Wireless terminal and method therefor

Abstract

This wireless terminal evaluates a criterion for triggering transmission of a report on a result obtained by inference using an artificial intelligence or machine learning (AI / ML) model for generating beam predictions in a first set pertaining to one or more beams on the basis of measurement results in a second set pertaining to a plurality of beams. The criterion is based on a comparison between the first set, which is obtained by executing inference one time, and the second set, which is used in order to execute inference. The wireless terminal transmits the report to a network on the basis of at least the evaluation of the criterion. This makes it possible to, e.g., provide details related to the report on the result of the inference made by the AI / ML model for the purpose of beam prediction.

Description

Wireless terminal and method thereof

[0001] The present disclosure relates to wireless communication networks, and more particularly to the application of artificial intelligence (AI) to wireless communication networks.

[0002] The 3rd Generation Partnership Project (3GPP®) is discussing the application or introduction of AI or machine learning (ML) to 5G. In particular, network-based AI / ML with User Equipment (UE) involvement and UE-based AI / ML are being considered for 3GPP Release 18 and beyond (see, for example, Non-Patent Documents 1 and 2).

[0003] AI / ML can be considered for both network internal functions and the air interface (i.e., Uu). Potential use cases of AI / ML for the air interface include Channel State Information (CSI) feedback compression, beam management (BM), and positioning accuracy enhancements (see, for example, Non-Patent Document 1).

[0004] Network-based AI / ML is also referred to as a network-side AI / ML model or network-side model. In network-based AI / ML, the network performs AI / ML inference. AI / ML inference refers to predictions or decisions using or based on a trained AI / ML model. The AI / ML inference function may be located in the Next Generation Radio Access Network (NG-RAN) (e.g., gNB). Alternatively, the AI / ML inference function may be located in a Near-Real-Time (Near-RT) RAN Intelligent Controller (RIC) coupled to the gNB. Training of the AI / ML model may be performed in the NG-RAN. Alternatively, an Operation, Administration, and Maintenance (OAM) server or a non-RT RIC may train the AI / ML model and provide the trained AI / ML model (e.g., trained parameters or an AI / ML application including the trained parameters) to the NG-RAN (e.g., gNB) or Near-RT RIC.

[0005] UE-based AI / ML is also referred to as a UE-side AI / ML model or UE-side model. In UE-based AI / ML, the UE performs AI / ML inference. In UE-based AI / ML, the UE runs an AI / ML model (i.e., a trained machine learning model) and obtains the AI ​​inference results locally. For example, the UE can predict future events or measurements based on past measurements. The UE can feed back the predicted results (e.g., mobility or beam predictions) to the network (e.g., gNB). Training of the AI / ML model for UE-based AI / ML may be performed by the UE or by the network (e.g., gNB, Near-RT RIC, Non-RT RIC, or other OAM server or controller).

[0006] In addition, a two-sided model may be used. A two-sided model is also called a two-sided AI / ML model. A two-sided model is a pair of AI / ML models in which joint inference is performed. Joint inference is AI / ML inference in which inference is performed jointly across the UE and the network. Specifically, the UE performs the first part of the inference and the gNB performs the rest, or vice versa.

[0007] One sub-use case for AI / ML BM involves downlink (DL) beam prediction in both UE-side and network-side models (see, for example, Non-Patent Documents 1 and 2). DL beam prediction includes spatial-domain beam prediction and temporal beam prediction. Temporal beam prediction may also be referred to as time-domain beam prediction. DL beam prediction includes prediction of DL transmission (Tx) beam, prediction of DL reception (Rx) beam, and prediction of beam pairs of DL Tx beam and DL Rx beam.

[0008] Spatial-domain DL beam prediction is a spatial-domain DL beam prediction of Set A of beams based on measurements of Set B of beams. Spatial-domain DL beam prediction is conveniently called BM-Case 1. Beam Set B is the set of beams whose measurements are taken as inputs to an AI / ML model. In spatial-domain DL beam prediction (i.e., BM-Case 1), Set A is different from Set B (i.e., Set B is not a subset of Set A), or Set B is considered to be a subset of Set A. There are four possible inputs to the AI / ML model for spatial domain DL beam prediction: Layer 1 (L1) Reference Signal Received Power (RSRP) measurements only based on set B; L1-RSRP measurements and assistance information based on set B; Channel Impulse Response (CIR) based on set B; L1-RSRP measurements based on set B and either or both of the corresponding DL Tx beam ID and Rx beam ID.

[0009] Temporal DL beam prediction is a temporal DL beam prediction of Set A of beams based on the historical measurement results of Set B of beams. Temporal DL beam prediction is conveniently called BM-Case 2. Beam Set B is the set of beams whose measurements are taken as inputs to the AI / ML model. In temporal DL beam prediction (i.e., BM-Case 2), Set A is considered to be different from Set B (i.e., Set B is not a subset of Set A), Set B is a subset of Set A (i.e., Set B is not identical to Set A), or Set A and Set B are considered to be the same. The input to the AI / ML model for temporal DL beam prediction is considered to be the measurement results of the K (K≧1) most recent measurement instances, with the following options: - Layer 1 (L1) Reference Signal Received Power (RSRP) measurements only based on set B; - L1-RSRP measurements based on set B and assistance information; - L1-RSRP measurements based on set B and one or both of the corresponding DL Tx beam ID and Rx beam ID.

[0010] Non-Patent Document 1 describes the life cycle management (LCM) of AI / ML models, particularly in Section 4.2. LCM includes, for example, model training, model deployment, model inference, model monitoring, and model update. More specifically, LCM considers the following aspects, including the definition (if necessary) and necessity of components: data collection (including related assistance information, if applicable); model training; functionality or model identification; model delivery or transfer; model inference operation; function or model selection, activation, deactivation, switching, and fallback operations; function or model monitoring; model update; and UE capabilities.

[0011] LCM procedures are studied for either or both cases where AI / ML models have a model ID and associated information (i.e., model ID-based LCM) or where predetermined functionality is provided by AI / ML operations (i.e., function-based LCM). AI / ML models identified by model IDs are logical, and how they map to physical AI / ML models may be implementation-dependent. Where a distinction is necessary for discussion, the term logical AI / ML model may be used to refer to a model that has been identified and assigned a model ID, and the term physical AI / ML model may be used to refer to the actual implementation of such a model.

[0012] For the UE-side model and the UE part of the two-sided model, in the case of AI / ML functionality identification, the legacy 3GPP framework of features is used as the starting point, and the UE indicates the supported functionalities for a given sub-use case. Here, the UE capability reporting is used as the starting point. In the case of AI / ML model identification, for the UE-side model and the UE part of the two-sided model, the model is identified by a model ID in the network, and the UE indicates the supported AI / ML models.

[0013] Non-Patent Document 1 describes performance monitoring of AI / ML models for beam management in Section 4.1.3. Regarding performance monitoring for BM-Case 1 and BM-Case 2, performance metrics include the following alternatives: Key Performance Indicators (KPIs) related to beam prediction accuracy, e.g., Top-1 or Top-K beam prediction accuracy; Link quality-related KPIs, e.g., throughput, L1-RSRP, L1 Signal-to-Interference-plus-Noise Ratio (SINR), hypothetical Block Error Rate (BLER); Performance metrics based on the AI / ML input / output data distribution; and L1-RSRP difference, evaluated by comparing the measured RSRP with the predicted RSRP.

[0014] Regarding performance monitoring of BM-Case 1 and BM-Case 2, benchmarks or references for performance comparison include: - The best beam obtained by measuring the beams included in the set indicated by the gNB (e.g., Beams from Set A), - Measurements of the predicted best beam corresponding to the model output (e.g., comparison of actual L1-RSRP with predicted RSRP of the predicted Top-1 / K beams).

[0015] Regarding performance monitoring of BM-Case 1 and BM-Case 2 with UE-side AI / ML models, Type 1 performance monitoring and Type 2 performance monitoring are considered.

[0016] Type 1 performance monitoring includes two options: network (NW)-side performance monitoring (option 1) and UE-assisted performance monitoring (option 2). In NW-side performance monitoring, the UE sends reports to the network, for example, for calculation of performance metrics in the network. In UE-assisted performance monitoring, the UE calculates performance metrics and reports them to the network, or reports events to the network based on the performance metrics. The network sends instructions for LCM operations to the UE.

[0017] In Type 2 performance monitoring, the UE makes the decision and indicates, requests, or reports the decision to the gNB. Indications, requests, or reports may not be required. In the case of UE-side model monitoring, the UE makes the model selection, activation, deactivation, switching, or fallback decision.

[0018] Regarding performance monitoring of BM-Case 1 and BM-Case 2 with UE-side AI / ML models, mechanisms are considered that facilitate the UE to detect whether a function or model is adequate or no longer adequate.

[0019] Non-Patent Document 3 describes the discussions and agreements on AI / ML in beam management at the 116bis meeting of the 3GPP Technical Specification Group (TSG) Radio Access Network (RAN) Working Group #1 (WG1) (RAN1). According to these, for the UE-side AI / ML models of BM-Case 1 and BM-Case 2, it was agreed that the current CSI framework should be used as a starting point for configuring Set A. It was also agreed that, at least for the UE-side model of BM Case 1, CSI-ReportConfig should be used to configure inference result reporting. Details of CSI-ReportConfig will be discussed further. In Option 1, one CSI-ResourceConfigId is configured for Set B. In this case, the method by which the UE determines information about Set A needs to be considered. In Option 2, one CSI-ResourceConfigId is configured for both Set A and Set B. In this case, the method by which the resource sets for Set A and Set B are configured in CSI-ResourceConfig needs to be considered.

[0020] Non-Patent Document 4 describes the discussions and agreements on AI / ML in beam management at the 117th 3GPP RAN1 meeting. According to these, regarding performance monitoring, it was agreed to support Type 1 performance monitoring, which includes two options: network-side performance monitoring (Option 1) and UE-assisted performance monitoring (Option 2), for BM-Case 1 and BM-Case 2 using UE-side AI / ML models. Whether one or both of Option 1 and Option 2 trigger reports based on events is further discussed. Support for Type 2 performance monitoring is also further discussed.

[0021] 3GPP TR 38.843 V18.0.0 (2023-12) "3rd Generation Partnership Project; Technical Specification Group Radio Access Network; Study on Artificial Intelligence (AI) / Machine Learning (ML) for NR air interface (Release 18)", January 2024 Qualcomm, "New WID on Artificial Intelligence (AI) / Machine Learning (ML) for NR Air Interface", RP-234039, 3GPP TSG RAN Meeting #102, Edinburgh, Scotland, December 11-15, 2023 Samsung, "FL summary #5 for AI / ML in beam management", R1-2403756, 3GPP TSG-RAN WG1 Meeting #116bis, Changsha, China, April 15-19, 2024 Samsung, "FL summary #5 for AI / ML in beam management", R1-2405680, 3GPP TSG-RAN WG1 Meeting #117, Fukuoka, Japan, May 20-24, 2024

[0022] The inventors have investigated the LCM of the UE-side AI / ML model and found various issues. Some of these issues relate to the reporting of inference results from the UE-side AI / ML model to the network (e.g., gNB). Reporting the inference results from the UE every time an inference is performed may increase signaling overhead. While it is conceivable for the gNB to configure the UE to periodically report the inference results, periodic reporting may also increase signaling overhead. Similar to the discussion regarding performance monitoring, basing inference result reporting on event triggers may be effective in reducing overhead. However, the definition of events (i.e., events that trigger a report or event entry conditions) suitable for reporting inference results from the UE-side AI / ML model has not yet been clearly defined.

