Electronic apparatus, artificial neural model learning method, and uniform multi-beam generation method

Abstract

An artificial neural network model learning method performed by an electronic device according to one embodiment includes the steps of: generating learning data by simulating digital intermediate frequency data that can be generated through digital signal processing after being received by an array antenna including a plurality of antennas according to frequency and azimuth angle; and training an artificial neural network model for broadband frequency uniform multi-beam generation using the generated learning data, wherein the weight vector of the artificial neural network model is regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for the generation of the desired multi-beam.

Description

Technology Field

[0001] The present invention relates to an electronic device comprising an artificial neural network model, a method for the artificial neural network model of the electronic device to learn, and a method for the electronic device comprising the artificial neural network model learned in this way to generate a frequency-uniform multibeam. Background Technology

[0002] Multiple Input Multiple Output (MIMO) systems can improve system performance and maximize communication capacity in a wireless communication environment while utilizing the same frequency resources and power.

[0003] Furthermore, receive beamforming, a type of beamforming, is performed to improve signal reception performance by matching the phase of the receiver to the received signal when receiving a signal by a multiple-input multiple-output (MIMO) system. This receive beamforming is modeled by multiplying the received signal by a receive beamforming weight vector.

[0004] A processing unit for receiving beamforming requires an array antenna section composed of multiple antennas, a multi-channel receiver, and a digital signal processing unit. In the array antenna section used for beamforming, the beam width is determined by the ratio of the wavelength of the received signal to the fixed antenna spacing; generally, the beam width formed narrows as the frequency changes from a low frequency band to a high frequency band. This characteristic implies that the azimuth range for simultaneous signal reception decreases as the frequency increases. Therefore, to effectively receive signals within a constant azimuth range across all required frequency bands, a signal processing technique for frequency-uniform beam synthesis is required to ensure a uniform beam width regardless of frequency variations.

[0005] A method for generating beamforming weight vectors for receiving beam synthesis according to conventional technology is calculated using a linear method such as maximum ratio combining (MRC) or minimum mean squared error (MMSE).

[0006] Conventional reception beam synthesis methods like this have limitations in generating a uniform beam across a wide frequency band because the beam width is determined by the ratio of the received signal wavelength to the fixed antenna spacing. To overcome these physical limitations and generate a uniformly synthesized reception beam width across a desired frequency band, research based on a data approach utilizing an artificial neural network structure has been conducted.

[0007] Research based on data approaches using artificial neural network structures concerns mathematical models and statistical learning algorithms created by analyzing and referencing the functions of neurons in the biological brain. It is widely used in many application fields, such as data interpolation, pattern recognition, time series modeling, control engineering, electronic equipment parameter modeling, 3-D object modeling, speech recognition, and channel equalization. In particular, it is advantageous for finding the optimal function that better represents the objective function compared to existing model-based methods, as it performs learning based on sufficient data rather than mathematical formulas and can implement non-linear functions by stacking layers containing non-linear activation functions. Prior art literature

[0008] Republic of Korea Published Patent Application No. 10-2022-0013906, Date of publication February 4, 2022. The problem to be solved

[0009] According to one embodiment, a method for training an artificial neural network model and an electronic device thereof are provided, which train the artificial neural network model to generate a broadband frequency uniform multiple beam using training data generated by simulating digital intermediate frequency data.

[0010] In addition, the present invention provides a method for generating a broadband frequency uniform multi-beam and an electronic device thereof by inputting digital intermediate frequency data, generated through digital signal processing after being received by an array antenna according to frequency and azimuth angle, into a pre-trained artificial neural network model.

[0011] The problems to be solved by the present invention are not limited to those mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art to which the present invention pertains from the description below. means of solving the problem

[0012] A method for training an artificial neural network model performed by an electronic device according to the first aspect comprises the steps of: generating training data by simulating digital intermediate frequency data that can be generated through digital signal processing after being received by an array antenna including a plurality of antennas according to frequency and azimuth angle; and training an artificial neural network model for generating a broadband frequency uniform multi-beam using the generated training data, wherein the weight vector of the artificial neural network model is trained recursively through a target beam pattern-based loss function and a target weight vector-based loss function required for generating a desired multi-beam.

[0013] An electronic device according to the second aspect includes a memory comprising an artificial neural network model for generating a broadband frequency uniform multi-beam, and a processor that performs the function of said artificial neural network model. The processor generates training data by simulating digital intermediate frequency data that can be generated through digital signal processing after being received by an array antenna comprising a plurality of antennas according to frequency and azimuth angle, and trains said artificial neural network model using said generated training data. The weight vector of said artificial neural network model is regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for generating a desired multi-beam.

