sending a request to update an AI / ML model based on satisfying a particular configured trigger condition

Abstract

A method of operating a communication device is provided. The method includes receiving, from an information processing server via a radio access network, trigger configuration information for triggering an update of an artificial intelligence (AI) model used by the communication device to perform a task. The trigger configuration information includes one or more trigger conditions for triggering the communication device to send a request to update the AI model. Each of the one or more trigger conditions is based on one or more performance indicators that are indicative of a suitability of the AI model to perform the task. The method includes calculating the one or more performance indicators for each of the one or more trigger conditions. The method includes determining, based on the one or more performance indicators calculated for each of the one or more trigger conditions, that the one or more trigger conditions have been satisfied. In response, the method includes sending, to the information processing server via the radio access network, the request to update the AI model.

Description

Technical Field

[0001] This disclosure relates to methods, communication devices, infrastructure equipment for radio access networks, and information processing servers.

[0002] This application claims Paris Convention priority to GB patent application number 2402279.0, filed on February 16, 2024. Background Technology

[0003] The “background” description provided herein is for the purpose of generally presenting the context of this disclosure. To the extent described in this background section, the work of the currently designated inventors and aspects of the description that may not be considered prior art at the time of filing are neither expressly nor implied to be considered prior art to this technology.

[0004] Previous-generation mobile telecommunications systems, such as those based on the UMTS and LTE architectures defined by 3GPP, can support a wider range of services than the simple voice and messaging services offered by previous-generation mobile telecommunications systems. For example, through the improved radio interface and increased data rates provided by LTE systems, users can enjoy high-data-rate applications, such as mobile video streaming and mobile video conferencing, which were previously only available via fixed-line data connections. Therefore, there is a strong demand for deploying such networks, and there is an expectation that the coverage areas of these networks (i.e., geographical locations where network access is available) will increase more rapidly.

[0005] Current and future wireless communication networks are expected to routinely and efficiently support communication with an increasingly diverse range of devices, whose associated data traffic characteristics and types are broader than those optimized for by existing systems. For example, future wireless communication networks are expected to efficiently support communication with various types of devices, including low-complexity devices, machine-type communication (MTC) devices, high-resolution video displays, virtual reality headsets, and extended reality (XR) devices. Some of these different device types can be deployed in large numbers, such as low-complexity devices supporting the Internet of Things (IoT), and are typically associated with relatively small data transmissions and relatively high latency tolerance. Other types of devices (such as devices supporting high-definition video streaming) may be associated with relatively large data transmissions and relatively low latency tolerance. Other types of devices (such as devices used for autonomous vehicle communication and other critical applications) may be characterized by low-latency, high-reliability data transmission over the network. Depending on the application being operated, individual device types may also be associated with different traffic characteristics / features. For example, when a smartphone is used for video streaming applications (high downlink data), the factors that need to be considered to efficiently support its data exchange may be different compared to when it is used for internet browsing applications (intermittent uplink and downlink data) or for voice communication by emergency responders in emergency scenarios (where data must meet strict reliability and latency requirements).

[0006] In light of this, it is expected that current wireless communication networks, such as what may be called 5G or New Radio (NR) systems / New Radio Access Technology (RAT) systems, or future 6G wireless communication and subsequent iterations / versions of existing systems, will be able to effectively support the connectivity of a wide range of devices associated with different applications and the data traffic characteristics and demands with different features.

[0007] One example of the new service is called Ultra Reliable Low Latency Communication (URLLC), which, as the name suggests, requires data units or packets to be transmitted with high reliability and low communication latency. Another example of the new service is Extended Reality (XR), which can be provided by various user devices such as wearable devices. Extended Reality (XR) combines the real world with virtual environments, encompassing aspects such as Augmented Reality (AR), Mixed Reality (MR), and Virtual Reality (VR), and therefore requires high quality and minimal interaction latency. Therefore, services such as URLLC and XR constitute challenging examples for LTE-type communication systems, 5G / New Radio (NR) communication systems, and future generations of communication systems.