[0023] Some other challenges relate to the UE's ability to notify or report to the network specific conditions, situations, or events related to the execution of inference using a UE-side AI / ML model. In one example, it may be desirable for UE1 to detect when a UE-side AI / ML model is not producing appropriate inference results and report this to the network. Alternatively, it may be desirable for UE1 to detect when inference using a UE-side AI / ML model is not beneficial and report this to the network. It is unclear what mechanism the UE uses to detect and report these specific conditions, situations, or events to the network.

[0024] One of the objectives to be achieved by the embodiments disclosed in this specification is to provide an apparatus, a method, and a program that contribute to solving at least one of multiple problems related to LCM of a UE-side AI / ML model, including the above-mentioned problem. It should be noted that this objective is only one of multiple objectives to be achieved by the multiple embodiments disclosed in this specification. Other objectives or objectives and novel features will become apparent from the description of this specification or the accompanying drawings.

[0025] In a first aspect, a wireless terminal is configured to evaluate criteria for triggering transmission of a report of results of inference using an AI / ML model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of beams, and to transmit the report to a network based at least on evaluation of the criteria, the criteria being based on a comparison of the first set obtained in a single inference run with the second set used for the inference run.

[0026] In a second aspect, a method performed by a wireless terminal includes (a) evaluating criteria for triggering transmission of a report of results from inference using an AI / ML model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of beams, and (b) transmitting the report to a network based at least on the evaluation of the criteria, wherein the criteria is based on a comparison of the first set obtained from a single inference run with the second set used for the inference run.

[0027] In a third aspect, a wireless terminal is configured to evaluate whether results of an inference using an AI / ML model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of beams satisfy a criterion, and the wireless terminal is configured to send a notification to a network regarding the AI / ML model based at least on the criterion being satisfied, the criterion being based on a comparison of the first set obtained from a single inference run with the second set used for the inference run.

[0028] In a fourth aspect, a method performed by a wireless terminal includes: (a) evaluating whether results of inference using an AI / ML model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of beams satisfy a criterion; and (b) the wireless terminal transmitting a notification to a network regarding the AI / ML model based at least on the criterion being satisfied, the criterion being based on a comparison of the first set obtained from a single inference run with the second set used for the inference run.

[0029] In a fifth aspect, a wireless terminal is configured to provide a Physical (PHY) layer and a Medium Access Control (MAC) layer. The PHY layer is configured to perform inference using an AI / ML model to generate beam predictions for one or more beams based on measurement results of a second set of beams, and to send an indication to the MAC layer each time a result of one inference run satisfies a criterion. The MAC layer is configured to count the number of times the indication is received to determine whether to send a notification regarding the AI / ML model to a network, and to send the notification to the network in response to determining that the notification needs to be sent.

[0030] In a sixth aspect, a method performed by a wireless terminal includes providing a PHY layer and a MAC layer, the PHY layer configured to perform inference using an AI / ML model to generate beam predictions for one or more beams based on measurement results of a second set of beams, and to send an indication to the MAC layer each time a result of one inference run satisfies a criterion, the MAC layer configured to count the number of times the indication is received to determine whether to send a notification regarding the AI / ML model to a network, and to send the notification to the network in response to determining that sending the notification is necessary.

[0031] In a seventh aspect, a program includes a group of instructions (software code) that, when loaded into a computer, causes the computer to perform a method according to any one of the above aspects.

[0032] According to the above-described aspects, it is possible to provide an apparatus, a method, and a program that contribute to solving at least one of a plurality of problems related to LCM of a UE-side AI / ML model, including the above-described problems.

[0033] 1 illustrates an example configuration of a wireless communication system according to one or more embodiments. 2 illustrates an example configuration of a wireless communication system according to one or more embodiments. 3 illustrates a flowchart illustrating an example operation of a UE according to one or more embodiments. 4 illustrates an example signaling between a UE and a RAN node according to one or more embodiments. 5 illustrates an example format of a CSI-ReportConfig information element according to one or more embodiments. 6 illustrates an example format of an EventTriggeredAIML information element according to one or more embodiments. 7 illustrates an example flowchart illustrating an example operation of a UE according to one or more embodiments. 8 illustrates an example format of a MAC Control Element (CE) according to one or more embodiments. 9 illustrates an example operation of a UE according to one or more embodiments. 10 illustrates an example signaling between a UE and a RAN node according to one or more embodiments. 11 illustrates an example format of an AIMLInference-FailureRecoveryConfig information element according to one or more embodiments. 12 illustrates an example configuration of a UE according to one or more embodiments. 13 illustrates an example configuration of a RAN node according to one or more embodiments.

[0034] Hereinafter, specific embodiments will be described in detail with reference to the drawings. In each drawing, the same or corresponding elements are designated by the same reference numerals, and for clarity of explanation, duplicate explanations will be omitted as necessary.

[0035] The multiple embodiments described below may be used independently, or two or more embodiments may be combined as appropriate. These multiple embodiments may have different novel features. Therefore, these multiple embodiments may contribute to achieving different objectives or solving different problems, and may contribute to achieving different effects.

[0036] Each drawing is merely an example for describing one or more embodiments. Each drawing may not relate to only one particular embodiment, but may also relate to one or more other embodiments. As will be understood by those skilled in the art, various features or steps described with reference to any one drawing can be combined with features or steps shown in one or more other drawings to create, for example, an embodiment not explicitly shown or described. Not all features or steps shown in any one drawing are necessary to describe an exemplary embodiment, and some features or steps may be omitted. The order of steps described in any drawing may be changed as appropriate.

[0037] Although the following embodiments are primarily described for the 3GPP 5th generation mobile communication system (5G system), these embodiments may also be applied to other wireless communication systems that support beam prediction, particularly AI / ML-based beam prediction.

[0038] As used herein, depending on the context, "if" may be interpreted to mean "when," "while," "at or around the time," "after," "upon," "in response to determining," "in accordance with a determination," or "in response to detecting." These expressions may be interpreted to have the same meaning, depending on the context.

[0039] First, the configurations and operations of several network elements common to several embodiments will be described. Fig. 1 shows an example configuration of a wireless communication system related to several embodiments. In the example of Fig. 1, the wireless communication system includes a wireless terminal (i.e., UE) 1 and a Radio Access Network (RAN) node (e.g., gNB) 2. Each element (network function) shown in Fig. 1 can be implemented, for example, as a network element on dedicated hardware, as a software instance running on dedicated hardware, or as a virtualized function instantiated on an application platform.

[0040] The UE 1 has at least one radio transceiver and is configured to perform wireless communication with the RAN node 2. The UE 1 is connected to the RAN node 2 via an air interface 101. The UE 1 may be referred to by other terms such as a radio terminal, a mobile terminal, a mobile station, or a wireless transmit receive unit (WTRU). The RAN node 2 manages a cell and is configured to perform wireless communication with multiple UEs, including the UE 1, using a cellular communication technology (e.g., an NR Radio Access Technology (RAT)). The RAN node 2 may be referred to by other terms such as a base station, a radio station, or an access point. The UE 1 may be simultaneously connected to multiple RAN nodes, including the RAN node 2, for dual connectivity (DC).

[0041] The RAN node 2 may be a Central Unit (CU) (e.g., gNB-CU) in a cloud RAN (C-RAN) deployment, or a combination of a CU and one or more Distributed Units (DUs) (e.g., gNB-DUs). Furthermore, the CU may include a Control Plane (CP) Unit (e.g., gNB-CU-CP) and one or more User Plane (UP) Units (e.g., gNB-CU-UP). Thus, the RAN node 2 may be a CU-CP or a combination of a CU-CP and a CU-UP. The CU may be a logical node that hosts the Radio Resource Control (RRC), Service Data Adaptation Protocol (SDAP), and Packet Data Convergence Protocol (PDCP) protocols of the gNB (or the RRC and PDCP protocols of the gNB). The DU may be a logical node that hosts the Radio Link Control (RLC), Medium Access Control (MAC), and Physical (PHY) layers of the gNB.

[0042] Specifically, as shown in Fig. 2, the RAN node 2 may include a CU 201 (e.g., gNB-CU) and one or more DUs 211 and 212 (e.g., gNB-DUs). The CU 201 is a logical node that controls the operation of the DUs 211 and 212. The CU 201 and each of the DUs 211 and 212 may also be referred to as a RAN node. Each of the DUs 211 and 212 is a logical node that hosts the RLC layer and MAC layer of the RAN node 2 and hosts part of the PHY layer of the RAN node 2, i.e., the upper PHY layer. The remaining signal processing of the PHY layer, i.e., the lower PHY layer, is located in Transmission Reception Points (TRPs) 231 to 235.

[0043] One DU may support one or more cells. One cell may be supported by only one DU. In the example of Figure 2, DU 211 is connected to TRPs 231 to 233, while DU 212 is connected to TRPs 234 and 235. TRPs 231 to 233 provide one cell 241, and TRPs 234 and 235 provide separate cells 242 and 243, respectively. In other words, DU 211 provides one cell 241, and TRPs 231 to 233 correspond to cell 241. DU 212 provides multiple cells 242 and 243, and TRPs 234 and 235 correspond to cells 242 and 243, respectively.

[0044] Each of the TRPs 231-235 can communicate with UE 1 using a beam. The TRPs 231-235 may also be called Radio Units (RUs), Remote Radio Heads (RRHs), access points (APs), or distributed antennas. Each TRP is a set of geographically co-located antennas (e.g., an antenna array with one or more antenna elements). Each TRP supports either or both Transmission Point (TP) and Reception Point (RP) functions.

[0045] The RAN node 2 may be connected to a RAN controller 3. The RAN controller 3 may be referred to by other terms, such as a control device or a control system. The RAN controller 3 may be integrated into the RAN node 2 (e.g., gNB). Alternatively, the RAN controller 3 may include one or both of a Non-RT RIC and a Near-RT RIC defined in the O-RAN Alliance technical specifications. In this case, the RAN node 2 may be connected to the RAN controller 3 via one or both of an O1 interface and an E2 interface. Alternatively, the RAN controller 3 may be an OAM server or another controller. In other words, the functions of the RAN controller 3 may be located in the RAN node 2, the Non-RT RIC, the Near-RT RIC, the OAM server, or another controller, or may be distributed across any combination thereof.