[0014] A method for generating a broadband frequency uniform multi-beam performed by an electronic device according to a third perspective includes the step of preparing digital intermediate frequency data generated through digital signal processing after receiving by frequency and azimuth angle at an array antenna including a plurality of antennas, and the step of inputting the digital intermediate frequency data into a pre-trained artificial neural network model to generate a broadband frequency uniform multi-beam as the output of the artificial neural network model, wherein the weight vector of the artificial neural network model is regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for generating a desired multi-beam.

[0015] An electronic device according to the fourth aspect includes a memory containing an artificial neural network model that has been trained for generating a broadband frequency uniform multi-beam, and a processor that performs the function of said artificial neural network model. The processor inputs digital intermediate frequency data generated through digital signal processing after being received by an array antenna including a plurality of antennas according to frequency and azimuth angle into said artificial neural network model to generate a broadband frequency uniform multi-beam as the output of said artificial neural network model, and the weight vector of said artificial neural network model is regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for generating a desired multi-beam.

[0016] According to the fifth aspect, a computer-readable recording medium storing a computer program, said computer program includes instructions for the processor to perform an artificial neural network model learning method performed by an electronic device when executed by a processor.

[0017] According to the sixth aspect, a computer program stored on a computer-readable recording medium, said computer program includes instructions for the processor to perform an artificial neural network model learning method performed by an electronic device when executed by a processor.

[0018] According to the seventh aspect, a computer-readable recording medium storing a computer program, said computer program includes instructions for the processor to perform a broadband frequency uniform multi-beam generation method performed by an electronic device when executed by a processor.

[0019] According to the eighth aspect, a computer program stored on a computer-readable recording medium, said computer program includes instructions for the processor to perform a broadband frequency uniform multi-beam generation method performed by an electronic device when executed by a processor. Effects of the invention

[0020] According to one embodiment, an artificial neural network model formed for generating a broadband frequency uniform multi-beam can be trained by simulating digital intermediate frequency data that can be generated through digital signal processing after being received by an array antenna according to frequency and azimuth angle. Then, by inputting the digital intermediate frequency data generated through digital signal processing after being received by the array antenna according to frequency and azimuth angle into the trained artificial neural network model, a broadband frequency uniform multi-beam with a guaranteed uniform beam width regardless of frequency variation can be synthesized as the output of the trained artificial neural network model. Brief explanation of the drawing

[0021] FIG. 1 is a configuration diagram of an electronic device capable of performing an artificial neural network model learning method and / or a frequency uniform multi-beam generation method according to an embodiment of the present invention. FIG. 2 is a flowchart illustrating an artificial neural network model learning method according to an embodiment of the present invention. FIGS. 3 to 5 are drawings illustrating various examples of learning weight vectors according to embodiments of the present invention. FIG. 6 is a conceptual diagram illustrating a frequency-uniform multi-beam generation method according to an embodiment of the present invention. FIG. 7 is a diagram for comparing different uniform beam patterns in an embodiment of the present invention. Specific details for implementing the invention

[0022] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims.

[0023] The terms used in this specification will be briefly explained, and the invention will be described in detail.

[0024] The terms used in this invention have been selected based on currently widely used general terms, taking into account their functions within the invention; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been arbitrarily selected by the applicant, and in such cases, their meanings will be described in detail in the relevant description of the invention. Therefore, the terms used in this invention should be defined not merely by their names, but based on their meanings and the overall content of the invention.

[0025] When a part of a specification is described as 'comprising' a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0026] Additionally, the term "part" as used in the specification refers to software or hardware components, such as FPGAs or ASICs, and the "part" performs certain roles. However, the meaning of "part" is not limited to software or hardware. The "part" may be configured to reside in an addressable storage medium or configured to run one or more processors. Thus, by example, the "part" 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" may be combined into a smaller number of components and "parts" or further separated into additional components and "parts."

[0027] Below, embodiments of the present invention are described in detail with reference to the attached drawings so that those skilled in the art can easily implement the invention. Additionally, parts of the drawings that are irrelevant to the description are omitted to clearly explain the invention.

[0028] FIG. 1 is a configuration diagram of an electronic device capable of performing an artificial neural network model learning method and / or a frequency uniform multi-beam generation method according to an embodiment of the present invention.

[0029] Referring to FIG. 1, the electronic device (100) may include memory (110) and a processor (120). The electronic device (100) may be implemented as various devices such as a desktop PC, a laptop PC, a tablet PC, a smartphone, a server device, etc.

[0030] Information related to various functions or instructions of the electronic device (100) can be stored in the memory (110). In addition to ROM and RAM, the memory (110) may include a hard disk, SSD, flash memory, etc.

[0031] One or more artificial intelligence models may be stored in the memory (110). The artificial intelligence model may be composed of multiple neural network layers. Each of the multiple neural network layers has multiple weight values ​​and performs neural network operations through the operation of the multiple weights and the operation of the previous layer. The multiple weights of the multiple neural network layers may be optimized by the learning results of the artificial intelligence model. This memory (110) may include an artificial neural network model for broadband frequency uniform multi-beam generation. This artificial neural network model will be described with reference to FIGS. 3 to 5. The functions of the artificial neural network model stored in this memory (110) can be performed through the processor (120) and the memory (110).

