Artificial intelligence auxiliary configuration method in wireless communication system

Abstract

The present disclosure relates generally to wireless communication systems and methods, and in particular to mechanisms for implementing an artificial intelligence framework for adaptively configuring an air communication interface of a wireless communication system. The AI network configuration functionality may be provided as a service for AI assisted network configuration (AI configuration service, or AICS). Such AICS may be requested and configured through various messaging and signaling mechanisms. AI model lifecycle management, including training, delivery, activation, inference, deactivation, handover, performance evaluation, and the like, may be configured, triggered, and otherwise provided by control messages and data messages and signaling communicated between various network elements in a wireless communication system.

Description

Technical Field

[0001] The present disclosure relates generally to wireless communication systems and methods, and more particularly to mechanisms for implementing an artificial intelligence framework for adaptively configuring an air communication interface of a wireless communication system. Background Art

[0002] In wireless communication systems, determining adaptive network configurations, particularly within air communication interfaces, may require lengthy measurement processes and / or significant amounts of computing power. These types of configurations may include, but are not limited to, beam management, channel state information (CSI) feedback compression and decompression, and wireless terminal positioning. The correlations between various network conditions and these adaptive configurations may be learned through artificial intelligence (AI) techniques and models. Therefore, it may be desirable to provide a mechanism for provisioning the lifecycles of various AI models and their application to assist in adaptively determining these network configurations. Summary of the invention

[0003] The present disclosure generally relates to wireless communication systems and methods, and more particularly to mechanisms for implementing an artificial intelligence framework for adaptively configuring an air communication interface of a wireless communication system. The AI ​​network configuration function can be provided as a service for AI-assisted network configuration (AI configuration service, AI configuration service, or AICS). Such AICS can be requested and configured through various messaging and signaling mechanisms. AI model lifecycle management can be configured, triggered, and otherwise provided through control messages and data messages and signaling transmitted between various network elements in a wireless communication system, including training, delivery, activation, inference, deactivation, switching, and performance evaluation, etc.

[0004] In some example embodiments, a method performed by at least one wireless network node is disclosed. The method may include: activating an artificial intelligence (AI) network configuration mode in response to receiving a request for an AICS (AI configuration service) from a wireless terminal device; determining a network-side AI model according to the request for the AICS; receiving a set of input data items for the network-side AI model from the wireless terminal device; determining a network configuration action (NCA) based on the inference of the network-side AI model according to the set of input data items; transmitting the NCA or an indication for the NCA to the wireless terminal device; receiving an AI feedback data item from the wireless terminal device, the AI ​​feedback data item being generated by the wireless terminal device according to the indication and based on the deployment result of the NCA; and performing at least one AI network configuration management task according to the AI ​​feedback data item.

[0005] In some other embodiments, a method performed by a wireless terminal device is disclosed. The method may include: transmitting a request for an AI configuration service (AICS) to at least one wireless network node; receiving a management configuration for the AICS; receiving an activation indication associated with a UE-side AI model; receiving a set of auxiliary information items associated with the UE-side AI model; in response to the activation indication, using the UE-side AI model to perform AI inference to generate an inference data item based on the set of auxiliary information items; transmitting the inference data item to the at least one wireless network node; receiving a network configuration behavior (NCA) determined by the at least one wireless network node based on the inference data item; and performing AI service management or assisting the at least one wireless network node in performing AI service management according to the management configuration for the AICS and the inference data item.

[0006] In some other embodiments, a wireless device including a processor and a memory is disclosed. The processor may be configured to read computer code from the memory to implement any of the above methods.

[0007] In some other embodiments, a computer program product is disclosed, the computer program product comprising a non-transitory computer-readable program medium on which a computer code is stored. When the computer code is executed by a processor, the processor may be prompted to implement any of the above methods.

[0008] Other aspects and alternatives to the above-described embodiments and their implementations are described in more detail in the following figures, detailed description, and claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 An example wireless communication network including a radio access network, a core network, and a data network is shown.

[0010] Figure 2 An example wireless access network is shown that includes a plurality of mobile stations or UEs and wireless access network nodes that communicate with each other via an over-the-air wireless communication interface.

[0011] Figure 3 Example functional blocks of a general AI platform for use in wireless communication networks are shown.

[0012] Figure 4 The data and logic flow for providing and managing offline trained AI models on the network side of a wireless communication network is shown.

[0013] Figure 5 The data and logic flow of an AI model for providing and managing online training or reinforcement training on the network side of a wireless communication network is shown.

[0014] Figure 6 The data and logic flow for providing and managing offline trained AI models on the terminal side of a wireless communication network is shown.

[0015] Figure 7 The data and logic flow of an AI model for providing and managing online training or reinforcement training on the terminal side of a wireless communication network is shown.

[0016] Figure 8 The data and logic flow for providing and managing offline trained AI models on the network side and terminal side of a wireless communication network is shown.

[0017] Fig. 9 The data and logic flow of an AI model for providing and managing online training or reinforcement training on the network side and terminal side of a wireless communication network is shown. DETAILED DESCRIPTION

[0018] The techniques and examples of the embodiments and / or examples described in the present disclosure can be used to facilitate adaptive and intelligent network configurations associated with, for example, air communication interfaces in wireless communication systems. The term "exemplary" is used to mean "an example of..." and, unless otherwise specified, does not imply an ideal or preferred example, embodiment, or embodiment. Section headings are used in the present disclosure to facilitate understanding of the disclosed embodiments and are not intended to limit the techniques disclosed in the sections to the corresponding sections. The disclosed embodiments may be further embodied in a variety of different forms, and therefore, the scope of the present disclosure or the claimed subject matter is intended to be interpreted as not being limited to any of the embodiments described below. Various embodiments may be embodied as methods, devices, components, systems, or non-transitory computer-readable media. Accordingly, the embodiments in the present disclosure may take the form of, for example, hardware, software, firmware, or any combination thereof.

[0019] The present disclosure relates to wireless communication systems and methods, and in particular to mechanisms for implementing an artificial intelligence framework for adaptively configuring an air communication interface of a wireless communication system. The AI ​​network configuration function can be provided as a service (AI configuration service, or AICS) for AI-assisted network configuration. Such AICS can be requested and configured through various messaging and signaling mechanisms. AI model lifecycle management can be configured, triggered, and otherwise provided through control messages and data messages and signaling transmitted between various network elements in a wireless communication system, including training, delivery, activation, inference, deactivation, switching, and performance evaluation.

[0020] Wireless Network Overview

[0021] Figure 1The example wireless communication network shown as 100 in FIG. 1 may include wireless terminal devices or user equipment (UE) 110, 111, and 112, a carrier network 102, various service applications 140, and other data networks 150. The carrier network 102 may include, for example, access networks 120 and 121 and a core network 130. The carrier network 110 may be configured to transmit voice, data, and other information (collectively referred to as data traffic) between the UEs 110, 111, 112, between the UEs and the service applications 140, or between the UEs and other data networks 150. The access networks 120 and 121 may be configured as various wireless access network nodes (WANN, alternatively referred to as base stations) to interact with the UEs on one side of a communication session and the core network 130 on the other side. The core network 130 may include various network nodes configured to control the communication session and perform network access management and traffic routing. The service applications 140 may be hosted by various application servers deployed outside the core network 130 but connected to the core network 130. Similarly, other data networks 150 may also be connected to the core network 130.

[0022] exist Figure 1 In the wireless communication network 100 of FIG. 1 , UEs may communicate with each other through a radio access network. For example, UEs 110 and 112 may be connected to the same access network 120 and communicate through the same access network 120. UEs may communicate with each other through an access network and a core network. For example, UE 110 may be connected to access network 120, and UE 111 may be connected to access network 121, and thus, UE 110 and UE 111 may communicate with each other through access networks 120 and 121 and core network 130. UEs may also communicate with service applications 140 and data network 150 through core network 130. In addition, as shown in 113, UEs may communicate directly with each other through sidelink communications.

[0023] Figure 2Also shown is an example system diagram of a wireless access network 120 including a WANN 202 serving UEs 110 and 112 via an air interface 204. The wireless transmission resources used for the air interface 204 include a combination of frequency resources, time resources, and / or space resources. Each of the UEs 110 and 112 may be a mobile terminal device or a fixed terminal device equipped with a mobile access unit (such as a SIM (Subscriber Identity Module) / USIM (Universal Subscriber Identity Module) module) for accessing the wireless communication network 100. Each of the UEs 110 and 112 may be implemented as a terminal device including, but not limited to, a mobile phone, a smart phone, a tablet computer, a laptop computer, an in-vehicle communication device, a roadside communication device, a sensor device, a smart appliance (such as a television, a refrigerator, and an oven), or other devices capable of wireless communication via a network. As Figure 2 As shown, each UE in the plurality of UEs (such as UE 112) may include a transceiver circuit 206 coupled to one or more antennas 208 to enable wireless communication with WANN 120 or with another UE (such as UE 110). Transceiver circuit 206 may also be coupled to processor 210, which may also be coupled to memory 212 or other storage device. Memory 212 may be transient or non-transitory and may store computer instructions or codes therein that, when read and executed by processor 210, cause processor 210 to implement the various methods described herein.

[0024] Similarly, WANN 120 may include a base station or other wireless network access point that can wirelessly communicate with one or more UEs and communicate with core network 130 via air interface 204. For example, WANN 120 may be implemented in, but not limited to, the following forms: 2G base station, 3G nodeB, LTE eNB, 4G LTE base station, 5GNR base station, 5G centralized unit base station, or 5G distributed unit base station. Each type of these WANNs may be configured to perform a corresponding set of wireless network functions. WANN 202 may include a transceiver circuit 214 coupled to one or more antennas 216 (antennas 216 may include various forms of antenna towers 218) to implement wireless communications with UEs 110 and 112. Transceiver circuit 214 may also be coupled to one or more processors 220, which may also be coupled to memory 222 or other storage devices. The memory 222 may be transient or non-transitory and may store therein instructions or codes that, when read and executed by the one or more processors 220 , cause the one or more processors 220 to implement various functions of the WANN 120 described herein.

[0025] Such as Figure 2 Data packets in the example wireless access network described in the figure can be transmitted as protocol data units (PDUs). The data therein can be packaged into PDUs wrapped with nested and / or layered protocol headers at various network layers. Once a connection (e.g., a radio link control (RRC) connection) is established between the sender and the receiver, the PDUs can be transmitted between the sending device or sender (the two terms are used interchangeably) and the receiving device or receiver (the two terms are also used interchangeably). Either the sending device or the receiving device can be a wireless terminal device, such as Figure 2 Devices 110 and 120; or wireless access network nodes, such as Figure 2 Node 202. Each device can be a sending device and a receiving device for two-way communication.

[0026] AI in Wireless Network Configuration

[0027] At the core of the general AI framework are various AI models. AI models typically contain a large number of model parameters that are determined through a training process in which correlations in the training dataset are learned and embedded into the trained model parameters. Therefore, the trained model parameters can be used to generate inference results from a set of input datasets that may not exist in the training dataset. AI models are particularly suitable for situations where there is little to no traceable deterministic or analytical derivation path between input data and output.

[0028] In wireless communication systems such as those described above, the determination of adaptive network configurations may rely on empirical characteristics and may also require a long measurement process and / or a large amount of computing power. These types of configurations may include, but are not limited to, air interface beam management, channel state information (CSI) feedback compression and decompression, and wireless terminal positioning. The correlation between various network conditions and these adaptive configurations can be learned through artificial intelligence (AI) technology. Therefore, the use of AI models to assist in network configuration may help reduce the amount of measurement and computing requirements, providing a more flexible network configuration. Therefore, it may be desirable to provide a mechanism for providing a life cycle and application of various AI models to assist in adaptively determining these network configurations.

[0029] For example, AI technology can be applied to beam management in air communication interfaces. In current implementations, beam management typically relies on comprehensive search beam scanning. In other words, the network (NW) can perform a full scan of the beam by sending a sufficient number of reference signals. The UE can be configured to monitor and measure each reference signal, and then report the measurement results to the NW for the NW to decide the best beam for the UE to switch to. However, this process is resource and power intensive. In order to accurately infer the best beam, few measurements (or fewer reference signals) may be required by embedding a trained AI model that learns the correlations between various network condition parameters. In some embodiments, the AI ​​model can help identify inferences about the best candidate beams using other network conditions, and then only scan and measure these candidate beams to select the beam used in the current communication. Additionally, since the beam configuration is closely related to the location of the UE, AI technology can also be used to infer or predict the UE trajectory or position, thereby indirectly helping to select the best beam.

[0030] For another example, AI technology can be applied to channel state information (CSI) feedback. Traditionally, CSI feedback can be implemented using a codebook known to the UE and NW. The UE can measure the CSI and obtain the measurement results, then map the measurement results to the closest vector of the codebook, and transmit the index of the vector to the NW to save air interface resource consumption. However, because the codebook is not unlimited or dynamically changes over time, there is always a mismatch, resulting in uncontrolled CSI feedback errors as the wireless environment changes. Therefore, AI can be applied to compression-decompression for CSI feedback. Specifically, the CSI report can be compressed by the UE-side AI model and decompressed by the corresponding NW-side AI model. Such AI models can be initially trained and continuously strengthened with the accumulation of time and network conditions.