[0046] The UE 1 may perform AI / ML inference locally. In other words, the UE 1 may support UE-based AI / ML. UE-based AI / ML is also referred to as a UE-side AI / ML model or a UE-side model. This AI / ML inference may relate to RAN optimization. The UE 1 may run AI inference on a trained AI / ML model and take one or more actions according to a prediction or decision based on the AI ​​inference. The AI / ML model may be any model known in the field of machine learning, including deep learning. The AI / ML model may be, for example, but not limited to, a neural network model, a support vector machine model, a decision tree model, a random forest model, or a K-nearest neighbor model.

[0047] By way of example and not limitation, the prediction or decision based on AI inference by UE1 and one or more actions triggered thereby may relate to beam management (BM) and include DL beam prediction. The one or more actions may include, for example, but not limited to, DL beam selection. The DL beam may include a Synchronization Signal (SS) / Physical Broadcast Channel (PBCH) block (SSB) beam, a Channel State Information (CSI) Reference Signal (CSI-RS) beam, or both. For example, the AI / ML model may output one or more candidate beams for DL ​​beam selection. Additionally or alternatively, the AI / ML model may predict or determine the timing of execution of an action for beam management. The DL beam prediction and selection by UE1 may be performed during a Beam Failure Recovery (BFR) procedure.

[0048] The training of the AI / ML model for AI / ML inference by the UE 1 may be performed by the UE 1 or by the network (e.g., RAN node 2, RAN controller 3, or OAM). The training method may be offline learning, online learning, or a combination thereof.

[0049] Similarly, the network, i.e., the RAN node 2 or the RAN controller 3, or other control system, or any combination thereof, may perform AI / ML inference. In other words, the network may support network-based AI / ML. Network-based AI / ML is also referred to as a network-side AI / ML model or network-side model. This AI / ML inference may relate to RAN optimization. The network may run AI inference on a trained AI / ML model and take one or more actions according to a prediction or decision based on the AI ​​inference. The AI / ML model may be any model known in the field of machine learning, including deep learning. The AI / ML model may be, for example, but not limited to, a neural network model, a support vector machine model, a decision tree model, a random forest model, or a k-nearest neighbor model.

[0050] By way of example and not limitation, the network-based AI inference-based prediction or decision and one or more actions triggered thereby may relate to beam management and include DL beam prediction. The one or more actions may include, for example and without limitation, DL beam selection. The DL beam may include an SSB beam, a CSI-RS beam, or both. For example, the AI / ML model may output one or more candidate beams for DL ​​beam selection. Additionally or alternatively, the AI / ML model may predict or determine when to perform an action for beam management. The network-based DL beam prediction and selection may be performed to determine the set of CSI-RS beams to be configured (or measured) by UE1.

[0051] Training of an AI / ML model for network-based AI / ML inference may be performed by any information processing system or computer system on the network side, and the training method may be offline learning, online learning, or a combination of these.

[0052] In addition, a two-sided model may be used. A two-sided model is also called a two-sided AI / ML model. A two-sided model is a pair of AI / ML models in which joint inference is performed. Joint inference is AI / ML inference in which inference is performed jointly across the UE and the network. Specifically, the UE performs the first part of the inference and the gNB performs the rest, or vice versa.

[0053] One sub-use case for AI / ML BM involves DL beam prediction in both the UE-side model and the network-side model. DL beam prediction includes spatial domain beam prediction and temporal beam prediction. Temporal beam prediction may also be referred to as time domain beam prediction. DL beam prediction includes DL Tx beam prediction, DL Rx beam prediction, and beam pair prediction of DL Tx beam and DL Rx beam.

[0054] Spatial-domain DL beam prediction is a spatial-domain DL beam prediction of Set A of beams based on measurements of Set B of beams. Spatial-domain DL beam prediction is conveniently called BM-Case 1. Beam Set B is the set of beams whose measurements are taken as inputs to an AI / ML model. In spatial-domain DL beam prediction (i.e., BM-Case 1), Set A is different from Set B (i.e., Set B is not a subset of Set A), or Set B is considered to be a subset of Set A. For example, there are four possible inputs to the AI / ML model for spatial domain DL beam prediction: Layer 1 (L1) Reference Signal Received Power (RSRP) measurements only based on set B; L1-RSRP measurements and assistance information based on set B; Channel Impulse Response (CIR) based on set B; L1-RSRP measurements based on set B and one or both of the corresponding DL Tx beam ID and Rx beam ID.

[0055] Temporal DL beam prediction is a temporal DL beam prediction of Set A of beams based on the historical measurement results of Set B of beams. Temporal DL beam prediction is conveniently called BM-Case 2. Beam Set B is the set of beams whose measurements are taken as inputs to the AI / ML model. In temporal DL beam prediction (i.e., BM-Case 2), Set A is considered to be different from Set B (i.e., Set B is not a subset of Set A), Set B is a subset of Set A (i.e., Set B is not identical to Set A), or Set A and Set B are considered to be the same. For example, the input to the AI / ML model for temporal DL beam prediction may be the measurement results of the K (K≧1) most recent measurement instances, with the following options: - Layer 1 (L1) Reference Signal Received Power (RSRP) measurements only based on set B; - L1-RSRP measurements based on set B and assistance information; - L1-RSRP measurements based on set B and one or both of the corresponding DL Tx beam ID and Rx beam ID.

[0056] The UE 1 and RAN node 2 support Life Cycle Management (LCM) of AI / ML models or AI / ML capabilities. LCM includes, for example, model training, model deployment, model inference, model monitoring, and model updates. More specifically, LCM may include one or more of the following: data collection (including associated assistance information, if applicable); model training; functionality or model identification; model delivery or transfer; model inference operations; feature or model selection, activation, deactivation, switching, and fallback operations; feature or model monitoring; model updates; and UE capabilities.

[0057] The LCM procedure relates to one or both of the cases where an AI / ML model has a model ID and related information (i.e., model ID-based LCM) and the case where predetermined functionality is provided by AI / ML operations (i.e., function-based LCM). The definitions and details of model ID-based LCM and function-based LCM are as explained in the Background Art section, so a duplicate explanation will be omitted here.

[0058] In BM-Case 1 and BM-Case 2 using the UE-side AI / ML model, UE1 supports reporting of inference results to the network (e.g., RAN node 2). The results of inference performed by the UE-side AI / ML model for DL ​​beam prediction can be reported by RAN node 2 for beam management. It is assumed that the inference results from the UE-side AI / ML model are reported from UE1 to RAN node 2 using CSI reporting, i.e., Layer 1 measurement reporting. At least in the UE-side AI / ML model for BM Case 1, CSI-ReportConfig may be used to configure inference result reporting. CSI reporting can be used by UE1 to report measurement results of one or both of SSB and CSI-RS to RAN node 2, and can also be used by UE1 to report inference results for one or both of SSB and CSI-RS to RAN node 2. In the case of BM-Case 1, set A may be a set of SSB beams, and set B may be a subset of the SSB beams that make up set A. Alternatively, for BM-Case 1, Set A may be a set of CSI-RS beams and Set B may be a set of SSB beams. For BM-Case 2, Set A may be a set of SSB beams and Set B may be the same as or a subset of Set A. Alternatively, for BM-Case 2, Set A may be a set of CSI-RS beams and Set B may be a set of SSB beams.

[0059] The content of the inference results of the UE-side AI / ML model may include, at least for BM-Case 1, any of the following (where K is 1 or more): Beam information on predicted Top K beam(s) among a set of beams; Beam information on predicted Top K beam(s) among a set of beams, and RSRP of predicted Top K beam(s) among a set of beams; Beam information on predicted Top K beam(s) among a set of beams, and RSRP of predicted Top K beam(s) among a set of beams; Beam information on predicted Top K beam(s) among a set of beams, and probability information of predicted Top K beam(s) among a set of beams; Beam information on predicted Top K beam(s) among a set of beams, and RSRP of predicted Top K beam(s) among a set of beams. beam(s) among a set of beams, RSRP of predicted Top K beam(s) among a set of beams, and confidence information of the RSRP).

[0060] According to the framework of CSI reporting, i.e., Layer 1 measurement reporting, the UE 1 may transmit the inference result report to the RAN node 2 via Layer 1 signaling. Specifically, the UE 1 may transmit the inference result report to the RAN node 2 using Uplink Control Information (UCI) transmitted in a Physical Uplink Control Channel (PUCCH) or a Physical Uplink Sheared Channel (PUSCH). Alternatively, the UE 1 may transmit the inference result report using Layer 2 signaling, for example, a MAC Control Element (CE). Alternatively, the UE 1 may transmit the inference result report using Layer 3 signaling, for example, an RRC message.

[0061] The transmission of the inference result report may be periodic, semi-persistent, or aperiodic. A periodic report may be an inference result report at a period set by the RAN node 2. The UE 1 may transmit a periodic inference result report every time periodic inference is performed. An aperiodic report may be transmitted only once by the UE 1 in response to a Layer 1 trigger signal (e.g., Downlink Control Information (DCI)) or a Layer 2 trigger signal (e.g., MAC CE) transmitted from the RAN node 2. A semi-persistent report may be transmitted periodically by the UE 1 in response to a Layer 1 trigger signal (e.g., DCI) or a Layer 2 trigger signal (e.g., MAC CE) transmitted from the RAN node 2.

[0062] Additionally or alternatively, the transmission of the inference result report may be event-triggered. In other words, the transmission of the inference result report may be triggered when a criterion for an inference result report event is met. In the case of an event-triggered report of the inference result, UE1 evaluates criteria for triggering the inference result report. The criteria may include an entry condition that is considered to start the inference result report. Furthermore, the criteria may include a leaving condition that is considered to stop the inference result report. The entry and leaving conditions may be defined using hysteresis. The entry and leaving conditions may be defined using an offset.

[0063] UE1 may use a time to trigger (TTT) in determining the entry and exit conditions associated with an inference result report event. For example, UE1 may initiate a procedure for transmitting an inference result report for the UE-side AI / ML model to RAN node 2 if the entry condition applied to the inference result report event is continuously satisfied for the inference result of the UE-side AI / ML model for a first TTT period.

[0064] UE1 may stop sending inference result reports for the UE-side AI / ML model if the withdrawal condition applied to the inference result reporting event is continuously satisfied for the inference results of the UE-side AI / ML model for a predetermined second TTT period, the length of which may be the same as or different from the length of the first TTT period.

[0065] Alternatively, UE1 may initiate a procedure for transmitting an inference result report to RAN node 2 if an entry condition applicable to the inference result reporting event is met a predetermined number of times within a predetermined first period of time. UE1 may stop transmitting the inference result report if an exit condition applicable to the inference result reporting event is met a predetermined number of times within a predetermined second period of time. The length of the second period may be the same as or different from the length of the first period of time.

[0066] Alternatively, the UE 1 may initiate a procedure for transmitting an inference result report to the RAN node 2 if an entry condition applied to the inference result reporting event is met a predetermined number of times in succession. The UE 1 may stop transmitting the inference result report if a departure condition applied to the inference result reporting event is met a predetermined number of times in succession. The predetermined number of times for the departure condition may be the same as or different from the predetermined number of times for the entry condition.

