Method and apparatus for site-specific and implicit AI model refinement for 3GPP systems

Abstract

The present disclosure relates to a 5G or 6G communication system for supporting a higher data transmission rate. Embodiments disclosed herein relate to methods (600) and systems (400) for refining an artificial intelligence (AI) model in a wireless system. A method (600) provides site-specific and implicit AI model refinement in which site-specific dataset collection is performed as part of a 3GPP process of operation, and the collected site-specific dataset is used to train a new model or refine an existing model for CSI prediction and CSI compression. The method (600) includes training CSI prediction and CSI compression models for UEs having the ability to download a new model from a base station (402) and for UEs not having the ability.

Description

TECHNICAL FIELD

[0001] Embodiments disclosed herein relate to wireless communication networks, and more specifically, to improving artificial intelligence (AI) models in wireless systems. BACKGROUND

[0002] 5G mobile communication technologies define wide frequency bands so that high transmission rates and new services are possible, and are implemented not only in "Sub 6 GHz" frequency bands, but also in "6 GHz and above" frequency bands (mmWave). Also, 6G mobile communication technologies are being discussed, which are implemented in terahertz bands (for example, 95 GHz to 3 THz bands) to achieve transmission rates higher than 5G mobile communication technologies.

[0003] At the time of starting the development of 5G mobile communication technologies, in order to support services and meet the performance requirements related to enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), and massive machine-type communication (mMTC), standardization has been made on the following: beamforming and massive MIMO for mitigating radio wave path loss and increasing radio wave transmission distances in mmWave, support of numerologies (for example, operating multiple subcarrier spaces) for efficient utilization of mmWave resources and time slots, initial access techniques for supporting multi-beam transmission and wideband, definition and operation of BWPs (bandwidth parts), new channel coding methods such as LDPC (low-density parity check) codes for large amounts of data transmission and polar codes for highly reliable transmission of control information, L2 pre-processing, and network slicing for providing a dedicated network dedicated to a specific service.

[0004] Currently, in view of services to be supported by 5G mobile communication technologies, discussions are under way on improvement and performance enhancement of initial 5G mobile communication technologies, and physical layer standardization has been made on technologies such as V2X (vehicle-to-everything) for assisting in driving determination of autonomous vehicles based on information about the position and state of vehicles transmitted by the vehicles and for enhancing user convenience, NR-U (new radio-unlicensed) targeting system operation conforming to various regulatory requirements in unlicensed bands, NR UE power saving, non-terrestrial networks (NTN) as UE satellite direct communication for providing coverage in areas where communication with terrestrial networks is unavailable, and positioning.

[0005] Further, in the air interface architecture / protocol, standardization has been ongoing regarding technologies such as Industrial Internet of Things (IIoT) for supporting new services through interworking and convergence with other industries, IAB (Integrated Access and Backhaul) for providing a node for network service area expansion by supporting a wireless backhaul link and an access link in an integrated manner, mobility enhancement including conditional handover and DAPS (Dual Active Protocol Stack) handover, and two-step random access (two-step RACH for NR) for simplifying a random access procedure. In the system architecture / service, standardization has also been ongoing regarding 5G baseline architecture (e.g., service based architecture or service based interface) for combining Network Functions Virtualization (NFV) and Software-Defined Networking (SDN) technologies, and mobile edge computing (MEC) for receiving services based on UE location.

[0006] As the commercialization of 5G mobile communication systems, exponentially increasing connected devices will be connected to communication networks, and thus it is expected that enhanced functions and performance of the 5G mobile communication system and integrated operations of connected devices will be necessary. To this end, new research has been scheduled in connection with the following: extended reality (XR) for efficiently supporting AR (Augmented Reality), VR (Virtual Reality), MR (Mixed Reality), etc., improvement of 5G performance and reduction of complexity by utilizing artificial intelligence (AI) and machine learning (ML), AI service support, metaverse service support, and drone communication.

[0007] Further, such development of the 5G mobile communication system will serve as a basis not only for developing new waveforms for providing coverage in a terahertz band for 6G mobile communication technology, multi-antenna transmission technologies such as Full Dimensional MIMO (FD-MIMO), array antennas, and large-scale antennas, metamaterial-based lenses and antennas for improving coverage of terahertz band signals, high-dimensional space multiplexing technology using OAM (Orbital Angular Momentum), and RIS (Reconfigurable Intelligent Surface), but also for developing full-duplex technology for improving frequency efficiency of 6G mobile communication technology and for improving system networks, AI-based communication technology for implementing system optimization by utilizing satellites and AI (Artificial Intelligence) from the design stage and internalizing end-to-end AI support functions, and next-generation distributed computing technology for implementing services by utilizing super-high-performance communication and computing resources at a complexity level exceeding the limitation of UE operation capability. SUMMARY

[0008] SOLUTION TO PROBLEM The disclosure relates to a wireless communication network, and more particularly, to a terminal in a wireless communication system and a communication method thereof.

[0009] According to one aspect of this disclosure, embodiments herein provide a method for managing artificial intelligence (AI) models in a wireless system. The method includes: receiving at least one UE capability information from at least one user equipment (UE) by a base station. The method includes: downloading one or more AI models by the base station based on the UE capability information. The method includes: selecting at least one AI model from the downloaded AI models by the base station and selecting a Channel State Information (CSI) configuration for the UE. Thereafter, the method includes: training the selected AI model by the base station based on the selected CSI configuration to perform at least one CSI operation.

[0010] Beneficial effects of the invention The aspects of this disclosure at least address the aforementioned problems and / or disadvantages, and at least provide the advantages described below. Therefore, one aspect of this disclosure will provide an efficient communication method in a wireless communication system. Attached Figure Description

[0011] The embodiments described herein are illustrated in the accompanying drawings, and the same reference numerals indicate corresponding portions throughout the drawings. The embodiments described herein will be better understood from the following description with reference to the following illustrative drawings. The embodiments described herein are illustrated by way of example in the accompanying drawings.

[0012] FIG. 1A and FIG. 1B It describes CSI prediction based on artificial intelligence (AI) according to existing technologies; FIG. 2A and FIG. 2B AI-based CSI compression based on existing technologies is described; FIG. 3A The typical deployment cycle is shown for non-AI approaches or in the case of a general AI-based model, based on existing technologies. FIG. 3B A typical deployment cycle for AI-based methods based on existing technologies is shown; FIG. 4 A system for managing AI models in a wireless system according to embodiments disclosed herein is described; FIG. 5 Multiple modules of a base station of a proposed system according to embodiments disclosed herein are described; FIG. 6 A method for managing AI models in a wireless system according to embodiments disclosed herein is described; FIG. 7 A message sequence diagram depicts a complexity-based AI model training performed using a site-specific dataset according to embodiments disclosed herein, based on the embodiments disclosed herein. FIG. 8A process for enhancing the performance of CSI compression by training a new model, according to embodiments disclosed herein, is described; FIG. 9 A process for enhancing the performance of CSI compression by training a decoder for a shortlisted new AI model, according to embodiments disclosed herein, is described. FIG. 10 A process for enhancing the performance of CSI prediction by training a new AI model, according to embodiments disclosed herein, is described; FIG. 11 A message sequence diagram for performing site-specific dataset collection according to embodiments disclosed herein is depicted; FIG. 12 A process for performing dataset collection conforming to implicit model refinement, according to embodiments disclosed herein, is described; FIG. 13 A flow chart depicts the deployment process for an AI-based method according to embodiments disclosed herein; FIG. 14 A process for refining the CSI compression model for a UE that supports AI model download, according to embodiments disclosed herein, is described; FIG. 15 A process for refining the CSI compression model for a UE that does not support AI model download, according to embodiments disclosed herein, is described. FIG. 16 A process for refining a CSI compression model for a UE that supports AI model downloading, according to embodiments disclosed herein, is described. FIG. 17 A process for refining the CSI compression model for a UE that does not support AI model download, according to embodiments disclosed herein, is described; FIG. 18 A process for improving the performance of CSI compression for UEs with model delivery capability by training a general model, according to embodiments disclosed herein, is described. FIG. 19 A process for improving the performance of CSI compression for UEs that do not have model delivery capability by training a general model, according to embodiments disclosed herein, is described. FIG. 20 A process for improving the performance of CSI prediction by training a predictor, according to embodiments disclosed herein, is described. FIG. 21 The structure of a base station according to an embodiment of the present disclosure is shown; and FIG. 22 The structure of a UE according to an embodiment of the present disclosure is shown.