[0031] As another example, AI technology can be applied to UE positioning. Traditional UE positioning methods depend on PRS (Positioning Reference Signal) or SRS (Sounding Reference Signal) (for example, DL (downlink) positioning reference signal and UL (uplink) sounding reference signal). Regardless of the alternative method, the LOS (Line-Of-Sight) beam is a key beam for identification in order to generate the most accurate position estimate through triangulation on the NW side. However, in most cases, it is difficult to identify the LOS beam from other non-line-of-sight (NLOS) beams, thereby providing inaccurate UE positioning. On the other hand, the trained AI model can identify various patterns and correlations in PRS and SRS to extract LOS information and provide more accurate UE positioning.

[0032] In some example embodiments disclosed herein, AI technology may be provided as a service including various configuration functions, each of which may be associated with one or more AI models, which may be trained offline or continuously online. Such AI functions may be requested by a terminal device. AI functions may be provided in an AI providing platform that supports several example aspects of AI functions, such as Figure 3 As shown and described below:

[0033] 1. AI model preparation may include at least one of the following parts:

[0034] a) Model training, including offline training, online training and reinforcement training.

[0035] i. Offline training: AI models can be pre-trained, or trained / validated / tested offline and stored on the NW side or UE side. For example, some AI models can be successfully trained offline.

[0036] And no reinforcement / enhancement is performed during the implementation of AI-based functions. In some embodiments, the term offline training on the UE side or NW side is used to indicate that the AI ​​model has been successfully trained and tested before the UE enters the RRC connected state, rather than being trained or reinforced during the RRC connected state.

[0037] ii. Online training. There may be two example cases of online training of AI models. In one example, the AI ​​model may be partially trained offline and then reinforced or continued training using an online data set. In another example, the AI ​​model is trained only using an online data set. In some example embodiments, online training on the UE side or the NW side means that the AI ​​model needs to be reinforced or trained after the UE enters the RRC connected state. In some example embodiments, due to its real-time nature, online training for the AI ​​model may be managed by an AI training timer to avoid the training process converging too long or failing to converge at all. For example, such an AI training timer may be started or restarted at the start of AI model training, and stopped when the training is successful, and when the timer expires, if the training has not converged sufficiently, the AI ​​model training may be considered to have failed.

[0038] b) AI model transmission, which involves how to deliver the trained model to the NW and / or

[0039] or UE.

[0040] c) One-side model vs. two-side model: For a specific AI function, it can be determined where the AI ​​model should reside for inference and / or training. The AI ​​model can reside on the NW side or the UE side for inference or training.

[0041] For some applications, AI models can reside on both sides for inference or training and collaboratively perform specific network functions. In some embodiments, for some AI models, the training location and inference location

[0042] The locations can be different. For example, an AI model can be trained on the network side but deployed on the UE side for inference.

[0043] d) Other information exchange between NW and UE for AI model training / inference. As described in other examples below, such information may include training data and other AI model auxiliary information.

[0044] Such information may be transmitted between the NW and the UE via one or more messages and signaling of various types.

[0045] 2. AI model implementation. Using the AI ​​model for inference may include at least one of the following aspects:

[0046] a) Input data generation, preparation, and transmission / delivery for AI model inference.

[0047] b) Transmission and processing of outputs inferred by AI models.

[0048] c) The location inferred by the AI ​​model, as mentioned above, on the NW side or the UE side.

[0049] 3. Lifecycle management. This can be achieved by, for example, activating and deactivating the overall or specific AI of multiple AI functions / models.

[0050] Functions / models to provide AI lifecycle management. Such lifecycle management can be based on one or more of the following schemes:

[0051] a) Time-based management. For example, an AI model-specific timer for an AI model can be configured

[0052] As part of the AI ​​lifecycle management configuration. For example, such a timer can be started when the AI ​​model is activated and when the UE / NW sends feedback information for AI model performance evaluation or when

[0053] When the UE / NW determines through performance evaluation that the AI ​​model is still feasible, the timer can be restarted, and when the AI ​​model is deactivated and / or switched to another AI model, the AI ​​model specific timer can be stopped.

[0054] The effect of an expiring model-specific timer may make the AI ​​model unavailable for inference unless

[0055] In some other embodiments, multiple AI functions can be controlled as a whole through a single

[0056] Timer management.

[0057] b) Event-based management. As described in more detail below, due to the occurrence of one or more events, an activated AI model will be deactivated and considered unavailable.

[0058] c) Management based on performance evaluation. The inference performance of AI models can be used to manage AI functions. For example, inference criteria can be predefined to evaluate the performance of AI models. If the performance evaluation indicates that a particular AI model is no longer suitable for the current physical / network environment, the AI ​​model will be considered infeasible.

[0059] and can be disabled or replaced.

[0060] In some embodiments, the AI ​​functionality may be provided as one or more configuration modes signaled in any suitable control message or any suitable control field in a control message. Various control message and signaling formats may be adapted to accommodate the implementation of the AI ​​functionality.

[0061] Example AI functions for configuration of air interface

[0062] The AI ​​model may be activated for determining various network configurations that may otherwise involve resource and power intensive processes for determination. While the following examples focus on AI-assisted determination of various aspects of air interface configurations, the underlying principles and architectures so described may be applicable to determination / selection of other network configurations in wireless network communication systems.

[0063] Example 1: AI-based beam management

[0064] Without AI assistance, traditional beam management and configuration involves the NW needing to scan all beams through various reference signals. The UE is correspondingly configured to perform measurements on all reference signals (RS) and then report the reference signal received power (RSRP) of all RS to the NW. The NW then decides the best beam and then adjusts / configures the UE's DL / UL beams.

[0065] AI-based beam management / configuration schemes can be implemented for spatial and / or temporal beam management / configuration. In AI-based spatial beam management, the NW may only need to scan part of the full beam, so the UE only needs to perform measurements for a reduced number of scanned beams. The RSRP of the reference signal associated with all beams including non-scanned beams can be derived from the RSRP of these partial beams using an AI model trained for beam inference. Such models can be enhanced to identify the correlation of the RSRP of various beams under various network conditions (which can also be part of the input to the AI ​​model for beam inference). The best beam can be selected based on the inference of the AI ​​model.

[0066] In AI-based time beam management applications, AI models can be trained and used to derive the beam pattern for the next time period based on the beam pattern for the current time period. Such AI models are trained to identify correlations between beam patterns between time instances.

[0067] Example 2: AI-based CSI feedback enhancement

[0068] In an example implementation, a two-sided AI approach may be applied. In other words, AI models are used on both the UE side and the NW side to generate inference results. Specifically, the UE-side AI model may be used to compress the CSI feedback to generate compressed CSI feedback, and the NW-side AI model may be used to decompress the compressed CSI feedback to generate decompressed CSI feedback. Such AI models may be inherently nonlinear and therefore support nonlinear compression and decompression, which is difficult to implement and lacking in traditional type II codebook-based CSI feedback schemes. Such nonlinear processes provide higher accuracy CSI feedback and significantly save resources.

[0069] As described in the various example embodiments below, the AI ​​network configuration functionality may be provided as a service for AI-assisted network configuration (AI configuration service, or AICS). Such AICS may be requested and configured through various messaging and signaling mechanisms. AI model lifecycle management may be configured, triggered, and otherwise provided through control and data messages and signaling transmitted between various network elements in a wireless communication system, including training, delivery, activation, inference, deactivation, switching, and performance evaluation, among others.

[0070] AICS classifies various example implementations according to options related to the location of the AI ​​model (on the NW side or the UE side, or on both the NW side and the UE side), the training process of the AI ​​model (pre-training / offline training, or online training or reinforcement training). These different options can be implemented as different AICS modes, in parallel with non-AI modes or in other relationships. These modes can be managed separately according to different types of network configuration functions, such as beam configuration and CSI feedback functions.

[0071] Case 1: NW side model, offline training

[0072] In this case, the AI ​​model can be trained offline on the NW side. These AI models can be developed by the NW vendor, or can be developed by other vendors but uploaded to the NW vendor before the AI ​​model is implemented, and the training process and model selection for AI-based functions may be transparent to any network specification. For model evaluation, it can also be up to the NW vendor or model developer to decide whether the model is accurate or feasible enough, which can also be transparent to the specification.

[0073] Figure 4 An example overall process of data and logic flow for requesting and providing information for AICS using an offline trained AI model is shown in FIG.

[0074] In some implementations, the NW side may also be configured with a separate function, entity, layer, or OAM (Operation, Administration, and Management function) to handle AICS provision and configuration. Such an entity may reside in Figure 1 in the core network shown in . Alternatively, such entity may be part of the access network. In some embodiments, such entity may reside in a base station (such as a gNB as a separate function). In some other embodiments, the AI ​​entity and the gNB function may be an integrated part of the base station. In such embodiments, the information exchange through the AI ​​interface described below will be irrelevant, and the operation of the AI ​​functional entity / layer will be considered part of the operation of the gNB. When the AI ​​entity and the gNB are implemented separately (functionally or physically), the communication interface (both data and control) between the AI ​​entity and the gNB may be referred to as the AI ​​interface, and the communication between the UE or wireless terminal device and the gNB or other base station may be referred to as the air interface, such as Figure 4 shown.

[0075] like Figure 4 As further shown, an example overall process in the air interface for Case 1 may include:

[0076] Step 0: The UE may send an indication to the NW requesting an AI-based function or service (or AICS). Alternatively, the NW determines that the AI-based function can be enabled.

[0077] • Step 1: If at least one AI model associated with or suitable for the requested AI-based function can be found, some operations are implemented by the NW to enable the AI-based function, otherwise, the process ends.

[0078] Step 2: The UE sends input data for generating model inference results to the NW.

[0079] Step 3: The NW performs model inference based on the received input data (either by processing the input data as is, or by additional processing through the AI ​​model), and then takes actions for the AI-based function based on the output of the model inference.

[0080] Step 4: UE sends feedback to NW to perform AI model life management or model operation. NW implements model operation based on the feedback information.

[0081] Step 5: NW implements model operations for AI-based functions.

[0082] Figure 4 An example overall process through the AI ​​interface may include:

[0083] Step 1: The gNB sends a request for AI-based functions to the AI ​​entity / layer.

[0084] Step 2: The AI ​​entity / layer selects an AI model for the requested AI-based functionality.

[0085] Step 3: If there is at least one AI model that can be selected and is suitable for the requested AI-based function

[0086] If yes, the AI ​​entity / layer sends an ACK (acknowledgement) to the gNB for the request. Otherwise,

[0087] The process ends.

[0088] Step 3: The gNB starts sending input data to the AI ​​functional layer / entity and obtains output data for AI-based functions from the AI ​​functional layer or entity.

[0089] Step 4: The gNB sends input data for model inference to the AI ​​entity / layer.

[0090] Step 5: The AI ​​entity / layer sends the output of the model inference to the gNB.

[0091] Step 6: The gNB sends feedback to the AI ​​entity / layer for AI model lifecycle management.

[0092] Step 7: The AI ​​entity / layer implements the model operations based on the feedback.

[0093] Step 8: The AI ​​entity / layer informs the gNB of the model operation to facilitate the operation of AI-based functions within the gNB air interface.

[0094] A further detailed description of the above steps implemented in the air interface is provided below.

[0095] For step 0 in the air interface:

[0096] The request for AI-based functionality (such as beam management) may be embedded / encapsulated or carried in at least one of the following messages / signaling / formats:

[0097] 1. Uplink (UL) media access layer (MAC) control element

[0098] element, CE), UL MAC CE;

[0099] 2. Uplink Control Information (UCI);

[0100] 3. Scheduling request (SR) / physical uplink control channel with format

[0101] (physical uplink control channel, PUCCH) signaling;

[0102] 4. UE Assistance information (UAI), for example, a form of uplink radio research control (RRC) signaling; or

[0103] 5.UE capability signaling.

[0104] In any of the alternatives for the request to AICS, the request may include at least one of the following information:

[0105] 1. AI-based function information: used to indicate the requested AI-based function from a function list (e.g., AI-based spatial beam management, AI-based temporal beam management, AI-based positioning, AI-based CSI feedback, etc.). Such AI-based functions may be represented by a predefined index or may be represented by a predefined AI function ID. In an implementation of AI-based beam management, the UE may include an AI-based function ID corresponding to the AI-based beam management in the request message.

[0106] 2. Auxiliary information for AI model selection / training: used to indicate what types of information the UE can provide for model selection or model training. For example, for AI-based beam management, such model selection or training information may include UE location information, speed information, direction information, or pointing information of the UE or wireless terminal device.

[0107] (orientation) information, etc.

[0108] For step 1 in the air interface:

[0109] In response to the AI-based function request from the UE, the NW (e.g., a base station) may then perform a set of operations for enabling the requested AI function. For example, but not limited to, a set of operations for enabling AI-based beam management (as an example request) may be performed. For the requested AI-based beam management, such NW operations may include one or more operations, including but not limited to the following operations:

[0110] 1. Change the CSI measurement configuration for non-AI-based beam management to another CSI measurement configuration for AI-based beam management. For example:

[0111] a) Activation of semi-periodic CSI reporting associated with a specific set of CSI resource sets for AI-based beam management through a DL MAC CE;

[0112] b) deactivating, through a DL MAC CE, semi-periodic CSI reporting associated with a CSI resource set for non-AI based beam management;

[0113] c) enabling or adding periodic CSI reporting configuration associated with a specific CSI resource set for AI-based beam management through RRC reconfiguration;

[0114] d) disabling or removing the periodic CSI reporting configuration associated with the CSI resource set for non-AI based beam management through RRC reconfiguration;

[0115] e) for a specific CSI measurement configuration (e.g., including a CSI resource set and a corresponding CSI report), changing from a CSI resource set for non-AI based beam management to a CSI resource set for AI based beam management, which may be achieved by at least one non-limiting operation in the following non-limiting set of operations:

[0116] i. For one specific CSI reporting configuration for non-AI based beam management, the CSI resource(s) associated with the CSI reporting configuration may be dynamically changed by using DL MAC CE.

