Terminal device in wireless communication system in which artificial intelligence (AI) / machine learning (ML) model is used, base station device, control method, and program

Abstract

This terminal device compliant with the cellular communication standard of the 3rd Generation Partnership Project (3GPP) receives, from a base station device the terminal device is connected to, information about a parameter indicating an environment to which artificial intelligence (AI) / a machine learning (ML) model is applied in relation to a beam formed by the base station device, determines whether or not an AI / ML model held by the terminal device is applicable to the environment indicated by the parameter and, if it is determined that the AI / ML model held by the terminal device is applicable, executes inference using the AI / ML model.

Description

Terminal equipment, base station equipment, control method, and program in a wireless communication system using artificial intelligence (AI) / machine learning (ML) models.

[0001] This invention relates to control technology for the appropriate use of artificial intelligence (AI) / machine learning (ML) models in cellular communication systems.

[0002] In the cellular communication system of the Third Generation Partnership Project (3GPP®), the use of artificial intelligence (AI) / machine learning (ML) estimation to improve communication efficiency is being considered (see Non-Patent Literature 1). For example, one method of applying AI / ML is being considered in which a terminal device inputs measured values ​​of the signal transmission characteristics of a specific beam into an AI / ML model to estimate the transmission characteristics of other beams (e.g., different in time or space). Another method of applying AI / ML is being considered in which a terminal device inputs measurement results from the terminal device for some of multiple beams into an AI / ML model to estimate the best beams (a predetermined number of the best ones) for communication of that terminal device.

[0003] 3GPP (Registered Trademark) TR38.843 V18.0.0, December 2023

[0004] If the characteristics of the beam formed by the base station equipment differ significantly from the characteristics of the beam handled by the AI / ML model, problems may arise in terminal equipment due to inconsistencies in the generation or updating of the model, or in estimations using that model, with the real environment.

[0005] The present invention provides a method for applying an artificial intelligence (AI) / machine learning (ML) model that is suitable for the environment in which a terminal device is placed.

[0006] A terminal device according to one aspect of the present invention is a terminal device compliant with the cellular communication standard of the Third Generation Partnership Project (3GPP), comprising: receiving means for receiving parameter information from a connected base station device indicating the environment to which an artificial intelligence (AI) / machine learning (ML) model relating to the beam formed by the base station device is applied; determining means for determining whether the AI / ML model held by the terminal device can be applied in the environment of the parameters; and executing means for performing inference using the AI / ML model when it is determined that the AI / ML model held by the terminal device can be applied.

[0007] A base station device according to one aspect of the present invention is a base station device compliant with the cellular communication standard of the Third Generation Partnership Project (3GPP), and includes: notification means for notifying a connected terminal device of parameter information indicating the environment in which an artificial intelligence (AI) / machine learning (ML) model relating to the beam formed by the base station device is applied; receiving means for receiving information from the terminal device indicating whether or not the AI / ML model held by the terminal device can be applied in the environment of the parameters; and instruction means for instructing the execution of inference using the AI / ML model when information indicating that the AI / ML model held by the terminal device can be applied in the environment of the parameters is received.

[0008] According to the present invention, it becomes possible to train an artificial intelligence (AI) / machine learning (ML) model that is suitable for the environment in which the terminal device is placed.

[0009] Other features and advantages of the present invention will become apparent from the following description with reference to the accompanying drawings. In the accompanying drawings, the same or similar components are given the same reference numeral.

[0010] The attached drawings are included in the specification and constitute a part thereof, illustrating embodiments of the present invention and are used to explain the principles of the present invention together with their description. Figure 1 is a diagram showing an example of a system configuration. Figure 2 is a diagram showing an example of the communication processing flow in the learning phase. Figure 3 is a diagram showing an example of information about a beam group. Figure 4 is a diagram showing an example of the communication flow when an external server is used for learning an AI / ML model. Figure 5 is a diagram showing an example of information about an AI / ML model. Figure 6A is a diagram showing an example of a resource block to which CSI-RS is transmitted. Figure 6B is a diagram showing an example of a resource block to which CSI-RS is transmitted. Figure 6C is a diagram showing an example of a resource block to which CSI-RS is transmitted. Figure 7 is a diagram showing an example of the processing flow when performing inference using an AI / ML model. Figure 8 is a diagram showing an example of the hardware configuration of a base station device and a terminal device. Figure 9 is a diagram showing an example of the functional configuration of a terminal device. Figure 10 is a diagram showing an example of the functional configuration of a base station device.

[0011] The embodiments will be described in detail below with reference to the attached drawings. Note that the following embodiments do not limit the invention as defined in the claims, and not all combinations of features described in the embodiments are essential to the invention. Two or more of the features described in the embodiments may be combined in any way. Furthermore, identical or similar configurations will be given the same reference numeral, and redundant descriptions will be omitted.

[0012] (System Configuration) Figure 1 shows an example of the configuration of the wireless communication system according to this embodiment. This wireless communication system is a wireless communication system that conforms to the cellular communication standards of the 3rd Generation Partnership Project (3GPP®), such as 5th generation (5G), or its successor standards. This wireless communication system is composed of communication devices including base station devices 101 to 104 and terminal device 111. In the following, if there is no need to particularly distinguish between any of the base station devices 101 to 104, the reference numerals will be omitted and they will be referred to as "base station devices," and the terminal device 111 will be referred to as "terminal devices" without specifying the reference numerals. A terminal device is a communication device that establishes a connection with a base station device, transmits user data on the uplink (link in the direction from the terminal device to the base station device), and receives user data on the downlink (link in the direction from the base station device to the terminal device). Similarly, a base station device is a communication device that transmits user data on the downlink to the connected terminal device and receives user data on the uplink. Note that while Figure 1 shows four base station devices and one terminal device for simplicity, the number of these devices is not limited. For example, there can naturally be many terminal devices.

[0013] In Figure 1, base station devices 101 to 104 can each form multiple beams. The terminal device measures the reference signals transmitted from these multiple beams (e.g., synchronization signal (SS) / physical broadcast channel (PBCH) block (SSB), channel state information-reference signal (CSI-RS), etc.) and reports the measurement results to the base station device. The base station device receives the report of the measurement results and can select a beam suitable for communication with the terminal device. However, if the number of beams formed by the base station device increases, the time it takes for the terminal device to complete the measurement may become longer. Therefore, in this embodiment, the terminal device uses an artificial intelligence (AI) / machine learning (ML) model to measure the reference signals for some of the beams and inputs the measurement results into the AI / ML model to obtain estimated values ​​corresponding to the measurement results for beams that have not been measured. This makes it possible to select a beam suitable for communication with the terminal device from all the formed beams while reducing the number of beams to be measured.

[0014] On the other hand, the beams formed by base station equipment are not always constant. For example, as shown in Figure 1, base station equipment 102 forms a narrow beam width over a narrower area compared to base station equipment 101. Also, base station equipment 103 forms fewer beams widths than base station equipment 101. In situations where the formed beams differ significantly, applying the same AI / ML model can result in large estimation errors, and in some cases, communication efficiency may actually decrease. Furthermore, if a common AI / ML model is trained in situations where the formed beams differ significantly, the training of that AI / ML model may not converge properly, or its estimation performance may be insufficient. For this reason, the training and application of the AI / ML model for the beam of base station equipment 101 should be performed for base station equipment 104 and other base station equipment that form beams similar to those of base station equipment 101.

[0015] In this embodiment, in view of these circumstances, we provide a technique for appropriately configuring an AI / ML model used in a terminal device to perform estimations related to the selection of beams to be used for communication from among the beams formed by the base station device, and for enabling appropriate training of such an AI / ML model.

[0016] (Operation in the Learning Phase) First, the learning (training) of the AI / ML model according to this embodiment will be described. In this embodiment, for example, the radio signals transmitted in each of the multiple beams formed by the base station device are measured by the terminal device. Then, for example, the measurement results of some beams from the measurement results of radio quality such as received power are used as input, and the measurement results of radio quality in all beams (beams other than the input) are used as training data to learn the AI / ML model. As a result, an AI / ML model is generated that takes the measurement results of radio quality in some beams as input and outputs estimated values ​​of radio quality in all beams (beams other than the input). Alternatively, the measured values ​​of radio quality for some beams may be used as input, and the index of a predetermined number of beams from the one with the best radio quality (for example, from the one with the highest received power) may be used as training data to learn the AI / ML model. In this case, an AI / ML model is generated that takes the measurement results of some beams as input and outputs the index of a predetermined number of beams from the one with the best radio quality. Furthermore, an AI / ML model may be trained using the measured wireless quality of some or all beams at a predetermined timing as input, and the wireless quality of some or all beams after a predetermined time as training data. In this case, an AI / ML model can be generated that takes the measured wireless quality of some or all beams as input and outputs an estimated value of the wireless quality of some or all beams after a predetermined time. In this embodiment, information is provided from the base station device to the terminal device that enables the terminal device to determine whether a beam group suitable for the model to be trained is formed by the base station device. This allows the terminal device to acquire appropriate input and training data for generating an AI / ML model suitable for a specific environment. Note that the training of the AI / ML model may be performed by the terminal device or by other devices such as an external server.