[0032] The processor (120) may be composed of one or more processors. For example, one or more processors may be a general-purpose processor such as a CPU (central process unit) or DSP (digital signal processor), a graphics-dedicated processor such as a GPU (graphic process unit) or VPU (vision process unit), or an artificial intelligence-dedicated processor such as an NPU (neural process unit).

[0033] Meanwhile, the electronic device (100) may further include a display unit for providing processing results by the processor (120). Alternatively, it may further include a communication unit capable of transmitting processing result data by the processor (120) to an external device.

[0034] FIG. 2 is a flowchart illustrating an artificial neural network model learning method according to an embodiment of the present invention.

[0035] Referring to FIG. 2, the artificial neural network model learning method according to the embodiment includes the step (S210) of generating learning data by simulating digital intermediate frequency data that can be generated through digital signal processing after being received by an array antenna including a plurality of antennas according to frequency and azimuth angle.

[0036] And, the artificial neural network model learning method according to the embodiment further includes the step (S220) of performing a complex correlation matrix operation on digital intermediate frequency data, which is simulated learning data.

[0037] And, the artificial neural network model training method according to the embodiment further includes the step (S230) of inputting training data into the artificial neural network model to train the artificial neural network model for broadband frequency uniform multi-beam generation.

[0038] To facilitate understanding of the artificial neural network model learning method according to the embodiment of the present invention, we will briefly refer to the conceptual diagram in FIG. 6 for explaining the frequency-uniform multi-beam generation method according to the embodiment of the present invention. FIG. 6 shows an example in which the weight vector of the artificial neural network model learned through FIG. 2 is applied to a known digital signal processing module for multi-beam generation. Let us examine frequency-uniform multi-beam generation with reference to FIG. 6.

[0039] High-frequency RF signals received by array antennas at various frequencies and azimuths are down-converted into intermediate frequency (IF) signals at a frequency suitable for signal processing and input into a digital signal processing module. The down-converted analog intermediate frequency signals input via multiple channels are sampled into digital intermediate frequency data for digital signal processing by an analog-to-digital converter (ADC). The sampled data then passes through a decimator and a filter to perform frequency conversion suitable for signal processing and to attenuate high-frequency noise.

[0040] Weight vector operations for ensuring time delay / gain matching per frequency and for receiving beam synthesis per frequency / azimuth angle may use an integer delay block for coarse time delay control, a fractional delay (FD) block for fine time delay control, and a phase shift weight block for applying beam synthesis weight vectors that vary per azimuth angle. To ensure time delay matching, correction data stored in the phase weight ROM table is referenced and applied to the FD block, and to ensure channel gain matching, correction data stored in the channel gain weight ROM table is referenced and applied to the phase shift weight block.

[0041] According to one embodiment of the present invention, the artificial neural network model learning method can learn the weight vector of the artificial neural network model that can be applied to the phase shift weight block through the artificial neural network model learning method of FIG. 2. In this way, when the previously learned weight vector is applied, a broadband frequency uniform multiple beam can be synthesized in which a uniform beam width is guaranteed regardless of frequency change.

[0042] FIGS. 3 to 5 are drawings illustrating various examples of learning weight vectors according to embodiments of the present invention.

[0043] FIG. 3 is an artificial neural network structure trained with a weight vector-based loss function for generating a broadband frequency uniform multi-beam based on an artificial neural network according to an embodiment of the present invention. For example, the artificial neural network structure can be a Shallow Neural Network based on a thin hidden layer and a Deep Neural Network based on a deep hidden layer. For artificial neural network training, instead of directly using the digital signal processing module of FIG. 6, digital intermediate frequency data received by the array antenna by frequency and azimuth angle and input / processed by the digital signal processing module is simulated / generated to utilize artificial neural network training data.

[0044] Multi-channel artificial neural network training data is processed into a complex correlation matrix (301) and input into an artificial neural network (302). The weight vector for generating a broadband frequency uniform multi-beam is regressively trained through a loss function based on a target weight vector required for generating the desired multi-beam (303). Through the artificial neural network that is updated and advanced through training, a weight vector with minimized error with the target weight vector is obtained. The target weight vector (304) may include weight vectors based on existing linear and non-linear techniques suitable for beam synthesis, and various techniques may be selected, thereby enabling the generation of a broadband frequency uniform multi-beam (305).

[0045] In the description of this embodiment, for example, a weight vector determined through the LCMV (linearly constrained minimum variance) algorithm for broadband frequency uniform multi-beam synthesis was selected as the target weight vector.