[0117] ii. For a specific CSI reporting configuration for non-AI based beam management, the CSI reporting period value can be dynamically changed by the DL MAC CE.

[0118] iii. For a specific CSI resource set configuration for non-AI based beam management, the CSI resource period can be dynamically changed by the DL MAC CE.

[0119] iv. For a specific CSI resource set configuration for AI-based beam management, the CSI resource period can be dynamically changed by the DL MAC CE.

[0120] For step 2 in the air interface:

[0121] The input data of the AI ​​model for inference can be transmitted from the UE to the NW. Specifically, in the example of AI-based beam management, such input data can be transmitted through any one or more of the following items: UCI on PUCCH; UCI on PUSCH (Physical Uplink Shared Channel); or UL MAC CE, etc.

[0122] For such AI-based beam management functions, any of the following datasets can be transmitted or signaled as input data to the beam management AI model on the NW side for inference via any of the above channels or messages:

[0123] 1. CSI measurement or SSB (synchronization signal block) measurement results;

[0124] 2. With CSI-RS (channel state information reference signal)

[0125] or time information associated with SSB measurement results;

[0126] 3.UE trajectory information;

[0127] 4. Beam information, such as CSI-RS ID and SSB ID;

[0128] 5.UE location information;

[0129] 6.UE pointing information;

[0130] 7.UE speed information.

[0131] For step 3 in the air interface:

[0132] Once the NW has performed inferences based on the above input data, actions may be determined based on the inferences. Such actions (such as network configuration actions), if necessary, may be performed by the UE or notified to the UE. In the example of AI-based beam management functionality, such actions based on inferences from the beam management AI model may be any one or more of the following non-limiting list of actions:

[0133] 1. Signaling used to switch UL / DL beams at a time.

[0134] 2. If there is only one entry in the pool, reconfigure the current TCI state pool.

[0135] 3. Signaling for switching UL / DL beams sequentially for a beam set within a preconfigured time period.

[0136] The above operation may be indicated to the UE through any of the following: a downlink (DL) MAC CE, a downlink control indication (DCI) message, and the like.

[0137] For step 4 in the air interface:

[0138] Feedback information for AI-based functions may be generated by the UE and transmitted to the NW.

[0139] For the case of AI-based beam management functionality, such feedback information to the NW may include one of the following:

[0140] 1. Inferred RSRP value of CSI-RS / SSB for beam management.

[0141] 2. The actual RSRP value of CSI-RS / SSB used for beam management.

[0142] 3. The inferred best beam pattern for the time period.

[0143] 4. The actual best beam pattern within the time period.

[0144] In some embodiments, the feedback information can be used by the NW to manage the life cycle of the AI ​​model or AI function. The life cycle management of the AI ​​function or AI model can be based on the feedback information and various other solutions.

[0145] In some example embodiments, the lifecycle management of the AI ​​function or AI model may be time-based. For example, an AI function management timer or an AI model management timer may be introduced and configured. Such timers may be maintained on the NW side (e.g., gNB or AI function layer / entity / OAM) that provides the AI ​​function or AI model. Such timers may operate in the following manner:

[0146] 1. When an indication to activate the AI ​​model or AI function is received (e.g. on the gNB side), or when the AI ​​function layer / entity / OAM of the NW sends an ACK to the gNB in ​​response to the AI ​​function request, or the NW side sends an ACK to the UE

[0147] When an ACK is sent, a timer is started. When the performance evaluation result showing the AI ​​model performance is better than a predefined threshold, the timer can be restarted.

[0148] 2. When the timer expires: For example, an indication may be sent by the gNB to the AI ​​functional layer / entity to deactivate the applied AI model or AI function. For another example, an indication may be sent from the AI ​​functional layer / entity to the gNB to disable the AI-based beam management function / configuration.

[0149] 3. When the AI ​​functional entity / layer receives an indication from the gNB to deactivate the AI ​​model or AI function, or when the gNB receives an indication from the AI ​​functional entity / layer to disable the AI-based beam management model or function (e.g. beam management), the timer is stopped.

[0150] In some other example implementations, the lifecycle management of AI functions or AI models may be event-based. In other words, lifecycle management behavior may be triggered by a preconfigured or scheduled occurrence of one or more events.

[0151] For example, these triggering events may include, but are not limited to, any one or more of the following:

[0152] 1. For AI-based beam management functionality, when beam failure is detected for (one or more) BWPs / serving cells for which AI-based beam management is being performed.

[0153] 2. For the AI-based beam management function, when a radio link failure is detected for (one or more) BWPs / serving cells for which AI-based beam management is being performed.

[0154] 3. Deactivation of BWP(s) / serving cell(s) based on AI functionality is ongoing.

[0155] 4. Reception of HO (Handover) command.

[0156] 5. MAC reset occurs.

[0157] When an indication of the above-mentioned triggering event is detected and / or received, the NW may perform operations to disable AI-based functions or models (such as AI-based beam management functions or models).

[0158] In yet other example embodiments, lifecycle management of an AI function or AI model may be an event-based performance evaluation of the AI ​​model or function. Performance evaluation of the AI ​​model may be implemented at the gNB or AI layer / entity by comparing the inferred output of the AI ​​model with the actual value corresponding to the inference. For example, for performance evaluation of an AI-based beam management function, RSRP (reference signal received power) values ​​inferred by the AI ​​model for one or more beams of the AI ​​model may be compared with actual measurement results of the same beams from traditional measurements to generate a performance evaluation.

[0159] Performance evaluation can be mainly used to measure the performance of the currently applied AI model. If the performance evaluation results show that the current AI model is not feasible / ideal enough in the current network environment, some AI model lifecycle management operations can be performed by the NW. For example, based on the performance evaluation results, the current AI model can be deactivated, the current AI model can be switched to a different AI model; the current AI model can be retrained, the current AI model can be reinforced, or the current AI model can be fine-tuned, etc.

[0160] In some embodiments, for the AI-based beam management function, in addition to the CSI resource / report configuration for AI-based beam management, the NW may additionally configure a CSI / SSB resource set and associated CSI report for performance evaluation. And in some embodiments, the period of the CSI resource / report may be longer than the period of the CSI resource / report configured for AI-based beam management. In some embodiments, in addition to the CSI resource / report configuration for AI-based management, the CSI resource / report configuration for non-AI-based beam management may also be adjusted for performance evaluation, for example, using a DL MAC CE and / or RRC reconfiguration to adjust the period of CSI reporting and / or CSI measurement for non-AI-based beam management.

[0161] For step 5 in the air interface

[0162] In some implementations, the lifecycle management operations of the AI-based functions on the UE side may include but are not limited to the following operations: disabling the AI-based functions; retaining the AI-based functions but adjusting them with some RRC reconfiguration, etc. For step 1 in the AI ​​interface:

[0163] Similar to step 0 described above for the air interface, the AI ​​function request sent from the gNB to the AI ​​function / entity / layer may include but is not limited to at least one of the following information:

[0164] 1. AI-based function information: used to indicate the requested AI-based function from the function set group by index or function ID. For example, in this case, the AI-based function identifier '0' can represent AI-based spatial beam management, the AI-based function identifier '1' can represent AI-based time beam management function, and so on.

[0165] 2. Auxiliary information for AI model selection / training: used to indicate what types of information the UE can provide for model selection or model training. For example, such model selection or training information may include UE location information, speed information, direction information or pointing information of the UE or wireless terminal device.

[0166] For step 3 in the AI ​​interface:

[0167] In response to the AI-based function request, the ACK from the AI ​​function / entity / layer to the gNB may include at least one of the following information:

[0168] 1. AI model information: An identifier used to indicate the AI ​​model used for the requested AI-based function.

[0169] For example, there may be more than one offline AI model for AI-based beam management functionality, and one of these models may be associated with a model ID, where the model ID is used to represent the AI ​​model for AI-based spatial beam management, while another model ID may represent another model for AI-based temporal beam management.

[0170] 2. Input data type information for AI models: used to indicate the input data type. For example, different AI

[0171] Models can take various input data types. For example, some AI models may require UE location information, while some other AI models may require UE pointing information.

[0172] 3. Model-inferred output data: In the case of AI-based spatial beam management, the model-inferred output data may be the top K beam information and associated RSRP values, where K is equal to or greater than 1. In another embodiment, the model-inferred output data may be the top K beams and associated RSRP values ​​for a time period, where the top K beams and associated RSRP values ​​exist for each time point in the time period, and the granularity of the time points in the time period is configurable. In some embodiments, K may be configurable.

[0173] 4. Performance evaluation related information, such as request information for performance evaluation of the selected AI model. Another example includes configuration of performance evaluation information. In the case of AI-based beam management, performance evaluation related information may include CSI / SSB RS resources and / or CSI report configuration.

[0174] For step 4 in the AI ​​interface:

[0175] The input data from the gNB to the AI ​​functional entity / layer for model inference is similar to the input data described in step 2 of the air interface above.

[0176] For step 5 in the AI ​​interface:

[0177] Once the inference output is generated by the selected AI model, the inference output is provided from the AI ​​functional entity / layer to the gNB. In some embodiments, the inference data may include at least one of the following information:

[0178] 1: Top K best candidate beams with RSPR for AI-based spatial beam management, K is equal to or greater than '1', and in some embodiments, K may be configurable.

[0179] 2: Beam pattern with RSRP for AI-based time beam management. In some embodiments, the beam pattern may be a set of candidate UL / DL beams, each beam may be associated with time information. In some embodiments, a time point may have more than one candidate beam.

[0180] For step 6 in the AI ​​interface:

[0181] For the AI-based beam management function, the feedback information may include at least one of the following:

[0182] 1. Inferred RSRP value of CSI-RS / SSB for beam management;

[0183] 2. The actual RSRP value of the CSI-RS / SSB used for beam management;

[0184] 3. The inferred optimal beam pattern for the time period;

[0185] 4. The actual optimal beam pattern for the time period.

[0186] For step 7 in the AI ​​interface:

[0187] The AI ​​model operations based on the feedback received in step 6 may include at least one of the following non-limiting list of operations:

[0188] 1. AI model switching: deactivate the applied AI model and activate another AI model;

[0189] 2. AI model deactivation: Deactivate the currently applied AI model.

[0190] Case 2: NW-side AI model, online training

[0191] In this case, the AI ​​model can be trained online on the NW side. These AI models can be built by the NW vendor, and the training process for the AI ​​model can be performed online using the current training data set. Figure 5An example overall process in terms of data and logical flow of information for this case is shown.

[0192] Similar to case 1, the NW side may also be configured with a separate function, entity, layer or OAM (operation, maintenance and management function) to handle AICS provision and configuration. Such an entity may reside in Figure 1 in the core network shown in . Alternatively, such entity may be part of the access network. In some embodiments, such entity may reside in a base station (such as a gNB as a separate function). In some other embodiments, the AI ​​entity and the gNB function may be an integrated part of the base station. In such embodiments, the information exchanged over the AI ​​interface will be irrelevant, and the operation of the AI ​​functional entity / layer will be considered as the operation of the gNB. When the AI ​​entity and the gNB are implemented separately (functionally or physically), the communication interface (both data and control) between the AI ​​entity and the gNB may be referred to as the AI ​​interface, and the communication between the UE or wireless terminal device and the gNB or other base station may be referred to as the air interface, such as Figure 5 shown.

[0193] like Figure 5 As further shown, an example overall process in the air interface for Case 2 may include:

[0194] Step 0: The UE may send an indication to the NW requesting an AI-based function or service. Alternatively, the NW determines that the AI-based function can be enabled.

[0195] Step 1: NW sends configuration for UE to report input data for model training.

[0196] Step 2: The UE reports the input data for model training to the NW for model training.

[0197] Step 3: NW can select an AI model for training / validation / testing through input data, or select an under-trained AI model for reinforcement training through input data. If the AI ​​model is successfully trained, NW

[0198] Operations for enabling AI-based functions can be implemented. At the same time, NW can initiate lifecycle management of AI models or AI functions.

[0199] Step 4: UE sends input data for model inference to gNB.

[0200] Step 5: gNB takes actions based on the output inferred by the model.

[0201] Step 6: UE feeds back information to gNB for AI model lifecycle management.

[0202] Step 7: Once model life management shows that the AI ​​model is no longer sufficiently viable, NW implements the

[0203] Operation of AI functions.

[0204] Figure 5 An example overall process through the AI ​​interface may include:

[0205] Step 0: The gNB sends a request to the AI ​​entity / layer to enable AI-based functions.

[0206] Step 1: When an AI-based function request is received and if the AI ​​entity / layer has an available AI model for the requested AI-based function, the AI ​​entity / layer may respond to the AI-based function request to

[0207] gNB sends ACK. Otherwise, the AI ​​entity / layer can send NACK (Negative ACK) to the gNB.

