Learning method

Abstract

To acquire an appropriate setting value of an antenna element in beamforming.SOLUTION: A learning method executes for each of a plurality of beams: first computation processing for outputting setting information to be set to an antenna element, which forms a beam, from first information identifying a radiation form of the beam; second computation processing for outputting radio wave radiation form information relating to a radio wave radiation form of the antenna element from the setting information, a form of the antenna element and coordinates of an installation position; and third computation processing for outputting reception power information relating to reception power of each of a plurality of terminal devices from the radio wave radiation form information and transmission line characteristics. The learning method also executes: fourth computation processing for outputting reception state information relating to a reception state of the beam for each terminal device from the reception power information, which is outputted for each of the plurality of beams, of each of the plurality of terminal devices; and learning processing for learning a first model, which executes the first computation processing, using the first information and the reception state information.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a learning method. [Background technology]

[0002] In recent years, Massive MIMO (mMIMO) has been attracting attention in wireless communications. In mMIMO, a base station device is configured with, for example, multiple antenna elements. The base station device performs beamforming (BF) by adjusting the phase of radio waves emitted from the multiple antenna elements and amplifying the radio wave strength in a predetermined direction. Beamforming improves the signal-to-noise ratio (SNR) in a predetermined direction, and can increase throughput compared to a single-antenna system.

[0003] However, the radiation pattern of beamformed radio waves radiates strong radio waves in a specific direction, while radiating unwanted radio waves in other directions. This can cause interference between beams transmitted to other terminal devices (inter-beam interference) or the radiation of unwanted radio waves to adjacent cells (inter-cell interference). In wireless communication systems that support beamforming, the phase and gain of radio waves radiated from multiple antenna elements can be adjusted to change the radiation pattern of the radio waves and perform control to suppress such interference.

[0004] Beamforming techniques are described in, for example, the following documents: [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2023-085217 [Patent Document 2] Special Publication No. 2022-527795 Summary of the Invention [Problem to be solved by the invention]

[0006] However, in the above-mentioned beamforming control, inter-cell interference and inter-beam interference occur, making it difficult to appropriately configure antenna elements to form beams in the desired direction and strength.

[0007] Therefore, one disclosure can provide a neural network learning method that can obtain appropriate setting values ​​for antenna elements in beamforming. [Means for solving the problem]

[0008] The system executes a first calculation process for each of a plurality of beams, which outputs setting information to be set in antenna elements that form the beam from first information that identifies the radiation shape of the beam; a second calculation process for outputting radio wave radiation shape information regarding the radio wave radiation shape of the antenna element from the setting information and the coordinates of the shape and installation position of the antenna element; and a third calculation process for outputting received power information regarding the received power of each of a plurality of terminal devices from the radio wave radiation shape information and transmission path characteristics.The system also executes a fourth calculation process for outputting reception status information regarding the reception status of the beam for each terminal device from the received power information for each of the plurality of terminal devices that has been output for each of the plurality of beams; and a learning process for learning a first model that executes the first calculation process using the first information and the reception status information. [Effects of the Invention]

[0009] One disclosure is capable of obtaining appropriate setting values ​​for antenna elements in beamforming. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram illustrating an example of the configuration of a wireless communication system 1. As shown in FIG. [Figure 2] FIG. 2 is a diagram illustrating an example of the configuration of the simulation device 300. As shown in FIG. [Figure 3]FIG. 3 is a diagram showing an example of a software configuration for executing the BF weight NN learning process S100. [Figure 4] FIG. 4 is a diagram showing an example of a processing flowchart of the BF weight NN learning process S100. [Figure 5] FIG. 5 is a diagram showing an example of a processing flowchart of the radial propagation NN and propagation loss NN learning process S200. [Figure 6] FIG. 6 is a diagram showing an example of the configuration of a radial propagation NN. [Figure 7] FIG. 7 is a diagram showing an example of a processing flowchart of the terminal position NN learning processing S300. [Figure 8] FIG. 8 is a diagram showing an example of the configuration of the terminal position NN. DETAILED DESCRIPTION OF THE INVENTION

[0011] [First embodiment] A first embodiment will be described.

[0012] <About Wireless Communication System 1> 1 is a diagram showing an example of the configuration of a wireless communication system 1. The wireless communication system 1 includes base station devices 200-1 to 200-3 and terminal devices 100-1 to 4. The wireless communication system 1 is, for example, a communication system that supports mMIMO.