[0017] Figure 2 shows an example of the processing flow performed between the base station device and the terminal device in relation to training. First, the terminal device establishes a connection (RRC (Radio Resource Control) connection) with the base station device, enabling it to communicate with the base station device individually (S201). Then, the terminal device notifies the base station device of information indicating that it has the capability to learn AI / ML models (S202). The information notified here may indicate that the terminal device can learn AI / ML models in general, or it may include information indicating the types of AI / ML models that the terminal device can learn. Here, for example, it may be indicated that the terminal device performs beam-related learning, such as estimating the radio quality of other beams based on radio quality information for a specific beam. It may also indicate the types of AI / ML models that an external server, not the terminal device, can learn. In other words, the terminal device may notify the base station device that it has the capability to collect data used for learning a specific type of AI / ML model. Then, the base station device provides the terminal device with information about the beam group formed by its own device (S203). Information regarding this beam group includes at least one of the following: the angular range covered by the entire beam, the number of beams, the beam width of each beam (the range in which the gain is halved from the peak (decreased by 3 dB)), the transmission power, and the EIRP (equivalent isotropically radiated power). In one example, in S202, the terminal device may notify the base station device that it has implemented functions related to acquiring learning data, such as requesting the start of transmission of the reference signal described later, measurement, and holding of measured values, as capability information. Then, in S203, the base station device may provide information regarding the beam group only to terminal devices that have implemented functions related to acquiring learning data. Note that the processing in S202 may be omitted.

[0018] The terminal device determines, based on information about the beam group, whether it is possible to acquire training data for training (learning) (S204). If the terminal device determines that the base station device is able to provide training data, it requests the base station device to transmit reference signals through each of the multiple beams for training the AI / ML model (S205).

[0019] In S204, when training an AI / ML model suitable for a specific environment, the terminal device determines whether the beam group formed by the base station device provides an environment identical or similar to that specific environment. For example, if the terminal device can acquire training data based on a reference signal transmitted from a beam whose beam width (half-power angle) is within a predetermined range (or is a predetermined value), it refers to the beam width (half-power angle) information in the beam group information notified by the base station device. If the value indicated by that information is within the predetermined range (or matches the predetermined value), the terminal device can determine that it can acquire training data from the reference signal transmitted by the beam formed by the base station device. Alternatively, the terminal device may determine whether it can acquire training data using a combination of multiple factors, such as determining that training data can be acquired when the angular range covered by the entire beam is within a certain range and the number of beams matches a predetermined number. An identifier may be assigned to each set of parameters in the beam group information. Hereinafter, this identifier will be referred to as the environment identifier.

[0020] Figure 3 shows an example of a correspondence table between environmental identifiers and information about beam groups. In Figure 3, environmental identifiers such as E1 to E6 are assigned to each of several sets of parameters related to beam groups. Note that Figure 3 is just one example, and it is naturally assumed that there are sets of parameters with values ​​not shown. Also, in Figure 3, information on the transmission angle, number of beams, beam width, transmission power, and EIRP for all beams is shown, but some of this information may be omitted or replaced with other elements, and further information elements not shown may be defined. In S203, the base station device may notify the terminal device of the environmental identifier of the set of parameters corresponding to the beams formed by the device. The terminal device may determine whether training data can be acquired based on the environmental identifier. In one example, an AI / ML model is associated with an environmental identifier, and the terminal device may determine that training data can be acquired when the base station device notifies it of the environmental identifier associated with the AI / ML model to be trained. Furthermore, for example, a terminal device may determine whether training data can be acquired by comparing the range of parameters required in the AI / ML model with the parameters associated with the environmental identifier notified by the base station device. Alternatively, a corresponding environmental identifier may be identified in advance based on the range of parameters required in the AI / ML model, and the terminal device may determine whether training data can be acquired by determining whether the environmental identifier notified by the base station device matches the pre-identified environmental identifier. For example, when training an AI / ML model in which training data can be acquired when the transmission angle of all beams is 90° ± 10° and the number of beams is 14 to 18, environmental identifiers E3 and E4 in Figure 3 are identified in advance as environmental identifiers corresponding to the parameters for which training data can be acquired. Then, when the terminal device receives environmental identifier E3 (or E4) from the connected base station device, it may determine that training data can be acquired and request the transmission of a reference signal through each of the multiple beams for training the AI / ML model.On the other hand, if the terminal device receives notification from the connected base station device of an environmental identifier different from E3 and E4, such as environmental identifier E1, it will not request the transmission of a reference signal through each of the multiple beams. Also, in situations where the environmental identifiers corresponding to the AI / ML model are predefined, the base station device may provide parameters such as the transmission angle of all beams, the number of beams, beam width, transmission power, and EIRP as information about the beam group. In this case, the terminal device may identify the environmental identifier based on these parameters and determine whether the base station device can provide an environment in which training data for the AI / ML model to be trained can be acquired.

[0021] A terminal device may request the transmission of a reference signal using an RRC message such as a User Equipment (UE) Assistance Information (UAI). Upon receiving this request, the base station device sends a notification to the terminal device to initiate the transmission of the reference signal using an RRC message such as an RRC Reconfiguration message (S206). This notification may include information about the resource block to which the reference signal for each beam is transmitted, as well as parameters such as the transmission period. Some or all of these parameters may be notified to the terminal device from the base station device in S203 as information about the beam group. Furthermore, the parameters transmitted in S203 may be notified again in S206, or they may not be notified in S206. Furthermore, the terminal device may request the base station device to transmit, for example, a transmission start notification, a transmission stop notification (described later), and information specifying the beam to which the reference signal will be transmitted, using a media access control / control element (MAC CE) or uplink control information (UCI). When the terminal device transmits a request using MAC CE or UCI to the base station device, it may specify the beam to which the reference signal will be transmitted using a logical identifier for each beam, an environmental identifier for identifying a group of beams, or other indexes that can identify beams. The base station device, for example, transmits a reference signal in each of the multiple beams that can be formed after transmitting a transmission start notification (S207). Then, based on the parameters obtained from the base station device, the terminal device measures the radio quality of each beam (e.g., reference signal received power (RSRP), reference signal received quality (RSRQ), signal-to-interference and noise ratio (SINR), channel quality indicator (CQI)) based on the reference signal transmitted in each beam, and obtains measurement results that will serve as learning data and training data (S208). The terminal device may request the base station device to stop transmitting the reference signal for acquiring the learning data after the measurement is completed (for example, after the learning is completed) (S209). The base station device will stop transmitting the reference signal for acquiring the learning data in response to the request and will notify the terminal device that the transmission of the reference signal has stopped (S210).Furthermore, the base station equipment may, for example, transmit a reference signal for all beams that can be formed after transmitting the transmission start notification in S206 and before transmitting the transmission stop notification in S210, and transmit a reference signal for only some beams during other periods. This can reduce the power consumption of the network.

[0022] The above describes an example in which a terminal device starts measuring radio quality for acquiring training data upon receiving a transmission start notification and ends the measurement upon receiving a communication stop notification, but it is not limited to this example. For example, a base station device may transmit a reference signal on all of the multiple beams that can be periodically formed (all beams for which radio quality used in training the AI / ML model needs to be acquired) without receiving a request from the terminal device, in which case the transmission start notification in S206 may be omitted. In this case, the transmission start request in S205 may also be omitted. In one example, in S203, parameters such as information specifying radio resources, such as the timing at which the reference signal is transmitted on all beams, may be transmitted from the base station device to the terminal device. Then, in response to the terminal device determining in S204 that training data can be acquired, it may start the measurement based on that information without sending a transmission start request or receiving a transmission start notification. Furthermore, the terminal device may autonomously decide to start or stop the measurement during the period in which the base station device transmits a reference signal on all beams for which radio quality used in training the AI / ML model needs to be acquired. For example, a terminal device may perform measurements only under specific conditions, such as obtaining specific wireless quality or specific terminal movement speed. The terminal device may collect training data or training data, and the AI / ML model may be trained after a certain amount of data has been collected. In this case, the terminal device may temporarily store the training data or training data until a certain amount of data has been collected. In this case, along with the wireless quality information that will serve as the training data or training data, the terminal device may also store information such as the measurement time, the state of the terminal device at the time of measurement (position, speed, acceleration, etc.), the base station device identifier, and the location of the base station device, for the purpose of classifying the AI / ML model.

[0023] Furthermore, if the reference signal mentioned above is CSI-RS, the index assigned to the CSI-RS (for example, nzp-CSI-RS-ResourceId, which is included in the parameter NZP-CSI-RS-Resource notified from the base station device to the terminal device) may be used as the logical identifier assigned to the beam. In the above example, when the terminal device establishes an RRC connection with the base station device, it may notify the base station device that the connection is being established to start the procedure for acquiring training data for the AI / ML model. This would allow for processing such as excluding the communication from billing, by explicitly indicating that it is a communication related to training an AI / ML model.

[0024] Furthermore, if the base station equipment includes multiple transmission / reception points (TRPs), the above processing may be performed for each TRP. In this case, the base station equipment may notify the terminal equipment of parameters related to the beam group (e.g., transmission angle of all beams, number of beams, beam width, transmission power, EIRP, etc.) or environmental identifiers for each TRP. The terminal equipment can then collect training data for the AI / ML model by sending a transmission start request to the base station equipment, for example, including information specifying the environmental identifier and TRPs. Furthermore, if CSI-RS is used for beam transmission for training, the base station equipment may reuse the information storage table format for parameters used to identify conventional CSI-RS resources (e.g., CSI-ResourceConfig or NZP-CSI-RS-Resource). This allows the base station equipment to efficiently notify the terminal equipment of information related to these beams without having to prepare a new information storage table format. Furthermore, while CSI-RS resource settings are normally notified to the terminal device in association with report settings, in the case of learning, measurement data is not reported to the base station device. Therefore, resource settings alone may be notified to the terminal device without being associated with report settings.

