a memory storing a process for beamforming based on a plurality of deep learning models scalable to the number of base station antennas and users; and a processor including the plurality of deep learning models and performing an operation according to the process. The processor infers an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on input incomplete channel information, channel error information, and finally, a power value assigned to the base station through the plurality of deep learning models, and selects at least one of the inferred expected power values ​​of the user in order of highest power value, such that the number of expected power values ​​to be selected is less than or equal to the number of base station antennas.

Description

Technology Field

[0001] The present disclosure relates to a general wireless communication system, and specifically to an apparatus and method for solving beamforming technology used in a wireless communication system through deep learning in a Multi User Multiple Input Single Output (MU-MISO) situation. Background Technology

[0002] The number of users in wireless communication is increasing explosively every year, and accordingly, usage is also growing exponentially.

[0003] In this context, Multiple Input Multiple Output (MIMO) systems have great potential to achieve high throughput in wireless systems. While such multiple inputs enable high data rate processing for many users simultaneously, interference between them also exists. Therefore, 5G adopts beamforming technology to effectively eliminate this interference and ensure optimal transmission.

[0004] Figure 1 is a diagram of the user beam forming process through conventional beamforming.

[0005] Referring to FIG. 1, beamforming technology is a technology that improves transmission efficiency by having multiple antennas of a base station recognize a user and form a beam that is focused to a specific user location. As shown in FIG. 1, by focusing the signal, it is a technology that can facilitate reception of radio waves over long distances without amplifying the transmitted power.

[0006] If Channel State Information (CSI) could be perfectly estimated at the base station, an optimal solution would also exist. However, the method for finding the optimal solution requires high complexity and time, which can be considered unsuitable for wireless communication channel conditions that change rapidly in real time.

[0007] Therefore, linear methods have been proposed to find local optimal solutions simply and quickly, and such beamforming algorithms include Zero Forcing (ZF) and Maximum Ratio Transmission (MRT). Although these algorithms do not find the optimal solution, they are methods that achieve beamforming through fast computation with low complexity; naturally, however, they have performance limitations.

[0008] To address this, algorithms such as Weighted Minimum Mean Squared Error (WMMSE), which find local optimal solutions using an iterative method, have been proposed. Although this method exhibits higher complexity than the previous two methods, it demonstrates good performance when calculating the sum rate of data by finding beamforming solutions through an iterative algorithm given channel conditions.

[0009] The algorithms introduced earlier are all based on the assumption that the base station has perfect Channel State Information (CSI). However, in actual wireless communication situations, it is impossible to know the channel conditions perfectly; only the estimated channel conditions obtained through channel estimation are known. This indirectly indicates that algorithms designed under the assumption of perfect channel knowledge suffer performance losses due to incomplete channel estimations. Therefore, it is necessary to develop beamforming algorithms based on Imperfect Channel State Information (ICSI) that mimic real-world wireless communication conditions.

[0010] In this situation, deep learning, which is being successfully applied in various industrial fields, can be used as an effective solution. In the field of beamforming as well, deep learning is being applied to demonstrate high performance similar to existing optimization approaches, while also showing good performance in real-time computation and trained situations.

[0011] However, existing beamforming models based on deep learning have several drawbacks. First, there is a lack of scalability due to the structure of the beamforming model. For instance, in the case of a model that takes a channel as input and a beamforming vector as output, the size of the input channel and the size of the output beamforming vector will be fixed. This results in a model that considers only a single case, as the channel and beamforming vector sizes based on the number of base stations and users cannot be changed. This can be considered unsuitable for wireless communication situations where the number of users and base station antennas varies.

[0012] The second point is that it relies on training data. This means that, as is the case with most deep learning models, it performs well for the data used for training but fails to perform well in other situations.