[0013] Base station devices 200-1 to 3 (hereinafter may be referred to as base station devices 200) are devices that perform wireless communication with terminal devices 100-1 to 4 (hereinafter may be referred to as terminal devices 100), and are, for example, eNodeBs or gNodeBs. Base station device 200 has, for example, one or more antenna elements, and performs beamforming by controlling setting values ​​for the antenna elements.

[0014] Furthermore, the base station device 200 configures a communication area (cell) in which the terminal device 100 performs wireless communication with the base station device 200. The base station devices 200-1 to 200-3 configure cells C200-1 to 200-3, respectively. In FIG. 1, the cells C200-1 to C200-3 do not overlap with each other, but they may overlap.

[0015] For example, terminal devices 100-1 and 100-2 are located within cell C200-1. Base station device 200-1 transmits beam B1 to terminal device 100-1 and transmits beam B2 to terminal device 100-2. In this way, one base station device 200 may transmit multiple beams. The multiple beams may be transmitted using, for example, antenna element groups each composed of different antenna elements, or may be transmitted using antenna element groups each composed of the same or some of the same antenna elements.

[0016] <Configuration example of simulation device 300> 2 is a diagram showing an example of the configuration of a simulation device 300. The simulation device 300 is, for example, a computer machine or a server machine, and is a simulator that trains a neural network (NN) that outputs setting values ​​(BF weights) of antenna elements. The simulation device 300 includes a CPU 310, a storage 320, and a memory 330.

[0017] The storage 320 is an auxiliary storage device such as a flash memory, HDD, or SSD that stores programs and data. The storage 320 stores a BF weight NN program 321 and a BF weight NN learning program 322.

[0018] The memory 330 is an area into which the programs stored in the storage 320 are loaded. The memory 330 may also be used as an area in which the programs store data.

[0019] The CPU 310 is a processor that loads the program stored in the storage 320 into the memory 330, executes the loaded program, constructs each part, and realizes each process.

[0020] By executing the BF weight NN program 321, the CPU 310 constructs the BF weight NN and executes the BF weight calculation process. The BF weight is a setting for the antenna element, for example, phase and gain. The BF weight NN is an AI that calculates the BF weight, for example, a NN.

[0021] By executing the BF weight NN learning program 322, the CPU 310 constructs the learning part and executes the BF weight NN learning process. The BF weight NN learning process is a process of learning the BF weight NN. The BF weight NN learning process includes a BF weight calculation process, a radiation propagation calculation process, a BF calculation process, a propagation loss calculation process, a SINR calculation process, and an optimization process.

[0022] <BF weight NN learning process> The BF weight NN learning process S100 is a process of learning the BF weight NN. The BF weight NN is an AI (artificial intelligence) that calculates the BF weight when a beam ID is input, for example, a neural network.

[0023] The beam ID is an identifier of a pattern (radiation pattern) regarding the radiation shape of the beam, for example, a number or a digit. There are multiple beam IDs for each beam (for each base station device). For example, a beam ID is assigned to each radiation pattern of the beam B2 in FIG. 1 and is represented as the beam B2ID. Hereinafter, the beam ID of beam n may be represented as beam nID.

[0024] The beam ID corresponds to, for example, the area to be radiated (the position of the terminal device 100). For example, beam 1 ID=1 corresponds to a certain area x, and beam 1 ID=2 corresponds to a certain area y (an area different from area x). In actual operation, the base station device 200 radiates, for example, a beam corresponding to the area where the terminal device 100 is located. Note that the beam ID may be, for example, information (first information) relating to the radiation shape of the beam.

[0025] The BF weight is a setting value (e.g., position and gain; hereinafter, the BF weight indicates the position and gain) for each antenna element that forms a beam. The BF weight NN calculates (outputs) the BF weight corresponding to the beam ID.

[0026] <Software configuration example> 3 is a diagram showing an example of a software configuration for executing the BF weight NN learning process S100. The software of the simulation device 300 is made up of BF weight NNs 10-1 to 10-n (n is an integer equal to or greater than 1), radiation propagation models 11-1 to 11-n, BF calculation units 12-1 to 12-n, propagation loss models 13-1 to 13-n, an SINR calculation unit 14, and an NN optimization unit 15. n indicates, for example, a beam number.

[0027] BF weight NN10-1 to n (hereinafter may be referred to as BF weight NN10), radiation propagation models 11-1 to n (hereinafter may be referred to as radiation propagation model 11), BF calculation units 12-1 to n (hereinafter may be referred to as BF calculation unit 12), and propagation loss models 13-1 to n (hereinafter may be referred to as propagation loss model 13) are provided for each beam number. In the following explanation, the units corresponding to beam 1 will be explained, but it is assumed that the units corresponding to beams 2 to n have the same configuration and the same inputs and outputs.

