Method and apparatus for performance monitoring

Abstract

A method according to an embodiment of the present specification comprises the steps of: receiving configuration information related to performance monitoring; and transmitting a report related to the performance monitoring. The report includes information indicating one of a plurality of states related to artificial intelligence / machine learning (AI / ML) functionality or an AI / ML model.

Description

Method and apparatus for performance monitoring

[0001] This specification relates to a method and apparatus for performance monitoring.

[0002] Mobile communication systems were developed to provide voice services while ensuring user mobility. However, mobile communication systems have expanded their scope to include data services as well as voice. Currently, due to the explosive increase in traffic leading to resource shortages and users demanding higher-speed services, more advanced mobile communication systems are required.

[0003] The requirements for next-generation mobile communication systems largely include the ability to accommodate explosive data traffic, a dramatic increase in transmission rates per user, a significantly increased number of connected devices, very low end-to-end latency, and high energy efficiency. To achieve this, various technologies are being researched, such as dual connectivity, massive multiple input multiple output (MMIMO), in-band full duplex, non-orthogonal multiple access (NOMA), super wideband support, and device networking.

[0004] Meanwhile, operations for performance monitoring of the AI / ML model can be performed. Based on the results of performance monitoring, operations for managing the AI / ML model (e.g., model life cycle management) (e.g., selection, (de)activation, switching, fallback) can be performed.

[0005] According to the existing methods for performance monitoring described above, the state related to an AI / ML model (or AI / ML functionality) is classified only as Normal or Abnormal. Consequently, the following problems may arise.

[0006] Actions such as model updates, switching, and (de)activates for model life cycle management may be applied inappropriately. Specifically, they may fail to reflect the gradual performance degradation or partial abnormal states of AI / ML models, leading to over- or under-response. This can result in model updates or switching being performed unnecessarily frequently, or necessary mitigation measures being delayed, causing performance degradation to accumulate. Consequently, it is difficult to guarantee that actions more appropriate for the level of performance degradation of AI / ML functionality or the AI / ML model are applied or executed.

[0007] In addition, system stability may be degraded by flapping, a phenomenon in which state transitions are repeated near the threshold.

[0008] In addition, rapid fluctuations in service quality, inefficient use of operational resources, and difficulties in applying differentiated recovery strategies based on the cause of failure may occur.

[0009] The purpose of this specification is to propose a method for solving the aforementioned problems.

[0010] The technical problems to be solved in this specification are not limited to those mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art to which this specification belongs from the description below.

[0011] A method according to an embodiment of the present specification for solving the aforementioned problem includes the steps of receiving configuration information related to performance monitoring and transmitting a report related to said performance monitoring. The report is characterized by including information indicating one of a plurality of states related to an AI / ML function (Artificial Intelligence / Machine Learning functionality) or an AI / ML model. Since the state related to the AI / ML function or AI / ML model is reported in a more detailed manner, an appropriate model management action corresponding to each state can be performed or applied.

[0012] According to the embodiments of this specification, life cycle management operations (e.g., parameter / threshold adjustment, application of correction logic, inference limiting, partial function reduction, gradual transition to an alternative model, rollback, retraining / redeployment triggers) corresponding to each of a plurality of states related to an AI / ML model or AI / ML function may be applied differentially. Accordingly, preemptive mitigation measures can be performed at the early stage of performance degradation of the AI / ML model (or AI / ML functionality) to prevent progression to failure.

[0013] Furthermore, service availability and stability can be improved by reducing unnecessary model switching or updates. Specifically, in situations of mild performance degradation, performing fine-tuning or partial updates can prevent service latency caused by model replacement and ensure continuity.

[0014] Furthermore, recovery time can be reduced by optimizing resource usage based on degradation levels. Specifically, by applying differential measures ranging from lightweight (e.g., parameter adjustment for minor degradation) to heavy (e.g., switching to a fallback model or updating the entire model for severe degradation) depending on the level of performance degradation, unnecessary network traffic and computing resource consumption can be reduced, and recovery time can be shortened.

[0015] In addition, conventionally, a single threshold is used to distinguish between two states (normal / abnormal). If performance metrics fluctuate slightly at the boundary of the single threshold between normal and abnormal states, a "ping-pong" (or flapping) phenomenon occurs, in which unnecessary model replacement and update operations occur continuously as the normal and abnormal states repeatedly switch. According to the present embodiment, multiple states are determined based on multiple thresholds (first / second thresholds). Therefore, system stability can be enhanced by preventing the aforementioned ping-pong (or flapping) phenomenon. Furthermore, it is possible to respond flexibly and adaptively to gradual data changes (e.g., Data Shift / Concept Drift) while maximizing the lifespan of existing deployed models.

[0016] The effects obtainable in this specification are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art to which this specification belongs from the description below.

[0017] Figure 1 is a diagram illustrating the overall functions from the perspective of an AI / ML model.

[0018] Figure 2 illustrates a general form of an AI / ML-related procedure performed between a network and a terminal.

[0019] Figure 3 illustrates an example of AI / ML-based beam management operation.

[0020] Figure 4 illustrates an example of an AI / ML-based CSI measurement / reporting operation.

[0021] Figure 5 illustrates an example of an AI / ML-based positioning operation.

[0022] Figure 6 is a flowchart showing an example of a CSI-related procedure.

[0023] FIG. 7 is a flowchart illustrating a method according to one embodiment of the present specification.

[0024] FIG. 8 is a flowchart illustrating a method according to another embodiment of the present specification.

[0025] FIG. 9 is a drawing showing the configuration of a first device and a second device according to an embodiment of the present specification.

[0026] In this specification, "A or B" may mean "only A," "only B," or "both A and B." Alternatively, in this specification, "A or B" may be interpreted as "A and / or B." For example, in this specification, "A, B or C" may mean "only A," "only B," "only C," or "any combination of A, B and C."

[0027] A slash ( / ) or a comma used in this specification may mean "and / or." For example, "A / B" may mean "A and / or B." Accordingly, "A / B" may mean "only A," "only B," or "both A and B." For example, "A, B, C" may mean "A, B or C."

[0028] In this specification, "at least one of A and B" may mean "only A," "only B," or "both A and B." Additionally, in this specification, the expressions "at least one of A or B" or "at least one of A and / or B" may be interpreted as synonymous with "at least one of A and B."

[0029] Additionally, in this specification, "at least one of A, B and C" may mean "only A," "only B," "only C," or "any combination of A, B and C." Also, "at least one of A, B or C" or "at least one of A, B and / or C" may mean "at least one of A, B and C."

[0030] Additionally, parentheses used in this specification may mean "for example." Specifically, when indicated as "control information (PDCCH)," "PDCCH" may be proposed as an example of "control information." In other words, "control information" in this specification is not limited to "PDCCH," and "PDCCH" may be proposed as an example of "control information." Furthermore, even when indicated as "control information (i.e., PDCCH)," "PDCCH" may be proposed as an example of "control information."

[0031] In the following explanation, 'when, if, in case of' can be replaced with 'based on'.

[0032] Technical features described individually within a single drawing in this specification may be implemented individually or simultaneously.

[0033] Hereinafter, preferred embodiments according to the present specification will be described in detail with reference to the accompanying drawings. The detailed description disclosed below, together with the accompanying drawings, is intended to describe exemplary embodiments of the present specification and is not intended to represent the only embodiment in which the present specification may be practiced. The following detailed description includes specific details to provide a complete understanding of the present specification.

[0034] In this specification, a terminal is a user-side device (user equipment, UE) or a consumer-side device, and may also be referred to as a first node that receives / transmits signals from / to a base station / second node / IAB node / Transmission-Reception Point (TRP). A terminal may correspond to a physical node or a logical node. A terminal may correspond to a user-side endpoint or an intermediate point between other endpoints. In communication between two points not limited to endpoints (including one-to-one / many-to-one / one-to-many / many-to-many communication), a terminal may correspond to a served node. A terminal may be a fixed-location node or a non-fixed-location (or mobile) node.

[0035] In this specification, a Base Station (BS) is a device on the network side and may also be referred to as a second node / IAB node / x-NodeB (x-NodeB, where x may be an abbreviation related to Radio Access Technology (RAT)) / Transmission-Reception Point (TRP). A Base Station may correspond to a physical node or a logical node. A Base Station may correspond to an endpoint on the network side or an intermediate point between other endpoints. In communication between two points not limited to endpoints (including one-to-one / many-to-one / one-to-many / many-to-many communication), a Base Station may correspond to a serving node. A Base Station may be a node with a fixed location or a node with an indefinite location.

[0036] In this specification, higher layer parameters may be set for the terminal, pre-set, or pre-defined. For example, a base station may transmit higher layer parameters to the terminal. For example, the terminal may transmit parameters such as capability to the base station as higher layer parameters. For example, higher layer parameters may be transmitted via RRC (radio resource control) signaling or MAC (medium access control) signaling.

[0037] In this specification, information / state / parameters being "configured" or "pre-configured" may be interpreted as the information / state / parameters being provided / pre-provided to the terminal through pre-defined signaling (e.g., SIB, MAC, RRC) from the base station. In this specification, information / state / parameters being "defined" or "pre-defined" may be interpreted as being known or stored in advance by the base station and the terminal without signaling between the base station and the terminal.

[0038] < AI / ML for Wireless Communication >

[0039] With the advancement of computing technology, artificial intelligence (AI) and machine learning (ML) are being adopted across various industries and technological fields. In the field of wireless communication, various discussions are underway regarding the application of AI models trained on ML; notably, the 3GPP standardization process refers to this as AI / ML. In this specification, we use the term "AI / ML" following the terminology currently in use during the 3GPP standardization discussions; however, "AI / ML" may be referred to by various other terms depending on the progress of standardization and implementation in the future. For example, it may be referred to as a "transmission / reception mode" or a "signal / channel / operation / transmission / reception configuration" set for AI / ML, but is not limited thereto. The meanings of the terms currently used in the 3GPP standardization process are briefly summarized as follows.

[0040] - AI / ML Model: Refers to a data-driven algorithm that applies AI / ML technology to generate a set of outputs containing predictive information and / or decision parameters based on a set of inputs.

[0041] - Data collection: This is the process of collecting data necessary for AI / ML model training, data analysis, and inference from network nodes, management entities, or terminals.

[0042] - AI / ML Training: An online or offline process of training an AI model by learning features and patterns that best represent data and acquire an AI / ML model trained for inference.

[0043] - Offline training: A process of training a model based on previously collected data sets, where the trained model is used or provided for future inference.

[0044] - Online Training: A method in which the model is trained in real-time upon the acquisition of new training sample data and used for inference.