[0013] Best mode for carrying out the invention Therefore, embodiments of this document provide a method for managing artificial intelligence (AI) models in a wireless system. The method includes: receiving at least one user equipment (UE) capability information from at least one UE by a base station. The method includes: downloading one or more AI models by the base station based on the UE capability information. The method includes: selecting at least one AI model from the downloaded AI models by the base station and selecting a channel state information (CSI) configuration for the UE. Thereafter, the method includes: training the selected AI model by the base station based on the selected CSI configuration to perform at least one CSI operation.

[0014] Therefore, embodiments of this document provide a base station. The base station includes a processor and a memory module. The processor is connected to the memory module. The processor is configured to: receive at least one UE capability information from at least one UE; download one or more AI models based on the UE capability information; select at least one AI model from the downloaded AI models and select a CSI configuration for the UE; and train the selected AI model based on the selected CSI configuration to perform at least one CSI operation.

[0015] These and other aspects of the exemplary embodiments herein will be better appreciated and understood when considered in conjunction with the following description and accompanying drawings. However, it should be understood that while exemplary embodiments and their many specific details are indicated, the following description is given by way of illustration rather than limitation. Many changes and modifications can be made within the scope of the exemplary embodiments herein without departing from their spirit, and the exemplary embodiments herein encompass all such modifications. Detailed Implementation

[0016] The following description, provided with reference to the accompanying drawings, is intended to aid in a full understanding of the various embodiments of this disclosure as defined by the claims and their equivalents. It includes various specific details to aid understanding, but these details are to be considered exemplary only. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the various embodiments described herein without departing from the scope and spirit of this disclosure. Furthermore, descriptions of well-known functions and constructions may be omitted for clarity and brevity.

[0017] The terms and words used in the following description and claims are not limited to their literal meaning, but are used by the inventors only to enable a clear and consistent understanding of this disclosure. Therefore, it will be apparent to those skilled in the art that the following description providing various embodiments of this disclosure is for illustrative purposes only and is not intended to limit the purpose of this disclosure as defined by the appended claims and their equivalents.

[0018] It should be understood that, unless the context clearly specifies otherwise, the singular forms “a,” “an,” and “the” include plural indicators. Thus, for example, a reference to “component surface” includes a reference to one or more such surfaces.

[0019] Before proceeding with the detailed description below, it may be advantageous to define certain words and phrases used throughout this patent document. The term “connection” and its derivatives refer to any direct or indirect communication between two or more elements, whether or not these elements are physically in contact with each other. The terms “transmit,” “receive,” and “communicate,” and their derivatives include both direct and indirect communication. The terms “comprise” and “include,” and their derivatives, mean including but not limited to. The term “or” is inclusive, meaning and / or. The phrase “associated with” and its derivatives mean including, being included in, connected to, interconnected with, containing, being contained within, connected to or connected to, linked to or connected to, able to communicate with, cooperate with, interleaved, juxtaposed, proximate, bound to or bound to, having, possessing the attributes of, having a relationship with, or being related to, etc. The term “controller” means any device, system, or part thereof that controls at least one operation. Such a controller may be implemented in hardware or a combination of hardware and software and / or firmware. The functionality associated with any particular controller may be centralized or distributed, whether local or remote. When used with a list of items, the phrase "at least one of..." means that different combinations of one or more of the listed items can be used, and only one item from the list may be required. For example, "at least one of A, B, and C" includes any of the following combinations: A, B, C, A and B, A and C, B and C, and A and B and C. Similarly, "at least one of A, B, or C" includes any of the following combinations: A, B, C, A and B, A and C, B and C, and A and B and C.

[0020] Furthermore, the various functions described below can be implemented or supported by one or more computer programs, each computer program being formed by computer-readable program code and embodied in a computer-readable medium. The terms "application" and "program" refer to one or more computer programs, software components, instruction sets, procedures, functions, objects, classes, instances, associated data, or portions thereof suitable for implementation in appropriate computer-readable program code. The phrase "computer-readable program code" includes any type of computer code, including source code, object code, and executable code. The phrase "computer-readable medium" includes any type of medium accessible by a computer, such as read-only memory (ROM), random access memory (RAM), hard disk drive, optical disc (CD), digital video disc (DVD), or any other type of storage. "Non-transitory" computer-readable media does not include wired, wireless, optical, or other communication links that transmit transient electrical or other signals. Non-transitory computer-readable media includes media that can permanently store data and media that can store data and subsequently rewrite it, such as rewritable optical discs or erasable memory devices.

[0021] The terminology used herein to describe embodiments of this disclosure is not intended to limit and / or restrict the scope of this disclosure. For example, unless otherwise defined, technical or scientific terms used herein should have the common meaning understood by one of ordinary skill in the art to which this disclosure pertains.

[0022] It should be understood that the terms “first,” “second,” and similar terms used in this disclosure do not indicate any order, quantity, or importance, but are only used to distinguish different components.

[0023] As used herein, any reference to “an example” or “example,” “an implementation” or “implementation,” “an embodiment” or “embodiment” means that a particular element, feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. The phrases “in one embodiment” or “in one example” appearing in different places in the specification do not necessarily refer to the same embodiment.

[0024] As used in this text, “a part” of something means “at least some” of that thing, and therefore can mean less than or all of that thing. Thus, “a part” of something includes the whole thing as a special case, that is, an example where the whole thing is a part of something.

[0025] As used in this article, the term "set" refers to one or more items. Therefore, a combination of items can be a single item or a collection of two or more items.

[0026] In this disclosure, expressions such as "greater than" or "less than" are used by way of example to determine whether a particular condition is met or achieved, and expressions such as "greater than or equal to" or "less than or equal to" are also applicable and not excluded. For example, a condition defined by "greater than or equal to" can be replaced by "greater than" (or vice versa), and a condition defined by "less than or equal to" can be replaced by "less than" (or vice versa), and so on.

[0027] It will also be understood that terms such as “including” or “comprising” mean that an element or object appearing before the term encompasses the listed elements or objects and their equivalents appearing after the term, but do not exclude other elements or objects. Terms such as “connected” or “connected” are not limited to physical or mechanical connections, but may include direct or indirect electrical connections. “Up,” “down,” “left,” and “right” are used only to indicate relative positional relationships, and the relative positional relationships may change accordingly when the absolute position of the objects changes.

[0028] The main objective of the embodiments described herein is to disclose methods and systems for site-specific and implicit AI model refinement in wireless systems.

[0029] Another objective of this embodiment is to disclose a method and system for site-specific AI model refinement for 3GPP systems, wherein dataset collection is performed as part of the 3GPP process of operation, and new models are trained or existing models are refined using the collected site-specific datasets for CSI prediction and CSI compression.

[0030] Another objective of the embodiments herein is to disclose separate methods for training CSI prediction and CSI compression models for UEs with the ability to download new models from base stations, as well as for UEs without this ability.

[0031] Another objective of the embodiments herein is to disclose a method and system for model delivery and execution based on UE capabilities for AI operations.

[0032] Introducing artificial intelligence (AI) into wireless systems promises to improve system capacity and efficiency in 3GPP (3rd Generation Partnership Project) systems. Due to the proven benefits of AI approaches, 3GPP has included a research project on introducing AI-based physical layer processes as part of Rel-18. Currently, in 3GPP AI research on Channel State Information (CSI), the following use cases are of particular focus: - AI-based CSI prediction (e.g.) FIG. 1A and FIG. 1B (as shown); and - AI-based CSI compression (e.g.) FIG. 2A and FIG. 2B (As shown).

[0033] Broadly speaking, the operating modes of AI can be classified as follows: 1) Single-sided operation: The AI ​​model is deployed independently at the base station (BS) or user equipment (UE); for example, CSI prediction.

[0034] 2) Two-sided operation: A pair of AI models, one deployed at the BS and its counterpart deployed at the UE, interactively perform operations; for example, CSI compression.

[0035] The downloaded model can be used on the UE side to perform AI deployment (e.g.) FIG. 1B (As shown). In this scenario, the AI ​​model (in whole or in part) is downloaded from the BS to the UE. In addition to the AI ​​models described above, the UE may already have AI models stored in its memory. Using CSI compression, in this operating mode, after attachment, the UE indicates its ability to perform AI-based operations if new model downloads are supported. The UE receives one or more AI models from the BS to perform CSI encoding. Furthermore, the BS indicates to the UE the AI ​​model to be used for CSI reporting via an AI Model Indicator (AMI). The UE deploys the selected AI model to perform CSI compression. CSI reports based on the selected AI model are shared with the BS until deactivation. In this scenario, the AI ​​model (in whole or in part) is available to the UE but downloading from the BS is not supported. Therefore, downloading from network entities is not involved.

