Communication relay equipment, remote control equipment, systems, methods and programs

Abstract

This invention provides an aerial relay type communication relay device that can achieve stable service link communication even when the aircraft is moving. [Solution] The communication relay device forms one or more cells toward a service area on land or at sea via a service link array antenna of a relay communication station located on an aircraft positioned in the air, and communicates wirelessly with a plurality of terminal devices located within the cells. The communication relay device includes a storage unit that stores information on the transition tolerance of the cells, which changes according to the amount of movement of the aircraft, and a calculation unit that intermittently performs a process to optimize the service area in which the cells are formed within the range of the transition tolerance of the cells, according to a predetermined period or predetermined conditions.

Description

[Technical Field]

[0001] This disclosure relates to the optimization of service areas in which cells are formed from an aerial relay type (airborne stationary type) communication relay device toward the ground or sea. [Background technology]

[0002] Conventionally, there is a known method of area optimization that optimizes the antenna parameters of a service link (e.g., the direction and width of the beam for cell formation) so that a desired communication quality (e.g., throughput) can be obtained over the entire service area (hereinafter simply referred to as "area") formed on the ground by communication relay equipment such as a High Altitude Platform Station (HAPS) (also called a "High Altitude Pseudo-Satellite") that can float and stay in the air.

[0003] Patent Document 1 discloses an aerial levitation type communication relay device comprising: an array antenna having a plurality of antenna elements that form a cell for wireless communication of a service link between terminal devices; an information acquisition unit that acquires information on at least one of the position and orientation of a communication relay device; and a control unit that controls the phase and amplitude of a plurality of transmit and receive signals transmitted and received through each of the plurality of antenna elements of the array antenna so as to fix the position of the cell footprint based on the information on at least one of the position and orientation of the communication relay device.

[0004] Patent Document 2 discloses an aerial relay type communication relay device that optimizes the antenna parameters of a service link array antenna to form the cell corresponding to the distribution of multiple terminal devices in a service area based on location information of multiple terminal devices, and controls the direction of the beam formed by the antenna or the main directivity direction of the antenna to fix the position of the cell footprint in the service area based on information of at least one of the aircraft's position and attitude.

[0005] Non-patent document 1 discloses a method for optimizing cell configurations that are independent of the actual antenna configuration by using a modeled beam pattern.

[0006] Non-patent document 2 discloses an antenna weight calculation method for generating a beam that approximates a desired beam. [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] Japanese Patent Publication No. 2020-036070 [Patent Document 2] Japanese Patent Publication No. 2025-067240 [Non-patent literature]

[0008] [Non-Patent Document 1] Y. Shibata, N. Kanazawa, M. Konishi, K. Hoshino, Y. Ohta, and A. Nagate, "System design of gigabit HAPS mobile communications," IEEE Access, vol. 8, pp. 157995-158007, 2020. [Non-Patent Document 2] K. Hoshino, S. Sudo, and Y. Ohta, "A study on antenna beamforming method considering movement of solar plane in HAPS system," in Proc. IEEE 90th Veh. Technol. Conf. (VTC-Fall), Sept. 2019, pp. 1-5. [Overview of the project]

[0009] A communication relay device according to an aspect of the present disclosure is an airborne relay type communication relay device that forms one or more cells toward a service area on the ground or at sea via an array antenna for service links of a relay communication station provided on an aircraft located in the air and wirelessly communicates with a plurality of terminal devices within the cells. This communication relay device includes a storage unit that stores information on the transition allowable capacity of the cells that changes according to the movement amount of the aircraft, and a calculation unit that intermittently executes a process of optimizing the service area in which the cells are formed within the range of the transition allowable capacity of the cells according to a predetermined period or predetermined conditions.

[0010] In the communication relay device, the calculation unit performs an operation of minimizing the value of a loss function reflecting the information on the transition allowable capacity of the cells in a mathematical model of service link communication set based on the information on the number of cells and the information on the antenna configuration, thereby obtaining an antenna weight for optimizing the service area within the range satisfying the transition allowable capacity of the cells, and may intermittently update the antenna weight according to a predetermined period or predetermined conditions.

[0011] In the communication relay device, the operation of minimizing the value of the loss function may be an optimization operation that calculates the gradient of the loss function and repeatedly updates the antenna weight by the gradient descent method based on the gradient a predetermined number of times.

[0012] The communication relay device may include a control unit that controls the directivity beam of the array antenna for service links by applying the optimized antenna weight to a plurality of antenna elements of the array antenna for service links.

[0013] In the communication relay device, the calculation unit dynamically optimizes the antenna weight by dynamically performing an operation of minimizing the value of the loss function using a learned AI model, and may include a control unit that controls the directivity beam of the array antenna for service links by applying the dynamically optimized antenna weight to a plurality of antenna elements of the array antenna for service links.

[0014] A remote control device according to another aspect of the present disclosure is a remote control device capable of communicating with an airborne relay-type communication relay device that forms one or more cells toward a ground or sea service area via an array antenna for a service link provided on an aircraft located in the air and wirelessly communicates with a plurality of terminal devices within the cells. This remote control device stores initial setting information including information on the number of cells formed in the service area and information on the antenna configuration of the service link array antenna, and information on the transition allowance capacity of the cells that changes according to the movement amount of the aircraft. In a mathematical model of service link communication set based on the information on the number of cells and the antenna configuration information, an operation is performed to minimize the value of a loss function reflecting the information on the transition allowance capacity of the cells, thereby obtaining an antenna weight that optimizes the service area within the range that satisfies the transition allowance capacity of the cells, and intermittently updating the antenna weight according to a predetermined cycle or predetermined conditions. An arithmetic unit, for a plurality of conditions in which the position, attitude, or both of the aircraft are different, a teacher data generation unit that generates a plurality of sets of teacher data consisting of a combination of information on the position, attitude, or both of the aircraft and the optimized antenna weight, which is used for machine learning of an AI model for dynamically estimating the optimized antenna weight, and a learning unit that executes machine learning by applying the plurality of sets of teacher data to the AI model to create a learned AI model.

[0015] In the remote control device, the AI model may be a neural network.

[0016] In the remote control device, the operation of minimizing the value of the loss function may be an optimization operation that calculates the gradient of the loss function and repeatedly performs a process of updating the antenna weight by the gradient descent method based on the gradient a predetermined number of times.

[0017] In the communication relay device and the remote control device, let the antenna weight be W, the loss function be E, and the index for maximizing the communication capacity in the service area be f capacityLet (W) be the penalty function to prevent the coverage of the service area by the cell from falling below the required amount, and E coverage Let (W) be the value of the transmission power or power consumption in the service link communication of the relay station, and let E be the penalty function to prevent the transmission power or power consumption from exceeding a standard value. power Let (W) be the penalty function to prevent fast cell transitions, and let E be the penalty function to prevent fast cell transitions. shift Let (W) be the coefficients indicating the influence rate, and let λ1, λ2, λ3, and λ4 be the coefficients indicating the influence rate. The loss function E may be expressed by the following equation (1).

number

[0018] A system according to yet another aspect of the present invention comprises any of the above-mentioned remote control devices and an aerial relay type communication relay device that forms one or more cells toward a ground or sea service area via a service link array antenna of a relay communication station located on an aircraft in the air and communicates wirelessly with a plurality of terminal devices located in the cells.

[0019] In the above system, the calculation unit of the communication relay device dynamically optimizes the antenna weights by dynamically performing calculations to minimize the value of the loss function using a trained AI model transferred from the remote control device, and the communication relay device may include a control unit that controls the directional beam of the service link array antenna by applying the dynamically optimized antenna weights to a plurality of antenna elements of the service link array antenna.

[0020] A method relating to yet another aspect of the present disclosure is a method for optimizing a ground or sea service area in which one or more cells are formed via a service link array antenna of a relay communication station located on an aircraft in the air. The method includes storing information on the transition tolerance of the cells, which changes according to the amount of movement of the aircraft, and intermittently performing a process to optimize the service area in which the cells are formed within the range of the transition tolerance of the cells, according to a predetermined period or predetermined conditions.

[0021] A method relating to yet another aspect of the present disclosure is a method for optimizing a ground or sea service area in which one or more cells are formed via a service link array antenna of a relay communication station located on an aircraft positioned in the air. The method includes storing initial setting information including information on the number of cells to be formed in the service area and information on the antenna configuration of the service link array antenna, and information on the cell transition tolerance which changes according to the amount of movement of the aircraft; and determining an antenna weight that optimizes the service area within a range that satisfies the cell transition tolerance by performing a calculation to minimize the value of a loss function that reflects the cell transition tolerance information in a mathematical model of service link communication set based on the information on the number of cells and the antenna configuration information, and updating the antenna weight intermittently according to a predetermined period or predetermined conditions.

[0022] A program relating to yet another aspect of this disclosure is a program executed on a computer or processor provided in an airborne relay type communication relay device that forms one or more cells toward a ground or sea service area via a service link array antenna of a relay communication station provided on an aircraft located in the air, and communicates wirelessly with a plurality of terminal devices located within the cells. The program includes program code for storing information on the transition tolerance of the cells, which changes according to the amount of movement of the aircraft, and program code for intermittently executing a process to optimize the service area in which the cells are formed, within the range of the transition tolerance of the cells, according to a predetermined period or predetermined conditions.

[0023] A program relating to yet another aspect of this disclosure is a program executed on a computer or processor installed in an airborne relay-type communication relay device that forms one or more cells toward a ground or sea service area via a service link array antenna of a relay communication station installed on an aircraft located in the air, and communicates wirelessly with a plurality of terminal devices located within the cells. The program includes storing initial setting information including information on the number of cells to be formed in the service area and information on the antenna configuration of the service link array antenna, and information on the cell transition tolerance which changes according to the amount of movement of the aircraft, and performing a calculation to minimize the value of a loss function that reflects the cell transition tolerance information in a mathematical model of service link communication set based on the information on the number of cells and the antenna configuration information, thereby determining an antenna weight that optimizes the service area within a range that satisfies the cell transition tolerance, and intermittently updating the antenna weight according to a predetermined period or predetermined conditions.

[0024] The program may include, in whole or in part, a trained model created by machine learning. [Brief explanation of the drawing]

