Method and apparatus for artificial intelligence (AI) / machine learning (ML) operation in wireless networks

Abstract

A method performed by a User Equipment (UE) for performing an Artificial Intelligence (AI) / Machine Learning (ML) operation is provided. The method activates an AI / ML functionality. The method detects a failure in response to determining that at least one of the following conditions is satisfied: (i) the AI / ML functionality is not applicable; (ii) a beam failure or a radio link failure is detected based on an inference result of the AI / ML functionality; or (iii) a prediction accuracy of the AI / ML functionality falls outside of a predetermined acceptable range. The method transmits, to a Base Station (BS), a failure report associated with the detected failure.　

Description

METHOD AND APPARATUS FOR ARTIFICIAL INTELLIGENCE (AI) / MACHINE LEARNING (ML) OPERATION IN WIRELESS NETWORKS

[0001] The present disclosure is related to wireless communication and, more specifically, to a User Equipment (UE), Base Station (BS), and method for performing an Artificial Intelligence (AI) / Machine Learning (ML) operation in the wireless communication networks.

[0002] Various efforts have been made to improve different aspects of wireless communication for the cellular wireless communication systems, such as the 5thGeneration (5G) New Radio (NR), by improving data rate, latency, reliability, and mobility. The 5G NR system is designed to provide flexibility and configurability to optimize network services and types, accommodating various use cases, such as enhanced Mobile Broadband (eMBB), massive Machine-Type Communication (mMTC), and Ultra-Reliable and Low-Latency Communication (URLLC). As the demand for radio access continues to grow, however, there exists a need for further improvements in the next-generation wireless communication systems, such as improvements in AI / ML operations.

[0003] The present disclosure is related to a UE, a BS, and a method for performing an AI / ML operation in the wireless communication networks.

[0004] In a first aspect of the present disclosure, a UE for performing an AI / ML operation is provided. The UE includes at least one processor and at least one non-transitory computer-readable medium that is coupled to the at least one processor and that stores one or more computer-executable instructions. The computer-executable instructions, when executed by the at least one processor, cause the UE to: activate an AI / ML functionality; detect a failure in response to determining that at least one of the following conditions is satisfied: (i) the AI / ML functionality is not applicable; (ii) a beam failure or a radio link failure is detected based on an inference result of the AI / ML functionality; or (iii) a prediction accuracy of the AI / ML functionality falls outside of a predetermined acceptable range; and transmit, to a BS, a failure report associated with the detected failure.

[0005] In some implementations of the first aspect, the AI / ML functionality is determined to be not applicable in response to determining at least one of the following conditions is satisfied: (a) an additional condition (AC) associated with the AI / ML functionality has changed; or (b) a hardware shortage has been detected in the UE.

[0006] In some implementations of the first aspect, the prediction accuracy of the AI / ML functionality is determined to fall outside of the predetermined acceptable range in response to determining at least one of the following conditions is satisfied: (a) the prediction accuracy of the AI / ML functionality is lower than a first threshold; (b) a difference between a predicted Reference Signal Received Power (RSRP) of a specific beam and a measured RSRP of the specific beam is larger than a second threshold; or (c) a probability that predicted beams belong to measured top-K beams is lower than a third threshold. K is a positive integer greater than or equal to one.

[0007] In some implementations of the first aspect, the one or more computer-executable instructions, when executed by the at least one processor, further cause the UE to receive, from the BS, a set of Beam Failure Recovery (BFR) parameters specifically for the AI / ML functionality. The beam failure is determined to be detected based on the set of BFR parameters specifically for the AI / ML functionality.

[0008] In some implementations of the first aspect, the one or more computer-executable instructions, when executed by the at least one processor, further cause the UE to deactivate the AI / ML functionality in response to the detected failure.

[0009] In some implementations of the first aspect, the one or more computer-executable instructions, when executed by the at least one processor, further cause the UE to activate a second AI / ML functionality for handover prediction in response to detecting the beam failure based on the inference result of the AI / ML functionality.

[0010] In some implementations of the first aspect, the failure report includes a time stamp indicating when the failure occurred.

[0011] In a second aspect of the present application, a BS for performing an AI / ML operation is provided. The BS includes at least one processor and at least one non-transitory computer-readable medium that is coupled to the at least one processor and that stores one or more computer-executable instructions. The computer-executable instructions, when executed by the at least one processor, cause the BS to: receive, from a UE, a message indicating an AI / ML functionality supported by the UE; and receive, from the UE, a failure report associated with a failure detected by the UE. The UE activates the AI / ML functionality. The UE detects the failure in response to determining that at least one of the following conditions is satisfied: (i) the AI / ML functionality is not applicable; (ii) a beam failure or a radio link failure is detected based on an inference result of the AI / ML functionality; or (iii) a prediction accuracy of the AI / ML functionality falls outside of a predetermined acceptable range.

[0012] In some implementations of the second aspect, the UE determines that the AI / ML functionality is not applicable in response to determining at least one of the following conditions is satisfied: (a) an AC associated with the AI / ML functionality has changed; or (b) a hardware shortage has been detected in the UE.

[0013] In some implementations of the second aspect, the UE determines that the prediction accuracy of the AI / ML functionality falls outside of the predetermined acceptable range in response to determining at least one of the following conditions is satisfied: (a) the prediction accuracy of the AI / ML functionality is lower than a first threshold; (b) a difference between a predicted RSRP of a specific beam and a measured RSRP of the specific beam is larger than a second threshold; or (c) a probability that predicted beams belong to measured top-K beams is lower than a third threshold. K is a positive integer greater than or equal to one.

[0014] In some implementations of the second aspect, the one or more computer-executable instructions, when executed by the at least one processor, further cause the BS to transmit, to the UE, a set of BFR parameters specifically for the AI / ML functionality. The UE determines that the beam failure is detected based on the set of BFR parameters specifically for the AI / ML functionality.

[0015] In some implementations of the second aspect, the UE deactivates the AI / ML functionality in response to the detected failure.

[0016] In some implementations of the second aspect, the UE activates a second AI / ML functionality for handover prediction in response to detecting the beam failure based on the inference result of the AI / ML functionality.

[0017] In some implementations of the second aspect, the failure report includes a time stamp indicating when the failure occurred.

[0018] In a third aspect of the present application, a method performed by a UE for performing an AI / ML operation is provided. The method includes activating an AI / ML functionality; detecting a failure in response to determining that at least one of the following conditions is satisfied: (i) the AI / ML functionality is not applicable; (ii) a beam failure or a radio link failure is detected based on an inference result of the AI / ML functionality; or (iii) a prediction accuracy of the AI / ML functionality falls outside of a predetermined acceptable range; and transmitting, to a BS, a failure report associated with the detected failure.

[0019] Aspects of the present disclosure are best understood from the following detailed disclosure when read with the accompanying drawings. Various features are not drawn to scale. Dimensions of various features may be arbitrarily increased or reduced for clarity of discussion.

[0020] FIG. 1 is a block diagram illustrating a functional framework for AI / ML for NR air interface, according to an example implementation of the present disclosure.

[0021] FIG. 2 is a block diagram illustrating a functional framework for handling link or channel failures without AI / ML techniques, according to an example implementation of the present disclosure.

[0022] FIG. 3 is a flowchart illustrating a procedure for life cycle management (LCM) of the AI / ML functionality under failure handling, according to an example implementation of the present disclosure.

[0023] FIG. 4 is a flowchart illustrating a method / process performed by a UE for an AI / ML operation, according to an example implementation of the present disclosure.

[0024] FIG. 5 is a flowchart illustrating a method / process performed by a BS for an AI / ML operation, according to an example implementation of the present disclosure.

[0025] FIG. 6 is a block diagram illustrating a node for wireless communication, according to an example implementation of the present disclosure.

[0026] The following contains specific information related to implementations of the present disclosure. The drawings and their accompanying detailed disclosure are merely directed to implementations. However, the present disclosure is not limited to these implementations. Other variations and implementations of the present disclosure will be obvious to those skilled in the art.

[0027] Unless noted otherwise, like or corresponding elements among the drawings may be indicated by like or corresponding reference numerals. Moreover, the drawings and illustrations in the present disclosure are generally not to scale and are not intended to correspond to actual relative dimensions.

[0028] For the purposes of consistency and ease of understanding, like features may be identified (although, in some examples, not illustrated) by the same numerals in the drawings. However, the features in different implementations may be different in other respects and may not be narrowly confined to what is illustrated in the drawings.

