Communication method and user device

Abstract

A communication method according to the present invention, in which a machine learning technology is applied to wireless communication between a user device and a base station in a mobile communication system, includes: a step in which one communication device of the user device and the base station transmits, to the other communication device of the user device and the base station, a notification indicating at least one of the fact that the one communication device has an unlearned model, the fact that the one communication device has a model being learned, and the fact that the one communication device has a learned model the inspection of which has been completed; and a step in which the one communication device receives a response corresponding to the notification from the other communication device.

Description

Technical Field

[0001] The present disclosure relates to a communication method and a user device used in a mobile communication system.

Background Art

[0002] In recent years, in 3GPP (Third Generation Partnership Project) (registered trademark; the same shall apply hereinafter), which is a standardization project for mobile communication systems, studies have been made to apply artificial intelligence (AI) technology, particularly machine learning (ML) technology, to the wireless communication (air interface) of mobile communication systems.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

[0004] The communication method according to the first aspect is a method of applying machine learning technology to wireless communication between a user device and a base station in a mobile communication system. The communication method includes a step in which one of the communication devices, i.e., the user device and the base station, transmits a notification indicating at least one of having an unlearned model, having a model under learning, and having a learned model for which inspection has been completed, to the other communication device, i.e., the user device and the base station; and a step in which the one communication device receives a response corresponding to the notification from the other communication device.

[0005] A second aspect of the communication method is a method for applying machine learning technology to wireless communication between a user device and a base station in a mobile communication system. The communication method includes the steps of: one of the communication devices, the user device and the base station, performing inference processing using a trained model obtained by training the model; the one communication device monitoring the performance of the trained model to determine the need to retrain the model; and the one communication device, in response to determining that retraining is necessary, transmitting a notification indicating the need for retraining to the other communication device, the user device and the base station.

[0006] A third aspect of the communication method is a method for applying machine learning technology to wireless communication between a user device and a base station in a mobile communication system. The communication method includes the steps of: one of the user device and the base station receiving configuration information from the other of the user device and the base station, which includes information indicating the time at which a dataset for monitoring the performance of a trained model is provided; and the one of the communication devices receiving the dataset from the other communication device at the time and performing monitoring processing to monitor the performance of the trained model using the dataset. [Brief explanation of the drawing]

[0007] [Figure 1] This diagram shows the configuration of a mobile communication system according to an embodiment. [Figure 2] This diagram shows the configuration of the UE (User Equipment) according to the embodiment. [Figure 3] This diagram shows the configuration of the gNB (base station) according to the embodiment. [Figure 4] This diagram shows the protocol stack configuration of the user plane wireless interface that handles data. [Figure 5] This diagram shows the protocol stack configuration of the wireless interface of the control plane that handles signaling (control signals). [Figure 6]This diagram shows the functional block configuration of AI / ML technology (machine learning technology) in a mobile communication system according to the embodiment. [Figure 7] This diagram shows an overview of the operation for each operation scenario according to the embodiment. [Figure 8] This is a diagram showing the first operation scenario according to the embodiment. [Figure 9] This figure shows a first example of reducing CSI-RS according to the embodiment. [Figure 10] This figure shows a second example of reducing CSI-RS according to the embodiment. [Figure 11] This is an operation flow diagram showing a first operation pattern related to the first operation scenario according to the embodiment. [Figure 12] This is an operation flow diagram showing a second operation pattern related to the first operation scenario according to the embodiment. [Figure 13] This is an operation flow diagram showing the third operation pattern related to the first operation scenario according to the embodiment. [Figure 14] This figure shows a second operation scenario according to the embodiment. [Figure 15] This is an operation flowchart showing an example of operation related to the second operation scenario according to the embodiment. [Figure 16] This figure shows a third operation scenario according to the embodiment. [Figure 17] This is an operation flowchart showing an example of operation related to the third operation scenario according to the embodiment. [Figure 18] This figure shows a first operation pattern relating to model transfer according to the embodiment. [Figure 19] This figure shows an example of a configuration message including a model and additional information according to the embodiment. [Figure 20] This figure shows a second operation pattern related to model transfer according to the embodiment. [Figure 21] This figure shows a third operation pattern related to model transfer according to the embodiment. [Figure 22] This figure shows an example of model management according to the embodiment. [Figure 23]This is a diagram showing the details of model management according to an embodiment. [Figure 24] This is a diagram showing the first operation pattern related to model monitoring according to an embodiment. [Figure 25] As an example of the second operation pattern related to model monitoring according to an embodiment, this is a diagram showing an example of applying AI / ML technology to CSI feedback. [Figure 26] As another example of the second operation pattern related to model monitoring according to an embodiment, this is a diagram showing an example of applying AI / ML technology to positioning. [Figure 27] This is a diagram showing CSI feedback and beam management to which AI / ML technology according to an embodiment is applied. [Figure 28] This is a diagram for explaining beam management to which AI / ML technology according to an embodiment is applied. [Figure 29] This is a diagram showing a specific example of CSI feedback to which AI / ML technology according to an embodiment is applied. [Figure 30] This is a diagram showing a specific example of beam management to which AI / ML technology according to an embodiment is applied.

Mode for Carrying Out the Invention

[0008] When attempting to apply machine learning technology to a mobile communication system, the technology regarding how to specifically utilize machine learning processing has not yet been established.

[0009] Therefore, an object of the present disclosure is to enable the utilization of machine learning processing in a mobile communication system.

[0010] A mobile communication system according to an embodiment will be described while referring to the drawings. In the description of the drawings, the same or similar parts are denoted by the same or similar reference numerals.

[0011] (Configuration of Mobile Communication System) First, the configuration of the mobile communication system according to the embodiment will be described. Figure 1 is a diagram showing the configuration of the mobile communication system 1 according to the embodiment. The mobile communication system 1 conforms to the 5th Generation System (5GS) of the 3GPP standard. In the following description, 5GS will be used as an example, but the mobile communication system may also have at least a portion of the LTE (Long Term Evolution) system applied to it. The mobile communication system may also have at least a portion of the 6th Generation (6G) system applied to it.

[0012] The mobile communication system 1 comprises User Equipment (UE) 100, a 5G radio access network (NG-RAN) 10, and a 5G core network (5GC) 20. Hereafter, NG-RAN 10 may be simply referred to as RAN 10, and 5GC 20 may be simply referred to as core network (CN) 20.

[0013] UE100 is a mobile wireless communication device. UE100 can be any device used by a user. For example, UE100 can be a mobile phone terminal (including a smartphone) or tablet terminal, a notebook PC, a communication module (including a communication card or chipset), a sensor or device attached to a sensor, a vehicle or device attached to a vehicle (Vehicle UE), or an aircraft or device attached to an aircraft (Aerial UE).

[0014] NG-RAN10 includes base stations (referred to as "gNBs" in 5G systems) 200. The gNBs 200 are interconnected via the Xn interface, which is an inter-base station interface. Each gNB 200 manages one or more cells. The gNB 200 performs wireless communication with UEs 100 that have established a connection with its own cell. The gNB 200 has radio resource management (RRM) functions, user data routing functions (hereinafter simply referred to as "data"), measurement and control functions for mobility control and scheduling, etc. "Cell" is used as a term to indicate the smallest unit of a wireless communication area. "Cell" is also used as a term to indicate a function or resource that performs wireless communication with the UE 100. One cell belongs to one carrier frequency (hereinafter simply referred to as "frequency").

[0015] Furthermore, gNBs can also connect to the EPC (Evolved Packet Core), which is the core network of LTE. LTE base stations can also connect to 5GCs. LTE base stations and gNBs can also be connected via an inter-base station interface.

[0016] The 5GC20 includes the AMF (Access and Mobility Management Function) and the UPF (User Plane Function) 300. The AMF performs various mobility controls for the UE100. The AMF manages the mobility of the UE100 by communicating with it using NAS (Non-Access Stratum) signaling. The UPF controls data transfer. The AMF and UPF are connected to the gNB200 via the NG interface, which is the base station-core network interface.

[0017] Figure 2 shows the configuration of UE100 (user device) according to an embodiment. UE100 comprises a receiving unit 110, a transmitting unit 120, and a control unit 130. The receiving unit 110 and the transmitting unit 120 constitute a communication unit that performs wireless communication with gNB200. UE100 is an example of a communication device.

[0018] The receiving unit 110 performs various types of reception under the control of the control unit 130. The receiving unit 110 includes an antenna and a receiver. The receiver converts the radio signal received by the antenna into a baseband signal (received signal) and outputs it to the control unit 130.

[0019] The transmitting unit 120 performs various types of transmissions under the control of the control unit 130. The transmitting unit 120 includes an antenna and a transmitter. The transmitter converts the baseband signal (transmission signal) output by the control unit 130 into a wireless signal and transmits it from the antenna.

[0020] The control unit 130 performs various control and processing operations in the UE 100. Such processing includes processing in each layer described later. The control unit 130 includes at least one processor and at least one memory. The memory stores programs executed by the processor and information used for processing by the processor. The processor may include a baseband processor and a CPU (Central Processing Unit). The baseband processor performs modulation, demodulation, encoding, and decoding of baseband signals. The CPU executes programs stored in memory and performs various processing operations.

[0021] Figure 3 shows the configuration of a gNB200 (base station) according to an embodiment. The gNB200 comprises a transmitter 210, a receiver 220, a control unit 230, and a backhaul communication unit 240. The transmitter 210 and receiver 220 constitute a communication unit that performs wireless communication with the UE100. The backhaul communication unit 240 constitutes a network communication unit that communicates with the CN20. The gNB200 is another example of a communication device.

[0022] The transmitting unit 210 performs various types of transmissions under the control of the control unit 230. The transmitting unit 210 includes an antenna and a transmitter. The transmitter converts the baseband signal (transmission signal) output by the control unit 230 into a radio signal and transmits it from the antenna.

[0023] The receiving unit 220 performs various types of reception under the control of the control unit 230. The receiving unit 220 includes an antenna and a receiver. The receiver converts the radio signal received by the antenna into a baseband signal (received signal) and outputs it to the control unit 230.

[0024] The control unit 230 performs various control and processing in the gNB200. Such processing includes processing in each layer described later. The control unit 230 includes at least one processor and at least one memory. The memory stores programs executed by the processor and information used for processing by the processor. The processor may include a baseband processor and a CPU. The baseband processor performs modulation, demodulation, encoding, and decoding of baseband signals. The CPU executes programs stored in memory and performs various processing.

[0025] The backhaul communication unit 240 is connected to an adjacent base station via the Xn interface, which is an inter-base station interface. The backhaul communication unit 240 is connected to the AMF / UPF300 via the NG interface, which is an inter-base station-core network interface. The gNB200 may consist of a central unit (CU) and distributed units (DU) (i.e., functionally separated), and the two units may be connected by the F1 interface, which is a fronthaul interface.

[0026] Figure 4 shows the configuration of the protocol stack for the user plane's wireless interface that handles data.

[0027] The user plane radio interface protocol comprises a physical (PHY) layer, a media access control (MAC) layer, a radio link control (RLC) layer, a packet data convergence protocol (PDCP) layer, and a service data adaptation protocol (SDAP) layer.