[0045] - AI / ML Inference: This is the process of making predictions or deriving decisions based on collected data and AI models using trained AI models. Meanwhile, depending on whether the AI / ML model is configured on both the transmitting and receiving devices or on only one, it can be classified into (i) two-sided models and (ii) one-sided models. In the case of (i) two-sided models, cooperative inference is performed through paired AI / ML models. Cooperative inference refers to cooperation between the network and the UE, where one side performs part of the inference and the other performs the remainder. (ii) One-sided models are divided into UE-side models and network-side models. In the case of one-sided models, inference is performed entirely by the UE / network-side models.

[0046] 1. Life Cycle Management (LCM) for AI / ML models

[0047] For AI / ML models, LCM is a concept that encompasses all overall procedures for the model, such as data collection, model training, model deployment, model inference, model monitoring, and model updates.

[0048] LCMs for AI / ML models can be broadly classified into functionality-based LCMs and model-ID-based LCMs. In functionality-based LCMs, the network can direct activation, deactivation, fallback, or switching for specific functions; in this case, the target AI / ML model may not be identified by the network. In model-ID (identifier)-based LCMs, the network can direct activation, deactivation, selection, or switching for AI / ML models identified based on their AI / ML model IDs.

[0049] Figure 1 is a diagram illustrating the overall functions from the perspective of an AI / ML model.

[0050] Referring to FIG. 1, a general AI / ML functional framework can be configured to include a data collection function (10), a model training function (20), a management function (30), an inference function (40), and a model storage function (50).

[0051] The Data Collection function (10) is a function that provides input data to the Model Training function (20), Management function (30), and Inference function (40). The Data Collection function (10) performs data preparation and can provide input data processed through data preparation.

[0052] Here, training data (11) refers to data required as input for the AI / ML model training function (20). monitoring data (12) refers to data required as input for the management (30) of the AI / ML model or AI / ML function. inference data (13) refers to data required as input for the AI / ML inference function (30).

[0053] The Model Training function (20) is a function that performs AI / ML model training, validation, and testing, and can generate model performance metrics that can be used as part of the AI / ML model testing procedure. If necessary, the Model Training function (20) can perform data preparation (e.g., data pre-processing and cleaning, forming and transformation) based on the Training Data (11) delivered from the Data Collection function (10).

[0054] Trained / Updated Model (21): If there is a Model Storage function (50), it is used to transfer trained, validated, and tested AI / ML models to the Model Storage function (50) or to transfer updated versions of the models to the Model Storage function (50).

[0055] The Management function (30) is a function that monitors the operation of an AI / ML model or an AI / ML function. Additionally, the Management function (30) may make decisions to ensure appropriate inference operations based on data received from the Data Collection function (10) (i.e., Monitoring Data (12)) and / or data received from the Inference function (40) (i.e., Inference Output (41)).

[0056] Management Instruction (32) is information required as input to manage the Inference function (40). The relevant information may include the selection / (de)activation / switching of an AI / ML model or an AI / ML-based function, and may also include a fallback to a non-AI / ML operation (i.e., not relying on the inference process).

[0057] A Model Transfer / Delivery Request (33) can be used to request model(s) from Model Storage (50).

[0058] Performance Feedback / Retraining Request (31) refers to information required as input to Model Training function (20) (e.g., for the purpose of retraining or updating the model).

[0059] The inference function (40) is a function that provides output from the process of applying an AI / ML model or AI / ML function using data (i.e., inference data (13)) provided by the data collection (10) as input. Data preparation (e.g., data preprocessing and cleaning, formatting and transformation) may also be performed based on the inference data (13) delivered by the data collection (10). If necessary, the inference function (40) may also perform data preparation (e.g., data pre-processing and cleaning, forming and transformation) based on the inference data (13) provided by the data collection function (10).

[0060] Inference Output (41) is data used in the Management function (30) to monitor the performance of an AI / ML model or AI / ML function. Inference Output (41) may include the inference output of an AI / ML model generated by the Inference function (30), and the details of the inference output may vary depending on the use case.

[0061] The Model Storage function (50) is a function that stores a trained / updated model that can be used to perform the Inference function (40). The Model Storage function (50) exemplified in FIG. 1 can be used as a reference point (if any) applicable to protocol termination, model transmission / delivery, and related processes. Additionally, the Model Storage function (50) is an example and is not intended to restrict the storage location of the actual AI / ML model, and may be omitted.

[0062] Model Transfer / Delivery (51) is used to transfer an AI / ML model to an inference function.

[0063] 2. General AI / ML related procedures between the network and the terminal

[0064] Figure 2 illustrates the general form of AI / ML-related procedures performed between a network and a terminal. While Figure 1 examined the LCM from the perspective of an AI / ML model, Figure 2 describes the general form of procedures performed between a terminal and a network from the perspective of signaling / protocols.

[0065] (1) AI / ML related setup procedure

[0066] Referring to FIG. 2, an AI / ML-related configuration procedure may be performed between the network and the terminal (B05). The AI / ML-related configuration procedure may include information exchange through at least one upper-layer signaling between the terminal and the network, and / or prior preparation / subsequent operations at the terminal / network respectively before / after the upper-layer signaling.

[0067] Specifically, the configuration procedure related to AI / ML may include, but is not limited to, at least one of the following: (i) reporting the capability of the AI / ML-related terminal, (ii) data collection, (iii) model training, (iv) model delivery / transmission, (v) selection of AI / ML functions / models, and (vi) configuration of various operations performed based on the AI / ML model (e.g., AI / ML-based CSI / Positioning / Beam Management).

[0068] (i) The terminal can inform the network of its capabilities, such as models and functions related to AI / ML, that it supports through UE Capability reporting. The network can provide AI / ML-related settings to the terminal based on the terminal's capabilities related to AI / ML reported by the terminal.

[0069] (ii) AI / ML-related configuration procedures may include data collection related to the training / inference of AI / ML models and / or the provision of configuration information regarding data collection. The configuration information regarding data collection may relate to how to configure the method / operation of data collection.

[0070] (iii) AI / ML-related configuration procedures may include training AI / ML models online or offline and / or providing configuration information for AI / ML model training. The configuration information for AI / ML model training may relate to how to configure the method / behavior, etc., of training the AI / ML model.

[0071] (iv) AI / ML-related configuration procedures may include transmitting / transmitting configuration information for a model. The configuration information for a model may include parameters that constitute the AI / ML model and / or an identifier (ID) for the AI / ML model.

[0072] The provided AI / ML model may be a model trained by the network or a model that requires self-training at the terminal. Even when a model trained by the network is provided, the terminal may perform fine-tuning or retraining as necessary. Meanwhile, if a model trained by the network is provided, the terminal may provide data for training to the network.

[0073] Meanwhile, AI / ML models can be classified into Type A models, which can be identified without OTA (over-the-air) signaling, and Type B models, which are identified through OTA signaling. A model ID may be assigned during the model identification process, and this process can be subdivided into methods initiated by the terminal and methods identified by the network.

[0074] (v) The configuration procedure related to AI / ML may include the configuration of how to select AI / ML Functionality / models and / or the selection process for AI / ML Functionality / models. In UE-side AI / ML models or two-sided AI / ML models, the selection of the UE part may be performed through instructions / signaling from the network or the terminal may select it itself. The selection of AI / ML Functionality / models may be performed when multiple AI / ML Functionality / models are configured / provided.

[0075] (vi) The configuration procedure related to AI / ML may include configuration information for various inference operations performed based on AI / ML models, e.g., AI / ML-based CSI measurement / reporting, AI / ML-based positioning, and / or AI / ML-based beam management.

[0076] (2) Operation based on inference by AI / ML models

[0077] Referring again to FIG. 2, the network and / or terminal can perform inference of the AI / ML model through the trained AI / ML model and perform various operations based on the inference of the AI / ML model (B10). If the AI / ML model is a one-sided model, the inference of the AI / ML model can be performed at either the network or the terminal where the AI / ML model is configured. If the AI / ML model is a two-sided model, each part of the inference of the AI / ML model can be performed at the network and the terminal, and depending on the implementation, such inference can be performed cooperatively between the network and the terminal.

[0078] (i) Actions performed based on the inference of an AI / ML model may include AI / ML-based CSI measurement / reporting. AI / ML-based CSI measurement / reporting is intended to improve CSI feedback and may be related to overhead reduction / CSI compression, accuracy improvement, and / or CSI prediction.

[0079] (ii) Actions performed based on the inference of an AI / ML model may include AI / ML-based beam management. AI / ML-based beam management may be related to beam prediction in the time domain, reduction of overhead / latency in the spatial domain, and / or improvement of beam selection accuracy.

[0080] (iii) Actions performed based on the inference of an AI / ML model may include AI / ML-based positioning. AI / ML-based positioning may be relevant to improving positioning accuracy in various scenarios, for example, in non-line-of-sight environments.

[0081] (3) Procedures for AI / ML management

[0082] The network and / or terminal can perform procedures for the management of AI / ML Functionality / model or the settings therefor (B15).

[0083] The network and / or terminal may perform monitoring of AI / ML Functionality / model during the AI / ML model inference or operation based thereon (B10) for the management procedure (B15).

[0084] The management procedure may include, for example, at least one of activation / deactivation, switching, model update, and / or fallback operation for AI / ML Functionality / model. For the signaling of the management procedure, various 3GPP signaling schemes, such as RRC, MAC-CE, DCI, etc., may be used.

[0085] As an example of model switching, multiple model groups are configured, and switching between them can be performed based on models having a common model structure or partially common substructures, and models within the same group may be associated with different input / output formats or processing.

[0086] Model updating involves modifying the parameters used by the model to suit channel conditions that change over time, and fine-tuning is an example of model updating.

[0087] Fallback: In a wireless communication system using an AI / ML model, this may refer to the operation of not using the AI / ML model or operating in a pre-configured / defined default mode when the reliability of the AI / ML model decreases due to internal or external environmental factors.

[0088] For example, the decision on whether to perform a management procedure can be made by the network. For instance, the network may decide to perform the management procedure upon network initiation, or the network may decide to perform the management procedure upon terminal initiation and request.

[0089] As another example, the decision on whether to perform a management procedure can be made by the terminal. For instance, the terminal's decision on the management procedure may be triggered when an event condition set by the network is satisfied, performed by reporting the terminal's decision to the network, or performed autonomously by the terminal.

[0090] 3. Specific operation examples based on AI / ML model inference

[0091] (1) Beam management

[0092] Figure 3 illustrates an example of AI / ML-based beam management operation.