[0046] To explain the LCMV algorithm, having an arbitrary form array antenna composed of dog sensors Received as an azimuth Consider multiple signal sources (including signals and interference). Output of the array beam synthesizer It can be expressed as follows.

[0047]

[0048] Optimal weight The problem of finding can be summarized as a constraint-type LMS (Least Mean Squares) problem as follows.

[0049]

[0050] Here, is the constraint matrix, and is a constraint vector. For example, Open azimuth If it is necessary to generate unit gain and form nulls for other azimuths, the constraint matrix and constraint vector can be expressed as follows.

[0051]

[0052]

[0053] Based on mathematical equations (2) and (3), the optimal weight vector of the LCMV beam synthesizer can be obtained using Lagrange multipliers as follows.

[0054]

[0055] FIG. 7 is a result illustrated for the comparison of uniform beam patterns according to an embodiment of the present invention, and FIG. 7(a) is a result illustrating multiple beams by frequency and azimuth when the optimal weight vector of the LCMV beam synthesizer is applied to the phase-shifting weight block of the receiver structure of FIG. 1. Such a conventional linear technique-based multiple beam exhibits a characteristic in which the width of the synthesized beam narrows as the frequency increases.

[0056] According to an embodiment, an artificial neural network structure trained with a weight vector-based loss function was implemented using the optimal weight vector of an LCMV beam synthesizer as the target weight vector for generating the broadband frequency uniform multiple beams of FIG. 3. The loss function used for artificial neural network training can be expressed as follows.

[0057]

[0058] In mathematical formula 5 Is It refers to an artificial neural network function parameterized as , and the target used follows mathematical formula 4. At this time is an artificial neural network input vector, and represents the number of data.

[0059] In the results illustrated for the comparison of uniform beam patterns according to the embodiment of the present invention in Fig. 7, Fig. 7(b) shows the results of applying the weight vector obtained by training the target weight vector using the optimal weight vector of the LCMV beam synthesizer to the phase-shifting weight block of the receiving structure exemplified in Fig. 6, thereby illustrating the multiple beams by frequency and azimuth angle. Unlike Fig. 7(a), it can be seen that a frequency-uniform multiple beam pattern is observed in the frequency band of 1.5 GHz or higher. However, in the low frequency band of 1.5 GHz or lower, it was confirmed that the beam width of the multiple beam gradually widens, resulting in a phenomenon similar to the results caused by the limitations of the existing linear method.

[0060] FIG. 4 is an artificial neural network structure trained with a uniform beam pattern-based loss function for generating a broadband frequency uniform multiple beam based on an artificial neural network according to an embodiment of the present invention. In this case, the artificial neural network structure can be a Shallow Neural Network based on a thin hidden layer and a Deep Neural Network based on a deep hidden layer. For artificial neural network training, instead of directly using the digital signal processing module exemplified in FIG. 6, digital intermediate frequency data received by an array antenna according to frequency and azimuth angle and input / processed by a digital signal processing unit is simulated / generated to utilize artificial neural network training data.

[0061] Similar to the embodiment of FIG. 3, multi-channel artificial neural network training data is processed into a complex correlation matrix (301) and input into an artificial neural network (302). A weight vector for generating a broadband frequency uniform multi-beam is regressively trained (303, 401) through a target beam pattern-based loss function required for generating the desired multi-beam. Through the artificial neural network that is updated and advanced through training, a weight vector is obtained that can form a beam (305) with a minimized error with the target beam pattern beam. As for the target beam pattern, a desired beam width shape can be selected and obtained mathematically, then copied across the entire band and used as a target for training.

[0062] To train the network structure according to the embodiment, an LCMV weight vector of a specific frequency band capable of satisfying the beam width of the uniform multi-beam was used as a target beam pattern for generating the broadband frequency uniform multi-beam of FIG. 4, and through this, an artificial neural network structure trained with a uniform beam pattern-based loss function was implemented. The beam pattern loss function formula used is as follows.

[0063]

[0064] In mathematical formula 6 Is It refers to an artificial neural network function parameterized as , and the target used represents the target beam radiated over the entire band after forming a beam using an LCMV weighting vector of a specific frequency band capable of forming a desired beam width. In this case, similar to the preceding Equation 5, is the input to the artificial neural network, and represents the number of data.

[0065] In the results illustrated for the comparison of uniform beam patterns according to the embodiment of the present invention in Fig. 7, Fig. 7(c) shows the results of applying the weight vector obtained by training with a target beam pattern at a frequency of 1.1 GHz using the optimal weight vector of an LCMV beam synthesizer to the phase shift weight block of the receiver structure in Fig. 1, showing multiple beams by frequency and azimuth angle. Unlike Figs. 7(a) and 7(b), it can be seen that Fig. 7(c) exhibits a frequency-uniform multiple beam pattern in most frequency bands. However, when training a network structure based on such a uniform beam pattern, the process proceeds with the weight vector value being learned according to the beam pattern, resulting in a relatively slow learning speed. Although it is improved from the perspective of the uniform beam waveform, the reception bit error rate performance through demodulation may not be guaranteed.