[0008] 5G NR continues to evolve, and current work plans include 5G-NR-Advanced, which is expected to achieve several further enhancements, especially to support new use cases / scenarios with higher requirements. The need to support these new use cases and scenarios presents new challenges to efficiently handling communications in wireless communication systems, challenges that need to be addressed. Summary of the Invention

[0009] This disclosure can help resolve or mitigate at least some of the aforementioned problems.

[0010] The appended claims define various aspects and features of this disclosure.

[0011] It should be understood that the foregoing general description and the following detailed description are exemplary and not intended to limit the technology. The described embodiments and other advantages will be best understood by referring to the following detailed description taken in conjunction with the accompanying drawings. Attached Figure Description

[0012] This disclosure will become more readily understood by referring to the following detailed description and taking into consideration the accompanying drawings, in which the same reference numerals indicate the same or corresponding parts in multiple views, and wherein: Figure 1 Some aspects of an LTE-type wireless telecommunications system that can be configured to operate according to some embodiments of this disclosure are illustrated schematically; Figure 2 Some aspects of an NR-type wireless telecommunications system that can be configured to operate according to some embodiments of this disclosure are illustrated schematically; Figure 3 This is a schematic block diagram of exemplary infrastructure equipment and communication apparatus that can be configured to operate according to some embodiments of this disclosure; Figure 4 The diagram illustrates a lifecycle management (LCM) architecture for artificial intelligence models. Figure 5 This illustration shows the deployment of AI / ML positioning. Figure 6A This is a partially illustrative, partial message flow diagram illustrating communication between a communication device, infrastructure equipment of a radio access network, and an information processing server according to an exemplary embodiment. Figure 6B This is a partially illustrative, partial message flow diagram illustrating communication between a communication device, infrastructure equipment of a radio access network, and an information processing server according to an exemplary embodiment. Figure 7 This is a flowchart illustrating model management according to an exemplary implementation; Figure 8 This is a signaling diagram illustrating model management according to an exemplary implementation; Figure 9 This is a signaling diagram illustrating model management according to an exemplary implementation. Detailed Implementation

[0013] Advanced Long Term Evolution (4G) wireless access technology

[0014] Figure 1 A schematic diagram is provided illustrating some basic functions of a mobile telecommunications network / system 6 that typically operates according to LTE principles, but it may also support other radio access technologies and can be adapted to implement embodiments of this disclosure as described herein. Figure 1 The various components and certain aspects of their respective operating modes are well known and defined in the relevant standards managed by the 3GPP (RTM) organization, as well as described in numerous books on the subject (e.g., Holma H. ​​and Toskala A [1]). It should be understood that aspects of telecommunications network operation not specifically described herein (e.g., those related to specific communication protocols and physical channels used for communication between different components) can be implemented according to any known technology, such as in accordance with relevant standards, and proposed modifications and additions to the relevant standards known thereto.

[0015] Network 6 includes multiple base stations 1 connected to core network 2. Each base station provides coverage area 3 (i.e., a cell) within which it can transmit data to and receive data from communication device 4. Although each base station 1... Figure 1 While shown as a single entity, those skilled in the art will understand that some functions of a base station may be performed by different interconnecting elements, such as antennas (antenna assemblies), remote radio heads, amplifiers, etc. One or more base stations may together form a radio access network.

[0016] Data is transmitted from base station 1 to communication device 4 within their respective coverage area 3 via radio downlink (DL). Data is transmitted from communication device 4 to base station 1 via radio uplink (UL). Core network 2 routes data to and from communication device 4 via each base station 1, and provides functions such as authentication, mobility management, and billing. Terminal devices may also be referred to as mobile stations, user equipment (UE), user terminals, mobile radio stations, communication devices, etc. Services provided by core network 2 may include connectivity to the Internet or external telephone services. Core network 2 can further track the location of communication device 4 to efficiently contact (i.e., call) communication device 4, thereby sending downlink data to communication device 4.