[0036] The downloaded model can be used on the UE side to perform AI deployment (e.g.) FIG. 2B (As shown). In this case, the AI ​​model (in whole or in part) is downloaded from the BS to the UE. In addition to the AI ​​models mentioned above, the UE may already have an AI model stored in its memory. In this case, the AI ​​model (in whole or in part) is available to the UE, but downloading from the BS is not supported. Therefore, downloading from network entities is not involved. After attachment, the UE indicates its ability to perform AI-based operations if it supports the selection of the stored AI model and the stored AI model information. The BS selects the AI ​​model based on UE measurements and shares instructions for selecting the AI ​​model in the UE's memory. The UE selects the model indicated by the BS in its memory and deploys the model to perform CSI coding. AI-based CSI reports are shared with the BS until deactivation.

[0037] Using public datasets to pre-train a general model makes it transparent to site-specific behaviors. This general model provides a good public foundation or starting point for AI-based deployments. It will be deployed in the initial phase of AI-based network deployments, either at the UE storage or BS side. Furthermore, training the general model using public and large datasets helps capture problem-specific features during AI training. FIG. 3AThe diagram illustrates a typical deployment cycle for both non-AI methods and general AI-based models. However, models trained on general data may not capture the full benefits of AI because they lack performance tailored to specific deployment scenarios.

[0038] Compared to using a general-purpose model, AI model customization through the deployment of a dedicated model yields significant gains. Therefore, after deploying AI at a site, the network and UE can continue to fine-tune the AI ​​model in a site-specific manner to improve its performance (e.g., ...). FIG. 3B (As shown). FIG. 3B This illustrates a typical deployment cycle for AI-based methods. Therefore, modem refinement of the AI ​​model is necessary after deployment. Consequently, the training and dataset collection processes are crucial to supporting this refinement process.

[0039] Data collection using dedicated processes incurs additional resource and computational overhead at the business unit (BS). Furthermore, the performance of trained models is correlated with the size of the dataset. Therefore, efficient data collection on a large scale is challenging.

[0040] There are challenges in training AI models and delivering them to UEs. While some UEs are expected to support downloading models from the BS, this capability may not be supported by all UEs in the network. Therefore, such UEs lacking AI model download support may not benefit from site-specific customization.

[0041] Managing the complexity of AI models used for inference at the UE presents challenges. Depending on the UE's capabilities in executing the AI ​​model, a UE attached to the BS may not be able to execute more complex AI models. Therefore, site-specific optimizations may have to take UE complexity into account. Furthermore, AI model complexity is often related to its performance.

[0042] Therefore, there is a need in this field for solutions to overcome the aforementioned shortcomings.

[0043] The embodiments described herein implement the methods and systems necessary for using AI-based operations in deployment scenarios within wireless systems. Reference is now made to the accompanying drawings, and more specifically to... FIG. 4 to FIG. 20 An embodiment is shown, wherein similar reference numerals are consistently used to denote corresponding features throughout the figures.

[0044] FIG. 4 A system 400 is depicted for managing an artificial intelligence (AI) model in a wireless system. System 400 includes a base station 402 and a user equipment (UE) 404. However, the components of UE 404 are not limited to these. For example, UE 404 may include more or fewer components than those described above. Additionally, UE 404 corresponds to... FIG. 22The UE is also included in the base station 402. The base station 402 further includes a processor 406, a communication module 408, and a memory module 410. However, the components of the base station 402 are not limited to these. For example, the base station 402 may include more or fewer components than those described above. Additionally, the base station 402 corresponds to... FIG. 21 The base station.

[0045] In the embodiments described herein, processor 406 can perform UE-based model delivery and execution for AI operations. Processor 406 can perform efficient site-specific dataset collection resulting in minimal resource overhead. Processor 406 can also provide customized AI models to UEs lacking the ability to dynamically deploy newer models. AI models can be customized by training model blocks executed in base station 402. Processor 406 also includes a model download module 502, a model selection module 504, a training module 506, and a CSI decoder module 508, such as... FIG. 5 As shown.

[0046] In the embodiments described herein, the model download module 502 may receive at least one UE capability information from at least one UE 404. The UE capability information may include, but is not limited to, the computational complexity capability of each AI-based function, and the ability to download and deploy new AI models from the base station 402 during at least one CSI operation. The AI-based functions may include, but are not limited to, AI compression, AI prediction, AI model download, and AI model selection. The model download module 502 may download one or more AI models based on the received UE capability information.

[0047] In embodiments described herein, model selection module 504 may select at least one AI model from one or more downloaded AI models. Additionally, model selection module 504 may select a Channel State Information (CSI) configuration for at least one UE 404. The AI ​​model and CSI configuration for UE 404 are selected based on at least one of UE capability information, site-specific dataset collection, one or more scheduling requirements, and one or more AI model training requirements. Scheduling requirements include one or more reporting cycles for UE 404 for AI-based CSI prediction scenarios. AI model training requirements include at least one feedback bit configuration for UE 404 for CSI reporting under AI compression scenarios. In embodiments described herein, model selection module 504 may select at least one new CSI configuration for refinement. In embodiments described herein, model selection module 504 may select at least one existing CSI configuration for refinement. Model selection module 504 selects an existing CSI configuration based on higher accuracy of CSI reports created using site-specific datasets to meet one or more dataset quality requirements. Dataset quality requirements may be site-specific to ensure that the dataset quality requirements meet the minimum accuracy for unimpeded data transmission.

[0048] In the embodiments described herein, if the UE 404 lacks the ability to download new AI models, the model selection module 504 may instruct the UE 404 to select one or more encoder models stored in the UE memory. The model selection module 504 may then select one or more new decoder models from existing encoder-decoder model pairs based on the selected encoder models. When higher accuracy in the CSI report is desired, the model selection module 504 selects a high-complexity level decoder. When moderate accuracy is desired in the CSI report, the model selection module 504 selects a default complexity level decoder.

[0049] In the embodiments described herein, training module 506 can train a selected AI model to perform at least one CSI operation based on the selected CSI configuration. In the embodiments described herein, training module 506 can use site-specific dataset collection to train the selected high-complexity AI model. Training module 506 can also transfer the trained AI model to UE 404 for execution, depending on UE 404's ability to download and deploy one or more new AI models.

[0050] In the embodiments described herein, training module 506 may use site-specific dataset collection to train a portion of a selected AI model executed on the network side. Training module 506 may also deploy the trained portion of the AI ​​model executed on the network side during AI operation, while keeping the AI ​​model on the UE side unchanged. In the embodiments described herein, training module 506 may use site-specific dataset collection to train at least one new AI model for a new CSI configuration. In the embodiments described herein, training module 506 may use site-specific dataset collection to train a new AI model for an existing CSI configuration.

[0051] In the embodiments described herein, training module 506 may identify a set of AI model configurations and corresponding complexity levels for training one or more new AI models based on UE distribution and throughput requirements. The set of AI model configurations by complexity level is a list of the complexity levels and payload sizes of the new AI models to be trained. The complexity level of the new AI model defines the maximum AI model complexity supported by UE 404. A complexity level is defined for at least one of the AI ​​models downloaded from base station 402 and the stored AI models residing in the UE's memory. The complexity level of the new AI model is mapped based on at least one of floating-point operation (FLOP) complexity, inference latency, and combinations of complexity and inference latency.

[0052] In the embodiments described herein, training module 506 can verify whether UE 404 has the capability to download a new AI model based on the received UE capability information. If UE 404 has the capability to download a new AI model, training module 506 can train the encoder and decoder pair of the selected AI model for the UE 404's inclusion complexity and payload size. If the selected CSI configuration is existing CSI compression and UE 404 has the capability to download a new AI model, training module 506 can retrain the encoder and decoder pair of the existing AI model using a site-specific dataset. If the selected CSI configuration is existing UE-side CSI prediction and UE 404 has the capability to download a new AI model, training module 506 can retrain the existing AI model and transmit the retrained existing AI model to UE 404. If the selected CSI configuration is new CSI compression and UE 404 has the capability to download a new AI model, training module 506 can train a new encoder and decoder pair for the new AI model. If the selected CSI configuration is a new UE-side CSI prediction, and UE 404 has the ability to download a new AI model, then training module 506 can train a new CSI predictor.

[0053] In the embodiments described herein, the training module 506 may train one or more new decoder models for one or more corresponding encoder models stored in the UE memory, wherein the new decoder models are selected by the model selection module 504.

[0054] In the embodiments described herein, if the selected CSI configuration is existing CSI compression and the UE lacks the ability to download a new AI model, the training module 506 can use a site-specific dataset to retrain a new decoder model of the existing AI model. If the selected CSI configuration is base station (BS)-side CSI prediction and the UE lacks the ability to download a new AI model, the training module 506 can train a new AI model. If the selected CSI configuration is existing CSI compression and the UE 404 lacks the ability to download a new AI model, the training module 506 can use a site-specific dataset to train a new decoder model for at least one of the given CSI configuration and AI encoder. If the selected CSI configuration is BS-side CSI prediction and the UE 404 lacks the ability to download a new AI model, the training module 506 can use a site-specific dataset to train a new CSI prediction model to perform predictions corresponding to new CSI application time instances.