[0067] In response to the entry condition applied to the inference result reporting event being satisfied, UE1 may transmit the inference result report periodically at predetermined intervals until the exit condition applied to the inference result reporting event is satisfied. Alternatively, in response to the entry condition applied to the inference result reporting event being satisfied, UE1 may transmit the inference result report only once or a predetermined number of times.

[0068] In addition, in the case of BM-Case 1 and BM-Case 2 using the UE-side AI / ML model, UE1 supports performance monitoring, which includes one or both of Type 1 performance monitoring and Type 2 performance monitoring.

[0069] Type 1 performance monitoring includes two options: NW-side performance monitoring (option 1) and UE-assisted performance monitoring (option 2). In NW-side performance monitoring, the UE 1 sends a report to the network (e.g., RAN node 2, RAB controller 3, or OAM) for calculation of performance metrics in the network. In UE-assisted performance monitoring, the UE 1 calculates performance metrics and reports them to the network, or reports events to the network based on the performance metrics. The network sends instructions for LCM operations to the UE 1.

[0070] In Type 2 performance monitoring, the UE 1 makes a decision and indicates, requests, or reports the decision to the RAN node 2. An indication, request, or report may not be necessary. In the case of UE-side model monitoring, the UE 1 makes the model selection, activation, deactivation, switching, or fallback decision.

[0071] The performance metrics for BM-Case 1 and BM-Case 2 may include one or any combination of the following alternatives: Key Performance Indicators (KPIs) related to beam prediction accuracy, e.g., Top-1 or Top-K beam prediction accuracy; Link quality related KPIs, e.g., throughput, L1-RSRP, L1 Signal-to-Interference-plus-Noise Ratio (SNIR), hypothetical Block Error Rate (BLER); Performance metrics based on AI / ML input / output data distribution; L1-RSRP difference assessed by comparing measured RSRP with predicted RSRP.

[0072] More specifically, the performance metrics for BM-Case 1 and BM-Case 2 may include one or any combination of the following alternatives: Alternative 1-1: Statistical results on beam prediction accuracy related KPIs, e.g., Top-K / 1 beam prediction accuracy; Alternative 1-2: Hypothetical on beam prediction accuracy related KPIs, e.g., Top-K / 1 beam prediction accuracy, based on configured resource(s); Alternative 2-1: Measured L1-RSRP of configured resource(s); Alternative 2-2: Hypothetical L1-RSRP based on the configured resource(s); Option 3-1: Probability information of the predicted beam to be the Top 1; Option 3-2: A confidence interval or prediction interval associated with predicted L1-RSRPs at a specific confidence level (e.g., 95%)., 95%)); Option 4-1: The L1-RSRP difference between the measured [L1-]RSRP and predicted RSRP according to beam(s) in the same target Set A resources; Option 4-2: The L1-RSRP difference between the measured [L1-]RSRP of the current beam and the predicted RSRP of the Top 1 beam.

[0073] In Option 1 of Type 1 performance monitoring, i.e., NW-side performance monitoring, UE1 may report one or any combination of the following to assist the network in calculating metrics: - Option 1-1: Predicted Top 1 or Top K beams, and ground truth of the target Set A resources (and at the target time instance(s) for BM Case-2); - Option 2-1: Measured L1-RSRP of the configured resource(s); - Option 3-1: Probability information of the predicted beam to be the Top 1; - Option 4-1: Measured L1-RSRP, and the predicted RSRP; Option 4-2: Measured current (L1-)RSRP and predicted RSRP of the predicted Top 1 beam (measured [L1-]RSRP of current and predicted RSRP of the predicted Top 1 beam).

[0074] In option 2 of Type 1 performance monitoring, i.e., UE-assisted performance monitoring, UE1 may report any or any combination of the above-mentioned metric options 1-1, 1-2, 2-1, 2-2, 3-1, 3-2, 4-1, and 4-2.

[0075] The UE1 may send the performance monitoring report to the RAN node 2 via Layer 1 signaling, e.g., a CSI report. Alternatively, the UE1 may send the performance monitoring report using Layer 2 signaling, e.g., a MAC CE. Alternatively, the UE1 may send the performance monitoring report using Layer 3 signaling, e.g., an RRC message.

[0076] The transmission of performance monitoring reports may be periodic, semi-persistent, or aperiodic. Additionally or alternatively, the transmission of performance monitoring reports may be event-triggered.

[0077] For example, in Type 1 performance monitoring for a UE-side AI / ML model, a performance monitoring report from UE1 to RAN node 2 may be triggered based on an event. In this case, UE1 evaluates criteria for triggering a performance monitoring report. The criteria may include an entry (or entering) condition that is considered to start the performance monitoring report. Furthermore, the criteria may include a leaving (or exit) condition that is considered to stop the performance monitoring report. The entry and leaving conditions may be defined with a hysteresis. The entry and leaving conditions may be defined with an offset.

[0078] UE1 may use a time to trigger (TTT) in determining the entry and exit conditions associated with a performance monitoring reporting event. For example, UE1 may initiate a procedure for transmitting a performance monitoring report for the UE-side AI / ML model to RAN node 2 if the entry condition applicable to the performance monitoring reporting event is continuously satisfied for the UE-side AI / ML model for a first TTT period.

[0079] UE1 may stop transmitting performance monitoring reports for the UE-side AI / ML model if the withdrawal condition applicable to the performance monitoring reporting event is continuously satisfied for the UE-side AI / ML model for a predetermined second TTT period, the length of which may be the same as or different from the length of the first TTT period.

[0080] Alternatively, UE1 may initiate a procedure for transmitting a performance monitoring report to RAN node 2 if an entry condition applicable to the performance monitoring reporting event is met a predetermined number of times within a predetermined first period of time. UE1 may stop transmitting a performance monitoring report if an exit condition applicable to the performance monitoring reporting event is met a predetermined number of times within a predetermined second period of time. The length of the second period may be the same as or different from the length of the first period of time.

[0081] Alternatively, UE1 may initiate a procedure for transmitting a performance monitoring report to RAN node 2 if an entry condition applicable to the performance monitoring reporting event is met a predetermined number of times in succession. UE1 may stop transmitting a performance monitoring report if a departure condition applicable to the performance monitoring reporting event is met a predetermined number of times in succession. The predetermined number of times for the departure condition may be the same as or different from the predetermined number of times for the entry condition.

[0082] In response to the entry condition applicable to the performance monitoring reporting event being satisfied, UE1 may transmit the performance monitoring report periodically at a predetermined interval until the exit condition applicable to the performance monitoring reporting event is satisfied. Alternatively, in response to the entry condition applicable to the performance monitoring reporting event being satisfied, UE1 may transmit the performance monitoring report only once or a predetermined number of times.

[0083] UE1 and RAN node 2 may support at least one of the following events as reporting events for Type 1 performance monitoring: Event-1: The measured Top-1 or Top K beam(s) of Set A and the predicted Top-1 or Top K beam(s) of Set A are different; Event-2: The measured L1-RSRP of one set of beams is lower than a threshold; Event-3: The probability information of Top-1 or Top K beam of Set A is lower than a threshold; Event-4: The L1-RSRP difference between the measured Top-1 or Top K beam(s) of Set A and the predicted Top-1 or Top K beam(s) of Set A are larger than a threshold value.

[0084] In the case of Type 1 performance monitoring, particularly NW-side performance monitoring (Option 1), at least a portion of the performance monitoring report from UE 1 to RAN node 2 can be integrated or common with the inference result report. Specifically, the transmission of inference results that can be used for performance metric calculation by the network can also serve as part of the performance monitoring report. Inference results that can be used for performance metric calculation include, for example, information on the predicted top 1 or top K beams, probability information on the beam predicted to be top 1, predicted (L1-)RSRP, or any combination thereof.

[0085] First Embodiment A configuration example of a wireless communication system according to this embodiment is similar to the configuration example described with reference to Figures 1 and 2. This embodiment provides details related to reporting inference results of a UE-side AI / ML model for beam prediction.

[0086] FIG. 3 illustrates an example of the operation of UE1. In step 301, UE1 evaluates criteria to trigger the transmission of a report of inference results using a UE-side AI / ML model. The UE-side AI / ML model is a model for generating beam predictions for a first set (i.e., set A) of one or more beams based on measurement results for a second set (i.e., set B) of beams. Here, the criteria are based on a comparison of the first set (i.e., set A) obtained from a single inference run with the second set (i.e., set B) used for that inference run. In step 302, UE1 transmits an inference result report to the network (e.g., RAN node 2) based at least on the evaluation of the criteria.

[0087] The term “criteria” may be rephrased as one or more conditions, one or more trigger conditions, etc. The term “criteria for triggering the sending of a report of an inference result” may be rephrased as criteria for an inference result reporting event, criteria for triggering an inference result reporting event, criteria that trigger an inference result report, or criteria for an event-triggered inference result reporting.

[0088] The criteria evaluated in step 301 may include entry conditions defined for a particular (inference result report) event. As described above, UE1 may use the TTT in determining the entry conditions. For example, UE1 may initiate a procedure for transmitting an inference result report for the UE-side AI / ML model to RAN node 2 if the entry conditions applied to the inference result report event are continuously met for the inference result of the UE-side AI / ML model for the TTT period.

[0089] The criterion evaluated in step 301 may be based at least on a comparison between the predicted beam quality of the first set (i.e., set A) obtained by the inference execution and the measured beam quality of the second set (i.e., set B) used for the inference execution. UE1 may determine that the criterion is met based at least on the quality of the best-quality beam (i.e., top 1 beam) of the first set obtained by the inference execution being worse than the quality of the best-quality beam (i.e., top 1 beam) of the second set used for the inference execution by at least an offset. Alternatively, UE1 may determine that the criterion is met based at least on the quality of the best-quality beam of the first set obtained by the inference execution not being better than the quality of the best-quality beam of the second set used for the inference execution by more than an offset. The beam quality may be L1-RSRP or L1-SINR.

[0090] Additionally or alternatively, the criteria evaluated in step 301 may take into account at least whether a beam belonging to the first set obtained by the inference execution is included in the second set used for the inference execution. The criteria may take into account at least whether a beam with the highest quality (i.e., a top 1 beam) from the first set obtained by the inference execution is included in the second set used for the inference execution. UE1 may determine that the criteria are met based at least on the fact that the beam with the highest quality (i.e., a top 1 beam) from the first set obtained by the inference execution is included in the second set used for the inference execution.

[0091] The criteria evaluated in step 301 may include entry conditions for any of the inference result reporting events defined below, Event AIML5, Event AIML6, and Event AIML7: Event AIML5: The predicted L1-RSRP of the predicted Top-1 beam is offset worse than the measured L1-RSRP of the measured Top-1 beam. Event AIML6: The predicted L1-RSRP of the predicted Top-1 beam is not offset better than the measured L1-RSRP of the measured Top-1 beam. Event AIML7: The predicted Top-1 beam of Set A of beams belongs to Set B of beams.