[0025] [Figure 1] Figure 1 is an explanatory diagram showing an example of the overall configuration of a communication system according to the embodiment. [Figure 2] Figure 2(a) shows a cell configuration consisting of 3 cells. Figure 2(b) shows a cell configuration consisting of 7 cells. Figure 2(c) shows a cell configuration consisting of 12 cells. Figure 2(d) shows a cell configuration consisting of 19 cells. [Figure 3] Figure 3 shows an example of the results of optimizing the cell configuration in the reference example. [Figure 4] Figure 4 shows an example of antenna tilt angle, horizontal half-width, and vertical half-width as antenna parameters used to optimize cell configuration. [Figure 5]Figure 5 is a characteristic diagram showing an example of a modeled beam pattern for the directional beam of a service link array antenna in the optimization of a cell configuration related to a reference example. [Figure 6] Figure 6 is a characteristic diagram showing an example of the gap between the modeled beam pattern and the actual beam pattern of a service link array antenna (a 4-element linear array antenna) in the optimization of the cell configuration for reference. [Figure 7] Figure 7(a) shows a SINR heatmap based on the modeled beam pattern of the directional beam of the service link array antenna in the optimization of the cell configuration in the reference example. Figure 7(b) shows a SINR heatmap based on the actual beam pattern of the directional beam of the service link array antenna (4×20 element cylinder array antenna) in the optimization of the cell configuration in the reference example. [Figure 8] Figures 8(a) to 8(d) show the cell configurations in the optimization of the cell configuration for the reference example, respectively. [Figure 9] Figure 9 is a characteristic diagram showing the modeled beam pattern of the directional beam of the service link array antenna in the optimization of the cell configuration for the reference example. [Figure 10] Figure 10 shows the optimized antenna parameters in the cell configuration optimization for a reference example. [Figure 11] Figure 11 shows the (3,9) cell configuration optimized using a genetic algorithm in the cell configuration optimization example. [Figure 12] Figure 12(a) shows the calculation of the horizontal antenna weights required to achieve the desired horizontal beam in the cell configuration optimization for the reference example. Figure 12(b) is a characteristic diagram showing the horizontal beam pattern when the same horizontal antenna weights are applied. [Figure 13]Figure 13(a) shows the calculation of vertical antenna weights to achieve the desired vertical beam in cell configuration optimization according to the reference example. Figure 13(b) is a characteristic diagram showing the vertical beam pattern when the same vertical antenna weights are applied. [Figure 14] Figure 14 is a block diagram showing an example of the main components of a relay communication station in a communication relay device (HAPS) according to an embodiment. [Figure 15] Figure 15 is a flowchart showing an example of antenna weight optimization processing in a system consisting of a communication relay device (HAPS) and a remote control device according to an embodiment. [Figure 16] Figure 16(a) is an external perspective view showing an example of a service link array antenna of a communication relay device (HAPS) according to an embodiment. Figure 16(b) is a diagram showing an example of a model of the antenna configuration of the same service link array antenna. Figure 16(c) is a diagram showing another example of a model of the antenna configuration of the service link array antenna. Figure 16(d) is a diagram showing yet another example of a model of the antenna configuration of the service link array antenna. [Figure 17] Figure 17 shows an example of the cell configuration and initial settings of the antenna weights in the antenna weight optimization process according to the embodiment. [Figure 18] Figure 18(a) is a characteristic diagram showing an example of the initial setting of the vertical beam pattern of a service link array antenna in the antenna weight optimization process according to the embodiment. Figure 18(b) is a characteristic diagram showing an example of the initial setting of the horizontal beam pattern of the same service link array antenna. [Figure 19] Figure 19(a) shows an example of direct optimization of antenna weights using the gradient descent method according to the embodiment. Figure 19(b) shows an example of cell configuration before the antenna weight optimization process according to the embodiment. Figure 19(c) shows an example of cell configuration after the antenna weight optimization process according to the embodiment. [Figure 20]Figure 20 shows an example of a system model (mathematical model) used in the antenna weight optimization process according to the embodiment. [Figure 21] Figure 21 shows an example of a configuration model of the transmitting section of a relay station in a High-Area Communication System (HAPS) in the system model (mathematical model) of Figure 20. [Figure 22] Figure 22 shows an example of iterative calculation in the antenna weight optimization process according to the embodiment. [Figure 23] Figure 23(a) shows an example of the coordinate definition of the attitude of a High-Area Communication System (HAPS) in the generation of training data for machine learning of an AI model used in the dynamic optimization process of antenna weights according to the embodiment. Figure 23(b) shows an example of the attitude change of the service link array antenna model of the HAPS. [Figure 24] Figure 24(a) shows an example of defining the yaw angle coordinates of a High-Area Connectivity System (HAPS) in the generation of training data for machine learning of an AI model. Figure 24(b) shows an example of sweeping the yaw angle of a High-Area Connectivity System (HAPS) during machine learning of an AI model. [Figure 25] Figure 25(a) shows an example of defining the coordinates of the pitch angle of a High-Area Connectivity System (HAPS) in the generation of training data for machine learning of an AI model. Figure 25(b) shows an example of sweeping the pitch angle of a High-Area Connectivity System (HAPS) during machine learning of an AI model. [Figure 26] Figure 26(a) shows an example of defining the coordinates of the roll angle of a High-Area Connectivity System (HAPS) in the generation of training data for machine learning of an AI model. Figure 26(b) shows an example of sweeping the roll angle of a High-Area Connectivity System (HAPS) during machine learning of an AI model. [Figure 27] Figure 27 shows an example of the configuration of a machine learning model of a neural network (AI model) according to the embodiment. [Figure 28]Figure 28(a) shows the results of a computer simulation of the SINR distribution using a conventional method. Figure 28(b) shows an example of the results of a computer simulation of the SINR distribution when the antenna weights are optimized using a machine learning-prepared neural network (AI model) in Scenario 1, to which the proposed method of the embodiment is applied. [Figure 29] Figure 29 shows an example of the results of a computer simulation of the CDF curve of SINR when the antenna weights are optimized using a machine learning-prepared neural network (AI model) in Scenario 1, to which the proposed method of the embodiment is applied. [Figure 30] Figure 30 shows an example of the results of a computer simulation of the CDF curve for frequency utilization efficiency when the antenna weights are optimized using a machine learning-prepared neural network (AI model) in Scenario 1, to which the proposed method of the embodiment is applied. [Figure 31] Figure 31 shows an example of the results of a computer simulation of the CDF curve of the footprint fixing rate when the antenna weights are optimized using a machine learning-prepared neural network (AI model) in scenario 2 of the communication relay device (HAPS) according to the embodiment. [Figure 32] Figure 32 shows an example of the results of a computer simulation of the CDF curve of the frequency utilization efficiency percentile when the antenna weights are optimized using a machine learning-prepared neural network (AI model) in scenario 2 of the communication relay device (HAPS) according to the embodiment. [Figure 33] Figure 33(a) shows an example of measured time variation of the yaw angle of a communication relay device (HAPS) in a computer simulation of scenario 3 of the HAPS according to the embodiment. Figure 33(b) shows an example of measured time variation of the pitch angle and roll angle of the communication relay device (HAPS) in a computer simulation of the same scenario 3. [Figure 34]Figure 34 shows an example of the results of a computer simulation of the time variation of the footprint fixing rate when the antenna weights are optimized using a machine learning-prepared neural network (AI model) in Scenario 3 of the communication relay device (HAPS) according to the embodiment. [Figure 35] Figure 35 shows an example of the results of a computer simulation of the time variation of the frequency utilization efficiency percentile when the antenna weights are optimized using a machine learning-prepared neural network (AI model) in scenario 3 of the communication relay device (HAPS) according to the embodiment. [Figure 36] Figures 36(a) and 36(b) show examples of computer simulation results for a SINR heatmap in a reference example where footprint fixing is not performed when the HAPS is orbiting in the air, and an example of the results of a computer simulation for an embodiment (scenario 3), respectively. Figures 36(c) and 36(d) show examples of computer simulation results for a reference example where footprint fixing is performed when the HAPS is orbiting in the air, and an example of the results of a computer simulation for an embodiment (scenario 3), respectively. [Figure 37] Figure 37 shows an example of attitude change in a model of a service link array antenna for a High-Area Communication System (HAPS). [Figure 38] Figure 38 shows an example of extending the communication relay device (HAPS) according to the embodiment to a user distribution-aware neural network (AI model). [Figure 39] Figure 39 shows another example of the extension of the communication relay device (HAPS) according to the embodiment to a user distribution-aware neural network (AI model). [Figure 40] Figure 40 shows an example of the movement of a communication relay device (HAPS) according to this embodiment. [Figure 41] Figure 41 shows an example of area optimization when a communication relay device (HAPS) according to this embodiment is in motion. [Figure 42]Figure 42 shows an example of cell transition area. [Figure 43] Figure 43 shows an example of a cell configuration formed in a service area by a communication relay device (HAPS) according to this embodiment. [Figure 44] Figure 44 is a block diagram showing an example of the main components of a relay communication station in a communication relay device (HAPS) according to an embodiment. [Figure 45] Figure 45 is a flowchart showing an example of area optimization processing at a relay station of a communication relay device (HAPS) according to an embodiment. [Figure 46] Figure 46 shows an example of the distribution of actual users (terminal devices) in multiple cells formed in a service area by a communication relay device (HAPS) according to the embodiment. [Figure 47] Figure 47 shows an example of the distribution of virtual users (terminal devices) in the system model (mathematical model) used for the antenna weight optimization process according to the embodiment. [Figure 48] Figure 48 shows an example of the configuration of the transmitting section of a relay station in a High-Area Communication System (HAPS) in the system model (mathematical model) of Figure 47. [Figure 49] Figure 49 shows an example of a definition of time in the minute movement of a communication relay device (HAPS) according to the embodiment. [Figure 50] Figure 50 shows an example of the cell transition area during minute movement of a communication relay device (HAPS) according to the embodiment. [Figure 51] Figure 51 shows an example of antenna weight calculation using a machine learning-prepared neural network (AI model) according to the embodiment. [Modes for carrying out the invention]

[0026] Embodiments of this disclosure will be described below with reference to the drawings. Note that each drawing is merely a schematic representation of the shape, size, positional relationships, correspondences, configuration, processing, steps, etc., to the extent that the contents of this disclosure can be understood. Therefore, this disclosure is not limited to the shapes, sizes, positional relationships, correspondences, configurations, processing, steps, and steps exemplified in each drawing. Furthermore, the numerical values ​​exemplified in this disclosure are merely preferred examples, and therefore, this disclosure is not limited to the numerical values ​​exemplified.

[0027] One embodiment of this disclosure optimizes antenna weights applied to a service link array antenna that forms multiple cells in a service area from an overhead relay type communication relay device (HAPS) as an aerial platform to the ground or sea. In one embodiment, the cell configuration of the multiple cells in the service area and the antenna weights applied to the service link array antenna are optimized simultaneously. In yet another embodiment, the antenna weights are directly optimized using gradient descent. In yet another embodiment, dynamic antenna weight optimization is achieved using a machine learning-based AI model (e.g., a neural network).

[0028] In other embodiments of this disclosure, area optimization is performed to intermittently (continuously) optimize the service area while smoothly transitioning the cells formed in the service area from an overhead relay type communication relay device (HAPS) as an aerial platform toward the ground or sea. In one embodiment, area optimization is performed intermittently at a high frequency to achieve stable service link communication even when the aircraft is moving. In one embodiment, antenna weights applied to a service link array antenna forming multiple cells are optimized. In another embodiment, the cell configuration of multiple cells in the service area and the antenna weights applied to the service link array antenna are optimized simultaneously. In yet another embodiment, the antenna weights are directly optimized using gradient descent. In yet another embodiment, dynamic antenna weight optimization is achieved using a machine learning-prepared AI model (e.g., a neural network).

[0029] Figure 1 is an explanatory diagram showing an example of the overall configuration of a communication system according to one embodiment of this disclosure. The communication system according to this embodiment is suitable for realizing a 3D network of fifth-generation or later mobile communications that can handle simultaneous connection to a large number of terminal devices (hereinafter referred to as "UEs") 61 and low latency. Furthermore, the mobile communications standards applicable to the communication system, radio relay station, base station, repeater and UE disclosed herein include fifth-generation mobile communications standards and standards for fifth-generation and subsequent generations of mobile communications.

[0030] As shown in Figure 1, the communication system includes a High Altitude Platform Station (HAPS) 10 (also called a "High Altitude Pseudo-Satellite" or "Stratospheric Platform") which is an aerial relay type communication relay device (wireless relay device) that constitutes the aerial platform. The HAPS 10 is an aerial stationary or airborne communication relay device that is located in the airspace at a predetermined altitude and forms a 3D cell (3D area) in the target airspace at a predetermined altitude toward the target service area (for example, an ultra-wide area with a radius of about 100 km) 20A.

[0031] HAPS10 is a flying or floating vehicle 100 equipped with a relay communication station 110, which is controlled by autonomous or external control to float or fly in high-altitude airspace (float airspace) at an altitude of 100 km or less above the ground or sea surface. The airspace in which HAPS10 is located may be, for example, stratospheric airspace with an altitude H of 18 km or more and 50 km or less. This airspace may also be airspace with relatively stable weather conditions at an altitude of 15 km or more and 25 km or less, and in particular, airspace at an altitude of approximately 20 km.

[0032] The target airspace for cell formation in HAPS10 may be a predetermined altitude range (for example, an altitude range of 50[m] or less, or an altitude range of 50[m] or more but 1000[m] or less) located between the airspace where HAPS10 is located and the near-ground cell formation area covered by conventional base stations such as macrocell base stations (e.g., LTE eNodeB or next-generation gNodeB).

[0033] The target airspace for cell formation may be over the sea, a river, or a lake. Furthermore, the three-dimensional cells formed by HAPS10 may be configured to reach the ground or sea surface to enable communication with UE61 located on land or at sea.