[0093] Referring to FIG. 3, the network / terminal can perform a configuration procedure related to AI / ML-based beam management (C05). The network / terminal can perform a configuration procedure for an AI / ML model to be used for AI / ML-based beam management, and an exchange of configuration information for upper-layer signaling for AI / ML-based beam management. For example, at least one of information related to model inference, configuration for a first set / second set beam, monitoring performance, and assistance information for data collection and beam measurement may be signaled.

[0094] The network / terminal can perform measurements on the first set of beams (C10). The beam measurements may be related to RSRP measurements.

[0095] A network / terminal can obtain information about a second set of beams based on measurement results for a first set of beams (C15). For example, the network / terminal can perform AI / ML inference by using the measurement results for the first set of beams as AI / ML input data. Beam ID information may also be additionally provided as AI / ML input data. Information about the second set of beams may correspond to AI / ML output data. The AI / ML output data may be related to, for example, the probability that each beam will become a top-N beam, the predicted RSRP, etc., for predicting future beam quality, but is not limited thereto.

[0096] According to an embodiment, the network / terminal can transmit and receive information about the acquired second set of beams.

[0097] Specifically, AI / ML-based beam management operations may include at least one of the following BM-Case 1 and BM-Case 2.

[0098] - BM-Case 1: Prediction of the second set of DL beams in the spatial domain through the first set of beam measurements

[0099] - BM-Case 2: Prediction of the second set of DL beams in the time domain through the first set of beam measurements

[0100] In BM-Case 1 and / or 2, both AI / ML model training and inference may be performed on the network or on the terminal. The first set of beams and the second set of beams may be different beams. Or the first set of beams may be a subset of the second set of beams. Or, particularly in BM-Case 2, the first set of beams and the second set of beams may be the same beam.

[0101] The report corresponding to the inference of the UE-side model for BM-Case 1 may relate to the RSRP for the predicted top N beams. The report may include, for example, the predicted RSRP values, and as an example, the predicted RSRP values ​​may be reported together with the actual measured RSRP.

[0102] UE-side AI / ML model inference for BM-Case 2 can report inference results for N future time points through a single report. The report for each time point can correspond to the report in BM-Case 1.

[0103] For performance monitoring of the UE-side model for BM-Case 1 / 2, (i) network-side performance monitoring and / or (ii) UE-assisted performance monitoring may be supported. (i) For network-side performance monitoring, the terminal may report information necessary for the network to calculate performance metrics, for example, by reporting measurement results (e.g., RSRP) and / or RS index for a set of resources for monitoring. (ii) For UE-assisted performance monitoring, the terminal may calculate performance metrics.

[0104] With respect to the NW-side model for BM-Case 1 / 2, quantization of the reported RSRP may be supported, for example, differential RSRP reporting may be supported along existing quantization steps and ranges. The reported content may include information on the RSRP and the corresponding upper N beam, where N can be set by the network.

[0105] With respect to the configuration of the first set of beams and the second set of beams of the UE-side model of BM Case-1, two resource sets may be configured separately for each of the first set and the second set, and the resource sets may be provided through CSI reporting settings. The terminal may perform inference / measurement on the resource set of the first set of beams. The terminal may not be expected to perform measurement / inference on the resource set of the second set of beams. The beam information in the inference report may include resource set information for the first set.

[0106] In relation to the UE-side model, the associated ID may be provided through the CSI framework. The terminal may assume identical / similar characteristics for DL ​​transmit beams / sets (lists) for the same associated ID.

[0107] Regarding UE-assisted performance monitoring for the UE-side models of BM-Case 1 and 2, the following methods may be considered.

[0108] i) Compare prediction results based on resources for monitoring and use the top 1 or top K beam prediction accuracy.

[0109] ii) Use RSRP difference information based on RSRP measurements of resources for monitoring and actual RSRP measurements for at least one of the top N prediction beams.

[0110] iii) Use the difference information between the measured RSRP and the predicted RSRP for the corresponding beam of the resources for monitoring.

[0111] iv) Probability information that the predicted beam will become one of the top 1 or N beams

[0112] For reporting inference results for the UE-side model, quantization of RSRP may be supported, and differential RSRP with existing quantization steps may be supported. The scope of RSRP reporting is such that differential RSRP among multiple beams is supported in the case of BM-case 1, and differential RSRP among multiple beams at multiple time points is supported in the case of BM-case 2.

[0113] For BM-Case 2 of the UE-side model, the network can be configured to report inferences about N future times to the terminal.

[0114] (2) CSI prediction and / or compression

[0115] Figure 4 illustrates an example of an AI / ML-based CSI measurement / reporting operation.

[0116] Referring to FIG. 4, the network / terminal can perform a configuration procedure related to AI / ML-based CSI (D05). The network / terminal can perform a configuration procedure for an AI / ML model to be used for AI / ML-based CSI, and for the exchange of configuration information for upper-layer signaling for AI / ML-based CSI measurement / reporting. For example, at least one of information related to model inference, settings for RS / resources to be used for CSI measurement, monitoring performance, data collection, and conditions / resources for CSI reporting may be signaled.

[0117] The terminal can perform CSI measurements based on AI / ML model inference (D10). The AI / ML model used by the terminal for CSI measurements may be a UE-side AI / ML model corresponding to a one-side AI / ML model, or an AI / ML model corresponding to the terminal part of a two-side AI / ML model.

[0118] The terminal may report CSI to the network based on the results of CSI measurements (D15). CSI reporting may be performed periodically or non-periodically depending on the configuration, and in the case of non-period CSI reporting, network instructions (not shown), such as DCI, that trigger it may be additionally signaled. CSI reporting may include AI / ML-based CSI content and may additionally include legacy CSI content (e.g., non-AI / ML-based RI, PMI, CQI, etc.) (depending on the configuration / scheduling). AI / ML-based CSI content may be related to at least one of 1) CSI compression to reduce the overhead of CSI reporting and 2) CSI prediction for future time points in the time domain.

[0119] The network can acquire CSI based on the terminal's CSI report.

[0120] If a two-sided AI / ML model is configured, the network can reconstruct the CSI by using the terminal's CSI report as input data to the AI / ML model configured in the network (D20). The inference (output) of the AI / ML model configured in the network may be the reconstructed CSI. In such a two-sided AI / ML model, the terminal-side AI / ML model part can be understood as a CSI encoder, and the network-side AI / ML model part can be understood as a concept similar to a CSI decoder.

[0121] CSI compression is CSI compression in the spatial-frequency domain and can primarily be based on two-sided AI / ML models. CSI prediction can primarily be based on one-sided, specifically UE-side AI / ML models.

[0122] In CSI compression based on a two-sided AI / ML model, AI / ML model training may include at least one of the following: (i) Type 1, in which the two-sided AI / ML model is jointly trained at either the terminal or the network; (ii) Type 2, in which the terminal and the network each jointly train the corresponding parts of the two-sided AI / ML model; and (iii) Type 3, in which the terminal and the network each separately train the corresponding parts of the two-sided AI / ML model, wherein the training of the terminal is mainly related to CSI generation and the training of the network is mainly related to CSI reconstruction. Joint training means that the CSI generation / reconstruction models are trained in the same loop for forward / backward delays, and separate training may mean a sequential method in which one of the terminals or the network starts training first, and then the other performs training.

[0123] (3) Positioning

[0124] Figure 5 illustrates an example of an AI / ML-based positioning operation.

[0125] Referring to FIG. 5, the network / terminal can perform a configuration procedure related to AI / ML-based positioning (E05). The network / terminal can perform a configuration procedure for an AI / ML model to be used for AI / ML-based positioning, and an exchange of configuration information for upper-layer signaling for AI / ML-based positioning. For example, at least one of information related to model inference, configuration for positioning RS, monitoring performance, and assistance information for data collection and positioning measurement may be signaled.

[0126] The network / terminal can perform a measurement for positioning (E10). The measurement for positioning may be related to PRS and / or SRS measurements.

[0127] The network / terminal can obtain information about terminal positioning based on measurement results (E15). For example, the network / terminal can perform AI / ML inference by using measurement results for PRS / SRS as AI / ML input data. Information about terminal positioning may correspond to AI / ML output data. The AI / ML output data may be, for example, terminal location or assistance information that serves as the basis for determining terminal location, but is not limited thereto.

[0128] According to an embodiment, the network / terminal can transmit and receive information regarding the acquired terminal positioning.

[0129] Specifically, AI / ML-based positioning may include at least one of (i) direct AI / ML positioning and / or (ii) AI / ML-assisted positioning. In (i) direct AI / ML positioning, the output of the AI / ML model includes the UE location. In (ii) AI / ML-assisted positioning, the output of the AI / ML model may be a new measurement result and / or an improvement on an existing measurement (e.g., LoS / NLoS identification, timing and / or angle of the measurement, likelihood of the measurement, etc.).

[0130] The data sample collected for training data collection may include at least one of the following parts.

[0131] - Part A: Channel Measurement, Quality Indicators of Channel Measurement, Time Stamps of Channel Measurement

[0132] - Part B: ground truth label (or its approximation), label quality indicators, label timestamp

[0133] The training data samples in Part A and Part B are for the same terminal (e.g., PRU or Non-PRU UE) and may be for the same location associated with Part B.

[0134] Measurements for positioning can be classified into sample-based measurements and path-based measurements.

[0135] - In sample-based measurements, the measurement consists of Nt' samples of the channel response estimated in the time domain. Timing information for the Nt' samples is reported as the timing granularity T, where T = 2k x Tc. k represents the timing reporting granularity factor, and Tc represents the base time unit of the corresponding wireless communication system. Nt' and k may be signaling parameters. Timing information can be defined as a value relative to a reference time.

[0136] - Path-based measurement refers to a measurement defined in existing wireless communication systems.

[0137] AI / ML-based positioning may be related to at least one of the following detailed cases.

[0138] - (i) Case 1: UE-based positioning using a UE-side model, in the case of direct AI / ML or AI / ML assisted positioning

[0139] - (ii) Case 2a: UE-assisted / LMF-based positioning using a UE-side model, in the case of AI / ML-assisted positioning

[0140] - (iii) Case 2b: UE-assisted / LMF-based positioning using an LMF-side model, in the case of direct AI / ML positioning

[0141] - (iv) Case 3a: NG-RAN node assisted positioning using a gNB-side model, in the case of AI / ML assisted positioning

[0142] - (v) Case 3b: NG-RAN node-assisted positioning using an LMF-side model, in the case of direct AI / ML positioning

[0143] (i) Regarding model performance monitoring in Case 1, the following options may be considered for calculating model performance metrics in label-based model monitoring.