[0055] In the embodiments described herein, the CSI decoder module 508 may receive a CSI report from the UE 404 while performing at least one CSI operation. The CSI decoder module 508 may decode the CSI report for scheduling purposes. The CSI decoder module 508 may preprocess the decoded CSI report to extract one or more features required for the dataset. The CSI decoder module 508 may also store the preprocessed data in the dataset for future implicit training. The CSI decoder module 508 may repeat the steps from decoding the CSI report to performing site-specific dataset collection until one or more dataset quality requirements are met.

[0056] In the embodiments described herein, processor 406 can process and execute data from multiple modules of base station 402. Processor 406 can be configured to execute instructions stored in memory module 410. Processor 406 may include one or more of a microprocessor, circuitry, and other hardware configured for processing. Processor 406 may be at least one of a single processor, multiple processors, multiple homogeneous or heterogeneous cores, multiple central processing units (CPUs) of different types, a microcontroller, special media, and other accelerators. Processor 406 may be an application processor (AP), a graphics-only unit (such as a graphics processing unit (GPU), a visual processing unit (VPU)), and / or an artificial intelligence (AI)-specific processor (such as a neural processing unit (NPU)).

[0057] In the embodiments described herein, multiple modules of the processor 406 of base station 402 can communicate via communication module 408. Communication module 408 can be in the form of a wired network or a wireless communication network module. The wireless communication network can include, but is not limited to, Global Positioning System (GPS), Global System for Mobile Communications (GSM), Wi-Fi, Bluetooth Low Energy, Near Field Communication (NFC), etc. Depending on the usage environment, the wireless communication can also include one or more of Bluetooth, ZigBee, short-range wireless communication (such as Ultra Wideband (UWB)) and mid-range wireless communication (such as Wi-Fi) or long-range wireless communication (such as 3G / 4G / 5G / 6G and non-3GPP technologies or WiMAX).

[0058] In embodiments herein, memory module 410 may include one or more volatile and non-volatile memory components capable of storing data and instructions for a module of base station 402 to be executed. Examples of memory module 410 may be, but are not limited to, NAND, embedded multimedia card (eMMC), secure digital card (SD), universal serial bus (USB), serial advanced technology accessory (SATA), solid-state drive (SSD), etc. Memory module 410 may also include one or more computer-readable storage media. Examples of non-volatile storage elements may include magnetic hard disks, optical disks, floppy disks, flash memory, or electrically programmable memory (EPROM) or electrically erasable programmable memory (EEPROM). Additionally, in some examples, memory module 410 may be considered a non-transitory storage medium. The term "non-transitory" may indicate that the storage medium is not embodied in a carrier wave or propagating signal. However, the term "non-transitory" should not be construed as meaning that memory module 410 is non-removable. In some examples, non-transitory storage media may store data that can change over time (e.g., in random access memory (RAM) or cache).

[0059] FIG. 4 Example modules of base station 402 are shown, but it should be understood that other embodiments are not limited thereto. In other embodiments, base station 402 may include fewer or more modules. Furthermore, the labels or names of modules are for illustrative purposes only and do not limit the scope of the invention. One or more modules may be combined together to perform the same or substantially similar functions in base station 402.

[0060] FIG. 6 A method 600 for managing AI models in a wireless system is described. Method 600 includes: receiving at least one AI-based UE capability information from at least one UE 404 by a base station 402, as shown in step 602. Method 600 includes: downloading one or more AI models by the base station 402 based on the UE capability information, as shown in step 604. Method 600 includes: selecting at least one AI model by the base station 402 from the downloaded one or more AI models, and selecting a CSI configuration for the UE 404, as shown in step 606. Thereafter, method 600 includes: training the selected AI model by the base station 402 based on the selected CSI configuration to perform at least one CSI operation, as shown in step 608.

[0061] The various actions in method 600 can be performed in the order they are presented, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 6 Some of the actions listed in the document.

[0062] FIG. 7This diagram depicts a message sequence diagram illustrating the complexity-based training of AI models for each CSI configuration using a site-specific dataset. For each CSI configuration, more than one AI model is trained using the site-specific dataset. Based on UE capability information, one or more appropriate AI models are delivered during CSI operations.

[0063] For example, for bilateral CSI compression operations, such as FIG. 7 As shown, once the Radio Resource Control (RRC) configuration request is sent from base station 402 to UE 404 as shown in step 7-2, UE 404 completes the RRC configuration as shown in step 7-4. Shortly thereafter, base station 402 sends a UE capability query to UE 404, as shown in step 7-6. UE 404 responds to base station 402 with a UE capability response, as shown in step 7-8, where the UE capability response includes the computational complexity capability for each AI-based function that enables UE 404 to download and deploy new AI models from base station 402 during at least one CSI operation.

[0064] In the embodiments described herein, the training configuration is selected by base station 402. AI model inference complexity is strongly correlated with inference performance. Therefore, this means that base station 402 should be notified of its AI-related capabilities and complexity information in advance by UE 404 via signaling. For CSI compression and prediction, the embodiments described herein define the following capability elements: AI Compression: The UE supports AI-based CSI compression. AI Prediction: UE supports AI-based CSI prediction AI model download: The UE supports downloading AI models from base station 402.

[0065] AI model selection: The UE supports the selection of at least one AI model from the AI ​​models stored in the UE memory by the base station 402.

[0066] AI Model Complexity Level: The complexity level of the UE 404 used for AI operation execution. This indicates that the maximum AI model complexity supported by the UE404 for the AI ​​model can be: - Downloaded by UE 404 from base station 402; or - Already exists in UE memory AI models are mapped to complexity levels based on the following: -Complexity in FLOP (floating-point arithmetic); - Inference delay (the time required for inference); and - A combination of complexity and latency.

[0067] like FIG. 7As shown, AI compression is indicated as supported, AI prediction is indicated as supported, AI model download is indicated as supported, AI model selection is indicated as supported, and the AI ​​model complexity level is indicated as X. Individual AI models pre-trained to perform the same operation at each UE complexity level are defined as AI model complexity levels. Based on the UE's computational complexity capabilities, base station 402 selects an AI model encoder and transmits it to UE 404 for CSI compression, as shown in steps 7-10 to 7-16.

[0068] The AI ​​model download instruction is sent from base station 402 to UE 404, as shown in steps 7-12. Additionally, CSI configuration and AI model selection are sent from base station 402 to UE 404, as shown in steps 7-14. Channel State Information Reference Signal (CSI-RS) is also sent from base station 402 to UE 404, as shown in steps 7-16. Later, a CSI report is sent from UE 404 to base station 402, as shown in steps 7-18. Base station 402 can train a high-complexity AI model counterpart existing at UE 404 to improve performance. For two-sided models, the complexity distribution performed between UE 404 and base station 402 is balanced or trade-offed as much as possible to optimize performance.

[0069] The embodiments of this paper disclose a method for selecting AI model configurations for training a new model by taking complexity into account. For training a new model, base station 402 first identifies a set of AI model configurations and corresponding complexity levels for which the new model should be trained. This configuration is defined based on a CSI accuracy metric that should be met at the end of a CSI compression or CSI prediction operation.

[0070] The embodiments in this document define the AI ​​model configuration and its identifier for training new AI models. For CSI compression, the requirements depend on: - Payload size A larger payload size improves reporting accuracy, but also increases overhead. -Complexity Level Higher complexity levels improve reporting accuracy, but at the cost of higher UE computational complexity per CSI report.

[0071] -Based on UE distribution and cell throughput requirements, the BS identifies sets of tuples ranked by complexity level. This is a list of complexity levels and payload sizes, which the model will train on.

[0072] For CSI prediction, the requirements depend on the complexity level. (Similar to CSI compression) and future CSI application time examples .For example, This means that we should predict the future. CSI. Similar to CSI compression, base station 402 identifies the set of configurations to be trained for the cell based on cell UE distribution and throughput requirements. .

[0073] FIG. 8 The process 800 for enhancing the performance of CSI compression by training a new model is described. For UEs with model delivery capabilities, a new encoder-decoder pair is trained. This applies to UEs that support the download of AI encoder modules for CSI compression. Therefore, model training involves targeting the enumeration complexity and payload level. To train new encoders and decoders, base station 402 first determines the rank of the AI ​​model based on complexity and payload size to meet CSI reporting accuracy requirements. Then, a separate model is trained for each shortlisted configuration using dataset collection, where an implicit dataset collection procedure is used to gather the dataset. During CSI compression operations, base station 402 uses one of the newly trained compression models with improved performance. Prior to its use, the corresponding encoder model is downloaded by the respective UE for CSI encoding.