[0034] HAPS10 communicates wirelessly with UE61 via a service link antenna (also called an "SL antenna") 111 of a relay communication station 110 installed on an aircraft 100 such as a flying object or floating object located in the air. HAPS10 can fly on electricity, for example, by being equipped with at least one of a battery and a solar power generation system. HAPS10 may be a solar-powered plane type HAPS as shown in the figure, or an airship type HAPS. Furthermore, the HAPS10 on which the relay communication station 110 is installed may be an artificial satellite (e.g., a communications satellite), a balloon, or an unmanned aerial vehicle (UAV) such as a drone or UAS (Unmanned Aircraft Systems). HAPS10 may also fly using at least one of a battery and an engine as a power source. The UAV may be, for example, an unmanned aircraft that flies on fuel, or a drone that flies on a battery or the like.

[0035] The relay communication station 110 includes a service link array antenna (SL antenna) 111 and a feeder link antenna (hereinafter also referred to as "FL antenna") 112. The relay communication station 110 can communicate with the UE61 via the SL antenna 111 over the service link SL.

[0036] The SL antenna 111 is a beamforming-controllable array antenna that can control the direction and width of each of the multiple beams 20B(1) to 20B(7) that form multiple cells 20C(1) to 20C(7) in a target service area (for example, an ultra-wide area with a radius of about 100 km) 20A. The relay communication station 110 performs beamforming control, for example, to control the phase and amplitude of multiple transmit and receive signals transmitted and received through each of the antenna elements constituting the SL antenna (array antenna) 111. The area through which the beams 20B(1) to 20B(7) pass in the cell formation target airspace may be a three-dimensional cell. Multiple beams adjacent to each other in the cell formation target airspace may partially overlap. Multiple cells 20C(1) to 20C(7) formed on land (or at sea, etc.) by multiple beams 20B(1) to 20B(7) are each also called footprints.

[0037] The relay station 110 may apply beam control to each of the radio resources (time-frequency resources) common to each cell in the wireless communication of the service links of multiple cells 20C(1) to 20C(7) (for example, in a 4G system). The relay station 110 may also apply beam control to specific radio resources (time-frequency resources) in a cell in the wireless communication of the service links of multiple cells 20C(1) to 20C(7). For example, in the case of massive MIMO in a 5G system, synchronization signals (SS) and broadcast channels (PBCH) need to reach all terminal devices 61 in the cell. Therefore, the relay station 110 may define an SS / PBCH block (SSB) as a single unit and perform control (beam sweeping) that transmits while sequentially switching beams for each SSB.

[0038] In the illustrated example, seven cells 20C(1) to 20C(7) are formed via the SL antenna 111, but the number of cells 20C can be single, two to six, or eight or more.

[0039] The SL antenna 111 is, for example, a single or multiple array antenna in which multiple antenna elements are arranged two-dimensionally or three-dimensionally, and which is capable of forming multiple beams toward the ground. The SL antenna 111 may also be a Massive antenna in which a large number of antenna elements are arranged two-dimensionally and the beam directivity in the horizontal and vertical directions can be controlled.

[0040] The body of the SL antenna 111 may have any shape. For example, the body shape of the SL antenna 111 may be a cylindrical shape (cylinder shape) or a prism shape in which multiple antenna elements are arranged in the circumferential direction (horizontal plane direction) and axial direction (vertical plane direction) of the outer surface. Alternatively, the body shape of the SL antenna 111 may be a partial cone shape or a partial pyramidal shape with a plane formed on the downward-facing apex side. Furthermore, the body shape of the SL antenna 111 may be a hemispherical shape with a downward-facing sphere, or a sphere.

[0041] The relay communication station 110 can communicate via feeder link FL with a gateway device for HAPS (also called a "feeder station"; hereinafter referred to as a "GW station") 70 located on land (or at sea, etc.) via the FL antenna 112. The FL antenna 112 is, for example, an array antenna whose directivity (direction of the directional beam) can be controlled. The FL antenna 112 is, for example, one or more array antennas in which multiple antenna elements are arranged two-dimensionally or three-dimensionally. The FL antenna 112 may also be a Massive antenna in which many antenna elements are arranged two-dimensionally and whose horizontal and vertical directivity can be controlled. In the figure, feeder link FL(F) is a forward link from GW station 70 to UE61 via HAPS 10, and feeder link FL(R) is a reverse link from UE61 to GW station 70 via HAPS 10.

[0042] The relay communication station 110 mounted on the HAPS10 aircraft 100 may be a repeater-type relay communication station that relays the transmitted and received signals without regenerating them, or it may be a base station-type relay communication station that has base station equipment that regenerates the transmitted and received signals, remodulates the regenerated signals, and relays them.

[0043] The repeater-type relay station 110 functions as a repeater slave unit corresponding to a repeater master unit consisting of GW stations 70, and is, for example, a wireless relay device (hereinafter also referred to as a "frequency conversion repeater") that converts the frequency of a service link and a feeder link with a different frequency. In the downlink, the relay station 110 converts the frequency of the feeder link transmitted from the base station equipment 80 via the GW station 70 to the frequency of the service link and transmits it to the UE 61. On the other hand, in the uplink, the relay station 110 converts the frequency of the service link transmitted from the UE 61 to the frequency of the feeder link and transmits it to the base station equipment 80 via the GW station 70.

[0044] A repeater-type relay communication station 110 includes, for example, a repeater and a frequency converter. The repeater includes, for example, a low-noise amplifier that amplifies the received signal of service link SL received via the SL antenna 111, and a power amplifier that amplifies the transmitted signal of service link SL transmitted via the SL antenna 111. The frequency converter performs conversion between the frequency of service link SL and the frequency of feeder link FL.

[0045] The base station type relay station 110 has base station equipment and a frequency converter. The base station equipment includes a baseband processing unit for processing the service link baseband signal, a communication interface unit for communicating with the core network of the mobile communication network 90 via a backhaul line through the GW station 70, and the like. The frequency converter converts between the frequency of the service link signal input and output to the base station equipment within the relay station 110 and the frequency of the feeder link signal transmitted and received via the FL antenna 112.

[0046] In the embodiments described below, the case where the relay communication station 110 mounted on the HAPS10 unit 100 is a repeater-type relay communication station (repeater slave unit) will be mainly explained.

[0047] The UE (User Equipment) 61 is a terminal device (hereinafter also referred to as "User") used by a user on land or at sea. The UE 61 is, for example, a mobile phone, a smartphone, a portable personal computer with mobile communication capabilities, and is also called a mobile terminal, mobile station, mobile device, or portable communication terminal. The UE 61 may be a modular mobile station incorporated into a mobile device such as a vehicle, a drone such as a small, remotely controlled helicopter, or a terminal device for IoT (Internet of Things) devices.

[0048] HAPS10 may autonomously control its own levitation movement (flight) and processing and control at the relay communication station 110 by executing a control program in a control unit consisting of an internally built-in computer or processor. For example, HAPS10 can autonomously perform area optimization control and FP fixed control as described later. In addition, HAPS10 may acquire its own current location information (e.g., GNSS (Global Navigation Satellite System) location information such as GPS location information), pre-stored location control information (e.g., flight schedule information), location information of other HAPS located in the vicinity, and autonomously control its levitation movement (flight) and processing and control at the relay communication station 110 based on this information.

[0049] The position and attitude information of HAPS10 may be obtained based on the output of a GPS receiver, gyro sensor, accelerometer, inertial sensor, etc., incorporated into HAPS10. For example, the position and attitude information of HAPS10 may be obtained based on the output of a GNSS / INS system, which combines a GNSS system and an inertial measurement unit (IMU) incorporated into HAPS10.

[0050] Furthermore, the levitation and movement (flight) of the HAPS10, as well as the processing and control at the relay communication station 110, may be controlled by a remote control device 95 located at a communication center or the like of the mobile communication network 90. ​​The remote control device 95 can be configured, for example, as a computer device such as a PC or a server. The HAPS10 may incorporate a control communication terminal device (e.g., a mobile communication module) so that it can receive control information from the remote control device 95 and transmit various information such as monitoring information to the remote control device 95, and may be assigned terminal identification information (e.g., an IP address, a telephone number, etc.) so that it can be identified by the remote control device 95. The MAC address of the communication interface may be used to identify the control communication terminal device.

[0051] The remote control device 95 may, for example, perform antenna weight optimization, area optimization, and FP fixed control, as described later, by coordinating with the HAPS 10.

[0052] Furthermore, HAPS10 may transmit monitoring information, such as information regarding the lift-off movement (flight) of itself or surrounding HAPS and processing at the relay communication station 110, the position information of HAPS10, information regarding the status of HAPS10, and observation data acquired by various sensors, to a predetermined transmission destination such as the remote control device 95. The control information may include information on the target flight route of the HAPS. The monitoring information may include at least one piece of information such as the current position of HAPS10, flight route history information, airspeed, ground speed and thrust direction, wind speed and wind direction of the airflow around HAPS10, and atmospheric pressure and temperature around HAPS10.

[0053] The duplexing scheme for the uplink and downlink of the wireless communication between the relay station 110 and the UE61 is not limited to a specific scheme, and may be, for example, a Time Division Duplex (TDD) scheme or a Frequency Division Duplex (FDD) scheme. Furthermore, the access scheme for the wireless communication between the relay station 110 and the UE61 is not limited to a specific scheme, and may be, for example, an FDMA (Frequency Division Multiple Access) scheme, a TDMA (Time Division Multiple Access) scheme, a CDMA (Code Division Multiple Access) scheme, or an OFDMA (Orthogonal Frequency Division Multiple Access) scheme. In addition, the wireless communication may utilize MIMO (Multi-Input and Multi-Output) technology, which has functions such as diversity coding, transmit beamforming, and spatial division multiplexing (SDM), and can increase the transmission capacity per unit frequency by simultaneously using multiple antennas for both transmission and reception. Furthermore, the MIMO technology may be SU-MIMO (Single-User MIMO) technology, in which one base station transmits multiple signals to one UE at the same time and frequency, or MU-MIMO (Multi-User MIMO) technology, in which one base station transmits signals to multiple different UEs at the same time and frequency, or multiple different base stations transmit signals to one UE at the same time and frequency.

[0054] In the communication system with the above configuration, for example, signals from the base station 80 are relayed by the GW station 70 and HAPS 10, enabling communication services to be provided to the UE 61 on the ground. In particular, according to the communication system of this embodiment, the HAPS 10, which functions as an aerial relay type communication relay device acting as an aerial platform, can directly provide ultra-wide-area mobile communication services from the stratosphere at altitudes of 18 km or higher and 50 km or lower (especially around 20 km) to the UE (mobile terminal) 61 on the ground. Furthermore, the aerial platform consisting of the HAPS 10 is attracting attention as a new communication form for use in large-scale disasters and other similar situations.

[0055] In the communication system of this embodiment, area optimization is performed to optimize the cell configuration and the antenna weights applied to each antenna element of the SL antenna 111 so that a desired communication quality (e.g., throughput) can be obtained throughout the service area 20A, which is composed of multiple cells 20C(1) to 20C(7) formed by an aerial platform such as HAPS10 facing the ground. For example, in this embodiment, on an aerial platform such as HAPS10 that covers the service area 20A with multiple cells 20C(1) to 20C(7), optimization processing is performed to simultaneously optimize the cell configuration and the antenna weights applied to each antenna element of the SL antenna 111 according to the population distribution (or user distribution, UE distribution, etc.).

[0056] In the communication system of this embodiment, it is necessary to accommodate a large number of users 61 on an aerial platform such as a single HAPS 10 in order to cover an ultra-wide service area 20A, and therefore an improvement in service link communication capacity is necessary. To improve communication capacity, a multi-cell HAPS system (also called a "HAPS service link system") that forms multiple cells in the service area 20A is effective.

[0057] Figures 2(a), 2(b), 2(c), and 2(d) are diagrams showing multi-cell configurations composed of 3, 7, 12, and 19 cells 20C in a service area 20A with a diameter of about 200 km, respectively. In such a HAPS service link system with a multi-cell configuration, a single HAPS 10 radiates a plurality of cell beams 20B, divides the service area (coverage area) 20A into a plurality of cells, improves the communication capacity according to the number of cells, and further improves the antenna gain and link budget by increasing the beam directivity of each cell. The overall communication capacity and frequency utilization efficiency of the service area (coverage area) 20A greatly depend on the cell configuration.