[0028] The BF weight NN10 is an NN that calculates the BF weights of the antenna elements that constitute a beam (for example, some or all of the antenna elements of the base station device 200-1). When a beam ID is input (S1), the BF weight NN10 outputs a BF weight corresponding to the beam ID (S2).

[0029] The radiation propagation model 11 is a model that calculates the radiation shape of a beam from each antenna element, and executes, for example, a calculation formula prepared in advance. When the BF weight corresponding to the beam ID is input to the radiation propagation model 11 (S2), the model calculates the radiation shape from each antenna element from the physical shape and coordinate information of the antenna, taking into account the influence of the transmission path, and outputs the phase and gain for each radiation direction (angle) from each antenna element for each antenna element (S3). The output S3 is, for example, a list of angles and corresponding phases and gains for each antenna element.

[0030] The BF calculation unit 12 is a model that calculates the radiation shape of a beam obtained by combining radiation from each antenna element in space, and executes, for example, a calculation formula prepared in advance. When the phase and gain of radiation for each angle from each antenna element are input to the BF calculation unit 12 (S3), the BF calculation unit 12 calculates the radiation shape of the beam by calculating the interference of the radiation radio waves from each antenna element, and outputs the power (gain) for each direction of the beam (S4). The output S4 is, for example, a list of angles (directions) and corresponding gains.

[0031] The propagation loss model 13 is a model that calculates propagation loss in space from the radiation shape and terminal position, and calculates the received power at the terminal device 100, and for example executes a calculation formula prepared in advance. When the position information of the terminal device 100 and the power for each direction of the beam are input (S4), the propagation loss model 13 calculates the propagation loss and outputs the received power for each terminal device 100 (S5). Note that the received power is output as, for example, the power of each of the desired wave and the interfering wave.

[0032] The number of terminal devices 100 is, for example, a number corresponding to the number of beams; if there are n beams, there are also n terminal devices 100. Note that one beam may be frequency-divided to simultaneously communicate with multiple terminal devices 100, so the number of terminal devices 100 may be more than n. Also, the terminal devices 100 are assumed to be located in a region corresponding to the beam ID of each beam. For example, when beam 1 ID = x, the terminal device 100 that is the target of beam 1 radiation is assumed to be located at a first position in a first region corresponding to beam 1 ID = x, and when beam 2 ID = y, the terminal device 100 that is the target of beam 2 radiation is assumed to be located at a second position in a second region corresponding to beam 2 ID = y.

[0033] The terminal position information may be a value actually measured by the terminal device 100 in an actual field where the terminal device 100 is used. The terminal position information may also be created by setting a random or specific distribution in a simulated space.

[0034] The SINR calculation unit 14 calculates the SINR (Signal to Interference plus Noise Ratio) for each terminal device 100. When the received power of the terminal device 100 corresponding to the beam ID of each beam (for example, desired signal and interference signal) is input to the SINR calculation unit (S5 and the received power of the terminal device 100 corresponding to the beam IDs of beams 2 to n), the SINR calculation unit calculates and outputs the SINR of each terminal device 100 (S6). For example, if the position of the terminal device 100-1 corresponds to beam 1 ID=1 of beam 1, the received power of beam 1 of the terminal device 100-1 is the desired signal, and the received power from other beams is the interference signal. The noise component is calculated from, for example, thermal noise determined from temperature. The SINR calculation unit calculates the desired signal and interference signal for each terminal device 100, and calculates the SINR taking the noise component into consideration.

[0035] The NN optimization unit 15 updates (calculates) the weights of each of the BF weight NNs 10-1 to n from the SINR value, and trains (updates) the BF weight NNs 10-1 to n using, for example, Deep Learning. The NN optimization unit 15 updates the weights of the NNs so that, for example, the SINR increases. An increase in the SINR means, for example, that the average value of the SINR of each terminal device 100 increases. Also, for example, additional conditions may be set such that the SINR of all terminal devices (or a predetermined number or more of terminal devices) does not fall below a reference value.

[0036] <BF Weight NN Learning Process> FIG. 4 is a diagram showing an example of a processing flowchart of the BF weight NN learning process S100. In the following description, the processing corresponding to beam 1 will be described, but the same applies to the processing corresponding to beams 2 to n.

[0037] The simulation device 300 inputs a beam ID and executes the BF weight calculation process S10. The BF weight calculation process is a process executed by the BF weight NN 10 and is a process of calculating the BF weights of the antenna elements constituting the beam corresponding to the beam ID.