[0144] 1) Option A. Calculate monitoring metrics on the target terminal side

[0145] - Option A-1: ​​At least part of the information regarding the ground truth label of the target terminal is generated in the LMF and provided to the target terminal. For example, the target terminal and / or base station sends measurement results to the LMF, and the LMF can derive information regarding the ground truth label from this.

[0146] - Option A-2: At least some of the information regarding location calculation assistance data is provided from the LMF to the target UE.

[0147] - Option A-3: As a method to reuse assistance data previously provided from the LMF to the target terminal, the PRU measurement results and the corresponding PRU location information are provided from the LMF to the target terminal.

[0148] - Option A-4: PRU measurement and PRU location are provided from the PRU to the target UE.

[0149] 2) Option B. LMF calculates monitoring metrics

[0150] - Option B-1: At least the inference result of the target terminal (i.e., the model output corresponding to the channel measurement of the target terminal) can be transmitted to the LMF by the target terminal.

[0151] - Option B-2: Channel measurements of the PRU are provided to the target terminal through the LMF, and inference results (i.e., model output corresponding to the channel measurements of the PRU) can be transmitted from the target terminal to the LMF.

[0152] (iv) Regarding the generation of training data in Case 3a, at least LMF can generate labels and related data (e.g., time stamp).

[0153] (iv) For calculating model performance monitoring metrics in label-based model monitoring in Case 3a, the following options A and B may be considered.

[0154] - Option A: The NG-RAN node performs monitoring metric calculations for its model

[0155] - Option B: LMF performs monitoring metric calculations for models located on the NG-RAN node.

[0156] (iv) For generating training data for Case 3a and (v) Case 3b, measurements and related data (e.g., timestamps) may be generated at TRP / base stations.

[0157] (v) For base station channel measurements reported to the LMF (location management server) in Case 3b, timing information can be expressed as a value relative to the UL RTOA reference time T0+tSRS.

[0158] (iii) Time domain channel measurements supported for reporting in Case 2b and (v) Case 3b may include timing information and / or power information corresponding to the timing information.

[0159] For the definition of a sample-based measurement, Nt' samples may be selected from a list of consecutive Nt samples, and the Nt samples may have a particle size at time T.

[0160] In relation to sample-based measurements, the Nt' samples selected for network measurement may be those with the highest power.

[0161] Regarding positioning based on Case 3b, for sample-based measurements, LMF can signal parameter values ​​such as Nt, Nt', and k to the base station.

[0162] < CSI Related Operations >

[0163] Channel state information (CSI) may include at least one of a channel quality indicator (CQI), a precoding matrix indicator (PMI), a CSI-RS resource indicator (CRI), an SS / PBCH block resource indicator (SSBRI), a layer indicator (LI), a rank indicator (RI), Layer 1-RSRP (Layer 1-Reference Signal Received Power), and / or Layer 1-SINR (Layer 1-Signal-to-Interference-plus-Noise Ratio).

[0164] In the case of the CSI prediction described below, the CSI related to the prediction may include at least one of the predicted CQI (predicted CQI, P-CQI), predicted PMI (predicted PMI, P-PMI), predicted CRI (predicted CRI, P-CRI), predicted SSBRI (predicted SSBRI, P-SSBRI), predicted LI (predicted LI, P-LI), predicted RI (predicted RI, P-RI), predicted L1-RSRP (predicted L1-RSRP, P-L1-RSRP) and / or predicted L1-SINR (predicted L1-SINR, P-L1-SINR).

[0165] When monitoring performance / accuracy related to the AI / ML model or AI / ML functionality described below, the CSI related to prediction accuracy may include a Prediction Accuracy Indicator (PAI). For example, the PAI may indicate the accuracy of predicted downlink reference signal(s) (e.g., predicted CRI(s) and / or predicted SSBRI(s)), and the PAI may be interpreted / replaced as a Reference Signal-Prediction Accuracy Indicator (RS-PAI). For example, the PAI may indicate the accuracy of predicted CSI (e.g., predicted PMI), and the PAI may be interpreted / replaced as a Channel State Information-Prediction Accuracy Indicator (CSI-PAI).

[0166] Figure 6 is a flowchart showing an example of a CSI-related procedure.

[0167] Referring to FIG. 6, to perform one of the uses of CSI-RS, a terminal (e.g., user equipment, UE) receives configuration information related to CSI from a base station (e.g., general Node B, gNB) via radio resource control (RRC) signaling (S610).

[0168] The configuration information related to the above CSI may include at least one of CSI-IM (interference management) resource information, CSI measurement configuration information, CSI resource configuration information (e.g., M≥1 CSI-ResourceConfig resource setting), CSI-RS resource information, or CSI report configuration information (e.g., N≥1 CSI-ReportConfig reporting setting). As an example, the configuration information may include at least one of one or more CSI resource settings and / or one or more CSI reporting settings.

[0169] For example, the configuration information may include a first CSI resource setting for measurement and a second CSI resource setting for prediction. As a specific example, the measurement related to the prediction of CSI described below may be performed based on the first CSI resource setting. The terminal may perform L1-RSRP measurements on CSI-RS resources or SS / PBCH block resources associated with the first CSI resource setting. As a specific example, the prediction of CSI described below may be performed based on the second CSI resource setting. Based on the L1-RSRP measurements, the terminal may perform predictions on CSI-RS resources or SS / PBCH block resources associated with the second CSI resource setting. In other words, using L1-RSRPs as measurement metrics, the best CRI / best SSBRI (e.g., P-CRI(s), P-SSBRI(s)) may be predicted.

[0170] For example, the above configuration information may include a first CSI reporting setting related to prediction and a second CSI reporting setting related to prediction accuracy.

[0171] Information related to CSI resource configuration can be expressed as CSI-ResourceConfig IE. Information related to CSI resource configuration defines a group including at least one of an NZP (non-zero power) CSI-RS resource set, a CSI-IM resource set, or a CSI-SSB resource set. That is, the information related to CSI resource configuration includes a CSI-RS resource set list, and the CSI-RS resource set list may include at least one of an NZP CSI-RS resource set list, a CSI-IM resource set list, or a CSI-SSB resource set list. A CSI-RS resource set is identified by a CSI-RS resource set ID, and one resource set includes at least one CSI-RS resource. Each CSI-RS resource is identified by a CSI-RS resource ID.

[0172] Information related to CSI report configuration (e.g., CSI-ReportConfig IE) includes a reportConfigType parameter representing time domain behavior and a reportQuantity parameter representing the CSI-related quantity to be reported. The time domain behavior may be periodic, aperiodic, or semi-persistent.

[0173] The above reportQuantity parameter includes the channel quality indicator (CQI), precoding matrix indicator (PMI), CSI-RS resource indicator (CRI), SS / PBCH block resource indicator (SSBRI), layer indicator (LI), rank indicator (RI), L1-RSRP (Layer 1-Reference Signal Received Power), L1-SINR (Layer 1-Signal-to-Interference-plus-Noise Ratio), predicted CQI (P-CQI), predicted PMI (P-PMI), predicted CRI (P-CRI), predicted SSBRI (P-SSBRI), predicted LI (P-LI), predicted RI (P-RI), predicted L1-RSRP (P-L1-RSRP), and / or predicted L1-SINR (predicted It can be set to a value representing at least one of L1-SINR, P-L1-SINR) and / or Prediction Accuracy Indicator (PAI) (or RS-PAI).

[0174] For example, the reportQuantity parameter can be set to cri, ssb-Index, cri-RSRP, or ssb-Index-RSRP. cri represents the CSI-RS resource indicator (CRI). RSRP represents the L1-RSRP (Layer 1-Reference Signal Received Power). ssb-Index represents the SS / PBCH block resource indicator (SSBRI).

[0175] For example, the reportQuantity parameter can be set to p-cri, p-ssb-index, p-cri-RSRP, or p-ssb-index-RSRP. p-cri represents the predicted CRI (predicted CRI, P-CRI). p-ssb-index represents the predicted SSBRI (predicted SSBRI, P-SSBRI). p-cri-RSRP represents the predicted CRI (predicted CRI, P-CRI) and predicted L1-RSRP (predicted L1-RSRP, P-L1-RSRP). p-ssb-index-RSRP represents the predicted SSBRI (predicted SSBRI, P-SSBRI) and predicted L1-RSRP (predicted L1-RSRP, P-L1-RSRP).

[0176] For example, the reportQuantity parameter can be set to pai (or rs-pai). pai (or rs-pai) represents PAI (or RS-PAI).

[0177] The measurement resource may include settings for downlink signals and / or downlink resources for which the terminal will perform measurements to determine feedback information. The measurement resource may be set as a set of ZP and / or NZP CSI-RS resources associated with a CSI reporting setting. The NZP CSI-RS resource set may include a CSI-RS set or an SSB set. For example, L1-RSRP may be measured against a CSI-RS set or against an SSB set.

[0178] The terminal measures the CSI based on configuration information related to the above CSI (S620). The CSI measurement may include (1) a process of receiving the terminal's CSI-RS (S621) and (2) a process of computing the CSI through the received CSI-RS (S622). The terminal reports the CSI to the base station (S630).

[0179] resource setting

[0180] Each CSI resource setting 'CSI-ResourceConfig' contains a configuration for S≥1 CSI resource sets (given by the higher layer parameter csi-RS-ResourceSetList). The CSI resource setting corresponds to the CSI-RS-resourcesetlist, where S represents the number of configured CSI-RS resource sets. Here, the list of S≥1 CSI resource sets includes either or both of the NZP CSI-RS resource set(s) and the SS / PBCH block (SSB) set(s) used for L1-RSRP computation, or includes CSI-IM resource set(s).

[0181] Next, one or more CSI resource settings for channel measurement (CM) and interference measurement (IM) are established through higher layer signaling.

[0182] - CSI-IM resource for interference measurement.

[0183] - NZP CSI-RS resources for interference measurement.

[0184] - NZP CSI-RS resources for channel measurement.

[0185] That is, the CMR (channel measurement resource) may be an NZP CSI-RS for CSI acquisition, and the IMR (Interference measurement resource) may be an NZP CSI-RS for CSI-IM and IM.

[0186] Here, CSI-IM (or ZP CSI-RS for IM) is primarily used for inter-cell interference measurements.

[0187] Also, the NZP CSI-RS for IM is mainly used for intra-cell interference measurement from multi-users.

[0188] A UE can assume that the CSI-RS resource(s) for channel measurement set for one CSI reporting and the CSI-IM / NZP CSI-RS resource(s) for interference measurement (when the NZP CSI-RS resource(s) are used for interference measurement) have a QCL relationship with respect to 'QCL-TypeD' on a resource-by-resource basis.

[0189] As examined, resource setting can refer to a resource set list.