[0029] References to “one implementation,” “an implementation,” “example implementation,” “various implementations,” “some implementations,” “implementations of the present application,” etc., may indicate that the implementation(s) of the present application so described may include a particular feature, structure, or characteristic, but not every possible implementation of the present application necessarily includes the particular feature, structure, or characteristic. Further, repeated use of the phrase “In some implementations,” or “in an example implementation,” “an implementation,” do not necessarily refer to the same implementation, although they may. Moreover, any use of phrases like “implementations” in connection with “the present application” are never meant to characterize that all implementations of the present application must include the particular feature, structure, or characteristic, and should instead be understood to mean “at least some implementations of the present application” includes the stated particular feature, structure, or characteristic. The term “coupled” is defined as connected, whether directly or indirectly through intervening components, and is not necessarily limited to physical connections. The term “comprising,” when utilized, means “including, but not necessarily limited to”; it specifically indicates open-ended inclusion or membership in the so-described combination, group, series, and the equivalent.

[0030] The expression “at least one of A, B and C” or “at least one of the following: A, B and C” means “only A, or only B, or only C, or any combination of A, B and C.” The terms “system” and “network” may be used interchangeably. The term “and / or” is only an association relationship for describing associated objects and represents that three relationships may exist such that A and / or B may indicate that A exists alone, A and B exist at the same time, or B exists alone. The character “ / ” generally represents that the associated objects are in an “or” relationship.

[0031] For the purposes of explanation and non-limitation, specific details, such as functional entities, techniques, protocols, and standards, are set forth for providing an understanding of the disclosed technology. In other examples, detailed disclosure of well-known methods, technologies, systems, and architectures are omitted so as not to obscure the present disclosure with unnecessary details.

[0032] Persons skilled in the art will immediately recognize that any network function(s) or algorithm(s) disclosed may be implemented by hardware, software, or a combination of software and hardware. Disclosed functions may correspond to modules which may be software, hardware, firmware, or any combination thereof.

[0033] A software implementation may include computer-executable instructions and / or Artificial Intelligence (AI) / Machine Learning (ML) module(s) stored on a computer-readable medium, such as memory or other type of storage devices. One or more microprocessors or general-purpose computers with communication processing capability may be programmed with corresponding computer-executable instructions and perform the disclosed network function(s), AI / ML module(s), or algorithm(s). The AI / ML module(s) may be implemented with a supervised learning approach, a semi-supervised learning approach, an unsupervised learning approach (e.g., Transductive approach and Inductive approach), a federated learning approach, or a reinforcement learning (RL) approach, but the present disclosure is not limited thereto. The computer-executable instructions associated with the AI module(s) and / or the ML module(s) may include but are not limited to, data management instructions (e.g., collection instructions, validation instructions…etc.), model monitoring and management instructions (e.g., Network (NW) key performance indicators (KPIs) monitoring, model input / output monitoring, model selection / switching / update / upload / download, model (de)activation, model identification, functionality selection…etc.), and / or pre-process input instructions.

[0034] The microprocessors or general-purpose computers may include Application-Specific Integrated Circuits (ASICs), programmable logic arrays, Central Processing Units (CPUs), Tensor Processing Units (TPUs), Graphics Processing Units (GPUs), General-purpose computing on GPUs (GPGPU, or less often GPGP), and / or using one or more Digital Signal Processors (DSPs). Although some of the disclosed implementations are oriented to software installed and executing on computer hardware, alternative implementations implemented as firmware, as hardware, or as a combination of hardware and software are well within the scope of the present disclosure. The computer-readable medium may include, but is not limited to, Random Access Memory (RAM), Dynamic Random Access Memory (DRAM), High Bandwidth Memory (HBM), Magnetoresistive Random Access Memory (MRAM), Ferroelectric Random Access Memory (FRAM), Resistive Random Access Memory (RRAM), Read-Only Memory (ROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), flash memory (or other memory technology), Compact Disc Read-Only Memory (CD-ROM) , Digital Versatile Disks (DVD) (or other optical disk storage), magnetic cassettes, magnetic tape, magnetic disk storage (or other magnetic storage devices), or any other equivalent medium capable of storing computer-readable instructions.

[0035] A radio communication network architecture such as a Long-Term Evolution (LTE) system, an LTE-Advanced (LTE-A) system, an LTE-Advanced Pro system, or a 5G NR Radio Access Network (RAN), 5G-Advanced (5G-A) system, or an open radio access network (O-RAN) may typically include at least one base station (BS), at least one UE, and one or more optional network elements that provide connection within a network. The BS and one or more optional network elements enable the UE to access a radio network. The UE may communicate with the network, such as a Core Network (CN), an Evolved Packet Core (EPC) network, an Evolved Universal Terrestrial RAN (E-UTRAN), a Next-Generation Core (NGC), a 5G Core (5GC), or an internet via a RAN established by one or more BSs and the network elements / functions.

[0036] A UE may include, but is not limited to, a mobile station, a mobile terminal or device, or a user communication radio terminal. The UE may be a portable radio equipment that includes, but is not limited to, a mobile phone, a tablet, a wearable device, a sensor, a vehicle, a virtual reality (VR) device, an augmented (AR) device, an Internet of Things (IoT) device, an unmanned aerial vehicle (UAV), or a Personal Digital Assistant (PDA) with wireless communication capability. The UE may be configured to receive and transmit signals over an air interface to one or more cells in a RAN. In some implementations, the UE may be an AI / ML-enabled device and / or an AI / ML capable device that is equipped with AI module(s) and / or ML module(s).

[0037] The BS may be configured to provide communication services according to at least a Radio Access Technology (RAT), such as Worldwide Interoperability for Microwave Access (WiMAX), Global System for Mobile communications (GSM) that is often referred to as 2G, GSM Enhanced Data rates for GSM Evolution (EDGE) RAN (GERAN), General Packet Radio Service (GPRS), Universal Mobile Telecommunication System (UMTS) that is often referred to as 3G based on basic Wideband-Code Division Multiple Access (W-CDMA), High-Speed Packet Access (HSPA), LTE, LTE-A, evolved / enhanced LTE (eLTE) that is LTE connected to 5GC, NR (often referred to as 5G), 5G-A, and / or LTE-A Pro. However, the scope of the present disclosure is not limited to these protocols.

[0038] The BS may include, but is not limited to, a node B (NB) in the UMTS, an evolved node B (eNB) in LTE or LTE-A, a radio network controller (RNC) in UMTS, a BS controller (BSC) in the GSM / GERAN, a next-generation eNB (ng-eNB) in an Evolved Universal Terrestrial Radio Access (E-UTRA) BS in connection with 5GC, a next-generation Node B (gNB) in the 5G-RAN or in the 5G Access Network (5G-AN), or any other apparatus capable of controlling radio communication and managing radio resources within a cell. The BS may serve one or more UEs via a radio interface. Although the gNB is used as an example in some implementations within the present disclosure, it should be noted that the disclosed implementations may also be applied to other types of base stations. In some implementations, the BS may be an AI / ML-enabled device and / or an AI / ML capable device that is equipped with AI module(s) and / or ML module(s).

[0039] The BS may be operable to provide radio coverage to a specific geographical area using multiple cells forming the RAN. The BS may support the operations of the cells. Each cell may be operable to provide services to at least one UE within its radio coverage.

[0040] Each cell (may often referred to as a serving cell) may provide services to one or more UEs within the cell’s radio coverage, such that each cell schedules the downlink (DL) (and optionally uplink (UL) resources) to at least one UE within its radio coverage for DL (and optionally UL packet transmissions from the UE). The BS may communicate with one or more UEs in the radio communication system via the cells.

[0041] A cell may allocate sidelink (SL) resources for supporting the Proximity Services (ProSe), LTE SL services, LTE / NR sidelink communication services, LTE / NR sidelink discovery services, and / or LTE / NR Vehicle-to-Everything (V2X) services. In addition, a cell may allocate DL and / or UL resources for supporting Multicast / Broadcast Service (MBS) services, Non-Terrestrial Networks (NTN) services, positioning services, power serving services and / or Network Energy Saving (NES) services.