[0028] The PHY layer performs encoding / decoding, modulation / demodulation, antenna mapping / demapping, and resource mapping / demapping. Data and control information are transmitted between the UE100's PHY layer and the gNB200's PHY layer via a physical channel. The UE100's PHY layer receives downlink control information (DCI) transmitted from the gNB200 over the physical downlink control channel (PDCCH). Specifically, the UE100 performs blind decoding of the PDCCH using a Radio Network Temporary Identifier (RNTI) and acquires the successfully decoded DCI as the DCI addressed to its own UE. The DCI transmitted from the gNB200 has a CRC parity bit added, which is scrambled by the RNTI.

[0029] In NR, the UE100 can use a bandwidth narrower than the system bandwidth (i.e., the cell bandwidth). The gNB200 configures the UE100 with a bandwidth portion (BWP) consisting of consecutive PRBs (Physical Resource Blocks). The UE100 sends and receives data and control signals in the active BWP. The UE100 may have up to four BWPs configured, for example. Each BWP may have a different subcarrier spacing. The frequencies of these BWPs may overlap. If multiple BWPs are configured for the UE100, the gNB200 can specify which BWP to apply by controlling the downlink. This allows the gNB200 to dynamically adjust the UE bandwidth according to the amount of data traffic on the UE100, thereby reducing UE power consumption.

[0030] The gNB200 can, for example, configure up to three control resource sets (CORESETs) for each of up to four BWPs on a serving cell. A CORESET is a radio resource for control information that the UE100 should receive. The UE100 may have up to 12 or more CORESETs configured on a serving cell. Each CORESET may have an index from 0 to 11 or more. A CORESET may consist of six resource blocks (PRBs) and one, two, or three consecutive OFDM (Orthogonal Frequency Division Multiplex) symbols in the time domain.

[0031] The MAC layer performs data priority control, retransmission processing using Hybrid Automatic Repeat request (HARQ), and random access procedures. Data and control information are transmitted between the MAC layer of the UE100 and the MAC layer of the gNB200 via the transport channel. The MAC layer of the gNB200 includes a scheduler. The scheduler determines the transport format for the up and down links (transport block size, modulation and coding scheme (MCS)) and the resource blocks to be allocated to the UE100.

[0032] The RLC layer transmits data to the receiving RLC layer using the functions of the MAC layer and PHY layer. Data and control information are transmitted between the UE100's RLC layer and the gNB200's RLC layer via a logical channel.

[0033] The PDCP layer performs header compression / decompression, encryption / decryption, etc.

[0034] The SDAP layer maps IP flows, which are the units under which the core network performs QoS (Quality of Service) control, to wireless bearers, which are the units under which the access layer (AS: Access Stratum) performs QoS control. Note that if the RAN is connected to the EPC, the SDAP layer may not be necessary.

[0035] Figure 5 shows the configuration of the protocol stack of the wireless interface of the control plane that handles signaling (control signals).

[0036] The protocol stack of the control plane's wireless interface includes a Radio Resource Control (RRC) layer and a Non-Access Stratum (NAS) layer, instead of the SDAP layer shown in Figure 4.

[0037] RRC signaling for various settings is transmitted between the RRC layer of the UE100 and the RRC layer of the gNB200. The RRC layer controls the logical channel, transport channel, and physical channel in response to the establishment, re-establishment, and release of the radio bearer. If there is a connection (RRC connection) between the RRC of the UE100 and the RRC of the gNB200, the UE100 is in the RRC connected state. If there is no connection (RRC connection) between the RRC of the UE100 and the RRC of the gNB200, the UE100 is in the RRC idle state. If the connection between the RRC of the UE100 and the RRC of the gNB200 is suspended, the UE100 is in the RRC inactive state.

[0038] The NAS, located above the RRC layer, handles session management and mobility management, among other things. NAS signaling is transmitted between the UE100's NAS and the AMF300A's NAS. The UE100 also has application layers in addition to the wireless interface protocol. Layers below the NAS are called AS (Access Stratum).

[0039] (Overview of AI / ML technology) Next, the AI / ML technology according to the embodiment will be described. Figure 6 is a diagram showing the functional block configuration of the AI / ML technology (also referred to as "machine learning technology") in the mobile communication system 1 according to the embodiment.

[0040] The functional block configuration shown in Figure 6 includes a data acquisition unit A1, a model learning unit A2, a model inference unit A3, and a data processing unit A4.

[0041] The data collection unit A1 collects input data, specifically training data and inference data, outputs the training data to the model training unit A2, and outputs the inference data to the model inference unit A3. The data collection unit A1 may acquire data from its own device as input data. The data collection unit A1 may acquire data from another device as input data.

[0042] Model learning unit A2 performs model learning. Specifically, model learning unit A2 optimizes the parameters of the learning model (hereinafter also referred to as "model" or "AI / ML model") using machine learning with training data, derives (generates, updates) the trained model, and outputs the trained model to model inference unit A3. The model is a data-driven algorithm that applies AI / ML technology to generate a set of outputs based on a set of inputs. For example, y = ax + b In this context, a (slope) and b (intercept) are parameters, and optimizing these parameters constitutes machine learning. Generally, machine learning includes supervised learning, unsupervised learning, and reinforcement learning. Supervised learning is a method that uses correct data as training data. Unsupervised learning is a method that does not use correct data as training data. For example, in unsupervised learning, feature points are memorized from a large amount of training data, and correct answers (range estimation) are determined. Reinforcement learning is a method that assigns a score to the output results and learns how to maximize the score.

[0043] Model inference unit A3 performs model inference. Specifically, model inference unit A3 uses a pre-trained model to infer an output from the inference data and outputs the inference result data to data processing unit A4. For example, y = ax + b In this context, x represents the inference data, and y represents the inference result data. Note that "y = ax + b" is the model. A model with optimized slope and intercept, such as "y = 5x + 3", is a trained model. Here, there are various modeling methods, including linear regression analysis, neural networks, and decision tree analysis. The above "y = ax + b" can be considered a type of linear regression analysis. The model inference unit A3 may provide model performance feedback to the model learning unit A2.

[0044] Data processing unit A4 receives the inference result data and performs processing that utilizes the inference result data.

[0045] Figure 7 is a diagram showing an overview of the operation for each operation scenario according to the embodiment. In Figure 7, one of UE100 and gNB200 corresponds to the first communication device, and the other corresponds to the second communication device.

[0046] In step S1, UE100 transmits or receives control data related to machine learning techniques to or from gNB200. The control data may be an RRC message, which is signaling at the RRC layer (i.e., Layer 3). The control data may be a MAC CE (Control Element), which is signaling at the MAC layer (i.e., Layer 2). The control data may be downlink control information (DCI), which is signaling at the PHY layer (i.e., Layer 1). The downlink signaling may be UE-specific signaling. The downlink signaling may be broadcast signaling. The control data may be a control message at a control layer specialized for artificial intelligence or machine learning (e.g., an AI / ML layer).

[0047] (First action scenario) Figure 8 shows a first operation scenario according to the embodiment. In the first operation scenario, the data acquisition unit A1, the model learning unit A2, and the model inference unit A3 are located in the UE100 (for example, the control unit 130), and the data processing unit A4 is located in the gNB200 (for example, the control unit 230). In other words, model learning and model inference are performed on the UE100 side.

[0048] In the first operational scenario, machine learning techniques are introduced to the channel status information (CSI) feedback from UE100 to gNB200. The CSI (CSI feedback information) sent (feeded back) from UE100 to gNB200 is information about the channel status of the downlink between UE100 and gNB200. The CSI includes at least one of the following: channel quality indicator (CQI), precoding matrix indicator (PMI), and rank indicator (RI). Based on the CSI feedback from UE100, gNB200 performs, for example, downlink scheduling.

[0049] The gNB200 transmits a reference signal for the UE100 to estimate the channel state of the downlink. Such a reference signal may be, for example, a CSI reference signal (CSI-RS). Such a reference signal may also be a demodulation reference signal (DMRS). For example, let's assume the reference signal is CSI-RS.

[0050] Firstly, during model training, UE100 (receiver 110) receives a first reference signal from gNB200 using a first resource. Then, UE100 (model training unit A2) uses training data, including the first reference signal, to derive a trained model for inferring CSI from the reference signal. Such a first reference signal is sometimes referred to as full CSI-RS.

[0051] For example, UE100 (CSI generation unit 131) performs channel estimation using the received signal (CSI-RS) received by the receiver unit 110 from gNB200 and generates a CSI. UE100 (transmitter unit 120) transmits the generated CSI to gNB200. Model learning unit A2 uses multiple sets of received signals (CSI-RS) and CSIs as training data to perform model learning and derives a trained model for inferring CSI from received signals (CSI-RS).

[0052] Secondly, in model inference, UE100 (receiver 110) receives a second reference signal from gNB200 using a second resource which is less resource than the first resource. Then, UE100 (model inference unit A3) uses a trained model to infer the CSI as inference result data from the inference data including the second reference signal. In the description of the first operation scenario, such a second reference signal may be referred to as a partial CSI-RS or a punctured CSI-RS.

[0053] For example, UE100 (model inference unit A3) uses the received signal (CSI-RS) received by the receiver unit 110 from gNB200 as inference data, and uses a trained model to infer the CSI from the received signal (CSI-RS). UE100 (transmitter unit 120) transmits the inferred CSI to gNB200.

[0054] This allows the UE100 to feed back the accurate (complete) CSI to the gNB200 from the small amount of CSI-RS (partial CSI-RS) received from the gNB200. For example, the gNB200 can intentionally reduce (puncture) the CSI-RS to reduce overhead. It also allows the UE100 to cope with situations where wireless conditions deteriorate and some CSI-RS cannot be received properly.

[0055] Figure 9 shows a first example of reducing CSI-RS according to the embodiment. In the first example, the gNB200 reduces the number of antenna ports that transmit CSI-RS. For example, in the mode in which the UE100 performs model learning, the gNB200 transmits CSI-RS from all antenna ports on the antenna panel. On the other hand, in the mode in which the UE100 performs model inference, the gNB200 reduces the number of antenna ports that transmit CSI-RS and transmits CSI-RS from half of the antenna ports on the antenna panel. Note that antenna ports are an example of resources. This reduces overhead, improves the utilization efficiency of antenna ports, and reduces power consumption.

[0056] Figure 10 shows a second example of reducing CSI-RS according to the embodiment. In the second example, the gNB200 reduces the number of radio resources, specifically time-frequency resources, used to transmit CSI-RS. For example, in the mode in which the UE100 performs model learning, the gNB200 transmits CSI-RS using a predetermined amount of time-frequency resources. On the other hand, in the mode in which the UE100 performs model inference, the gNB200 transmits CSI-RS using a smaller amount of time-frequency resources than the predetermined amount. This reduces overhead, improves the efficiency of radio resource utilization, and reduces power consumption.