[0017] A base station (as an example of a network infrastructure device) may also be referred to as a transceiver station, nodeB, e-nodeB, eNB, g-nodeB, gNB, etc. A base station may also be referred to as a radio access network (RAN) node (e.g., EUTRAN, NG RAN). In this respect, different generations of wireless telecommunication systems are typically associated with different terms for elements providing substantially equivalent functionality. However, some embodiments of this disclosure can be equivalently implemented in different generations of wireless telecommunication systems, and for the sake of brevity, specific terms may be used without being limited to the underlying network architecture. That is, the use of specific terms in some exemplary implementations is not intended to indicate that these implementations are limited to a particular generation of networks most closely associated with that specific term.

[0018] New radio access technology (5G)

[0019] Figure 2 An exemplary configuration of a wireless communication network is shown, which uses some of the terminology proposed and used for NR and 5G. Figure 2 In this configuration, multiple transmission and reception points (TRPs) 10 are connected to distributed control units (DUs) 41, 42 via a connection interface represented by line 16. Each TRP 10 is arranged to transmit and receive signals within the available radio frequency bandwidth of the wireless communication network via a wireless access interface. Therefore, each TRP 10 has a coverage area, as indicated by circle 12, within the range of radio communication via the wireless access interface. Thus, wireless communication devices 14 within the coverage area 12 of each TRP 10 can transmit and receive signals with the TRP 10 via the wireless access interface. Each of the distributed units 41, 42 is connected to a central unit (CU) 40 (which may be referred to as a control node) via interface 46. The central unit 40 is then connected to a core network 20, which may contain all other functions required for data transmission during bidirectional communication with wireless communication devices, and the core network 20 may be connected to other networks 30.

[0020] Figure 2 The components of the wireless access network shown can be configured according to the relevant provisions. Figure 1 The example describes a similar operation to the corresponding components of an LTE network. It should be understood that, in Figure 2 The operational aspects of the telecommunications network represented, and the operational aspects of other networks according to embodiments of this disclosure discussed herein, wherein the parts not specifically described (e.g., regarding specific communication protocols and physical channels for communication between different components) may be implemented according to any known technology, such as methods currently used for implementing such operations of wireless telecommunications systems, for example, according to relevant standards.

[0021] Figure 2 The TRP 10 in the new RAT network may partially possess functions corresponding to a base station, eNodeB, or gNB in ​​an LTE network. Similarly, the communication device 14 may possess functions corresponding to a UE device 4 known for operation with an LTE or NR network. Therefore, it should be understood that operational aspects of the new RAT network (e.g., specific communication protocols and physical channels for communication between different components) may differ from known operations in LTE, NR, or other known mobile telecommunications standards. However, it should also be understood that each of the core network components, base stations, and communication devices in the new RAT network will functionally resemble, respectively, the core network components, base stations, and communication devices of an LTE or NR wireless communication network.

[0022] In terms of a wide range of high-level functions, connecting to Figure 2 The core network 20 of the new RAT telecommunications system can be roughly considered to be the same as... Figure 1 Corresponding to the core network 2, and each central unit 40 and its associated distributed unit / TRP 10 can be considered to provide a similar service to... Figure 1 At least some of the functions corresponding to base station 1. The term "network infrastructure equipment / access node" can be used to encompass these elements as well as elements of more traditional base station types in wireless telecommunications systems. Depending on the specific application, the responsibility for scheduling transmissions arranged on the radio interfaces between the distributed units and communication devices may be borne by the control node / central unit and / or the distributed unit / TRP. Figure 2 In this context, communication device 14 is represented within the coverage area of ​​the first coverage area 12. Therefore, communication device 14 can exchange signaling with the first central unit 40 within the first coverage area 12 via one of the distributed units / TRP10 associated with the first coverage area 12.