[0038] Next, the simulation device 300 inputs the BF weights and executes the radiation propagation calculation process S11. The radiation propagation calculation process S11 is a process executed by the radiation propagation model 11 and is a process of calculating the phase and gain for each angle of radiation from each antenna element.

[0039] Next, the simulation device 300 inputs the phase and gain for each angle of radiation from each antenna element and executes the BF calculation process S12. The BF calculation process S12 is a process executed by the BF calculation unit 12 and is a process of calculating the interference of the radiated radio waves from each antenna and calculating the power for each direction.

[0040] Next, the simulation device 300 inputs the power for each beam direction and terminal position information, and executes a propagation loss calculation process S13. The propagation loss calculation process S13 is a process executed by the propagation loss model 13, and is a process of calculating the received power for each terminal device 100 from the propagation loss.

[0041] Next, the simulation device 300 inputs the received power of the terminal device 100 and executes the SINR calculation process S14. The SINR calculation process S14 is a process executed by the SINR calculation unit 14, and is a process for calculating the SINR for each terminal device 100.

[0042] Next, the simulation device 300 inputs the SINR for each terminal device 100 and executes optimization processing S15. The optimization processing S15 is processing executed by the NN optimization unit 15, and is processing for updating the weight of each NN of the BF weights NN10-1 to NN10-n according to the value of the SINR (S16).

[0043] As described above, the simulation device 300 sets the position of the terminal device 100 within the area corresponding to the beam ID for each beam, performs each calculation process for each beam, calculates the SINR taking into account the received power of all beams and the gain of the beam in all directions (including interference waves), and trains the BF weight NN10 so that the SINR becomes large.

[0044] The simulation device 300 changes the beam ID of each beam, changes the terminal position information within the corresponding area, and repeats the above learning process to improve the learning level of the BF weight NN. The learning process is repeated until it is detected that the amount of change in the weight of the BF weight NN has become smaller, or the amount of change in the BF weight output by the BF weight NN has become smaller. The number of repetitions may also be determined in advance.

[0045] <Reflection in actual operation> For example, the base station device 200 in Fig. 1 stores a trained BF weight NN. In actual operation, the base station device 200 selects a beam ID corresponding to the position of the terminal device 100, inputs the beam ID to the trained BF weight NN, obtains BF weights, and controls the antenna elements. Alternatively, the base station device 200 may store, as data, a list of beam IDs and corresponding BF weights output using the trained BF weight NN instead of the trained BF weight NN.

[0046] <Modification> Although the BF weight NN10 is updated in accordance with the SINR, for example, the transmission power of the base station device 200 may also be taken into consideration. The simulation device 300 calculates the transmission power of the base station device 200 from the gain of the BF weight and outputs the calculated power to the NN optimization unit 15. The NN optimization unit 15 updates the weight of the NN taking into consideration both the transmission power and the SINR. For example, the NN optimization unit 15 updates the weight of the NN so that the average value of the transmission power is as small as possible and the average value of the SINR is as large as possible.

[0047] The beam IDs of other beams may also be input as information to the BF weight NN 10. In this case, the learned BF weight NN 10 receives the beam IDs of the other beams as input information in addition to the beam ID of the own beam, and outputs the BF weight of the own beam.

[0048] Also, instead of SINR, other information (numerical values) regarding the communication state may be used. For example, throughput may be used instead of or in addition to SINR during optimization. SINR is an example of a numerical value indicating the communication state, and any index other than SINR may be used as long as it indicates the communication state.

[0049] [Second embodiment] A second embodiment will be described. In the second embodiment, the simulation device 300 causes an AI to execute a radiative propagation calculation process and a propagation loss calculation process. The simulation device 300 causes a radiative propagation NN that performs the radiative propagation calculation process and a propagation loss NN that performs the propagation loss calculation process to learn each of them. The radiative propagation NN and the propagation loss NN are, for example, neural networks. Note that the radiative propagation NN and the propagation loss NN are constructed, for example, by the CPU 310 of the simulation device 300 executing a stored program.

[0050] <Radiation propagation NN and propagation loss NN learning process> 5 is a diagram showing an example of a processing flowchart of the radiation propagation NN and propagation loss NN learning process S200. In the following explanation, the processing corresponding to beam 1 will be described, but the processing corresponding to beams 2 to n will be similar.

[0051] The simulation device 300 receives a beam ID and executes a BF weight calculation process S20. In the first embodiment, the BF weight calculation process S10 is executed by the BF weight NN. The BF weight calculation process S20 may be executed by the trained BF weight NN in the first embodiment, or may be executed by another NN or model (a processing unit that calculates a predetermined formula).