[0057] Next, the first operation pattern related to the first operation scenario will be described. In this first operation pattern, the gNB200 sends a switching notification to the UE100 as control data to notify it of a mode switch between the model learning mode (hereinafter also referred to as the "learning mode") and the model inference mode (hereinafter also referred to as the "inference mode"). The UE100 receives the switching notification and performs a mode switch between the learning mode and the inference mode. This makes it possible to perform a mode switch between the learning mode and the inference mode appropriately. The switching notification may be setting information that sets the mode for the UE100. The switching notification may also be a switching command that instructs the UE100 to switch modes.

[0058] In this first operation pattern, when model training is complete, UE100 sends a completion notification to gNB200 as control data indicating that model training is complete. gNB200 receives the completion notification. This allows gNB200 to understand that model training has been completed on the UE100 side.

[0059] Figure 11 is an operation flow diagram showing a first operation pattern related to a first operation scenario according to the embodiment. This flow may be performed after UE100 has established an RRC connection with the gNB200 cell. In the following operation flow diagram, optional steps are indicated by dashed lines.

[0060] In step S101, the gNB200 may notify or set the input data pattern in inference mode, for example, the CSI-RS transmission pattern (puncture pattern) in inference mode, to the UE100 as control data. For example, the gNB200 notifies the UE100 of the antenna ports and / or time-frequency resources that transmit or do not transmit CSI-RS in inference mode.

[0061] In step S102, the gNB200 may send a switch notification to the UE100 to initiate learning mode.

[0062] In step S103, UE100 enters learning mode.

[0063] In step S104, gNB200 transmits a full CSI-RS. UE100 receives the full CSI-RS and generates a CSI based on the received CSI-RS. In learning mode, UE100 can perform supervised learning using the received CSI-RS and its corresponding CSI. UE100 may derive and manage learning results (trained models) for each of its communication environments, for example, for each reception quality (RSRP, RSRQ, SINR) and / or mobile speed.

[0064] In step S105, UE100 sends (feeds back) the generated CSI to gNB200.

[0065] Subsequently, in step S106, when model training is complete, UE100 sends a completion notification to gNB200 indicating that model training is complete. UE100 may also send a completion notification to gNB200 when the derivation (generation, update) of the trained model is complete. Here, UE100 may also send a notification indicating the completion of training for each of its communication environments (e.g., mobile speed, reception quality). In this case, UE100 includes information in the notification indicating which communication environment the completion notification pertains to.

[0066] In step S107, the gNB200 sends a switch notification to the UE100 to switch from learning mode to inference mode.

[0067] In step S108, UE100 switches from learning mode to inference mode in response to receiving the switch notification in step S107.

[0068] In step S109, gNB200 transmits a partial CSI-RS. Upon receiving the partial CSI-RS, UE100 uses a trained model to infer the CSI from the received CSI-RS. UE100 may also select a trained model corresponding to its own communication environment from among the trained models managed for each communication environment, and use the selected trained model to perform CSI inference.

[0069] In step S110, UE100 sends (feeds back) the inferred CSI to gNB200.

[0070] In step S111, if UE100 determines that model training is necessary, it may send a notification to gNB200 as control data indicating that model training is necessary. For example, UE100 may send such a notification to gNB200 if it determines that the accuracy of the inference results can no longer be guaranteed when it moves, when its movement speed changes, when the reception quality at its location changes, when the cell it is located in changes, or when the bandwidth portion (BWP) it uses for communication changes.

[0071] Next, the second operation pattern related to the first operation scenario will be described. This second operation pattern may be used in combination with the operation pattern described above. In this second operation pattern, gNB200 sends a completion condition notification, indicating the completion conditions for model learning, to UE100 as control data. UE100 receives the completion condition notification and determines the completion of model learning based on the completion condition notification. This allows UE100 to appropriately determine the completion of model learning. The completion condition notification may also be configuration information that sets the completion conditions for model learning in UE100. The completion condition notification may also be included in a switching notification that notifies (instructs) the switching to learning mode.

[0072] Figure 12 is an operation flow diagram showing a second operation pattern related to the first operation scenario according to the embodiment.

[0073] In step S201, the gNB200 sends a completion condition notification, indicating the completion conditions for model training, to the UE100 as control data. The completion condition notification may include at least one of the following completion condition information.

[0074] • Tolerance for error relative to the correct data: For example, this refers to the acceptable range of error between the CSI generated using a standard CSI feedback calculation method and the CSI inferred by model inference. Once a certain level of training has been completed, UE100 can infer the CSI using the trained model at that point, compare it to the correct CSI, and determine that training is complete based on whether the error is within an acceptable range.

[0075] • Number of training data points: This refers to the number of data points used for training; for example, the number of CSI-RS signals received corresponds to the number of training data points. The UE100 can determine that training is complete when the number of CSI-RS signals received in training mode reaches the notified (configured) number of training data points.

[0076] • Number of training trials: This is the number of times the model has been trained using the training data. The UE100 can determine that training is complete when the number of training sessions in training mode reaches the notified (configured) number.

[0077] • Output score threshold: For example, consider a score in reinforcement learning. UE100 can determine that learning is complete based on whether the score reaches the notified (set) score.

[0078] UE100 continues learning based on full CSI-RS until it determines that learning is complete (steps S203 and S204).

[0079] In step S205, when UE100 determines that model training is complete, it may send a completion notification to gNB200 indicating that model training is complete.

[0080] Next, the third operation pattern related to the first operation scenario will be described. This third operation pattern may be used in combination with the operation pattern described above. If you want to improve the accuracy of CSI feedback, you can use not only CSI-RS but also other types of data, such as the reception characteristics of a physical downlink shared channel (PDSCH), as training data and inference data. In this third operation pattern, the gNB200 transmits data type information, which specifies at least the type of data to be used as training data, to the UE100 as control data. That is, the gNB200 specifies to the UE100 what the training data and inference data should be (type of input data). The UE100 receives the data type information and performs model training using the specified type of data. This allows the UE100 to perform appropriate model training.

[0081] Figure 13 is an operation flow diagram showing the third operation pattern related to the first operation scenario according to the embodiment.

[0082] In step S301, UE100 may send capability information to gNB200 as control data indicating which types of input data UE100 can handle using machine learning. Here, UE100 may also provide additional information such as the accuracy of the input data.

[0083] In step S302, gNB200 transmits data type information to UE100. The data type information may be setting information that sets the type of input data to UE100. Here, the input data type may be the reception quality and / or UE moving speed for CSI feedback. The reception quality may be the reference signal received power (RSRP), reference signal received quality (RSRQ), signal-to-interference noise ratio (SINR), bit error rate (BER), block error rate (BLER), analog / digital converter output waveform, etc.

[0084] Furthermore, assuming UE positioning as described later, the input data may include GNSS (Global Navigation Satellite System) location information (latitude, longitude, altitude), RF fingerprint (cell ID and its reception quality, etc.), angle of arrival (AoA) of the received signal, reception level, reception phase, and reception time difference (OTDOA) for each antenna, roundtrip time, and reception information for short-range wireless networks such as Wi-Fi (Local Area Network).

[0085] The gNB200 may specify the input data type independently for training data and inference data. The gNB200 may also specify the input data type independently for CSI feedback and UE positioning.

[0086] (Second action scenario) Next, we will explain the second operation scenario, focusing primarily on the differences from the first operation scenario. In the first operation scenario, we mainly explained the downlink reference signal (i.e., downlink CSI estimation). In the second operation scenario, we will explain the uplink reference signal (i.e., uplink CSI estimation). In the explanation of the second operation scenario, we will assume that the uplink reference signal is the sounding reference signal (SRS), but it may also be the uplink DMRS, etc.

[0087] Figure 14 shows a second operation scenario according to the embodiment. In the second operation scenario, the data acquisition unit A1, model learning unit A2, model inference unit A3, and data processing unit A4 are arranged in the gNB200 (for example, the control unit 230). That is, model learning and model inference are performed on the gNB200 side.

[0088] In the second operating scenario, machine learning techniques are introduced to the CSI estimation performed by the gNB200 based on the SRS from the UE100. Therefore, the gNB200 (e.g., the control unit 230) has a CSI generation unit 231 that generates CSI based on the SRS received by the receiver unit 220 from the UE100. This CSI is information indicating the channel state of the uplink between the UE100 and the gNB200. The gNB200 (e.g., the data processing unit A4) performs, for example, uplink scheduling based on the CSI generated from the SRS.

[0089] Firstly, during model training, the gNB200 (receiver 220) receives a first reference signal from the UE100 using a first resource. Then, the gNB200 (model training unit A2) uses training data, including the first reference signal, to derive a trained model for inferring the CSI from the reference signal (SRS). In the description of the second operating scenario, such a first reference signal may be referred to as the full SRS.

[0090] For example, the gNB200 (CSI generation unit 231) performs channel estimation using the received signal (SRS) received by the receiver unit 220 from the UE100 and generates a CSI. The model learning unit A2 uses multiple sets of received signals (SRS) and CSIs as training data to train the model and derives a trained model for inferring the CSI from the received signal (SRS).

[0091] Secondly, in model inference, the gNB200 (receiver 220) receives the second reference signal from the UE100 using a second resource that is less resource-intensive than the first resource. The UE100 (model inference unit A3) then uses a trained model to infer the CSI from the inference data, including the second reference signal, as inference result data. In the description of the second operating scenario, such a second reference signal may be referred to as a partial SRS or a punctured SRS. The same puncture patterns as in the first operating scenario can be used for the SRS (see Figures 9 and 10).

[0092] For example, the gNB200 (model inference unit A3) uses the received signal (SRS) received by the receiver unit 220 from the UE100 as inference data, and uses a trained model to infer the CSI from the received signal (SRS).

[0093] This allows the gNB200 to generate an accurate (complete) CSI from the small amount of SRS (partial SRS) received from the UE100. For example, the UE100 can intentionally reduce (puncture) the SRS to reduce overhead. It also allows the gNB200 to handle situations where wireless conditions deteriorate and some SRS cannot be received properly.

[0094] In this type of operation scenario, it is possible to replace "CSI-RS" with "SRS", "gNB200" with "UE100", and "UE100" with "gNB200" in the operation of the first operation scenario described above.

[0095] In the second operating scenario, the gNB200 transmits reference signal type information as control data to the UE100, specifying which type of reference signal (either a first reference signal (full SRS) or a second reference signal (partial SRS)) to transmit to the UE100. The UE100 receives the reference signal type information and transmits the specified SRS to the gNB200. This ensures that the appropriate SRS is transmitted to the UE100.

[0096] Figure 15 is an operation flowchart showing an example of operation related to the second operation scenario according to the embodiment.

[0097] In step S501, gNB200 configures UE100 for SRS transmission.

[0098] In step S502, the gNB200 starts learning mode.

[0099] In step S503, UE100 sends the full SRS to gNB200 according to the settings in step S501. gNB200 receives the full SRS and performs model training for channel estimation.

[0100] In step S504, the gNB200 identifies the SRS transmission pattern (puncture pattern) to be input as inference data to the trained model, and sets the identified SRS transmission pattern in the UE100.

[0101] In step S505, the gNB200 transitions to inference mode and starts model inference using the trained model.