[0042] In Multi-RAT Dual Connectivity (MR-DC) cases, the primary cell of a Master Cell Group (MCG) or a Secondary Cell Group (SCG) may be referred to as a Special Cell (SpCell). A Primary Cell (PCell) may include the SpCell of an MCG. A Primary SCG Cell (PSCell) may include the SpCell of an SCG. MCG may include a group of serving cells associated with the Master Node (MN), including the SpCell and optionally one or more Secondary Cells (SCells). An SCG may include a group of serving cells associated with the Secondary Node (SN), including the SpCell and optionally one or more SCells.

[0043] As discussed above, the frame structure for NR may support flexible configurations for accommodating various next generation (e.g., 5G) communication requirements, such as Enhanced Mobile Broadband (eMBB), Massive Machine Type Communication (mMTC), and Ultra-Reliable and Low-Latency Communication (URLLC), while fulfilling high reliability, high data rate, and low latency requirements. The Orthogonal Frequency-Division Multiplexing (OFDM) technology in the third generation partnership project (3GPP) may serve as a baseline for an NR waveform. The scalable OFDM numerology, such as adaptive sub-carrier spacing, channel bandwidth, and Cyclic Prefix (CP), may also be used.

[0044] Two coding schemes may be considered for NR, specifically, Low-Density Parity-Check (LDPC) code and Polar Code. The coding scheme adaption may be configured based on channel conditions and / or service applications.

[0045] At least the DL transmission data, a guard period, and UL transmission data should be included in a transmission time interval (TTI) of a single NR frame. The respective portions of the DL transmission data, the guard period, and the UL transmission data should also be configurable based on, for example, the network dynamics of NR. SL resources may also be provided in an NR frame to support ProSe services or V2X services.

[0046] Any two or more than two of the following paragraphs, (sub)-bullets, points, actions, behaviors, terms, or claims described in the present disclosure may be combined logically, reasonably, and properly to form a specific method.

[0047] Any sentence, paragraph, (sub)-bullet, point, action, behaviors, terms, or claims described in the present disclosure may be implemented independently and separately to form a specific method.

[0048] Dependency, e.g., “based on”, “more specifically”, “preferably”, “in one embodiment”, “in some implementations”, etc., in the present disclosure is just one possible example which would not restrict the specific method.

[0049] In some implementations, all the designs / embodiment / implementations introduced within this disclosure are not limited to be applied for dealing with the problems discussed within this disclosure. For example, the described embodiments may be applied to solve other problems that exist in the RAN of wireless communication systems. In some implementations, all of the numbers listed within the designs / embodiment / implementations introduced within this disclosure are just examples and for illustration, for example, of how the described methods are executed.

[0050] The terms, definitions, and abbreviations given in the present disclosure are either imported from existing documentation (e.g., European Telecommunications Standards Institute (ETSI), International Telecommunication Union (ITU), or elsewhere) or newly created by 3GPP experts whenever the need for precise vocabulary is identified.

[0051] In 3GPP Rel-19, the application of AI / ML techniques to NR air interface has been introduced. The objectives of the specification mainly include (1) AI / ML general framework for one-sided AI / ML models, (2) beam management and DL Tx beam prediction for both UE-sided model and NW-sided model, and (3) direct AI / ML positioning or AI / ML-assisted positioning for positioning accuracy enhancements. With the motivation, it may be needed to investigate and specify the necessary measurements, signaling, or mechanism to facilitate the Life Cycle Management (LCM) operations subject to beam management and positioning. In addition, legacy mechanisms (e.g., beam management, and positioning without AI / ML techniques) may still be required when the corresponding or affected AI / ML functionalities do not apply, thereby ensuring that the UE continues to operate normally in current 5G access network.

[0052] AI / ML-based Beam Management (BM)

[0053] The beam management may include DL Tx beam prediction for both the UE-sided model and the NW-sided model. The beam prediction may include predicting a Set A of beams based on a Set B of beams and an AI / ML model. For example, the AI / ML model may receive the Set B of beams as input and generate the Set A of beams as output. In some implementations, the measurement results may correspond to the inference / predicted results.

[0054] BM-Case1: BM-Case1 may refer to a spatial-domain DL Tx beam prediction for Set A of beams based on measurement results of Set B of beams. In some implementations, the Set B of beams may include beams received from one direction and the Set A of beams may include beams received from another direction. In some implementations, the Set B of beams may include wide beams (e.g., Synchronization Signal Block (SSB)) and the Set A of beams may include narrow beams (e.g., Channel State Information-Reference Signal (CSI-RS)). In some implementations, the AI / ML model training and inference may be performed at the NW side. In some implementations, the AI / ML model training and inference may be performed at the UE side. In some implementations, Set A and Set B may be different. In some implementations, Set B may be a subset of Set A.

[0055] BM-Case2: BM-Case2 may refer to a temporal DL Tx beam prediction for Set A of beams based on the historic measurement results of Set B of beams. In some implementations, the Set B of beams may include beams that were received in the past and the Set A of beams may include beams that are expected to be received in the future. In some implementations, the AI / ML model training and inference may be performed at the NW side. In some implementations, the AI / ML model training and inference may be performed at the UE side. In some implementations, Set A and Set B may be different. In some implementations, Set B may be a subset of Set A. In some implementations, Set A and Set B may be the same.

[0056] Functional framework for AI / ML for NR air interface

[0057] FIG. 1 is a block diagram 100 illustrating a functional framework for AI / ML for NR air interface, according to an example implementation of the present disclosure. As illustrated in FIG. 1, data collection 102 is a function that provides input data to the model training, management, and inference functions. Training data may include data needed as input for the AI / ML model training function. Monitoring data may include data needed as input for the management of AI / ML models or AI / ML functionalities. Inference data may include data needed as input for the AI / ML inference function.

[0058] Model training 104 is a function that performs AI / ML model training, validation, and testing, which may generate model performance metrics that may be used as part of the model testing procedure. The model training function is also responsible for data preparation (e.g., data pre-processing and cleaning, formatting, and transformation) based on training data delivered by a data collection function, if required. In some implementations, there may be a model storage 106 in the framework. The model training 104 may deliver a trained / updated model to the model storage 106. In some implementations, the trained / updated model may include trained, validated, and tested AI / ML models. In some implementations, the trained / updated model may include an updated version of the model.

[0059] Management 108 is a function that oversees the operation (e.g., selection / (de)activation / switching / fallback) and monitoring (e.g., performance) of AI / ML models or AI / ML functionalities. The management 108 is also responsible for making decisions to ensure the proper inference operation based on data received from the data collection function and the inference function. In some implementations, a management instruction represents essential input information to manage the inference function. Concerning information may include selection / (de)activation / switching of AI / ML models or AI / ML-based functionalities, fallback to non-AI / ML operation (e.g., not relying on inference process), and so on. In some implementations, a model transfer / delivery request is used to request model(s) to the model storage 106. In some implementations, performance feedback and retraining request represent essential input information for the model training 104 (e.g., for model (re)training or updating purposes).

[0060] Inference 110 is a function that provides outputs from the process of applying AI / ML models or AI / ML functionalities, using the data that is provided by the data collection 102 (e.g., inference data) as an input. The inference 110 is also responsible for data preparation (e.g., data pre-processing and cleaning, formatting, and transformation) based on the inference data delivered by the data collection 102, if required. In some implementations, inference output serves as data for the management 108 to monitor the performance of AI / ML models or AI / ML functionalities.

[0061] Model storage 106 is a function responsible for storing trained / updated models that may be used to perform the inference function. The model storage 106 may transfer / deliver an AI / ML model to the inference 110.

[0062] Target Scenarios

[0063] The present disclosure describes implementations regarding beam management and handling of link / channel failures. FIG. 2 is a block diagram 200 illustrating a functional framework for handling link or channel failures without AI / ML techniques, according to an example implementation of the present disclosure.

[0064] Beam management 202 is a function in which the BS (e.g., a gNB) may request the UE to report beams measurement results (e.g., signal strength) and attempt to maintain the Quasi Co-Location (QCL) and Transmission Configuration Indicator (TCI) state among the Tx and Rx beams, thereby ensuring appropriate Tx / Rx Beam refinement.