[0066] To overcome such limitations, the embodiment may apply a hybrid learning method that uses a target beam pattern to generate a broadband frequency uniform multiple beam and a target weight vector to guarantee reception bit error rate performance.

[0067] FIG. 5 is an artificial neural network structure trained with a uniform beam pattern-based loss function and a weight vector-based loss function for generating a broadband frequency uniform multiple beam based on an artificial neural network according to an embodiment of the present invention. In this case, it is possible to apply a Shallow Neural Network based on a thin hidden layer and a Deep Neural Network based on a deep hidden layer to the artificial neural network structure. For artificial neural network training, instead of directly using the receiver of FIG. 1 described above, digital intermediate frequency data received by an array antenna according to frequency and azimuth angle and input / processed by a digital signal processing module is simulated / generated to utilize artificial neural network training data.

[0068] In order to train the network structure according to an embodiment of the present invention, a hybrid artificial neural network training method was implemented in which a uniform beam pattern-based loss function is applied using an LCMV weight vector of a specific frequency band capable of satisfying the beam width of the uniform multi-beam as a target beam pattern for generating the broadband frequency uniform multi-beam of FIG. 5, and at the same time, a weight vector-based loss function is trained using the optimal weight vector of the LCMV beam synthesizer as the target weight vector, and the following mathematical formula was applied as the loss function to reflect both training methods.

[0069]

[0070] Here is the total loss, and Equation 5 applied is the loss for the target weight vector, and Equation 6 applied is the loss for the target beam pattern. is the weight of the relative loss for the target beam pattern relative to the target weight vector, and can have a value between 0 and 1. Weight As the value approaches 0, the uniform beam pattern generation performance decreases, but the received bit error rate decreases, and As the value approaches 1, the uniform beam pattern generation performance increases, but the received bit error rate also increases. Therefore By adjusting the value, the trade-off between uniform beam pattern performance and the received bit error rate can also be controlled.

[0071] According to the embodiment of FIG. 5, it is possible to effectively receive a signal within a constant azimuth range by securing a uniform beam width regardless of frequency variation while satisfying the reception bit error rate, and this can be utilized as a method to secure spatial selection freedom for the signal. A hybrid learning method that uses the target beam pattern described above and a target weight vector to guarantee reception bit error rate performance It was experimentally confirmed that when applied through Equation 7, it is possible to generate a wideband frequency uniform multi-beam while simultaneously securing the required reception bit error rate performance.

[0072] In the result illustrated for the comparison of uniform beam patterns according to an embodiment of the present invention in FIG. 7, FIG. 7(d) shows a hybrid artificial neural network learning method that applies a uniform beam pattern-based loss function to an LCMV beam synthesizer target beam pattern at a frequency of 1.1 GHz and simultaneously learns with a weight vector-based loss function for an LCMV beam synthesizer target weight vector. The result shows the multi-beams plotted by frequency and azimuth angle when the weight vector obtained by applying the method is applied to the phase shift weight block of the receiving structure exemplified in Fig. 6. In the case of Fig. 7(d), it can be seen that it shows a pattern that is intermediate between the result of Fig. 7(b) and the result of Fig. 7(c).

[0073] FIG. 6 is a conceptual diagram illustrating a frequency-uniform multi-beam generation method according to an embodiment of the present invention.

[0074] Referring to FIG. 6, another frequency uniform multi-beam generation method of an embodiment includes the step (S610) of preparing digital intermediate frequency data generated through digital signal processing after being received by an array antenna including a plurality of antennas according to frequency and azimuth angle.

[0075] In addition, the frequency uniform multi-beam generation method of another embodiment includes the step (S620) of inputting digital intermediate frequency data into a pre-trained artificial neural network model to generate a broadband frequency uniform multi-beam as the output of the artificial neural network model.

[0076] In this frequency uniform multi-beam generation method, the artificial neural network model trained for frequency uniform multi-beam generation may be trained by the artificial neural network model training method described with reference to FIGS. 3 to 5. When the artificial neural network model is trained by the artificial neural network model training method described through FIG. 5, it is possible to generate a broadband frequency uniform multi-beam while simultaneously securing the required reception bit error rate performance.

[0077] According to the embodiment, to generate weight vectors for broadband frequency uniform multi-beam generation, an artificial neural network based on supervised learning can be trained, and an artificial neural network based on unsupervised learning and reinforcement learning can also be trained.