[0052] Next, the simulation device 300 inputs the BF weights and executes the radiation propagation calculation process S21. The radiation propagation calculation process S21 is executed by the radiation propagation NN.

[0053] Next, the simulation apparatus 300 inputs the phase and gain for each angle of radiation from each antenna element and executes BF calculation processing S22. The BF calculation processing S22 is similar to, for example, the BF calculation processing S12 in the first embodiment.

[0054] Next, the simulation device 300 inputs the power for each direction of the beam and terminal position information, and executes the propagation loss calculation process S33. The propagation loss calculation process S13 is a process executed by the propagation loss NN.

[0055] Next, the simulation device 300 inputs the received power of the terminal device 100 and executes an SINR calculation process S24. The SINR calculation process S24 is similar to the SINR calculation process S14 in the first embodiment, for example.

[0056] Next, the simulation device 300 inputs the SINR for each terminal device 100 and executes optimization processing S25. The optimization processing S25 uses, for example, a communication report from the terminal device 100. The communication report is, for example, an actual measurement value of the terminal device 100, and includes received power and SINR. In the optimization processing S25, the simulation device 300 updates the weights of the NNs in the radiation propagation NN and the propagation loss NN according to the difference between the received power and SINR actually measured by the terminal device 100 and the received power and SINR calculated in each calculation processing so that the difference becomes small (to 0), and causes the radiation propagation NN and the propagation loss NN to learn (S26, S27).

[0057] Then, the simulation device 300 changes the beam ID of each beam, changes the terminal position information within the corresponding area, and repeats the above learning process to improve the degree of learning of the radiation propagation NN and the propagation loss NN.

[0058] <Example of radiation propagation NN configuration> The radial propagation NN may be configured, for example, by one neural network, or may be configured by combining one or more neural networks and circuits.

[0059] 6 is a diagram showing an example of the configuration of a radiation propagation NN. The radiation propagation NN has element radiation NN models 20-1 to m (m is an integer equal to or greater than 1; hereinafter, they may be referred to as element radiation models NN 20), a random circuit 25, an averaging unit 21, an amplitude mask limiting unit and phase range limiting unit 22, a filter unit 23, and a gain / phase setting error correcting NN model 24.

[0060] The element radiation NN model 20 includes NNs for all antenna elements and calculates, for example, a radiation pattern. The element radiation NN model 20 estimates and outputs the radiation pattern of each antenna element (for example, the gain and phase of the radiated radio wave for each direction) corresponding to the BF weight of each antenna element, for example, from the actual measurement values ​​of the SINR and / or received power of the terminal device 100. The element radiation NN model 20 is used by adding a random phase to the phase via a random circuit 25. Adding a random phase allows the element radiation NN model 20 to learn without correlation.

[0061] The averaging unit 21, for example, averages the output from the element radiation NN model 20 of each antenna element. By performing averaging, it is possible to suppress bias in the estimation result. Note that if there is only one element radiation NN model 20, the averaging unit 21 does not need to be installed.

[0062] The amplitude mask limiting section and phase range limiting section 22, for example, limits prominent amplitudes to a predetermined value or less and limits prominent phases to within a predetermined range.

[0063] The gain / phase setting error correction NN model 24 is a model that inputs, for example, gain and phase setting values ​​for an antenna element and outputs the error between the setting value of the antenna element and the actual output value. The gain / phase setting error correction NN model 24 is a model that estimates and learns the error between the setting value of the antenna element and the actual output value so as to reduce the error between the received power / SINR calculated from the setting value of the antenna element and the received power / SINR actually measured by the terminal device. By outputting the error, it is possible to absorb individual differences between the antenna elements and / or the base station device 200.

[0064] The filter section 23 is provided for the purpose of smoothing out any abrupt changes in the output of the amplitude mask limiting section and the phase range limiting section 22, since this would hinder optimization.

[0065] As described above, it is possible to construct a radiation propagation NN by installing an NN that learns the radiation shape of each antenna element and learns the error in the setting values.

[0066] Note that the filter unit 23, the amplitude mask limiting unit and phase range limiting unit 22, and the random circuit 25 may not be installed depending on individual differences between antenna elements and the manufacturing accuracy of the antenna (base station device 200).

[0067] [Third embodiment] A third embodiment will be described. In the third embodiment, the simulation device 300 causes an AI to estimate terminal location information. The simulation device 300 causes the AI ​​to learn a terminal location NN. The terminal location NN is an NN that estimates terminal location information. The terminal location NN is constructed, for example, by the CPU 310 of the simulation device 300 executing a stored program.