[0092] For Event AIML5, UE1 may consider the entry condition for this event to be met if the following formula AIML5-1 is satisfied: where P1 is the predicted beam quality (e.g., L1-RSRP) of the top 1 beam predicted by inference, P2 is the measured beam quality (e.g., L1-RSRP) of the top 1 beam in Set B used for inference, Hys is the hysteresis parameter for the event, and Off is the offset parameter for the event.

[0093] For Event AIML5, UE1 may consider the leave condition for this event to be met if the following formula AIML5-2 is satisfied:

[0094] With respect to Event AIML6, UE1 may consider the entry condition for this event to be met if the following formula AIML6-1 is satisfied:

[0095] With respect to Event AIML6, UE1 may consider the condition for leaving for this event to be met if the following formula AIML6-2 is satisfied:

[0096] With respect to Event AIML7, UE1 may consider the entry condition for this event to be met if the following formula AIML7-1 is satisfied: where I1 is the index of the top 1 beam predicted by inference.

[0097] With respect to Event AIML7, UE1 may consider the leave condition for this event to be met if the following formula AIML7-2 is satisfied:

[0098] The TTT may be used to determine the entry and departure conditions in Event AIML5, Event AIML6, and Event AIML7. For example, when Event AIML5 is used, UE1 may initiate a procedure for transmitting an inference result report for the UE-side AI / ML model to RAN node 2 if the entry condition (formula AIML5-1) is continuously satisfied for the inference result of the UE-side AI / ML model for the TTT period.

[0099] The operations described with reference to Fig. 3 enable UE1 to transmit an event-triggered inference result report. This may be used in cases where the measurement reports of Set B are used as a baseline and event-triggered inference results are reported on top of the measurement reports of Set B. For example, the measurement results of Set B may be reported, for example periodically, similar to existing CSI reports, while the results of performing inference using the measurement results of Set B may be reported based on an event trigger.

[0100] According to the operation described with reference to FIG. 3, for example, an inference result report can be triggered when the quality of beam set A obtained by performing inference is not sufficiently high compared to the quality of beam set B used for the inference. Alternatively, an inference result report can be triggered when the difference between beam set A obtained by performing inference and beam set B used for the inference is not sufficient. An inference result report triggered by such an event can be used to inform the network that the UE-side AI / ML model is not providing appropriate inference results. Alternatively, an inference result report triggered by such an event can be used to inform the network that inference using the UE-side AI / ML model is not beneficial.

[0101] The following provides an example of a configuration for event-triggered inference result reporting. Figure 4 shows an example of signaling between UE1 and RAN node 2. In step 401, RAN node 2 sends a configuration for event-triggered inference result reporting to UE1. RAN node 2 may send the configuration to UE1 using an RRC message. The RRC message may be an RRC Reconfiguration message. UE1 receives the configuration for event-triggered inference result reporting and reports the inference result by the UE-side AI / ML model according to the received configuration.

[0102] The setting for event-triggered inference result reporting indicates the event type to be configured for inference result reporting. The event type may indicate one or more of the above-described Event AIML5, Event AIML6, and Event AIML7. The inference result reporting setting may further include values ​​for one or more parameters for the specified event. These may include offset, hysteresis, TTT, or any combination thereof.

[0103] CSI-ReportConfig may be used for the configuration for event-triggered inference result reporting. Fig. 5 shows an example of a format of CSI-ReportConfig improved to include the configuration for event-triggered inference result reporting. In the example of Fig. 5, the CSI-ReportConfig information element (IE) includes an eventTriggeredAIML field (501). The eventTriggeredAIML field indicates the EventTriggerConfigAIML IE.

[0104] FIG. 6 shows an example of the format of the EventTriggerConfigAIML IE. In the example of FIG. 6, the EventTriggerConfigAIML IE indicates which of Event AIML5, Event AIML6, and Event AIML7 is applied. When Event AIML5 is used, the EventTriggerConfigAIML IE includes an eventAIML5 field (601). The eventAIML5 field (601) includes an aIML5-Offset field, a reportOnLeave field, a hysteresis field, and a timeToTrigger field. The reportOnLeave field indicates whether UE1 should initiate an inference result reporting procedure when a leave condition is met. When Event AIML6 is used, the EventTriggerConfigAIML IE includes an eventAIML6 field (602). The structure of the eventAIML6 field is similar to that of the eventAIML5 field. When Event AIML7 is used, the EventTriggerConfigAIML IE includes an eventAIML7 field (603). The eventAIML7 field (603) includes a reportOnLeave field and a timeToTrigger field.

[0105] Second Embodiment A configuration example of a wireless communication system according to this embodiment is similar to the configuration example described with reference to Figures 1 and 2. This embodiment provides details related to monitoring of a UE-side AI / ML model.

[0106] 7 illustrates an example of the operation of UE1. In step 701, UE1 evaluates whether the results of inference using a UE-side AI / ML model satisfy a criterion or criteria. The UE-side AI / ML model is a model for generating beam predictions for a first set (i.e., set A) of one or more beams based on measurement results of a second set (i.e., set B) of multiple beams. Here, the criteria are based on a comparison of the first set (i.e., set A) obtained from a single inference run with the second set (i.e., set B) used for that inference run.

[0107] In step 702, UE1 sends a notification regarding the UE-side AI / ML model to the network (e.g., RAN node 2) based at least on the criteria being met. UE1 may send the notification to the network via Layer 2 signaling (e.g., MAC CE) or an RRC message (e.g., RRC message). The notification may indicate to the network that the UE-side AI / ML model is not suitable for UE1's current situation. The notification may indicate to the network that the UE-side AI / ML model is not providing appropriate inference results. The notification may indicate to the network that inference using the UE-side AI / ML model is not beneficial. Alternatively, upon receiving the notification from UE1, the network may understand that the UE-side AI / ML model is not providing appropriate inference results or that inference using the UE-side AI / ML model is not beneficial.

[0108] The term "criterion" may be replaced with other terms such as one or more conditions, one or more trigger conditions, etc. UE 1 may receive configuration information from the network to define the criteria. RAN node 2 may send this configuration information to UE 1 via an RRC message (e.g., RRC Reconfiguration).

[0109] The criterion evaluated in step 701 may be based at least on a comparison between the predicted beam quality of the first set (i.e., set A) obtained by the inference execution and the measured beam quality of the second set (i.e., set B) used for the inference execution. UE1 may determine that the criterion is met based at least on the quality of the best-quality beam (i.e., top 1 beam) of the first set obtained by the inference execution being worse than the quality of the best-quality beam (i.e., top 1 beam) of the second set used for the inference execution by at least an offset. Alternatively, UE1 may determine that the criterion is met based at least on the quality of the best-quality beam of the first set obtained by the inference execution not being better than the quality of the best-quality beam of the second set used for the inference execution by more than an offset. The beam quality may be L1-RSRP or L1-SINR.

[0110] Additionally or alternatively, the criteria evaluated in step 701 may take into account at least whether a beam belonging to the first set obtained by the inference execution is included in the second set used for the inference execution. The criteria may take into account at least whether a beam with the highest quality (i.e., a top 1 beam) from the first set obtained by the inference execution is included in the second set used for the inference execution. UE1 may determine that the criteria are met based at least on the fact that the beam with the highest quality (i.e., a top 1 beam) from the first set obtained by the inference execution is included in the second set used for the inference execution.

[0111] In step 701, UE1 may perform a determination similar to the determination of the entry conditions of one or any combination of Event AIML5, Event AIML6, and Event AIML7 described in the first embodiment.

[0112] UE1 may determine whether to send a notification by counting the number of times the criteria are met for multiple runs of inference. For example, UE1 may determine to send a notification in response to detecting that the number of times the criteria are met for multiple runs of inference is equal to or greater than a predetermined value within a predetermined period of time. UE1 may receive configuration information from the network to define the predetermined value and the predetermined period of time. RAN node 2 may send this configuration information to UE1 via an RRC message (e.g., RRC Reconfiguration).

[0113] In some implementations, the PHY layer of UE1 may evaluate the criteria (step 701), and the MAC layer of UE1 may transmit a notification of the criteria (step 702). The PHY layer of UE1 may transmit an indication to the MAC layer of UE1 each time the result of one inference run satisfies the criteria. The MAC layer may determine whether to transmit the notification by counting the number of times the indication is received. The MAC layer may determine to transmit the notification in response to detecting that the number of times the indication is received is equal to or greater than a predetermined value within a predetermined period. If there are no uplink resources for transmitting the notification, the MAC layer of UE1 may trigger the PHY layer of UE1 to transmit a Scheduling Request to RAN node 2.

[0114] According to the operations described with reference to Figure 7, UE1 can send a notification to the network (e.g., RAN node 2) regarding the UE-side AI / ML model based on a comparison between the first beam set obtained from one inference run and the second beam set used for the inference run. This enables UE1 to provide notifications to the network about specific conditions, situations, or events related to the execution of inference using the UE-side AI / ML model, for example.

[0115] 7, for example, a notification from UE1 to the network can indicate that the UE-side AI / ML model is not suitable for the current situation of UE1, or the notification can indicate to the network that the UE-side AI / ML model is not providing appropriate inference results, or the notification can indicate to the network that a situation exists in which inference using the UE-side AI / ML model is not beneficial.

[0116] FIG. 8 shows an example of the format of a notification sent from UE1 to the network. In the example of FIG. 8, the notification is sent using MAC CE 800. For example, but not limited to, the name of MAC CE 800 may be AI / ML inference failure MAC CE. UE1 can simultaneously activate multiple UE-side AI / ML models. The maximum number of simultaneously activated UE-side AI / ML models may be eight. In the example of FIG. 8, MAC CE 800 has a length of eight bits. Each bit corresponds to a respective one of the activated UE-side AI / ML models. If a criterion is met for one of the activated UE-side AI / ML models, UE1 sets the corresponding bit in MAC CE 800 to 1; otherwise, the corresponding bit is set to 0. Conversely, if a criterion is met for one of the activated UE-side AI / ML models, UE1 may set the corresponding bit in MAC CE 800 to 0; otherwise, the corresponding bit is set to 1.

[0117] 9 shows an example of the operation of the PHY and MAC layers of UE 1. In step 910, the UE MAC layer 11 initiates and continues the AI / ML inference failure detection and recovery procedure. The UE MAC layer 11 manages an inference failure indication counter (InferenceFaiIureIndication_COUNTER) and an inference failure detection timer (aIMLInferenceFailureDetectionTimer) for each activated UE-side AI / ML model. The term "failure" may be replaced with other terms such as deterioration, poor performance, uselessness, or ineffectiveness.