[0078] During the supervised learning process, the similarity between input intermediate frequency data can cause a significant decrease in reception bit error rate performance at specific angle combinations and frequency bands; it is possible to apply clustering using unsupervised learning by leveraging this similarity. T-SNE unsupervised learning, which proceeds with learning by mapping high-dimensional data to 2 dimensions, can be utilized and can serve as a basis for determining the beamforming azimuth angle intervals of intermediate frequency data used as input to artificial neural networks. To implement this, for example, an autoencoder network can be constructed by stacking multiple fully connected layers, and feature vectors of the input data expressed in latent space can be obtained through the output of the encoder. Subsequently, T-SNE unsupervised learning can be applied to these feature vectors to determine the similarity between input data. When examining the clustering results of the data sets input for training based on T-SNE unsupervised learning, based on the target azimuth angle interval, it was found that the clusters did not overlap with each other, and through this, it was confirmed that it is possible to correctly generate a broadband frequency uniform multibeam using the angle interval used in the experiment.

[0079] Additionally, the Neural Architecture Search (NAS) technique, which searches for and applies the optimal combination of hyperparameters and structures for artificial neural network configuration and training for reinforcement learning, can be applied to the generation of weight vectors for creating broadband frequency uniform multiple beams. These hyperparameters are configured to be found by code rather than by humans, and Optuna, a Python open source tool that automates hyperparameter tuning for machine learning algorithms, can be utilized to find the combination that yields the most optimized performance within a set range. In this case, the environment for reinforcement learning is the entire space where hyperparameters can exist, the agent is the artificial neural network model itself, and the set total loss The goal is to obtain a state representing the optimal combination by performing an action to select a hyperparameter combination in a direction that minimizes it.

[0080] Meanwhile, each step included in the artificial neural network model learning method and / or frequency uniform multibeam generation method according to the above-described embodiment may be implemented in a computer-readable recording medium that records a computer program programmed to perform such steps.

[0081] In addition, each step included in the artificial neural network model learning method and / or frequency uniform multibeam generation method according to the above-described embodiment may be implemented in the form of a computer program stored on a computer-readable recording medium programmed to perform such steps.

[0082] Combinations of each step of each flowchart attached to the present invention may be performed by computer program instructions. Since these computer program instructions may be loaded into the processor of a general-purpose computer, a computer for special purposes, or other programmable data processing equipment, the instructions performed through the processor of the computer or other programmable data processing equipment create means for performing the functions described in each step of the flowchart. Since these computer program instructions may also be stored in a computer-available or computer-readable recording medium that can be oriented toward the computer or other programmable data processing equipment to implement the function in a specific manner, the instructions stored in the computer-available or computer-readable recording medium may also produce a manufactured item containing instruction means for performing the function described in each step of the flowchart. Since computer program instructions can be loaded onto a computer or other programmable data processing equipment, instructions that execute a computer or other programmable data processing equipment by performing a series of operation steps on the computer or other programmable data processing equipment to create a process executed by the computer can also provide steps for executing the functions described in each step of the flowchart.

[0083] Additionally, each step may represent a module, segment, or part of code containing one or more executable instructions for executing a specified logical function(s). Also, it should be noted that in some alternative embodiments, the functions mentioned in the steps may occur out of order. For example, two steps described in succession may actually be performed substantially simultaneously, or the steps may sometimes be performed in reverse order according to the corresponding function.

[0084] The above description is merely an illustrative explanation of the technical concept of the present invention, and those skilled in the art to which the present invention pertains will be able to make various modifications and variations within the scope of the essential quality of the present invention. Accordingly, the embodiments disclosed in the present invention are intended to explain, not limit, the technical concept of the present invention, and the scope of the technical concept of the present invention is not limited by such embodiments. The scope of protection of the present invention shall be interpreted by the claims below, and all technical concepts within the equivalent scope shall be interpreted as being included within the scope of rights of the present invention. Explanation of the symbols