[0190] For aperiodic CSI, each trigger state set using the higher layer parameter CSI-AperiodicTriggerState is associated with one or more CSI-ReportConfigs, and each CSI-ReportConfig is linked to a periodic, semi-persistent, or aperiodic resource setting.

[0191] One reporting setting (e.g., CSI-ReportConfig) can be associated with up to three resource settings (e.g., CSI-ResourceConfig). For example, one CSI reporting setting may include the ID (e.g., CSI-ResourceConfigId) of at least one CSI resource setting. The at least one CSI resource setting may include a CSI resource setting associated with a measurement.

[0192] Beam Management (BM)

[0193] BM procedures are L1 (layer 1) / L2 (layer 2) procedures for acquiring and maintaining a set of base station (e.g., gNB, TRP, etc.) and / or terminal (e.g., UE) beams that can be used for downlink (DL) and uplink (UL) transmission / reception, and may include the following procedures and terms.

[0194] - Beam measurement: An operation in which a base station or UE measures the characteristics of a received beamforming signal.

[0195] - Beam determination: The operation in which a base station or UE selects its transmit beam (Tx beam) / receive beam (Rx beam).

[0196] - Beam sweeping: An operation that covers a spatial area using transmitting and / or receiving beams for a set time interval in a predetermined manner.

[0197] - Beam report: An operation in which the UE reports information about the beam-formed signal based on beam measurements.

[0198] The BM procedure can be divided into (1) a DL BM procedure using an SS (synchronization signal) / PBCH (physical broadcast channel) Block or CSI-RS, and (2) a UL BM procedure using an SRS (sounding reference signal).

[0199] In addition, each BM procedure may include Tx beam sweeping to determine the Tx beam and Rx beam sweeping to determine the Rx beam.

[0200] DL BM

[0201] The DL BM procedure may include (1) transmission to beamformed DL RS (reference signals) of the base station (e.g., CSI-RS or SS Block (SSB)) and (2) beam reporting of the terminal.

[0202] Here, beam reporting may include preferred DL RS ID(identifier)(s) and the corresponding L1-RSRP(Reference Signal Received Power).

[0203] The above DL RS ID may be SSBRI (SSB Resource Indicator) or CRI (CSI-RS Resource Indicator).

[0204] An example of beam forming using SSB and CSI-RS will be examined in detail below.

[0205] SSB beams and CSI-RS beams can be used for beam measurement. The measurement metric is L1-RSRP per resource / block. SSB is used for coarse beam measurement, while CSI-RS can be used for fine beam measurement. SSB can be used for both Tx beam sweeping and Rx beam sweeping.

[0206] Rx beam sweeping using SSBs can be performed as the UE changes the Rx beam across multiple SSB bursts for the same SSBRI. Here, one SS burst includes one or more SSBs, and one set of SS bursts includes one or more SSB bursts.

[0207] The DL BM procedure is examined below.

[0208] Configuration for beam reporting using SSB is performed during CSI / beam configuration in the RRC connected state (or RRC connected mode).

[0209] - The terminal receives configuration information from the base station. As a specific example, the terminal receives from the base station a CSI-ResourceConfig IE containing a CSI-SSB-ResourceSetList containing SSB resources used for BM.

[0210] Table 1 shows an example of CSI-ResourceConfig IE. As shown in Table 1, BM configuration using SSB is not defined separately, and SSB is configured like a CSI-RS resource.

[0211]

[0212] In Table 1, the csi-SSB-ResourceSetList parameter represents a list of SSB resources used for beam management and reporting in a single CSI-RS resource set. Here, the SSB resource set can be set to {SSBx1, SSBx2, SSBx3, SSBx4, …}. For example, the SSB index can be defined from 0 to 63.

[0213] - The terminal receives a DownLink Reference Signal (DL RS) from the base station. As a specific example, the terminal receives an SSB from the base station based on the CSI-SSB-ResourceSetList.

[0214] - The terminal transmits a beam report to the base station. As a specific example, if a CSI-ReportConfig related to reporting on SSBRI (SSB Resource Indicator) and L1-RSRP is configured, the terminal reports the best SSBRI and the corresponding L1-RSRP to the base station.

[0215] That is, if the reportQuantity of the above CSI-ReportConfig IE is set to 'ssb-Index-RSRP', the terminal reports the best SSBRI and the corresponding L1-RSRP to the base station.

[0216] And, if the terminal has a CSI-RS resource configured in the same OFDM symbol(s) as the SSB (SS / PBCH Block) and 'QCL-TypeD' is applicable, the terminal can assume that the CSI-RS and SSB are quasi-co-located in terms of 'QCL-TypeD'.

[0217] Here, the above QCL Type D may mean that antenna ports are QCL-connected in terms of spatial Rx parameters. When a terminal receives multiple DL antenna ports that are in a QCL Type D relationship, it is acceptable to apply the same receiving beam. Additionally, the terminal does not expect CSI-RS to be configured in an RE that overlaps with the RE of the SSB.

[0218] The configuration for beam reporting using CSI-RS is performed in the same manner as the configuration for beam reporting using SSB described above, so a redundant explanation is omitted. The operation of the beam reporting procedure using CSI is described below.

[0219] - The terminal receives configuration information from the base station. As a specific example, the terminal receives from the base station a CSI-ResourceConfig IE containing a CSI-SSB-ResourceSetList containing CSI resources used for BM (e.g., NZP CSI-RS resource set IE).

[0220] - The terminal receives CSI-RS resources within the NZP CSI-RS resource set through different Tx beams (DL spatial domain transmission filters) of the base station.

[0221] - The terminal selects (or determines) the best beam.

[0222] - The terminal reports the ID and associated quality information (e.g., L1-RSRP) for the selected beam to the base station. In this case, the reportQuantity of the CSI report config can be set to 'cri-RSRP'.

[0223] The symbols / abbreviations / terms used in this specification are as follows.

[0224] BM: beam management

[0225] CQI: channel quality indicator

[0226] CRI: CSI-RS (channel state information - reference signal) resource indicator

[0227] CSI: channel state information

[0228] CSI-IM: channel state information - interference measurement

[0229] CSI-RS: channel state information - reference signal

[0230] DMRS: demodulation reference signal

[0231] FDM: frequency division multiplexing

[0232] FFT: fast Fourier transform

[0233] IFDMA: interleaved frequency division multiple access

[0234] IFFT: inverse fast Fourier transform

[0235] L1-RSRP: Layer 1 reference signal received power

[0236] L1-RSRQ: Layer 1 reference signal received quality

[0237] MAC: medium access control

[0238] NZP: non-zero power

[0239] OFDM: orthogonal frequency division multiplexing

[0240] PDCCH: physical downlink control channel

[0241] PDSCH: physical downlink shared channel

[0242] PMI: precoding matrix indicator

[0243] RE: resource element

[0244] RI: Rank indicator

[0245] RRC: radio resource control

[0246] RSSI: received signal strength indicator

[0247] Rx: Reception

[0248] QCL: quasi co-location

[0249] SINR: signal to interference and noise ratio

[0250] SSB (or SS / PBCH block): synchronization signal block (including primary synchronization signal, secondary synchronization signal and physical broadcast channel)

[0251] TDM: time division multiplexing

[0252] TRP: transmission and reception point

[0253] TRS: tracking reference signal

[0254] Tx: transmission

[0255] UE: user equipment

[0256] ZP: zero power

[0257] Due to advancements in computational processing technology and AI / ML technologies, the nodes and terminals constituting wireless communication networks are becoming more intelligent and sophisticated. In particular, as a result of this network intelligence, it is expected that various network decision parameter values ​​(e.g., transmit / receive power of each base station, transmit power of each terminal, precoder / beam of base stations and terminals, time / frequency resource allocation for each terminal, duplex method of each base station, etc.) can be rapidly optimized, derived, and applied based on diverse network environment parameters (e.g., distribution / location of base stations, distribution / location / material of buildings / furniture, location / movement direction / speed of terminals, weather information, etc.).

[0258] This specification proposes a method related to Life Cycle Management (LCM) for stably maintaining the performance of the entire communication system and improving the generalization performance of the AI / ML model in a changing channel environment when wireless communication services are supported based on AI / ML.

[0259] In this document, ' / ' means 'and', 'or', or 'and / or' depending on the context.

[0260] The functional framework of AI / ML includes data collection, model training, model storage, management, and inference functions. The management function determines the model's operations (e.g., selection / (de)activation / switching / fallback) and monitors performance. Monitoring methods can be classified into UE-side monitoring, NW-side monitoring, and hybrid monitoring depending on the entity calculating the monitoring metrics. It is important for the management function to make decisions that ensure proper inference operations based on the data received from the data collection and inference functions. The description of the management function (30) in the AI / ML functional framework of FIG. 1 described above is as follows.

[0261] The Management function (30) is a function that monitors the operation of an AI / ML model or an AI / ML function. Additionally, the Management function (30) may make decisions to ensure appropriate inference operations based on data received from the Data Collection function (10) (i.e., Monitoring Data (12)) and / or data received from the Inference function (40) (i.e., Inference Output (41)).

[0262] Management Instruction (32) is information required as input to manage the Inference function (40). The relevant information may include the selection / (de)activation / switching of an AI / ML model or an AI / ML-based function, and may also include a fallback to a non-AI / ML operation (i.e., not relying on the inference process).

[0263] A Model Transfer / Delivery Request (33) can be used to request model(s) from Model Storage (50).

[0264] Performance Feedback / Retraining Request (31) refers to information required as input to Model Training function (20) (e.g., for the purpose of retraining or updating the model).

[0265] UE-side monitoring is a method in which a UE directly monitors the performance of an AI / ML model. The UE analyzes the model's input data and output results in real time to calculate monitoring metrics, determines monitoring outputs based on these metrics, and reports them to the NW. This approach enables the UE to independently evaluate the model's performance and respond quickly.

[0266] NW-side monitoring is a method in which the network directly calculates monitoring metrics based on data collected from terminals, determines monitoring outputs based on these metrics, and issues instructions to the terminals. Specifically, the network evaluates the performance of a model based on data collected from terminals and updates the terminal's model or performs functional changes as necessary. This NW-side monitoring method enables the maintenance of consistent performance and efficient resource management of the entire system.

[0267] Hybrid monitoring is a method in which the terminal and the network cooperate to monitor and optimize the performance of AI / ML models. The terminal performs initial monitoring based on local data, calculates monitoring metrics, and transmits them to the network. Subsequently, the network determines monitoring output values ​​based on the monitoring metrics received from the terminal and instructs the terminal.