[0102] In step S506, UE100 transmits a partial SRS according to the SRS transmission settings in step S504. gNB200 inputs this SRS as inference data into the trained model to obtain channel estimation results, and then uses these channel estimation results to perform uplink scheduling for UE100 (for example, controlling uplink transmission weights). Note that gNB200 may reconfigure UE100 to transmit a full SRS if the inference accuracy of the trained model deteriorates.

[0103] (Third action scenario) Next, the third operation scenario will be explained, primarily focusing on its differences from the first and second operation scenarios. The third operation scenario is an embodiment in which the position estimation of UE100 (so-called UE positioning) is performed using federated learning. Figure 16 is a diagram showing the third operation scenario according to this embodiment. In such an example of federated learning application, the following procedure is performed, for example.

[0104] First, the model is sent from the location server 400 to the UE100.

[0105] Secondly, UE100 performs model learning on the UE100 (model learning unit A2) side using the data present in UE100. The data present in UE100 includes, for example, the positioning reference signal (PRS) received by UE100 from gNB200 and / or the output data of the GNSS receiver 140. The data present in UE100 may also include position information (including latitude and longitude) generated by the position information generation unit 132 based on the PRS reception result and / or the output data of the GNSS receiver 140.

[0106] Thirdly, UE100 applies the trained model, which is the learning result, to UE100 (model inference unit A3), and also transmits the variable parameters included in the trained model (hereinafter also referred to as "trained parameters") to the position server 400. In the example above, the optimized a (slope) and b (intercept) correspond to the trained parameters.

[0107] Fourth, the position server 400 (federated learning unit A5) collects trained parameters from multiple UE100s and integrates them. The position server 400 may also transmit the trained model obtained through integration to the UE100s. Based on the trained model obtained through integration and the measurement reports from the UE100s, the position server 400 can estimate the position of the UE100s.

[0108] In the third operation scenario, gNB200 sends trigger setting information as control data to UE100, which sets the transmission trigger conditions for UE100 to send the learned parameters. UE100 receives the trigger setting information and sends the learned parameters to gNB200 (location server 400) when the set transmission trigger conditions are met. This allows UE100 to send the learned parameters at the appropriate time.

[0109] Figure 17 is an operation flowchart showing an example of operation related to the third operation scenario according to the embodiment.

[0110] In step S601, gNB200 may notify UE100 of the base model to be trained. Here, the base model may be a model that has been trained in the past. As described above, gNB200 may also send data type information to UE100 indicating what the input data should be.

[0111] In step S602, gNB200 instructs UE100 to train the model and sets the timing (trigger condition) for reporting trained parameters. The set reporting timing may be periodic. The set reporting timing may also be triggered (i.e., event triggered) when the training proficiency meets a condition.

[0112] In the case of periodic timing, gNB200 sets a timer value to UE100, for example. UE100 starts the timer when learning begins (step S603), and reports the learned parameters to gNB200 (location server 400) when it expires (step S604). Alternatively, gNB200 may specify to UE100 the radio frame or time to be reported. The radio frame may be specified as an absolute value, for example, SFN=512. The radio frame may be calculated by modulo operation. For example, gNB200 sets N as the set value and reports the learned parameters to UE100 in an SFN such that "SFN mod N=0" (step S604).

[0113] In the case of an event trigger, the completion conditions described above are set in UE100. When these completion conditions are met, UE100 reports the trained parameters to gNB200 (location server 400) (step S604). UE100 may also trigger the reporting of trained parameters, for example, when the accuracy of the model inference is better than the previously transmitted model. UE100 may also introduce an offset here and trigger when "current accuracy > previous accuracy + offset". UE100 may also trigger the reporting of trained parameters, for example, when training data has been input (trained) N or more times. Such an offset and / or value of N may be set in UE100 from gNB200.

[0114] In step S604, when the reporting timing conditions are met, UE100 reports the learned parameters at that point to the network (gNB200).

[0115] In step S605, the network (location server 400) integrates the learned parameters reported from multiple UE100s.

[0116] (First operational pattern for model transfer) Figure 18 shows a first operation pattern relating to model transfer according to the embodiment. In the drawings referenced in the following embodiments, non-essential processes are shown with dashed lines. In the following embodiments, the communication device 501 is mainly assumed to be UE100, but the communication device 501 may be gNB200 or AMF300A. Similarly, the communication device 502 is mainly assumed to be gNB200, but the communication device 502 may be UE100 or AMF300A.

[0117] As shown in Figure 18, in step S701, gNB200 sends a capability query message to UE100 requesting the transmission of a message containing information elements indicating the ability to perform machine learning processing. The capability query message is an example of a transmission request that requests the transmission of a message containing information elements indicating the ability to perform machine learning processing. UE100 receives the capability query message. However, gNB200 may also send the capability query message when it decides to perform machine learning processing (when it determines that it will perform it).

[0118] In step S702, UE100 sends a message to gNB200 containing informational elements indicating the execution capability for machine learning processing (or, from another perspective, the execution environment for machine learning processing). gNB200 receives the message. The message may be an RRC message, for example, a "UE Capability" message as defined in the RRC technical specifications, or a newly defined message (for example, a "UE AI Capability" message). Alternatively, if the communication device 502 is an AMF300A, the message may be a NAS message. Alternatively, if a new layer for executing or controlling machine learning processing (AI / ML processing) is defined, the message may be a message for that new layer. The new layer will be appropriately referred to as the "AI / ML layer".

[0119] An information element indicating the ability to perform machine learning processing is at least one of the following information elements (A1) to (A3).

[0120] • Information element (A1) Information element (A1) is an information element that indicates the processor's capacity for performing machine learning processing and / or the memory's capacity for performing machine learning processing.

[0121] The information element indicating the processor's capabilities for executing machine learning processing may also indicate whether or not the UE100 has an AI processor. If the UE100 has such a processor, the information element may include the AI ​​processor part number (model number). The information element may also indicate whether or not the UE100 can utilize a GPU (Graphics Processing Unit). The information element may also indicate whether or not machine learning processing must be performed on the CPU. By transmitting the information element indicating the processor's capabilities for executing machine learning processing from the UE100 to the gNB200, the network can determine, for example, whether or not the UE100 can utilize a neural network model as a model. The information element indicating the processor's capabilities for executing machine learning processing may also indicate the processor's clock frequency and / or the number of parallel executions possible.

[0122] The information element indicating the memory capacity for executing machine learning processing may also be an information element indicating the memory capacity of volatile memory (e.g., RAM: Random Access Memory) within the UE100's memory. This information element may also be an information element indicating the memory capacity of non-volatile memory (e.g., ROM: Read Only Memory) within the UE100's memory. This information element may also include both. The information element indicating the memory capacity for executing machine learning processing may be defined for each type, such as model storage memory, AI processor memory, GPU memory, etc.

[0123] Information element (A1) may be defined as an information element for inference processing (model inference). Information element (A1) may be defined as an information element for learning processing (model learning). Information element (A1) may be defined as both an information element for inference processing and an information element for learning processing.

[0124] ·Information element (A2) Information element (A2) is an information element that indicates the ability to perform inference processing. Information element (A2) may also be an information element that indicates the models supported by the inference processing. This information element may also be an information element that indicates whether or not a deep neural network model is supported. In that case, this information element may include at least one of the following: information indicating the number of layers (stages) of the neural network that can be supported, information indicating the number of neurons that can be supported (which may be the number of neurons per layer), and information indicating the number of synapses that can be supported (which may be the number of input or output synapses per layer or per neuron).

[0125] Information element (A2) may also be an information element indicating the execution time (response time) required to execute the inference process. Information element (A2) may also be an information element indicating the number of concurrent inference processes (for example, how many inference processes can be executed in parallel). Information element (A2) may also be an information element indicating the processing capacity of the inference process. For example, if the processing load of a certain standard model (standard task) is fixed at 1 point, the information element indicating the processing capacity of the inference process may be information indicating how many points its own processing capacity is.

[0126] • Information element (A3) Information element (A3) is an information element that indicates the ability to perform the learning process. Information element (A3) may also be an information element that indicates the learning algorithm supported by the learning process. The learning algorithms indicated by this information element include supervised learning (e.g., linear regression, decision trees, logistic regression, k-nearest neighbors, support vector machines, etc.), unsupervised learning (e.g., clustering, k-means algorithm, principal component analysis, etc.), reinforcement learning, and deep learning. If UE100 supports deep learning, this information element may include at least one of the following: information indicating the number of layers (stages) of the neural network that can be supported, information indicating the number of neurons that can be supported (which may be the number of neurons per layer), and information indicating the number of synapses that can be supported (which may be the number of input or output synapses per layer or per neuron).

[0127] Information element (A3) may also be an information element indicating the execution time (response time) required to execute the learning process. Information element (A3) may also be an information element indicating the number of simultaneous executions of the learning process (for example, how many learning processes can be executed in parallel). Information element (A3) may also be an information element indicating the processing capacity of the learning process. For example, if the processing load of a certain standard model (standard task) is fixed at 1 point, the information element indicating the processing capacity of the learning process may be information indicating how many points its own processing capacity is. Regarding the number of simultaneous executions, since the processing load of the learning process is generally higher than that of the inference process, the information may also be the number of simultaneous executions of the inference process and the learning process (for example, 2 inference processes and 1 learning process).

[0128] In step S703, gNB200 determines the model to set (deploy) on UE100 based on the information elements contained in the message received in step S702. This model may be a trained model used by UE100 in inference processing. This model may also be an untrained model used by UE100 in training processing.

[0129] In step S704, gNB200 sends a message to UE100 containing the model determined in step S703. UE100 receives the message and uses the model contained in the message to perform machine learning processing (training and / or inference). A specific example of step S704 will be explained in the following second operation pattern.

[0130] (Second operating pattern for model transfer) Figure 19 shows an example of a configuration message including a model and additional information according to the embodiment. The configuration message may be an RRC message sent from gNB200 to UE100, for example, an "RRC Reconfiguration" message as defined in the RRC technical specifications, or a newly defined message (for example, an "AI Deployment" message or an "AI Reconfiguration" message). Alternatively, the configuration message may be a NAS message sent from AMF300A to UE100. Or, if a new layer for executing or controlling machine learning processing (AI / ML processing) is defined, the message may be a message for that new layer.

[0131] In the example in Figure 19, the configuration message includes three models (Model #1 to #3). Each model is included as a container for the configuration message. However, the configuration message may include only one model. The configuration message further includes, as additional information, three individual additional pieces of information (Info #1 to #3) corresponding to each of the three models (Model #1 to #3), and common additional pieces of information (Meta-Info) that are associated with all three models (Model #1 to #3). Each of the individual additional pieces of information (Info #1 to #3) contains information specific to the corresponding model. The common additional pieces of information (Meta-Info) contain information common to all models within the configuration message.

[0132] Figure 20 shows a second operation pattern relating to model transfer according to the embodiment.

[0133] In step S711, gNB200 sends a configuration message to UE100 that includes the model and additional information. UE100 receives the configuration message. The configuration message includes at least one of the following information elements (B1) to (B6):

[0134] (B1 model) The "model" may be a pre-trained model used by UE100 in inference processing. The "model" may also be an untrained model used by UE100 in training processing. In configuration messages, the "model" may be encapsulated (contained). If the "model" is a neural network model, the "model" may be represented by the number of layers (stages), the number of neurons in each layer, and the synapses (weights) between each neuron. For example, a pre-trained (or untrained) neural network model may be represented by a combination of matrices.