[0208] Acknowledgment, negative acknowledgment), and then the process ends.

[0209] Step 2: The gNB sends input data for AI model training to the AI ​​entity / layer.

[0210] Step 3: The AI ​​entity / layer starts training / enhancing / validating / testing the AI ​​model. If the AI ​​model training is successful, proceed to step 4. Otherwise, the AI ​​entity / layer sends an indication of training failure to the gNB, and the process ends.

[0211] Step 4: The AI ​​entity / layer sends an indication that the AI ​​model has been trained successfully.

[0212] Step 5: The gNB sends input data for model inference to the AI ​​entity / layer.

[0213] Step 6: The AI ​​entity / layer performs inference using the input data received from the gNB via the activated AI model.

[0214] Step 7: The AI ​​entity / layer sends the inference output to the gNB for AI-based beam management.

[0215] Step 8: The gNB sends feedback information to the AI ​​functional entity / layer for AI model performance evaluation.

[0216] Step 9: The AI ​​entity / layer determines the AI ​​model operation (e.g., activation / deactivation, handover, retraining, fine-tuning) upon receiving an indication of AI model failure from the gNB.

[0217] Step 10: The AI ​​entity / layer sends an indication of the AI ​​model operation to the gNB to facilitate the operation of the AI-based functions of the gNB.

[0218] A further detailed description of the above steps implemented in the air interface is provided below.

[0219] For step 0 in the air interface:

[0220] The request for AI-based functionality (such as beam management) may be embedded / encapsulated or carried in at least one of the following messages / signaling / formats:

[0221] 1. Uplink (UL) media access layer (MAC) control element (CE), UL MAC CE;

[0222] 2. Uplink control information (UCI);

[0223] 3. Scheduling Request (SR) / Physical Uplink Control Channel (PUCCH) signaling with format;

[0224] 4. UE Assistance Information (UAI), e.g., a form of uplink Radio Research Control (RRC) information

[0225] order; or

[0226] 5.UE capability signaling.

[0227] In any of the alternatives for the request to AICS, the request may include at least one of the following information:

[0228] 1. AI-based function information: used to indicate the requested AI-based function from a function list (e.g., AI-based beam management, AI-based spatial beam management, AI-based temporal beam management, AI-based positioning, AI-based CSI feedback, etc.). Such AI-based functions may be represented by a predefined index or may be represented by a predefined AI function ID.

[0229] 2. Auxiliary information for AI model selection / training: used to indicate what types of information the UE can provide for model selection or model training. For example, such model selection or training information may include UE location information, speed information, direction information, and pointing information.

[0230] In step 1 of the air interface:

[0231] The NW can configure the following settings for the UE:

[0232] 1. CSI resources for beam management configuration. In some implementations, the CSI resources are used for non-AI based beam management.

[0233] 2. CSI report for beam management configuration. In some embodiments, the CSI report is used for non-AI based

[0234] beam management.

[0235] For step 2 in the air interface:

[0236] In step 2, after configuring the CSI measurement configuration in step 1, the UE reports the CSI measurement results for beam management and some other auxiliary information to the NW as model training input.

[0237] For step 3 in the air interface:

[0238] To enable AI-based beam management, as an example, the operations from the NW side may include at least one of the operations in the list of step 1 in the air interface of case 1 above.

[0239] For step 4 in the air interface:

[0240] The input data for AI model inference for beam management may be sent through, for example, UCI on PUCCH, UCI on PUSCH, or UL MAC CE.

[0241] For such AI-based beam management functions, any of the following datasets can be sent or signaled as input data to the beam management AI model on the NW side for inference through the above channels or messages:

[0242] 1. CSI measurement or SSB (synchronization signal block) measurement results;

[0243] 2. Time information associated with CSI-RS or SSB measurement results;

[0244] 3.UE trajectory information;

[0245] 4. Beam information, used to indicate the beam with CSI-RS ID or SSB ID;

[0246] 5.UE location information;

[0247] 6.UE pointing information;

[0248] 7.UE speed information.

[0249] For step 5 in the air interface: similar to step 3 in the air interface in case 1.

[0250] In steps 6 and 7 of the air interface:

[0251] These steps of Case 2 are similar to Step 4 and Step 5 in the air interface in Case 1.

[0252] In some example embodiments, the lifecycle management of the AI ​​function or AI model may be time-based. For example, an AI function management timer or an AI model management timer may be introduced and configured. Such timers may be maintained on the NW side (e.g., gNB or AI function layer / entity / OAM) that provides the AI ​​function or AI model. Such timers may operate in the following manner:

[0253] 1. When an indication that an AI model or AI function is successfully trained is received (e.g., on the gNB side), or when an indication of successful training is sent to the gNB (e.g., on the AI ​​function layer / entity side), a timer is started.

[0254] When the AI ​​model is successfully fine-tuned, the timer can be restarted. When the performance evaluation result showing that the AI ​​model performance is better than a predefined threshold, the timer can be restarted.

[0255] 2. When the timer expires: If the timer is located on the gNB side, the gNB may send, for example, an indication to the AI ​​functional layer / entity to deactivate the applied AI model or AI function.

[0256] An indication may be sent from the AI ​​functional layer / entity to the gNB to disable the AI ​​based beam management function / configuration.

[0257] 3. When the AI ​​function entity / layer receives an indication from the gNB to deactivate or retrain the AI ​​model or AI function, or when the AI ​​function entity / layer determines that the AI ​​model or function is to be deactivated or retrained, or when the gNB receives an indication from the AI ​​function entity / layer to disable or retrain the AI-based beam management model or function (e.g., beam management), or when the gNB determines that the AI ​​model should be deactivated based on some events (such as a received beam failure recovery request, SCell (Secondary cell) / BWP activation / deactivation, handover preparation, MAC reset, etc.), the timer is stopped.

[0258] In some other example implementations, the lifecycle management of AI functions or AI models may be event-based. In other words, lifecycle management behavior may be triggered by a preconfigured or scheduled occurrence of one or more events.

[0259] For example, these triggering events may include, but are not limited to, any one or more of the following:

[0260] 1. For AI-based beam management functionality, when beam failure is detected for (one or more) BWPs / serving cells for which AI-based beam management is being performed.

[0261] 2. For the AI-based beam management function, when a radio link failure is detected for (one or more) BWPs / serving cells for which AI-based beam management is being performed.

[0262] 3. Deactivation of BWP(s) / serving cell(s) based on AI functionality is ongoing.

[0263] 4. Reception of HO command.

[0264] 5. MAC reset occurs.

[0265] Upon detecting the above triggering event, the NW may perform operations to disable AI-based functions or models (such as AI-based beam management functions or models).

[0266] In yet other example embodiments, lifecycle management of an AI function or AI model may be an event-based performance evaluation of the AI ​​model or function. Performance evaluation of the AI ​​model may be implemented at the gNB or AI layer / entity by comparing the inference of the AI ​​model and the actual measurement corresponding to the inference. For example, for performance evaluation of an AI-based beam management function, RSRP (reference signal received power) values ​​inferred by the AI ​​model for one or more beams of the AI ​​model may be compared with actual measurement results of the same beams from traditional measurements to generate a performance evaluation.

[0267] Performance evaluation can be mainly used to measure the performance of the currently applied AI model. If the performance evaluation results show that the current AI model is not feasible / ideal enough in the current network environment, some AI model lifecycle management operations can be performed by the NW. For example, based on the performance evaluation results, the current AI model can be deactivated, the current AI model can be switched to a different AI model; the current AI model can be retrained, the current AI model can be reinforced, or the current AI model can be fine-tuned, etc.

[0268] In some embodiments, for AI-based beam management functionality, in addition to the CSI resource / report configuration for AI-based beam management, the NW may additionally configure a CSI / SSB resource set and associated CSI report for performance evaluation. And in some embodiments, the period of the CSI resource / report may be longer than the period of the CSI resource / report configured for AI-based beam management. In some embodiments, in addition to the CSI resource / report configuration for AI-based management, the CSI resource / report configuration for non-AI-based management may also be adjusted for performance evaluation, for example, using a DL MAC CE and / or RRC reconfiguration to adjust the period of CSI reporting and / or CSI measurement for non-AI-based functionality.

[0269] In some embodiments, lifecycle management operations of AI-based functions may include but are not limited to the following operations: disabling AI-based functions; retaining AI-based functions but adjusting some RRC configurations to adapt to new AI models, for AI model reinforcement training or fine-tuning, etc.

[0270] In step 0 in the AI ​​interface:

[0271] The AI ​​function request includes at least one of the following information:

[0272] 1. AI-based function information: used to indicate an AI-based function from a function set group. For example, the AI-based function identifier '0' is used to indicate AI-based spatial beam management. The AI-based function identifier

[0273] '1' is used to indicate AI-based time beam management functionality, and so on.

[0274] 2. Auxiliary information for AI model selection / training: used to indicate what types of information the UE can provide for model selection or model training. For example, such model selection or training information may include UE location information, speed information, direction information, and pointing information.

[0275] 3. AI model information: used to indicate what kind of information the UE / gNB can provide for AI model training / inference.

[0276] In step 1 in the AI ​​interface:

[0277] The ACK to the request may include at least one of the following information:

[0278] AI model information: an identifier of an AI model, indicating an AI model in an AI model set.

[0279] Input data type for training: In the case of beam management, the input data type for training may include the information set required for model training. For an information set, at least one of the following information should be included:

[0280] a) CSI-RS / SSB measurement results for non-AI based beam management.

[0281] b) Time information associated with CSI-RS / SSB measurement results.

[0282] c) Location information associated with CSI-RS / SSB measurement results.

[0283] d) Speed ​​information associated with CSI-RS / SSB measurement results.

[0284] e) Pointing information associated with CSI-RS / SSB measurement results.

[0285] Input data type for inference: In the case of beam management, the input data type for training can include the information set required for model training. For an information set, at least one of the following information should be included:

[0286] a) CSI-RS / SSB measurement results for AI-based beam management.

[0287] b) Time information associated with CSI-RS / SSB measurement results.

[0288] c) Location information associated with CSI-RS / SSB measurement results.

[0289] d) Speed ​​information associated with CSI-RS / SSB measurement results.

[0290] e) Pointing information associated with CSI-RS / SSB measurement results.

[0291] Model-inferred output data: In the case of AI-based spatial beam management, the model-inferred output data may be the top K beam information and associated RSRP values, where K is equal to or greater than 1. In another embodiment, the model-inferred output data may be the top K beams and associated RSRP values ​​for a time period, where the top K beams and associated RSRP values ​​exist for each time point in the time period, and the granularity of the time points in the time period is configurable. In some embodiments, K may be configurable.

[0292] Performance evaluation related information, such as request information for performance evaluation of the selected AI model. Another example includes configuration of performance evaluation information. In the case of AI-based beam management, performance evaluation related information may include CSI / SSB RS resources and / or CSI report configuration.

[0293] In step 2 in the AI ​​interface:

[0294] In order to train the beam management AI model, the input data for online training may include, for example, at least one of the following information:

[0295] 1. CSI-RS / SSB measurement results for beam management.

[0296] 2. Time information associated with CSI-RS / SSB measurement results.

[0297] 3. Location information associated with CSI-RS / SSB measurement results.

[0298] 4. Speed ​​information associated with CSI-RS / SSB measurements.

[0299] 5. Pointing information associated with CSI-RS / SSB measurement results.

[0300] In step 3 in the AI ​​interface:

[0301] For model training, in order to terminate endless model training, the AI ​​entity / layer can have a timer to train the model, and the timer can be associated with at least one of the following operations:

[0302] 1. Start: Model training begins. In one embodiment, the timer may be started at the end of the ACK message to the request, or may be started when the first input data for online training is successfully received.

[0303] 2. Stop: Model training has terminated successfully.

[0304] 3. Expiration: Considering that the current model cannot converge, trigger the deactivation of the AI ​​model and send a message to the gNB and / or AI entity /

[0305] The layer sends an indication that model training failed. The process ends.

[0306] In steps 5 and 6 in the AI ​​interface:

[0307] For the example AI-based beam management model, the inferred input data includes at least one of the following information:

[0308] 1. CSI-RS / SSB measurement results for AI-based beam management.

[0309] 2. Time information associated with CSI-RS / SSB measurement results.

[0310] 3. Location information associated with CSI-RS / SSB measurement results.

[0311] 4. Speed ​​information associated with CSI-RS / SSB measurements.

[0312] 5. Pointing information associated with CSI-RS / SSB measurement results.

[0313] The input data can then be processed by the trained AI model to generate inference results.

[0314] In step 7 in the AI ​​interface:

[0315] For the example AI-based beam management model, the inferred data may be at least one of:

[0316] 1. The top K best candidate beams with RSRP for AI-based spatial beam management, where K

[0317] Equal to or greater than '1', in some embodiments, K may be configurable.

[0318] 2. A beam pattern with RSRP values ​​in a time period for AI-based time beam management. In some embodiments, the beam pattern may be a set of candidate UL / DL beams, and each beam may be associated with time information in a time period. In some embodiments, at one point in time, there is more than one candidate beam.

[0319] In step 8 in the AI ​​interface:

[0320] For details, please refer to the lifecycle management section in case 1 above.