[0025] Furthermore, in order to prevent overfitting during AI / ML model training, the acquired training data may be classified into training data and validation data.

[0026] The terminal device performs training on an AI / ML model based on the measurement results in S208. For example, machine learning may be performed using the measurement results of the radio quality of the other beams as training data, in order to train an AI / ML model that estimates the radio quality of other beams by taking the radio quality of some of the beams that the base station device can form as input. Also, if the base station device can form a first beam group with wide beamwidths per beam and a second beam group with narrow beamwidths per beam, machine learning may be performed using the measurement results of the radio quality of each beam in the second beam group as training data, in order to train an AI / ML model that estimates the radio quality of each beam in the second beam group by taking the measured radio quality of each beam in the first beam group as input data. Furthermore, similar training may be performed on an AI / ML model that takes the measured radio quality of some or all of the beams that the base station device can form at a specific timing as input data, and outputs an estimated value of the radio quality of all of the formable beams at a timing predetermined time after that specific timing. In other words, the terminal device measures a first radio quality for each beam that the base station device can form at a specific timing, and also measures a second radio quality for each beam that the base station device can form after a predetermined time has elapsed since that specific timing. The first radio quality can then be used as input, and the second radio quality can be used as training data to train an AI / ML model. The beam width and other characteristics of the beams formed may differ between the beam group when the first radio quality is measured and the beam group when the second radio quality is measured or estimated. Furthermore, an AI / ML model that takes the radio quality of some of the beams that the base station device can form as input and estimates a predetermined number of beams from among those beams based on their radio quality can also be trained by providing training data corresponding to the measurement results. In addition to learning about beam quality, an AI / ML model that estimates the position of the terminal device may be generated, for example, based on the measurement results of some of the beams.In this case, the terminal device can perform machine learning by taking the measurement results for some beams as input, and using the location of its own device, identified, for example using the Global Positioning System (GPS), as training data. The machine learning can be terminated, for example, when the magnitude of the difference between the output data and the training data falls below a predetermined value, or when a predetermined number of learning iterations have been performed.

[0027] Furthermore, AI / ML model training may be performed by an external server (for example, on the internet) rather than by the terminal device. When AI / ML model training (construction) is performed by the server, the terminal device transmits the acquired training data (including training data) to the server via the base station device and receives the AI / ML model as a result of the training from the server. This makes it possible to efficiently generate AI / ML models using a high-performance external server without being limited by the processing power of the terminal device. Regarding the generation of AI / ML models by such an external server, training data is provided from the terminal device to the server, but this data may be treated differently from normal user data. For example, data communication related to the generation of AI / ML models may be subject to special treatment, such as being exempt from billing. In this case, the terminal device may notify the base station device that it is sending and receiving data for the generation of an AI / ML model before sending and receiving the data. An example of the processing flow in this case is shown in Figure 4. First, the terminal device requests the base station device to send training data for the generation and updating of the AI / ML model (S401). The base station device decides to permit the transmission of training data, for example, when the amount of user data traffic is below a predetermined amount, and sends a response containing the details of that decision to the terminal device (S402). The base station device may refuse to transmit training data if the amount of data that should take priority over training data, such as user data, exceeds a predetermined amount. The base station device may also permit the transmission of training data only during predetermined periods, such as at night. The base station device may also notify the terminal device in advance of whether or not the transmission of training data is permitted, and the terminal device may send a training data transmission request to the base station device only if the transmission of training data is permitted. In one example, the base station device may include information indicating whether or not the transmission of training data is permitted in system information such as System Information Block Type 1 (SIB1).For example, the regulatory information of Unified Access Control (UAC) included in SIB1 is used to indicate whether or not a terminal device is permitted to provide a specific service. By defining and notifying the type of training data in this UAC, the terminal device can be notified of whether or not it is permissible to transmit the training model by utilizing the functions of conventional outgoing transmission restrictions. Note that the training data may be treated as ordinary user data, in which case the terminal device does not need to make any special inquiries to the base station device, and S401 and S402 (and S408 and S409 described later) may be omitted.

[0028] When the terminal device receives permission to transmit training data from the base station device, it then sends a training data transmission request to the server to notify the server of the transmission of training data (S403). When the server receives the training data transmission request, it determines whether to accept the request, and if it accepts, it sends permission to transmit to the terminal device (S404). The server may, for example, determine whether it can perform training depending on the load status, accept the request if training is possible, and reject the request if training is not possible. If the request is accepted, the terminal device transmits the training data to the server (S405). The server then performs machine learning using the acquired training data and generates and updates the AI / ML model (S406). The server may be configured to always be able to receive training data, in which case the processes in S403 and S404 may be omitted. When the terminal device transmits training data to the server, it may also notify information to identify the type of training data. Information for identifying the type of training data may, for example, indicate the type of AI / ML model constructed from the training data, such as the estimation of beam radio quality or the estimation of the terminal device's location. Furthermore, information for identifying the type of training data related to the estimation of beam radio quality may include detailed information such as whether it is training data for a first model that outputs estimated radio quality values ​​for other beams based on measurement results for some beams, training data for a second model that outputs estimated radio quality values ​​for a narrow beam group from measurement results for a wide beam group, or training data for a third model that outputs estimated radio quality values ​​after a predetermined time from measurement results at a predetermined timing. This information may also include information indicating the environment in which the training data was acquired. This information indicating the environment may be information about the beam group provided by the base station device in S203 described above. In one example, the above-mentioned environment identifier may be used as this information indicating the environment. That is, the terminal device may transmit information about the environment identifier at the time of acquisition of the training data to the server along with the training data.Since the server can collect learning data from a plurality of terminal devices, the server can identify learning data acquired by the plurality of terminal devices in a common environment and perform machine learning.

[0029] After completing the generation and update of the AI / ML model, the server notifies the terminal device that the model can be transferred (S407). Then, the terminal device requests the base station device to perform communication for acquiring the model from the server (S408), and when permission is received (S409), requests the server to transfer the AI / ML model (S410). Note that, similarly to the case of S402, the base station device can determine whether to permit the transfer of the AI / ML model, and transmit information indicating the determination to the terminal device. When receiving a model transfer request, the server transmits the generated or updated AI / ML model to the terminal device (S411). Note that when providing the AI / ML model, the server can notify the terminal device of information of the AI / ML model (such as input target data, output target data, and information (environment identifier) related to an acquisition environment of learning data). An example of this information is shown in FIG. 5. This information can be stored, for example, in association with a model identifier. In FIG. 5, x k represents the radio quality of a beam assigned with an index k (1≤k≤the number of beams), and the data acquisition environment is represented by an environment identifier. The terminal device can determine whether to use the AI / ML model (which AI / ML model to use when a plurality of AI / ML models are generated) based on the information of the AI / ML model. In this way, when the server generates or updates an AI / ML model using data collected by the terminal device, and the terminal device acquires the AI / ML model, special control such as charging exemption can be performed on communication related to the AI / ML model.

[0030] Furthermore, in the above description, the processing related to the learning phase of the AI / ML model has been described. Similarly, a test may be performed to determine whether the generated AI / ML model is in an appropriate state (whether an update is necessary or not). That is, the terminal device determines, based on the information related to the beam group provided by the base station device, whether test data suitable for the AI / ML model to be tested can be acquired. Then, when such test data can be acquired, the terminal device measures reference signals transmitted in each of a plurality of beams from the base station device. The terminal device uses a part of the measurement results as input data to the AI / ML model, and acquires estimated values of radio quality for other beams for which no input data is provided as output data. Here, since the terminal device has acquired measurement results for the other beams, it uses the measurement results as correct data and compares the correct data with the output data. Then, when the magnitude of the difference is equal to or less than a predetermined value, the terminal device can determine that the AI / ML model is operating properly. On the other hand, when the magnitude of the difference between the output data and the correct data exceeds a predetermined value, the terminal device can determine that an update of the AI / ML model is necessary, and can execute an update process (can re-execute the processing of the learning phase).