[0058] Figure 3 is a diagram showing an example of the result of optimizing the cell configuration according to a reference example. Figure 4 is a diagram showing an example of the antenna tilt angle, horizontal half-power beamwidth, and vertical half-power beamwidth as antenna parameters used for optimizing the cell configuration. Conventionally, by using a modeled beam pattern, optimization of the cell configuration independent of the actual antenna configuration has been actively carried out (see, for example, Non-Patent Document 1 mentioned above). In the conventional optimization of the cell configuration, using the beam pattern model of Figure 4, as shown in Figure 3, optimization of beam parameters for optimizing four parameters (horizontal beam direction (azimuth angle φ str ) and horizontal beamwidth (φ 3dB ) in the horizontal plane, and vertical beam direction (elevation angle (depression angle) θ str ) and vertical beamwidth (θ 3dB ) in the vertical plane) per cell beam is carried out. Here, the vertical beam direction's (elevation angle (depression angle) θ str ) is also called the tilt angle. Also, the horizontal beamwidth (φ 3dB ) and vertical beamwidth (θ 3dB ) are the angles of the beamwidth at a level 3 dB lower than the peak gain of the beam in the horizontal and vertical planes, respectively.

[0059] Figure 5 is a characteristic diagram showing an example of a modeled beam pattern of a directional beam of a service link array antenna in the optimization of a cell configuration related to a reference example. In this example of a modeled beam pattern, a model is used that approximates a planar array antenna with patch antennas (antenna elements) arranged in two dimensions, based on the IMT HAPS recommendation (ITU-R M.1456). By setting the width of the directional beam of the service link array antenna, the beam pattern G(Ψ), which is the profile of the antenna gain G in each direction (Ψ), is determined as shown in equation (2) below (see, for example, Figure 5). Research has been conducted under the assumption that this beam G(Ψ) can be directed in any direction.

number

[0060] In the optimization of the cell configuration in the above example, a difference (gap) arises between the modeled beam pattern and the beam pattern that can actually be realized. For example, as shown in Figure 6, a difference (gap) arises between the beam pattern C111 of the desired beam modeled for a 4-element linear array antenna and the actual beam pattern C112. As a result, as shown in Figures 7(a) and 7(b), a difference (gap) also arises in the SINR heatmap, which is a two-dimensional distribution of the SINR (signal-to-noise ratio) of the received signal in the service area 20A based on the beam pattern of the directional beam of the SL antenna. For this reason, after optimizing the beam parameters in the cell configuration optimization in the example, it is necessary to calculate antenna weights that can generate a beam that approximates the desired beam. In other words, a two-stage design is required, in which the antenna weights for generating the desired beam are calculated after the optimization of the beam direction and width. Furthermore, although an example of a method for calculating antenna weights to generate a beam that approximates the desired beam is proposed in the aforementioned Non-Patent Literature 2, no matter what calculation method is used, the desired beam cannot be perfectly reproduced, and a difference (gap) arises between the beam pattern of the desired beam and the actual beam pattern. In particular, the above difference (gap) becomes more pronounced when the number of antenna elements in an SL antenna is small. Due to the gap between the desired beam and the actual beam, the cell configuration optimized based on the modeled beam pattern is not necessarily the optimal cell configuration in reality.

[0061] The two-stage optimization of the cell configuration and antenna weights in the example can be performed, for example, by following the steps A1 to A4 below.

[0062] A1. First, you need to select and decide on the number of cells and cell configuration to be used for optimization. For example, you can choose and decide on one of the cell numbers and cell configurations shown in Figures 8(a) to 8(c). Note that the numbers within service area 20A in the figures are numbers that identify multiple cells.

[0063] A2. Next, the beam width and beam direction are optimized for the multiple beams of the SL antenna using the modeled beam pattern of the SL antenna. For example, using the modeled beam patterns C121 to C124 in Figure 9, the vertical beam width (θ) is optimized for each of the multiple beams that form multiple cells in the SL antenna as shown in Figure 10. 3dB ), vertical beam direction (elevation angle (depression angle) θ str ), horizontal beam width (φ 3dB ), horizontal beam direction (azimuth φ str Figure 11 shows the results of optimizing the system using a genetic algorithm (beam direction and width, and SINR heatmap) for a (3,9) cell configuration where there are 12 cells in total, with 3 cells in the central part (first layer) and 9 cells in the peripheral part (second layer) of the service area 20A.

[0064] Next, based on the optimization results of the beam width and beam direction of the multiple beams of the SL antenna 111, the antenna configuration (for example, the overall shape of the array antenna and the number and arrangement of antenna elements) is determined. For example, the antenna configuration of the array antenna 111 is determined to have the main body shape such as the aforementioned cylindrical shape, prism shape, partial cone shape, partial pyramidal shape, hemispherical shape, or spherical shape, and a predetermined number and arrangement of antenna elements.

[0065] Next, for the antenna configuration of the SL antenna 111 determined above, the weights to be applied to each signal of the multiple antenna elements 111e of the SL antenna 111 (hereinafter also referred to as "antenna weights") are calculated. For example, the antenna weights are calculated using the least-squares solution for the desired beam that realizes the optimized cell configuration.

[0066] In the case of a cylindrical SL antenna 111, the antenna weights for the horizontally arranged antenna elements 111e and the antenna weights for the vertically arranged antenna elements 111e are calculated independently of each other. For example, as shown in Figure 12(a), for a plurality of antenna elements 111e arranged at predetermined intervals (e.g., 0.6λ (wavelength)) in the horizontal direction (circumferential direction) of the SL antenna 111, the horizontal antenna weights are calculated to form a desired horizontal beam C131 as shown in Figure 12(b). Similarly, as shown in Figure 13(a), for a plurality of antenna elements 111e arranged at predetermined intervals (e.g., 0.5λ (wavelength)) in the vertical direction (axial direction) of the SL antenna 111, the vertical antenna weights are calculated to form a desired vertical beam C141 as shown in Figure 13(b).

[0067] Here, the vector a of the desired beam can be expressed by equation (3) below, for example, using the array response matrix F and the weight vector w of the antenna weights, and the weight vector w of the antenna weights can be calculated by equation (4) below.

number

number

[0068] In equation (4) above

number

[0069] If the HAPS10 aircraft changes, the direction of the beam of each cell formed by the SL antenna 111 is updated and the antenna weight is recalculated according to the change in the aircraft's attitude. This enables footprint fixation.

[0070] In the above example, the optimization process involves two stages: first optimizing the cell configuration, and then optimizing the antenna configuration and antenna weight of the SL antenna 111. In the first stage, the cell configuration (beam parameters) is optimized using a modeled beam pattern without considering the antenna configuration and antenna weight. In the second stage, the antenna weight is calculated to realize the optimized cell configuration obtained in the first stage. However, in an actual system using HAPS10, the cell configuration (beam parameters) formed in the service area 20A and the calculation of the antenna configuration and antenna weight of the SL antenna 111 influence each other.

[0071] In the embodiments of this disclosure, in light of the fact that the cell configuration (beam parameters) of the service area 20A and the antenna configuration and antenna weight calculation of the SL antenna 111 influence each other, the antenna configuration is considered from the outset, and the cell configuration and antenna weight (including the transmission power distribution for each cell) are optimized simultaneously. In one example of this embodiment, instead of a two-stage optimization process of optimizing the cell configuration (beam parameters) and then calculating the optimized antenna weight, the antenna weight is optimized directly. For example, by incorporating the cell configuration information into the antenna weight, simultaneous optimization of the cell configuration and antenna weight is achieved.

[0072] Direct optimization of antenna weights in the embodiments of this disclosure has the following advantages, for example, B1 to B4. B1. Cell configuration can be optimized considering the antenna configuration. B2. Can be applied to any array antenna. B3. Optimization can be performed flexibly by specifying only the number of cells, without limiting the cell structure. B4. By assigning power meaning to the norm of the antenna weight, it is possible to simultaneously optimize the antenna weight and the transmission power distribution between cells.

[0073] In another example of this embodiment, the antenna weights are dynamically optimized using a machine learning-based AI model (e.g., a neural network) in response to changes in the attitude of the HAPS10 aircraft, thereby achieving footprint fixation for the optimal cell configuration.

[0074] Figure 14 is a block diagram showing an example of the main components of a relay communication station 110 of a communication relay device (HAPS) 10 according to an embodiment. In Figure 14, the relay communication station 110 includes an SL antenna 111, an information acquisition unit 113, a storage unit 114, an antenna weight calculation unit 115, and a control unit 116.

[0075] The information acquisition unit 113 acquires information on the number of cells 20C formed in the service area 20A and information on the antenna configuration of the SL antenna (service link array antenna) 111. The information acquisition unit 113 may also acquire information on the position and orientation of the HAPS (communication relay device) 10, as well as time information. The information acquisition unit 113 may also acquire information on the distribution (user distribution) of terminal devices (UEs) 61 in the service area 20A.

[0076] The memory unit 114 stores initial setting information, which includes information on the number of cells 20C to be formed in the service area 20A and information on the antenna configuration of the SL antenna 111.

[0077] The antenna weight calculation unit 115 includes a calculation unit 1151 that optimizes the antenna weight by performing calculations to minimize the value of the loss function, which is a function of the antenna weight applied to the SL antenna 111, in a mathematical model of service link communication set based on the information of the number of cells 20C and the information of the antenna configuration.

[0078] In the calculation unit 1151, the loss function may be a function of the frequency utilization efficiency vector. The frequency utilization efficiency vector may be a function of the cell connection matrix representing the cells 20C to which each of the multiple terminal devices (UEs) 61 is connected.

[0079] The calculation unit 1151 may approximate the non-differentiable operator argmax with the differentiable operator softmax if the cell connection matrix includes the non-differentiable operator argmax. The calculation to minimize the value of the loss function may be an optimization calculation that calculates the gradient of the loss function and updates the antenna weights by gradient descent based on the gradient, repeating this process a predetermined number of times. The loss function may have a correction term for fixing the cell footprint when the position, attitude, or both of the aircraft change.

[0080] The calculation unit 1151 may repeatedly perform the calculation of the loss function and the calculation to minimize the value of the loss function a predetermined number of times (M times) using backpropagation. Furthermore, the calculation unit 1151 may repeat the calculation of the loss function and the calculation to minimize the value of the loss function a predetermined number of times (M times) for each of a plurality of distribution patterns in which the distribution of terminal devices (UEs) 61 in cell 20C is different from each other, for a further predetermined number of times (N times).

[0081] The antenna weight calculation unit 115 may include a machine learning-based AI model (e.g., a neural network (NN)) 1154.

[0082] The control unit 116 controls the directional beam of the SL antenna 111 by applying optimized antenna weights to multiple antenna elements 111e of the SL antenna 111.

[0083] Figure 15 is a flowchart showing an example of the antenna weight optimization process (S100) in a system consisting of a communication relay device (HAPS) 10 and a remote control device according to an embodiment. In Figure 15, first, the relay communication station 110 determines and sets the antenna configuration of the actual SL antenna 111 and the number of cells 20C to be formed in the service area 20A (S101). For example, as the antenna configuration of the SL antenna 111, the antenna configuration of a cylindrical array antenna 111 having the appearance shown in Figure 16(a) and the antenna model shown in Figure 16(b) is set. The antenna configuration may be a flat-plate-shaped array antenna 111 as shown in Figure 16(c), or a hemispherical-shaped array antenna 111 as shown in Figure 16(d). Also, the number of cells is set to, for example, 12 cells (see, for example, Figure 2(c)).

[0084] Next, the relay communication station 110 sets the initial value of the antenna weight w to be applied to the multiple antenna elements 111e of the SL antenna 111 having the above antenna configuration (S102). As the initial value of the antenna weight w, it is possible to set the result of the two-stage optimization according to the above reference example as is. For example, as shown in Figure 17, the cell configuration is optimized so that there are 3 cells in the Layer 1 area inside the service area 20A and 9 cells in the Layer 2 area outside. The initial value of the antenna weight w is set so that forming control is performed to form multiple beams 20B from the 12 cells of that cell configuration. For example, the initial value of the antenna weight w is set so that forming control is performed to form multiple beams 20B having the vertical beam pattern C151 shown in Figure 18(a) and the horizontal beam pattern C161 shown in Figure 18(b).

[0085] Next, the relay station 110 uses the initial value of the antenna weight w described above to directly optimize the antenna weight W using the gradient descent method (S103).