[0068] <Terminal position NN learning process> 7 is a diagram showing an example of a processing flowchart of the terminal position NN learning process S300. In the following explanation, the processing corresponding to beam 1 will be described, but the processing corresponding to beams 2 to n will be assumed to be similar.

[0069] The simulation device 300 receives the beam ID and executes the BF weight calculation process S30. The BF weight calculation process S30 is similar to the BF weight calculation process S30 in the second embodiment, for example.

[0070] Next, the simulation device 300 inputs the BF weights and executes the radiative propagation calculation process S31. The radiative propagation calculation process S31 is similar to, for example, the radiative propagation calculation process S11 in the first embodiment.

[0071] Next, the simulation apparatus 300 inputs the phase and gain for each angle of radiation from each antenna element and executes BF calculation processing S32. The BF calculation processing S32 is similar to, for example, the BF calculation processing S12 in the first embodiment.

[0072] Next, the simulation device 300 inputs the power for each beam direction and terminal position information, and executes a propagation loss calculation process S33. The propagation loss calculation process S33 is similar to the propagation loss calculation process S12 in the first embodiment, for example.

[0073] Terminal position information is input to the propagation loss calculation process S33. The terminal position information is estimated information, and is calculated in the terminal position estimation process S37 executed by the terminal position NN.

[0074] Next, the simulation device 300 inputs the received power of the terminal device 100 and executes an SINR calculation process S34. The SINR calculation process S34 is similar to the SINR calculation process S14 in the first embodiment, for example.

[0075] Next, the simulation device 300 inputs the SINR for each terminal device 100 and executes optimization processing S35. The optimization processing S35 uses, for example, a communication report from the terminal device 100. The communication report is, for example, an actual measurement value of the terminal device 100, and includes received power and SINR. In the optimization processing S35, the simulation device 300 updates the weight of the terminal position NN according to the difference between the SINR actually measured by the terminal device 100 and the SINR calculated in each calculation processing so that the difference becomes small (to 0), and causes the terminal position NN to learn (S36).

[0076] Then, the simulation device 300 changes the beam ID of each beam and repeats the above learning process to improve the degree of learning of the terminal position NN.

[0077] <Example of terminal location NN configuration> The terminal position NN may be configured, for example, by one neural network, or by a combination of one or more neural networks and circuits.

[0078] 8 is a diagram showing an example of the configuration of a terminal position NN. The terminal position NN has selectors 30 and 31, beam ID NNs 32-1 to 32-z (z is the number of beam IDs, hereinafter, may be referred to as beam ID NNs 32), and a range limiting unit 33.

[0079] The selectors 30 and 31 select a beam ID NN 32 for each beam ID. For example, when the beam ID is 1, the beam ID NN for beam ID = 1 is selected. Note that the selectors 30 and 31 do not need to be installed when there is only one beam ID NN 32.

[0080] The beam ID NN32 outputs a position corresponding to the beam ID (estimated position of the terminal device). The beam ID NN32 is provided for each beam ID, and receives the received power, SINR, and beam IDs of other beams as input, outputting an estimated position of the terminal device. Note that the received power, SINR, and beam IDs of other beams are examples of input data, and may be part of the data or may include other data.

[0081] The range limiting unit 33 limits the output position information to a predetermined range, excludes any prominent abnormal ranges, and performs processing to improve accuracy.

[0082] [Other embodiments] The first to third embodiments may be combined in the learning stage and in actual operation. The trained NN may be stored in the base station device 200 in actual operation, or the results output by the trained NN may be compiled into a list and stored in the base station device. Furthermore, the learning of the NN may be performed during actual operation.

[0083] For example, the BF weight NN learning process and the radial propagation NN and propagation loss NN learning processes may be executed simultaneously.

[0084] Furthermore, for example, the radial propagation NN and propagation loss NN learning process and the terminal position NN learning process may be executed simultaneously, or may be alternately and repeatedly learned.

[0085] Also, for example, in the BF weight NN learning process, trained radial propagation NNs and propagation loss NNs may be used.

[0086] Also, for example, for a terminal device 100 whose terminal position is unknown in actual operation, the position of the terminal device 100 may be estimated using a learned terminal position NN.

[0087] Furthermore, for example, in actual operation, learning processes for the BF weight NN, the radial propagation NN, the propagation loss NN, and the terminal position NN may be executed. In this case, learning of one type of NN or simultaneous learning of multiple types of NN may be executed.