[0321] In step 9 in the AI ​​interface:

[0322] Model operations may include at least one of the following operations:

[0323] Deactivate the model

[0324] Retrain the model

[0325] Switch the model to another model

[0326] Fine-tune the model

[0327] Case 3: UE-side model, offline training

[0328] In this case, the AI ​​models can be trained offline on the UE side. These AI models can be developed by the UE or other vendors. Figure 6 An example overall process in terms of data and logical flow of information for this case is shown in FIG.

[0329] like Figure 6 As further shown, an example overall process for Case 3 may include:

[0330] Step 0: The UE sends a request to the NW to enable AI-based functions.

[0331] Step 1: The NW configures the AI-based function to the UE via RRC signaling.

[0332] • Step 2: The NW sends signaling to activate the AI ​​model for the AI-based function.

[0333] Step 3: NW sends information for UE to perform model inference.

[0334] Step 4: The UE obtains input for model inference according to the information in step 3, and then obtains inference results from the AI ​​model.

[0335] Step 5: UE sends the inference result to NW for action.

[0336] Step 6: The NW takes actions based on the received inference results.

[0337] If the model performance evaluation of the AI ​​model is performed on the UE side, then:

[0338] Step 7a: NW and UE start operations for AI model lifecycle management.

[0339] Step 8a: The UE processes the performance evaluation.

[0340] Step 9a: The UE sends an indication to the NW for notifying the operation of the applied AI model according to the evaluation result.

[0342] If the performance evaluation of the AI ​​model is performed on the NW side, then:

[0343] Step 7b: NW and UE start operations for AI model lifecycle management.

[0344] Step 8b: The NW processes the performance evaluation according to the information received from the UE.

[0345] Step 9b: The NW sends an indication to the UE indicating the operation of the applied AI model according to the evaluation result.

[0347] In step 0:

[0348] The request message may be through at least one of the following:

[0349] 1.UCI;

[0350] 2.UE auxiliary information;

[0351] 3.UL MAC CE;

[0352] 4.UE capability signaling.

[0353] For the above message types, the following information may be included:

[0354] 1. AI-based function information, such as an AI-based function identifier, for indicating what the AI-based function is for requesting, in which case the AI-based function may be requested AI-based beam management.

[0355] 2. AI model information, such as:

[0356] a) AI model identifier;

[0357] b) Processing time for AI model inference;

[0358] c) The value of the time period T of the AI ​​model’s life cycle;

[0359] d) Input data types used for model inference and / or training;

[0360] e) Output data type used for model inference;

[0361] f) Recommendations for RRC configuration of input data types, such as the periodicity of CSI resources, the periodicity of CSI reports, etc.

[0362] In step 1:

[0363] The RRC configuration for AI-based beam management may include at least one of the following:

[0364] 1. RRC reconfiguration for CSI measurement and reporting for non-AI based beam management.

[0365] 2. RRC configuration for CSI measurement and reporting for AI-based beam management.

[0366] In step 2:

[0367] Activation of the AI ​​model for AI-based beam management can be done by:

[0368] 1.DL MAC CE. In the DL MAC CE, the AI-based function ID and AI model ID should be included.

[0369] 2. RRC message, used to activate the AI ​​model through an enable flag, which can be incorporated into step 1.

[0370] In step 3:

[0371] The information used by the UE to perform model inference is the CSI-RS / SSB reference signal for AI-based beam management.

[0372] In step 4:

[0373] The UE makes an inference based on the measurement result of the CSI-RS / SSB reference signal in step 3.

[0374] In steps 5 and 6:

[0375] The output of model inference can be at least one of the following information:

[0376] 1. The top K best beams and / or associated RSRP values, K is configurable and can be equal to and greater than 1.

[0378] 2. All beam information and / or associated RSRP values.

[0379] 3. A series of top K beams and / or associated RSRP values ​​for the time period, where K is configurable and may be equal to and greater than 1.

[0380] a) The time period may be indicated by M consecutive CSI resources and / or CSI reporting periods, where M may be configurable and may be equal to or greater than 1.

[0381] In some embodiments, if K>1, then the behavior on the inferred data includes the following:

[0382] 1. Send a DL MAC to the UE to dynamically adjust the UL / DL beam based on the received inference information

[0383] CE.

[0384] 2. Based on the received inference information, send a DL MAC CE to the UE to dynamically adjust a series of UL / DL beams within a time period.

[0385] a) The time period may be indicated by M consecutive CSI resources and / or CSI reporting periods, where M may be configurable and may be equal to or greater than 1.

[0386] In some embodiments, if K=1, the behavior on the inferred data includes the following, and the behavior on the inferred data is used to adjust the UL / DL beam at the NW side.

[0387] Step 7a, Step 7b, Step 8a, Step 8b for lifecycle management:

[0388] For timer-based lifecycle management, a timer is configured for the UE for lifecycle management of the AI ​​model for AI-based beam management. The timer can be started / restarted on at least one of the following events:

[0389] 1. Receive activation signaling for the AI ​​model.

[0390] 2. Transmission of feedback information for performance evaluation of AI models.

[0391] 3. An indication of a performance evaluation is received, indicating or determining that the performance evaluation is that the current AI model is still suitable. In some embodiments, whether the AI ​​model is suitable is determined based on the difference between the inferred average / best RSRP value of one or more beams and the actual average / best RSRP value of the same beam. If the difference is less than a preconfigured / defined threshold, the AI ​​model can be considered suitable. If the difference is equal to or greater than a preconfigured threshold, the AI ​​model can be considered unsuitable.

[0392] In some embodiments, the timer should be stopped upon at least one of the following events: deactivation signaling for the AI ​​model is received, and / or the AI ​​model is deemed to be deactivated.

[0393] In some implementations, when the timer expires, the AI ​​model is considered deactivated.

[0394] In some embodiments, AI model lifecycle management may be event-based. For example, based on at least one of the following events, the applied AI model for the AI-based function will be deemed infeasible.

[0395] 1. Beam failure is detected for (one or more) BWPs / serving cells to which the AI ​​model is applied.

[0396] 2. Radio link failure is detected for the BWP(s) / serving cell(s) to which the AI ​​model is applied.

[0397] 3. Deactivation of BWP(s) / serving cell(s) applying the AI ​​model.

[0398] 4. Reception of HO command.

[0399] 5.MAC reset.

[0400] For the above events, the AI ​​model is considered deactivated, and / or the timer associated with the AI ​​model (if any) is considered expired or should be stopped.

[0401] In some embodiments, AI model lifecycle management can be based on AI performance evaluation. Performance evaluation is used to evaluate the feasibility of the AI ​​model, and the feasibility evaluation of the AI ​​model can be performed on the NW side or the UE side. In an example embodiment, the inferred RSRP value of one or more beams is compared with the actual RSRP value of the same beam.

[0402] In one embodiment, the AI ​​model may be deemed feasible if the difference between the inferred RSRP value of one or more beams and the actual RSRP value of the same beam is less than a preconfigured / defined threshold. If the difference is equal to or greater than the preconfigured threshold, the AI ​​model may be deemed infeasible. In some embodiments, the determination of the feasibility of the AI ​​mode may not rely on one comparison. It may be configured with a value 'N' as the maximum number of times. When N consecutive differences between the inferred average / best RSRP value of one or more beams and the actual average / best RSRP value of the same beam are equal to or greater than the preconfigured threshold, the AI ​​model may be deemed failed / inappropriate.

[0403] If performance evaluation is performed on the NW side, the UE may need to report feedback information for performance evaluation. The feedback information includes the inferred RSRP value of one or more beams and the actual RSRP value of the same beam. The inferred RSRP value of one or more beams can be obtained from the output of the model inference. The actual RSRP value of the same beam can be obtained from the CSI measurement / report for non-AI based beam management.

[0404] If the performance evaluation shows that the AI ​​model is no longer feasible, the NW may send deactivation signaling for the current AI model to the UE, or the NW may send an indication to the UE to switch the applied AI model to another AI model.

[0405] If performance evaluation is performed on the UE side, in addition to the SSB / CSI-RS used for AI-based beam management, the NW may need to configure the SSB / CSI RS associated with the beam for the UE to perform measurements in order to obtain the actual RSRP value.

[0406] If the performance evaluation shows that the current AI model is no longer feasible, the UE will deactivate the AI ​​model, and then the UE sends an indication to the NW to disable the AI-based function. In another embodiment, the UE sends an indication of the evaluation result to the NW. The AI ​​model operation can be determined by the NW, and then the NW sends an AI model operation indication to the UE. Regarding the AI ​​model operation, at least one of the following may be included:

[0407] 1. Deactivate the applied AI model.

[0408] 2. Switch the applied AI model to another AI model.

[0409] Case 4: UE-side model, online training

[0410] In this case, the AI ​​models can be trained online on the UE side. These AI models can be developed by the UE vendor. Figure 7 An example overall process in terms of data and logical flow of information for this case is shown in FIG.

[0411] like Figure 7 As further shown, an example population for Case 4 may include:

[0412] Step 0: The UE sends a request to the NW to enable AI-based functions and / or AI model training.

[0413] Step 1: The NW sends an indication to the UE to activate the AI ​​model.

[0414] Step 2: The NW configures the UE with the input data for model training via RRC signaling.

[0415] Step 3: NW sends signaling of input data for model training.

[0416] Step 4: UE starts model training using input data.

[0417] Step 5: The UE sends an indication of successful training of the AI ​​model to the NW.

[0418] Step 6: NW sends information about input data for model inference.

[0419] Step 7: The UE obtains input for model inference according to the information in step 6, and then obtains inference results from the AI ​​model.

[0420] Step 8: UE sends the inference result to NW for action.

[0421] Step 9: The NW takes actions based on the received inference results.

[0422] If the performance evaluation of the AI ​​model is performed on the UE side, then:

[0423] Step 10a: NW and UE start operations for AI model lifecycle management.

[0424] Step 11a: The UE processes the performance evaluation.

[0425] Step 11a: The UE sends an indication to the NW for notifying the operation of the applied AI model according to the evaluation result.

[0427] If the performance evaluation of the AI ​​model is performed on the NW side, then:

[0428] Step 10b: NW and UE start operations for AI model lifecycle management.

[0429] Step 11b: The NW processes the performance evaluation according to the information received from the UE.

[0430] Step 12b: The NW sends an indication of the operation of the applied AI model to the UE according to the evaluation result. In step 0:

[0431] The request message may include at least one of the following information:

[0432] 1.UCI;

[0433] 2.UE auxiliary information;

[0434] 3.UL MAC CE;

[0435] 4.UE capability signaling.

[0436] For the above message types, the following information may be included:

[0437] 1. AI-based functional information, such as an AI-based functional identifier.

[0438] 2. AI model information, such as:

[0439] a) AI model identifier.

[0440] b) Time requirements for AI model inference.

[0441] c) Time requirements for AI model training.

[0442] d) The time value of the AI ​​model’s life cycle.

[0443] e) The type of input data used for model inference and / or training.

[0444] f) Output data type for model inference.

[0445] g) Recommended configuration for input data types (e.g., inference and / or training), e.g., CSI resources

[0446] The period of the source, the period of the CSI report and the number of inferred periods, etc.

[0447] In steps 1, 2, and 3:

[0448] The RRC configuration of the input data for model training includes the RRC configuration of CSI measurement and reporting for beam management, and therefore, the input data for model training is the measurement result of CSI / SSB for beam management. In some embodiments of step 1, the NW may activate the AI ​​model for training through a DL MAC CE or an RRC message according to the request information in step 0.

[0449] In step 4:

[0450] For model training, a timer can be introduced for model training. The timer is used to avoid endless training caused by the AI ​​model failing to converge for a long time. The operation of the timer is as follows:

[0451] 1. Upon receiving an indication to activate the AI ​​model for online training / retraining / fine-tuning, start a timer.

[0452] 2. Upon receiving an instruction to deactivate the AI ​​model for online training, or determining that the AI ​​model training is successful

[0453] When the timer is stopped (for example, an indication of successful model training is sent to the NW), the timer is stopped.

[0454] 3. If the timer expires, the AI ​​model training is considered to have failed, and in some implementations, the UE needs to send a message to the NW to indicate the model training failure.

[0455] In step 5

[0456] When the AI ​​model training is successful, the UE sends an indication to the NW. The indication can be in at least one of the following formats:

[0457] 1: UL MAC CE;

[0458] 2: UCI;

[0459] 3: UE auxiliary information.

[0460] In step 6:

[0461] In the example of AI-based beam management, the input data for model inference is the CSI-RS / SSB reference signal for AI-based beam management.

[0462] In step 7, step 8, and step 9:

[0463] The output of model inference can be at least one of the following information:

[0464] 1. The first K beams and associated RSRP values, K is configurable and can be equal to and greater than 1.

[0465] 2. All beam information and associated RSRP values.

[0466] 3. A list of the first K beams and associated RSRP values ​​for the time period, where K is configurable and can be equal to and greater than 1.

[0467] a) The time period may be indicated by M consecutive CSI resources and / or CSI reporting periods, where M may be configurable and may be equal to or greater than 1.

[0468] In some implementations, if K>1, the behavior at NW for inferred data may include:

[0469] 1. Based on the received inference information, a DL signal is sent to the UE for dynamically adjusting the UL / DL beam.

[0470] MAC CE or DCI.