[0074] As shown in the figure, the set of CSI compression configurations (i.e., the AI ​​model configuration set) (c, p) for training the new model is calculated, as shown in step 802. Base station 402 checks whether training has been performed on all configurations, as shown in step 804. If training has been performed, base station 402 performs CSI compression using the model encoder used during UE model transmission, as shown in step 818. If training has not been performed, a model configuration for training is selected based on the calculated set of AI model configurations, as depicted in step 806. Later, the encoder-decoder pair of the AI ​​model is trained, and the parameters of both the encoder and decoder are updated, as shown in step 808.

[0075] Subsequently, base station 402 checks whether the enhancement criteria are met, as shown in step 810. If the enhancement criteria are not met, base station 402 continues training, as shown in step 812. If the enhancement criteria are met, base station 402 confirms that training is complete, as shown in step 814. Base station 402 stores the new encoder-decoder model pair, as shown in step 816. After storing the new encoder-decoder model pair, base station 402 checks whether training is complete, as shown in step 804. Base station 402 uses the model encoder from the UE model transmission for CSI compression, as shown in step 818. Subsequently, at base station 402, the decoder is used with updated parameters for CSI compression operations, as shown in step 820.

[0076] The various actions in process 800 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 8 Some of the actions listed in the document.

[0077] FIG. 9 A process 900 is described for enhancing the performance of CSI compression by training a decoder with a shortlisted new AI model. For UE 404, which lacks model delivery capability, a new decoder block can be trained. This new model training is suitable for machines lacking the ability to download new models from base station 402. Therefore, new model training involves: - Identify existing encoder-decoder model pairs for which you want to train new decoder blocks.

[0078] Note that it is assumed that the corresponding encoder model exists in the UE memory and therefore cannot be trained. - Choose the set of complexity and payload to train the new decoder for. .

[0079] In this scenario, base station 402 trains only the decoders of the shortlisted AI models. During CSI compression operations, base station 402 instructs UE 404 to select an encoder. Base station 402 also selects one of the decoders based on complexity level and CSI reporting accuracy requirements. For example, for a given payload, when higher accuracy in CSI reporting is desired, base station 402 may select a high-complexity-level decoder. When moderate accuracy in CSI reporting is desired, base station 402 may select a default-complexity-level decoder.

[0080] As shown in the figure, the set of CSI compression configurations (i.e., the AI ​​model configuration set) (c, p) for training the new model is calculated, as shown in step 902. Base station 402 checks whether training has been performed on all configurations, as shown in step 904. If training has been performed, base station 402 selects the correct decoder based on complexity level and CSI reporting requirements during the CSI compression operation, as shown in step 922. If training has not been performed, the encoder model configuration for training the new decoder is selected based on the calculated set of AI model configurations, as shown in step 906.

[0081] Base station 402 checks whether training has been performed for all decoder complexity levels for the encoder, as shown in step 908. Base station 402 selects the next decoder architecture for training, as shown in step 910. A new decoder design is trained for the model, and the parameters of both the encoder and decoder are updated, as shown in step 912. Thereafter, base station 402 checks whether the enhancement criteria are met, as shown in step 914. If the enhancement criteria are not met, base station 402 continues training, as shown in step 916. If the enhancement criteria are met, base station 402 confirms that training is complete, as shown in step 918. Base station 402 stores the new decoder model pair, as shown in step 920. After storing the new decoder model pair, base station 402 checks whether training is complete, as shown in step 904. After storing the new decoder model pair, as shown in step 922, base station 402 selects the correct decoder based on the complexity level and CSI reporting requirements during the CSI compression operation.

[0082] The various actions in process 900 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 9 Some of the actions listed in the document.

[0083] FIG. 10 A process 1000 for enhancing CSI prediction performance by training a new AI model is described. For a UE 404 with model delivery capability, a new predictor is trained. In this mode, the base station 402 initially selects a set (c, m) of CSI prediction model configurations to be trained, where c represents the set of complexity levels and m represents the set of future CSI application time instances. The base station 402 trains a new CSI predictor model corresponding to the enumerated complexity and payload size. After completion, the newly trained predictor is delivered to the UE 404 to perform UE-side CSI prediction.

[0084] The embodiments described herein can refine predictions without UE model download support. Since CSI prediction is a one-sided AI model performed at UE 404, prediction model updates without UE model updates are not feasible. During CSI operation, base station 402 receives UE capability information to know the complexity levels supported by UE 404. Taking into account the complexity levels and future CSI application window requirements, base station 402 transmits at least one of newly trained predictor models to be used during CSI prediction. Then, for CSI reporting, base station 402 instructs UE 404 that UE 404 should use it to perform the CSI reporting AI by appropriately selecting the complexity level and CSI application time instance for prediction.

[0085] As shown in the figure, the set of CSI prediction configurations (i.e., the AI ​​model configuration set) (c, m) for training the new model is calculated, as shown in step 1002. Base station 402 checks whether training has been performed on all configurations, as shown in step 1004. If training has been performed, base station 402 transmits the newly trained predictor to UE 404 to perform UE-side CSI prediction, as shown in step 1016. If training has not been performed, a new predictor model configuration corresponding to the ingress complexity and payload size is selected, as shown in step 1006.

[0086] Subsequently, base station 402 checks whether the enhancement criteria are met, as shown in step 1008. If the enhancement criteria are not met, base station 402 continues training, as shown in step 1010. If the enhancement criteria are met, base station 402 confirms that training is complete, as shown in step 1012. Base station 402 stores the new predictor model, as shown in step 1014. After storing the predictor model, base station 402 checks whether training is complete, as shown in step 1004. After storing the new predictor model, base station 402 transmits the newly trained predictor to UE 404 to perform UE-side CSI prediction, as shown in step 1016.

[0087] The various actions in process 1000 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 10 Some of the actions listed in the document.

[0088] FIG. 11 A message sequence diagram for performing site-specific dataset collection is depicted. Data sets can be collected using standard 3GPP procedures. For example, consider selecting CSI measurement and reporting configurations based on both scheduler requirements and AI model refinement requirements. As depicted, in the example, CSI compression and reporting are performed periodically by UE 404 according to the requested configuration of base station 402.

[0089] As shown in step 11-2, once the RRC configuration request is sent from base station 402 to UE 404, UE 404 completes the RRC configuration as shown in step 11-4. Shortly afterward, base station 402 sends a UE capability query to UE 404, as shown in step 11-6. UE 404 responds to base station 402 with a UE capability response, as shown in step 11-8. Base station 402 selects the CSI report configuration based on scheduling requirements and dataset collection requirements, as shown in step 11-10.

[0090] In addition, a CSI report configuration is sent from base station 402 to UE 404, as shown in steps 11-12. A CSI-RS signal is also sent from base station 402 to UE 404, as shown in steps 11-14. Later, a CSI report is sent from UE 404 to base station 402, as shown in steps 11-16. The CSI report is decoded as shown in steps 11-18. The decoded CSI report is also used for scheduling and site-specific model refinement, as shown in steps 11-20.

[0091] FIG. 12 A process 1200 for performing dataset collection conforming to implicit model refinement is described. Embodiments herein disclose a process for selecting a valid CSI configuration set for dataset collection using a conventional CSI procedure on the base station 402 side. Base station 402 selects a valid CSI configuration set that is effective for the implicit refinement procedure. This selection is based on the high accuracy of CSI reports created for datasets that meet the dataset quality requirements, for use in further refinement purposes.

[0092] For example, let This is the squared generalized cosine similarity (SGCS) requirement used for training. In a standard CSI codebook... In this context, base station 402 can include a set of configurations that satisfy the squared generalized cosine similarity threshold.

[0093] Similarly, when AI-based compression configurations are supported, When deployed, base station 402 includes subsets, such as:

[0094] For example, in the Rel-16 Type II CSI codebook, paramCombination In the case of composition, it can be obtained from paramCombination 6 and 8 were chosen because of their high-precision reporting performance. .

[0095] The embodiments herein disclose a process for performing implicit model refinement-compliant dataset collection using a conventional CSI procedure on the base station 402 side. When configuring CSI reporting to UE 404 during RRC configuration, base station 402 enables... The shortlist is then determined. Next, a CSI report configuration that meets the accuracy requirements for training, in addition to the mandatory scheduling requirements, is selected. During the reception of UECSI reports by base station 402, the CSI reports are decoded and used for scheduling purposes. Preprocessing is then applied to the received CSIs to extract the features needed for the dataset; for example, a transform is applied to the received pre-encoded data. The preprocessed data is then stored in the dataset for future implicit training. Base station 402 continues the dataset collection process until the dataset requirements are met; for example, the number of samples and their distribution in the training dataset.