[0013] In addition, the fact that most deep learning approaches to beamforming are based on CSI models also means that they are models that do not match actual wireless communication situations. Prior art literature

[0014] Korean Published Patent Application No. 10-2022-0029039 (Publication Date: March 8, 2022) The problem to be solved

[0015] The present disclosure performs user selection simultaneously for a model scalable to base station antennas and the number of users. User selection stems from the fact that performance is lower when the number of users exceeds the number of available antennas for the base station. Therefore, if the number of base station antennas is less than the number of users, data must be transmitted by selecting a number of users equal to or less than the number of base station antennas to maximize performance. Such user selection is strictly dependent on channel conditions. Furthermore, the computation for performing such user selection requires complexity and time. Additionally, while optimal results can be obtained when user selection is performed through perfect CSI, this may not be the case with ICSI. However, most deep learning-based beamforming algorithms are models that consider only the case where the number of users is smaller than the number of antennas.

[0016] Therefore, the present disclosure proposes a model that can be effectively scalable even when the number of users exceeds the number of antennas in an ICSI situation, by utilizing this beamforming model for effective user selection even in the case of ICSI.

[0017] Based on the above circumstances, the present disclosure addresses the incomplete channels of ICSI using a deep learning-based model and provides a generalized beamforming model that is not specialized for specific situations (e.g., number of users, number of base station antennas, channel incompleteness) through a beamforming model that is scalable to the number of users and base station antennas.

[0018] That is, the present disclosure can provide a deep learning-based beamforming device and method for optimizing the total user data rate through a model that is scalable to the number of base station antennas and users and robust to channel uncertainty through beamforming via deep learning in an ICSI MU-MISO wireless communication system.

[0019] The problems that this disclosure aims to solve are not limited to those mentioned above, and other unmentioned problems will be clearly understood by a person skilled in the art from the description below. means of solving the problem

[0020] A deep learning-based beamforming device according to the present disclosure, which is scalable to the number of base station antennas and users for achieving the aforementioned technical problem, comprises: a memory storing a process for beamforming based on a plurality of deep learning models that is scalable to the number of base station antennas and the number of users; and a processor that includes the plurality of deep learning models and performs an operation according to the process, wherein the processor infers an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on input incomplete channel information, channel error information, and finally, a power value assigned to the base station through the plurality of deep learning models, and selects at least one of the inferred expected power values ​​of the user in order of highest power value, such that the number of expected power values ​​to be selected is less than or equal to the number of base station antennas.

[0021] At this time, the plurality of deep learning models includes a first model for generating a user message, and the first model can generate a user message by receiving a power value assigned to each user terminal, an antenna message of the corresponding base station, incomplete channel information, channel error information, and finally, a power value assigned to the base station, and merging them in a concat manner.

[0022] In addition, the plurality of deep learning models includes a second model for generating base station antenna messages, and the second model reduces the generated user message by the size of the user, generates an antenna interference message by concating the result of the reduce sum of the existing base station antenna message and the remaining antenna messages excluding the user, and can generate a base station antenna message by receiving the antenna interference message, the reduced sum user message, the incomplete channel information, the channel error information, and the power value finally assigned to the base station as input and merging them by concating.

[0023] In addition, the plurality of deep learning models includes a third model for determining beam power, and the third model can infer an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on the antenna interference message and the result of reducing the sum for the corresponding base station antenna.

[0024] Additionally, the processor may select at least one of the inferred user's expected power values ​​in order of highest power value based on the output values ​​of the first model, the second model, and the third model, such that the number of expected power values ​​to be selected is less than or equal to the number of antennas of the base station.

[0025] In addition, a deep learning-based beamforming method according to the present disclosure, which is scalable to the number of base station antennas and users for achieving the aforementioned technical problem, may include: a step of inferring an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on input incomplete channel information, channel error information, and finally, a power value assigned to the base station through a plurality of deep learning models; and a step of selecting at least one of the inferred expected power values ​​of the user in order of highest power value, wherein the number of expected power values ​​to be selected is less than or equal to the number of antennas of the base station.

[0026] In addition, a computer program stored on a computer-readable recording medium for executing a method for implementing the present disclosure may be further provided.