[0471] 2. Based on the received inference information, send to the UE a message for dynamically adjusting a series of UL / DL

[0472] DL MAC CE or DCI of the beam.

[0473] In some embodiments, the time period may be indicated by M consecutive CSI resources and / or CSI reporting periods, where M may be configurable and may be equal to or greater than 1.

[0474] In some embodiments, if K=1, it means that the behavior on both the NW side and the UE side for the inferred data is to automatically adjust the UL / DL beam for one time or a period of time according to the inferred output.

[0475] In steps 10, 11, and 12 for lifecycle management:

[0476] For timer-based lifecycle management, a timer is configured to the UE for lifecycle management of the AI ​​model of the AI-based function. The timer should be started / restarted on at least one of the following events:

[0477] 1. Receive activation signaling for AI model inference and / or training for AI-based beam management.

[0478] 2. Feedback information is sent to NW for performance evaluation of the activated AI model.

[0479] In some cases, an indication of a performance evaluation may be received that shows that the current AI model is still suitable. Alternatively, a performance evaluation of the activated AI model determines that the current AI model is still suitable. In some embodiments, whether the AI ​​model is suitable may be determined based on the difference between the inferred average / best RSRP value of one or more beams and the actual average / best RSRP value of the corresponding beam. If the difference is less than a preconfigured / defined threshold, the AI ​​model may be deemed suitable. If the difference is equal to or greater than a preconfigured threshold, the AI ​​model may be deemed unsuitable.

[0480] In some embodiments, the timer may be stopped when at least one of the following events occurs:

[0481] 1. Receive deactivation signaling for the AI ​​model.

[0482] 2. The AI ​​model is seen as being deactivated by some event.

[0483] In some implementations, when the timer expires, the AI ​​model is considered deactivated.

[0484] In some implementations, AI model lifecycle management may be event-based, and an applied AI model for an AI-based function may be deemed infeasible based on at least one of the following events:

[0485] 1. Beam failure is detected for (one or more) BWPs / serving cells to which the AI ​​model is applied.

[0486] 2. Radio link failure is detected for the BWP(s) / serving cell(s) to which the AI ​​model is applied.

[0487] 3. Deactivation of BWP(s) / serving cell(s) applying the AI ​​model.

[0488] 4. Reception of HO command.

[0489] 5.MAC reset.

[0490] In some embodiments, for the above events, the AI ​​model is considered deactivated, and / or the timer associated with the AI ​​model (if any) is considered expired.

[0491] In some embodiments, AI model lifecycle management may be based on AI model performance. For the example case of AI-based beam management, performance evaluation may be used to evaluate the feasibility of the applied AI model by comparing the inferred RSRP value of one or more beams with the actual RSRP value of the same beam. The feasibility evaluation of the applied AI model may be performed on the NW side or the UE side.

[0492] In one embodiment, if the difference between the inferred RSRP value of one or more beams and the actual RSRP value of the same beam is less than a preconfigured / defined threshold, the AI ​​model may be deemed feasible. If the difference is equal to or greater than the preconfigured threshold, the AI ​​model may be deemed infeasible. In some embodiments, the determination of the feasibility of the AI ​​mode may not rely on one comparison, which may be configured with a value 'N' as the maximum number of times, and the AI ​​model may be deemed failed / inappropriate when N consecutive differences between the inferred average / best RSRP value of one or more beams and the actual average / best RSRP value of the same beam are equal to or greater than the preconfigured threshold.

[0493] If performance evaluation is performed on the NW side, the UE may need to report feedback information for performance evaluation. The feedback information may include the inferred RSRP value of one or more beams and the actual RSRP value of the same beam. The inferred RSRP value of one or more beams may be obtained from the output of the model inference. The actual RSRP value of the same beam may be obtained from the CSI measurement / report for beam management.

[0494] If the performance evaluation shows that the AI ​​model is no longer viable, the NW can:

[0495] 1. Send deactivation signaling for the current AI model to the UE, or

[0496] 2. The NW may send an indication to the UE to switch the applied AI model to another AI model; or

[0497] 3. The NW may send instructions to the UE to fine-tune the applied AI model; or

[0498] 4. The NW may send an indication to the UE to retrain the applied AI model.

[0499] In some embodiments, if performance evaluation is performed on the UE side, in addition to the SSB / CSI-RS used for AI-based beam management, the NW needs to configure the SSB / CSI RS associated with the beam for the UE to perform measurements to obtain the actual RSRP value.

[0500] If the performance evaluation shows that the current AI model is no longer feasible, the UE will deactivate the AI ​​model, and then the UE sends an indication to the NW to disable the AI-based function. In another embodiment, the UE sends an indication of the evaluation result to the NW, the AI ​​model operation can be determined by the NW, and then the NW sends an AI model operation indication to the UE. Regarding the AI ​​model operation, the following items may be included:

[0501] 1. Deactivate the applied AI model.

[0502] 2. Switch the applied AI model to another AI model.

[0503] 3. Retrain the applied AI model.

[0504] 4. Fine-tune the applied AI model.

[0505] Case 5: Two-sided model, only offline training

[0506] In this case, the AI ​​model can be located on both the UE side and the NW side, and the model can be trained offline. Figure 8 An example overall process of data and logical flow of information is shown in FIG.

[0507] In some embodiments, the NW side may also be configured with a separate function, entity, layer or OAM (operation, maintenance and management function) to handle AICS provisioning and configuration. Such an entity may reside in Figure 1 in the core network shown. Alternatively, such entity may be part of the access network. In some embodiments, such entity may reside in a base station (such as a gNB as a separate function). In some other embodiments, the AI ​​entity and the gNB function may be an integrated part of the base station. In such embodiments, the information exchanged over the AI ​​interface will be irrelevant, and the operation of the AI ​​functional entity / layer will be considered as the operation of the gNB. When the AI ​​entity and the gNB are implemented separately (functionally or physically), the communication interface (both data and control) between the AI ​​entity and the gNB may be referred to as the AI ​​interface, and the communication between the UE or wireless terminal device and the gNB or other base station may be referred to as the air interface, such as Figure 8 shown.

[0508] like Figure 8 As further shown, an example overall process in the air interface for Case 5 may include:

[0509] Step 0: The UE sends a request to enable AI-based CSI feedback enhancement.

[0510] Step 1: NW selects two-side AI models (e.g., AI model pair) for CSI feedback enhancement and sends

[0511] The UE sends AI model information for compressing CSI feedback, or sends an indication to the UE for activating an AI model for compressing CSI feedback.

[0512] Step 2: The UE may activate a pre-stored corresponding AI model for compressing CSI feedback according to the indication received in step 1, or install the AI ​​model after receiving the AI ​​model information from the NW.

[0513] Step 3: The NW sends the CSI measurement / reporting configuration to the UE for AI-based CSI feedback enhancement.

[0514] Step 4: The NW sends the CSI-RS for measurement to the UE.

[0515] Step 5: The UE performs measurements on the CSI-RS and compresses the CSI feedback by inferring the UE-side AI model for compressing the CSI feedback.

[0516] Step 6: The UE sends compressed CSI feedback to the NW.

[0517] Step 7: The UE sends compressed CSI feedback and uncompressed CSI feedback to the NW in a specific manner for the performance evaluation of the two-side model.

[0518] Step 8: NW sends AI model actions to UE based on the evaluation results (if necessary).

[0519] In the case where the NW-side AI model resides in the AI ​​entity, the overall process for the AI ​​interface includes:

[0520] Step 0: The gNB sends a request to enable AI-based CSI feedback enhancement.

[0521] Step 1: The AI ​​entity / layer selects the two-sided AI models for the requested functionality.

[0522] Step 2: The AI ​​entity / layer sends the AI ​​model information to the gNB.

[0523] Step 3: The gNB sends the compressed CSI feedback to the AI ​​entity as input for model inference.

[0524] Step 4: The AI ​​entity obtains the compressed CSI feedback and infers the compressed CSI through the NW-side AI model

[0525] Feedback is decompressed.

[0526] Step 5: The AI ​​entity sends the decompressed CSI feedback to the gNB.

[0527] Step 6: The AI ​​entity receives feedback information from the gNB for performance evaluation.

[0528] Step 7: The AI ​​entity performs performance evaluation based on the feedback information including the received compressed CSI feedback and the uncompressed CSI feedback to generate an evaluation result.

[0529] Step 8: The AI ​​entity sends AI model actions (if required) to the gNB based on the evaluation results.

[0530] For step 0 in the air interface:

[0531] The request message for enabling AI-based CSI feedback may be through at least one of the following:

[0532] 1.UAI;

[0533] 2.UCI;

[0534] 3.UL MAC CE;

[0535] 4.UE capability signaling.

[0536] Based on the above information, the request message may include at least one of the following information:

[0537] 1. AI-based function information: used to identify the AI-based function from the AI-based function set. In this case, the AI-based function information is used to represent the AI-based CSI feedback.

[0538] 2. At least one of the following AI model-related information:

[0539] a) AI model ID.

[0540] b) AI model complexity requirements.

[0541] c) AI model types, including DNN (deep-learning neural network), CNN (convolutional neural network), etc.

[0542] d) Information related to the input data used for the AI ​​model.

[0543] e) Information related to the output data used for AI models.

[0544] f) Whether online reinforcement training or fine-tuning is supported.

[0545] For step 1 and step 2 of the air interface:

[0546] The derivation of the AI ​​model can be on the UE side. There may be two example solutions:

[0547] 1. Direct AI model transfer.

[0548] a) In step 2, the UE needs to download the AI ​​model from the NW for compressing the CSI-RS feedback.

[0549] 2. Based on the indication received in step 1, activate the AI ​​model for compressing CSI feedback from the pre-stored AI model list.

[0550] a) In step 2, the UE needs to activate the indicated UE-side AI model for compressed CSI feedback based on the information received from the NW. In some embodiments, the indication may be included in

[0551] DL MAC CE or DCI.

[0552] Step 7 for air interface:

[0553] In some implementations where the UE sends compressed CSI feedback and / or uncompressed CSI feedback, at least one of the following methods may be used:

[0554] 1: Time periodicity, there is a CSI reporting configuration for uncompressed CSI feedback, which is configured with a specific time period 'T' configured for the UE to send uncompressed CSI feedback to the NW. In this embodiment, the period value for reporting uncompressed CSI feedback can be an integer multiple of the period value for reporting compressed CSI feedback.

[0555] 2: Half-duration periodicity, in one embodiment, there are more than one CSI reporting configurations, which are configured with different time periods 'T' configured for the UE to send uncompressed CSI feedback to the NW, and each CSI reporting configuration can be activated / deactivated by the DL MAC CE. In another embodiment, there is one CSI reporting configuration for the UE to send uncompressed CSI feedback, and in such CSI

[0556] The time period configured in the reporting configuration may be dynamically adjusted through DL MAC CE or DCI.

[0557] 3: Time aperiodicity, the NW can send DCI and / or DL ​​MAC CE to trigger compressed CSI reporting and / or uncompressed CSI reporting, so that the UE can send uncompressed CSI feedback at any time.

[0558] To perform lifecycle management, the UE may need to send uncompressed CSI feedback as well as compressed CSI feedback to the NW in time period T for the NW to process the performance evaluation of the applied AI model. The AI ​​model operation may be determined by the NW based on the difference between the decompressed CSI feedback and the uncompressed CSI feedback.

[0559] In some embodiments, the performance evaluation of the applied AI model is processed. If the difference between the uncompressed CSI feedback and the decompressed CSI feedback is equal to or greater than a predefined / configured threshold for N or N consecutive times within a specific time period M, the two-sided AI model is deemed no longer feasible. In one embodiment, N may be configurable and the value of N may be equal to or greater than 1. In one embodiment, the time period M is configurable. Otherwise, the two-sided AI model is deemed feasible.

[0560] In some example embodiments, the lifecycle management of the AI ​​function or AI model may be time-based. For example, an AI function management timer or an AI model management timer may be introduced and configured. Such timers may be maintained on the NW side (e.g., gNB or AI function layer / entity / OAM) that provides the AI ​​function or AI model. Such timers may operate in the following manner:

[0561] 1. When an indication to activate the AI ​​model or AI function is received (e.g. on the gNB side), or when the AI ​​function layer / entity / OAM of the NW sends an ACK to the gNB in ​​response to the AI ​​function request, or the NW side sends an ACK to the UE

[0562] When an ACK is sent, a timer is started. When the performance evaluation result showing the AI ​​model performance is better than a predefined threshold, the timer can be restarted.

[0563] 2. When the timer expires: For example, an indication may be sent by the gNB to the AI ​​functional layer / entity to deactivate the applied AI model or AI function. For another example, an indication may be sent from the AI ​​functional layer / entity to the gNB to disable the AI-based CSI feedback function / configuration.

[0564] 3. When the AI ​​function entity / layer receives an indication from the gNB to deactivate the AI ​​model or AI function, or when the gNB receives an indication from the AI ​​function entity / layer to disable the AI-based CSI feedback model or function (e.g., CSI feedback), the timer is stopped.

[0565] Step 8 for the air interface:

[0566] The NW may consider AI model related operations based on the evaluation results based on the received decompressed CSI feedback and compressed CSI feedback. If the performance evaluation shows that the applied AI model is not feasible / suitable, the AI

[0567] A model operation is at least one of the following:

[0568] a) Deactivate the AI ​​models on both sides;

[0569] b) Switch the two-sided AI model to another two-sided AI model, and then implement the new two-sided AI model from step 1 in the air interface.