[0065] Layer 1 (L1) may provide the Medium Access Control (MAC) layer indications of beam failure instances (BFIs). The MAC layer may count the indications and declare failure when the configured maximum number of BFI indications has been reached. The number of BFI indications reaching the configured maximum number may be referred to as a Beam Failure Detection (BFD) trigger condition, as illustrated in FIG. 2. When the UE has detected a beam failure, the UE may perform a procedure of Beam Failure Recovery (BFR) 204. In the BFR 204, the UE may be provided with a set of resources for the recovery procedure, and the UE may select the best one among them to perform BFR request accordingly. If the BFR 204 is successful, the UE may perform beam management 202 on the same cell.

[0066] For the measurement trigger conditions in FIG. 2, the BS (e.g., a gNB) may configure one or more measurement events and corresponding reporting procedures to monitor the channel quality of the UE. Unlike L1 detection of BFI, the measurement in this context represents a cell-level channel quality assessment. The UE may perform beam consolidation and apply L3 filtering to the consolidation results to derive an indication of the overall cell quality. When the quality of the serving cell degrades or when the quality of a neighboring cell exceeds that of the serving cell under certain conditions, the gNB may transmit an Radio Resource Control (RRC) reconfiguration with sync message to trigger the handover process to a new serving cell. The RRC reconfiguration with sync 206 may also be referred to as a handover (HO). After the handover or the RRC reconfiguration with sync 206, the UE may perform beam management on the new serving cell 208.

[0067] The gNB may configure a set of parameters, and the UE may monitor whether the quality of the current serving cell remains below a configured threshold for a duration specified by a timer (e.g., a T310 timer). If this condition is satisfied, the UE may determine that a Radio Link Failure (RLF) 210 has occurred. The UE may attempt to recover from the RLF 210, and upon successful recovery, may resume performing the beam management 202 procedure on the same serving cell. The UE may initiate an RRC reestablishment procedure to connect to a new serving cell. The UE may further transmit a corresponding report to the new serving cell to indicate the failure. Upon successfully connecting to the new serving cell, the UE may perform beam management on the new serving cell 208.

[0068] Similarly, if the UE is instructed to perform a handover or the RRC reconfiguration with sync 206 but is unable to camp on the indicated serving cell due to an interruption exceeding a configured timer (e.g., a T304 timer), or if the random access procedure fails, the UE may also perform the RRC reestablishment procedure and attempt to identify a new serving cell. This condition may be referred to as HO Failure (HOF) 212. Upon successfully connecting to the new serving cell, the UE may perform beam management on the new serving cell 208.

[0069] With AI / ML techniques, the training and inference may be performed by the NW side and / or the UE side. When the NW side applies AI / ML techniques to beam management, the NW may predict and identify a good beam for the UE to maintain good beam-level and cell-level channel quality. Hereafter, the failure handling may also be performed by the NW side. On the contrary, the UE may detect and select a new beam or new cell when the UE side AI / ML model is running, but the UE may still be required to report the inference results to the gNB and await further instructions. In the present disclosure, both scenarios are considered, and corresponding designs are proposed.

[0070] AI / ML Failure Management

[0071] Under AI / ML techniques applicable to BM-Case1 and BM-Case2, the model may predict a set A (as an inference result) based on the measurement of a set B. However, such prediction may result in a poor channel quality, and it remains unclear how the prediction cooperates with existing procedures or whether additional procedures need to be designed. In particular, the definition of a failure under the AI / ML models for BM-Case1 and BM-Case2 may need to be clarified, so that the system may apply appropriate procedures following such classification.

[0072] For example, as illustrated in FIG. 2, BFR will be applied for beam quality detection, beam monitoring, and beam recovery. It shall be clarified whether these behaviors will keep same procedure and / or condition to trigger while AI / ML techniques was introducing. Moreover, what AI / ML can additionally improve during the procedure (if still required) needs further investigation.

[0073] For example, as illustrated in FIG. 2, BFR may be applied for beam quality detection, beam monitoring, and beam recovery. It may need to be clarified whether these procedures and / or the conditions for their triggering remain the same when AI / ML techniques are introduced. Furthermore, it may require further investigation as to what additional improvements AI / ML techniques may provide during such procedures, if the procedures are still required.

[0074] In the present disclosure, failures under activated AI / ML functionalities may be classified as below.

[0075] Type 1 failure: When the network-side or UE-side additional condition (AC) changes and the activated AI / ML functionalities cannot re-associate with any provided ACs, the activated AI / ML functionalities (e.g., inference generation based on the corresponding AI / ML training, or AI / ML functionalities already enabled for performing inference) may fail to satisfy the consistency requirements subject to the updated AC. Another scenario that may also be considered a type 1 failure is a shortage of device hardware resources (e.g., memory or UE battery), which may prevent the UE from supporting the execution of AI / ML techniques, thereby preventing inference generation. When a type 1 failure occurs, even if performance monitoring results and channel link quality remain adequate, the AI / ML functionalities may be unable to generate feasible inferences in a timely manner. Consequently, the UE may declare the failure to the gNB or NW and fall back to conventional mechanisms. Type 1 failures may occur only for UE-side models, and the gNB or NW may provide failure handling configurations to instruct the UE how to fall back and respond accordingly.

[0076] Type 2 failure: The AI / ML functionality is applicable and activated, but the inference results fail to achieve satisfactory beam or cell channel quality. Type 2 failure may be further subdivided into type 2-a failure (e.g., beam failure) and type 2-b failure (e.g., link failure). In type 2-a failure, either beam failure instances or the beam inference results may be used for failure detection. In type 2-b failure, failure detection may rely on existing Radio Resource Management (RRM) performance criteria or on the consolidated beam inference results. Upon detection of a type 2 failure, the UE may need to ensure that performance monitoring results remain adequate before the detection outcome is applied for failure management.

[0077] Type 3 failure: The AI / ML functionality is applicable and activated, but the prediction accuracy performance, which may be verified through a UE-assisted performance monitoring operation, is determined to fall outside of an acceptable range (e.g., failure detection). The out-of-acceptance condition may include at least one of the following options:

[0078] Option 1-1: Top-1 or Top-K beam prediction accuracy (with or without a margin), determined by comparing the prediction results with the Top-1 or Top-K beam based on measurements from a resource set or resource, is lower than a specified threshold. In some implementations, the criteria may be configured such that if one or at least N beam predictions (or measurement instances based on the resource set or resource) among the K beam predictions falls below the specified threshold, the UE recognizes the AI / ML functionality performance as being out of acceptance and triggers corresponding type 3 failure handling. K and N may be positive integers. In some implementations, a time interval T may be configured such that, if any of the Top-K beam prediction accuracy values fall below the specified threshold for a duration longer than the interval T, the UE recognizes the AI / ML functionality as being out of acceptance and triggers corresponding type 3 failure handling.

[0079] Option 1-2: The L1-RSRP difference, determined based on an actual measurement of the L1- Reference Signal Received Power (RSRP) of one or more of the Top-K predicted beams and L1-RSRP measurements from a resource set or resource, exceeds a specified threshold.

[0080] In some implementations, the network may configure a number N of L1-RSRP difference instances. If more than N of the L1-RSRP differences among the Top-K predicted beams exceed the specified threshold, the UE may recognize the AI / ML functionality performance as being out of acceptance and trigger type 3 failure handling.

[0081] In some implementations, a time interval T may be configured. If any of the Top-K beams exhibit an L1-RSRP difference larger than the specified threshold for a duration longer than the interval T, the UE may recognize the AI / ML functionality as being out of acceptance and trigger type 3 failure handling.

[0082] In some implementations, both a number N (e.g., the number of unacceptable L1-RSRP differences per monitoring instance) and a time interval T may be configured. If more than N of the L1-RSRP differences among the Top-K predicted beams exceed the specified threshold, and this condition persists longer than the interval T, the UE may recognize the AI / ML functionality as being out of acceptance and trigger type 3 failure handling.

[0083] In some implementations, a time interval T may be configured on a per-beam basis. If a Top-K beam with an L1-RSRP difference exceeding the specified threshold continues to remain in the Top-K set for longer than the interval T, the UE may recognize the AI / ML functionality as being out of acceptance and trigger type 3 failure handling.

[0084] In some implementations, a time interval T may be configured as a holding time. If, within the holding time, the total number of measurement instances among the Top-K beams whose L1-RSRP difference exceeds the specified threshold is greater than a configured number N, the UE may recognize the AI / ML functionality as being out of acceptance and trigger type 3 failure handling.