[0096] like FIG. 12 As shown, base station 402 shortlists Ψ (the set of CSI configurations valid for the implicit refinement process) as shown in step 1202. Base station 402 checks whether UE 404 is configured for CSI reporting, as shown in step 1204. If UE 404 is configured for CSI reporting, base station 402 selects a CSI reporting configuration based on scheduling, as shown in step 1206, and simultaneously checks for implicit refinement, as shown in step 1208. Later, CSI reporting is configured for UE 404, as shown in step 1212. If implicit refinement is performed, base station 402 selects a CSI reporting configuration from Ψ (based on scheduling and AI training), as shown in step 1210, and configures a CSI report for UE 404, as shown in step 1212. If UE 404 is not configured for CSI reporting and / or if implicit refinement is not performed, base station 402 continuously checks whether UE 404 is configured for CSI reporting.

[0097] Subsequently, base station 402 checks whether a CSI report has been received from UE 404, as shown in step 1214. If a CSI report is received from UE 404, the CSI report is decoded as shown in step 1216, and implicit refinement is checked as shown in step 1218. If no CSI report is received and / or implicit refinement is not performed, base station 402 continuously checks whether a CSI report has been received from UE 404. If implicit refinement is performed, base station 402 stores the received CSI in the dataset as shown in step 1220, and simultaneously completes dataset collection as shown in step 1222. After the dataset has been fully collected, base station 402 checks whether a CSI report has been received.

[0098] The various actions in process 1200 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 12 Some of the actions listed in the document.

[0099] FIG. 13A flow 1300 is described for the deployment process of the AI-based method. Embodiments herein can use site-specific data to train improved AI models to refine model performance for existing CSI configurations, train AI models for new CSI configurations, and improve the performance of UEs that do not support downloading the new model, thereby enhancing the AI ​​block owned by base station 402.

[0100] Consider CSI compression for two-sided AI model operations as an example. In this case, UE 404 holds a set of AI encoder models in its memory, and base station 402 holds a corresponding set of AI decoders. FIG. 13 As shown, the site-specific dataset is collected by base station 402, as shown in step 1302. Base station 402 selects an existing CSI compression model for refinement, as shown in step 1304.

[0101] As shown in step 1306, the selected model is retrained by base station 402 using a site-specific dataset. Furthermore, a new CSI configuration is selected for training the new model, as shown in step 1308. As shown in step 1310, the selected model is trained by base station 402 using a site-specific dataset. Therefore, base station 402 uses the site-specific dataset to train a new AI model for the new CSI configuration or to improve the performance of the existing CSI configuration.

[0102] The various actions in process 1300 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 13 Some of the actions listed in the document.

[0103] FIG. 14 and FIG. 15 This paper describes the process of refining the CSI compression model for UEs that support and do not support AI model downloading. Model refinement refers to improving AI model performance using site-specific datasets by retraining available models for existing CSI configurations and training new AI models for existing or new CSI configurations. The embodiments in this paper can retrain available models for existing CSI configurations. This depends on the UE's ability to download new models.

[0104] Therefore, leveraging UE model transmission support, for CSI compression, the performance of the existing encoder and decoder pair can be enhanced by retraining the encoder and decoder pair and fine-tuning the existing AI model using a site-specific dataset. For UE-side CSI prediction, the AI ​​model is retrained at base station 402 and then transmitted to UE 404.

[0105] In the absence of UE model delivery support, for CSI compression, performance enhancements can be achieved by retraining the decoder portion of an existing AI model using a site-specific dataset, while retaining the same encoder portion already available at UE 404. For CSI prediction, a new AI model can be trained to support CSI prediction only on the base station side.

[0106] FIG. 14 The process 1400 for refining the CSI compressed model for a UE that supports AI model download is described. In the first step 1402, base station 402 loads the CSI compressed model. Base station 402 also loads a site-specific dataset, as shown in step 1404. Base station 402 checks whether all models have been refined, as shown in step 1406, and selects the model for refinement, as shown in step 1408. The encoder-decoder model is then retrained, as shown in step 1410.

[0107] Subsequently, base station 402 selects a retrained AI model for CSI compression, as shown in step 1412. Base station 402 transmits the selected encoder model to UE 404, as shown in step 1414. Later, base station 402 performs CSI reporting, as shown in step 1416.

[0108] The various actions in process 1400 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 14 Some of the actions listed in the document.

[0109] FIG. 15 The process 1500 for refining the CSI compressed model for a UE that does not support AI model download is described. In the first step 1502, base station 402 loads the CSI compressed model. Base station 402 also loads a site-specific dataset, as shown in step 1504. Base station 402 checks whether all models have been refined, as shown in step 1506, and selects the model to be refined, as shown in step 1508. Base station 402 trains only the decoder model, as shown in step 1510.

[0110] Subsequently, base station 402 selects a retrained AI model for CSI compression, as shown in step 1512. Base station 402 transmits the selected encoder model to UE 404, as shown in step 1514. Later, base station 402 performs CSI reporting, as shown in step 1516.

[0111] The various actions in process 1500 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 15 Some of the actions listed in the document.

[0112] The embodiments described herein can train new AI models for existing or new CSI configurations. Depending on the UE model download capability, existing and new CSI configurations can be included for new model training, provided UE model delivery is supported. For CSI compression, new encoder-decoder pairs can be trained for new CSI configurations. For existing CSI configurations, such as payload size, the encoder-decoder, decoder-only, or different complexity levels can be retrained. For UE-side CSI prediction, new CSI predictors can be trained for new CSI configurations, such as new CSI application time instances. For existing CSI configurations, predictors with different complexity levels can be trained.

[0113] In the absence of UE model transmission support, for CSI compression, a new decoder with the same or different complexity is trained using site-specific data for a given CSI configuration or AI encoder. For base station-side CSI prediction, a new CSI prediction model can be trained using site-specific data to perform predictions corresponding to new CSI application time instances.

[0114] FIG. 16 The process 1600 for refining the CSI compression model for a UE that supports AI model download by training a new AI model is described. In the first step 1602, base station 402 identifies the CSI configuration for training. Base station 402 also loads a site-specific dataset, as shown in step 1604. Base station 402 checks whether all models have been refined, as shown in step 1606, and selects the model for refinement, as shown in step 1608. The encoder-decoder model is retrained, as shown in step 1610.

[0115] Subsequently, base station 402 selects a retrained AI model for CSI compression, as shown in step 1612. Base station 402 transmits the selected encoder model to UE 404, as shown in step 1614. Later, base station 402 performs CSI reporting, as shown in step 1616.

[0116] The various actions in process 1600 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 16 Some of the actions listed in the document.

[0117] FIG. 17The process 1700 for refining the CSI compression model for UEs that do not support AI model download is described. In the first step 1702, base station 402 identifies the CSI configuration for training. Base station 402 also loads a site-specific dataset, as shown in step 1704. Base station 402 checks whether all models have been refined, as shown in step 1706, and selects the model for refinement, as shown in step 1708. Base station 402 trains only the decoder model, as shown in step 1710.

[0118] Subsequently, base station 402 selects a retrained AI model for CSI compression, as shown in step 1712. Base station 402 transmits the selected encoder model to UE 404, as shown in step 1714. Later, base station 402 performs CSI reporting, as shown in step 1716.

[0119] The various actions in process 1700 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 17 Some of the actions listed in the document.

[0120] FIG. 18 A process 1800 is described for improving the performance of CSI compression for UEs with model-transmitting capabilities by training a general model. For UEs with model-transmitting capabilities, an encoder-decoder (enc-dec) pair is trained. In this process, base station 402 refines the performance of a pre-trained AI model using a site-specific dataset collected using an implicit data collection process. For CSI compression, both the encoder and decoder pairs are trained with the aim of ultimately transmitting the refined encoder to UE 404 to improve CSI compression.

[0121] First, base station 402 was shortlisted as an AI model for further training. For example, the ensemble of models. This keeps the uplink CSI control overhead within a threshold, i.e. The payload size of each model in the dataset is smaller than the predetermined size. Then, for For each AI model, BS loads the model and trains both the encoder and decoder to improve its performance. During training, intermediate KPIs such as squared generalized cosine similarity (SGCS) are used to track performance improvements.

[0122] Model refinement can be performed in the following ways: -Execute at the base station when the AI ​​training server is unavailable; and - After the model and dataset are sent to the server, the training is performed on the training server.

[0123] During the CSI compression operation, an updated encoder is transmitted to UE 404 for the selected CSI payload size or CSI configuration, and an updated decoder is selected at base station 402.