[0088] To summarise the above, the following is added:

[0089] (Appendix 1) a first calculation process for outputting setting information to be set on antenna elements that form the beam from first information that identifies the radiation shape of the beam; a second calculation process for outputting radio wave radiation shape information relating to the radio wave radiation shape of the antenna element from the setting information and the coordinates of the shape and installation position of the antenna element; a third calculation process for outputting received power information relating to the received power of each of a plurality of terminal devices from the radio wave radiation shape information and the transmission path characteristics, for each of the plurality of beams; a fourth calculation process of outputting reception state information regarding the reception state of the beam for each terminal device from the reception power information of each of the plurality of terminal devices output for each of the plurality of beams; a learning process for learning a first model that executes the first calculation process using the first information and the reception state information; How to learn.

[0090] (Appendix 2) the first model comprises a neural network; In the learning process, the connection weights of the neural network are optimized. Study method described in Appendix 1.

[0091] (Appendix 3) The learning process updates the weights of the connections so that the reception state of the beams for each of the plurality of terminal devices is improved. Study method described in Appendix 2.

[0092] (Appendix 4) The setting information includes a phase and a gain to be set to the antenna element. Study method described in Appendix 1.

[0093] (Appendix 5) the radiation shape of each of the plurality of beams corresponds to a respective region, and the plurality of terminal devices are located within the respective regions; the third calculation process calculates a received power of each of the plurality of terminal devices using terminal position information relating to positions of the plurality of terminal devices corresponding to the radiation shape of each of the plurality of beams; The fourth calculation process acquires the received power of a desired wave and an interference wave in each of the plurality of terminal devices from the received power of each of the plurality of beams of the plurality of terminal devices, and calculates an SINR (Signal to Interference plus Noise Ratio) of each of the plurality of terminal devices as the reception state; The learning process updates the coupling weights so that the SINR of each of the plurality of terminal devices increases. Study method described in Appendix 2.

[0094] (Appendix 6) The reception state further includes a transmission power of the antenna element; The learning process updates the coupling weights according to the SINRs and the transmission powers of the plurality of terminal devices. Study method described in Appendix 5.

[0095] (Appendix 7) the second computation is performed by a second model; The second model is a trained model trained using the received power and SINR actually measured by the terminal device. Study method described in Appendix 1.

[0096] (Appendix 8) the second model comprises one or more element radiation models; the element radiation model is a model that estimates a radiation pattern of each antenna element from an actual measurement value of SINR and / or received power, When there are multiple element radiation models, an averaging unit is installed to average the outputs and determine the radiation shape of each antenna element. Study method described in Appendix 7.

[0097] (Appendix 9) The second model has a phase random circuit that randomizes the phase at the input of the element radiation model. Study method described in Appendix 8.

[0098] (Appendix 10) the second model includes a gain and phase setting error correction model; The gain and phase setting error correction model is a model that receives gain and phase setting values ​​for an antenna element as input and outputs an error between the setting value and an actual output value of the antenna element. 10. A learning method as described in appendix 7 to 9.

[0099] (Appendix 11) the second model has an amplitude mask and a phase range limiter; the amplitude mask limits the gain of the output of the radiation shape of each of the antenna elements within a predetermined range; The phase range limiting unit limits the phase of the output of the radiation pattern of each of the antenna elements within a predetermined range. The study method described in Appendix 8 or 9.

[0100] (Appendix 12) the second model has a filter section at its output section; The filter unit smoothes the output when a value steeper than a predetermined value exists. 10. A learning method as described in appendix 7 to 9.

[0101] (Appendix 13) the third calculation process is performed by a third model; The learning process further uses the calculated reception state calculated for each of the plurality of terminal devices and the measured reception state actually measured by each of the terminal devices, causing the third model to learn about the propagation path characteristics; Study method described in Appendix 1.

[0102] (Appendix 14) the third calculation process is performed by a third model; Using the first information and the reception state information, the third model is caused to perform learning related to estimation of the plurality of terminal devices. Study methods described in Appendix 7 or 13.

[0103] (Appendix 15) the third model includes a model for one or more beam IDs; The beam ID model outputs an estimated position of the terminal device according to the first information. Study method described in Appendix 13.

[0104] (Appendix 16) the third model has a selector connected to the beam ID model, When a plurality of beam ID models are installed, the selector selects the beam ID model from the plurality of beam ID models. Study method described in Appendix 15.