[0085] Option 1-3: The RSRP difference between a predicted RSRP and the measured L1-RSRP of corresponding beam(s) of a resource set or resources exceeds (or falls below) a specified threshold. Similar to Options 1-1 and 1-2, the NW may configure at least one of the number of beams or the time interval. Based on the configuration, the UE may determine whether the AI / ML functionality is out of acceptance and trigger corresponding type 3 failure handling.

[0086] Option 1-4: The probability information associated with the predicted beam(s) being the Top-1 or Top-K beam is lower than a specified threshold. Similar to the previous options, the NW may configure at least one of the number of samples or the time interval during which the probability remains below the threshold. Based on the configuration, the UE may determine that the AI / ML functionality is out of acceptance and trigger corresponding type 3 failure handling.

[0087] Considering that the failure reasons and the corresponding resolutions may differ, the NW and the UE may include a mechanism to detect the respective failure types as described above, and to determine whether to fall back to conventional functions. Otherwise, performance degradation may occur due to the use of an inappropriate AI / ML algorithm. In addition, it may need to be clarified how the state of the AI / ML functionality is reflected and changed, in order to ensure that the overall LCM still operates appropriately. FIG. 3 is a flowchart illustrating a procedure 300 for life cycle management of the AI / ML functionality under failure handling, according to an example implementation of the present disclosure.

[0088] For one or more AI / ML functionalities 302, the NW may initiate the procedure 300 by transmitting a UE Capability Enquiry message to the UE. The UE may report its supported / supportable AI / ML functionalities 304. After the UE indicates the supportable AI / ML functionalities 304, the UE may determine what functionalities are applicable based on the given NW / UE-side additional condition(s), model availability, and / or device hardware situation. Consequently, the UE may respond with the applicable AI / ML functionalities 306 to the NW, for example, in response to a request from the NW to report the applicable functionalities or after receiving particular configurations.

[0089] After receiving, from the UE, response related to the applicable AI / ML functionalities 306, the NW may explicitly or implicitly activate the applicable AI / ML functionalities, for example, by sending instruction / configuration to the UE. For the activated AI / ML functionalities 308, the UE may then adopt the AI / ML techniques for beam management and / or RRM management.

[0090] In some implementations, if the UE does not receive instructions or configurations from the NW to explicitly or implicitly activate its applicable functionalities after responding with its applicable functionalities, the UE may identify the applicable functionalities as deactivated functionalities 310 and may perform conventional functions 324 to handle the failure (e.g., fallback to conventional functions).

[0091] In some implementations, if the UE receives instructions or configurations from the NW to explicitly or implicitly deactivate its applicable functionalities, the UE may identify the applicable functionalities as deactivated functionalities 310 and may perform conventional functions 324 to handle the failure (e.g., fallback to conventional functions).

[0092] In some implementations, the default activation state of the AI / ML functionalities may be deactivated after the UE reports its applicable functionalities 306.

[0093] When a Type 1 failure 312 occurs or is detected, the UE may change the corresponding AI / ML functionality to a non-applicable state 318 and deactivate the AI / ML functionality accordingly. The UE may report the Type 1 failure to the NW via UE Assistance Information (UAI) or an RRC message, and may fall back to conventional functions 324.

[0094] When a Type 2 failure 314 occurs or is detected, the UE may perform new failure handling functions 320, which may depend on whether a Type 2-a failure or Type 2-b failure is detected. When a Type 3 failure 316 occurs or is detected, the UE may deactivate the AI / ML functionality 322 and fall back to conventional functions 324.

[0095] With respect to the Type 1 failure 312, the fallback operation, such as reverting to the conventional functions 324, may be performed at one of the following timings:

[0096] Option 2-1: The fallback may be performed immediately after the UE identifies the non-applicability of AI / ML functionalities (e.g., when a Type 1 failure occurs and is detected).

[0097] The UE may declare the Type 1 failure while operating with conventional functions. In some implementations, the UE may fall back to conventional beam management procedures, such as procedure P1 (e.g., initial beam establishment), procedure P2 (e.g., transmit beam refinement), or procedure P3 (e.g., receive beam refinement). In addition, the UE may request uplink (UL) resources (e.g., by transmitting a Scheduling Request (SR) or Buffer Status Report (BSR) in accordance with legacy specifications) to deliver UAI, where the UAI may include the failure declaration information.

[0098] The failure declaration information may include at least one of the following elements: the failure type (e.g., Type 1 in this case); the failure occurrence time (e.g., identified by System Frame Number (SFN) or Coordinated Universal Time (UTC)); the failure cause (e.g., indicated by a cause ID); the AI / ML functionality state information, including applicability (e.g., applicable or non-applicable) and activation / deactivation status; and the availability of the generated inference (e.g., a single bit indicating whether to release or keep the inference). The cause ID may be indicated in a pre-defined table. A failure event (e.g., out-of-memory, low battery level) may be mapped to a corresponding ID. The UE may indicate the cause ID corresponding to the occurred event to the NW / gNB according to the pre-defined table. Furthermore, if the UE indicates that the generated inference is to be kept, the NW may, depending on NW implementation, query the UE for the inference report or directly apply the inference when the AI / ML functionality becomes applicable and activated again. The UE may also assume that stored configurations (e.g., CSI-RS configuration, Positioning Reference Signal (PRS) configuration, CSI report configuration, or TCI state-related configuration) remain valid and may apply such configurations for conventional functions without modification.

[0099] By default, if the inference result (e.g., a beam prediction) was kept, the UE may not use the inference result as an actual measurement result during the initiation of conventional functions, unless the NW configures the applicability during the AI / ML functionality configuration.

[0100] Option 2-2: The fallback may be performed after transmitting the UAI or RRC message to the NW without the confirmation from the NW / gNB.

[0101] By default, the UE may fallback to conventional functions (e.g., a BFR procedure or an RLF procedure) immediately after transmitting the UAI or RRC message to the gNB / NW for failure declaration, if the gNB / NW does not configure a fallback time offset value in the AI / ML functionality configuration. If the gNB / NW configures a fallback time offset value in the AI / ML functionality configuration, the UE may fallback to conventional functions after a time offset relative to the transmission time T of the UAI or RRC message for the failure declaration, where the fallback occurs at T + the configured time offset value.

[0102] Similar to Option 2-1, the failure declaration information may include at least one of the following elements: the failure type (e.g., Type-1 in this case), the failure occurrence time (e.g., indicated by SFN or UTC time), the failure cause, a state report of the AI / ML functionality (including applicability and activation / deactivation status), and the availability of the generated inference (e.g., whether to release or keep).

[0103] If the UE is unable to transmit the UAI or RRC message within a duration (e.g., a time value configured by the gNB / NW via RRC signaling) to declare the failure, or if the UE does not receive any further configuration from the gNB / NW after transmitting the UAI or RRC message for failure declaration, the UE may initiate a cell reselection procedure.

[0104] Option 2-3: The fallback may be performed after receiving confirmation from the NW / gNB in response to the UAI or RRC message corresponding to a failure declaration.

[0105] The failure declaration content may include at least one of the following: the failure type (e.g., Type-1 in this case), the failure occurrence time (e.g., indicated by SFN or UTC time), the failure cause, and / or the availability of the generated inference (e.g., release or keep). The confirmation may be an acknowledgement of UAI reception or an RRC response message. In some implementations, the confirmation may include a conventional function re-configuration information element (IE), which the UE applies while performing conventional functions. In some implementations, the confirmation may indicate a substitute functionality (e.g., a functionality ID or an associated ID related to the substitute functionality), and the UE may activate the indicated substitute functionality upon receiving the configuration message.

[0106] Option 2-4: Upon detecting a Type-1 failure, the UE may select a substitute functionality that is applicable but not yet activated. The UE may then transmit a UAI or RRC message to inform the NW of the functionality change. The UAI or RRC message may include at least one of the following: the failure type (e.g., Type-1), the failure occurrence time (e.g., indicated by SFN or UTC time), the failure cause (indicated by a cause ID), the availability of the generated inference (e.g., a single bit indicating release or keep), and an indication of the substitute functionality (e.g., a functionality ID or an associated ID related to the substitute functionality).

[0107] With respect to a Type 2-a failure, the AI / ML functionalities may be operating properly (e.g., the UE and NW can verify performance via performance monitoring), but the beam predicted by the AI / ML algorithm fails to satisfy the link performance requirements. Specifically, the UE may detect this failure by checking BFI and / or relevant trigger conditions. Several implementation options related to the detection of the Type 2-a failure are provided below.