[0086] The remote control device 95 may extend the direct optimization of the antenna weight W using the gradient descent method to perform dynamic antenna weight optimization, as will be described in detail later (S104-S106). For example, the remote control device 95 may generate a dataset (training data) of attitude angles and optimal antenna weight w (S104), perform machine learning on an AI model (neural network (NN)) using this dataset (training data) (S105), and then use the trained AI model (neural network (NN)) to perform dynamic antenna weight optimization when the attitude of the communication relay device (HAPS) 10 changes (S106).

[0087] In the direct optimization of the antenna weight w described above, the computation flow for obtaining a predetermined objective function (e.g., frequency utilization efficiency c) from the antenna weight w is described using tensor computation, and gradient descent can be performed quickly and on a large scale by automatic differentiation that makes maximum use of GPU (Graphics Processing Unit) acceleration (GPGPU).

[0088] In one example of direct optimization of the antenna weight W in this embodiment, a loss function E1 is defined using the frequency utilization efficiency c, which is a function of the antenna weight w, as shown in the following set of mathematical formulas, and a calculation is performed to optimize the antenna weight w that minimizes this loss function E1.

number

[0089] In the above set of mathematical formulas,

number

[0090] Tables 1 to 3 list the vectors, variables, and constants used in the above calculations.

[0091] [Table 1]

[0092] [Table 2]

[0093] [Table 3]

[0094] In the calculation of this embodiment, as shown in Figure 19(a), the antenna weight w is optimized by repeatedly performing the calculation to minimize the loss function E1 using backpropagation. For example, Figure 19(b) shows the cell configuration (SINR heatmap) when using the initial value of the antenna weight calculated by the two-stage optimization of the reference example described above. On the other hand, Figure 19(c) shows the cell configuration (SINR heatmap) when using the optimized antenna weight calculated by the iterative calculation using backpropagation of this embodiment.

[0095] The matrices and vectors used in the calculation to minimize the above loss function E1 can be defined, for example, using the system model (mathematical model) including HAPS10 in Figure 20 and the configuration model of the transmitter section of HAPS10 in Figure 21, as shown in equations (5) to (10) below. Note that in Figure 20, information on the attitude of HAPS10 (e.g., yaw angle, pitch angle, roll angle) can be detected and acquired, for example, by a gyro sensor (gyroscope) 120 mounted on HAPS10.

[0096]

number

number

number

number

number

number

[0097] Here, the element a of the cell connection matrix A above contains the non-differentiable operator argmax, but it may be approximated by the differentiable operator softmax as shown in equation (11). With this approximation,

number

[0098] Through the above approximation, the cell connection matrix A can be expressed using a differentiable operation as shown in equation (9)'. As a result, all the matrices and vectors in equations (6) to (10) can be expressed using differentiable operations, and the calculation flow for determining the frequency utilization efficiency c from the antenna weight W can be described using tensor calculations.

number

[0099] As a result, all operations on the matrices and vectors described above, which are forward algorithms, become combinations of basic tensor operations (tensor product, scalar multiplication, component-wise function application, tensor addition and subtraction). For example, the penalty function E of the transmit power limit is a function of the antenna weight W. Pt (W) and the loss function E1(W) can be calculated using the following differentiable equations (12) and (13).

number

number

[0100] This enables automatic differentiation with GPGPU acceleration using neural network (NN) frameworks such as PyTorch, TensorFlow, and JAX (see reference: T. Wadayama and L. Wei, "Gradient flow decoding," IEEE Access, vol. 13, pp. 131937-131956, 2025). Furthermore, the antenna weight W can be calculated by performing gradient descent to minimize the value of the loss function E1(W), as shown in equation (14).

number

[0101] Figure 22 shows an example of iterative calculations in the antenna weight optimization process (S200) according to the embodiment. In the example in Figure 22, the calculation of the loss function E1 and the calculation to minimize the value of the loss function E1 (gradient descent) are repeatedly performed a predetermined number of times (M times) using the backpropagation method. In the illustrated example, the calculation process (S210) including the calculation of the frequency utilization efficiency vector c and the loss function E1 (S211) and the calculation and update of the antenna weight W using the backpropagation method (S212) is repeatedly performed about 100 times.

[0102] Furthermore, the distribution of terminal devices (UEs) 61 in each cell 20C (distribution of users) may be set or information thereof may be acquired (S220), and for each of the multiple distribution patterns in which the user distributions differ from one another, the calculation of the loss function and the calculation to minimize the value of the loss function may be repeated a predetermined number of times (M times) (approximately 100 loops in the illustrated example), and then repeated a further predetermined number of times (N times). For example, in the illustrated example, for each of the multiple distribution patterns, the calculation process (S210) which includes the calculation of the frequency utilization efficiency vector c and the loss function E1 (S211) and the calculation and update of the antenna weight W by backpropagation (S212) is repeated a predetermined number of times (N times), and then repeated approximately 10,000 times.

[0103] In embodiments of this disclosure, the antenna weight W may be optimized to accommodate the attitude changes and footprint fixing of the HAPS10 aircraft, as shown below. In this case, the optimization of the antenna weight W includes generating training data by warm-starting the HAPS100 (S104 in Figure 15 above), machine learning an AI model (NN) using the training data (S105 in Figure 15 above), and dynamic optimization and updating of the antenna weight W using the trained AI model (NN) (S106 in Figure 15 above).

[0104] In generating training data, the antenna weight W is first optimized when the HAPS10 aircraft is in a state with no attitude variation (yaw, pitch, roll) = (0, 0, 0). This optimization of the antenna weight W applies the direct weight optimization method using gradient descent described above. The optimized antenna weight W is then set as the initial value.

[0105] Next, the optimization of the antenna weight W is continued while slowly inflicting attitude changes on the HAPS10 aircraft (warm start). In this warm start optimization of the antenna weight W, the direct weight optimization using the gradient descent method described above is also applied, but in order to accommodate footprint fixing, the loss function E2 calculated by equations (15) and (16) is used.

number

number

[0106] In equation (15)

number

number

[0107] Then, by performing a differential operation using gradient descent to minimize the value of the loss function E2(W) as shown in equation (17), the antenna weight W corresponding to footprint fixation when there are attitude fluctuations of the HAPS10 aircraft can be calculated. Hereafter, the optimization of the antenna weight W corresponding to footprint fixation will also be called "FP fixed optimization".

number

[0108] Next, the attitude of the HAPS10 aircraft is changed to cover the three-dimensional attitude angle grid 300 in Figure 23(a), and the yaw angle, pitch angle, and roll angle of the SL antenna 111 are swept as shown in SL antennas 111(1) to 111(3) in Figure 23(b). Here, if the front-to-back direction of the HPAS10 (SL antenna 111) (forward direction is positive) is the X-axis, the left-to-right direction of the HPAS10 (SL antenna 111) (right direction is positive) is the Y-axis, and the up-and-down direction of the HAPS10 (SL antenna 111) (down direction is positive) is the Z-axis, then the yaw angle is the angle that indicates yawing (symmetrical movement) around the Z-axis, the pitch angle is the angle that indicates pitching (vertical movement) around the Y-axis, and the roll angle is the angle that indicates rolling (lateral movement) around the X-axis.

[0109] By performing a warm start from the HAPS10's reference attitude and changing the HAPS10's aircraft attitude, the attitude angles (yaw angle, pitch angle, roll angle) of the SL antenna 111 are swept while FP fixed optimization is continued, generating a dataset of optimal antenna weights labeled by attitude angles (yaw angle, pitch angle, roll angle). This dataset is used as a training dataset for machine learning of an AI model (NN).

[0110] For example, in yaw angle sweeping, the angle of the yaw angle axis (the vertical axis in the figure) in the 3D attitude angle grid 310 in Figure 24(a) is changed by 1 degree downwards (positive direction) from the origin, as shown in Figure 24(b), and FP fixed optimization is repeatedly performed to calculate the antenna weight W. Here, the learning rate per degree of yaw angle is 1 × 10⁻⁶ -3 Therefore, the number of users (terminal devices) is 1 × 10 4 The number of trials was 2,000.

[0111] Furthermore, in pitch angle sweeping, the angle of the pitch angle axis (the axis from the front left to the back right in the figure) in the 3D attitude angle grid 320 shown in Figure 25(a) is changed by 1 degree at a time from the origin towards the back right (negative direction), as shown in Figure 25(b), and FP fixed optimization is repeatedly performed to calculate the antenna weight W. Here, the learning rate per degree of pitch angle is 2 × 10⁻⁶ -3 Therefore, the number of users (terminal devices) is 1 × 10 4 The number of trials was 1,000.

[0112] Furthermore, in roll angle sweeping, the angle of the roll angle axis (the left-right axis in the figure) in the 3D attitude angle grid 330 shown in Figure 26(a) is changed by 1 degree at a time from the origin to the right (positive direction), as shown in Figure 26(b), and FP fixed optimization is repeatedly performed to calculate the antenna weight W. Here, the learning rate per degree of roll angle is 5 × 10⁻¹⁰ -3 Therefore, the number of users (terminal devices) is 1 × 10 4 The number of trials was 200.

[0113] Figure 27 shows an example of the machine learning configuration of the neural network (AI model) 400 according to the embodiment. The dataset of optimal antenna weights labeled by attitude angles (yaw angle, pitch angle, roll angle) generated by sweeping the yaw angle, pitch angle, and roll angle of the HAPS10 aircraft (SL antenna 111) is stored in the database of the memory unit 1102 as a training dataset for machine learning of the AI ​​model (NN) 400. In machine learning of the AI ​​model (NN) 400, the datasets are read one by one from the memory unit 1102 and set in the input layer 401 and output layer 402 of the AI ​​model (NN) 400 to perform machine learning. For example, the input layer 401 of the AI ​​model (NN) 400 is set with the sine function values ​​and cosine function values ​​of the attitude angles (yaw angle, pitch angle, roll angle) of the dataset. The output layer 402 of the AI ​​model (NN) 400 is set with the real and imaginary parts of the antenna weight W of the dataset.

[0114] Next, we will describe the results of a computer simulation of the optimization of the antenna weight W during changes in the attitude of the HAPS10 aircraft using the machine learning-prepared neural network (AI model) 400 of this embodiment (hereinafter also referred to as the "proposed method"). The computer simulation was performed for the following three scenarios 1, 2, and 3 in the system model shown in Figure 20 above.

[0115] Scenario 1: The test was conducted under the following conditions: HAPS10 had no attitude changes, HAPS10 was located at the origin (0,0), and users (terminal devices) were uniformly located within a 500m x 500m grid service area. Scenario 2: HAPS10 experiences random attitude changes within a predetermined range of yaw angle ψ, pitch angle θ, and roll angle φ (-180°≦ψ≦+180°, -10°≦θ≦+10°, -10°≦φ≦+10°), HAPS10 is located 20km above the origin (0,0), and the number of trials is 10 5 The user (terminal device) was resampled uniformly for each trial. Scenario 3: Using log data from a field test in a real environment, the HAPS10 aircraft circled within a radius of approximately 2.5 km, the HAPS10's altitude was fixed at 20 km (dummy data), and the user (terminal device) was positioned on a 500 m grid.

[0116] Tables 4, 5, and 6 are lists of the parameters used in the computer simulation. Note that "Section IV-A" in Table 5 refers to parameters specific to the optimization of antenna weight W using loss function E1, which does not support the aforementioned fixed FP configuration. "Section IV-B" refers to parameters specific to the optimization of antenna weight W using loss function E2, which does support the aforementioned fixed FP configuration.

[0117] [Table 4]

[0118] [Table 5]

[0119] [Table 6]

[0120] [Simulation results for Scenario 1] Figure 28(a) shows the results of a computer simulation of the SINR distribution of the two-stage optimization method (hereinafter referred to as the "conventional method") described in the reference example above. Figure 28(b) shows an example of the results of a computer simulation of the SINR distribution when the antenna weights are optimized using a machine learning-prepared neural network (AI model) in Scenario 1 to which the proposed method of the embodiment is applied. Compared to the SINR distribution (heatmap) of the conventional method, the SINR distribution (heatmap) of the proposed method achieves a higher SINR across the entire service area 20A by optimizing the antenna weights W, and the shape of the cell configuration (cell pattern) is also improved. In particular, the inner first layer cells 20C(1) to 20C(3) extend into the boundary of the adjacent outer second layer cells 20C(4) to 20C(12).