[0031] (AI / ML model generation processing) Next, an outline of processing when an AI / ML model is generated from learning data will be described. For example, the radio quality x2 of beam 2 is expressed by the formula x2 = a1x1 + a3x3 + a5x5... + a 15 x 15 and the radio quality x4 of beam 4 is expressed by the formula x4 = b1x1 + b3x3 + b5x5... + b 15 x 15 It is assumed that the first AI / ML model is generated as a linear model estimated respectively by the above formulas. In this case, the coefficients a1, a2,... a 15 , b1, b2,... b 15 are determined from the learning data. Similarly, for the radio quality x6, x8,... x 16 of beams 6, 8,... 16, the coefficients in the above formulas are also determined by machine learning based on the learning data. Here, x1 to x 16If the measurement or estimation of radio quality in the beam up to CSI-RS is performed, then the radio quality x1 to x 16 This represents the measured or estimated values ​​of the radio quality for CSI-RS from CSI-RS resource index = 1 to 16. Furthermore, a second AI / ML model different from the first AI / ML model described above may be generated using the same training data. For example, suppose the second AI / ML model is a linear model that estimates the radio quality of 12 beams using the radio quality of 4 beams as input. In this case, equation x² = c1x¹ + c5x⁵ + c9x⁹ + c 13 x 13 , x3=d1x1+d5x5+d9x9+d 13 x 13 The second AI / ML model is defined in a format such as the following. Then the coefficients c1, c5, c9, c 13 d1, d5, d9, d 13 ... are determined by machine learning based on training data. Furthermore, using the same training data, a third AI / ML model different from the first and second AI / ML models described above can be generated, which takes the radio quality of four beams as input and estimates the single beam with the best radio quality. In the third AI / ML model, for example, radio quality x1, x5, x9, and x 13 From the measured values, the measured or estimated values ​​of wireless quality are x1 to x 16 The system then estimates which of the beams is the best. For example, a decision tree is built from training data, and by inputting measured values ​​into the constructed decision tree, the single beam with the best wireless quality is estimated.

[0032] In the above example, we described an example of a base station device capable of providing 16 beams, but the number of beams that a base station device can form may be other numbers, such as 8 or 32. Also, in the example above, we showed cases where the measured values ​​of radio quality input to the AI / ML model are 8 or 4, but this number may be more or less, and accordingly, the number of beams for which radio quality is estimated is not limited to 8 or 12. Furthermore, the estimated values ​​by the AI / ML model are not limited to the above example. For example, instead of the third AI / ML model described above, an AI / ML model may be formed in which two or four beams are estimated from the one with the best radio quality. Alternatively, without using a decision tree, the estimated values ​​of the radio quality of the beams may be identified, and then a predetermined number of beams may be estimated from the one with the best radio quality by sorting the estimated values ​​and measured values. In addition, in the above example, we described an example where the AI / ML model is a linear model, but a nonlinear model may also be used.

[0033] After an AI / ML model is generated, its generalization performance can be verified using test data. In the example above, for the first and second AI / ML models, the generalization performance can be evaluated based on the error between the estimated value and the measured value. If this error falls below a predetermined threshold, the AI / ML model is determined to be operational. For the third AI / ML model described above, the generalization performance can be evaluated by dividing the number of correct answers obtained by a predetermined number of estimations by that predetermined number (the accuracy rate of the beam with the best wireless quality). If the accuracy rate exceeds a predetermined threshold, the AI / ML model is determined to be operational. An AI / ML model determined to be operational is set up for use in a terminal device (for example, if a server generates an AI / ML model, the server provides that AI / ML model to the terminal device), and its use is initiated, for example, in response to instructions from a base station device.

[0034] (Determination of Applicability of AI / ML Models) As a method for determining whether a particular AI / ML model is applicable to a terminal device, it is conceivable to use an identifier (for example, model identifiers M1 to M5 shown in Figure 5) to identify that particular AI / ML model. In this method, the base station device can identify one or more AI / ML models that can be applied to the terminal device based on the characteristics of the beams that it can form, and notify the terminal device of a list of identifiers for the identified AI / ML models. For example, if the base station device transmits a reference signal using only four of the beams that can be formed, it may determine that it cannot use an AI / ML model that requires eight beams of radio quality as input. For this reason, it may identify an AI / ML model that only requires four beams of radio quality as input and notify the terminal device of a list of its identifiers. If the terminal device finds that the notified list of AI / ML model identifiers includes an identifier corresponding to an AI / ML model it holds, it may report the identifier of the AI / ML model it holds to the base station device. In one example, the base station device sends an instruction to the terminal device to use the reported AI / ML model. This allows the terminal device to apply the AI / ML model and perform efficient communication using estimation.

[0035] The method using identifiers described above assumes that there is common knowledge about AI / ML models between the base station equipment and the terminal equipment, and is particularly effective when only a relatively small number of AI / ML models are available. On the other hand, since it is assumed that there are many types of AI / ML models, and furthermore, that AI / ML models are updated over time, it may not be easy to share common knowledge about AI / ML models between the base station equipment and the terminal equipment. For this reason, the following provides a method for determining whether an AI / ML model is applicable to a terminal equipment without using identifiers to identify the AI / ML model. For example, the base station equipment may provide parameters related to the AI / ML model to the terminal equipment, and the terminal equipment may determine whether it can apply the AI / ML model it holds based on the received parameters.

[0036] The base station device provides the terminal device with information indicating the communication environment it provides, for example, as a parameter for this determination. In one example, the base station device may notify the terminal device of an environmental identifier (as shown in Figure 3). The base station device may also provide the terminal device with information regarding the beam to be measured (the beam from which measured values ​​such as radio quality, which are inputs to the AI / ML model, are to be acquired) and information regarding the beam to be estimated (the beam from which estimated values ​​such as radio quality, which are outputs to the AI / ML model, are to be acquired). Furthermore, the base station device may provide the terminal device with setting information for reporting measured and estimated values. These pieces of information may also be combined and provided from the base station device to the terminal device. In one example, the base station device may provide an environmental identifier (e.g., "E3" as shown in Figure 3), the beam to be measured (e.g., radio quality x1, x3, ... x 15 (8 beams corresponding to the same beams) and the beam to be estimated (e.g., wireless quality x2, x4, ... x) 16 Information regarding the corresponding 8 beams, along with configuration information for reporting measured and predicted values, can be combined and notified to the terminal device. The information notified here may be referred to as the inference decision parameter below. In addition, for example, an AI / ML model may be associated with a type, and the base station device may notify the terminal device of information indicating the type of AI / ML model in addition to the inference decision parameter. In this case, the terminal device can determine whether it has an applicable AI / ML model from among the AI / ML models it holds that corresponds to the notified type, based on the inference decision parameter.

[0037] The environmental identifier notified here is identified based on the beam formed by the base station equipment. That is, the base station equipment may notify the terminal equipment of the environmental identifier corresponding to the beam formed by its own equipment as an inference decision parameter. Based on the received information, the terminal equipment determines whether or not it can apply the AI / ML model it holds. For example, if the terminal equipment holds an AI / ML model that corresponds only to environmental identifier = E1, and the base station equipment notifies it of environmental identifier = E3, it may determine that the AI / ML model cannot be applied. Also, if the terminal equipment holds an AI / ML model that corresponds to environmental identifier = E3, and the base station equipment notifies it of environmental identifier = E3, it may determine that the AI / ML model can be applied. In this case, the terminal equipment may determine whether the environment is actually suitable for applying the AI / ML model based on further conditions, such as those described later, or it may decide to apply the AI / ML model without considering such conditions.

[0038] Information about the beam to be measured is, for example, the wireless quality x1, x3, ... x used as input in Model M1 in Figure 5. 15 This is information that identifies the beams to be acquired. For example, this could be information about the CSI-RS used to acquire the radio quality of these beams (e.g., transmission period, location of resource blocks used for transmission, etc.). In addition, information about the beams to be estimated could be, for example, the radio quality x2, x4, ... x, which are considered to be the output in Model M1 in Figure 5. 16This information identifies the beams to be acquired. For example, it may be information about the assumed CSI-RS for these beams (e.g., assumed transmission period, location of resource blocks used for transmission). Note that even if CSI-RS is not actually transmitted for the beams to be estimated, the AI / ML model may output the radio quality assuming that CSI-RS was measured at a specific transmission period or resource block location. This information prevents, for example, the CSI-RS being treated as an estimate for a resource block location far away from the actual location, even though the estimate is based on transmission at a specific resource block location. Furthermore, when a terminal device needs to obtain an estimate for a specific frequency range, it can use this CSI-RS information to check whether the resource block location of the CSI-RS deviates significantly from the frequency range for which the estimate should be obtained, thereby determining whether the AI / ML model can obtain the necessary estimate. Furthermore, information on the transmission period can be used, for example, to determine whether the timing difference between obtaining a measurement value for the beam being measured and obtaining an estimated value for the beam being estimated is suitable for the environment in which the AI / ML model is applied. For example, in an environment where the AI / ML model is applied, an estimated value is required at time Δt after the measurement value is obtained, and based on the CSI-RS transmission period information for both the beam being measured and the beam being estimated, the time from obtaining the measurement value to obtaining the estimation value is αΔt. In this case, if the coefficient α falls within a predetermined range close to 1, the AI / ML model is determined to be applicable, and if the coefficient α does not fall within that predetermined range, the AI / ML model may be determined to be unapplicable. In addition, the information on the beam being measured and the information on the beam being estimated may include information for relating the respective beams being measured and estimated to the input and output values ​​of the AI / ML model, respectively, in the inference environment. For example, in the AI / ML model M3 (4 input beams, 12 prediction beams) in Figure 5, the inference is x² = a¹x¹ + a⁵x⁵ + a⁹x⁹ + a⁹x⁹ + a 13x 13Such formulas may be used. Information to identify which beam corresponds to each of the radio quality x1 and radio quality x2 in the inference environment may be notified from the base station equipment to the terminal equipment. For example, information indicating that CSI-RS index=1 corresponds to the beam associated with radio quality x1, and CSI-RS index=2 corresponds to the beam associated with radio quality x2 may be notified from the base station equipment to the terminal equipment. Alternatively, TCI=1 may be notified from the base station equipment to the terminal equipment as information about the beam associated with radio quality x1, and TCI=2 as information about the beam associated with radio quality x2. The base station equipment can efficiently notify the terminal equipment of this beam information without having to prepare a new information storage table format by reusing the format of the parameter information storage table used for identifying conventional CSI-RS resources. For example, this can be notified to the terminal equipment using NZP-CSI-RS-Resource. In other words, the NZP-CSI-RS-Resource of the beam to be measured and the NZP-CSI-RS-Resource of the beam to be estimated are notified to the terminal device from the base station device. The terminal device can use the CSI-RS-ResourceMapping contained in each NZP-CSI-RS-Resource to identify the location of the resource block of the beam to be measured and the beam to be estimated, and use periodicityAndOffset to identify the transmission timing of the reference signal for each beam. In one example, the CSI-RS-ResourceMapping may include information for identifying the resource block to which CSI-RS is transmitted, such as information for identifying the time domain location of the resource block, information for identifying the frequency domain location of the resource block, and information for identifying the number of resource blocks used.In Figure 6A, No. CDM (Code Division Multiplexing) may be specified by CSI-RS-ResourceMapping for identifying the time domain position = 2, the frequency domain positions = 1 and 5, and the number of resource blocks used = 1. In Figure 6B, FD (Frequency Domain)-CDM2 may be specified by CSI-RS-ResourceMapping for identifying the time domain position = 1, the frequency domain positions = 1 and 5, and the number of resource blocks used = 2 (multiplexed and linked in the frequency direction). In Figure 6C, CDM4 may be specified by CSI-RS-ResourceMapping for identifying the time domain position = 3, the frequency domain positions = 1 and 5, and the number of resource blocks used = 4. Furthermore, the terminal device can use all or part of the above-mentioned information for identifying resource blocks to determine, for example, whether the frequency domain position falls within a specific range, and whether the estimated value deviates significantly from the frequency range from which it should be obtained.