[0108] Option 3-1: Separate BFR parameters (e.g., beamFailureInstanceMaxCount and beamFailureDetectionTimer) may be configured for AI / ML functionalities. The UE may apply the respective parameters while the AI / ML functionalities are activated. For example, smaller values of the beamFailureInstanceMaxCount and beamFailureDetectionTimer may trigger a faster BFR, which could be supported by AI / ML techniques that are able to identify a new candidate beam without performing actual measurements. The AI / ML-specific BFR parameters may be configured together with the conventional BFR configuration message, or may be configured together with the AI / ML functionality configuration.

[0109] Option 3-2: Common BFR parameters (e.g., beamFailureInstanceMaxCount and beamFailureDetectionTimer) may be applied regardless of whether the AI / ML functionalities are activated or deactivated. It should be noted that if the UE does not receive any AI / ML-specific BFR parameters, the UE may assume that the common BFR parameters are to be applied.

[0110] Option 3-3: Different BFD mechanisms may be applied for Type 2-a failure detection such that the UE does not rely on the BFI and the existing count and timer mechanisms to detect a beam failure. Instead, when the AI / ML functionalities are activated and the quality of N consecutive predicted beams is worse than a threshold, the physical (PHY) layer may notify the MAC layer, and a recovery procedure may be initiated. The consecutive number N and the threshold may be configured by the NW.

[0111] Additionally, the recovery procedure may be performed via the CSI reporting framework (e.g., through Uplink Control Information (UCI)), where the UE may indicate the failure to the source cell when the conditions are met. In some implementations, this behavior may also be achieved for AI / ML BM Case-2, in which the NW configures the UE with N future time instances for inference, and the UE may report those N future time instances in a report when all of them are worse than the quality threshold. The reporting configuration may indicate the threshold. The UE may transmit the aperiodic report only when the condition is satisfied.

[0112] The conventional BFD mechanism and the AI / ML-specific BFD mechanism may run independently. However, the NW / gNB may configure the priority such that one mechanism may overwrite the other. In some implementations, the conventional BFD mechanism may be suspended when the AI / ML functionalities are activated and the AI / ML-specific BFD mechanism is configured and applied.

[0113] The UE may transmit an event-triggered report when a Type 2 failure is detected. The event may be based on the quality of a measured beam (set B of beams) or a predicted beam (set A of beams). If the predicted beam has an L1-RSRP value lower than a threshold, or if the predicted beam meets the failure trigger condition, the UE may transmit a beam failure indication without NW scheduling (e.g., by using Configured Grant (CG) resources or UCI).

[0114] Option 3-4: When the NW / gNB activates applicable UE functionalities for beam management, the UE may be configured, via an RRC parameter, to perform an AI / ML-based beam failure recovery procedure. The RRC parameter may be used to enable the AI / ML-based beam failure recovery. In some implementations, if the UE is enabled for AI / ML-based beam failure recovery, the UE may be further configured with a set of reference signals for measurement reporting or inference.

[0115] If the UE detects a beam failure during beam failure detection, the UE may perform model inference based on the configured set of reference signals. Then the UE may transmit a scheduling request (SR) or buffer status report (BSR) to request resources for sending the inference report, or may send the inference report via UCI to the NW / gNB. After receiving the inference report, the NW / gNB may indicate a new beam (e.g., from set A or set B) via DCI.

[0116] In some implementations, if the UE is enabled for AI / ML-based beam failure recovery, the UE may be further configured with a set of reference signals for measurement reporting or inference. Upon detecting a beam failure, the UE may request resources (e.g., by SR / BSR) and report the measurement results related to the configured reference signals, where the measurement results may be used for inference of NW-side model. After the UE sends the measurement results, the NW / gNB may indicate a new beam (e.g., from the corresponding reference signal set, set A or set B) via DCI.

[0117] While the AI / ML functionalities are activated and BFR is triggered, the UE may apply new failure handling functions (e.g., the new failure function 320 in FIG. 3) for Type 2-a failure. The functions may include at least one of the following:

[0118] (a) The UE may keep concurrent AI / ML functionalities activated and continuously perform BM Case-1 and / or Case-2. Specifically, the UE may measure candidate RS from set B while indicating a new inference (candidate RS) from set A. To support this association, in the BFR MAC control element (CE), the reserved bit “R” may be set to “1” as an association indication. The NW may interpret the indicated candidate RS ID, which is configured and provided from set B, as being associated with an inference from set A. Otherwise, the candidate RS ID may be interpreted as the configured candidate beam from set B. In some implementations, the association ID may be appended in the BFR MAC CE while the candidate RS ID corresponds to an inference from set A. The association ID may be identical to an ID within the CSI framework or may be separately configured during BFR-related configurations. The candidate RS ID may correspond to the reference signal set associated with the association ID indicated in the BFR MAC CE.

[0119] (b) The UE may declare a Type 2-a failure, for example, by sending a report to the NW / gNB for this failure declaration. The declaration information may include at least a time stamp and an identification of the failed beam. The time stamp may be represented by SFN information or UTC time. The failed beam may be represented by, for example, a beam index, an SSB index, and / or a CSI-RS index.

[0120] (c) The UE may log the failure and the corresponding interval during user data collection, and may upload the logged data to the NW when requested by the NW or when a reporting condition is triggered. The log of failure interval may include labels indicating when the failure starts and when recovery is successfully completed.

[0121] (d) The UE may initiate the AI / ML functionalities for mobility management, such as activating the functionalities and using them for handover prediction. Conditional Handover (CHO) and L1 / L2 Triggered Mobility (LTM) mechanisms may also be initiated, and the UE may start the relevant measurements of neighboring cell SSBs and CSI-RSs.

[0122] For a Type 2-b failure, the activated AI / ML functionalities may be operating correctly (e.g., as verified by performance monitoring by the UE and NW), but the selected cell (e.g., after beam consolidation based on the AI / ML algorithm) may not meet the required cell performance criteria. Specifically, the UE may detect this link failure by monitoring consecutive “out-of-sync” indications. The UE may start timer T310 upon receiving an out-of-sync indication and may trigger a radio link failure (RLF) if the timer expires. It is also possible to apply the same or different RLF parameters while the AI / ML functionalities are activated. If no AI / ML-specific RLF configurations are provided, the UE may assume that the conventional RLF parameters are to be applied.

[0123] While an RLF occurs, the UE may perform cell reselection with new failure handling functions (e.g., the new failure function 320 in FIG. 3). The functions may include at least one of the following:

[0124] (a) The UE may deactivate concurrent AI / ML functionalities upon recognizing a Type 2-b failure. After completing RRC connection establishment with a new serving cell, the UE may be requested to reactivate the AI / ML functionalities when the new serving cell provides the NW-side additional condition and sends the corresponding configuration to the UE. The UE may also autonomously reactivate the AI / ML functionalities (e.g., without receiving configuration from the new serving cell) if the NW-side additional condition (e.g., ID of the additional condition or an associated ID) is the same as that of the previous serving cell.

[0125] (b) The UE may declare a Type 2-b failure, for example, by sending a report to the NW. The declaration information may include at least a time stamp (e.g., SFN or UTC time) and an identification of the failed cell (e.g., candidate cell ID, physical cell ID (PCI), or additionalPCIIndex). The UE may declare the Type 2-b failure by the same message as RLF reporting.

[0126] (c) The UE may log the failure and the corresponding interval during user data collection, and may upload the logged data to the NW when requested by the NW or when a reporting condition is triggered. The log of failure interval may include labels indicating when the failure starts and when recovery is successfully completed.

[0127] During performance monitoring, the UE may evaluate the AI / ML functionalities and determine whether the accuracy of the activated functionalities meets at least one specified performance metric. If one or more performance metrics (e.g., prediction accuracy) fail to satisfy the specified requirements, the UE may recognize a Type 3 failure and perform at least one of the following actions:

[0128] (a) The UE may report the Type 3 failure to the NW via L1 or L2 signaling. The reporting content may include the failure cause, such as which performance metric was not fulfilled, and the failure occurrence, indicating when the failure happened and which inference was unstable.

[0129] (b) The UE may automatically deactivate all related AI / ML functionalities (e.g., deactivate the AI / ML functionality 322 in FIG. 3) and fallback to conventional functions (e.g., conventional functions 324 in FIG. 3). The deactivation and fallback may occur prior to sending the Type 3 failure report to the NW.