[0121] Figures 29 and 30 show examples of computer simulation results (curves C172 and C182 in the figures) of the CDF (Cumulative Distribution Function) curves for SINR and frequency utilization efficiency when the antenna weight is optimized using a machine learning-prepared neural network (AI model) in Scenario 1, applying the proposed method of the embodiment. Curves C171 and C181 in the figures are the results of computer simulations using the conventional method. According to the simulation results of the embodiment, both SINR and frequency utilization efficiency are superior to the conventional method. This advantage is obtained by directly optimizing the antenna weight W of the SL antenna (array antenna) 111, and contributes to the realization of beamforming and cell-specific transmission power distribution that takes the SL antenna (array antenna) 111 into consideration. Furthermore, the median frequency utilization efficiency in Figure 30 is 3.45 [bps / Hz] for the conventional method and 4.19 [bps / Hz] for the embodiment, showing a 21.4% improvement in the embodiment.

[0122] [Simulation results for Scenario 2] Figures 31 and 32 show examples of computer simulation results for the CDF curve of the footprint fixation rate and frequency utilization efficiency percentile when the antenna weights are optimized using a machine learning-prepared neural network (AI model) in Scenario 2 of the embodiment, respectively. The footprint fixation rate in Figure 31 is 10 5These are the results from conducting the trial. Curves C191, C201, and C202 in the figure represent the results of computer simulations using the conventional method. According to the simulation results of the embodiment, as shown by the steeply rising CDF in Figure 31, the footprint fixation rate (C192) clearly surpasses that of the conventional method (C191), and the footprint fixation rate rarely fell below 99% during random attitude fluctuations. Furthermore, regarding the frequency utilization efficiency percentiles in Figure 32, the 5th and 50th percentiles of the frequency utilization efficiency for each user were calculated in each trial, and their CDFs were plotted over all trials. The nearly vertical curves C201 to C204 indicate that both the embodiment and the conventional method maintain a nearly constant frequency utilization efficiency regardless of attitude. However, the embodiment consistently achieves higher values.

[0123] [Simulation results for Scenario 3] Figure 33(a) shows an example of measured time variation of the yaw angle of the HAPS 10 in a computer simulation of Scenario 3 of the embodiment. Figure 33(b) shows an example of measured time variation of the pitch angle and roll angle of the HAPS 10 in a computer simulation of the same Scenario 3. According to the measured example of attitude change of the HAPS 10, the yaw angle (curve C211 in Figure 33(a)) shows a smooth variation (sweep), while the pitch angle and roll angle (curves C212 and C213 in Figure 33(b)) show slight variations (sweeps) due to wind disturbances.

[0124] Figures 34 and 35 show examples of computer simulation results for the time variation of the footprint fixation rate and frequency utilization efficiency percentile when antenna weight optimization is performed using a machine learning-prepared neural network (AI model) in Scenario 3 of the embodiment, respectively. Curves C211, C231, and C232 in the figures represent the results of computer simulations using the conventional method. According to the results in Figure 34, the embodiment (curve C222) consistently achieves a high footprint fixation rate of over 95% over almost the entire section. Thus, even though the embodiment only corrects for attitude changes and does not consider the movement of HAPS10, the impact on the handover of HTA (Heavier Than Air) type HAPS that stays in a circular flight path with a radius of several kilometers is minimal. Furthermore, according to the results in Figure 35, the embodiment (curves C233, C234) also consistently achieves high frequency utilization efficiency throughout. At the 50th percentile, the embodiment is 21.5% higher on average than the conventional method.

[0125] Figures 36(a) and 36(b) show examples of computer simulation results of the SINR heatmap for a reference example and embodiment (scenario 3) when the communication relay device (HAPS) 10 is orbiting in the air and footprint fixing is not performed, respectively. Figures 36(c) and 36(d) show examples of computer simulation results of the SINR heatmap for a reference example and embodiment (scenario 3) when the communication relay device (HAPS) 10 is orbiting in the air and footprint fixing is performed, respectively. When the HAPS 10 orbits in the air, the SL antenna 111 tilts and its attitude changes as shown in Figure 37, and the yaw angle, pitch angle, and roll angle of the HAPS 10 (SL antenna 111) vary over time as shown by curves C211, C212, and C213 in Figures 33(a) and 33(b) above. If footprint fixing is not performed when the attitude of HAPS10 changes during rotation, the positions of the multiple cells formed in the service area 20A will change, as shown in Figures 36(a) and 36(b) in both the reference example and the embodiment. If footprint fixing is performed, the multiple cells will be formed in predetermined positions as shown in Figure 36(c) in the reference example, but the overall SINR will be lower across the entire service area 20A. On the other hand, if the antenna weights are optimized using a machine learning-prepared neural network (AI model) in Scenario 3 of the embodiment, a higher SINR can be achieved across the entire service area 20A even when HAPS10 is rotating in the air, and the shape of the cell configuration (cell pattern) can also be improved.

[0126] Furthermore, in this embodiment (Scenario 3), as shown in Figure 34 above, the footprint fixation rate (C222) exceeded that of the conventional method (C221), and the footprint fixation rate rarely fell below 95% during attitude fluctuations when HAPS10 was turning.

[0127] Furthermore, in this embodiment (Scenario 3), as shown by curves C233 and C234 in Figure 35, a consistently high frequency utilization efficiency was achieved throughout the entire process.

[0128] Figure 38 shows an example of an extension to a user distribution-aware neural network (AI model) in the communication relay device (HAPS) 10 according to the embodiment. In the extension example of Figure 38 (S300), a user density matrix corresponding to the distribution of users (UE) 61 in the service area 20A is created (S310), and the hypernetwork structure has a machine learning function that determines the weights Ω of a CNN (convolutional neural network) having convolutional layers and pooling layers using the user density matrix, the aircraft's attitude angle information, and corresponding antenna weight training data (S320). The CNN inference result outputs a set of NN (neural network) weights used for optimizing the antenna weights W with fixed FP. This makes it possible to dynamically create a lightweight NN with low computational cost for the optimization of antenna weights W with fixed FP (S340), which is executed repeatedly every few milliseconds, according to the user distribution.

[0129] Figure 39 shows another example of the extension of the communication relay device (HAPS) 10 according to the embodiment to a user distribution-considering neural network (AI model). The extension example in Figure 39 is an example of extension to a two-stage NN. In the extension example in Figure 39 (S400), a user density matrix corresponding to the distribution of users (UE) 61 in the service area 20A is created (S410). By applying this user density matrix to a CNN (convolutional neural network) that performs the first stage optimization when there is no attitude change of HAPS 10, the antenna weight W when there is no attitude change of HAPS 10, taking into account the distribution of users (UE) 61, is obtained. nominal Generates (S420). Antenna weight W using this CNN. nominal The generation of (S420) is computationally intensive, so it is performed at longer time intervals (for example, once every few tens of minutes). Next, the antenna weight W is calculated when there is no attitude change output from the CNN. nominalThe sine and cosine values ​​of the HAPS10 attitude angle (yaw angle, pitch angle, roll angle) vectors are input to a lightweight NN that performs a second-stage optimization process for FP fixing, generating antenna weights W with FP fixing (S440). This generation of antenna weights W with FP fixing by the second-stage NN (S440) requires less computation than the first-stage CNN, so it is executed at shorter time intervals (for example, once every few milliseconds). The results of the optimization of antenna weights W with FP fixing, which is repeatedly performed by the lightweight NN (S440), are returned to the CNN by backpropagation (S450) and used for CNN machine learning.

[0130] In the embodiments of this disclosure, the remote control device 95 may also include the aforementioned storage unit 114 and antenna weight calculation unit (calculation unit) 115. In the remote control device 95, the storage unit stores initial setting information for one or more HAPS 10, including information on the number of cells 20C to be formed in the service area 20A and information on the antenna configuration of the SL antenna 111. The antenna weight calculation unit optimizes the antenna weight W for one or more HAPS 10 by performing a calculation to minimize the value of the loss function, which is a function of the antenna weight W applied to the SL antenna 111, in a mathematical model of service link communication set based on the information on the number of cells and the antenna configuration information.

[0131] Furthermore, the remote control device 95 may include a training data generation unit that generates multiple sets of training data consisting of a combination of information on the position, attitude, or both of the HAPS 10 and the optimized antenna weight, for use in machine learning of an AI model to dynamically estimate the optimized antenna weight W for multiple conditions in which the position, attitude, or both of the HAPS 10 differ for one or more HAPS 10s, and a learning unit that performs machine learning by applying the multiple sets of training data to the AI ​​model and creates a trained AI model.

[0132] In the remote control device 95, the AI ​​model may be a neural network (NN).

[0133] In the remote control device 95, the loss function may include a correction term for fixing the footprint of cell 20C when the position, attitude, or both of the HAPS 10 aircraft change.

[0134] In the remote control device 95, the loss function may be a function of the frequency utilization efficiency vector. Here, the frequency utilization efficiency vector may be a function of the cell connection matrix, which represents the cells to which each of the multiple terminal devices (UEs) 61 is connected.

[0135] In the remote control device 95, if the cell connection matrix includes the non-differentiable operator argmax, the antenna weight calculation unit (calculation unit) may approximate the non-differentiable operator argmax with the differentiable operator softmax.

[0136] In the remote control device 95, the calculation to minimize the value of the loss function may be an optimization calculation that calculates the gradient of the loss function and updates the antenna weight W by gradient descent based on the gradient, repeating this process a predetermined number of times.

[0137] Next, other embodiments of this disclosure will be described.

[0138] Figure 40 shows an example of the movement of a communication relay device (HAPS) 10 according to an embodiment. As shown in Figure 40, the communication relay device (HAPS) 10, which is an aerial platform that does not remain stationary, changes the service area 20A of the communication service provided as the aircraft moves. HAPS10' and service area 20A' in the figure show the state after movement. Service areas 20A and 20A' include areas with different characteristics, such as rural and urban areas. Furthermore, in order to design an appropriate area considering the geographical characteristics and population distribution within the service area that change moment by moment, the use of area optimization technology is effective.

[0139] As described above, HAPS10, an aerial platform that does not remain stationary, makes it fundamentally impossible to completely fix the footprint of each cell relative to the ground. Furthermore, as the aircraft moves, the geographical features and population distribution within the service area change dynamically, so the optimal cell shape also changes in real time. Therefore, for example, the following C1 and C2 challenges exist. C1. Performing cell optimization at a high frequency within a short period of time makes it difficult to provide a stable communication service because the cell shape and serving cells change in real time. C2. When the number of cells formed in the service area is large, the number of parameters that need to be optimized becomes enormous, making real-time optimization difficult.

[0140] Conventional area optimization technologies assume a stationary aerial platform, meaning the geographical features of the service area as seen from the aircraft do not change over time. Furthermore, footprint fixing technology cannot be applied to address the aforementioned challenges.

[0141] In other embodiments of this disclosure, under the above background, stable communication is achieved even when the aircraft is moving by continuously performing area optimization at a high frequency. Furthermore, the service area is continuously optimized by gradually transitioning the cells formed on the ground.

[0142] Figure 41 shows an example of area optimization when a communication relay device (HAPS) 10 according to the embodiment moves. As shown in Figure 41, when moving from HAPS 10 to HAPS 10', cell transition beamforming control is performed at a high frequency for minute-time movements of HAPS 10. Specifically, for example, as shown in Figure 41, when transitioning from cell 20C in service area 20A before movement to cell 20C' in service area 20A' after movement, a cell transition of a certain area determined according to the amount of movement of the HAPS 10 is allowed. For example, for cell transitions from multiple cells 20C(1) to 20C(7) in service area 20A before movement, shown by the dashed line in Figure 42, to multiple cells 20C(1) to 20C(7) in service area 20A' after movement, shown by the solid line in Figure 42, a cell transition of a certain area determined according to the amount of movement of the HAPS 10 is allowed. Then, a small amount of area optimization is performed within the range of the above allowable amount.

[0143] Furthermore, as shown in Figure 41, where the movement from HAPS10' to HAPS10'' is observed, for long-term, large-scale movements of HAPS10, the area is optimized by gradually transitioning between cells, as shown in cell 20'' of service area 20'' in the figure.