[0268] Each monitoring method is performed based on the inputs and / or outputs of AI / ML models, enabling the evaluation and maintenance of model performance. In particular, it plays a role in detecting discrepancies between the data trained on the AI / ML models and the actual operating environment, and responding in real-time to prevent model performance degradation.

[0269] The key performance indicators (KPIs) to consider when performing model monitoring are as follows.

[0270] Accuracy: A factor that evaluates how well the prediction and inference results provided by AI / ML models match reality.

[0271] Overhead: Additional computational costs and network resource consumption incurred during the model monitoring process

[0272] Complexity: A factor evaluating how complex the algorithms and processes performing model monitoring are.

[0273] Latency: The time difference between the model's input and the monitoring result, a factor that ensures real-time responsiveness.

[0274] In AI / ML-based NR systems, it is essential to continuously monitor and manage AI / ML models to maintain optimal performance. Generally, monitoring techniques based on the input and output data distributions of AI / ML models are used to evaluate their performance. Additionally, monitoring AI / ML models utilizing anomaly detection or performance prediction technologies are also employed. These monitoring models operate by predicting the performance of operational AI / ML models or calculating the similarity between the data distribution during the training phase and the actual data distribution encountered. This enables the early detection of performance degradation in operational AI / ML models and supports appropriate responses.

[0275] Existing monitoring techniques based on the input / output data of AI / ML models may have limitations in that real-time monitoring and immediate action are difficult because analysis and evaluation are only possible after a certain amount of data has been collected. Furthermore, in the case of performance prediction models, if the channels experienced by the AI / ML model in a real-time changing network environment differ from the trained channel data, the accuracy of the performance prediction model can deteriorate rapidly. This makes the application of performance prediction models difficult and can be a factor in increasing management overhead due to prediction errors. Monitoring techniques utilizing 2-Level Anomaly Detection operate by classifying AI / ML models into only two states—Normal or Abnormal—making it difficult to determine appropriate response methods when an abnormal state is detected. Consequently, existing monitoring techniques may cause inappropriate model updates, switching, or (De)activation, leading to unnecessary resource waste and increased management overhead.

[0276] This specification proposes a new monitoring technique that solves the problems faced by existing monitoring techniques, such as real-time monitoring, applicability of AI / ML models in environments different from the training channel environment, and determination of appropriate model management operations.

[0277] Specifically, this specification proposes a 3-level monitoring technique that detects real-time channel changes by adding an intermediate point called a "novel" (new state) between the normal and abnormal states based on the input / output data of an AI / ML model. According to this method, channel changes can be detected and the model managed in real time without the need to collect a specific number of data samples, and the model can be operated stably even in channel environments different from the training data. Furthermore, because the model is managed using three states—normal, novel, and abnormal—when the AI / ML model's channel environment differs from the training channel environment during inference, appropriate model management procedures are performed to reduce unnecessary model switching and fallback, thereby minimizing model management overhead. The method described above will be explained in more detail below.

[0278] Proposal 1

[0279] To monitor AI / ML model performance, methods utilizing monitoring-related configuration information and monitoring outputs may be considered. For example, the monitoring output may represent one of at least three states.

[0280] In the following, states related to an AI / ML model or AI / ML functionality are described by classifying / defining them into three categories (three states). However, this is merely for the convenience of explanation and is not intended to limit the technical concept according to the embodiments of this specification to three categories (three states). For example, states related to an AI / ML model or AI / ML functionality may be classified / defined into multiple states (more subdivided than the three states mentioned above). As a specific example, at least one of the first category (first state) to the third category (third state) may be subdivided into multiple categories (multiple states) (e.g., first-1 state, first-2 state... / second-1 state, second-2 state... / third-1 state, third-2 state...).

[0281] In the above Proposal 1, the monitoring output value may be configured to represent one of three states, Category 1, Category 2, and Category 3, according to criteria pre-set by the base station. The classification of the three states may be based on a monitoring metric value (e.g., a performance metric). Each category indicates how similar the input / output data of the AI / ML model during inference is to the input / output data of the AI / ML model during training.

[0282] Category 1 (State 1) relates to a situation where the input / output data of an AI / ML model exhibits high similarity to the input / output data of the AI / ML model during training. Specifically, Category 1 (State 1) signifies that the environment during inference of the AI / ML model is operating under conditions similar to the training environment. In Category 1 (State 1), no performance degradation of the AI / ML model can be expected.

[0283] Category 2 (Second State) relates to a situation where the input / output data of an AI / ML model exhibits a moderate level of similarity to the input / output data during training. Specifically, Category 2 (Second State) signifies that the AI / ML model is operating in an environment that, while showing a certain level of similarity to the training environment during inference, still differs to some extent. In Category 2 (Second State), a certain level of performance degradation of the AI / ML model can be expected.

[0284] Category 3 (Third State) relates to a situation where the input / output data of an AI / ML model exhibits a low level of similarity to the input / output data during training. Specifically, Category 3 (Third State) signifies that the AI / ML model is operating in an environment different from the training environment during inference. Significant performance degradation of the AI / ML model can be expected in Category 3 (Third State).

[0285] In the above Proposal 1, a scenario may be considered in which a separate AI / ML model for monitoring is utilized in addition to monitoring-related configuration information to monitor the performance of the AI / ML model. In this case, whether to use a separate AI / ML model for monitoring may be determined based on the capability of the terminal or NW. The monitoring model may be trained based on the training dataset used to train the AI / ML model. The monitoring model calculates and outputs a similarity value indicating how similar the input / output data of the AI / ML model during inference is to the input / output data of the AI / ML model during training. For example, the similarity value or information related to the similarity (e.g., monitoring metric or performance metric) may include a numerical value or indicator indicating how much the input data (e.g., input / output data of the AI / ML model during inference) deviates from a normal pattern (e.g., input / output data of the AI / ML model during training). As a specific example, the information related to the similarity (e.g., monitoring metric or performance metric) may include a value indicating the degree of abnormality (e.g., anomaly score), SGCS (e.g., Squared Generalized Cosine Similarity or Structural Graphical Correlation Score) or NMSE (Normalized Mean Squared Error).

[0286] Based on the corresponding similarity value, the monitoring output can be determined as one of three states for the current state of the AI / ML model. The aforementioned proposal is not limited to cases where an AI / ML model is utilized for monitoring. For example, (in cases where the use of an AI / ML model for monitoring is restricted or not supported), the similarity between the input / output data of the AI / ML model during inference and the input / output data of the AI / ML model during training can be calculated by utilizing conventional techniques for the AI / ML model during inference. In other words, the aforementioned similarity can be calculated even for terminals / base stations (NW) where a dedicated monitoring model is not implemented, or for terminals / base stations (NW) where a dedicated monitoring model is not supported.

[0287] For example, in a two-sided model scenario, a terminal that has received monitoring settings / instructions from a base station may transmit additional information for monitoring (e.g., use case / functionality ID, model ID, input / output statistics of the inference model, value range, etc.) to the NW side for the monitoring.

[0288] Proposal 1-1

[0289] The above-described states can be classified / expressed as i) a state requiring a change to the model / functionality (e.g., the third state), ii) a state requiring no change to the model / functionality (e.g., the first state), and iii) a state requiring adjustment to the model / functionality (e.g., the second state).

[0290] In Proposal 1 above, the monitoring output can be determined as at least one of the three states. The AI / ML model management procedure for each state / category follows the following procedure.

[0291] In Category 1 (High Similarity), the input / output data of the AI / ML model exhibits high similarity to the input / output data of the AI / ML model during training. This implies that the model is operating under conditions similar to the environment in which it was trained. Since the likelihood of model performance degradation is low in Category 1, it is appropriate to maintain the existing model without additional measures (e.g., changes to the model / functionality). Furthermore, because there is a high probability that the performance of the AI / ML model will remain stable in Category 1, the terminal can request a change in the monitoring cycle from the base station (NW). This reduces unnecessary computations and optimizes resource usage for both the terminal (UE) and the NW. During the process of adjusting the monitoring cycle, the terminal can transmit a request for a monitoring cycle update to the NW along with the monitoring metric value. Subsequently, the NW can determine the monitoring cycle based on the information received from the terminal and issue instructions to the terminal.

[0292] In Category 2 (Moderate Similarity), the input and output data of the AI / ML model show a certain level of difference from the input and output data during previous training, so there is a possibility that the AI / ML model's performance will gradually degrade. Therefore, adjustments to the model and functionality are necessary in Category 2. In this case, a method to improve the performance of the AI / ML model by updating the existing model(s) can be applied.

[0293] For example, updates can be performed by improving performance while maintaining existing AI / ML model(s). Instead of replacing existing model(s) with new ones, the following actions may be performed or applied. Specifically, the parameters of the model(s) may be fine-tuned while maintaining the current model(s). Another specific example is that a short-term model update may be performed using new channel data currently experienced by the terminal while maintaining the current model(s). In this case, the entity performing the model update may be the terminal (UE) or the network.

[0294] When a model update occurs, it may be necessary to modify the monitoring model or related configuration information that calculates the similarity between the input / output data of the inference AI / ML model and the input / output data of the AI / ML model during training. A signaling procedure may be performed for this purpose. Specifically, the network (or terminal) may configure or instruct the terminal (or network) to additionally update monitoring-related configuration information or the monitoring model regarding the new data.

[0295] In Category 3 (Low Similarity), there is a high probability that the performance of the AI / ML model will degrade rapidly because the input / output data has characteristics that are completely different from those of the data used during training. In Category 3 (Low Similarity), it may be difficult to sufficiently recover performance through short-term model updates alone, so changes to the model and functionality are necessary.

[0296] The third category can be divided into the following three cases.

[0297] First, if the UE possesses appropriate AI / ML model(s) or functionality(s), the UE immediately performs AI / ML model(s) or functionality(s) switching and reports the Switching ID to the NW. At this time, the UE can perform the switching after utilizing Model Buffering information to verify whether there is an appropriate model(s) or functionality(s) among the stored models or functionality(s).

[0298] Second, if the Network has appropriate AI / ML model(s) or functionality(s), the UE sends a switching request for the model(s) or functionality(s) to the Network. The switching is performed by the Network distributing the new model(s) or functionality(s) to the UE.

[0299] Third, if neither the NW nor the UE possesses appropriate AI / ML model(s) or functionality(s), the NW instructs the terminal to perform a fallback operation. During the execution of the fallback operation, a time window can be secured to train or update a new AI / ML model. Along with the instruction for the fallback operation, the NW may transmit a request for dataset collection and configuration information for this purpose to the terminal, and the terminal may continuously collect fine-tuning datasets for training the AI / ML model. Once the terminal has completed collecting the amount of data set by the NW, it may transmit the corresponding fine-tuning dataset to the NW. Based on this, the NW can perform a long-term model update. In this specification, a model update in this third category is referred to as a long-term model update. For example, a long-term model update may be performed by the UE or the NW. Specifically, a long-term model update may require high computing resources because the dataset is generally large and the amount of training computation is substantial. Considering this, long-term model updates may be performed in a network with sufficient training computing resources.