[0027] In addition, a computer-readable recording medium for recording a computer program for executing a method for implementing the present disclosure may be further provided. Effects of the invention

[0028] According to the aforementioned means for solving the problem of the present disclosure, through an ICSI-based deep learning model similar to actual wireless communication situations compared to existing CSI-based algorithms, the sum rate in ICSI situations can be effectively increased compared to existing ones, thereby providing the effect of showing a higher transmission rate in actual situations.

[0029] The effects of the present disclosure are not limited to those mentioned above, and other unmentioned effects will be clearly understood by a person skilled in the art from the description below. Brief explanation of the drawing

[0030] Figure 1 is a diagram of the user beam forming process through conventional beamforming. FIG. 2 is a diagram showing a base station that performs deep learning-based beamforming scalable to the number of base station antennas and users according to the present disclosure. FIG. 3 is a diagram showing a deep learning structure used for beamforming according to the present disclosure. FIG. 4 is a diagram illustrating algorithm 1 for ICSI beam power selection according to the present disclosure. FIG. 5 is a diagram illustrating algorithm 2 for training an ICSI beam power selection model according to the present disclosure. Figure 6 is a graph comparing the sum rate between a conventional algorithm and a model according to the present disclosure. Figure 7 shows graphs comparing scalability with respect to channel inaccuracy. Figure 8 shows graphs comparing scalability for base station power values. Specific details for implementing the invention

[0031] The advantages and features of the present disclosure and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present disclosure is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to make the present disclosure complete and to fully inform those skilled in the art of the scope of the present disclosure, and the present disclosure is defined only by the scope of the claims.

[0032] The terms used in this specification are for describing embodiments and are not intended to limit the disclosure. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. The terms “comprises” and / or “comprising” as used in this specification do not exclude the presence or addition of one or more other components in addition to the components mentioned. Throughout the specification, the same reference numerals refer to the same components, and “and / or” includes each of the mentioned components and all combinations of one or more. Although terms such as “first,” “second,” etc., are used to describe various components, these components are not limited by these terms. These terms are used merely to distinguish one component from another. Accordingly, the first component mentioned below may be the second component within the technical scope of this disclosure.

[0033] Unless otherwise defined, all terms used herein (including technical and scientific terms) may be used in a meaning commonly understood by those skilled in the art to which this disclosure pertains. Additionally, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0034] Throughout this disclosure, the same reference numerals denote the same components. This disclosure does not describe all elements of the embodiments, and general content in the art to which this disclosure pertains or content that overlaps between embodiments is omitted. As used in the specification, the terms “part” or “module” refer to hardware components such as software, FPGAs, or ASICs, and the “part” or “module” performs certain roles. However, the term “part” or “module” is not limited to software or hardware. The “part” or “module” may be configured to reside in an addressable storage medium or may be configured to run one or more processors. Accordingly, by example, the “part” or “module” includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and "parts" or "modules" may be combined into a smaller number of components and "parts" or "modules," or further separated into additional components and "parts" or "modules."

[0035] Throughout the specification, when a part is described as being "connected" to another part, this includes not only cases where they are directly connected but also cases where they are indirectly connected, and indirect connections include connections made via a wireless communication network.

[0036] Furthermore, when it is stated that a part "includes" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0037] Throughout the specification, when it is stated that a component is located "on" another component, this includes not only cases where a component is in contact with another component, but also cases where another component exists between the two components.

[0038] Terms such as "first," "second," etc., are used to distinguish one component from another, and the components are not limited by the aforementioned terms.

[0039] Singular expressions include plural expressions unless there is an obvious exception in the context.

[0040] In each step, identification codes are used for convenience of explanation and do not describe the order of the steps; the steps may be performed differently from the specified order unless a specific order is clearly indicated in the context.

[0041] FIG. 2 is a diagram showing a base station that performs deep learning-based beamforming scalable to the number of base station antennas and users according to the present disclosure.

[0042] FIG. 3 is a diagram showing a deep learning structure used for beamforming according to the present disclosure.