[0144] Figure 43 shows an example of a cell configuration formed in a service area 20A by a HAPS (High-Area Communication System) 10 to which area optimization according to other embodiments of the present disclosure can be applied. In Figure 43, in the left-hand area of ​​the service area 20A, such as urban areas, there are many UE61s, so multiple cells 20C (9 cells in the illustrated example) of the service link SL are formed. On the other hand, in the right-hand area of ​​the service area 20A, such as mountainous areas and suburban areas, there are few UE61s, so a single cell 20C of the service link SL is formed.

[0145] Figure 44 is a block diagram showing an example of the main components of a relay station 110 of a communication relay device (HAPS) 10 according to another embodiment of the present disclosure. In Figure 44, components common to Figure 14 are denoted by the same reference numerals and their descriptions are omitted. In Figure 44, the relay station 110 comprises an SL antenna 111, an information acquisition unit 113, a storage unit 114, an antenna weight calculation unit (calculation unit) 115, and a control unit 116.

[0146] The information acquisition unit 113 includes a position / attitude / time information acquisition unit 1131 and a ground user distribution information acquisition unit 1132. The position / attitude / time information acquisition unit 1131 acquires position and attitude information and time information of the HAPS (communication relay device) 10. The ground user distribution information acquisition unit 1132 acquires information on the distribution (user distribution) of terminal devices (UEs) 61 in the service area 20A.

[0147] The information acquisition unit 113 may include a terrain / radio wave propagation information acquisition unit 1133 that acquires terrain and radio wave propagation information of the service area 200A, and a ground relay station / adjacent PF information acquisition unit 1134 that acquires information about ground base stations located in the service area 200A and information about adjacent platforms such as other HAPS located near HAPS 10.

[0148] The memory unit 114 stores information about the transition tolerance of cell 20C, which changes according to the amount of movement of the HAPS 10 aircraft.

[0149] The antenna weight calculation unit 115 functions as an optimization processing unit that intermittently executes a process to optimize the service area 20A in which cell 20C is formed, within the range of the transition tolerance of cell 20C.

[0150] The antenna weight calculation unit (calculation unit) 115 calculates an antenna weight W that optimizes the service area 20A within a range that satisfies the transition tolerance of cell 20C, by performing a calculation to minimize the value of the loss function that reflects the transition tolerance information of cell 20C in a service link communication model (mathematical model) set based on the information of the number of cells 20C in the service area 20A and the antenna configuration information of the service link array antenna (SL antenna) 111, and may update the antenna weight W intermittently according to a predetermined period or predetermined conditions.

[0151] The operation to minimize the value of the loss function may be an optimization operation in which the process of calculating the gradient of the loss function and updating the antenna weight W by gradient descent based on the gradient is repeated a predetermined number of times.

[0152] The control unit 116 controls the directional beam of the SL antenna 111 by applying optimized antenna weights to multiple antenna elements 111e of the SL antenna 111.

[0153] The antenna weight calculation unit (calculation unit) 115 may dynamically optimize the antenna weight W by dynamically performing calculations to minimize the value of the loss function using a trained AI model. In this case, the control unit 116 controls the directional beam of the SL antenna 111 by applying the dynamically optimized antenna weight W to the multiple antenna elements 111e of the SL antenna 111.

[0154] Figure 45 is a flowchart showing an example of area optimization processing (S500) in a relay communication station 110 of a communication relay device (HAPS) 10 according to another embodiment of the present disclosure. In Figure 45, the area optimization processing (S500) includes information input processing (S510) and antenna weight calculation processing (S520).

[0155] In the information input process (S510), the next cell connection matrix, the in-area user ID vector, and the antenna weight information for time t-Δt, which is Δt [seconds] before the current time t, are read from the storage unit 114 and input to the antenna weight calculation unit 115. Optionally, user distribution information for time t-Δt may also be input.

number

[0156] Furthermore, in the information input process (S510), the aircraft position, aircraft attitude, user distribution information, and the in-area user ID vector information for the current time t are read from the storage unit 114 and input to the antenna weight calculation unit 115.

number

[0157] In the antenna weight calculation process (S520), the antenna weight W is calculated using, for example, the following method (A) or (B). (A) Gradient descent method (B) A method based on an AI model (e.g., a neural network (NN)) that has been trained using the results of the method in (A) above.

[0158] For example, in the antenna weight calculation process (S520), the antenna weight Wt at the current time t that minimizes the loss function E given by equation (18) is calculated. Here, f capacity (W) is an indicator for maximizing communication capacity in service area 20A. coverage (W) is a penalty function to prevent the coverage of service area 20A by cell 20C from falling below the required amount. power (W) is a penalty function to prevent the transmission power or power consumption in the service link communication of the relay communication station 110 from exceeding a standard value. shift(W) is a penalty function to prevent fast cell transitions. λ1, λ2, λ3, and λ4 are coefficients that show the influence of each indicator and the penalty function.

number

[0159] In the area optimization process (S500) shown in Figure 45, the information input process (S510) and the antenna weight calculation process (S520) are repeatedly executed at predetermined time intervals (Δt).

[0160] [Method using gradient descent] In the gradient descent method described above, the matrices and vectors used in the calculation to minimize the loss function E can be defined, for example, using the system model (mathematical model) including HAPS10 in Figures 46 and 47 and the configuration model of the HAPS10 transmitter in Figure 48, as shown in equations (19) to (24) below. The system model (mathematical model) in Figure 46 is a model that assumes the distribution of real users (UEs) 61. In Figure 46, the distribution (user distribution) of real users (UEs) 61 located in multiple 20C areas of the service area 20A may be obtained by some method, or it may be an estimated distribution. The system model (mathematical model) in Figure 47 places uniformly distributed virtual users 61V for measuring spatial characteristics on a grid defined at regular intervals in the service area 20A. Figure 49 shows the definition with respect to time t, and Figure 50 shows the cell transitions in a predetermined time interval (Δt). The vectors, variables, and constants used in calculating the loss function E are the same as those in Tables 1 to 3 above.

[0161]

number

number

number

number

number

number

[0162] Here, the desired power vector μ does not necessarily have to be assumed to be a full buffer. Also, although element a of the cell connection matrix A above contains the non-differentiable operator argmax, it may be approximated by the differentiable operator softmax as shown in equation (25).

number

[0163] Through the above approximation, the cell connection matrix A can be expressed using differentiable operations as shown in equation (23)'. As a result, all the matrices and vectors in equations (19) to (24) can be expressed using differentiable operations, and the calculation flow for determining the frequency utilization efficiency c from the antenna weight W can be described using tensor calculations.

number

[0164] As a result, all operations on the matrices and vectors described above are combinations of the basic tensor operations (tensor product, scalar multiplication, component-wise function application, tensor addition and subtraction).

[0165] Using the above matrix and vector, the loss function E used to calculate the antenna weight Wt can be expressed as shown in equation (26).

number

[0166] The above penalty function E shift(W) included in

Number

Number

Number

[0167] By using the loss function E with the above matrices and vectors, automatic differentiation with GPGPU acceleration using NN (neural network) frameworks such as PyTorch, TensorFlow, and JAX becomes possible.

[0168] Also, as shown in the following equation (27), by performing a differential operation using the gradient descent method to minimize the value of the loss function E(W), the antenna weight W can be calculated. In particular, by using W t-Δt as the initial value, the calculated value converges in a few trials, enabling rapid and frequent area optimization.

Number

Number

[0170] There are various variations in the implementation of the input and output combinations of the NN 410. For example, in the example of FIG. 51, the output of the NN 410 may be the antenna weight W, and the input may be the following four types.

Number

[0171] In the example of FIG. 51, for example, in the following, the antenna W at time t - Δt t-Δt is used as the input to obtain the displacement of the aircraft

Number

Number

[0172] In other embodiments of this disclosure, the remote control device 95 may also include the aforementioned storage unit 114 and antenna weight calculation unit (calculation unit) 115. In the remote control device 95, the storage unit stores initial setting information including information on the number of cells 200C to be formed in the service area 20A and information on the antenna configuration of the SL111 of the HAPS10, as well as information on the transition tolerance of cells 20C which changes according to the amount of movement of the HAPS10. The antenna weight calculation unit (calculation unit) calculates an antenna weight W that optimizes the service area 20A within a range that satisfies the transition tolerance of cells 20C by performing a calculation to minimize the value of the loss function E that reflects the information on the transition tolerance of cells 20C in a mathematical model of service link communication set based on the information on the number of cells 20C and the antenna configuration information of the HAPS10, and updates the antenna weight W intermittently according to a predetermined period or predetermined conditions.

[0173] Furthermore, the remote control device 95 may include a training data generation unit that generates multiple sets of training data consisting of a combination of information on the position, attitude, or both of the HAPS10 aircraft and the optimized antenna weight W, for use in machine learning of an AI model to dynamically estimate the optimized antenna weight for multiple conditions in which the position, attitude, or both of the HAPS10 aircraft differ; and a learning unit that performs machine learning by applying the multiple sets of training data to the AI ​​model and creates a trained AI model. Here, the AI ​​model is, for example, a neural network (NN).

[0174] In the remote control device 95, the loss function E may be the function represented by equation (30) described above. Furthermore, the calculation to minimize the value of the loss function E may be an optimization calculation that calculates the gradient of the loss function E and updates the antenna weight W by gradient descent based on the gradient, repeating this process a predetermined number of times.

[0175] As described above, according to one embodiment of the present disclosure, it is possible to optimize the antenna weight W applied to the SL antenna 111 that forms a plurality of cells 20C in a service area 20A, extending from an overhead relay type communication relay device (HAPS) 10 to the ground or sea.

[0176] Furthermore, according to one embodiment of this disclosure, the cell configuration of multiple cells 20C in the service area 20A and the antenna weight W applied to the SL antenna 111 can be optimized simultaneously.

[0177] Furthermore, according to one embodiment of this disclosure, the antenna weight can be directly optimized using the gradient descent method.

[0178] Furthermore, according to one embodiment of this disclosure, dynamic antenna weight optimization can be achieved using a machine learning-based AI model (e.g., a neural network).

[0179] According to other embodiments of the present disclosure, area optimization can be performed to optimize the service area 20A intermittently (continuously) while smoothly transitioning cells 20C formed in the service area 20A toward the ground or sea from the overhead relay type communication relay device (HAPS) 10.

[0180] According to other embodiments of this disclosure, area optimization can be performed intermittently at a high frequency, enabling stable service link communication even when the HAPS10 unit moves.

[0181] According to other embodiments of the present disclosure, the antenna weight W applied to the SL antenna 111 forming a plurality of cells 20C is optimized.

[0182] The HAPS, remote control device, and system of the present disclosure can perform area optimization that optimizes the service area intermittently (continuously) while gently transitioning the cells formed in the service area from the high-altitude relay communication relay device (HAPS) toward the ground or sea, and thus can contribute to the achievement of Sustainable Development Goal (SDG) 9, "Build the infrastructure for industry and innovation."

[0183] As used herein, AI includes generative AI (generative system AI), generative AI, language models (LLM / SLM), GPT (registered trademark), Gemini (registered trademark), Claude (registered trademark), Llama (registered trademark), and other language models, and refers to artificial intelligence (including AGI or ASI) that generates content such as text, images, audio, video, etc. using deep learning (deep neural network) technologies such as transformers, self-attention, autoregressive networks, etc. Expansion technologies for generative AI include retrieval-augmented generation (RAG), memory-augmented generation, hybrid search using vector databases, chunking / chunk processing, knowledge graph linkage, entity linking, frameworks such as AutoGen, AOG, LangChain, etc. Performance improvement technologies for generative AI include fine-tuning by RLHF / RLAIF, PEFT, LoRA, etc., distillation, quantization, weight sharing, continuous learning, collaborative learning, in-context learning, etc. Also, the operating environment of generative AI can utilize any environment of on-premises / cloud / edge (on-device), and parallel / distributed learning and inference using GPUs, TPUs, NPUs, IPUs, ASICs, FPGAs are also possible.

[0184] Note that the processing steps described in this specification and the components such as the high-altitude stay-type communication relay device, remote control device, system, memory unit, arithmetic unit, teacher data generation unit, learning unit, neural network (NN), etc. can be implemented by various means. For example, these steps and components may be implemented by hardware, firmware, software, or a combination thereof.