[0300] When AI / ML model / functionality switching is performed, existing monitoring-related configuration information or monitoring models are also changed. To this end, the network (or terminal) can configure or instruct the terminal (or network) to set monitoring-related configuration information or a monitoring model that matches the changed AI / ML model / functionality.

[0301] Proposal 1-2

[0302] The monitoring-related configuration information in Proposal 1 above may include information related to multiple thresholds (e.g., two or more thresholds). The multiple thresholds may be used to distinguish three states.

[0303] For example, information related to two or more thresholds may be included in the monitoring configuration information to distinguish the three possible states of the monitoring output value. These thresholds can be utilized to distinguish the three states. Specifically, the first threshold can be used to distinguish between the first category (first state) and the second and third categories (second / third states). The third threshold can be used to distinguish between the first / second categories (first / second states) and the third category (third state). The above monitoring configuration information can be set or instructed by the network to the terminal. If the monitoring model is not utilized, the network may additionally set or instruct the terminal to the values ​​of the input / output data or the distribution ID during the training process. Based on this, it can be used to distinguish which category the input / output data of the AI / ML model belongs to during inference.

[0304] Proposal 1-3

[0305] The determination of the above-described state (e.g., one of the first to third states) can be determined based on the ratio / range of the data.

[0306] To prevent erroneous monitoring results, whether data belongs to a specific category (specific state) can be determined by considering the distribution characteristics of the data (e.g., the aforementioned similarity value, monitoring metric, or performance metric). Data may be classified into the corresponding category if its ratio or range exceeds a certain ratio or range. Specifically, the terminal can determine that data belongs to a specific category (specific state) only if the ratio or range of data corresponding to that category exceeds a specific threshold during a pre-set time window. The network can set or instruct the terminal to provide monitoring-related configuration information, including information such as the time window and threshold.

[0307] Proposal 1-4

[0308] The terminal may perform reporting to the NW only when it belongs to some of the states described above. For example, the terminal may transmit a report containing information indicating the state to the base station based on the fact that the state related to the AI / ML model or AI / ML functionality is a second state. For example, the terminal may transmit a report containing information indicating the state to the base station based on the fact that the state related to the AI / ML model or AI / ML functionality is a second state or a third state.

[0309] According to one embodiment, model management based on a plurality of states that are more finely subdivided than the three states described above may be performed. For example, threshold information for distinguishing the plurality of states (e.g., a first threshold, a second threshold, a third threshold, etc.) may be included in the monitoring-related configuration information.

[0310] The operation procedure of the base station and the terminal based on the aforementioned Proposal 1 is described in detail below.

[0311] Terminal operation:

[0312] Step 1: The terminal reports capability information regarding monitoring functions that the terminal can support to the base station, such as the type of monitoring model the terminal can support and whether the terminal has a model buffering function.

[0313] Step 2: The terminal receives monitoring model and / or monitoring-related configuration information from the base station.

[0314] Step 3: The terminal calculates a monitoring metric based on the input data of the UE-side model.

[0315] Case A: Case where the terminal determines the monitoring output and subsequent action

[0316] Step 3A-1: The terminal reports monitoring metric and / or monitoring output to the base station.

[0317] Case B: Where the base station determines the monitoring output and subsequent action

[0318] Step 3B-1: The terminal reports the monitoring metric to the base station.

[0319] Step 3B-2: The terminal receives information about the monitoring output and subsequent action from the base station.

[0320] Step 4: The terminal performs the above subsequent action and reports the result of the execution to the base station.

[0321] Base station operation:

[0322] Step 1: The base station receives capability information from the terminal regarding monitoring functions that the terminal can support, such as the type of monitoring model that the terminal can support and whether the terminal has a model buffering function.

[0323] Step 2: The base station transmits the monitoring model and / or monitoring-related configuration information to the terminal.

[0324] Step 3: The base station receives a monitoring metric from the terminal.

[0325] Case A: Case where the terminal determines the monitoring output and subsequent action

[0326] Step 3A-1: The base station receives a monitoring metric and / or monitoring output from the terminal.

[0327] Case B: Where the base station determines the monitoring output and subsequent action

[0328] Step 3B-1: The base station receives a monitoring metric from the terminal.

[0329] Step 3B-2: The base station determines the monitoring output and subsequent action and transmits information about the monitoring output and subsequent action to the terminal.

[0330] Step 4: The base station receives the result of performing a subsequent action from the terminal.

[0331] Herein, the operations described in this specification may be described separately for convenience, but unless specifically stated otherwise, each operation may be combined with others.

[0332] In terms of implementation, operations of a base station / terminal according to the embodiments described above (e.g., operations based on at least one of proposal 1, proposal 1-1, proposal 1-2, proposal 1-3 and / or proposal 1-4) can be processed by the device of FIG. 9 (e.g., the processor (110, 210) of FIG. 9).

[0333] In addition, the operations of the base station / terminal according to the above-described embodiment (e.g., operations based on at least one of proposal 1, proposal 1-1, proposal 1-2, proposal 1-3 and / or proposal 1-4) may be stored in memory (e.g., 140, 240 of FIG. 9) in the form of instructions / programs (e.g., instruction, executable code) for driving at least one processor (e.g., 110, 210 of FIG. 9).

[0334] In a first aspect, a method is provided that includes the step of performing an operation described in the present specification, in a method performed by a terminal (or a first node).

[0335] In a second aspect, a terminal (or first node) of a wireless communication system is provided, comprising: at least one transceiver; at least one processor configured to perform the operation described herein; and at least one computer memory operably connected to the at least one processor.

[0336] In a third aspect, an apparatus is provided comprising at least one processor; and at least one computer memory operably connected to the at least one processor and storing instructions that cause a terminal (or first node) to perform the operation described herein based on execution by the at least one processor.

[0337] In a fourth aspect, a non-transitory computer-readable storage medium is provided that stores instructions for a terminal (or first node) to perform the operation described herein, based on execution by at least one processor.

[0338] In a fifth aspect, a method is provided that includes the step of performing the operation described in the present specification, in a method performed by a base station (or a second node).

[0339] In a sixth aspect, a base station (or second node) of a wireless communication system is provided, comprising: at least one transceiver; at least one processor configured to perform the operation described herein; and at least one computer memory operably connected to the at least one processor.

[0340] In a seventh aspect, an apparatus is provided comprising at least one processor; and at least one computer memory operably connected to the at least one processor and storing instructions that cause a base station (or a second node) to perform the operation described herein based on execution by the at least one processor.

[0341] In an eighth aspect, a non-transient computer-readable storage medium is provided that stores instructions for a base station (or a second node) to perform the operation described herein, based on execution by at least one processor.

[0342] The embodiments described above will be explained in detail below with reference to FIGS. 7 and FIGS. 8 regarding the operation of the terminal and base station. The methods described below are distinguished only for convenience of explanation, and it is understood that a part of one method may be substituted with a part of another method or combined with one another and applied.

[0343] FIG. 7 is a flowchart illustrating a method according to one embodiment of the present specification.

[0344] Referring to FIG. 7, a method according to one embodiment of the present specification may include a step of receiving setting information related to performance monitoring (S710) and a step of transmitting a report related to performance monitoring (S720).

[0345] In S710, the terminal receives configuration information related to performance monitoring from the base station. For example, the configuration information may include information based on at least one of the above-described proposals 1, 1-1, 1-3, and 1-4.

[0346] In S720, the terminal transmits a report related to the above performance monitoring to the base station.

[0347] According to one embodiment, the report may include information indicating a state among a plurality of states related to an AI / ML function (Artificial Intelligence / Machine Learning functionality) or an AI / ML model. This embodiment may be based on at least one of the above-described proposals 1, 1-1, 1-3, and 1-4.

[0348] According to one embodiment, the plurality of states may include i) a first state in which no change is required for the AI / ML function or the AI / ML model, ii) a second state in which adjustment is required for the AI / ML function or the AI / ML model, and iii) a third state in which a change is required for the AI / ML function or the AI / ML model. This embodiment may be based on Proposal 1 and Proposal 1-1.

[0349] According to one embodiment, the setting information may include a first threshold and a second threshold. The first threshold may be related to the distinction between i) the first state and ii) the second state and the third state. The second threshold may be related to the distinction between i) the first state and the second state and ii) the third state. This embodiment may be based on Proposal 1-2.

[0350] According to one embodiment, the ratio of a performance metric associated with the state during a set time window may be greater than or equal to a specific threshold. This embodiment may be based on Proposal 1-3.

[0351] For example, the above performance metric may be based on the similarity between the distribution related to the first data and the distribution related to the second data.

[0352] For example, the first data may be input data related to the training of the AI / ML model. The second data may be input data related to the inference of the AI / ML model.

[0353] For example, the first data may be output data related to the training of the AI / ML model. The second data may be output data related to the inference of the AI / ML model.

[0354] According to one embodiment, the report may be transmitted based on the fact that the state is the second state or the third state. This embodiment may be based on Proposal 1-4.

[0355] According to one embodiment, based on the fact that the state associated with the AI / ML function or the AI / ML model is the first state, the report may include a request related to a change in the monitoring cycle. This embodiment may be based on Proposal 1-1.

[0356] According to one embodiment, the method may further include the step of receiving information indicating an action related to the state associated with the AI / ML function or the AI / ML model. Specifically, the terminal may receive information from a base station indicating an action (e.g., a subsequent action) associated with the state associated with the AI / ML function or the AI / ML model. This embodiment may be based on Proposal 1-1.

[0357] For example, based on the fact that the above state is the second state, the information may include information related to a short-term update of the AI / ML function or the AI / ML model.

[0358] For example, based on the fact that the above state is the third state, the information may include i) an ID indicating an AI / ML function or AI / ML model to be switched, ii) information for deploying a new AI / ML function or a new AI / ML model, or iii) information indicating a fallback operation.

[0359] As a specific example, it may be assumed that the terminal possesses / is equipped with AI / ML model(s) or functionality(s) for model / function switching. The information may include an ID representing the AI / ML function or AI / ML model to be switched.

[0360] As a specific example, it may be assumed that a base station (NW) possesses / is equipped with AI / ML model(s) or functionality(s) for model / function switching. The information may include information for deploying new AI / ML functions or new AI / ML models.

[0361] As a specific example, it may be assumed that there are no AI / ML model(s) or functionality(s) for model / function switching at the terminal and base station. The above information may include information indicating a fallback operation.