[0043] FIG. 4 is a diagram illustrating algorithm 1 for ICSI beam power selection according to the present disclosure.

[0044] FIG. 5 is a diagram illustrating algorithm 2 for training an ICSI beam power selection model according to the present disclosure.

[0045] Figure 6 is a graph comparing the sum rate between a conventional algorithm and a model according to the present disclosure.

[0046] Figure 7 shows graphs comparing scalability with respect to channel inaccuracy.

[0047] Figure 8 shows graphs comparing scalability for base station power values.

[0048] With reference to FIGS. 2 through 8, a deep learning-based beamforming process scalable to a number of base station antennas and users according to the present disclosure will be described in detail.

[0049] First, the base station (100) includes a memory (not shown), a communication module (110), and a processor (120), and the processor (120) may include a first model (130) for generating a user message, a second model (140) for generating a base station antenna message, and a third model (150) for determining beam power.

[0050] The memory may be configured to store various information related to the present disclosure. In the present disclosure, the memory may be provided in the base station itself according to the present disclosure. Alternatively, at least a portion of the memory may mean at least one of a database and a cloud storage (or cloud server). That is, the memory is sufficient as a space where information necessary for the device and method according to the present disclosure is stored, and it can be understood that there are no restrictions on the physical space.

[0051] The memory may store a plurality of application programs (or applications) running on the base station (100), data for the operation of the base station (100), and instructions. At least some of these application programs may be downloaded from an external server via wireless communication. Meanwhile, the application programs may be driven to perform operations (or functions) by at least one processor stored in the memory through the processor (120). That is, according to the present disclosure, the memory stores a plurality of deep learning model-based beamforming processes that are scalable to the number of antennas of the base station and the number of users, and the processor (120) may control beamforming operations according to the processes. Meanwhile, the memory may be a flash memory type, a hard disk type, a multimedia card micro type, a card type memory (e.g., SD or XD memory, etc.), RAM (Random Access Memory), SRAM (Static Random Access Memory), ROM (Read-Only Memory), or EEPROM (Electrically Erasable Programmable Memory). It may include at least one type of storage medium among Read-Only Memory, PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk, and optical disk. In addition, the memory may store information temporarily, permanently, or semi-permanently, and may be provided as an embedded or removable type.

[0052] The communication module (110) includes an antenna and can communicate with at least one of each user terminal (h1, h2, h3).

[0053] Next, the processor (120) may be configured to control the overall operation of the device related to the present disclosure. The processor (120) may process signals, data, information, etc. that are input or output through the components described above.

[0054] The processor (120) includes at least one CPU (Central Processing Unit) and can perform the functions according to the present disclosure.

[0055] Specifically, the processor (120) infers an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on input incomplete channel information, channel error information, and finally, a power value assigned to a base station through the first to third models (130, 140, 150), and selects at least one of the inferred expected power values ​​of the user in order of highest power value, such that the number of expected power values ​​to be selected is less than or equal to the number of antennas of the base station.

[0056] In detail, the first model (130) is a deep learning model for generating a user message, and can generate a user message by receiving a power value assigned to each user terminal, an antenna message of the corresponding base station, the incomplete channel information, channel error information, and finally, a power value assigned to the base station, and merging them in a concat manner.

[0057] Additionally, the second model (140) is a deep learning model for generating base station antenna messages, and can generate an antenna interference message by reducing the generated user message by the size of the user, and by concatenating the result of reducing the existing base station antenna message and the remaining antenna messages excluding the user, and can generate a base station antenna message by receiving the antenna interference message, the reduced sum user message, the incomplete channel information, the channel error information, and the power value finally assigned to the base station as input and merging them in a concatenated manner.

[0058] Additionally, the third model (150) is a deep learning model for determining beam power, and can infer an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on the antenna interference message and the result of reducing the sum for the corresponding base station antenna.

[0059] And, the processor (120) can select at least one of the inferred user's expected power values ​​in order of highest power value based on the output values ​​of the first to third models (130, 140, 150), and select such that the number of expected power values ​​to be selected is less than or equal to the number of antennas of the base station.