[0185] With respect to hardware implementation, means such as processing units used to realize the above processes and components in a physical entity (e.g., various wireless communication devices, Node B, terminals, hard disk drive devices, or optical disc drive devices) may be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, electronic devices, other electronic units designed to perform the functions described herein, computers, or combinations thereof.

[0186] Furthermore, with respect to the firmware and / or software implementation, means such as processing units used to realize the above-mentioned components may be implemented in the form of a program (e.g., code such as procedures, functions, modules, instructions, etc.) that performs the functions described herein. Generally, any computer / processor-readable medium that clearly embodies the firmware and / or software code may be used to implement means such as processing units used to realize the above-mentioned processes and components as described herein. For example, the firmware and / or software code may be stored in memory in a control device, for example, and executed by a computer or processor. That memory may be implemented inside the computer or processor, or it may be implemented outside the processor. Also, the firmware and / or software code may be stored in a computer or processor-readable medium such as random access memory (RAM), read-only memory (ROM), non-volatile random access memory (NVRAM), programmable read-only memory (PROM), electrically erasable PROM (EEPROM), flash memory, floppy disks, compact disks (CDs), digital versatile disks (DVDs), magnetic or optical data storage devices, etc. The code may be executed by one or more computers or processors, and the computers or processors may be made to perform functional embodiments as described herein.

[0187] Furthermore, the medium may be a non-temporary recording medium. Also, the program code may be readable and executable by a computer, processor, or other device or machine, and its format is not limited to a specific format. For example, the program code may be source code, object code, or binary code, or it may be a mixture of two or more of these codes.

[0188] Furthermore, the descriptions of embodiments disclosed herein are provided to enable those skilled in the art to manufacture or use the disclosure. Various modifications to the disclosure will be readily apparent to those skilled in the art, and the general principles defined herein are applicable to other variations without departing from the spirit or scope of the disclosure. Therefore, the disclosure is not limited to the examples and designs described herein, but should be accepted in the broadest sense that conforms to the principles and novel features disclosed herein. [Explanation of symbols]

[0189] 10: HAPS (High-Aerial Relay System) 20A: Service Area 20C: Cell 61: UE (Terminal Device) 70: GW Station 71: Wireless relay device 80:Base station equipment 90: Mobile communication network (core network) 95: Remote control device 100: Aircraft 110: Relay communication station 111: Service link array antenna (SL antenna) 111e: Antenna elements 112: Antenna for feeder link (FL antenna) 113: Information acquisition department 114: Storage section 115: Antenna weight calculation unit 1151: Arithmetic unit 1152: Training Data Generation Unit 1153: Learning Department 116: Control Unit

Claims

1. An aerial relay type communication relay device that forms one or more cells toward a ground or sea service area via a service link array antenna of a relay communication station installed on an aircraft located in the air, and communicates wirelessly with multiple terminal devices located within the cells, A storage unit that stores information on the transition tolerance of the cell, which changes according to the amount of movement of the aircraft, A calculation unit intermittently performs a process at a predetermined interval or under predetermined conditions to ensure that a desired communication quality is obtained throughout the service area in which the cells are formed, within the range of the cell transition tolerance stored in the storage unit. A communication relay device equipped with the following features.

2. In the communication relay device according to claim 1, The aforementioned arithmetic unit, In a mathematical model of service link communication set based on the information of the number of cells and the antenna configuration information of the service link array antenna, an optimized antenna weight is determined so that the desired communication quality can be obtained throughout the service area within the range that satisfies the cell transition tolerance, by performing a calculation to minimize the value of the loss function that reflects the information of the cell transition tolerance, and the antenna weight is updated intermittently according to a predetermined period or predetermined conditions. Communication relay device.

3. In the communication relay device according to claim 2, Let W be the antenna weight, E be the loss function, and f be the index for maximizing the communication capacity in the service area. ｃａｐａｃｉｔｙ Let (W) be a penalty function E to prevent the coverage of the service area by the cell from falling below the required amount. ｃｏｖｅｒａｇｅ Let (W) be the value of the transmission power or power consumption in the service link communication of the relay station, and let E be the penalty function to prevent the transmission power or power consumption from exceeding a standard value. ｐｏｗｅｒ Let (W) be the penalty function to prevent fast cell transitions, and E be the penalty function to prevent fast cell transitions. ｓｈｉｆｔ Let (W) be the coefficient that indicates the influence rate, and let λ be the coefficient that indicates the influence rate. １ λ2, λ ３ , λ ４ In this case, the loss function E is expressed by the following equation (1): Communication relay device. [Math 1]

4. In the communication relay device according to claim 2, The operation to minimize the value of the loss function is an optimization operation that calculates the gradient of the loss function and updates the antenna weights by gradient descent based on the gradient, repeating this process a predetermined number of times. Communication relay device.

5. In the communication relay device of claim 3, The operation to minimize the value of the loss function is an optimization operation that calculates the gradient of the loss function and updates the antenna weights by gradient descent based on the gradient, repeating this process a predetermined number of times. Communication relay device.

6. In a communication relay device according to any one of claims 2 to 5, The service link array antenna is equipped with a control unit that controls the directional beam of the service link array antenna by applying the optimized antenna weights to a plurality of antenna elements of the service link array antenna. Communication relay device.

7. In the communication relay device according to claim 2 or 3, The calculation unit dynamically optimizes the antenna weight to minimize the value of the loss function by using a trained AI model that takes as input the displacement and attitude change of the aircraft from a predetermined time (Δt) before the current time (t) and the antenna weight at the time (t-Δt) prior to the predetermined time (Δt), estimates and outputs the optimized antenna weight at the current time (t), thereby dynamically optimizing the antenna weight so that the desired communication quality can be obtained throughout the service area. The service link array antenna is equipped with a control unit that controls the directional beam of the service link array antenna by applying the dynamically optimized antenna weights to a plurality of antenna elements of the service link array antenna. Communication relay device.

8. A remote control device capable of communicating with an aerial relay type communication relay device that forms one or more cells toward a ground or sea service area via a service link array antenna of a relay communication station installed on an aircraft located in the air, and wirelessly communicates with multiple terminal devices located within the cells, A storage unit that stores initial setting information including information on the number of cells formed in the service area and information on the antenna configuration of the service link array antenna, and information on the cell transition tolerance which changes according to the amount of movement of the aircraft, A calculation unit that, in a mathematical model of service link communication set based on the information of the number of cells and the information of the antenna configuration, performs a calculation to minimize the value of the loss function that reflects the information of the cell transition tolerance stored in the memory unit, thereby determining an optimized antenna weight so that the desired communication quality can be obtained throughout the entire service area within the range that satisfies the cell transition tolerance, and updates the antenna weight intermittently according to a predetermined period or predetermined conditions, A training data generation unit generates multiple sets of training data consisting of combinations of information on the position, attitude, or both of the aircraft and the optimized antenna weight, which are used in machine learning of an AI model for dynamically estimating the optimized antenna weight for multiple conditions in which the position, attitude, or both of the aircraft differ. A learning unit that performs machine learning by applying the multiple sets of training data to the AI ​​model and creates a trained AI model, A remote control device equipped with the following features.

9. In the remote control device of claim 8, The aforementioned AI model is a neural network. Remote control device.

10. In the remote control device of claim 8, Let the antenna weight be \(W\), the loss function be \(E\), and the index for maximizing the communication capacity in the service area be \(f\). ｃａｐａｃｉｔｙ Let it be \(E_{ ｃａｐａｃｉｔｙ}(W)\) as the penalty function for preventing the coverage of the service area by the cell from falling below the required amount. ｃｏｖｅｒａｇｅ Let it be \(E_{ ｃｏｖｅｒａｇｅ}(W)\) as the penalty function for preventing the transmission power or power consumption amount in the service link communication of the relay communication station from exceeding the reference value. ｐｏｗｅｒ Let it be \(E_{ ｐｏｗｅｒ}(W)\) as the penalty function for preventing fast cell transition. ｓｈｉｆｔ Let the coefficient indicating the influence rate be \(\lambda\). １ , \(\lambda_2\), \(\lambda\). ３ , \(\lambda\). ４ When this is the case, the loss function \(E\) is represented by the following formula (1). Remote control device. [Math 2]

11. In the remote control device of claim 10, The operation to minimize the value of the loss function is an optimization operation that calculates the gradient of the loss function and updates the antenna weights by gradient descent based on the gradient, repeating this process a predetermined number of times. Remote control device.

12. In the remote control device of claim 9, When the antenna weight is W, the loss function is E, the index for maximizing the communication capacity in the service area is fcapacity(W), the penalty function for preventing the coverage of the service area by the cell from falling below the required amount is Ecooverage(W), the penalty function for preventing the transmission power or power consumption in the service link communication of the relay station from exceeding a standard value is Epower(W), the penalty function for preventing high-speed cell transitions is Eshift(W), and the coefficients indicating the rate of influence are λ1, λ2, λ3, and λ4, the loss function E is expressed by the following equation (1): Remote control device. [Math 3]

13. In the remote control device of claim 12, The operation to minimize the value of the loss function is an optimization operation that calculates the gradient of the loss function and updates the antenna weights by gradient descent based on the gradient, repeating this process a predetermined number of times. Remote control device.

14. A system comprising a remote control device according to any one of claims 8 to 13 and a communication relay device according to claim 2.

15. In the system of claim 14, The calculation unit of the communication relay device dynamically optimizes the antenna weights so that the desired communication quality can be obtained throughout the service area by dynamically performing calculations to minimize the value of the loss function using the trained AI model transferred from the remote control device. The communication relay device includes a control unit that controls the directional beam of the service link array antenna by applying the dynamically optimized antenna weights to a plurality of antenna elements of the service link array antenna. system.

16. A method for obtaining desired communication quality throughout a ground or sea service area where one or more cells are formed via a service link array antenna of a relay communication station installed on an aircraft located in the air, The information of the cell transition tolerance, which changes according to the amount of movement of the aircraft, is stored in the storage unit. Within the range of the cell transition tolerance stored in the memory unit, a process is performed intermittently at predetermined intervals or under predetermined conditions to ensure that the desired communication quality is obtained throughout the service area where the cell is formed. Methods that include...

17. A method for obtaining desired communication quality throughout a ground or sea service area where one or more cells are formed via a service link array antenna of a relay communication station installed on an aircraft located in the air, The storage unit stores initial setting information including information on the number of cells formed in the service area and information on the antenna configuration of the service link array antenna, and information on the cell transition tolerance which changes according to the amount of movement of the aircraft. In a mathematical model of service link communication set based on the information of the number of cells and the information of the antenna configuration, a calculation is performed to minimize the value of the loss function that reflects the information of the cell transition tolerance stored in the memory unit, thereby determining an optimized antenna weight so that the desired communication quality can be obtained throughout the entire service area within the range that satisfies the cell transition tolerance, and the antenna weight is updated intermittently according to a predetermined period or predetermined conditions. Methods that include...

18. A program executed on a computer or processor installed in an airborne relay-type communication relay device that forms one or more cells toward a ground or sea service area via a service link array antenna of a relay communication station installed on an aircraft located in the air, and communicates wirelessly with multiple terminal devices located within the cells, A program code for storing information in a storage unit about the transition tolerance of the cell, which changes according to the amount of movement of the aircraft, A program code for intermittently executing, according to a predetermined period or predetermined conditions, a process to ensure that a desired communication quality is obtained throughout the service area where the cells are formed, within the range of the cell transition tolerance stored in the memory unit, A program that includes this.

19. A program executed on a computer or processor installed in an airborne relay-type communication relay device that forms one or more cells toward a ground or sea service area via a service link array antenna of a relay communication station installed on an aircraft located in the air, and communicates wirelessly with multiple terminal devices located within the cells, A program code for storing in a storage unit initial setting information including information on the number of cells formed in the service area and information on the antenna configuration of the service link array antenna, and information on the cell transition tolerance which changes according to the amount of movement of the aircraft, In a mathematical model of service link communication set based on the information of the number of cells and the information of the antenna configuration, a program code for determining an optimized antenna weight so that the desired communication quality can be obtained throughout the service area within a range that satisfies the cell transition tolerance, by performing a calculation to minimize the value of the loss function that reflects the information of the cell transition tolerance stored in the memory unit, and for intermittently updating the antenna weight according to a predetermined period or predetermined conditions, A program that includes this.

Images