[0570] Case 6: Two-sided model, online training

[0571] In this case, the AI ​​model can be located on both the UE side and the NW side, and the model can be trained online. Fig. 9 An example overall process of data and logical flow of information is shown in FIG.

[0572] For two-sided model training, in some embodiments, the NW or UE performs two-sided model training (e.g., AI model for decompression, AI model for compression) using input training data, and after the two-sided AI model training is successful, if the NW performs two-sided model training, the NW should transmit the UE-side AI model to the UE, or if the UE performs two-sided model training, the UE should transmit the NW-side AI model to the NW.

[0573] In some embodiments, the NW and the UE start training AI models (e.g., an AI model for decompression, an AI model for compression) at the same time.

[0574] In some embodiments, the NW and the UE start training the corresponding AI models in sequence, for example, the UE first starts training the UE-side AI model, and after the UE successfully trains the AI ​​model, the UE can provide two methods for the NW to train the NW-side AI model.

[0575] In some embodiments, the UE-side AI model is transmitted to the NW, and the UE sends the input value inferred by the UE-side AI model to the NW for NW-side AI model training, and the NW will perform NW-side AI model training using the output inferred by the UE-side model and the input inferred by the UE-side model.

[0576] In the second method, the UE always sends the input of the UE-side model inference and the output of the UE-side model inference to the NW to train the NW-side AI model until the NW-side AI model training is successful on the NW side.

[0577] In some embodiments, a timer can be configured to control AI model training on each side. The timer can be started when the AI ​​model starts training, the timer can be stopped when the AI ​​model training is successful, and the model training will be considered failed due to the expiration of the timer.

[0578] In some embodiments, the NW side may also be configured with a separate function, entity, layer or OAM (operation, maintenance and management function) to handle AICS provisioning and configuration. Such an entity may reside in Figure 1 in the core network shown. Alternatively, such entity may be part of the access network. In some embodiments, such entity may reside in a base station (such as a gNB as a separate function). In some other embodiments, the AI ​​entity and the gNB function may be an integrated part of the base station. In such embodiments, the information exchanged over the AI ​​interface will be irrelevant, and the operation of the AI ​​functional entity / layer will be considered as the operation of the gNB. When the AI ​​entity and the gNB are implemented separately (functionally or physically), the communication interface (both data and control) between the AI ​​entity and the gNB may be referred to as the AI ​​interface, and the communication between the UE or wireless terminal device and the gNB or other base station may be referred to as the air interface, such as Fig. 9 shown.

[0579] like Fig. 9 As further shown, an example overall process in the air interface for Case 5 may include:

[0580] Step 0: The UE sends a request to enable AI-based CSI feedback enhancement.

[0581] Step 1: The NW selects the two-side AI model for AI-based CSI feedback enhancement according to the UE request information.

[0582] Step 2: NW configures the CSI configuration for AI model training.

[0583] Step 3: The NW sends the CSI-RS for measurement to the UE.

[0584] Step 4: UE reports uncompressed CSI feedback to NW, and then NW starts to use uncompressed CSI

[0585] CSI feedback is used to train the AI ​​models on both sides.

[0586] Step 5: After successfully training the two-sided AI model for CSI feedback enhancement on the NW side, the NW sends the UE-side AI model of the two-sided AI model to the UE. The UE installs the received model. The NW sends the CSI RS for feedback to the UE to generate input data for the UE-side model.

[0587] Step 6: The UE compresses the CSI-RS feedback using the UE-side AI model as the model inference output.

[0588] Step 7: The UE sends compressed CSI feedback to the NW, and the NW regards the compressed CSI feedback as NW

[0589] The UE also sends uncompressed CSI feedback information in a specific way for AI model performance evaluation.

[0590] Step 8: NW performs performance evaluation and obtains AI model operation based on the evaluation result, and then sends

[0591] UE sends AI model operation message (if necessary).

[0592] In the case where the NW-side AI model resides in the AI ​​entity, the overall process of the AI ​​interface includes:

[0593] Step 0: The gNB sends a request for AI-based CSI feedback to the AI ​​entity / layer.

[0594] Step 1: The AI ​​entity selects an AI model for training.

[0595] Step 2: The AI ​​entity sends the selected AI model information to the gNB, and the AI ​​entity starts model training.

[0596] Step 3: After the AI ​​model is trained successfully, the gNB sends compressed CSI feedback to the AI ​​entity / layer.

[0597] Step 4: The AI ​​entity decompresses the CSI feedback using the decompression AI model.

[0598] Step 5: The AI ​​entity sends the decompressed CSI feedback to the gNB.

[0599] Step 6: The gNB sends feedback to the AI ​​entity for performance evaluation.

[0600] Step 7: The AI ​​entity performs performance evaluation on the applied AI model.

[0601] Step 8: The AI ​​entity performs operations on the AI ​​models on both sides based on the evaluation results.

[0602] Step 0 for the air interface:

[0603] The request message for enabling AI-based CSI feedback may be through at least one of the following:

[0604] 1.UAI;

[0605] 2.UCI;

[0606] 3.UL MAC CE;

[0607] 4.UE capability signaling.

[0608] Based on the above information, the request message may include at least one of the following information:

[0609] 1. AI-based function information: used to identify the AI-based function from the AI-based function set. In this case, the AI-based function information is used to represent the AI-based CSI feedback.

[0610] 2.AI model ID.

[0611] 3.AI model complexity requirements.

[0612] 4. AI model types, including DNN, CNN, etc.

[0613] 5. Information about the input data used for the AI ​​model.

[0614] 6. Information related to the output data used for AI models.

[0615] 7. Whether online reinforcement training or fine-tuning is supported.

[0616] Step 7 for air interface:

[0617] In some implementations where the UE sends compressed CSI feedback and / or uncompressed CSI feedback, at least one of the following methods may be used:

[0618] 1: Time periodicity, there is a CSI reporting configuration for uncompressed CSI feedback, which is configured with a specific time period 'T' configured for the UE to send uncompressed CSI feedback to the NW. In this embodiment, the period value for reporting uncompressed CSI feedback can be an integer multiple of the period value for reporting compressed CSI feedback.

[0619] 2: Half-duration periodicity, in one embodiment, there are more than one CSI reporting configurations, which are configured with different time periods 'T' configured for the UE to send uncompressed CSI feedback to the NW, and each CSI reporting configuration can be activated / deactivated by the DL MAC CE. In another embodiment, there is one CSI reporting configuration for the UE to send uncompressed CSI feedback, and in such CSI

[0620] The time period configured in the reporting configuration may be dynamically adjusted through DL MAC CE or DCI.

[0621] 3: Time aperiodicity. The NW may send DCI and / or MAC CE to trigger compressed CSI reports and / or uncompressed CSI reports for the UE to send uncompressed CSI feedback at any time.

[0622] For model life cycle management, within a time period T, the UE may need to send uncompressed CSI feedback and compressed CSI feedback to the NW for the NW to process the performance evaluation of the applied AI model. The operation of the AI model can be determined by the NW based on the difference between the decompressed CSI feedback and the uncompressed CSI feedback.

[0623] In some embodiments, the performance evaluation of the applied AI model is processed. If within a specific time period M, the difference between the uncompressed CSI feedback and the decompressed CSI feedback is equal to or greater than a predefined / configured threshold for N consecutive times or N times, the AI models on both sides are considered no longer viable. In one embodiment, N can be configurable and the value of N can be equal to or greater than 1. Otherwise, the AI models on both sides are considered viable.

[0624] In some embodiments, there may be more than one threshold, which are predefined and / or preconfigured for determining different operations of the AI model. For example, there are two thresholds A and B, and A < B. If the difference is greater than threshold A but less than threshold B, the AI model can be fine-tuned online and the AI-based function can remain as it is. If the value of the difference is greater than threshold B, the AI model should be retrained.

[0625] In some exemplary embodiments, the life cycle management of the AI function or the AI model can be time-based. For example, an AI function management timer or an AI model management timer can be introduced and configured. Such timers can be maintained on the NW side that provides the AI function or the AI model (e.g., gNB or AI function layer / entity / OAM). Such timers can operate in the following manner:

[0626] 1. When receiving an indication to activate the AI model or the AI function (e.g., on the gNB side), or when the AI function layer / entity / OAM of the NW sends an ACK to the gNB in response to an AI function request or the NW side sends an ACK to the UE

[0627] the timer is started. When the performance evaluation result showing the performance of the AI model is better than a predefined threshold, the timer can be restarted.

[0628] 2. When the timer expires: For example, an indication may be sent by the gNB to the AI ​​functional layer / entity to deactivate the applied AI model or AI function. For another example, an indication may be sent from the AI ​​functional layer / entity to the gNB to disable the AI-based CSI feedback function / configuration.

[0629] 3. When the AI ​​function entity / layer receives an indication from the gNB to deactivate the AI ​​model or AI function, or when the gNB receives an indication from the AI ​​function entity / layer to disable the AI-based bCSI feedback model or function (e.g., CSI feedback), the timer is stopped.

[0630] Step 8 for the air interface:

[0631] NW can consider the relevant operations of the AI ​​models on both sides based on the evaluation results for CSI feedback enhancement. The AI ​​model operations are as follows:

[0632] 1. Deactivate the AI ​​models on both sides.

[0633] 2. Switch the current two-side AI model to another two-side AI model.

[0634] 3. Fine-tune the AI ​​models on both sides.

[0635] 4. Retrain the AI ​​models on both sides.

[0636] The above description and accompanying drawings provide specific example embodiments and implementations. However, the described subject matter can be implemented in various different forms, and therefore, the subject matter covered or claimed is intended to be interpreted as not being limited to any example embodiment set forth herein. It is intended to provide a reasonably wide scope for the subject matter claimed or covered. In particular, for example, the subject matter can be implemented as a method, device, component, system, or non-transient computer-readable medium for storing computer code. Therefore, the embodiment can, for example, take the form of hardware, software, firmware, storage medium, or any combination thereof. For example, the above method embodiment can be implemented by a component, device, or system including a memory and a processor by executing a computer code stored in the memory.

[0637] Throughout the specification and claims, terms may have nuanced meanings that are suggested or implied by the context, beyond those explicitly stated. Similarly, the phrase "in one embodiment / implementation" as used herein does not necessarily refer to the same embodiment, and the phrase "in another embodiment / implementation" as used herein does not necessarily refer to a different embodiment. For example, the claimed subject matter is intended to include, in whole or in part, a combination of the example embodiments.

[0638] In general, terms can be understood at least in part according to the usage in the context. For example, terms such as "and", "or", or "and / or" used in this article can include various meanings, which can depend at least in part on the context in which these terms are used. Typically, if "or" is used to associate a list such as A, B, or C, it is intended to represent A, B, and C (here for inclusive meanings), and A, B, or C (here for exclusive meanings). In addition, at least in part depending on the context, the term "one or more" as used in this article can be used to describe any feature, structure, or characteristic in a singular sense, or can be used to describe a combination of features, structures, or characteristics in a plural sense. Similarly, at least in part depending on the context, terms such as "one", "an", or "the" can be understood to convey singular usage or to convey plural usage. In addition, also at least in part depending on the context, the term "based on" can be understood as not necessarily intended to convey a set of factors that are exclusive, but allows the presence of additional factors that are not necessarily explicitly described.

[0639] References to features, advantages, or similar language throughout the specification do not imply that all features and advantages that may be achieved using the present solution should be or are included in any single implementation thereof. Rather, language referencing such features and advantages is understood to mean that a particular feature, advantage, or characteristic described in connection with an embodiment is included in at least one embodiment of the present solution. Thus, discussion of features and advantages and similar language throughout the specification may, but need not, refer to the same embodiment.

[0640] In addition, the features, advantages, and characteristics of the present solution may be combined in any suitable manner in one or more embodiments. Based on the description herein, one of ordinary skill in the relevant art will recognize that the present solution may be practiced without one or more of the specific features or advantages of a particular embodiment. In other cases, additional features and advantages may be recognized in certain embodiments that may not be present in all embodiments of the present solution.

Claims

1. A method performed by at least one wireless network node, comprising: In response to receiving a request for AICS (AI configuration service) from a wireless terminal device, activating an artificial intelligence (AI) network configuration mode; Determine the network-side AI model according to the request to AICS; Receiving an input data item set for the network-side AI model from the wireless terminal device; Determine a network configuration behavior (NCA) based on the inference of the network-side AI model according to the set of input data items; transmitting the NCA or an indication of the NCA to the wireless terminal device; receiving an AI feedback data item from the wireless terminal device, the AI ​​feedback data item being generated by the wireless terminal device according to the indication and based on a deployment result of the NCA; and At least one AI network configuration management task is performed according to the AI ​​feedback data item.

2. The method according to claim 1, wherein: The request to AICS is implemented by one of the following: an uplink medium access control (MAC) control element (CE) from the wireless terminal device; an uplink control information (UCI) message from the wireless terminal device; a physical uplink control channel (PUCCH) scheduling request (SR) from the wireless terminal device; User equipment (UE) assistance information in a radio resource control (RRC) message from the wireless terminal device; or Capability signaling of the wireless terminal device.