[0060] That is, the present disclosure assumes a downlink MU-MISO situation in which data is transmitted from a base station (100) to a user, with multiple users having a single antenna and a base station (100) having multiple antennas.

[0061] at N class K represents the base station antenna and the number of users, respectively. And the user k It receives signals from N antennas of the base station, and this It undergoes the channel vector. At this time, the transmit beamforming vector is Defined as. means a complex number. And actually, the user k The signal received by can be represented as shown in the following mathematical formula 1.

[0062]

[0063] If represents an additional noise signal to the received signal, the mean is 0 and the variance is am. is actually the user k It refers to a data symbol for, and is data transmitted by the base station (100). And for each user, a channel Beamforming vector for forming a beam according to Creates.

[0064] Here k Signal-to-Interference-plus-Noise Ratio (SINR) of the th user It can be expressed as shown in the following mathematical formula 2.

[0065]

[0066] And users according to SINR k The data rate of can be expressed as shown in the following mathematical formula 3.

[0067]

[0068] That is, the present disclosure aims to maximize the sum rate of the data of the system model as shown in Equation 4 below.

[0069]

[0070] Here, the constraint P is the maximum transmit power available for beamforming at the base station.

[0071] And in this disclosure, assuming an ICSI situation, the channel received by the base station (100) can be represented as Equation 5 below.

[0072]

[0073] Here is a random variable and possesses statistical characteristics such as a probability distribution. Furthermore, the base station performs beamforming based on a channel such as Mathematics 5, rather than perfect CSI. Also, to assume an ICSI situation in this disclosure, channel error It follows a Gaussian variance with a mean of 0 and variance Follows, and the variance It follows. This is an error ratio factor representing the ratio of CSI error in terms of channel gain.

[0074] The following is a description of the deep learning model used. First, to effectively train the deep learning model, the optimal beamforming structure formula is used as shown in Equation 6 below.

[0075]

[0076] At this time, is user k Indicates the download beamforming power for. And, represents virtual uplink power. And, It satisfies. Based on this formula, the deep learning model receives incomplete channel information as input and Infer the users' downlink beamforming power and virtual uplink power through this.

[0077] That is, as shown in FIG. 3, the first to third models (130, 140, 150) are composed of three deep neural networks (DNN), each being a User Neural Network, an Antenna Neural Network, and a Power Decision Neural Network.

[0078] Referring to Algorithm 1 for ICSI beam power selection in Fig. 4, Each represents the user power allocation and base station antenna message that are initially set arbitrarily. With a size, M represents the Embedding Size, which is a hyperparameter that can be arbitrarily set. And, is of mathematical formula 6 It is a beamforming allocation power parameter that is primarily learned and inferred in a deep learning model composed of

[0079] represent the User Message Generation Neural Network, the Antenna Message Generation Neural Network, and the Beam Power Decision Neural Network, respectively. And, It uses Reduce Sum as the pooling operator. Reduce Sum is an operation that adds all N data points to form a single data point. Finally, is a function that merges two or more arrays.

[0080] First, the first model (130) corresponding to number 1 of Algorithm 1 has incomplete channel information ( ) and channel error information( It receives ) and, finally, dB, which is the power P allocated to the base station, as input. No. 1 of Algorithm 1 represents the generation of a user message by a deep learning model. The neural network located at the very front of Fig. 3 ( As the input at the t-th position, ) The user, base station antenna message and It receives all of its information as input. Therefore, for example, the size becomes M + 6 [User message 2, antenna message M, channel information 2 (separated into real and imaginary values ​​for deep learning processing), channel error information 1, power information 1]. The new user that emerges after passing through two layers in this way The message size of is M.

[0081] The second model (140) corresponding to number 2 of Algorithm 1 generates an antenna message of the base station. First, the second model (140) Reduce Sum by the user size.

[0082] And, the second model (140) is It generates an antenna interference message saying... The above antenna interference message is the existing base station antenna message. and users k The Reduce Sum of the remaining antenna messages excluding It consists of.