[0124] As shown in step 1802, base station 402 selects a set of CSI compression models for refinement. Base station 402 checks whether all models have been refined, as shown in step 1804. If refinement is complete, base station 402 performs CSI compression using the encoder with updated parameters during UE model transmission, as shown in step 1818. If refinement is not performed, a model is selected for refinement, as shown in step 1806. Later, the encoder-decoder pair of the AI ​​model is trained, and the parameters of both the encoder and decoder are updated, as shown in step 1808.

[0125] Subsequently, base station 402 checks whether the enhancement criteria are met, as shown in step 1810. If the enhancement criteria are not met, base station 402 continues training, as shown in step 1812. If the enhancement criteria are met, base station 402 confirms that training is complete, as shown in step 1814. Base station 402 stores the updated encoder-decoder model pairs, as shown in step 1816. After storing the updated encoder-decoder model pairs, base station 402 checks whether all models have been refined, as shown in step 1804. Base station 402 performs CSI compression using the model encoder with updated parameters from the UE model transmission, as shown in step 1818. Subsequently, CSI compression is performed at base station 402 using the decoder with updated parameters, as shown in step 1820.

[0126] The various actions in process 1800 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 18 Some of the actions listed in the document.

[0127] FIG. 19 A process 1900 is described for improving the performance of CSI compression for UEs without model transmission capabilities by training a general model. For UEs without model transmission capabilities, only the decoder portion is trained. This variant of CSI compression refinement is for UEs lacking support for downloading new models from the network. Therefore, only the decoder of the existing CSI compression model is trained. First, base station 402 is shortlisted as an AI model for further training.

[0128] For example, a collection of models This keeps the uplink CSI control overhead within a threshold, i.e. The payload size of each model in the dataset is smaller than the predetermined size. Then, for For each AI model in the system, base station 402 loads the model and updates only the decoder parameters during training. The encoder parameters are not updated as part of the training. During training, intermediate KPIs such as squared generalized cosine similarity (SGCS) are used to track performance improvements. The training location can be base station 402 or a third-party parameter server. During CSI compression operations, base station 402 uses the updated decoder model to improve CSI compression performance. However, the encoder at UE 404 remains unchanged.

[0129] As shown in step 1902, base station 402 selects a set of CSI compression models for refinement. Base station 402 checks whether all models have been refined, as shown in step 1904. If refinement is complete, base station 402 performs CSI compression at base station 402 using a decoder with updated parameters, as shown in step 1918. If refinement has not been performed, a model is selected for refinement, as shown in step 1906. Later, the encoder-decoder pair of the AI ​​model is trained, and only the decoder is updated, as shown in step 1908.

[0130] Subsequently, base station 402 checks whether the enhancement criteria are met, as shown in step 1910. If the enhancement criteria are not met, base station 402 continues training, as shown in step 1912. If the enhancement criteria are met, base station 402 confirms that training is complete, as shown in step 1914. Base station 402 stores the updated decoder for the model, as shown in step 1916. After storing the updated decoder, base station 402 checks whether all models have been refined, as shown in step 1904. Base station 402 performs CSI compression at base station 402 using the model decoder with updated parameters, as shown in step 1918.

[0131] The various actions in process 1900 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 19 Some of the actions listed in the document.

[0132] FIG. 20 The process described in section 2000 for improving CSI prediction performance by training a predictor is illustrated. The predictor can be trained for a UE with model delivery capability. In this case, the predictor model is trained at base station 402 using an additional dataset. The enhanced predictor is then delivered to the UE to support improved CSI prediction. First, the BS is shortlisted for CSI prediction configurations for further training. A subset is selected from the set of CSI application time instances that support prediction. Then, for For each predictor AI model in the system, base station 402 loads the model and trains it by updating the predictor's parameters during training. To track performance improvements, intermediate KPIs such as squared generalized cosine similarity (SGCS) are used. During CSI prediction operations: For the CSI prediction configuration, base station 402 first transmits the updated predictor to UE 404. UE 404 then uses the predictor to generate a prediction-based report, which is then shared with base station 402.

[0133] As depicted in step 2002, base station 402 selects a set of CSI prediction models for refinement. Base station 402 checks whether all models have been refined, as shown in step 2004. If refinement is complete, base station 402 performs CSI prediction using a predictor with updated parameters during UE model transmission, as shown in step 2018. If refinement is not performed, a model is selected for refinement, as shown in step 2006. Later, as shown in step 2008, the predictor for the AI ​​model is trained by updating the parameters of the AI ​​model.

[0134] Subsequently, base station 402 checks whether the enhancement criteria are met, as shown in step 2010. If the enhancement criteria are not met, base station 402 continues training, as shown in step 2012. If the enhancement criteria are met, base station 402 confirms that training is complete, as shown in step 2014. Base station 402 stores the updated predictor for the model pair, as shown in step 2016. After storing the updated decoder, base station 402 checks whether all models have been refined, as shown in step 2004. Base station 402 uses the predictor with updated parameters during UE model transmission to perform CSI prediction, as shown in step 2018.

[0135] The various actions in process 2000 can be performed in the presented order, in different orders, or simultaneously. Furthermore, in some embodiments, actions can be omitted. FIG. 20 Some of the actions listed in the document.

[0136] Based on simulations, the ML CSI prediction gain of CSI compression utilizing temporal correlation based on machine learning (ML) is depicted. Table 1 shows the performance of the CSI compression model trained with a site-specific dataset collected using an existing AI model. In the example shown, the dataset was collected using a CSI configuration with a payload size of 564 bits. Using this dataset, base station 402 trained a new model with a payload size of 272 bits. The performance of the newly trained model is approximately 0.9356 SGCS, which is similar to the performance of a model with a centralized dataset collection process.

[0137] Table 1

[0138] Therefore, AI models trained with site-specific data outperform those trained with general datasets. The proposed method provides UE-capability-based model delivery and execution for AI operations. This method offers efficient site-specific dataset collection with minimal resource overhead. Furthermore, to bring the benefits of customized AI models to UEs lacking the ability to dynamically deploy newer models, the proposed method customizes AI models by training only model blocks executed in base station 402. This method is applicable to various scenarios discussed in the 3GPP (3rd Generation Partnership Project), such as CSI prediction, CSI compression, location estimation, and network handover optimization.

[0139] The embodiments disclosed herein can be implemented by at least one software program that runs on at least one hardware device and performs network management functions to control network elements. FIG. 4 The network element shown includes a block that may be at least one of a hardware device or a combination of a hardware device and a software module.

[0140] FIG. 21 The structure of a base station according to an embodiment of the present disclosure is shown.

[0141] like FIG. 21 As shown, the base station according to the embodiment may include a transceiver 2110, a memory 2120, and a processor 2130. The transceiver 2110, memory 2120, and processor 2130 of the base station can operate according to the communication method of the base station described above. However, the components of the base station are not limited thereto. For example, the base station may include more or fewer components than those described above. Furthermore, the processor 2130, transceiver 2110, and memory 2120 may be implemented as a single chip. Furthermore, the processor 2130 may include at least one processor. Additionally, FIG. 21 The base station corresponds to FIG. 4 Base station 402.

[0142] Transceiver 2110 is collectively referred to as a base station receiver and a base station transmitter, and can transmit / receive signals to / from a terminal (UE) or network entity. Signals transmitted to or received from a terminal or network entity may include control information and data. Transceiver 2110 may include an RF transmitter for up-converting and amplifying the transmitted signal, and an RF receiver for low-noise amplification and down-converting the received signal. However, this is only an example of transceiver 2110, and the components of transceiver 2110 are not limited to RF transmitters and RF receivers.

[0143] In addition, transceiver 2110 can receive signals via a wireless channel and output them to processor 2130, and can also transmit signals output from processor 2130 via a wireless channel.

[0144] The memory 2120 can store programs and data required for the operation of the base station. Furthermore, the memory 2120 can store control information or data included in signals acquired by the base station. The memory 2120 can be a storage medium, such as a read-only memory (ROM), random access memory (RAM), hard disk, CD-ROM, and DVD, or a combination of storage media.

[0145] The processor 2130 can control a series of processes to enable the base station to operate as described above. For example, the transceiver 2110 can receive data signals including control signals transmitted by the terminal, and the processor 2130 can determine the result of receiving the control signals and data signals transmitted by the terminal.

[0146] FIG. 22 The structure of a UE according to an embodiment of the present disclosure is shown.