[0085] 100: Electronic device 110: Memory 120: Processor

Claims

Claim 1 A method for training an artificial neural network model performed by an electronic device comprises: a step of generating training data by simulating digital intermediate frequency data that can be generated through digital signal processing after being received by an array antenna including multiple antennas according to frequency and azimuth angle; and a step of training an artificial neural network model for broadband frequency uniform multi-beam generation using the generated training data, wherein the weight vector of the artificial neural network model is regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for the generation of a desired multi-beam, and the step of generating the training data comprises: a step of preparing data to be input to the artificial neural network model by performing operations on the digital intermediate frequency data as a complex correlation matrix; a step of obtaining a feature vector of the digital intermediate frequency data expressed in a latent space through an encoder of an autoencoder network composed of stacking multiple fully connected layer layers; and a step of determining the similarity between data by applying T-SNE unsupervised learning to the feature vector. An artificial neural network model training method comprising: a step of checking whether clusters do not overlap among angle combinations resulting from clustering based on a target azimuth angle interval, determining a beamforming azimuth angle interval for broadband frequency uniform multi-beam generation, and proceeding with training. Claim 2 In claim 1, the weight vector is a weight of a relative loss function for a target weight vector relative to a target beam pattern having a value between 0 and 1 ( By applying ) the above target beam pattern-based loss function ( ) and the above target weight vector-based loss function ( ) is a mathematical expression A method for training a combined artificial neural network model using Claim 3 delete Claim 4 delete Claim 5 An artificial neural network model training method according to claim 1, wherein, in the training step, hyperparameters for obtaining the weight vector are tuned based on the search results through neural network structure search. Claim 6 The apparatus includes a memory containing an artificial neural network model for generating broadband frequency uniform multi-beams, and a processor that performs the functions of the artificial neural network model. The processor generates training data by simulating digital intermediate frequency data that can be generated through digital signal processing after being received by an array antenna including multiple antennas according to frequency and azimuth angle. The artificial neural network model is trained using the generated training data. The weight vectors of the artificial neural network model are regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for generating the desired multi-beams. In the process of generating the training data, the processor prepares data to be input to the artificial neural network model by performing operations on the digital intermediate frequency data as a complex correlation matrix. It obtains feature vectors of the digital intermediate frequency data expressed in the latent space through an encoder of an autoencoder network composed of stacked fully connected layer layers. It determines the similarity between data by applying T-SNE unsupervised learning to the feature vectors and determines whether clusters do not overlap between angle combinations resulting from clustering based on the desired azimuth angle interval. An electronic device that performs learning by determining the beamforming azimuth angle interval for broadband frequency uniform multi-beam generation by verifying. Claim 7 In claim 6, the weight vector is a weight of the relative loss function for a target weight vector relative to a target beam pattern having a value between 0 and 1 ( By applying ) the above target beam pattern-based loss function ( ) and the above target weight vector-based loss function ( ) is a mathematical expression Electronic device combined using Claim 8 delete Claim 9 delete Claim 10 An electronic device according to claim 6 that, when training, tunes hyperparameters for obtaining the weight vector based on the search results through neural network structure search. Claim 11 A method for generating a broadband frequency uniform multiple beam performed by an electronic device comprises: a step of preparing digital intermediate frequency data generated through digital signal processing after receiving data by frequency and azimuth angle at an array antenna including multiple antennas; and a step of inputting the digital intermediate frequency data into a pre-trained artificial neural network model to generate a broadband frequency uniform multiple beam as the output of the artificial neural network model, wherein the weight vector of the artificial neural network model is regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for generating a desired multiple beam, wherein the artificial neural network model uses data obtained by operating the digital intermediate frequency data as a complex correlation matrix as input data during training, determines the similarity between data by applying T-SNE unsupervised learning to feature vectors in latent space obtained through an encoder of an autoencoder network composed of stacked fully connected layer layers, and verifies whether clusters do not overlap between angle combinations resulting from clustering based on a desired azimuth angle interval, and is a frequency uniform multiple beam model trained on training data generated with a beamforming azimuth angle interval determined by checking whether the resulting angle combinations do not overlap. Creation method. Claim 12 In claim 11, the weight vector is a weight of the relative loss function for a target weight vector relative to a target beam pattern having a value between 0 and 1 ( By applying ) the above target beam pattern-based loss function ( ) and the above target weight vector-based loss function ( ) is a mathematical expression A method for generating a combined frequency uniform multi-beam using Claim 13 delete Claim 14 delete Claim 15 In claim 11, the artificial neural network model is a frequency uniform multi-beam generation method in which hyperparameters for obtaining the weight vector are tuned based on the search results through neural network structure search. Claim 16 The system includes a memory containing a pre-trained artificial neural network model for generating broadband frequency uniform multiple beams, and a processor that performs the functions of said artificial neural network model. The processor inputs digital intermediate frequency data, generated through digital signal processing after being received by an array antenna including multiple antennas according to frequency and azimuth angle, into said artificial neural network model to generate broadband frequency uniform multiple beams as the output of said artificial neural network model. The weight vector of said artificial neural network model is regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for generating the desired multiple beams. During training, said artificial neural network model uses data obtained by operating the digital intermediate frequency data as a complex correlation matrix as input data. It determines the similarity between data by applying T-SNE unsupervised learning to feature vectors in latent