[0130] (c) The UE may be requested to perform model transfer and delivery procedures to forward the AI / ML related setting and parameters.

[0131] (d) While falling back to conventional functions, the UE may refrain from using any inference results or previously collected measurement data. All measurement samples may be cleared, and the UE may perform new measurements and evaluate beam and cell performance by complying with the configurations for conventional functions.

[0132] (e) The UE may consider the related AI / ML functionalities to be non-applicable if the functionality declared with the Type 3 failure is a UE-side functionality.

[0133] (f) In some implementations, Option 2-4 (e.g., selecting a substitute functionality) may be applied to the Type 3 failure.

[0134] (g) In some implementations, Type 2 and Type 3 failures may be applied together. Upon identifying a Type 2 failure, the UE may further determine whether the failure is due to a beam failure or due to poor quality of the input (set B of beams) leading to a degraded output (set A of beams). Accordingly, when the UE transmits an indication of the Type 2 failure, the UE may additionally check the monitoring metrics within a specified time interval. If the results of the monitoring metrics fall below a threshold value, the UE may perform recovery for the Type 3 failure; otherwise, the UE may perform recovery for the Type 2 failure.

[0135] In some embodiments, the UE may need to suspend one or more AI / ML functionalities to retrain or update the associated model(s). If any AI / ML functionality is identified as suspended due to the UE’s internal model management, the UE may recognize a Type 1 failure and perform the corresponding actions as disclosed in the present disclosure.

[0136] FIG. 4 is a flowchart illustrating a method / process 400 performed by a UE for an AI / ML operation, according to an example implementation of the present disclosure. In the action 402, the process 400 may start by activating an AI / ML functionality. In the action 404, the process 400 may detect a failure in response to determining that at least one of the following conditions is satisfied: (i) the AI / ML functionality is not applicable; (ii) a beam failure or a radio link failure is detected based on an inference result of the AI / ML functionality; or (iii) a prediction accuracy of the AI / ML functionality falls outside of a predetermined acceptable range. These three conditions may correspond, respectively, to the Type 1 failure, Type 2 failure, and Type 3 failure as described in the present disclosure. In the action 404, the process 400 may transmit, to a BS, a failure report associated with the detected failure. In some implementations, the failure report may be transmitted via UAI or an RRC message. In some implementations, the failure report may be transmitted via L1 or L2 signaling. The process 400 may then end.

[0137] The steps / actions shown in FIG. 4 should not be construed as necessarily order dependent. The order in which the process is described is not intended to be construed as a limitation. Moreover, some of the actions shown in FIG. 4 may be omitted in some implementations and one or more actions shown in FIG. 4 may be combined.

[0138] The technical problem addressed by the method illustrated in FIG. 4 is how to determine when a failure associated with an activated AI / ML functionality occurs. By clearly defining the types of failures associated with the AI / ML functionalities, the UE is able to detect such failures and report them to the network. Consequently, the UE and / or the NW may adopt corresponding recovery procedures to ensure proper operation for continuous communication, either using the AI / ML functionalities or relying on conventional functions.

[0139] In some implementations, the AI / ML functionality may be determined to be not applicable (e.g., the non-applicable state 318 in FIG. 3) in response to determining at least one of the following conditions is satisfied: (a) an additional condition (AC) associated with the AI / ML functionality has changed; or a hardware shortage has been detected in the UE. For example, when the AC has changed, the activated AI / ML functionalities may fail to satisfy the consistency requirements subject to the updated AC. When the hardware shortage has been detected in the UE, the UE may fail to support the execution of AI / ML techniques, thereby failing to generate inference results.

[0140] In some implementations, the prediction accuracy of the AI / ML functionality may be determined to fall outside of the predetermined acceptable range in response to determining at least one of the following conditions is satisfied: (a) the prediction accuracy of the AI / ML functionality is lower than a first threshold; (b) a difference between a predicted Reference Signal Received Power (RSRP) of a specific beam and a measured RSRP of the specific beam is larger than a second threshold; or (c) a probability that predicted beams belong to measured top-K beams is lower than a third threshold. K may be a positive integer greater than or equal to one. In some implementations, the out-of-acceptance condition related to the Type 3 failure may include at least one of Options 1-1 to 1-4 as described in the present disclosure.

[0141] In some implementations, the UE may receive, from the BS, a set of BFR parameters specifically for the AI / ML functionality. The beam failure may be determined to be detected based on the set of BFR parameters specifically for the AI / ML functionality. In some implementations, the set of BFR parameters specifically for the AI / ML functionality may include beamFailureInstanceMaxCount and beamFailureDetectionTimer. The values of the BFR parameters may differ between AI / ML functionalities and non-AI / ML functionalities (e.g., conventional parameters). For example, the values of the beamFailureInstanceMaxCount and beamFailureDetectionTimer for AI / ML functionalities may be smaller than the corresponding values for non-AI / ML functionalities.

[0142] In some implementations, the UE may deactivate the AI / ML functionality in response to the detected failure. In some implementations, the UE may fallback to convention functions (e.g., the conventional functions 324 in FIG. 3) in response to deactivating the AI / ML functionality.

[0143] In some implementations, the UE may activate a second AI / ML functionality for handover prediction in response to detecting the beam failure based on the inference result of the AI / ML functionality. For example, the UE may initiate the second AI / ML functionality for mobility management. In some implementations, the UE may initiate a CHO or LTM procedure.

[0144] In some implementations, the failure report may include a time stamp indicating when the failure occurred. The time stamp may be indicated by SFN or UTC time. In some implementations, the failure report may further include a failure type (e.g., Type 1 failure, Type 2 failure, or Type 3 failure) and / or a failure cause (e.g., indicated by a cause ID).

[0145] FIG. 5 is a flowchart illustrating a method / process 500 performed by a BS for an AI / ML operation, according to an example implementation of the present disclosure. In the action 502, the process 500 may start by receiving, from a UE, a message indicating an AI / ML functionality supported by the UE. In the action 504, the process 500 may receive, from the UE, a failure report associated with a failure detected by the UE. The UE may activate the AI / ML functionality. The UE may detect the failure in response to determining that at least one of the following conditions is satisfied: (i) the AI / ML functionality is not applicable; (ii) a beam failure or a radio link failure is detected based on an inference result of the AI / ML functionality; or (iii) a prediction accuracy of the AI / ML functionality falls outside of a predetermined acceptable range. These three conditions may correspond, respectively, to the Type 1 failure, Type 2 failure, and Type 3 failure as described in the present disclosure. The process 500 may then end. The method illustrated in FIG. 5 is similar to that in FIG. 4, except that it is described from the perspective of the BS (instead of the UE).

[0146] FIG. 6 is a block diagram illustrating a node 600 for wireless communication, according to an example implementation of the present disclosure. As illustrated in FIG. 6, a node 600 may include a transceiver 620, a processor 628, a memory 634, one or more presentation components 638, and at least one antenna 636. The node 600 may also include a radio frequency (RF) spectrum band module, a BS communications module, a network communications module, and a system communications management module, Input / Output (I / O) ports, I / O components, and a power supply (not illustrated in FIG. 6).

[0147] Each of the components may directly or indirectly communicate with each other over one or more buses 640. The node 600 may be a UE or a BS that performs various functions disclosed with reference to FIGS. 1 through 5.

[0148] The transceiver 620 has a transmitter 622 (e.g., transmitting / transmission circuitry) and a receiver 624 (e.g., receiving / reception circuitry) and may be configured to transmit and / or receive time and / or frequency resource partitioning information. The transceiver 620 may be configured to transmit in different types of subframes and slots including, but not limited to, usable, non-usable, and flexibly usable subframes and slot formats. The transceiver 620 may be configured to receive data and control channels.

[0149] The node 600 may include a variety of computer-readable media. Computer-readable media may be any available media that may be accessed by the node 600 and include volatile (and / or non-volatile) media and removable (and / or non-removable) media.

[0150] The computer-readable media may include computer-storage media and communication media. Computer-storage media may include both volatile (and / or non-volatile media), and removable (and / or non-removable) media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules, AI / ML module(s), or data.