[0147] like FIG. 22 As shown, the UE according to the embodiment may include a transceiver 2210, a memory 2220, and a processor 2230. The transceiver 2210, memory 2220, and processor 2230 of the UE can operate according to the communication method of the UE described above. However, the components of the UE are not limited thereto. For example, the UE may include more or fewer components than those described above. Furthermore, the processor 2230, transceiver 2210, and memory 2220 may be implemented as a single chip. Furthermore, the processor 2230 may include at least one processor. Additionally, FIG. 22 The UE corresponds to FIG. 4 UE 404.

[0148] Transceiver 2210 is collectively referred to as a UE receiver and a UE transmitter, and can transmit signals to or receive signals from a base station or network entity. Signals transmitted to or received from a base station or network entity may include control information and data. Transceiver 2210 may include an RF transmitter for up-conversion and amplification of transmitted signals, and an RF receiver for low-noise amplification and down-conversion of received signals. However, this is only an example of transceiver 2210, and the components of transceiver 2210 are not limited to RF transmitters and RF receivers.

[0149] In addition, transceiver 2210 can receive signals via a wireless channel and output them to processor 2230, and can also transmit signals output from processor 2230 via a wireless channel.

[0150] The memory 2220 can store programs and data required for the operation of the UE. Furthermore, the memory 2220 can store control information or data included in signals received by the UE. The memory 2220 can be a storage medium, such as a read-only memory (ROM), random access memory (RAM), hard disk, CD-ROM, and DVD, or a combination of storage media.

[0151] Processor 2230 can control a series of processes to enable the UE to operate as described above. For example, transceiver 2210 can receive data signals including control signals transmitted by a base station or network entity, and processor 2230 can determine the result of receiving the control signals and data signals transmitted by the base station or network entity.

[0152] Those skilled in the art will understand that the various illustrative logic blocks, modules, circuits, and steps described herein can be implemented in hardware, software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps are generally described above as sets of functions. Whether such sets of functions are implemented in hardware or software depends on the specific application and the design constraints imposed on the system as a whole. Those skilled in the art may implement the described sets of functions in different ways for each specific application, but such design decisions should not be construed as departing from the scope of this application.

[0153] In the embodiments described above in this disclosure, all operations and messages may be selectively performed or omitted. Furthermore, the operations in each embodiment do not need to be performed sequentially, and the order of operations can vary. Messages do not need to be transmitted in order, and the transmission order of messages can be changed. Each operation and transmission for each message can be performed independently.

[0154] Although the accompanying drawings illustrate different examples of user equipment, various changes can be made to the drawings. For example, the user equipment may include any number of each component in any suitable arrangement. Generally, the drawings do not limit the scope of this disclosure to any particular configuration. Furthermore, although the drawings illustrate operating environments in which the various user equipment features disclosed in this patent document can be used, these features can be used in any other suitable system.

[0155] The various illustrative logic blocks, modules, and circuits described in this application may be implemented or executed by a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic, discrete hardware component, or any combination thereof, designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but alternatively, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0156] The steps of the methods or algorithms described in this application can be directly embodied in hardware, software modules executed by a processor, or a combination thereof. The software modules can reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, or any other form of storage medium known in the art. The storage medium is coupled to the processor to enable the processor to read and write information to / from the storage medium. Alternatively, the storage medium can be integrated into the processor. The processor and storage medium can reside in an ASIC. The ASIC can reside in the user terminal. In an alternative, the processor and storage medium can reside as discrete components in the user terminal.

[0157] In one or more designs, these functions can be implemented using hardware, software, firmware, or any combination thereof. If implemented in software, each function can be stored or transmitted as one or more instructions or lines of code on or through a computer-readable medium. Computer-readable media include computer storage media and communication media, the latter including any medium that facilitates the transfer of a computer program from one place to another. Storage media can be any available medium accessible by a general-purpose or special-purpose computer.

[0158] Although this disclosure has been shown and described with reference to various embodiments thereof, those skilled in the art will understand that various changes in form and detail may be made therein without departing from the spirit and scope of this disclosure as defined by the appended claims and their equivalents.

Claims

1. A method performed by a base station (402) in a wireless communication system, the method comprising: receiving at least one UE capability information from at least one user equipment, UE (404); obtaining an AI model based on the at least one UE capability information; selecting at least one AI model among the AI models, and a channel state information, CSI, configuration for the at least one UE (404); and training the at least one AI model for at least one CSI operation based on the CSI configuration.

2. The method (600) of claim 1, wherein the at least one UE capability information comprises at least one of a computational complexity capability for each AI-based function, or a capability to download and deploy a new AI model from the base station (402) during the at least one CSI operation, and wherein the AI-based function comprises at least one of AI compression, AI prediction, AI model download, and AI model selection.

3. The method (600) of claim 1, wherein the at least one AI model and the CSI configuration for the at least one UE (404) are selected based on at least one of the at least one UE capability information, a site-specific dataset collection, at least one scheduling requirement, or at least one AI model training requirement, wherein the at least one scheduling requirement comprises at least one reporting periodicity for the at least one UE (404) for an AI-based CSI prediction scenario, and wherein the at least one AI model training requirement comprises at least one feedback bit configuration for CSI reporting for the at least one UE (404) in an AI compression scenario.

4. The method (600) of claim 1, the method further comprising: training the at least one AI model of high complexity using a site-specific dataset collection; and and transmitting the trained at least one AI model to the at least one UE (404) for execution depending on a capability of the at least one UE (404) to download and deploy at least one new AI model.

5. The method (600) of claim 1, wherein the method (600) comprising: training a portion of the at least one AI model executed at a network side using a site-specific dataset collection; and deploying the trained portion of the at least one AI model executed at the network side during an AI operation while leaving the at least one AI model at a UE side intact.

6. The method (600) of claim 1, the method further comprising: selecting at least one new CSI configuration for refinement; and training at least one new AI model for the at least one new CSI configuration using a site-specific dataset collection.

7. The method (600) of claim 1, the method further comprising: selecting at least one existing CSI configuration for refinement, wherein the base station (402) selects the at least one existing CSI configuration based on a higher accuracy of CSI reporting for a site-specific dataset creation that satisfies at least one dataset quality requirement; and using a site-specific dataset to collect training of at least one new AI model.

8. The method (600) of claim 6, the method further comprising: identifying a set of AI model configurations and corresponding complexity levels for training at least one new AI model based on UE distribution and throughput requirements, wherein the set of AI model configurations by complexity level is a list of complexity levels and payload sizes of the at least one new AI model to train.

9. A base station (402) in a wireless communication system, the base station (402) comprising: a transceiver; and at least one processor (406) coupled with the transceiver and configured to: receive at least one user equipment, UE, (404) capability information from at least one UE (404), obtain AI models based on the at least one UE capability information, select at least one AI model from among the AI models downloaded, and a channel state information, CSI, configuration for the at least one UE (404), and train the at least one AI model for at least one CSI operation based on the CSI configuration.

10. The base station (402) according to claim 9, wherein the at least one UE capability information comprises at least one of a computational complexity capability for each AI-based function, or a capability to download and deploy a new AI model from the base station (402) during the at least one CSI operation, and wherein the AI-based function comprises at least one of AI compression, AI prediction, AI model download, and AI model selection.

11. The base station (402) according to claim 9, wherein the at least one AI model and the CSI configuration for the at least one UE (404) are selected based on at least one of the at least one UE capability information, a site-specific dataset collection, at least one scheduling requirement, and at least one AI model training requirement, wherein the at least one scheduling requirement comprises at least one reporting periodicity for the at least one UE (404) for an AI-based CSI prediction scenario, and wherein the at least one AI model training requirement comprises at least one feedback bit configuration for CSI reporting for the at least one UE (404) in an AI compression scenario.

12. The base station (402) according to claim 9, wherein the at least one processor (406) is further configured to: train the at least one AI model of high complexity using a site-specific dataset collection, and depending on a capability of the at least one UE (404) to download and deploy at least one new AI model, transmit the trained at least one AI model to the at least one UE (404) for execution.

13. The base station (402) according to claim 9, wherein the at least one processor (406) is further configured to: train a portion of the at least one AI model to be executed on a network side using a site-specific dataset collection, and deploy the trained portion of the at least one AI model to be executed on the network side during an AI operation while leaving the at least one AI model on a UE side intact.

14. The base station (402) according to claim 9, wherein the at least one processor (406) is further configured to: selecting at least one new CSI configuration for refinement, and collecting, for the at least one new CSI configuration, site-specific datasets to train at least one new AI model.

15. The base station (402) according to claim 13, wherein The at least one processor (406) is further configured to: select at least one existing CSI configuration for refinement, wherein the base station (402) selects the at least one existing CSI configuration based on a higher accuracy of CSI reporting used to satisfy site-specific dataset creation of at least one dataset quality requirement, and collecting, for the at least one existing CSI configuration, site-specific datasets to train at least one new AI model.

Images