3. The method according to claim 2, wherein: The request to AICS includes at least one of the following: an AI function indicator for indicating an AI configuration function among a plurality of AI configuration functions; or AI-assisted information for network-side AI model selection or training, The multiple AI configuration functions include at least: AI-based spatial beam management configuration function, AI-based time beam management configuration capabilities, and AI-based CSI feedback configuration function; or The AI-assisted information includes at least one of the following: The location information of the wireless terminal device, speed information of the wireless terminal device, Direction information of the wireless terminal device, or The device pointing information of the wireless terminal device.

4. The method according to claim 1, wherein: Activating the AI ​​network configuration mode includes one of the following: Switching from a non-AI channel state measurement configuration to an AI-based channel state measurement configuration is accomplished by at least one of the following: Deactivate the semi-periodic CSI reporting configuration associated with the non-AI based beam managed CSI resource set, Activate the semi-periodic CSI reporting configuration associated with the predefined AI-based beam-managed CSI resource set, Enable or add periodic CSI reporting configuration associated with a predefined CSI resource set for AI-based beam management through RRC reconfiguration, or disabling or removing, through RRC reconfiguration, the periodic CSI reporting configuration associated with the CSI resource set for non-AI based beam management; or For a specific channel state measurement configuration, switching from a non-AI channel state resource set to an AI-based channel state resource set is performed by at least one of the following: Dynamically switch the CSI resources associated with the CSI reporting configuration from AI to non-AI using the downlink MAC CE, Dynamically changing the CSI reporting period value associated with the specific channel state measurement configuration through the downlink MAC CE, or The CSI resource period associated with the specific channel state measurement configuration is dynamically changed through the downlink MAC CE.

5. The method according to claim 1, wherein: The input data item set to the network-side AI model includes at least one of the following: Channel State Information Reference Signal (CSI-RS) measurement results or Synchronization Signal Block (SSB) measurement results, time information associated with the CSI-RS measurement result or the SSB measurement result, Track information of the wireless terminal device, Beam information, The location information of the wireless terminal device, device pointing information of the wireless terminal device, or Speed ​​information of the wireless terminal device; and The input data item set to the network-side AI model is transmitted from the wireless terminal device by one of the following: UCI on PUCCH, UCI on the Physical Uplink Shared Channel (PUSCH), or Uplink MAC CE.

6. The method according to claim 1, wherein: The NCA includes one of the following: Switching uplink beam or downlink beam; switching a plurality of uplink beams or downlink beams in sequence; or If there is only one TCI state in the current transmission configuration information (TCI) state pool, reconfigure the current TCI state pool; and The NCA or the indication for the NCA is transmitted to the wireless terminal device via a downlink MAC CE or a downlink control information (DCI) message.

7. The method according to claim 1, wherein: The at least one AI network configuration management task: is time-based according to an AI timer, the AI ​​timer being enabled by the at least one radio network node for setting a time period for the AI ​​network configuration mode; is triggered by detecting one of the following network events, the network event comprising one of: beam failure, radio link failure, deactivation of a bandwidth part (BWP) or serving cell associated with the AICS, a handover command, or a MAC reset; or It is based on a performance evaluation of the network-side AI model, wherein the performance evaluation includes a comparison between a reference signal received power (RSRP) inferred from the network-side AI model and an actual measurement result.

8. The method according to claim 7, wherein: Additional CSI resource sets and reports are configured for the performance evaluation of the network-side AI model; and The additional CSI resource set and reporting period are longer than a management period associated with the AICS.

9. The method according to claim 7, wherein: The at least one AI network configuration management task includes at least one of the following: Switch to a different network-side AI model; Retraining the network-side AI model; Strengthen the training of the network-side AI model; Fine-tuning the network-side AI model; Deactivating the network-side AI model; disable said AICS; or The AICS is retrained but reconfigured.

10. The method according to claim 1, wherein: The AI ​​feedback data item includes at least one of the following: AI-inferred RSRP value for CSI-RS or SSB; The actual RSRP value of CSI-RS or SSB; The AI-inferred optimal beam sequence for a predefined or preconfigured time period; or The actual best beam sequence for a predefined or preconfigured time period.

11. The method according to claim 1, wherein: According to the request to AICS, determining the network-side AI model includes: Get a pre-trained AI model; or The network-side AI model is trained in the following ways: Transmitting AI model input data configuration to the wireless terminal device; Receiving a training data set from the wireless terminal device, where the training data set is generated by the wireless terminal device according to the AI ​​model input data configuration; and A training or reinforcement training process is performed based on the training data set to generate the network-side AI model.

12. The method according to claim 11, wherein: The AI ​​model input data configuration includes at least one of the following: a CSI resource set for beam management configuration or a CSI report configuration for beam management configuration.

13. The method according to claim 11, wherein: The training data set includes at least one of the following: CSI-RS / SSB measurement results for beam management; Time information associated with the CSI-RS / SSB measurement results; or Location information associated with CSI-RS / SSB measurements.

14. The method according to claim 11, wherein: The at least one wireless network node is configured with an AI model training timer; When the training or reinforcement training process starts, the AI ​​model training timer starts; If the training or reinforcement training process is successful, the AI ​​model training timer is stopped before expiration; and If the training or reinforcement training process is unsuccessful, the AI ​​model training timer expires after a preconfigured time period.

15. The method of claim 1, wherein: The at least one wireless network node includes a wireless access network node and an AI functional entity; The request for the AICS from the wireless terminal device is received at the wireless network node and relayed to the AI ​​functional entity; The method further includes: in response to receiving the request for the AICS from the wireless access network node, the AI ​​function entity transmits a confirmation message, wherein the confirmation message includes at least one of the following: Information associated with the network-side AI model; or AI model performance evaluation information.

16. The method according to claim 15, wherein: The information associated with the network-side AI model includes at least one of the following: input type information of the input data item set of the network-side AI model, or output type information of the inference of the network-side AI model.

17. The method according to claim 16, wherein: The output type information for the inference of the network-side AI model includes at least one of the following: RSRP values ​​associated with the first K beams, where K is a positive integer; RSRP values ​​associated with the first K beams in a preconfigured time period, wherein the RSRP values ​​associated with the first K beams exist in each time instance with a configurable granularity in the preconfigured time period; time information of the preconfigured time period; or A time point having the configurable granularity in the preconfigured time period.

18. The method of claim 1, further comprising: Selecting a UE-side AI model according to the request to the AICS; and activating the UE-side AI model in the wireless terminal device or downloading the UE-side AI model to the wireless terminal device, Among them, the input data item set for the network-side AI model includes an output set of the UE-side AI model generated by the wireless terminal device and transmitted to the at least one wireless network node.

19. The method according to claim 18, further comprising: Transmitting a CSI-RS to the wireless terminal device to generate a CSI-RS measurement result, wherein: The output set of the UE-side AI model is generated by processing the CSI-RS measurement results into compressed CSI feedback using the UE-side AI model; The network-side AI model includes a CSI feedback decompression model; and The inference from the network-side AI model includes decompressed CSI feedback.

20. The method according to claim 19, wherein: The AI ​​feedback data item includes both the compressed CSI feedback and the uncompressed CSI feedback.

21. The method according to claim 20, wherein: The at least one AI network configuration management task is based on a performance evaluation of the network-side AI model and the UE-side AI model, wherein the performance evaluation is based on the uncompressed CSI feedback and the decompressed CSI feedback.

22. The method according to claim 19, wherein: Both the network-side AI model and the UE-side AI model are trained after the wireless terminal device generates the CSI-RS measurement result and before being used for inference.

23. A method performed by a wireless terminal device, comprising: transmitting a request for an AI configuration service (AICS) to at least one wireless network node; Receiving a management configuration for the AICS; Receiving an activation indication associated with a UE-side AI model; Receiving a set of auxiliary information items associated with the UE-side AI model; In response to the activation indication, based on the set of auxiliary information items, using the UE-side AI model to perform AI inference to generate inference data items; transmitting the inferred data item to the at least one wireless network node; receiving a network configuration action (NCA) determined by the at least one radio network node based on the inferred data item; as well as Perform AI service management or assist the at least one wireless network node in performing AI service management according to the management configuration for the AICS and the inferred data item.

24. The method according to claim 23, wherein: The request to AICS is transmitted by one of the following: an uplink medium access control (MAC) control element (CE) from the wireless terminal device; an uplink control information (UCI) message from the wireless terminal device; a physical uplink control channel (PUCCH) scheduling request (SR) from the wireless terminal device; UE assistance information in a radio resource control (RRC) message from the wireless terminal device; or Capability signaling of the wireless terminal device.

25. The method according to claim 24, wherein: The request to AICS includes at least one of the following: an AI function indicator for indicating an AI configuration function among a plurality of AI configuration functions; or AI model information, The multiple AI configuration functions include at least: AI-based spatial beam management configuration function, AI-based time beam management configuration function, AI-based CSI feedback configuration function, and AI-based wireless terminal positioning configuration; or The AI ​​model information includes at least one of the following: An identifier of the UE-side AI model; The processing time for inference of the UE-side AI model; Lifecycle information for the UE-side AI model; Input data type for the UE-side AI model; Output data type for the UE-side AI model; or A recommended radio resource configuration (RRC) for an input data item of the UE-side AI model.

26. The method according to claim 25, wherein: The management configuration for the AICS is received via a radio resource configuration (RRC) message and includes one of: RRC configuration or RRC reconfiguration of CSI measurement and reporting for AI-based beam management.

27. The method according to claim 25, wherein: The activation indication associated with the UE-side AI model is received at the wireless terminal device by one of: a downlink MAC CE, the downlink MAC CE comprising a first identifier associated with the UE-side AI model and a second identifier associated with the AI ​​configuration function; or An RRC message including an activation flag.

28. The method of claim 25, wherein: The auxiliary information item set includes a CSI-RS or SSB reference signal for AI-based beam management; and Performing the AI ​​inference is based on the measurement result of the CSI-RS or SSB reference signal.

29. The method according to claim 25, wherein: The inferred data item includes at least one of the following: The first K beams and their associated RSRP values, where K is a positive integer; All beam information and associated RSRP values; or A series of the top K beams and associated RSRP values ​​in a time period indicated by M consecutive CSI resources or reporting periods, where M is a configurable positive integer.

30. The method of claim 29, wherein: When K is greater than 1, the NCA is received via a downlink MAC CE based on the inferred data item, the downlink MAC CE involving a single dynamic adjustment of an uplink / downlink beam or a series of dynamic adjustments of an uplink / downlink beam within a preconfigured time period; and When K is equal to 1, the NCA includes adjusting the uplink / downlink beams on the network side.

31. The method of claim 23, wherein: The timer is configured to perform the AI ​​service management or assist the at least one wireless network node in performing the AI ​​service management; The timer is configured to start or restart when at least one of the following occurs: receiving the activation instruction; or Transmitting feedback information for a performance evaluation showing that the UE-side AI model is still appropriate; The timer is configured to stop upon receiving deactivation signaling for the UE-side AI model; and Expiration of the timer causes the UE-side AI model to be deactivated.

32. The method of claim 23, wherein: Performing the AI ​​service management or assisting the at least one wireless network node in performing the AI ​​service management according to the management configuration for the AICS and the inferred data item includes: deactivating the UE-side AI model when beam failure, radio link failure, deactivation of a bandwidth part (BWP) or a serving cell associated with the AICS, a handover command, or a MAC reset is detected.

33. The method of claim 23, further comprising: performing a performance evaluation on the UE-side AI model and transmitting the performance evaluation to the at least one radio network node, or receiving a performance evaluation performed by the at least one radio network node, The management configuration for the AICS indicates executing the AI ​​service management or assisting the at least one wireless network node in executing the AI ​​service management according to the performance evaluation.

34. The method of claim 33, wherein: The inferred data item comprises inferred RSRP values ​​for one or more beams; and The performance evaluation of the UE-side AI model includes comparing the difference between the inferred RSRP value of the one or more beams and the actual RSRP value of the same one or more beams with a preconfigured difference threshold.

35. The method of claim 33, wherein: The AI ​​service management includes at least one of the following: Switch to a different UE-side AI model; Retraining the UE-side AI model; Strengthen the training of the UE-side AI model; Fine-tuning the UE-side AI model; Deactivating the UE-side AI model; disable said AICS; or The AICS is retained but reconfigured.

36. The method of claim 33, further comprising: receiving a training data set sent by the at least one wireless network node in response to the request for the AICS; as well as The UE-side AI model is trained or enhanced using the training data set before performing the AI ​​inference.

37. The method of claim 33, wherein: The UE-side AI model includes a CSI feedback compression model; The method further includes: receiving a CSI-RS from the at least one radio network node for measurement, and measuring the CSI-RS to obtain a CSI-RS measurement result; The inferred data item includes compressed CSI feedback generated by processing the CSI-RS measurement result using the CSI feedback compression model.

38. The method of claim 37, further comprising: In addition to the compressed CSI feedback, uncompressed CSI feedback is transmitted to the at least one radio network node for the performance evaluation.

39. The at least one wireless network node or the wireless terminal device comprises a memory for storing instructions and a processor for executing the instructions, so as to implement any one of claims 1 to 38.

40. A computer-readable non-transitory medium, wherein: The computer-readable non-transitory medium is used to store computer instructions, and when the computer instructions are executed by a processor of a wireless network device or a wireless terminal device, any one of claims 1 to 38 is implemented.

Images