[0135] A single configuration message may contain multiple "models." In this case, the multiple "models" may be included in the configuration message in list format. Multiple "models" may be configured for the same purpose, or they may be configured for different purposes. Details on the uses of models will be described later.

[0136] (B2) Model Index The "model index" is an example of additional information (e.g., individual additional information). The "model index" is the index (index number) assigned to a model. In the activation command and deletion message described later, you can specify the model using the "model index". You can also specify the model using the "model index" when changing the model's settings.

[0137] (B3 model application) "Model Use" is an example of additional information (individual or common additional information). "Model Use" specifies the function to which the model applies. For example, functions to which a model applies include CSI feedback, beam management (beam estimation, overhead latency reduction, beam selection accuracy improvement), positioning, modulation / demodulation, coding / decoding (CODEC), and packet compression. The content of the model use and its index (identifier) ​​may be predefined in the 3GPP technical specification, and "Model Use" may be specified by index. For example, CSI feedback is specified by use index #A, beam management by use index #B, and so on, with the model use and its index (identifier) ​​being defined. The UE100 deploys the model with the specified "Model Use" to the function block corresponding to the specified use. Note that "Model Use" may also be an information element that specifies the input and output data of the model.

[0138] (B4) Model execution requirements "Model execution requirements" is an example of additional information (e.g., individual additional information). "Model execution requirements" are information elements that indicate the performance (required performance) necessary to apply (execute) the model, such as processing delay (required latency).

[0139] (B5) Model Selection Criteria The "model selection criteria" is an example of additional information (individual additional information or common additional information). The UE100 applies (executes) the corresponding model depending on whether the criteria specified in the "model selection criteria" are met. The "model selection criteria" may be the UE100's movement speed. In that case, the "model selection criteria" may be specified as a speed range such as "slow movement" or "fast movement". The "model selection criteria" may also be specified as a threshold for movement speed. The "model selection criteria" may also be the radio quality measured by the UE100 (e.g., RSRP / RSRQ / SINR). In that case, the "model selection criteria" may be specified as a range for radio quality. The "model selection criteria" may also be specified as a threshold for radio quality. The "model selection criteria" may also be the UE100's location (latitude / longitude / altitude). The "model selection criteria" may also be set to follow sequential notifications from the network (activation commands described later). The "model selection criteria" may also specify the UE100's autonomous selection.

[0140] (B6) Whether or not learning processing is necessary The "Necessity of Learning Process" is an information element indicating whether or not learning (or retraining) is necessary for the corresponding model. If learning is required, the type of parameters to be used for learning may be further set. For example, in the case of CSI feedback, CSI-RS and UE movement speed are set to be used as parameters. If learning is required, the method of learning, such as supervised learning, unsupervised learning, reinforcement learning, or deep learning, may be further set. Whether or not to execute the learning process immediately after the model is set may also be further set. If not to execute immediately, the execution of learning may be controlled by the activation command described later. For example, in the case of Federated learning, whether or not to notify gNB200 of the results of the learning process of UE100 may be further set. If it is necessary to notify gNB200 of the results of the learning process of UE100, UE100 may encapsulate the trained model or trained parameters after the learning process is executed and send them to gNB200 via an RRC message or the like. The information element indicating "whether or not a learning process is necessary" may also include an information element indicating whether or not the corresponding model is used only for model inference, in addition to whether or not a learning process is necessary.

[0141] In step S712, UE100 determines whether the model configured in step S711 is deployable (executable). UE100 may perform this determination when activating the model, as described later, and step S713, described later, may be a message notifying of an error at the time of activation. Furthermore, UE100 may perform this determination while the model is in use (while machine learning processing is being executed), rather than at deployment or activation. If the model is determined to be undeployable (step S712: NO), i.e., an error occurs, in step S713, UE100 sends an error message to gNB200. The error message may be an RRC message sent from UE100 to gNB200, for example, a "Failure Information" message defined in the RRC technical specifications, or a newly defined message (for example, an "AI Deployment Failure Information" message). The error message may also be a UCI (Uplink Control Information) defined in the physical layer or a MAC CE (Control Element) defined in the MAC layer. Alternatively, the error message may be a NAS message sent from UE100 to AMF300A. Or, if a new layer (AI / ML layer) for performing machine learning processing (AI / ML processing) is defined, the message may be a message for that new layer.

[0142] An error message includes at least one of the following information elements (C1) through (C3):

[0143] (C1) Model Index This is the model index for models that were determined to be undeployable.

[0144] (C2) Usage Index This is an index of the uses of models that were determined to be undeployable.

[0145] (C3) Error Cause This is an information element regarding the cause of the error. The "cause of the error" may be, for example, "unsupported model," "processing capacity exceeded," "phase in which the error occurred," or "other error." "Unsupported model" may include, for example, the UE100 not being able to support a neural network model, or not being able to support machine learning processing (AI / ML processing) of a specified function. "Processing capacity exceeded" may include, for example, overload (processing load and / or memory load exceeding capacity), inability to satisfy the requested processing time, or interrupt processing or priority processing by the application (higher layer). "Phase in which the error occurred" is information indicating when the error occurred. "Phase in which the error occurred" may be categorized as deployment (configuration), activation, or operation. "Phase in which the error occurred" may be categorized as inference processing or training processing. "Other errors" are other causes.

[0146] UE100 may automatically delete the corresponding model if an error occurs. UE100 may also delete the model when it confirms that an error message has been received by gNB200, for example, when an ACK is received at a lower layer. gNB200 may recognize that the model has been deleted when it receives an error message from UE100.

[0147] On the other hand, if it is determined that the model configured in step S711 is deployable (step S712: YES), that is, if no error occurs, in step S714, UE100 deploys the model according to the configuration. "Deployment" may mean making the model applicable. "Deployment" may also mean actually applying the model. In the former case, the model is not applied simply by deploying it; it is applied when the model is activated by the activation command described later. In the latter case, once the model is deployed, it enters a state of being in use.

[0148] In step S715, UE100 sends a response message to gNB200 in response to the completion of model deployment. gNB200 receives the response message. UE100 may also send a response message when model activation is completed by the activation command described later. The response message may be an RRC message sent from UE100 to gNB200, for example, the "RRC Reconfiguration Complete" message specified in the RRC technical specification, or a newly defined message (for example, the "AI Deployment Complete" message). The response message may also be a MAC CE specified in the MAC layer. Alternatively, the response message may be a NAS message sent from UE100 to AMF300A. Alternatively, if a new layer for performing machine learning processing (AI / ML processing) is defined, the message may be a message for that new layer.

[0149] In step S716, UE100 may send a measurement report message, which is an RRC message containing the measurement results of the wireless environment, to gNB200. gNB200 receives the measurement report message.

[0150] In step S717, gNB200 selects a model to activate, for example based on a measurement report message, and sends an activation command (selection command) to UE100 to activate the selected model. UE100 receives the activation command. The activation command may be a DCI, MAC CE, RRC message, or an AI / ML layer message. The activation command may include a model index indicating the selected model. The activation command may also include information specifying whether UE100 will perform inference processing or training processing.

[0151] The gNB200 selects models to deactivate based on, for example, a measurement report message, and sends a deactivation command (selection command) to the UE100 to deactivate the selected models. The UE100 receives the deactivation command. The deactivation command may be a DCI, MAC CE, RRC message, or an AI / ML layer message. The deactivation command may include a model index indicating the selected models. Upon receiving the deactivation command, the UE100 may deactivate (cancel application) the specified models without deleting them.

[0152] In step S718, UE100 applies (activates) the specified model upon receiving the activation command. UE100 then performs inference and / or training using the activated model from among the deployed models.

[0153] Subsequently, in step S719, gNB200 sends a delete message to UE100 to delete the model. UE100 receives the delete message. The delete message may be a MAC CE, RRC message, NAS message, or AI / ML layer message. The delete message may include the model index of the model to be deleted. Upon receiving the delete message, UE100 deletes the specified model.

[0154] (Third operational pattern regarding model transfer) In this third operating pattern, UE100 notifies the network of the load status of the machine learning processing (AI / ML processing). This allows the network (e.g., gNB200) to determine how many more models can be deployed (or activated) on UE100 based on the notified load status. This third operating pattern does not necessarily presuppose the first operating pattern regarding model transfer described above. This third operating pattern may presuppose the first operating pattern.

[0155] Figure 21 is a diagram showing a third operation pattern related to model transfer according to the embodiment.

[0156] In step S751, gNB200 sends a message to UE100 that includes a request for information on the AI / ML processing load status or a setting for reporting the AI / ML processing load status. UE100 receives the message. The message may be a MAC CE, RRC message, NAS message, or an AI / ML layer message. The setting for reporting the AI / ML processing load status may include information that sets a report trigger (send trigger), for example, "Periodic" or "Event triggered". "Periodic" sets the reporting period, and UE100 reports at that period. "Event triggered" sets a threshold that is compared to a value indicating the AI / ML processing load status in UE100 (processing load value and / or memory load value), and UE100 reports when the value meets the condition of the threshold. Here, the threshold may be set for each model. For example, the message may associate a model index with a threshold.

[0157] In step S752, UE100 sends a message (report message) to gNB200 containing information indicating the AI / ML processing load status. This message may be an RRC message, for example, a "UE Assistance Information" message or a "Measurement Report" message. This message may be a newly defined message (for example, an "AI Assistance Information" message). This message may be a NAS message or an AI / ML layer message.

[0158] The message includes "processing load status" and / or "memory load status". "Processing load status" may indicate what percentage of processing power (processor capacity) is being used, or what percentage remains available. Alternatively, "processing load status" may express the load in points as described above, notifying how many points are being used and how many points remain available. UE100 may notify "processing load status" for each model. For example, UE100 may include at least one set of "model index" and "processing load status" in the message. "Memory load status" may include memory capacity, memory usage, or remaining memory. UE100 may notify "memory load status" by type, such as model storage memory, AI processor memory, GPU memory, etc.

[0159] In step S752, if UE100 wants to discontinue using a particular model for reasons such as high processing load or poor efficiency, it may include information (model index) indicating the model it wishes to delete or deactivate in the message. UE100 may also send alert information to gNB200 in the message when its own processing load becomes critical.

[0160] In step S753, gNB200 determines whether to change the model settings based on the message received from UE100 in step S752, and sends a message to UE100 for the model setting change. This message may be a MAC CE, RRC message, NAS message, or AI / ML layer message. gNB200 may also send the activation command or deactivation command described above to UE100.

[0161] (Model management) Figure 22 shows an example of model management according to this embodiment.

[0162] In step S801, the communication device 501 performs AI / ML processing (machine learning processing). This machine learning processing is one of the steps shown in Figure 23, which will be described later.

[0163] In step S802, communication device 501 transmits a notification regarding machine learning processing as control data to communication device 502. Communication device 502 receives the notification.

[0164] In step S802, the communication device 501 sends a notification to the communication device 502 indicating, for example, that it has an untrained model, a model that is being trained, and a trained model that has been tested.