[0083] And, the base station antenna message is the antenna interference message and user messages Neural network for generating base station antennas that takes all of them as input The input size is 3M + 4. And the output comes out The size of is M. Based on this, the second model (140) is a new antenna interference message. Creates.

[0084] And, number 3 of Algorithm 1 is the beam power decision neural network It represents. That is, the third model (150) is an antenna interference message. Performed Reduce Sum on base station antenna i It uses as input. The final output of the neural network resulting from this is as It means.

[0085] The third model (150) is a new beamforming vector through the t-th beam power thus generated and Equation 6. It is possible to generate and, through this, obtain the data rate through mathematical formulas 2 and 3.

[0086] And, number 4 of Algorithm 1 is an algorithm performed by a processor (120) related to user selection. That is, the present disclosure is an algorithm designed so that the processor (120) can make a selection while performing iterations for effective processing in ICSI.

[0087] In other words, if there are actually more users than the number of antennas of the base station (100), only users with good channels are selected and sent based on the estimated channels. Therefore, the maximum number of users (100) that can be selected without performance degradation becomes equal to the number of antennas of the base station. Of course, this method of selecting and sending users with good channels also requires complexity and additional time, and this also needs to be processed within deep learning.

[0088] To solve this, the present disclosure infers that first to third models (130, 140, 150) infer an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on incomplete channel information, channel error information, and finally, a power value assigned to a base station, and a processor (120) selects at least one of the inferred expected power values ​​of the user in order of highest power value, such that the number of expected power values ​​to be selected is less than or equal to the number of antennas of the base station.

[0089] That is, in the case where the number of base station antennas is large, the upper beam power in the beam power according to the inference results of the first to third models (130, 140, 150) By selecting only users with good channels, the Sum rate is improved. And, since maximizing the Sum rate transmitted is optimal when the number of users is equal to or less than the number of antennas, as the iteration proceeds, if the number of selected users and the number of antennas are the same, the more optimal case is selected by comparing it with the case where the number of users is one less.

[0090] Most deep learning-based beamforming models use only data from channel datasets where user selection has already been completed, or consider only cases where the number of users is less than the number of antennas. However, since user selection based on ICSI can also lead to performance degradation, user selection suitable for such ICSI situations is also necessary.

[0091] The input and output sizes and detailed parameters of the first to third models (130, 140, 150) follow Table 1 below.

[0092]

[0093] Table 1 shows the detailed number of parameters of the model used.

[0094] In the first to third models (130, 140, 150), information regarding channel error and base station power was added to the inputs used to generate user and base station antenna messages. Through this method, the existing models, which were scalable only for the user and base station antenna, were made scalable for base station power and various channel error situations. Additionally, Layer Normalization was applied to each layer to ensure that information is not lost as it passes through large Dense layers for these various inputs.

[0095] Next, referring to Algorithm 2 for ICSI beam power selection in Fig. 5, represents the objective function used for learning. That is, It means.

[0096] In training, in particular, it must be trained scalable with respect to the number of existing users and base station antennas, but additionally, the total power P of the base stations and channel inaccuracy As expected, since there are significant variations in wireless communication channels, all these parameters must be able to operate scalably after training.

[0097] Therefore, during training, randomly generated Error channel information based on [the source] must be used, and dB, i.e., base station power, must also be randomly selected and included. And a Validation Set for learning. Also randomly generated It is created through this, and for accurate comparison, the Sum rate obtained by dividing the base station's allocated power in dB for the same channel into specific first power value (e.g., 10dB) and specific second power value (e.g., 25dB) cases, respectively, is added to obtain the final Get . This existing In the case of being larger, This new It becomes, and the model parameters at that time It stores. And the Loss Function for learning is a negative objective function. am.

[0098] As explained above, the present disclosure describes in detail the structure and learning method for a beamforming model that is scalable to the number of base station antennas and users in MU-MISO and effective in ICSI situations.