[0039] The configuration information for reporting measured and estimated values ​​includes, for example, period and offset information to specify the timing for reporting measured or estimated wireless quality values ​​to the base station equipment. For example, this configuration information specifies that measured and estimated values ​​should be reported at time T + ΔT based on the measured value at time T, and if the AI / ML model available to the terminal device requires more than ΔT for estimation, it cannot report the estimated value. Therefore, if such reporting configuration information is notified, the terminal device may determine that it cannot apply the AI / ML model it holds. Also, if configuration information requesting reporting at a time interval of 2ΔT is notified, the terminal device may determine that an AI / ML model that can be completed within 2ΔT for estimation is applicable, and that an AI / ML model that requires more than 2ΔT for estimation is not applicable. The configuration information for reporting can also reuse the format of the parameter information storage table used for reporting settings of conventional CSI-RS resources. This allows the base station equipment to efficiently notify the terminal device of the reporting configuration information without having to prepare a new information storage table format. Configuration information for reporting is notified to the terminal device using, for example, CSI-ReportConfig. For example, the CSI-ResourceConfigId included in CSI-ReportConfig specifies whether it is a beam to be measured or a beam to be estimated, and the reportConfigType specifies the type of report (periodic, semiPersistent, aperiodic).

[0040] The terminal device notifies the base station device of the result of its determination of whether or not the AI / ML model it holds is applicable based on the received inference decision parameters. The base station device may provide the terminal device with multiple inference decision parameters. In this case, a unique identifier may be assigned to each of the multiple inference decision parameters. The terminal device can then determine whether or not the AI / ML model it holds is applicable to each of the multiple inference decision parameters. The terminal device may then notify the base station device of information (e.g., an identifier) ​​that specifies the inference decision parameter that holds the corresponding AI / ML model. For example, a bitmap indicating whether or not the terminal device holds the corresponding AI / ML model for each of the multiple inference decision parameters may be transmitted from the terminal device to the base station device. For example, if four inference decision parameters are notified, a four-bit bitmap may be used, where the first bit is set to the inference decision parameter with the smallest identifier, the second bit to the inference decision parameter with the second smallest identifier, the third bit to the inference decision parameter with the third smallest identifier, and the fourth bit to the inference decision parameter with the largest identifier. For example, each bit can indicate that there is an applicable AI / ML model ("1") or that there is no applicable AI / ML model ("0"), respectively (conversely, "0" may indicate that there is an applicable AI / ML model). As an example, if four inference decision parameters, each assigned identifiers from 1 to 4, are provided to a terminal device, and the terminal device holds AI / ML models corresponding to the second and fourth inference decision parameters, the terminal device may notify the base station device of a bitmap named "0101". The size of the bitmap may be set to be variable depending on the number of inference decision parameters notified by the base station device to the terminal device. Here, information indicating the size of the bitmap may be notified from the terminal device to the base station device, or, since the base station device knows the number of inference decision parameters provided, such information indicating the size does not need to be transmitted.

[0041] The determination of model applicability based on inference determination parameters is performed for each base station device. When a terminal device changes the connected base station device due to a handover, the inference determination parameters provided by the handover destination base station device may be provided to the terminal device by the handover destination base station device via an RRC Reconfiguration message or the like. In one example, the handover destination base station device obtains the inference determination parameters from the handover destination base station device via a HANDOVER REQUEST ACKNOWLEDGE message and provides those parameters to the terminal device. The terminal device determines whether it holds the AI / ML model corresponding to those inference determination parameters and may notify the handover destination base station device of the determination result via an RRC Reconfiguration complete message upon completion of the handover or other messages after connection. In the case of conditional handover (CHO) or Layer 1 / Layer 2 Triggered Mobility (LTM), the terminal device receives inference decision parameters provided by the receiving base station from the handover base station and notifies the handover base station of the decision result before executing CHO or LTM. The handover base station can then notify the receiving base station of the decision result. In this way, the base station can determine whether the terminal device can communicate by applying the AI / ML model without sharing knowledge of the AI / ML model with the terminal device.

[0042] (Estimation using AI / ML models) When the base station device receives a report that a terminal device can apply an AI / ML model, it may instruct the terminal device to perform inference using that AI / ML model. In this case, the base station device transmits an RRC Reconfiguration message, MAC CE, or Downlink Control Information (DCI) containing information specifying the AI / ML model to be applied by the terminal device. The information specifying the AI / ML model to be applied may be information specifying the inference decision parameters corresponding to that AI / ML model. After transmitting the notification, the base station device transmits a reference signal in the beam under measurement to enable the terminal device to acquire the radio quality of the input data for inference. The terminal device measures the radio quality based on the transmitted reference signal and inputs the measured value into the specified AI / ML model to obtain an estimate of the radio quality for the beam under estimation. Prior to notifying the start of AI / ML model inference, the base station equipment may notify the terminal equipment of information regarding the period during which AI / ML model inference will be performed and the period during which inference will not be performed. In this case, the base station equipment may not transmit a reference signal over the beam to be estimated during the inference period, and may transmit a reference signal over the beam to be estimated during the period during which inference will not be performed. During the period during which inference will not be performed, the terminal equipment may also measure the radio quality of the beam to be estimated based on the reference signal. The terminal equipment may then acquire the measured radio quality for the beam to be estimated, perform estimation using the AI / ML model in parallel, and calculate the error between the measured value and the estimated value. The terminal equipment may then notify the base station equipment (or the server that generates or updates the AI / ML model) of this error information. The terminal equipment may also report to the base station equipment at least one of the estimated or measured values ​​of the radio quality (e.g., L1 (Layer 1) - RSRP) for the beam to be estimated. In this case, the report may be made in a format that allows the base station equipment to identify whether the reported value is an estimated value or a measured value. Furthermore, at least one of the estimated and measured values ​​of wireless quality, along with the error, may be reported to the base station equipment.When the base station equipment receives estimated and measured values ​​of wireless quality, it calculates the error by comparing the estimated and measured values. If the error becomes large, the base station equipment may decide to terminate the prediction by the AI / ML model in the terminal equipment or change the AI / ML model being used.

[0043] Conventional CSI reports can be reused for reporting estimated and measured values ​​of wireless quality. Roughly speaking, the settings used for conventional beam measurement reports are performed by a combination of reference signal resource settings and report settings, but the settings used for measurement reports for the beam to be estimated can also be notified from the base station equipment to the terminal equipment in a similar manner. That is, resource settings are given by CSI-ResourceConfig, and report settings are specified by CSI-ReportConfig. Resources are then set for both the beam to be measured and the beam to be estimated. In one example, ResourceID=1 may be assigned to the beam to be measured, and ResourceID=2 may be assigned to the beam to be estimated. In addition, separate setting items may be provided, such as CSI-ResourceConfig for the beam to be measured and CSI-ResourceConfig-Inference for the beam to be estimated. Report settings can be specified by CSI-ReportConfig. The reporting period and timing of measurement results for the beam being measured and the beam being estimated can be set by CSI-ReportConfig, respectively. For example, ResourceID=1 may be associated with the CSI-ReportConfig for the beam being measured, and ResourceID=2 may be associated with the CSI-ReportConfig for the beam being estimated. The CSI reporting method is specified by CSI-ReportConfig included in the RRC message, and there are three types: Periodic, Semi-persistent, and Aperiodic. Examples of processing when each method is used are described below. In the following section, we will describe an example in which a terminal device uses an AI / ML model that takes measured values ​​of the radio quality for a portion of the beam formed by the base station device as input and outputs estimated values ​​of the radio quality for the other beams.This is just one example; the terminal device may use an AI / ML model that takes measured radio quality values ​​for some beams as input and outputs information indicating a predetermined number of beams with good radio quality from among the beams formed by the base station device, or it may use an AI / ML model that estimates the position of the terminal device, for example.