[0165] In step S803, the communication device 502 transmits a response corresponding to the notification in step S802 to the communication device 501 as control data. The communication device 501 receives the response.

[0166] The notification in step S802 may be a notification indicating that the communication device 501 has an untrained model. In this case, step S803 may include at least one of the dataset and configuration parameters used for model training.

[0167] The notification in step S802 may be a notification indicating that the communication device 501 has a model being trained. In this case, the response in step S803 may include a dataset for continuing model training.

[0168] The notification in step S802 may be a notification indicating that the communication device 501 has a trained model for which the inspection has been completed. The response in step S803 may include information to initiate the use of the trained model for which the inspection has been completed.

[0169] Each of the notifications in step S802 and the responses in step S803 may include identification information to identify the index of the applicable model and / or the type or application of the applicable model (e.g., for CSI feedback, beam management, positioning, etc.). Hereafter, this information will also be referred to as "model application information, etc."

[0170] Figure 23 shows the model management according to the embodiment, specifically the details of step S801 in Figure 22.

[0171] In step S811, the communication device 501 performs the model deployment process. Here, the communication device 501 notifies the communication device 502 that it has an untrained model, that is, a model that needs to be trained. For example, the untrained model may be pre-installed when the communication device 501 is shipped. The untrained model may be obtained by the communication device 501 from the communication device 502. If the model training is not complete, for example, if it does not meet certain quality standards, the communication device 501 may notify the communication device 502 that it has an untrained model. For example, even if the model training was completed, if it has been moved to a different environment (for example, from indoors to outdoors), the quality of the model can no longer be guaranteed in monitoring. Based on this notification, the communication device 502 may provide the communication device 501 with a training dataset. The communication device 502 may perform any associated settings on the communication device 501. The communication device 502 may exclude the model from application, for example, by discarding it, deconfiguring it, or deactivating it.

[0172] In step S812, the communication device 501 executes the model training process. The communication device 501 notifies the communication device 502 that it is training the model. This notification may include information on the model's intended use, as described above. Based on this notification, the communication device 502 continues to provide the training dataset to the communication device 501. The communication device 502 may also recognize that the communication device 501 is using a conventional method that does not apply the model, if it receives a notification before or during training.

[0173] In step S813, the communication device 501 executes a model verification process. The model verification process is a sub-process of the model learning process. The model verification process evaluates the quality of the AI / ML model using a different dataset than the one used for model learning, and selects (adjusts) the model parameters. The communication device 501 may notify the communication device 502 that model learning is in progress or that model verification has been completed.

[0174] In step S814, the communication device 501 performs a model checking process. The model checking process is a sub-process of the model learning process. In the model checking process, the performance of the final AI / ML model is evaluated using a different dataset than the datasets used in model learning and model validation. Unlike model validation, model checking does not involve adjusting the model. The communication device 501 notifies the communication device 502 that it has a checked (i.e., a model that can guarantee a certain level of quality) model. This notification may include information on the model's intended use, as described above. Based on this notification, the communication device 502 performs a process to initiate the use of the model, such as setting up or activating the model. The communication device 502 may decide to provide an inference dataset and make the necessary settings for the communication device 501.

[0175] In step S815, the communication device 501 performs model sharing processing. For example, the communication device 501 sends (uploads) the trained model to the communication device 502.

[0176] In step S816, the communication device 501 performs a model activation process. The model activation process is a process that activates (enables) the model for a specific function. The communication device 501 may notify the communication device 502 that the model has been activated. This notification may include information such as the model's intended use, as described above.

[0177] In step S817, the communication device 501 performs model inference processing. Model inference processing is a process that uses a trained model to generate a set of outputs based on a set of inputs. The communication device 501 may notify the communication device 502 that it has performed model inference. This notification may include information such as the model's intended use, as described above.

[0178] In step S818, the communication device 501 performs model monitoring. Model monitoring is a process that monitors the inference performance of an AI / ML model. The communication device 501 may also send a notification regarding the model monitoring process to the communication device 502. This notification may include information on the model's intended use, as described above. Specific examples of this notification will be described later.

[0179] In step S819, the communication device 501 performs a model deactivation process. The model deactivation process is a process that deactivates (disables) a model for a specific function. The communication device 501 may notify the communication device 502 that the model has been deactivated. This notification may include information such as the model's intended use, as described above. The model deactivation process may also be a process that deactivates the currently active model and activates another model. This process is also called model switching.

[0180] (First operational pattern for model monitoring) In this first operation pattern, the communication device 501 performs inference processing using the trained model obtained by training the model. The communication device 501 monitors the performance of the trained model and determines whether the model needs to be retrained. If the communication device 501 determines that retraining is necessary, it sends a notification to the communication device 502 indicating the need for retraining. This makes it possible for the communication device 502 to provide the communication device 501 with training data to be used for retraining.

[0181] Figure 24 shows a first operation pattern relating to model monitoring according to the embodiment.

[0182] As shown in Figure 24, in step S821, the communication device 502 provides training data (training dataset) to the communication device 501. The training data is, for example, full C S I-RS stir.

[0183] In step S822, the communication device 501 generates a trained model by performing model training using training data.

[0184] In step S823, the communication device 501 may activate the model.

[0185] In step S824, the communication device 502 transmits configuration information, including parameters related to monitoring model performance (monitoring process), to the communication device 501 as control data. The communication device 501 receives the configuration information. The communication device 501 uses the parameters included in the configuration information to perform a model monitoring process to monitor the performance of the trained model (step S826).

[0186] The configuration information includes at least one of the following: the model index (or purpose) to be monitored, the monitoring threshold, and the conditions for sending monitoring result reports. The threshold (monitoring threshold) is a threshold that is compared with a value indicating the inference performance of the model in model monitoring, and is used to determine whether the model's performance meets a certain standard. The conditions for sending monitoring result reports may be, for example, periodic reporting or event triggers (such as when the monitoring threshold is no longer met).

[0187] In step S825, the communication device 501 performs model inference using the model.

[0188] In step S826, the communication device 501 starts monitoring the model.

[0189] In step S827, the communication device 501 determines whether the performance of the model meets a certain standard, that is, whether the model needs to be retrained.

[0190] If it is determined that the model's performance does not meet a certain standard, that is, that the model needs to be retrained (step S827: YES), in step S828, the communication device 501 sends a notification as control data to the communication device 502. This notification is a request for the provision of a training dataset, and for example, specifies the type of dataset to be provided (e.g., full C S The notification may include identification information to identify the I-RS). The notification is a notification that model retraining is required and may include identification information to identify, for example, the model index and / or the model type / purpose (CSI inference, beamforming inference, position inference, etc.).

[0191] In step S829, the communication device 502 provides the requested training dataset to the communication device 501. For example, the communication device 502 initiates full CSI-RS transmission.

[0192] In step S830, the communication device 501 modifies the trained model (for example, by adjusting the model parameters) by performing model training (retraining) using the training data.

[0193] (Second operational pattern for model monitoring) In this second operation pattern, the communication device 501 receives configuration information from the communication device 502, which includes monitoring parameters for monitoring the performance of the trained model. Based on this configuration information, the communication device 501 performs model monitoring processing using the monitoring dataset. This second operation pattern may be implemented in combination with the first operation pattern for model monitoring described above.

[0194] Here, the monitoring parameters may include time information indicating the time (i.e., monitoring period) during which the monitoring dataset is provided by the communication device 502. The communication device 501 may receive the monitoring dataset from the communication device 502 and perform model monitoring processing during the time indicated by the time information. The monitoring parameters may also include usage condition information indicating the conditions under which the monitoring dataset is reduced and used in order to monitor model performance in the model monitoring processing. The monitoring parameters may also include performance evaluation thresholds for monitoring model performance in the model monitoring processing.

[0195] Communication device 501 may transmit request information to communication device 502 requesting the provision of a monitoring dataset. Communication device 501 may receive a monitoring dataset transmitted from communication device 502 based on said request information.

[0196] The communication device 501 may perform model monitoring processing as follows. Specifically, the communication device 501 derives correct data (e.g., CSI feedback information) from a monitoring dataset (e.g., full CSI-RS) without using a trained model. The communication device 501 also takes partial data (e.g., partial CSI-RS) obtained by reducing the monitoring dataset as input and obtains inference result data (e.g., inferred CSI feedback information) output by the trained model.

[0197] The communication device 501 then evaluates the performance of the trained model by comparing the ground truth data with the inference result data. For example, the communication device 501 may determine that the inference performance of the model meets the criteria if the error of the inference result data with respect to the ground truth data is within a predetermined threshold range. On the other hand, the communication device 501 may determine that the inference performance of the model does not meet the criteria if the error of the inference result data with respect to the ground truth data is outside a predetermined threshold range.

[0198] Furthermore, if communication device 501 is UE100 and communication device 502 is gNB200, the monitoring dataset may include a reference signal transmitted from gNB200 to UE100. The reference signal may be CSI-RS or a positioning reference signal (PRS).

[0199] Figure 25 shows an example of applying AI / ML technology to CSI feedback as an example of a second operation pattern related to model monitoring according to the embodiment.

[0200] In step S851, the communication device 502 provides the communication device 501 with training data (training dataset). The training data is, for example, full C S I-RS stir.

[0201] In step S852, the communication device 501 generates a trained model by performing model training using training data.

[0202] In step S853, the communication device 501 may activate the model.

[0203] In step S854, if a monitoring dataset is required, the communication device 501 may send a request for the monitoring dataset to the communication device 502 as control data. This request may include the requested values ​​for the monitoring parameters set in step S855.

[0204] In step S855, the communication device 502 transmits configuration information, including monitoring parameters, to the communication device 501 as control data. The communication device 501 receives the configuration information.

[0205] The configuration information includes at least one of the following configuration parameters (D1) through (D5).

[0206] • (D1) Time information indicating the monitoring period The information indicates at least one of the following: the monitoring period, the monitoring timing, and the period for evaluating model performance. The information may also be a bitmap of slots (etc.). For example, communication device 502 transmits a full CSI-RS once per wireless frame (1 slot) and a partial CSI-RS at other times (other times). The monitoring period may be set dynamically. If communication device 502 is a gNB, DCI (or MAC CE) may be used to notify the start and / or stop of the transmission of monitoring data (or model monitoring processing by communication device 501).

[0207] • (D2) Resource information indicating monitoring resources This information may include information indicating the frequency resource from which the monitoring dataset is provided, such as the resource block start position and the number of resource blocks. The resource providing the monitoring dataset may be configured independently of the resource providing the inference data.

[0208] • (D3) Information indicating the purpose of the monitoring dataset This information indicates the intended use of the monitoring dataset, such as CSI feedback, beam management, or positioning. Here, we will assume that CSI feedback is the intended use of the monitoring dataset.

[0209] • (D4) Information indicating the usage conditions for the monitoring dataset This information indicates, for example, CSI-RS resources (or multiple patterns) that may cause the communication device 502 to become corrupted. During model monitoring, the communication device 501 (simulates) the corruption of the specified patterns to evaluate the model's performance.