[0118] In step 920, the UE PHY layer 12 detects a failure of inference by the UE-side AI / ML model. Specifically, the UE PHY layer 12 detects that the above-described criteria are met. The criteria are based on a comparison between a first set (i.e., Set A) obtained from a single inference run and a second set (i.e., Set B) used for that inference run. In response to this detection, in step 921, the UE PHY layer 12 sends an inference failure instance indication (AI / ML inference failure instance indication) to the UE MAC layer 11. The inference failure indication indicates the identifier of the AI / ML model for which the inference failure was detected. This makes it easy for the UE MAC layer 11 to determine which inference failure indication counters to increment and which inference failure detection timers to restart in response to the inference failure indication.

[0119] In step 922, in response to receiving the inference failure indication, the UE MAC layer 11 increments a corresponding inference failure indication counter and starts or restarts a corresponding inference failure detection timer. When the inference failure indication counter is equal to or greater than a configured value (aIMLInferenceFailureInstanceMaxCount), the UE MAC layer 11 detects or triggers an AI / ML inference failure for the associated AI / ML model. Meanwhile, the UE MAC layer 11 resets the inference failure indication counter in response to expiration of the inference failure detection timer. That is, the UE MAC layer 11 counts successively received inference failure indications at intervals shorter than the configured timer expiration time (aIMLInferenceFailureDetectionTimer). Then, when the count reaches a maximum value (aIMLInferenceFailureInstanceMaxCount), the UE MAC layer 11 detects or triggers an AI / ML inference failure.

[0120] The following provides an example of configuration for AI / ML inference failure detection. Figure 10 shows an example of signaling between UE1 and RAN node 2. In step 1001, RAN node 2 sends configuration information for AI / ML inference failure detection to UE1. RAN node 2 may send the configuration information to UE1 via UE-dedicated RRC signaling (e.g., RRC Reconfiguration message).

[0121] The setting information in step 1001 indicates the maximum value of the inference failure indication counter (aIMLInferenceFailureInstanceMaxCount) and the expiration time of the inference failure detection timer (aIMLInferenceFailureDetectionTimer). In addition, the setting information in step 1001 indicates the definition of the criteria (or events) set for inference failure detection (e.g., event type). The setting information in step 1001 may also indicate the value of the offset used in the criteria.

[0122] 11 shows an example of the format of configuration information for AI / ML inference failure detection. The AIMLInference-FailureRecoveryConfig IE shown in FIG. 11 includes an aIMLInferenceFailureInstanceMaxCount field 1101 and an aIMLInferenceFailureDetectionTimer field 1102. The aIMLInferenceFailureInstanceMaxCount field 1101 indicates the maximum value of the inference failure indication counter. The aIMLInferenceFailureDetectionTimer field 1102 indicates the expiration time of the inference failure detection timer.

[0123] Third Embodiment A configuration example of a wireless communication system according to this embodiment is similar to the configuration example described with reference to Figures 1 and 2. This embodiment provides details related to monitoring of the UE-side AI / ML model.

[0124] The mechanism in which the UE MAC layer 11 and the UE PHY layer 12 cooperate to detect an AI / ML inference failure described in the second embodiment can also be used to detect specific situations related to AI / ML inference based on criteria other than those described in the second embodiment, such as failure, degradation, performance degradation, uselessness, or inefficiency of AI / ML inference.

[0125] The UE PHY layer 12 performs inference using the AI / ML model to generate beam predictions for a first set of one or more beams (i.e., set A) based on measurement results for a second set of beams (i.e., set B). The UE PHY layer 12 sends an indication to the UE MAC layer 11 each time the results of one inference run meet a criterion.

[0126] The UE MAC layer 11 determines whether to send a notification regarding the AI / ML model to the network (e.g., RAN node 2) by counting the number of times the indication is received. In response to determining that a notification needs to be sent, the UE MAC layer 11 sends the notification to the network. The UE MAC layer 11 may determine to send the notification in response to detecting that the number of times the indication is received is equal to or greater than a predetermined value within a predetermined period. The notification sent by the UE MAC layer 11 may be a MAC CE.

[0127] Details of the operation and settings of the UE MAC layer 11 and the UE PHY layer 12 may be the same as those described in the second embodiment.

[0128] Next, exemplary configurations of a UE 1 and a RAN node 2 related to the above-described embodiments will be described. FIG. 12 is a block diagram showing an exemplary configuration of a UE 1. A radio frequency (RF) transceiver 1201 performs analog RF signal processing for communication with a RAN node 2. The RF transceiver 1201 may include multiple transceivers. The analog RF signal processing performed by the RF transceiver 1201 includes frequency up-conversion, frequency down-conversion, and amplification. The RF transceiver 1201 is coupled to an antenna array 1202 and a baseband processor 1203. The RF transceiver 1201 receives modulation symbol data (or orthogonal frequency-division multiplexing (OFDM) symbol data) from the baseband processor 1203, generates a transmit RF signal, and provides the transmit RF signal to the antenna array 1202. The RF transceiver 1201 also generates a baseband receive signal based on the receive RF signal received by the antenna array 1202 and provides the baseband receive signal to the baseband processor 1203. The RF transceiver 1201 may include an analog beamformer circuit for beamforming, which may include, for example, multiple phase shifters and multiple power amplifiers.

[0129] The baseband processor 1203 performs digital baseband signal processing (data plane processing) and control plane processing for wireless communications. Digital baseband signal processing includes (a) data compression / decompression, (b) data segmentation / concatenation, (c) transmission format (transmission frame) generation / decomposition, (d) transmission path coding / decoding, (e) modulation (symbol mapping) / demodulation, and (f) generation of OFDM symbol data (baseband OFDM signal) using Inverse Fast Fourier Transform (IFFT). Meanwhile, control plane processing includes communication management for Layer 1 (e.g., transmit power control), Layer 2 (e.g., radio resource management and hybrid automatic repeat request (HARQ) processing), and Layer 3 (e.g., signaling related to attachment, mobility, and call management).

[0130] For example, the digital baseband signal processing by the baseband processor 1203 may include signal processing of a PDCP layer, an RLC layer, a MAC layer, and a PH layer. Also, the control plane processing by the baseband processor 1203 may include processing of a Non-Access Stratum (NAS) protocol, an RRC protocol, MAC CEs, and DCIs.

[0131] The baseband processor 1203 may perform orthogonal frequency-division multiplexing (MIMO) encoding and precoding for beamforming.

[0132] The baseband processor 1203 may include a modem processor (e.g., a Digital Signal Processor (DSP)) that performs digital baseband signal processing and a protocol stack processor (e.g., a Central Processing Unit (CPU) or a Micro Processing Unit (MPU)) that performs control plane processing. In this case, the protocol stack processor that performs control plane processing may be shared with the application processor 1204, which will be described later.

[0133] The application processor 1204 is also referred to as a CPU, MPU, microprocessor, or processor core. The application processor 1204 may include multiple processors (multiple processor cores). The application processor 1204 executes a system software program (operating system (OS)) and various application programs (e.g., a calling application, a web browser, a mailer, a camera operation application, and a music playback application) read from the memory 1206 or other memories, thereby realizing various functions of the UE 1.

[0134] In some implementations, the baseband processor 1203 and the application processor 1204 may be integrated on a single chip, as indicated by the dashed line (1205) in Figure 12. In other words, the baseband processor 1203 and the application processor 1204 may be implemented as a single System on Chip (SoC) device 1205. An SoC device is sometimes called a system Large Scale Integration (LSI) or chipset.

[0135] The memory 1206 is volatile memory, nonvolatile memory, or a combination thereof. The memory 1206 may include multiple physically independent memory devices. The volatile memory may be, for example, static random access memory (SRAM), dynamic RAM (DRAM), or a combination thereof. The nonvolatile memory may be mask read only memory (MROM), electrically erasable programmable ROM (EEPROM), flash memory, a hard disk drive, or any combination thereof. For example, the memory 1206 may include an external memory device accessible from the baseband processor 1203, the application processor 1204, and the SoC 1205. The memory 1206 may also include an internal memory device integrated within the baseband processor 1203, the application processor 1204, or the SoC 1205. Furthermore, the memory 1206 may include memory within a Universal Integrated Circuit Card (UICC).

[0136] The memory 1206 may store one or more software modules (computer programs) 1207 containing instructions and data for processing by the UE 1. In some implementations, the baseband processor 1203 or the application processor 1204 may be configured to read and execute the software modules 1207 from the memory 1206 to perform the processing of the UE 1 described in one or more of the embodiments.

[0137] It should be noted that the control plane processing and operations performed by UE 1 described in the above embodiment can be realized by elements other than the RF transceiver 1201 and the antenna array 1202, namely, at least one of the baseband processor 1203 and the application processor 1204, and the memory 1206 storing the software module 1207.

[0138] FIG. 13 is a block diagram showing an example configuration of a RAN node 2. Referring to FIG. 13, the RAN node 2 includes an RF transceiver 1301, a network interface 1303, a processor 1304, and a memory 1305. The RF transceiver 1301 performs analog RF signal processing for communication with a UE 1. The RF transceiver 1301 may include multiple transceivers. The RF transceiver 1301 is coupled to an antenna array 1302 and a processor 1304. The RF transceiver 1301 receives modulation symbol data from the processor 1304, generates a transmit RF signal, and provides the transmit RF signal to the antenna array 1302. The RF transceiver 1301 also generates a baseband receive signal based on the receive RF signal received by the antenna array 1302 and provides the baseband receive signal to the processor 1304. The RF transceiver 1301 may include an analog beamformer circuit for beamforming. The analog beamformer circuit may include, for example, multiple phase shifters and multiple power amplifiers.

[0139] The network interface 1303 is used to communicate with network nodes (e.g., other RAN nodes, and control and forwarding nodes of the core network), and may include, for example, a network interface card (NIC) compliant with the IEEE 802.3 series.

[0140] The processor 1304 performs digital baseband signal processing (data plane processing) and control plane processing for wireless communication. The processor 1304 may include multiple processors. For example, the processor 1304 may include a modem processor (e.g., a Digital Signal Processor (DSP)) that performs digital baseband signal processing and a protocol stack processor (e.g., a CPU or MPU) that performs control plane processing. The processor 1304 may include a digital beamformer module for beamforming. The digital beamformer module may include a MIMO encoder and a precoder.

[0141] The memory 1305 is configured by a combination of volatile memory and non-volatile memory. The volatile memory is, for example, SRAM or DRAM, or a combination thereof. The non-volatile memory is, for example, MROM, EEPROM, flash memory, or a hard disk drive, or any combination thereof. The memory 1305 may include storage located remotely from the processor 1304. In this case, the processor 1304 may access the memory 1305 via the network interface 1303 or other I / O interface.

[0142] The memory 1305 may store one or more software modules (computer programs) 1306 containing instructions and data for processing by the RAN node 2. In some implementations, the processor 1304 may be configured to read and execute the software modules 1306 from the memory 1305 to perform the processing of the RAN node 2 described in one or more of the embodiments.