[0362] According to one embodiment, based on the fact that the state associated with the AI / ML function or the AI / ML model is the third state, the report may include a request for a new AI / ML function or a new AI / ML model. This embodiment may be based on Proposal 1-1.

[0363] The operation based on the steps of receiving information indicating an operation related to the state, as described above (S710 to S720), can be implemented by the device of FIG. 9. For example, referring to FIG. 9, the terminal (200) can control one or more transceivers (230) and / or one or more memories (240) to perform the operation based on the steps of receiving information indicating an operation related to the state, as described above (S710 to S720).

[0364] The embodiments described above will be explained in detail below in terms of base station operation.

[0365] The steps of transmitting information indicating an operation related to the state, S810 to S820 described below, correspond to the steps of receiving information indicating an operation related to the state, S710 to S720 described in FIG. 7. Considering the correspondence, redundant descriptions are omitted. That is, the specific description of the base station operation described below can be replaced by the description / embodiment of FIG. 7 corresponding to the operation.

[0366] FIG. 8 is a flowchart illustrating a method according to another embodiment of the present specification.

[0367] Referring to FIG. 8, a method according to another embodiment of the present specification may include a step of transmitting setting information related to performance monitoring (S810) and a step of receiving a report related to performance monitoring (S820).

[0368] In S810, the base station transmits configuration information related to performance monitoring to the terminal.

[0369] In S820, the base station receives a report related to the performance monitoring from the terminal.

[0370] According to one embodiment, the report may include information indicating a state among a plurality of states related to an AI / ML function (Artificial Intelligence / Machine Learning functionality) or an AI / ML model.

[0371] According to one embodiment, the method may further include the step of transmitting information indicating an action related to the state associated with the AI / ML function or the AI / ML model. Specifically, the base station may transmit information indicating an action (e.g., a subsequent action) associated with the state associated with the AI / ML function or the AI / ML model to the terminal. This embodiment may be based on Proposal 1-1.

[0372] The operation based on the steps of transmitting information indicating an operation related to the state, as described above, S810 to S820, can be implemented by the device of FIG. 9. For example, referring to FIG. 9, a base station (100) can control one or more transceivers (130) and / or one or more memories (140) to perform the operation based on the steps of transmitting information indicating an operation related to the state, as described above, S810 to S820.

[0373] Hereinafter, an apparatus to which the embodiments of the present specification can be applied (an apparatus implementing the method / operation according to the embodiments of the present specification) will be described with reference to FIG. 9.

[0374] FIG. 9 is a drawing showing the configuration of a first device and a second device according to an embodiment of the present specification.

[0375] The first device (100) may include a processor (110), an antenna unit (120), a transceiver (130), and a memory (140).

[0376] The processor (110) performs baseband-related signal processing and may include an upper layer processing unit (111) and a physical layer processing unit (115). The upper layer processing unit (111) may process operations of the MAC layer, RRC layer, or higher upper layers. The physical layer processing unit (115) may process operations of the PHY layer. For example, if the first device (100) is a base station device in base station-terminal communication, the physical layer processing unit (115) may perform uplink reception signal processing, downlink transmission signal processing, etc. For example, if the first device (100) is a first terminal device in terminal-terminal communication, the physical layer processing unit (115) may perform downlink reception signal processing, uplink transmission signal processing, sidelink transmission signal processing, etc. In addition to performing baseband-related signal processing, the processor (110) may also control the overall operation of the first device (100).

[0377] The antenna section (120) may include one or more physical antennas, and if it includes multiple antennas, it may support MIMO transmission and reception. The transceiver (130) may include an RF (Radio Frequency) transmitter and an RF receiver. The memory (140) may store information processed by the processor (110) and software, operating systems, applications, etc. related to the operation of the first device (100), and may include components such as a buffer.

[0378] The processor (110) of the first device (100) may be configured to implement the operation of the base station in base station-terminal communication (or the operation of the first terminal device in terminal-terminal communication) in the embodiments described in this disclosure.

[0379] The second device (200) may include a processor (210), an antenna unit (220), a transceiver (230), and a memory (240).

[0380] The processor (210) performs baseband-related signal processing and may include an upper layer processing unit (211) and a physical layer processing unit (215). The upper layer processing unit (211) may process operations of the MAC layer, RRC layer, or higher upper layers. The physical layer processing unit (215) may process operations of the PHY layer. For example, if the second device (200) is a terminal device in base station-terminal communication, the physical layer processing unit (215) may perform downlink reception signal processing, uplink transmission signal processing, etc. For example, if the second device (200) is a second terminal device in terminal-terminal communication, the physical layer processing unit (215) may perform downlink reception signal processing, uplink transmission signal processing, sidelink reception signal processing, etc. In addition to performing baseband-related signal processing, the processor (210) may also control the overall operation of the second device (210).

[0381] The antenna section (220) may include one or more physical antennas, and may support MIMO transmission and reception if it includes multiple antennas. The transceiver (230) may include an RF transmitter and an RF receiver. The memory (240) may store information processed by the processor (210) and software, operating systems, applications, etc. related to the operation of the second device (200), and may include components such as a buffer.

[0382] The processor (210) of the second device (200) may be configured to implement the operation of the terminal in base station-terminal communication (or the operation of the second terminal device in terminal-terminal communication) in the embodiments described in this disclosure.

[0383] In the operation of the first device (100) and the second device (200), the details described in the examples of the present disclosure regarding the base station and terminal (or the first terminal and the second terminal in terminal-to-terminal communication) in base station-to-terminal communication may be applied in the same way, and redundant descriptions are omitted.

[0384] Here, the wireless communication technology implemented in the device of the present disclosure may include LTE, NR, and 6G, as well as Narrowband Internet of Things (NB-IoT) for low-power communication. For example, NB-IoT technology may be an example of Low Power Wide Area Network (LPWAN) technology and may be implemented according to standards such as LTE Cat NB1 and / or LTE Cat NB2, but is not limited to the names mentioned above.

[0385] Additionally or alternatively, the wireless communication technology implemented in the device of the present disclosure may perform communication based on LTE-M technology. For example, LTE-M technology may be an example of LPWAN technology and may be referred to by various names such as eMTC (enhanced Machine Type Communication). For example, LTE-M technology may be implemented in at least one of various standards such as 1) LTE CAT 0, 2) LTE Cat M1, 3) LTE Cat M2, 4) LTE non-BL (non-Bandwidth Limited), 5) LTE-MTC, 6) LTE Machine Type Communication, and / or 7) LTE M, and is not limited to the names mentioned above.

[0386] Additionally or generally, the wireless communication technology implemented in the device of the present disclosure may include at least one of ZigBee, Bluetooth, and a Low Power Wide Area Network (LPWAN) for low-power communication, but is not limited to the names mentioned above. For example, ZigBee technology can create personal area networks (PANs) related to small / low-power digital communication based on various standards such as IEEE 802.15.4 and may be referred to by various names.

Claims

1. Regarding the method, A step of receiving configuration information related to performance monitoring from a base station by a terminal; and The method includes the step of transmitting a report related to performance monitoring to the base station by the terminal, A method characterized in that the above report includes information indicating a state among a plurality of states related to an AI / ML function (Artificial Intelligence / Machine Learning functionality) or an AI / ML model.

2. In Paragraph 1, A method characterized in that the plurality of states include i) a first state in which no change is required for the AI / ML function or the AI / ML model, ii) a second state in which adjustment is required for the AI / ML function or the AI / ML model, and iii) a third state in which a change is required for the AI / ML function or the AI / ML model.

3. In Paragraph 1, The above setting information includes a first threshold and a second threshold, and The above first threshold value is related to i) the first state and ii) the distinction between the second state and the third state, and A method characterized in that the second threshold is related to i) the first state and the second state and ii) the third state.

4. In Paragraph 1, A method characterized by the ratio of a performance metric associated with the state during a set time window being greater than or equal to a specific threshold.

5. In Paragraph 4, A method characterized in that the above performance metric is based on the similarity between the distribution related to the first data and the distribution related to the second data.

6. In Paragraph 4, A method characterized in that the first data is input data related to the training of the AI / ML model, and the second data is input data related to the inference of the AI / ML model.

7. In Paragraph 4, A method characterized in that the first data is output data related to the training of the AI / ML model, and the second data is output data related to the inference of the AI / ML model.

8. In Paragraph 1, A method characterized by the above report being transmitted based on whether the above state is the second state or the third state.

9. In Paragraph 2, A method characterized in that, based on the fact that the state related to the AI / ML function or the AI / ML model is the first state, the report includes a request related to a change in the monitoring cycle.

10. In Paragraph 2, A method further comprising the step of receiving information indicating an operation related to the state associated with the AI / ML function or the AI / ML model.

11. In Paragraph 10, A method characterized in that, based on the fact that the above state is the second state, the information includes information related to a short-term update of the AI / ML function or the AI / ML model.

12. In Paragraph 10, A method characterized by, based on the fact that the above state is the third state, the information including i) an ID representing an AI / ML function or AI / ML model to be switched, ii) information for deploying a new AI / ML function or a new AI / ML model, or iii) information representing a fallback operation.

13. In Paragraph 2, A method characterized in that, based on the fact that the state associated with the AI / ML function or the AI / ML model is the third state, the report includes a request for a new AI / ML function or a new AI / ML model.

14. In the terminal, One or more transmitters / receivers; One or more processors; and It includes one or more memories connected to the above one or more processors and storing instructions, A terminal characterized by the above instructions enabling the terminal to perform all steps of the method according to any one of claims 1 to 13, based on execution by the one or more processors.

15. A device comprising one or more memories and one or more processors connected to the one or more memories, An apparatus characterized in that the above one or more memories store instructions that cause the apparatus to perform all steps of the method according to any one of claims 1 to 13, based on execution by the above one or more processors.

16. In a non-transitory computer-readable storage medium for storing instructions, A non-transitory computer-readable storage medium characterized by instructions executable by one or more processors such that the terminal performs all steps of the method according to any one of claims 1 to 13.

17. Regarding the method, A step of transmitting configuration information related to performance monitoring to a terminal by a base station; and The method includes the step of receiving a report related to performance monitoring from the terminal by the base station, A method characterized in that the above report includes information indicating a state among a plurality of states related to an AI / ML function (Artificial Intelligence / Machine Learning functionality) or an AI / ML model.

18. Regarding base stations, One or more transmitters / receivers; One or more processors; and It includes one or more memories connected to the above one or more processors and storing instructions, A base station characterized by the above instructions, based on execution by one or more processors, having the base station perform all steps of the method according to claim 17.

Images