space obtained through an encoder of an autoencoder network composed of stacked fully connected layer layers, and checks whether clusters do not overlap between angle combinations resulting from clustering based on the desired azimuth angle interval, and is trained by training data generated with the determined beamforming azimuth angle interval. Model electronic device. Claim 17 In claim 16, the weight vector is a weight of the relative loss function for a target weight vector relative to a target beam pattern having a value between 0 and 1 ( By applying ) the above target beam pattern-based loss function ( ) and the above target weight vector-based loss function ( ) is a mathematical expression Electronic device combined using Claim 18 delete Claim 19 delete Claim 20 In claim 16, the artificial neural network model is an electronic device in which hyperparameters for obtaining the weight vector are tuned based on the search results through neural network structure search. Claim 21 A computer-readable recording medium storing a computer program, wherein the computer program, when executed by a processor, is an artificial neural network model learning method performed by an electronic device, comprising: a step of generating learning data by simulating digital intermediate frequency data that can be generated through digital signal processing after being received by an array antenna including a plurality of antennas according to frequency and azimuth angle; and a step of training an artificial neural network model for broadband frequency uniform multi-beam generation using the generated learning data, wherein the weight vector of the artificial neural network model is regressively learned through a target beam pattern-based loss function and a target weight vector-based loss function required for the generation of a desired multi-beam, and wherein the step of generating the learning data comprises: a step of preparing data to be input to the artificial neural network model by performing operations on the digital intermediate frequency data as a complex correlation matrix; a step of obtaining a feature vector of the digital intermediate frequency data expressed in a latent space through an encoder of an autoencoder network composed of stacking multiple fully connected layer layers; and a step of determining similarity between data by applying T-SNE unsupervised learning to the feature vector. A computer-readable recording medium comprising instructions for a processor to perform a method including: a step of determining whether clusters do not overlap among angle combinations resulting from clustering based on a desired azimuth angle interval, and determining the correct beamforming azimuth angle interval for broadband frequency uniform multi-beam generation and proceeding with learning. Claim 22 A computer program stored on a computer-readable recording medium, wherein the computer program, when executed by a processor, is an artificial neural network model learning method performed by an electronic device, comprising the steps of: generating learning data by simulating digital intermediate frequency data that can be generated through digital signal processing after being received by an array antenna including a plurality of antennas according to frequency and azimuth angle; and training an artificial neural network model for broadband frequency uniform multi-beam generation using the generated learning data, wherein the weight vector of the artificial neural network model is regressively learned through a target beam pattern-based loss function and a target weight vector-based loss function required for the generation of a desired multi-beam, and wherein the step of generating the learning data comprises: preparing data to be input to the artificial neural network model by performing operations on the digital intermediate frequency data as a complex correlation matrix; obtaining a feature vector of the digital intermediate frequency data expressed in a latent space through an encoder of an autoencoder network composed of stacking multiple fully connected layer layers; and determining the similarity between data by applying T-SNE unsupervised learning to the feature vector. A computer program comprising: a step of determining whether clusters do not overlap among angle combinations resulting from clustering based on a target azimuth angle interval, and determining the beamforming azimuth angle interval to generate a broadband frequency uniform multi-beam, and proceeding with learning. Claim 23 A computer-readable recording medium storing a computer program, wherein the computer program, when executed by a processor, is a broadband frequency uniform multibeam generation method performed by an electronic device, comprising: a step of preparing digital intermediate frequency data generated through digital signal processing after receiving by frequency and azimuth angle at an array antenna including a plurality of antennas; and a step of inputting the digital intermediate frequency data into a pre-trained artificial neural network model to generate a broadband frequency uniform multibeam as the output of the artificial neural network model, wherein the weight vector of the artificial neural network model is regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for the generation of a desired multibeam, wherein during training, data obtained by computation of the digital intermediate frequency data into a complex correlation matrix is ​​used as input data, and T-SNE unsupervised learning is applied to feature vectors in latent space obtained through an encoder of an autoencoder network composed of stacking multiple fully connected layer layers to determine similarity between data, and confirms whether clusters do not overlap between angle combinations resulting from clustering based on a desired azimuth angle interval. A computer-readable recording medium comprising instructions for the processor to perform a method of a model learned by training data generated at determined beamforming azimuth angle intervals. Claim 24 A computer program stored on a computer-readable recording medium, wherein the computer program, when executed by a processor, is a broadband frequency uniform multibeam generation method performed by an electronic device, comprising: a step of preparing digital intermediate frequency data generated through digital signal processing after receiving by frequency and azimuth angle at an array antenna including a plurality of antennas; and a step of inputting the digital intermediate frequency data into a pre-trained artificial neural network model to generate a broadband frequency uniform multibeam as the output of the artificial neural network model, wherein the weight vector of the artificial neural network model is regressively trained through a target beam pattern-based loss function and a target weight vector-based loss function required for the generation of a desired multibeam, wherein during training, data obtained by operating the digital intermediate frequency data as a complex correlation matrix is ​​used as input data, and T-SNE unsupervised learning is applied to feature vectors in latent space obtained through an encoder of an autoencoder network composed of stacked fully connected layer layers to determine similarity between data, and by checking whether clusters do not overlap between angle combinations resulting from clustering based on a desired azimuth angle interval A computer program that is a model trained by training data generated at determined beamforming azimuth angle intervals.

Images