[0143] It should be noted that the control plane processing and operations performed by the RAN node 2 described in the above embodiment can be realized by elements other than the RF transceiver 1301 and the antenna array 1302, namely the processor 1304 and the memory 1305 storing the software module 1306.

[0144] As described with reference to Figures 12 and 13, each of the processors included in the UE 1 and the RAN node 2 according to the above-described embodiments may execute one or more programs including instructions for causing a computer to perform the algorithms described with reference to the drawings. The programs include instructions (or software code) that, when loaded into a computer, cause the computer to perform one or more functions described in the embodiments. The programs may be stored on a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, computer-readable media or tangible storage media include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technology, CD-ROM, digital versatile disk (DVD), Blu-ray disc or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device. The programs may also be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, transitory computer-readable media or communication media include electrical, optical, acoustic, or other forms of propagated signals.

[0145] The above-described embodiments are merely examples of application of the technical ideas obtained by the inventors of the present invention. In other words, the technical ideas are not limited to the above-described embodiments, and various modifications are possible.

[0146] For example, some or all of the above embodiments may also be described as, but are not limited to, the following appendices. Some or all of the elements (e.g., configurations and functions) described in appendices directed to devices (e.g., wireless terminals) may naturally also be described as appendices directed to methods and programs. For example, some or all of the elements described in appendices 2-8, which are dependent on appendices 1, may also be described as appendices dependent on appendices 9 and 10, due to the same dependency relationship as appendices 2-8. Some or all of the elements described in any appendice may be applied to various hardware, software, recording means for recording software, systems, and methods.

[0147] (Supplementary Note 1) A wireless terminal comprising: means for evaluating a criterion for triggering transmission of a report of results of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of beams; and means for transmitting the report to a network based at least on the evaluation of the criterion, wherein the criterion is based on a comparison of the first set obtained by a single inference run with the second set used for the inference run. (Supplementary Note 2) The wireless terminal of Supplementary Note 1, wherein the criterion is based at least on a comparison of predicted beam qualities of the first set obtained by the inference run with measured beam qualities of the second set used for the inference run. (Supplementary Note 3) The wireless terminal of Supplementary Note 2, wherein the evaluating means is configured to determine that the criterion is met at least based on a quality of a best-quality beam of the first set obtained by the inference run being worse than a quality of a best-quality beam of the second set used for the inference run by at least an offset. (Supplementary Note 4) The wireless terminal of Supplementary Note 2, wherein the evaluating means is configured to determine that the criterion is satisfied based at least on the quality of the beam with the best quality of the first set obtained by the inference execution being not better than the quality of the beam with the best quality of the second set used for the inference execution by an offset or more. (Supplementary Note 5) The wireless terminal of Supplementary Note 1, wherein the criterion takes into account at least whether a beam belonging to the first set obtained by the inference execution is included in the second set used for the inference execution. (Supplementary Note 6) The wireless terminal of Supplementary Note 1, wherein the criterion takes into account at least whether the beam with the best quality of the first set obtained by the inference execution is included in the second set used for the inference execution. (Supplementary Note 7) The wireless terminal of Supplementary Note 6, wherein the evaluating means is configured to determine that the criterion is satisfied based at least on the quality of the beam with the best quality of the first set obtained by the inference execution being included in the second set used for the inference execution.(Supplementary Note 8) The wireless terminal of any one of Supplements 1 to 7, wherein the report of the result of the inference is transmitted as a Channel State Information (CSI) report. (Supplementary Note 9) A method performed by a wireless terminal, comprising: evaluating criteria for triggering transmission of a report of the result of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams, and transmitting the report to a network based at least on the evaluation of the criteria, wherein the criteria is based on a comparison of the first set obtained by a single inference run with the second set used for the inference run. (Supplementary Note 10) A program that causes a computer to perform a method, comprising: evaluating criteria for triggering transmission of a report of the result of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams, and transmitting the report to a network based at least on the evaluation of the criteria, wherein the criteria is based on a comparison of the first set obtained by a single inference run with the second set used for the inference run. (Supplementary Note 11) A wireless terminal comprising: means for evaluating whether a result of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams satisfies a criterion; and means for transmitting a notification to a network regarding the artificial intelligence or machine learning model based at least on the criterion being satisfied, wherein the criterion is based on a comparison of the first set obtained by a single inference run with the second set used for the inference run. (Supplementary Note 12) The wireless terminal of Supplementary Note 11, wherein the transmitting means is configured to determine whether to send the notification by counting the number of times the criterion is satisfied over multiple runs of the inference.(Supplementary Note 13) The wireless terminal of Supplementary Note 12, wherein the transmitting means is configured to determine to transmit the notification in response to detecting that the number of times the criterion is satisfied for multiple runs of the inference is equal to or greater than a predetermined value within a predetermined period of time. (Supplementary Note 14) The wireless terminal of Supplementary Note 13, further comprising: means for receiving configuration information from the network for defining the predetermined value and the predetermined period of time. (Supplementary Note 15) The wireless terminal of Supplementary Note 11, further comprising: means for providing a Physical (PHY) layer and a Medium Access Control (MAC) layer, the PHY layer including the evaluating means, the MAC layer including the transmitting means, the PHY layer configured to transmit an indication to the MAC layer each time a result of one run of the inference satisfies the criterion, and the MAC layer configured to determine whether to transmit the notification by counting the number of times the indication is received. (Supplementary Note 16) The wireless terminal of Supplementary Note 15, wherein the MAC layer is configured to determine to transmit the notification in response to detecting that the number of times the indication is received is equal to or greater than a predetermined value within a predetermined period. (Supplementary Note 17) The wireless terminal of Supplementary Note 16, further comprising: means for receiving configuration information from the network for defining the predetermined value and the predetermined period. (Supplementary Note 18) The wireless terminal of any one of Supplements 11 to 17, wherein the criterion is based at least on a comparison between predicted beam qualities of the first set obtained by performing inference and measured beam qualities of the second set used for performing inference. (Supplementary Note 19) The wireless terminal of Supplementary Note 18, wherein the evaluating means is configured to determine that the criterion is met at least based on a quality of a best quality beam of the first set obtained by performing inference being worse than a quality of a best quality beam of the second set used for performing inference by at least an offset.(Supplementary Note 20) The wireless terminal of Supplementary Note 18, wherein the evaluating means is configured to determine that the criterion is satisfied based at least on the quality of the beam with the best quality of the first set obtained by the inference execution being not better than the quality of the beam with the best quality of the second set used for the inference execution by an offset or more. (Supplementary Note 21) The wireless terminal of any one of Supplements 11 to 17, wherein the criterion takes into account at least whether a beam belonging to the first set obtained by the inference execution is included in the second set used for the inference execution. (Supplementary Note 22) The wireless terminal of any one of Supplements 11 to 17, wherein the criterion takes into account at least whether the beam with the best quality of the first set obtained by the inference execution is included in the second set used for the inference execution. (Supplementary Note 23) The wireless terminal of Supplementary Note 22, wherein the evaluating means is configured to determine that the criterion is satisfied based at least on the quality of the beam with the best quality of the first set obtained by the inference execution being included in the second set used for the inference execution. (Supplementary Note 24) The wireless terminal according to any one of Supplements 11 to 23, wherein the notification is a Medium Access Control (MAC) Control Element (CE). (Supplementary Note 25) The wireless terminal according to any one of Supplements 11 to 24, further comprising: means for receiving configuration information for defining the criteria from the network. (Supplementary Note 26) The wireless terminal according to any one of Supplements 11 to 25, wherein the notification indicates that the artificial intelligence or machine learning model is not suitable for a current situation of the wireless terminal.(Supplementary Note 27) A method performed by a wireless terminal, comprising: evaluating whether results of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams satisfy a criterion, and sending a notification to a network regarding the artificial intelligence or machine learning model based at least on the criterion being satisfied, wherein the criterion is based on a comparison of the first set obtained by a single inference run with the second set used for the inference run. (Supplementary Note 28) A program causing a computer to perform a method for a wireless terminal, comprising: evaluating whether results of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams satisfy a criterion, and sending a notification to a network regarding the artificial intelligence or machine learning model based at least on the criterion being satisfied, wherein the criterion is based on a comparison of the first set obtained by a single inference run with the second set used for the inference run. (Supplementary Note 29) A wireless terminal comprising: means for providing a Physical (PHY) layer and a Medium Access Control (MAC) layer, wherein the PHY layer is configured to: perform inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of a plurality of beams; and send an indication to the MAC layer each time a result of one inference run satisfies a criterion; and the MAC layer is configured to: determine whether to send a notification regarding the artificial intelligence or machine learning model to a network by counting the number of times the indication is received; and send the notification to the network in response to determining that sending the notification is necessary. (Supplementary Note 30) The wireless terminal according to Supplementary Note 29, wherein the MAC layer is configured to determine to send the notification in response to detecting that the number of times the indication is received is equal to or greater than a predetermined value within a predetermined period.(Supplementary Note 31) The wireless terminal according to Supplementary Note 30, further comprising means for receiving configuration information for defining the predetermined value and the predetermined period from the network. (Supplementary Note 32) The wireless terminal according to any one of Supplements 29 to 31, wherein the notification is a Medium Access Control (MAC) Control Element (CE). (Supplementary Note 33) The wireless terminal according to any one of Supplements 29 to 32, wherein the criterion is based on a comparison of the first set obtained by one inference run with the second set used for the inference run. (Supplementary Note 34) The wireless terminal according to any one of Supplements 29 to 33, wherein the criterion is based at least on a comparison of a predicted beam quality of the first set obtained by the inference run with a measured beam quality of the second set used for the inference run. (Supplementary Note 35) The wireless terminal of Supplementary Note 34, wherein the PHY layer is configured to determine that the criterion is satisfied based at least on the quality of the beam with the best quality of the first set obtained by the inference execution being worse than the quality of the beam with the best quality of the second set used for the inference execution by at least an offset. (Supplementary Note 36) The wireless terminal of Supplementary Note 34, wherein the PHY layer is configured to determine that the criterion is satisfied based at least on the quality of the beam with the best quality of the first set obtained by the inference execution being not better than the quality of the beam with the best quality of the second set used for the inference execution by a difference of at least an offset. (Supplementary Note 37) The wireless terminal of any one of Supplements 29 to 33, wherein the criterion at least takes into account whether a beam belonging to the first set obtained by the inference execution is included in the second set used for the inference execution. (Supplementary Note 38) The wireless terminal according to any one of Supplementary Notes 29 to 33, wherein the criterion takes into consideration at least whether the beam with the best quality among the first set obtained by the inference execution is included in the second set used for the inference execution.(Supplementary Note 39) The wireless terminal of Supplementary Note 38, wherein the PHY layer is configured to determine that the criterion is met based at least on the fact that a beam with the best quality from the first set obtained by performing the inference is included in the second set used for performing the inference. (Supplementary Note 40) The wireless terminal of any one of Supplements 29 to 39, further comprising: means for receiving configuration information from the network to define the criterion. (Supplementary Note 41) The wireless terminal of any one of Supplements 29 to 40, wherein the notification indicates that the artificial intelligence or machine learning model is not suitable for a current situation of the wireless terminal. (Supplementary Note 42) A method performed by a wireless terminal, comprising providing a Physical (PHY) layer and a Medium Access Control (MAC) layer, wherein the PHY layer is configured to: perform inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams; and send an indication to the MAC layer each time the result of one inference run satisfies a criterion; and the MAC layer is configured to: determine whether to send a notification regarding the artificial intelligence or machine learning model to a network by counting the number of times the indication is received; and send the notification to the network in response to determining that the notification needs to be sent.(Supplementary Note 43) A program causing a computer to perform a method for a wireless terminal comprising providing a Physical (PHY) layer and a Medium Access Control (MAC) layer, wherein the PHY layer is configured to: perform inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams; and send an indication to the MAC layer each time the result of one inference run satisfies a criterion; and the MAC layer is configured to: determine whether to send a notification regarding the artificial intelligence or machine learning model to a network by counting the number of times the indication is received; and send the notification to the network in response to determining that sending the notification is necessary.