[0044] • Periodic report The location of the radio resource (resource block) from which the reference signal is transmitted in the beam being measured, and the location of the radio resource (resource block) from which the reference signal is assumed to be transmitted in the beam being estimated, are provided to the terminal device, for example, in Resource Configuration or CSI-ResourceConfig. Here, the base station device transmits the reference signal in the beam being measured, but does not transmit the reference signal in the beam being estimated for at least a portion of the time. In one example, the reference signal in the beam being estimated is transmitted only when it is necessary to compare the predicted value with the measured value. The CSI-ResourceConfig of the beam being estimated may include information to identify the periods during which the reference signal is not transmitted and the periods during which the reference signal is transmitted. In one example, the CSI-ResourceConfig of the beam to be estimated may be associated with ResourceID=2 for periods when no reference signal is transmitted and ResourceID=3 for periods when a reference signal is transmitted. Based on this information, the terminal device identifies the period during which it measures the reference signal actually transmitted in the beam to be estimated during the period when a reference signal is transmitted. Information regarding the location of resource blocks related to the beam to be measured and the beam to be estimated may be provided to the terminal device in the procedure for determining the applicability of the AI / ML model described above. In this case, this information does not need to be provided to the terminal device again; only the start of reference signal transmission in the beam to be measured may be notified to the terminal device. That is, the terminal device can recognize the start of reference signal transmission in the beam to be measured by receiving the inference start notification. The base station device may start transmitting the reference signal in the beam to be measured prior to sending the inference start notification. Upon receiving notification to start inference, the terminal device begins measuring the reference signal transmitted in the beam to be measured, based on the CSI-ResourceConfig for the beam to be measured that the terminal device holds. The terminal device also begins estimating the radio quality in the beam to be estimated by inputting the measured radio quality values ​​in the beam to be measured into the AI / ML model (for example, in parallel).Furthermore, if the terminal device was instructed to report the measurement results of the radio quality before the inference start notification (for example, if the setting and instruction for reporting the CSI-RS measurement results were made while estimation was not being performed), it does not need to start a new measurement in response to the inference start notification. In other words, the terminal device can continue the measurement it was performing up to that point. Also, the base station device may notify the terminal device to stop the inference while the terminal device is performing inference. When the terminal device receives the inference stop notification, it stops the inference of the radio quality in the beam being estimated. The terminal device may also stop measuring the radio quality in the beam being measured when it stops the inference. Alternatively, when the terminal device receives the inference stop instruction, it may receive predetermined information from the base station device to determine whether or not to stop the measurement in the beam being measured, and may decide to stop the measurement in the beam being measured according to that predetermined information. In one example, if the terminal device is instructed to continue reporting the measurement results when it receives the inference stop instruction, it may continue measuring the beam being measured. Furthermore, the terminal device may continue measuring and estimating the radio quality when the transmission of a reference signal in the beam being estimated is started in response to a command to stop the radio quality estimation, in order to identify the error between the estimated value and the measured value of the radio quality for the beam being estimated.

[0045] - Semi-persistent report: Base station equipment can instruct terminal equipment to start and stop reporting (also called activate and deactivate) about the beam to be estimated, using MAC CE (e.g., SP CSI reporting on PUCCH Activation / Deactivation MAC CE) or DCI. In the case of semi-persistent reporting, unlike the periodic reporting described above, it is sufficient for inference about the beam to be estimated to be performed only during the period for which reporting is required (the period from when the start is instructed until when the stop is instructed). Therefore, the base station device may, for example, notify the terminal device of the Semi-periodic reporting setting in the CSI-ResourceConfig of the beam to be measured when notifying the start of inference, and then instruct the terminal device to start transmitting the reference signal for the beam to be measured prior to the start of reporting for the beam to be estimated. For example, the base station device may transmit this instruction using SP CSI-RS / CSI-IM Resource Set Activation MAC CE. In this case, the base station device may also notify the terminal device to stop transmitting the beam to be measured after the end of the period for reporting on the beam to be estimated. The base station device may transmit this notification using SP CSI-RS / CSI-IM Resource Set Deactivation MAC CE. If the beam being measured is always transmitted periodically, the transmission of the beam being measured will begin from the moment the Periodic reporting setting is notified in CSI-ResourceConfig, similar to the case of Periodic reports. Therefore, it is not necessary for the base station equipment to instruct the terminal equipment to stop transmitting the beam being measured each time.

[0046] - The Aperiodic base station device, for example, notifies the Aperiodic reporting settings in the CSI-ResourceConfig of the beam to be measured when notifying the start of inference. The terminal device, upon receiving the inference start notification, performs measurements on the beam to be measured for a specified period, obtains an estimated value of the radio quality of the beam to be estimated by inference using the measured values, and reports this estimated value to the base station device. In the case of Aperiodic, information regarding the location of resource blocks of the beam to be measured and the beam to be estimated may be provided in the procedure for determining the applicability of the AI / ML model described above, or it may be notified when the inference start notification is issued, similar to the case of Aperiodic.

[0047] Figure 7 shows an example of the processing flow when using an AI / ML model, from determining whether an AI / ML model is applicable based on inference determination parameters. Note that the details of the processing in Figure 7 are as described above, so they will not be repeated here. First, the base station device provides the terminal device with inference determination parameters (S701). The terminal device then determines whether the AI / ML model it holds is applicable based on these inference determination parameters (S702), and notifies the base station device of the determination result (S703). If the base station device has an usable AI / ML model, it instructs the terminal device to start inference using that AI / ML model (S704), and the terminal device performs inference according to that instruction (S705). While performing inference, the terminal device may provide, for example, error information to the base station device (S706). Subsequently, if the base station device decides to terminate the inference using the AI / ML model in the terminal device, it sends an instruction to terminate the inference to the terminal device (S707), and the terminal device terminates the inference according to that instruction.

[0048] (Modification 1) As described above, when a base station device is capable of forming a first beam group with wide beamwidths and a second beam group with narrow beamwidths, an AI / ML model can be used to estimate the radio quality of each beam in the second beam group, using the measured radio quality of each beam in the first beam group as input data. It is also conceivable that an AI / ML model can be used to estimate a predetermined number of beams in the second beam group from those with better radio quality, using the measured radio quality of each beam in the first beam group as input data. In this case, the environment for the beam being measured and the environment for the beam being estimated will be different. For example, a base station device can form multiple beams (a first beam group with wide beamwidths) identified by the environmental identifier Ea and transmit a reference signal, and form multiple beams (a second beam group with narrow beamwidths) identified by the environmental identifier Eb and perform data communication with a terminal device. In this case, the two environmental identifiers Ea and Eb can be notified from the base station device to the terminal device.

[0049] For example, the wireless quality in the first beam group is x k toshi, y l = Σ k a k,l x k The wireless quality y in the second beam group is as shown above. l This can be estimated. Note that although a linear model is shown as an example here, a nonlinear model may also be applied. Here, the coefficient a k,lHowever, this is a coefficient determined based on the training data. In order to generate such an AI / ML model, it is necessary to acquire training data in a way that associates the first beam group to be measured with the second beam group to be estimated. When acquiring training data, the base station device notifies the terminal device of information about the beam groups that can be transmitted. For example, the base station device notifies the terminal device of information about the beam groups that can be formed. The terminal device notifies the base station device of information about the beam groups (the first beam group and the second beam group) from among these beam groups on which the reference signal should be observed for the generation of the AI / ML model. The base station device notifies the terminal device of information about the two beam groups and setting information of the radio resources to which the reference signal is transmitted in each beam group (for example, the period and frequency position on which the reference signal is transmitted). Then, the base station device starts transmitting the reference signal in each beam included in those beam groups according to the setting information. When CSI-RS is used as the reference signal, the base station equipment notifies the terminal equipment of a list of nzp-CSI-RS-Resources specifying the CSI-RS resources and their periods for the beam to be measured, and a list of nzp-CSI-RS-Resources specifying the CSI-RS resources and their periods for the beam to be estimated. Note that SSB may be transmitted as the reference signal for the first beam group, CSI-RS for the second beam group, or the same reference signal, such as CSI-RS, may be transmitted for both beam groups. The terminal equipment measures the reference signal according to the list of nzp-CSI-RS-Resources, which is the setting information for the beam to be measured, and the list of nzp-CSI-RS-Resources, which is the setting information for the beam to be estimated, and records the measured value as training data. It is assumed that the base station equipment cannot transmit the reference signal for the first beam group and the second beam group simultaneously. For this reason, as an example, the base station equipment can alternately transmit reference signals in the first beam group and the second beam group. The terminal equipment measures the alternately transmitted reference signals of the first beam group and the second beam group, respectively, and records them in association as input data and output data (training data) pairs.Furthermore, in some cases, a reference signal can be transmitted simultaneously by combining the beams of the first beam group and the beams of the second beam group. In this case, the base station equipment may transmit the reference signal simultaneously for those beam combinations. In this case, the terminal equipment can measure each of the simultaneously transmitted reference signals. Also, when the terminal equipment saves training data, it may record the measurement time and the state of the terminal equipment at the time of measurement (position, velocity, acceleration, etc.) along with the wireless quality. When the terminal equipment is stationary, the positional relationship between the terminal equipment and the base station equipment does not change between the timing of input data acquisition and the timing of output data acquisition. In contrast, when the terminal equipment is moving, the positional relationship between the terminal equipment and the base station equipment may change between the timing of input data acquisition and the timing of output data acquisition. In this case, the movement speed of the terminal equipment may be taken into consideration in the AI / ML model.