[0099] That is, the present disclosure additionally uses channel error information and base station allocated power as inputs in the user message generation and base station antenna message generation of Algorithm 1 to create an effective beamforming model in an ICSI situation. Layer Normalization is applied to each layer so that information is not lost as it passes through large Dense layers for various inputs of the model used.

[0100] And in the case of Algorithm 2, randomly generated during training It aims for a generalized model rather than a model applicable only to specific situations by using error channel information based on and randomly selecting and including dB, i.e., base station power. Additionally, the Validation Set created to verify the effectiveness of the learning is a model that generalizes scalability with respect to dB by adding both cases of 10dB and 25dB for the same channel.

[0101] In addition, existing deep learning beamforming models select a good channel (a state of selecting users) based on channel information when there are more users than base station antennas. However, the present disclosure sequentially selects users with the highest beam power during deep learning iterations from among users who have not yet been selected in a situation where the channel is inaccurate. Consequently, the base station can perform user selection and beamforming simultaneously through the model without needing to pre-select users based on channel conditions. In particular, if user selection is also based on an inaccurate channel, this can lead to performance degradation. This allows the deep learning-based user selection technique to be effectively used in conjunction with beamforming in ICSI situations.

[0102] Referring to FIGS. 6 through 8, to demonstrate that the sum rate of other models in the present disclosure is effective, a user is randomly placed within a 100m cell. Small-scale fading of the channel follows a Rayleigh fading channel, and large-scale fading is Following, d k means the distance between the user and the base station and is. And, Thus, the SNR is equal to the base station power P. Also, M = 5 and the model's iteration T = 10. And, It was set to the case of [0.0, 0.1, 0.3, 0.5, 1.0].

[0103] And, the Adam Algorithm was used for training, and the learning rate was 0.0005. The maximum antenna and user count at = 1000 The model runs 50 mini-batch sets per epoch for 500 epochs.

[0104] 5,000 test sets were used to demonstrate actual performance.

[0105] As shown in Figure 6, the beamforming inference based on the ICSI information received by the actual base station demonstrates better performance in ICSI situations compared to existing beamforming models based on CSI.

[0106] This confirms that the model, having actually trained through inference using only potentially inaccurate ICSI information, demonstrates better performance compared to other CSI-based beamforming methods, and this effect becomes more pronounced as channel inaccuracy increases.

[0107] Figures 7 and 8 are performance graphs related to Scalable, and the solid line in Figure 7 is The dotted line represents a scalable model, while the dotted line represents a model trained specifically for that situation using only the corresponding channel data. Nevertheless, it can be seen that it exhibits nearly similar performance when compared to the dotted line. Furthermore, in ICSI scenarios where the number of users exceeds the number of antennas (N>K), it can be observed that user selection is performed normally without significant performance degradation.

[0108] Figure 8 also shows a comparison between the red line, which was scalably trained at dB, and the models trained at 10dB and 25dB, respectively. It was confirmed that there was no significant difference in performance between the scalably trained model and the models trained at 10dB and 25dB at almost all dB levels.

[0109] For reference, the lines marked in red in Figures 6 and 8 and the solid line in Figure 7 are all a single model consisting of the same parameters of the same model.

[0110] The deep learning-based beamforming method scalable to the number of base station antennas and users according to the present disclosure described above may be designed in the form of a program, and the program may include code coded in a computer language such as C, C++, JAVA, or machine language that can be read by the processor (CPU) of the computer through the device interface of the computer so that the computer reads the program and executes the methods implemented in the program. Such code may include functional code related to functions that define the necessary functions for executing the methods, and may include control code related to execution procedures necessary for the processor of the computer to execute the functions according to a predetermined procedure. Furthermore, such code may further include memory reference code regarding where (address) additional information or media necessary for the processor of the computer to execute the functions should be referenced in the internal or external memory of the computer. In addition, if the processor of the computer needs to communicate with any other computer or server located remotely in order to execute the above functions, the code may further include communication-related code regarding how to communicate with any other computer or server located remotely using the communication module of the computer, and what information or media to transmit or receive during communication.