[0148] This application claims priority based on Japanese Patent Application No. 2024-124313, filed on July 31, 2024, the disclosure of which is incorporated herein by reference in its entirety.

[0149] 1 UE 2 RAN node 3 RAN controller 201 CU 211, 212 DU 231-235 TRP 241-243 Cell 1203 Baseband processor 1204 Application processor 1206 Memory 1207 Modules 1304 Processor 1305 Memory 1306 Modules

Claims

1. A wireless terminal comprising: means for evaluating criteria to trigger the transmission of a report of the results of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams; and means for transmitting the report to a network based at least on the evaluation of the criteria, wherein the criteria are based on a comparison of the first set obtained by a single inference run with the second set used for the inference run.

2. The wireless terminal of claim 1, wherein the criterion is based at least on a comparison of predicted beam qualities of the first set obtained by the inference execution with measured beam qualities of the second set used for the inference execution.

3. The wireless terminal of claim 2, wherein the evaluating means is configured to determine that the criterion is met based at least on the quality of the best quality beam of the first set obtained by the inference execution being worse than the quality of the best quality beam of the second set used for the inference execution by at least an offset.

4. The wireless terminal of claim 2, wherein the evaluating means is configured to determine that the criterion is met based at least on the quality of the best quality beam of the first set obtained by the inference execution being no better than the quality of the best quality beam of the second set used for the inference execution by more than an offset.

5. The wireless terminal of claim 1, wherein the criterion takes into account at least whether a beam belonging to the first set obtained by the inference execution is included in the second set used for the inference execution.

6. The wireless terminal of claim 1, wherein the criterion takes into account at least whether the beam with the best quality among the first set obtained by the inference execution is included in the second set used for the inference execution.

7. The wireless terminal of claim 6, wherein the evaluating means is configured to determine that the criterion is met based at least on the fact that the best quality beam of the first set obtained by the inference execution is included in the second set used for the inference execution.

8. A wireless terminal according to any one of claims 1 to 7, wherein the report of the result of the inference is transmitted as a Channel State Information (CSI) report.

9. A method performed by a wireless terminal, comprising: evaluating criteria for triggering the transmission of a report of the results of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams; and transmitting the report to a network based at least on the evaluation of the criteria, wherein the criteria are based on a comparison of the first set obtained by a single inference run with the second set used for the inference run.

10. A program causing a computer to perform a method comprising: evaluating criteria for triggering the sending of a report of results from inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams; and sending the report to a network based at least on the evaluation of the criteria, wherein the criteria are based on a comparison between the first set obtained by a single inference run and the second set used for the inference run.

11. A wireless terminal comprising: means for evaluating whether results of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams satisfy a criterion; and means for sending a notification to a network regarding the artificial intelligence or machine learning model based at least on the criterion being satisfied, wherein the criterion is based on a comparison between the first set obtained by a single inference run and the second set used for the inference run.

12. The wireless terminal of claim 11, wherein said means for transmitting is configured to determine whether to transmit said notification by counting the number of times said criteria is met over multiple runs of said inference.

13. The wireless terminal of claim 12, wherein the transmitting means is configured to determine to transmit the notification in response to detecting that the number of times the criteria is satisfied for multiple executions of the inference is equal to or greater than a predetermined value within a predetermined period of time.

14. The wireless terminal of claim 13, further comprising: means for receiving configuration information from the network for defining said predetermined value and said predetermined period of time.

15. The wireless terminal of claim 11, further comprising means for providing a Physical (PHY) layer and a Medium Access Control (MAC) layer, wherein the PHY layer includes the means for evaluating, and the MAC layer includes the means for transmitting, the PHY layer configured to send an indication to the MAC layer each time a result of one inference execution satisfies the criterion, and the MAC layer configured to determine whether to send the notification by counting the number of times the indication is received.

16. The wireless terminal of claim 15, wherein the MAC layer is configured to determine to transmit the notification in response to detecting that the number of times the indication is received is equal to or greater than a predetermined value within a predetermined period of time.

17. The wireless terminal of claim 16, further comprising: means for receiving configuration information from the network to define said predetermined value and said predetermined period of time.

18. A wireless terminal according to any one of claims 11 to 17, wherein the criterion is based at least on a comparison of predicted beam qualities of the first set obtained by the inference execution with measured beam qualities of the second set used for the inference execution.

19. The wireless terminal of claim 18, wherein the evaluating means is configured to determine that the criterion is met based at least on the quality of the best quality beam of the first set obtained by the inference execution being worse than the quality of the best quality beam of the second set used for the inference execution by at least an offset.

20. The wireless terminal of claim 18, wherein the evaluating means is configured to determine that the criterion is met based at least on the quality of the best quality beam of the first set obtained by the inference execution being no better than the quality of the best quality beam of the second set used for the inference execution by more than an offset.

21. A wireless terminal according to any one of claims 11 to 17, wherein the criterion takes into consideration at least whether a beam belonging to the first set obtained by the inference execution is included in the second set used for the inference execution.

22. A wireless terminal according to any one of claims 11 to 17, wherein the criterion takes into consideration at least whether the beam with the best quality among the first set obtained by the inference execution is included in the second set used for the inference execution.

23. The wireless terminal of claim 22, wherein the evaluating means is configured to determine that the criterion is met based at least on the best quality beam of the first set obtained by the inference execution being included in the second set used for the inference execution.

24. The wireless terminal according to any one of claims 11 to 23, wherein the notification is a Medium Access Control (MAC) Control Element (CE).

25. A wireless terminal according to any one of claims 11 to 24, further comprising means for receiving configuration information from the network for defining said criteria.

26. The wireless terminal of any one of claims 11 to 25, wherein the notification indicates that the artificial intelligence or machine learning model is not suitable for the wireless terminal's current situation.

27. A method performed by a wireless terminal, comprising: evaluating whether results of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of beams satisfy a criterion; and sending a notification to the network regarding the artificial intelligence or machine learning model based at least on the criterion being satisfied, wherein the criterion is based on a comparison of the first set obtained by a single inference run with the second set used for the inference run.

28. A program causing a computer to perform a method for a wireless terminal, comprising: evaluating whether results of inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams satisfy a criterion; and sending a notification to a network regarding the artificial intelligence or machine learning model based at least on the criterion being satisfied, wherein the criterion is based on a comparison between the first set obtained by a single inference run and the second set used for the inference run.

29. A wireless terminal comprising: means for providing a Physical (PHY) layer and a Medium Access Control (MAC) layer; wherein the PHY layer is configured to: perform inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of beams; and send an indication to the MAC layer each time the result of one inference run satisfies a criterion; and the MAC layer is configured to: determine whether to send a notification regarding the artificial intelligence or machine learning model to a network by counting the number of times the indication is received; and send the notification to the network in response to determining that sending the notification is necessary.

30. The wireless terminal of claim 29, wherein the MAC layer is configured to determine to transmit the notification in response to detecting that the number of times the indication is received is equal to or greater than a predetermined value within a predetermined period of time.

31. The wireless terminal of claim 30, further comprising: means for receiving configuration information from the network to define said predetermined value and said predetermined period of time.

32. The wireless terminal according to any one of claims 29 to 31, wherein the notification is a Medium Access Control (MAC) Control Element (CE).

33. A wireless terminal according to any one of claims 29 to 32, wherein the criteria is based on a comparison of the first set obtained by a single inference run with the second set used for the inference run.

34. A wireless terminal according to any one of claims 29 to 33, wherein the criterion is based at least on a comparison of predicted beam qualities of the first set obtained by the inference execution with measured beam qualities of the second set used for the inference execution.

35. The wireless terminal of claim 34, wherein the PHY layer is configured to determine that the criterion is met based at least on the quality of the best quality beam of the first set obtained by the inference execution being worse than the quality of the best quality beam of the second set used for the inference execution by at least an offset.

36. The wireless terminal of claim 34, wherein the PHY layer is configured to determine that the criterion is met based at least on the quality of the best quality beam of the first set obtained by the inference execution being no better than the quality of the best quality beam of the second set used for the inference execution by more than an offset.

37. A wireless terminal according to any one of claims 29 to 33, wherein the criterion takes into consideration at least whether a beam belonging to the first set obtained by the inference execution is included in the second set used for the inference execution.

38. A wireless terminal according to any one of claims 29 to 33, wherein the criterion takes into consideration at least whether the beam with the best quality among the first set obtained by the inference execution is included in the second set used for the inference execution.

39. The wireless terminal of claim 38, wherein the PHY layer is configured to determine that the criterion is met based at least on the fact that the best quality beam from the first set obtained by the inference execution is included in the second set used for the inference execution.

40. A wireless terminal according to any one of claims 29 to 39, further comprising means for receiving configuration information from said network for defining said criteria.

41. The wireless terminal of any one of claims 29 to 40, wherein the notification indicates that the artificial intelligence or machine learning model is not suitable for the wireless terminal's current situation.

42. A method performed by a wireless terminal, comprising providing a Physical (PHY) layer and a Medium Access Control (MAC) layer, wherein the PHY layer is configured to: perform inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams; and send an indication to the MAC layer each time the result of one inference run satisfies a criterion; and the MAC layer is configured to: determine whether to send a notification regarding the artificial intelligence or machine learning model to a network by counting the number of times the indication is received; and send the notification to the network in response to determining that sending the notification is necessary.

43. A program causing a computer to perform a method for a wireless terminal comprising providing a Physical (PHY) layer and a Medium Access Control (MAC) layer, wherein the PHY layer is configured to: perform inference using an artificial intelligence or machine learning model to generate beam predictions for a first set of one or more beams based on measurement results of a second set of multiple beams; and send an indication to the MAC layer each time the result of one inference run satisfies a criterion; and the MAC layer is configured to: determine whether to send a notification regarding the artificial intelligence or machine learning model to a network by counting the number of times the indication is received; and send the notification to the network in response to determining that sending the notification is necessary.

Images