[0105] (Appendix 17) the third model has a range limiting portion at the rear of the beam ID model, The range limiting unit limits the output of the beam ID model to within a predetermined range. A learning method as described in Appendix 15 or 16. [Explanation of symbols]

[0106] 10: Wireless communication system 11: Radiation propagation model 12:BF calculation section 13: Propagation loss model 14:SINR calculation section 15: NN optimization section 20: Element radiation NN model 21: Average part 22: Phase range limiting section 23: Filter section 24: Phase setting error correction NN model 25: Random circuit 30: Selector 33: Range restriction section 100: Terminal device 200:Base station equipment 300: Simulation device 310: CPU 320: Storage 321:NN Program 322:NN Learning Program 330: Memory

Claims

1. a first calculation process for outputting setting information to be set on antenna elements that form the beam from first information that identifies the radiation shape of the beam; a second calculation process for outputting radio wave radiation shape information relating to the radio wave radiation shape of the antenna element from the setting information, the shape of the antenna element, and coordinates of the installation position; a third calculation process for outputting received power information relating to the received power of each of a plurality of terminal devices from the radio wave radiation shape information and the transmission path characteristics, for each of the plurality of beams; a fourth calculation process of outputting reception state information regarding the reception state of the beam for each terminal device from the reception power information of each of the plurality of terminal devices output for each of the plurality of beams; a learning process for learning a first model that executes the first calculation process using the first information and the reception state information; How to learn.

2. the first model comprises a neural network; In the learning process, the connection weights of the neural network are optimized. The learning method according to claim 1.

3. The learning process updates the weights of the connections so that the reception state of the beams for each of the plurality of terminal devices is improved. The learning method according to claim 2.

4. The setting information includes a phase and a gain to be set to the antenna element. The learning method according to claim 1.

5. the radiation shape of each of the plurality of beams corresponds to a respective region, and the plurality of terminal devices are located within the respective regions; the third calculation process calculates a received power of each of the plurality of terminal devices using terminal position information relating to positions of the plurality of terminal devices corresponding to the radiation shapes of each of the plurality of beams; The fourth calculation process acquires the received power of a desired wave and an interference wave in each of the plurality of terminal devices from the received power of each of the plurality of beams of the plurality of terminal devices, and calculates an SINR (Signal to Interference plus Noise Ratio) of each of the plurality of terminal devices as the reception state; The learning process updates the coupling weights so that the SINR of each of the plurality of terminal devices increases. The learning method according to claim 2.

6. The reception state further includes a transmission power of the antenna element; The learning process updates the coupling weights according to the SINR and the transmission power of each of the plurality of terminal devices. The learning method according to claim 5.

7. the second computation is performed by a second model; The second model is a trained model trained using the received power and SINR actually measured by the terminal device. The learning method according to claim 1.

8. the second model comprises one or more element radiation models; the element radiation model is a model that estimates a radiation pattern of each antenna element from actual measured values ​​of SINR and / or received power, When there are multiple element radiation models, an averaging unit is installed to average the outputs and determine the radiation shape of each antenna element. The learning method according to claim 7.

9. The second model has a phase random circuit that randomizes the phase at the input of the element radiation model. The learning method according to claim 8.

10. the second model includes a gain and phase setting error correction model; The gain and phase setting error correction model is a model that receives gain and phase setting values ​​for an antenna element as input and outputs an error between the setting value and an actual output value of the antenna element. A learning method according to any one of claims 7 to 9.

11. the second model has an amplitude mask and a phase range limiter; the amplitude mask limits the gain of the output of the radiation shape of each of the antenna elements within a predetermined range; The phase range limiting unit limits the phase of the output of the radiation pattern of each of the antenna elements within a predetermined range.

10. The learning method according to claim 8 or 9.

12. the second model has a filter section at its output; The filter unit smoothes the output when a value steeper than a predetermined value exists. A learning method according to any one of claims 7 to 9.

13. the third calculation process is performed by a third model; The learning process further uses the calculated reception state calculated for each of the plurality of terminal devices and the measured reception state actually measured by each of the terminal devices, causing the third model to learn about the propagation path characteristics; The learning method according to claim 1.

14. the third calculation process is performed by a third model; Using the first information and the reception state information, the third model is caused to perform learning related to estimation of the plurality of terminal devices.

14. The learning method according to claim 7 or 13.

15. the third model includes one or more models for beam ID; The beam ID model outputs an estimated position of the terminal device according to the first information. The learning method according to claim 13.

16. the third model has a selector connected to the beam ID model, When a plurality of beam ID models are installed, the selector selects the beam ID model from the plurality of beam ID models. The learning method according to claim 15.

17. the third model has a range limiting portion at the rear of the beam ID model, The range limiting unit limits the output of the beam ID model to within a predetermined range.

17. A learning method according to claim 15 or 16.

Images