[0210] • (D5) Thresholds for performance evaluation of model monitoring This threshold, for example, is a threshold that indicates the error range between the inference result data and the correct answer data.

[0211] In step S856, the communication device 502 provides the communication device 501 with monitoring data (monitoring dataset) during the monitoring period. The monitoring data is, for example, full C S I-RS stir.

[0212] In step S857, the communication device 501 performs model inference and model monitoring. Here, the communication device 501 derives ground truth data (e.g., CSI feedback information) from the monitoring dataset without using the trained model. The communication device 501 also takes partial data (e.g., partial CSI-RS) obtained by reducing the monitoring dataset as input and acquires inference result data (e.g., inferred CSI feedback information) output by the trained model. The communication device 501 then evaluates the performance of the trained model by comparing the ground truth data with the inference result data.

[0213] Here, we have described an example in which communication device 501 is UE100 and communication device 502 is gNB200, and AI / ML technology is applied to CSI feedback. However, communication device 501 may be gNB200 and communication device 502 may be UE100, and AI / ML technology may be applied to SRS transmission. In this case, model inference and model monitoring are performed by gNB200. For example, gNB200 sets a timing (i.e., monitoring period) for full SRS transmission to communication device 501. At that time, gNB200 may transmit at least one of the above (D1) to (D3) to UE100. UE100 transmits a full SRS at that timing and a puncture SRS at other timings. Here, UE100 may transmit full SRS periodically, such as a configured grant. UE100 may also transmit a full SRS in a one-shot manner by DCI (or MAC CE), such as in a PDCCH order. The gNB200 may cause the UE100 to transmit full SRS by notifying the UE100 via DCI (or MAC CE) of the start and / or stop of full SRS transmission.

[0214] Figure 26 shows an example of applying AI / ML technology to positioning (specifically, generating location information for the communication device 501) as another example of a second operation pattern relating to model monitoring according to the embodiment.

[0215] In step S871, the communication device 502 provides training data (training dataset) to the communication device 501. The training data may be, for example, a full PRS. The communication device 501 may derive location information from the full PRS and use the full PRS and location information as training data to generate a trained model that derives location information from the PRS (step S872). The training data may be a general reference signal or a PRS. If the communication device 501 has a GNSS receiver, it may use the reception state of a general reference signal or PRS (so-called RF fingerprint) and GNSS location information as training data to generate a trained model that derives location information from the RF fingerprint (step S872). Here, if the communication device 501 does not have a GNSS receiver, location information provided by a location server may be used instead of GNSS location information.

[0216] In step S873, the communication device 501 may activate the model.

[0217] In step S874, if a monitoring dataset is required, the communication device 501 may send a request for the monitoring dataset to the communication device 502 as control data. This request may include the requested values ​​for the monitoring parameters set in step S875.

[0218] In step S875, the communication device 502 transmits configuration information, including monitoring parameters, to the communication device 501 as control data. The communication device 501 receives the configuration information. The configuration information may include at least one of the above-mentioned configuration parameters (D1) to (D5), as well as parameters for setting the data source for the ground truth data during model monitoring (e.g., a GNSS receiver, a position server, etc.).

[0219] In step S876, the communication device 501 provides training data (training dataset) to the communication device 501 during the monitoring period. The training data may be, for example, full PRS or location information provided by the location server.

[0220] In step S877, the communication device 501 performs model inference and model monitoring. As a first example of model monitoring, when the location information (GNSS location information) from the GNSS receiver of the communication device 501 is used as the ground truth data, the communication device 501 performs model inference using the PRS reception status as input data, and evaluates the model's performance by comparing the error between the inference result data and the GNSS location information with a threshold.

[0221] As a second example of model monitoring, when location information provided by a location server (server-provided location information) is used as ground truth data, the communication device 501 performs model inference using the PRS reception status as input data, and evaluates the model's performance by comparing the error between the inference result data and the server-provided location information with a threshold. Here, the communication device 501 may also obtain server-provided location information from the location server by notifying the location server of the PRS reception status.

[0222] As a third example of model monitoring, when location information derived from the full PRS is used as ground truth data, the communication device 501 derives the location information as ground truth data from the full PRS without using the trained model. The communication device 501 also takes the partial PRS obtained by reducing the full PRS as input and acquires the inference result data (i.e., inferred location information) output by the trained model. The communication device 501 then evaluates the performance of the trained model by comparing the ground truth data with the inference result data.

[0223] (CSI feedback and beam management using AI / ML technology) Figure 27 shows CSI feedback and beam management applying AI / ML technology according to the embodiment.

[0224] In step S901, the communication device 501 receives radio signals transmitted through each of the multiple communication resources of the communication device 502. Assuming CSI feedback, these multiple communication resources are the multiple antenna ports of the communication device 502 (see Figure 9). Assuming beam management, these multiple communication resources are the multiple beams formed by the communication device 502 (e.g., gNB200) (see Figure 28).

[0225] In step S902, communication device 501 communicates (i.e., transmits and / or receives) information indicating a combination of communication resources having a predetermined correlation among the plurality of communication resources to communication device 502. Assuming CSI feedback, the information indicating the combination includes identification information for each antenna port constituting the combination. Assuming beam management, the information indicating the combination includes identification information for each beam constituting the combination.

[0226] In step S903, the communication device 501 performs AI / ML processing (machine learning processing) using the combination.

[0227] In this way, by having communication device 501 communicate information indicating a combination of communication resources having a predetermined correlation with communication device 502, it becomes possible to efficiently perform AI / ML processing (machine learning processing) using that combination. In the following, a combination of communication resources having a predetermined correlation may be referred to as a combination of antenna ports with a high correlation or a combination of beams with a high correlation.

[0228] In step S902, communication device 501 may receive a notification from communication device 502 as control data that includes information indicating the combination. In step S902, communication device 501 may transmit a notification to communication device 502 that includes information indicating the combination and / or information regarding communication resources that can stop transmitting radio signals.

[0229] The communication device 501 may identify a combination to communicate and acquire a trained model based on the identified combination and the radio signal received from the communication device 502. The trained model may be a model for deriving inference result data for another communication resource constituting the combination based on the reception status data of the radio signal of one of the communication resources constituting the combination. The communication device 501 may derive inference result data for the other communication resource based on the reception status data of the radio signal of the one communication resource.

[0230] Figure 29 shows a specific example of CSI feedback applying AI / ML technology according to the embodiment.

[0231] In step S911, communication device 502 may send a notification to communication device 501 as control data that includes information indicating a combination of antenna ports with a high correlation. Communication device 501 receives the notification. The combination may be a combination of antenna ports that communication device 501 is targeting for model learning.

[0232] In step S912, communication device 502 transmits full CSI-RS from multiple antenna ports. Communication device 501 transmits full C S The I-RS is received. If there is no notification in step S911, the communication device 501 may identify a combination of antenna ports with a high correlation. For example, the communication device 501 performs model training using the CSI-RS (full CSI-RS) transmitted from two antenna ports to generate a trained model (step S913). Then, the communication device 501 inputs the CSI-RS of one antenna port (partial CSI-RS) obtained by reducing (puncturing) the full CSI-RS as input data to the trained model, and if the inference result output from the trained model falls within a certain error with respect to the ground truth data, the two antenna ports may be determined to have a high correlation.

[0233] In step S914, the communication device 501 generates a trained model using a combination of antenna ports with high correlation. For example, the communication device 501 generates a trained model for estimating the CSI of antenna port #2 from the CSI-RS reception result of antenna port #1 using a combination of antenna ports #1 and #2 with high correlation.

[0234] In step S915, communication device 501 transmits a notification to communication device 502 as control data indicating that model learning is complete or a notification containing information on highly correlated antenna ports. Communication device 501 may also make the notification in step S915 when the generation of the trained model is completed in step S914, or when the inspection of the trained model is completed. The notification in step S915 may include at least one of the following: information indicating a combination of highly correlated antenna ports (e.g., the antenna port number of each highly correlated antenna port), information indicating a combination of antenna ports for which the generation of the trained model is complete, information indicating an antenna port from which CSI-RS transmission can be stopped (e.g., the antenna port number), or information indicating an antenna port from which CSI-RS transmission should continue.

[0235] In step S916, the communication device 502 identifies the antenna port on which to stop transmitting CSI-RS based on the notification in step S915.

[0236] In step S917, the communication device 502 stops transmitting some of the CSI-RS signals, thereby entering a state where it transmits only a portion of the CSI-RS signals. The communication device 501 receives these partial CSI-RS signals.

[0237] In step S918, the communication device 501 derives the CSI as inference result data based on the partial CSI-RS by model inference using the trained model generated in step S914.

[0238] In step S919, communication device 501 transmits the CSI feedback information, which is the inference result data, to communication device 502. Communication device 502 receives the CSI feedback information.

[0239] Figure 30 shows a specific example of beam management applying AI / ML technology according to the embodiment. Here, as shown in Figure 28, it is assumed that the communication device 502 (gNB200 in the illustrated example) forms multiple beams (e.g., beams #1 to #3). Beam management may be beam management for SSB-based beamforming. Such beam management includes, for example, SSB selection in the RRC idle state and beam monitoring and recovery in the RRC connected state. Beam management may also be beam management for CSI-RS-based beamforming (precoding). Such beam management includes, for example, management of PDSCH beamforming in the RRC connected state.

[0240] In this specific example, the communication device 501 uses AI / ML technology to estimate another beam (e.g., beam #2) based on the measurement of one beam (e.g., beam #1), specifically estimating the measurement result of the other beam. For such estimation to work, the beam control of beam #1 and beam #2 must be synchronized. If the correlation between the beams is low, for example, if the precoding differs for each slot, the above estimation cannot be performed.

[0241] As shown in Figure 30, in step S931, communication device 502 transmits a notification to communication device 501 that includes information indicating a combination of two or more beams to be controlled in conjunction. The notification may also indicate the correspondence between the beam to be used as inference data and the beam to be inferred. Communication device 501 receives the notification. The notification includes at least one of the following pieces of information (E1) to (E3).

[0242] • (E1) Beam identifier mapping information This is a set of two or more beam identifiers (beam indices). For example, communication device 502 notifies communication device 501 that beam #1 and beam #2 are associated (i.e., linked). Communication device 502 may also notify communication device 501 that beam #1 and beam #2 have a high correlation in the current location and propagation environment of communication device 501. It may also notify that estimation is possible between beam #1 and beam #2.

[0243] • (E2) Timing information for linked control For example, in slot #1, beam #1 and beam #2 are associated with time, and in slot #2, beam #1 and beam #3 are associated with time. One piece of time information may be individually associated with one set of beam identifiers. Alternatively, one piece of time information may be commonly associated with two or more sets of beam identifiers.

[0244] • (E3) Precoding (weight) information for linked beams For example, the communication device 501 infers the measurement result when received by beam #2 based on the precoding (weight) information of beam #2 and the measurement information of beam #1.

[0245] In step S932, the communication device 502 transmits a radio signal (e.g., SSB or PDSCH) by forming multiple beams. The communication device 501 receives and measures each beam.