[0050] As mentioned above, the training data may be transferred to a server on the internet, and the AI / ML model may be generated and updated on the server. When the training data is transferred to the server, information about the environment and conditions under which the training data was acquired is also sent to the server along with the training data. For example, each beam information and its transmission period are sent to the server in combination with the training data. The server may aggregate training data acquired in the same or similar environments (where some aspects of the environment are the same) to generate an AI / ML model.

[0051] Prior to starting inference, the terminal device determines, as described above, whether the AI / ML model it holds is applicable based on the inference determination parameters. The base station device notifies the terminal device of the inference determination parameters for this determination. As described above, in an AI / ML model that estimates the radio quality in the second beam group based on the radio quality in each beam of the first beam group, a different environmental identifier is associated with each beam group. For this reason, the base station device may notify the terminal device of, for example, the environmental identifier for each beam group, information to identify the radio resources and transmission timing related to the reference signal transmitted by the beams included in each beam group, and setting information for reporting measured and estimated values ​​as inference determination parameters. The terminal device may determine whether inference is possible using the AI / ML model it holds, based on other factors as needed (e.g., conditions related to the beam to be measured and the beam to be estimated), assuming that the input data acquisition environment for the AI / ML model corresponds to the first beam group and the output data acquisition environment corresponds to the second beam group. Thus, even when determining whether different AI / ML models are applicable to the beams to be measured and the beams to be estimated, multiple inference decision parameters may be notified from the base station device to the terminal device. In this case, as described above, the terminal device may notify the base station device of information indicating which of the multiple inference decision parameters it holds an AI / ML model that is applicable to.

[0052] When the base station device receives a report indicating that a terminal device possesses an AI / ML model applicable to the inference decision parameters, it instructs the terminal device to perform inference using that AI / ML model. Based on this instruction, the terminal device then performs inference using the AI / ML model it holds. In this case, the terminal device can compare the estimated value of the radio quality in the second beam group, obtained by inputting the measured radio quality of the first beam group into the AI / ML model, with the measured radio quality of the second beam group. Therefore, the base station device notifies the terminal device of information regarding the period T1 during which a reference signal is transmitted in the first beam group and the period T2 during which a reference signal is transmitted in the second beam group (for example, the period immediately following period T1). The terminal device then measures the radio quality of the first beam group during period T1 and the radio quality of the second beam group during T2, according to the received information. The terminal device inputs the radio quality measured in period T1 into an AI / ML model to obtain an estimated value of the radio quality for the second beam group output by the AI / ML model, and can compare the measured value of the radio quality for the second beam group measured in period T2 with that estimated value. Note that if the terminal device is stationary, the positional relationship between the terminal device and the base station device does not change between period T1 and period T2, but if the terminal device is moving, that positional relationship changes. For this reason, if the terminal device is moving, it may use an AI / ML model that obtains estimated values ​​of the radio quality for the second beam group in period T1 and period T2 from the measured value of the radio quality for the first beam group in period T1. Alternatively, in one example, the terminal device may obtain an estimated value of the radio quality for the second beam group in period T1 from the measured value of the radio quality for the first beam group in period T0 immediately preceding period T1, and also obtain an estimated value of the radio quality for the second beam group in period T2 from the measured value of the radio quality for the first beam group in period T1. The terminal device may also notify the base station device of estimated values ​​of the radio quality in the second beam group for periods T1 and T2.

[0053] (Modification 2) Consider an AI / ML model that takes measured values ​​of radio quality at a specific timing in some or all of the beams that a base station device can form as input data, and outputs an estimated value of the radio quality in all of the beams that can be formed at a predetermined time after that specific timing. In this AI / ML model, the location of the resource block to which the reference signal is transmitted is common to the beam to be measured and the beam to be estimated, and no reference signal is transmitted in the beam to be measured for a predetermined time interval, and the radio quality for that time interval is estimated by the AI / ML model held by the terminal device. In this AI / ML model, for example, the radio quality x of the k-th beam at time t k Using (t), the radio quality x of the l-th beam at time t + Δt. l (t + Δt) is x l (t + Δt) = Σ k a k,l x k (t) is estimated. Note that this AI / ML model is a linear model, but a nonlinear model may also be generated. When generating the AI / ML model, the training data is used to determine the variable a k,l This is determined. Note that variable a k,l This coefficient is multiplied by the measured wireless quality of the k-th beam at time t when estimating the wireless quality of the l-th beam at time t + Δt. Note that the AI / ML model may be constructed differently for each beam, each terminal device's movement speed, and each environment (distinguished, for example, by an environmental identifier), or a single model applicable to all cases may be constructed.

[0054] (Device Configuration) Next, an example of the configuration of a base station device and a terminal device will be described. Figure 8 shows an example of the hardware configuration of a base station device and a terminal device. In one example, the base station device and terminal device are configured to include a processor 801, ROM 802, RAM 803, storage device 804, and communication circuit 805. The processor 801 is a computer configured to include one or more processing circuits, such as a general-purpose CPU (Central Processing Unit) or ASIC (Application-Specific Integrated Circuit), and executes the overall control processing of the device and the above-mentioned processing by reading and executing programs stored in the ROM 802 and storage device 804. The ROM 802 is a read-only memory that stores information such as programs and various parameters related to the processing performed by the base station device and terminal device. The RAM 803 functions as a workspace when the processor 801 executes programs and is a random access memory that stores temporary information. The storage device 804 is configured to include, for example, a removable external storage device. The communication circuit 805 is composed of circuits for wireless communication compliant with cellular communication standards such as LTE and 5G, and their successor standards. Although Figure 8 shows one communication circuit 805, the base station equipment and terminal equipment may have multiple communication circuits. For example, the base station equipment and terminal equipment may have wireless communication circuits for LTE, 5G, and their successor standards, and an antenna common to these circuits. The base station equipment and terminal equipment may also have separate antennas suitable for each standard. Furthermore, the base station equipment may also have wired communication circuits used when communicating with other base station equipment or nodes in the core network. Furthermore, the terminal equipment may also have communication circuits compliant with wireless communication standards other than cellular communication standards, such as Wireless Local Area Network (LAN) and Bluetooth®. The base station equipment and terminal equipment may have separate communication circuits 805 for each of the multiple usable frequency bands, or they may have a common communication circuit 805 for at least a portion of those frequency bands.

[0055] Figure 9 shows an example of the functional configuration of a terminal device. The terminal device is composed of, for example, an environmental information receiving unit 901, a learning determination unit 902, a learning processing unit 903, an application determination unit 904, and an inference processing unit 905. Note that Figure 9 shows only the functions particularly relevant to this embodiment, and various other functions that the terminal device may have are omitted from the illustration. For example, the terminal device naturally has other functions that terminal devices compliant with LTE, 5G, and subsequent standards generally have. Also, the functional blocks in Figure 9 are shown schematically, and each functional block may be implemented as an integrated unit or further subdivided. Furthermore, each function in Figure 9 may be implemented, for example, by the processor 801 executing a program stored in the ROM 802 or storage device 804, or by a processor located inside the communication circuit 805 executing predetermined software. Note that the details of the processing performed by each functional unit are as described above, so only the general functions of the terminal device will be outlined here.

[0056] The environmental information receiving unit 901 receives environmental information (e.g., environmental identifier, beam group information, inference decision parameters, etc.) provided by the base station equipment. The learning decision unit 902 determines, based on the environmental information received by the environmental information receiving unit 901, whether or not the environment is suitable for collecting information for training the AI / ML model. If the learning decision unit 902 determines that the environment is suitable for collecting information for training the AI / ML model, the learning processing unit 903 requests the base station equipment to transmit a beam for training, if necessary, and obtains training data by measuring the reference signal transmitted in the beam formed by the base station equipment. The learning processing unit 903 uses the acquired training data to train (update) the AI / ML model. If the training (update) of the AI / ML model is performed by an external server, the learning processing unit 903 can transfer the acquired training data to the external server and obtain the AI / ML model of the training result, for example, by the process described in relation to Figure 4. Furthermore, the learning processing unit 903 may, for example, acquire test data similar to the training data while performing inference using the AI / ML model, and evaluate the performance of the AI / ML model. The learning processing unit 903 may then decide to update the AI / ML model if the estimation error in the AI / ML model exceeds a predetermined level, and may execute processing for retraining the AI / ML model. The applicability determination unit 904 determines whether the AI / ML model acquired by the learning processing unit 903 is applicable based on the inference determination parameters received by the environment information receiving unit 901. The applicability determination unit 904 may notify the base station device of the determination result. The inference processing unit 905 receives an instruction from the base station device indicating that it should use the AI / ML model that the applicability determination unit 904 has determined to be applicable, and performs inference using the specified AI / ML model according to that instruction. The inference processing unit 905 may also autonomously decide to use the AI / ML model that the applicability determination unit 904 has determined to be applicable. In other words, the inference processing unit 905 may use an applicable AI / ML model without receiving instructions from the base station equipment.In this case, the application determination unit 904 may notify the base station device of the determination result, but the inference processing unit 905 may notify the base station device of the AI / ML model to be used (for example, information indicating the corresponding inference determination parameters). This allows the terminal device to autonomously determine the AI / ML model to be used, while the base station device performs control such as transmitting a reference signal using a portion of the beam that can be formed in a format suitable for that AI / ML model.