[0111] The above-mentioned storage medium refers to a medium that stores data semi-permanently and is readable by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. Specifically, examples of the above-mentioned storage medium include, but are not limited to, ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage device. That is, the above-mentioned program may be stored on various recording media on various servers that the computer can access, or on various recording media on the user's computer. Additionally, the above-mentioned medium may be distributed across networked computer systems, and computer-readable code may be stored in a distributed manner.

[0112] The steps of the method or algorithm described in connection with the embodiments of the present disclosure may be implemented directly in hardware, implemented as a software module executed by hardware, or implemented by a combination thereof. The software module may reside in RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), Flash Memory, a hard disk, a removable disk, a CD-ROM, or any form of computer-readable recording medium well known in the art to which the present disclosure belongs.

[0113] Although embodiments of the present disclosure have been described above with reference to the attached drawings, those skilled in the art will understand that the present disclosure may be implemented in other specific forms without altering its technical concept or essential features. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive.

Claims

Claim 1 Memory storing processes for beamforming based on multiple deep learning models that are scalable to the number of base station antennas and the number of users; and a processor including the plurality of deep learning models and performing operations according to the process; wherein the plurality of deep learning models include a first model for generating a user message, a second model for generating a base station antenna message, and a third model for determining beam power; the processor receives a power value assigned to each user terminal, an antenna message of the corresponding base station, input incomplete channel information, channel error information, and finally, a power value assigned to the base station through the first model, and merges them in a concat manner to generate a user message; through the second model, it reduces the generated user message by the user size and generates an antenna interference message in a concat manner based on the result of the reduce sum of the existing base station antenna message and the remaining antenna messages excluding the corresponding user; receives the antenna interference message, the reduced sum user message, the incomplete channel information, the channel error information, and finally, a power value assigned to the base station, and merges them in a concat manner to generate a base station antenna message; and through the third model, it infers an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on the result of the reduce sum for the antenna interference message and the corresponding base station antenna; and the first A deep learning-based beamforming device scalable to base station antennas and the number of users, wherein at least one of the inferred user's expected power values ​​is selected in order of highest power value based on the output values ​​of the model, the second model, and the third model, and the number of expected power values ​​to be selected is less than or equal to the number of base station antennas. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 delete Claim 6 A deep learning-based beamforming method scalable to the number of base station antennas and users, performed by a processor of a device, comprising: a step in which the processor infers an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on input incomplete channel information, channel error information, and finally, a power value assigned to a base station, through a plurality of deep learning models; and a step in which the processor selects at least one of the inferred expected power values ​​of the user in order of highest power value, such that the number of selected expected power values ​​is less than or equal to the number of antennas of the base station.The plurality of deep learning models includes a first model for generating a user message, a second model for generating a base station antenna message, and a third model for determining beam power. The processor receives a power value assigned to each user terminal, an antenna message of the corresponding base station, input incomplete channel information, channel error information, and finally, a power value assigned to the base station through the first model, and merges them in a concat manner to generate a user message. Through the second model, it reduces the generated user message by the user size and generates an antenna interference message by a concat method on the result of the reduce sum of the existing base station antenna message and the remaining antenna messages excluding the user. It receives the antenna interference message, the reduced sum user message, the incomplete channel information, the channel error information, and finally, the power value assigned to the base station, and merges them in a concat manner to generate a base station antenna message. Through the third model, it infers an expected power value including a downlink beamforming power value and a virtual uplink power value for each user based on the antenna interference message and the result of the reduce sum for the corresponding base station antenna. Based on the output values ​​of the first model, the second model, and the third model, the A deep learning-based beamforming method scalable to the number of base station antennas and users, wherein at least one of the inferred user's expected power values ​​is selected in order of highest power value, such that the number of selected expected power values ​​is less than or equal to the number of base station antennas. Claim 7 delete Claim 8 delete Claim 9 delete Claim 10 delete

Images