[0246] In step S933, the communication device 501 performs model learning based on the notification in step S931. For example, the communication device 501 performs model learning to infer the measurement results of one beam from the measurement information (measurement results) of another beam.

[0247] In step S934, the communication device 501 generates a trained model using a combination of highly correlated beams. For example, the communication device 501 generates a trained model for estimating the measurement result of beam #2 from the measurement result of beam #1 using a combination of highly correlated beams #1 and #2.

[0248] In step S935, the communication device 502 forms multiple beams and transmits a radio signal (e.g., SSB or PDSCH). The communication device 501 receives and measures the beams.

[0249] In step S936, the communication device 501 performs beam measurements and infers the quality of other beams. The communication device 501 may use the inference results to estimate, for example, that the other beams are of higher quality than the current beam (pre-coded).

[0250] In step S937, the communication device 501 may feed back the inference result to the communication device 502, for example, as a CSI. Here, the communication device 501 may transmit CSI feedback information to the communication device 502 along with information indicating that it is an inferred CSI.

[0251] (Other embodiments) Although the above embodiments mainly described communication between UE100 and gNB200, the operation according to the above embodiments may also be applied to communication between gNB200 and AMF300A (i.e., communication between the base station and the core network). The above control data may be transmitted from gNB200 to AMF300A over the NG interface. The above control data may also be transmitted from AMF300A to gNB200 over the NG interface. Federated learning execution requests and / or federated learning results may be exchanged between AMF300A and gNB200. Each of the above operation scenarios may also be applied to communication between gNB200 and another gNB200 (i.e., communication between base stations). The above control data may be transmitted from gNB200 to another gNB200 over the Xn interface. Federated learning execution requests and / or federated learning results may be exchanged between gNB200 and another gNB200. Each of the above operations may be applied to communication between UE100 and another UE100 (i.e., communication between user devices). The above control data may be transmitted from UE100 to another UE100 over the sidelink. Requests for federated learning and / or learning results of federated learning may be exchanged between UE100 and another UE100.

[0252] Each of the above-described operation flows can be performed not only independently, but also in combination of two or more operation flows. For example, some steps of one operation flow may be added to another operation flow. Some steps of one operation flow may be replaced with some steps of another operation flow. It is not necessary to execute all steps in each flow; only some steps may be executed.

[0253] In the above embodiment, an example was described in which the base station is an NR base station (gNB), but the base station may also be an LTE base station (eNB). Furthermore, the base station may be a relay node such as an IAB (Integrated Access and Backhaul) node. The base station may also be a DU (Distributed Unit) of an IAB node. The user equipment (terminal equipment) may be a relay node such as an IAB node, or it may be an MT (Mobile Termination) of an IAB node.

[0254] A program may be provided that causes a computer to execute each process performed by the communication device (e.g., UE100 or gNB200). The program may be recorded on a computer-readable medium. Using a computer-readable medium, it is possible to install the program on a computer. Here, the computer-readable medium on which the program is recorded may be a non-transient recording medium. The non-transient recording medium is not particularly limited, but may be a recording medium such as a CD-ROM or DVD-ROM. Furthermore, the circuits that execute each process performed by the communication device may be integrated, and at least a part of the communication device may be configured as a semiconductor integrated circuit (chipset, SoC: System on a chip).

[0255] The terms "based on" and "depending on" used in this disclosure do not mean "based solely on" or "depending solely on" unless otherwise specified. The term "based on" means both "based solely on" and "at least partially on." Similarly, the term "depending on" means both "at least partially on" and "at least partially on." Furthermore, "obtain / acquire" may mean obtaining information from stored information, obtaining information from information received from other nodes, or obtaining information by generating it. The terms "include" and "comprise" do not mean to include only the listed items, but may include only the listed items, or may include additional items in addition to the listed items. Furthermore, the term "or" used in this disclosure is not intended to mean exclusive OR. In addition, any reference to elements using designations such as "first," "second," etc., used in this disclosure does not limit the quantity or order of those elements in general. These designations may be used herein as a convenient way to distinguish between two or more elements. Therefore, references to the first and second elements do not imply that only two elements may be employed therein, or that the first element must precede the second element in any way. In this disclosure, where articles are added by translation, such as a, an, and the in English, these articles shall be plural unless it is clearly indicated otherwise by the context.

[0256] Although the embodiments have been described in detail above with reference to the drawings, the specific configuration is not limited to those described above, and various design changes can be made without departing from the gist of the invention.

[0257] This application claims priority to Japanese Patent Application No. 2022-117532 (filed on July 22, 2022), and all of its contents are incorporated into the specification of this application.

[0258] (Supplementary Note) The features related to the above-described embodiments are noted below.

[0259] (Supplementary Note 1) A communication method for applying machine learning technology to wireless communication between a user device and a base station in a mobile communication system, wherein one of the communication devices, either the user device or the base station, transmits a notification indicating at least one of having an unlearned model, having a model being learned, and having a learned model for which inspection has been completed, to the other communication device between the user device and the base station; and the one communication device receives a response corresponding to the notification from the other communication device. Communication method.

[0260] [[ID=2º]] (Supplementary Note 2) The transmitting step is a step of transmitting the notification indicating having the unlearned model, and the receiving step includes receiving at least one of a dataset and setting parameters used for model learning as the response. The communication method according to Supplementary Note 1.

[0261] (Supplementary Note 3) The transmitting step is a step of transmitting the notification indicating having the model being learned, <ã and the receiving step includes receiving a dataset for continuing model learning as the response. The communication method according to Supplementary Note 1. [[ID=3º]]

[0262] (Supplementary Note 4) The transmitting step is a step of transmitting the notification indicating having the learned model for which inspection has been completed, and the receiving step includes receiving information for starting the use of the learned model for which inspection has been completed as the response. The communication method described in Appendix 1.

[0263] (Note 5) The notification includes an index of the model and / or identification information for identifying the type or use of the model. The communication method described in any of the appendices 1 to 4.

[0264] (Note 6) A communication method that applies machine learning technology to wireless communication between user equipment and a base station in a mobile communication system, The steps include: one of the user device and the base station performing inference processing using the trained model obtained by training the model; The steps include: one of the communication devices monitors the performance of the trained model to determine the need to retrain the model; The step of one of the communication devices determining that relearning is necessary, by transmitting a notification indicating the need for relearning to the other communication device among the user device and the base station. Communication method.

[0265] (Note 7) The steps include: one communication device receiving configuration information, including parameters related to monitoring the performance, from the other communication device; The aforementioned communication device further comprises the step of performing a monitoring process to monitor the performance of the trained model using the parameters included in the configuration information. The communication method described in Appendix 6.

[0266] (Note 8) The aforementioned notice includes information requesting the provision of training data to be used for the retraining. The communication method described in Appendix 6 or 7.

[0267] (Note 9) The notification includes identification information for identifying the type of training data. The communication method described in Appendix 8.

[0268] (Appendix 10) The notification includes the index of the model and / or identification information for identifying the type or use of the model The communication method according to any one of Appendices 6 to 9

[0269] (Appendix 11) A communication method for applying machine learning technology to wireless communication between a user device and a base station in a mobile communication system, wherein one of the communication devices, i.e., the user device or the base station, receives setting information including monitoring parameters for monitoring the performance of a learned model from the other communication device, and the one communication device performs the monitoring process using a monitoring dataset based on the setting information Communication method

[0270] (Appendix 12) The monitoring parameter includes time information indicating the time when the monitoring dataset is provided by the other communication device, The step of performing the monitoring process includes receiving the monitoring dataset from the other communication device at the time indicated by the time information and performing the monitoring process The communication method according to Appendix 11

[0271] (Appendix 13) The one communication device further has a step of transmitting request information for requesting the provision of the monitoring dataset to the other communication device, The step of performing the monitoring process includes receiving the monitoring dataset transmitted from the other communication device based on the request information The communication method according to Appendix 11 or 12

[0272] (Appendix 14) The monitoring parameters further include information indicating the conditions under which the monitoring dataset is reduced and used in order to monitor the performance during the monitoring process. The communication method described in any of the appendices 11 to 13.

[0273] (Note 15) The monitoring parameters further include performance evaluation thresholds for monitoring the performance in the monitoring process. The communication method described in any of the appendices 11 to 14.

[0274] (Note 16) The aforementioned monitoring step is, The steps include: deriving the correct data from the aforementioned monitoring dataset without using the aforementioned trained model; The steps include: obtaining inference result data output by the trained model using partial data obtained by reducing the aforementioned monitoring dataset as input; The step includes evaluating the performance of the trained model by comparing the ground truth data with the inference result data. The communication method described in any of the appendices 11 to 15.

[0275] (Note 17) The first communication device is the user device, and the second communication device is the base station. The monitoring dataset includes a reference signal transmitted from the base station to the user device. The communication method described in any of the appendices 11 to 16. [Explanation of Symbols]

[0276] 1: Mobile communication systems 100 :UE 110: Receiving unit 120: Transmitter 130: Control Unit 131:CSI generation unit 132: Location information generation unit 140: GNSS receiver 200 :gNB 210: Transmitter 220: Receiving unit 230: Control Unit 231:CSI generation unit 240: Backhaul Communications Department 400: Location Server 501: Communication device 502: Communication equipment A1: Data Collection Department A2: Model Learning Department A3: Model inference section A4: Data Processing Section A5: Joint Learning Department

Claims

1. A communication method that applies machine learning technology to wireless communication between a user device and a network node in a mobile communication system, The communication device of one of the user device and the network node transmits a notification to the other communication device of the user device and the network node indicating that it has at least one of the following: that it has an untrained model, that it has a model being trained, and that it has a trained model whose testing has been completed. The other communication device transmits control information to the first communication device for controlling the operation of the machine learning model in the first communication device based on the notification. The aforementioned communication device has the ability to receive the control information from the other communication device, The control information includes at least one of the following: a dataset used for model training, information for initiating the use of the machine learning model, information for changing the settings of the machine learning model, and information for deleting the machine learning model. Communication method.

2. The aforementioned transmission is the transmission of the notification indicating that the model has not been trained. The aforementioned receiving includes receiving at least one of the dataset and configuration parameters used for model training as control information. The communication method according to claim 1.

3. The transmission described above means transmitting the notification indicating that the model being trained is being transmitted. The aforementioned receiving includes receiving a dataset for continuing model learning as control information. The communication method according to claim 1.

4. The transmission described above is the transmission of the notification indicating that the trained model has been inspected and the inspection has been completed. The aforementioned receiving includes receiving information as control information that enables the use of the trained model after the inspection has been completed. The communication method according to claim 1.

5. The notification includes an index of the model and / or identification information for identifying the type or use of the model. The communication method according to any one of claims 1 to 4.

6. A user device that performs wireless communication with a network node to which machine learning technology is applied, A transmitting unit that sends a notification to the network node indicating that the user device has at least one of the following: that it has an untrained model, that it has a model being trained, and that it has a trained model for which testing has been completed. The device includes a receiving unit that receives control information from the network node for controlling the operation of a machine learning model in the user device, The control information includes at least one of the following: a dataset used for model training, information for initiating the use of the machine learning model, information for changing the settings of the machine learning model, and information for deleting the machine learning model. User device.

Images