[0057] Figure 10 shows an example of the functional configuration of a base station device. The base station device is composed of, for example, an environmental information notification unit 1001, a beam control unit 1002, and an inference instruction unit 1003. Note that Figure 10 shows only the functions particularly relevant to this embodiment, and various other functions that the base station device may have are omitted from the illustration. For example, the base station device naturally has other functions that base station devices compliant with LTE, 5G, and subsequent standards generally have. Also, the functional blocks in Figure 10 are shown schematically, and each functional block may be implemented as an integrated unit or further subdivided. Furthermore, each function in Figure 10 may be implemented, for example, by the processor 801 executing a program stored in the ROM 802 or storage device 804, or by a processor located inside the communication circuit 805 executing predetermined software. Note that the details of the processing performed by each functional unit are as described above, so only the general functions of the base station device will be outlined here.

[0058] The environmental information notification unit 1001 notifies the terminal device of environmental information (e.g., environmental identifier, beam group information, inference decision parameters, etc.). The beam control unit 1002 controls the transmission of a reference signal using at least a portion of the beams that can be formed. For example, when the terminal device collects training data for training an AI / ML model, the beam control unit 1002 controls the transmission of a reference signal in all of the beams that can be formed so that the terminal device can acquire the training data. The beam control unit 1002 may also control the transmission of a reference signal in some beams and not in other beams, for example, based on an inference decision parameter that indicates the terminal device holds a corresponding AI / ML model. The inference instruction unit 1003 sends an instruction to the terminal device to perform inference using an AI / ML model that is available to the terminal device. In one example, the inference instruction unit 1003 acquires information on AI / ML models available to the terminal device, transmits information to the terminal device specifying which AI / ML model to use, and instructs the terminal device to perform inference using that AI / ML model. The inference instruction unit 1003 may also notify the terminal device of inference determination parameters indicating that the terminal device possesses the corresponding AI / ML model, as information specifying which AI / ML model to use.

[0059] As described above, according to this embodiment, it is possible to appropriately train an AI / ML model or to have an appropriate AI / ML model used by the terminal device, depending on the environment in which the terminal device is placed. Therefore, it is possible to contribute to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), "Build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0060] The invention is not limited to the embodiments described above, and various modifications and changes are possible within the scope of the gist of the invention.

[0061] This application claims priority based on Japanese Patent Application No. 2025-052341, filed on 26 March 2025, and all of its contents are incorporated herein by reference.

Claims

1. A terminal device compliant with the cellular communication standards of the Third Generation Partnership Project (3GPP), comprising: a receiving means for receiving parameter information from a connected base station device indicating the environment to which an artificial intelligence (AI) / machine learning (ML) model relating to the beam formed by the base station device is applied; a determination means for determining whether or not the AI / ML model held by the terminal device can be applied in the environment of the parameters; and an execution means for performing inference using the AI / ML model when it is determined that the AI / ML model held by the terminal device can be applied.

2. The terminal device according to claim 1, wherein the parameters include information on beams that the base station device can form and information on inputs to and outputs from an AI / ML model.

3. The terminal device according to claim 2, wherein the parameters include information indicating a beam corresponding to a measured value of radio quality input to an AI / ML model, and a beam corresponding to an estimated value of radio quality output from the AI / ML model, among the beams that the base station device can form, and setting information for reporting the measured value and the estimated value.

4. The terminal device according to claim 1, further comprising a notification means for notifying the base station device of the result of the determination, wherein the execution means executes inference using the AI / ML model when it is determined that the AI / ML model held by the terminal device can be applied and when it receives an instruction from the base station device to perform inference using the AI / ML model.

5. The terminal device according to claim 4, wherein the receiving means receives information from the base station device regarding a second parameter indicating the environment in which the AI / ML model at the other base station device to be handed over is applied during the handover process; the determination means determines whether or not the AI / ML model held by the terminal device can be applied in the environment of the second parameter; and the notification means notifies the other base station device of the result of the determination by a message when the handover process is completed or by another message after the completion of the handover process.

6. The terminal device according to claim 4, wherein the receiving means receives information from the base station device regarding a second parameter indicating the environment in which the AI / ML model at the other base station device to which the connection is switched is applied during the processing of a conditional handover (CHO) or Layer 1 / Layer 2 Triggered Mobility (LTM); the determination means determines whether the AI / ML model held by the terminal device can be applied in the environment of the second parameter; and the notification means notifies the base station device of the result of the determination after the connection switching by the CHO or the LTM is performed.

7. The terminal device according to claim 4, wherein the receiving means receives the information about a plurality of parameters, and the determination means determines whether or not the AI / ML model held by the terminal device can be applied in each of the environments of the plurality of parameters.

8. The terminal device according to claim 7, wherein the instruction for inference using an AI / ML model from the base station device includes information specifying the parameters to which the AI / ML model has been determined to be applicable.

9. The terminal device according to claim 3, further comprising a reporting means for reporting at least one of an estimated value and an measured value for a beam to be estimated, wherein the reporting means makes the report in a format that allows the base station device to identify whether the reported value is an estimated value or a measured value.

10. The terminal device according to any one of claims 1 to 9, wherein the inference using the AI / ML model is an AI / ML model that takes measured values ​​of wireless quality as input and outputs an index of a predetermined number of beams from the one with the best wireless quality.

11. The terminal device according to any one of claims 1 to 9, wherein the AI / ML model is an AI / ML model that outputs an estimated value at a specific timing after acquiring a measurement value, and the parameter information indicating the environment to which the AI / ML model is applied includes information regarding timing for identifying the specific timing.

12. A base station device conforming to the cellular communication standards of the Third Generation Partnership Project (3GPP), comprising: notification means for notifying a connected terminal device of parameter information indicating the environment in which an artificial intelligence (AI) / machine learning (ML) model relating to the beam formed by the base station device is applied; receiving means for receiving information from the terminal device indicating whether or not the AI / ML model held by the terminal device can be applied in the environment of the parameters; and instruction means for instructing the execution of inference using the AI / ML model when information indicating that the AI / ML model held by the terminal device can be applied in the environment of the parameters is received.

13. The base station device according to claim 12, wherein the parameters include information on beams that the base station device can form and information on inputs to and outputs from an AI / ML model.

14. The base station device according to claim 13, wherein the parameters include information indicating a beam corresponding to a measured value of radio quality input to an AI / ML model, and a beam corresponding to an estimated value of radio quality output from the AI / ML model, among the beams that the base station device can form, and setting information for reporting the measured value and the estimated value.

15. The base station device according to claim 12, wherein the notification means, in the process of handover, notifies the terminal device of information of a second parameter indicating the environment in which the AI / ML model at the other base station device to be handed over is applied.

16. The base station device according to claim 12, wherein the notification means, in the processing of a conditional handover (CHO) or Layer 1 / Layer 2 Triggered Mobility (LTM), notifies from the base station device of information of a second parameter indicating the environment in which the AI / ML model at the other base station device to which the connection is switched is applied, and the receiving means receives from the terminal device before the connection switching by the CHO or the LTM is performed information indicating whether or not the AI / ML model held by the terminal device can be applied in the environment of the second parameter.

17. The base station device according to claim 12, wherein the notification means notifies the information regarding a plurality of parameters, the receiving means receives information indicating whether or not the terminal device holds an AI / ML model that can be applied in each of the environments of the plurality of parameters, and the instruction means instructs the terminal device to perform inference using the AI / ML model held by the terminal device.

18. The base station device according to claim 17, wherein the inference instructions include information specifying the parameters that indicate the terminal device can apply the AI / ML model.

19. A control method performed by a terminal device compliant with the cellular communication standards of the Third Generation Partnership Project (3GPP), comprising: receiving parameter information from a connected base station device indicating the environment to which an artificial intelligence (AI) / machine learning (ML) model relating to the beam formed by the base station device is applied; determining whether the AI / ML model held by the terminal device can be applied in the environment of the parameters; and, if it is determined that the AI / ML model held by the terminal device can be applied, performing inference using the AI / ML model.

20. A control method performed by a base station device compliant with the cellular communication standards of the Third Generation Partnership Project (3GPP), comprising: notifying a connected terminal device of parameter information indicating the environment in which an artificial intelligence (AI) / machine learning (ML) model relating to the beam formed by the base station device is applied; receiving information from the terminal device indicating whether or not an AI / ML model held by the terminal device can be applied in the environment of the parameters; and, when information indicating that an AI / ML model held by the terminal device can be applied in the environment of the parameters is received, instructing the execution of inference using the AI / ML model.

21. A program for causing a computer installed in a terminal device compliant with the cellular communication standards of the Third Generation Partnership Project (3GPP) to execute the control method described in claim 19.

22. A program for causing a computer installed in a base station device compliant with the cellular communication standards of the Third Generation Partnership Project (3GPP) to execute the control method described in claim 20.

Images