[0151] Computer-storage media may include RAM, ROM, EPROM, EEPROM, flash memory (or other memory technology), CD-ROM, Digital Versatile Disks (DVD) (or other optical disk storage), magnetic cassettes, magnetic tape, magnetic disk storage (or other magnetic storage devices), etc. Computer-storage media may not include a propagated data signal. Communication media may typically embody computer-readable instructions (e.g., computer-readable instructions related to AI module(s) and / or the ML module(s)), data structures, program modules, or other data in a modulated data signal, such as a carrier wave, or other transport mechanisms and include any information delivery media.

[0152] The term “modulated data signal” may mean a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. Communication media may include wired media, such as a wired network or direct-wired connection, and wireless media, such as acoustic, RF, infrared, and other wireless media. Combinations of any of the above listed components should also be included within the scope of computer-readable media.

[0153] The memory 634 may include computer-storage media in the form of volatile and / or non-volatile memory. The memory 634 may be removable, non-removable, or a combination thereof. Example memory may include solid-state memory, hard drives, optical-disc drives, etc. As illustrated in FIG. 6, the memory 634 may store a computer-readable and / or computer-executable instructions 632 (e.g., software codes) that are configured to, when executed, cause the processor 628 to perform various functions disclosed herein, for example, with reference to FIGS. 1 through 5. Alternatively, the instructions 632 may not be directly executable by the processor 628 but may be configured to cause the node 600 (e.g., when compiled and executed) to perform various functions disclosed herein.

[0154] The processor 628 (e.g., having processing circuitry) may include an intelligent hardware device, e.g., a Central Processing Unit (CPU), a microcontroller, an ASIC, etc. The processor 628 may include memory. The processor 628 may process the data 630 and the instructions 632 received from the memory 634, and information transmitted and received via the transceiver 620, the baseband communications module, and / or the network communications module. The processor 628 may also process information to send to the transceiver 620 for transmission via the antenna 636 to the network communications module for transmission to a CN.

[0155] One or more presentation components 638 may present data indications to a person or another device. Examples of presentation components 638 may include a display device, a speaker, a printing component, a vibrating component, etc.

[0156] In view of the present disclosure, it is obvious that various techniques may be used for implementing the disclosed concepts without departing from the scope of those concepts. Moreover, while the concepts have been disclosed with specific reference to certain implementations, a person of ordinary skill in the art may recognize that changes may be made in form and detail without departing from the scope of those concepts. As such, the disclosed implementations are to be considered in all respects as illustrative and not restrictive. It should also be understood that the present disclosure is not limited to the particular implementations disclosed and many rearrangements, modifications, and substitutions are possible without departing from the scope of the present disclosure.

Claims

1. A User Equipment (UE) for performing an Artificial Intelligence (AI) / Machine Learning (ML) operation, the UE comprising: 　　　　at least one processor; and 　　　　at least one non-transitory computer-readable medium coupled to the at least one processor and storing one or more computer-executable instructions that, when executed by the at least one processor, cause the UE to: 　　　　　　　　activate an AI / ML functionality; 　　　　　　　　detect a failure in response to determining that at least one of the following conditions is satisfied: 　　　　　　　　　　　　(i) the AI / ML functionality is not applicable; 　　　　　　　　　　　　(ii) a beam failure or a radio link failure is detected based on an inference result of the AI / ML functionality; or 　　　　　　　　　　　　(iii) a prediction accuracy of the AI / ML functionality falls outside of a predetermined acceptable range; and 　　　　　　　　transmit, to a Base Station (BS), a failure report associated with the detected failure.

2. The UE of claim 1, wherein the AI / ML functionality is determined to be not applicable in response to determining at least one of the following conditions is satisfied: 　　　　(a) an additional condition (AC) associated with the AI / ML functionality has changed; or 　　　　(b) a hardware shortage has been detected in the UE.

3. The UE of claim 1, wherein the prediction accuracy of the AI / ML functionality is determined to fall outside of the predetermined acceptable range in response to determining at least one of the following conditions is satisfied: 　　　　(a) the prediction accuracy of the AI / ML functionality is lower than a first threshold; 　　　　(b) a difference between a predicted Reference Signal Received Power (RSRP) of a specific beam and a measured RSRP of the specific beam is larger than a second threshold; or 　　　　(c) a probability that predicted beams belong to measured top-K beams is lower than a third threshold, 　　　　wherein K is a positive integer greater than or equal to one.

4. The UE of claim 1, wherein the one or more computer-executable instructions, when executed by the at least one processor, further cause the UE to: 　　　　receive, from the BS, a set of Beam Failure Recovery (BFR) parameters specifically for the AI / ML functionality, wherein: 　　　　the beam failure is determined to be detected based on the set of BFR parameters specifically for the AI / ML functionality.

5. The UE of claim 1, wherein the one or more computer-executable instructions, when executed by the at least one processor, further cause the UE to: 　　　　deactivate the AI / ML functionality in response to the detected failure.

6. The UE of claim 1, wherein the one or more computer-executable instructions, when executed by the at least one processor, further cause the UE to: 　　　　activate a second AI / ML functionality for handover prediction in response to detecting the beam failure based on the inference result of the AI / ML functionality.

7. The UE of claim 1, wherein: 　　　　the failure report comprises a time stamp indicating when the failure occurred.

8. A Base Station (BS) for performing an Artificial Intelligence (AI) / Machine Learning (ML) operation, the BS comprising: 　　　　at least one processor; and 　　　　at least one non-transitory computer-readable medium coupled to the at least one processor and storing one or more computer-executable instructions that, when executed by the at least one processor, cause the BS to: 　　　　　　　　receive, from a User Equipment (UE), a message indicating an AI / ML functionality supported by the UE; and 　　　　　　　　receive, from the UE, a failure report associated with a failure detected by the UE, wherein: 　　　　　　　　the UE activates the AI / ML functionality, and 　　　　　　　　the UE detects the failure in response to determining that at least one of the following conditions is satisfied: 　　　　　　　　　　　　(i) the AI / ML functionality is not applicable; 　　　　　　　　　　　　(ii) a beam failure or a radio link failure is detected based on an inference result of the AI / ML functionality; or 　　　　　　　　　　　　(iii) a prediction accuracy of the AI / ML functionality falls outside of a predetermined acceptable range.

9. The BS of claim 8, wherein the UE determines that the AI / ML functionality is not applicable in response to determining at least one of the following conditions is satisfied: 　　　　(a) an additional condition (AC) associated with the AI / ML functionality has changed; or 　　　　(b) a hardware shortage has been detected in the UE.

10. The BS of claim 8, wherein the UE determines that the prediction accuracy of the AI / ML functionality falls outside of the predetermined acceptable range in response to determining at least one of the following conditions is satisfied: 　　　　(a) the prediction accuracy of the AI / ML functionality is lower than a first threshold; 　　　　(b) a difference between a predicted Reference Signal Received Power (RSRP) of a specific beam and a measured RSRP of the specific beam is larger than a second threshold; or 　　　　(c) a probability that predicted beams belong to measured top-K beams is lower than a third threshold, 　　　　wherein K is a positive integer greater than or equal to one.

11. The BS of claim 8, wherein the one or more computer-executable instructions, when executed by the at least one processor, further cause the BS to: 　　　　transmit, to the UE, a set of Beam Failure Recovery (BFR) parameters specifically for the AI / ML functionality, wherein: 　　　　the UE determines that the beam failure is detected based on the set of BFR parameters specifically for the AI / ML functionality.

12. The BS of claim 1, wherein: 　　　　the UE deactivates the AI / ML functionality in response to the detected failure.

13. The BS of claim 8, wherein: 　　　　the UE activates a second AI / ML functionality for handover prediction in response to detecting the beam failure based on the inference result of the AI / ML functionality.

14. The BS of claim 8, wherein: 　　　　the failure report comprises a time stamp indicating when the failure occurred.

15. A method performed by a User Equipment (UE) for performing an Artificial Intelligence (AI) / Machine Learning (ML) operation, the method comprising: 　　　　activating an AI / ML functionality; 　　　　detecting a failure in response to determining that at least one of the following conditions is satisfied: 　　　　　　　　(i) the AI / ML functionality is not applicable; 　　　　　　　　(ii) a beam failure or a radio link failure is detected based on an inference result of the AI / ML functionality; or 　　　　　　　　(iii) a prediction accuracy of the AI / ML functionality falls outside of a predetermined acceptable range; and 　　　　transmitting, to a Base Station (BS), a failure report associated with the detected failure.

Images