[0023] It should also be understood that Figure 2 This is merely an example of the proposed architecture for a telecommunications system based on the new RAT, in which the methods described herein can be employed, and the functionality disclosed herein can also be applied to wireless telecommunications systems with different architectures.

[0024] Therefore, some embodiments of this disclosure discussed herein can be implemented according to various different architectures (such as, Figure 1 and Figure 2The exemplary architecture shown is implemented in a wireless telecommunications system / network. Therefore, it should be understood that the specific wireless telecommunications architecture is not of primary significance to the principles described herein in any particular implementation. In this regard, some embodiments of this disclosure are generally described in the context of communication between network infrastructure equipment / access nodes and communication devices, wherein the specific form of the network infrastructure equipment / access nodes and communication devices will depend on the network infrastructure in the specific implementation scenario. For example, in some scenarios, the network infrastructure equipment / access node may include a base station (such as...) suitable for providing functions consistent with the principles described herein. Figure 1 The LTE-type base station 1 shown, in other examples, network infrastructure equipment may include a central unit / control node 40 and / or a central unit / control node 40 adapted to provide functions consistent with the principles described herein. Figure 2 The TRP10 shown is of the type shown.

[0025] Figure 3 Provided Figure 2 A more detailed illustration of some components of the network is shown. Figure 3 In, such as Figure 2 The TRP 10 shown (in simplified form) includes a wireless transmitter 30, a wireless receiver 32, and a controller or control processor 34 operable to control the transmitter 30 and the wireless receiver 32 to transmit and receive radio signals to one or more UEs 14 within the coverage area 12 formed by the TRP 10. Figure 3 As shown, the exemplary UE 14 is shown to include a corresponding transmitter 49, receiver 48 and controller 44, the controller being configured to control the transmitter 49 and receiver 48 to transmit signals representing uplink data to the wireless communication network via the wireless access interface formed by TRP 10, and to receive downlink data (such as signals transmitted by transmitter 30 and received by receiver 48) in normal operation.

[0026] Transmitters 30, 49 and receivers 32, 48 (and other transmitters, receivers, and transceivers described in connection with the examples and embodiments of this disclosure) may include radio frequency filters, amplifiers, and signal processing components and means for transmitting and receiving radio signals, for example, according to the 5G / NR standard. Controllers 34, 44 (and other controllers described in connection with the examples and embodiments of this disclosure) may be, for example, microprocessors, CPUs, or dedicated chipsets, configured to execute instructions stored on a computer-readable medium, such as non-volatile memory. The processing steps described herein may be executed, for example, by a microprocessor in conjunction with random access memory, and the microprocessor operates according to instructions stored on the computer-readable medium. For ease of illustration, the transmitters, receivers, and controllers are shown in... Figure 3These are schematically shown as independent components. However, it should be understood that the functions of these components can be implemented in a variety of different ways, such as using one or more appropriately programmed programmable computers, or one or more appropriately configured application-specific integrated circuits / circuit systems / chips / chipsets. It should be understood that infrastructure equipment / TRP / base stations and UE / communication devices typically include a variety of other components related to their operational functions.

[0027] like Figure 3 As shown, TRP 10 also includes a network interface 50, which is connected to DU 42 via physical interface 16. Therefore, network interface 50 provides a communication link for transmitting data and signaling from TRP 10 to core network 20 via DU 42 and CU 40.

[0028] The interface 46 between DU 42 and CU 40 is referred to as the F1 interface, which can be a physical interface or a logical interface. The F1 interface 46 between CU and DU can operate according to 3GPP TS 38.470, 3GPP TS 38.473, and 3GPP TS 38.401 specifications and can be formed by fiber optic or other wired or wireless high-bandwidth connections. In one example, the connection 16 from TRP 10 to DU 42 is implemented via fiber optic. The connection between TRP 10 and core network 20 is typically referred to as backhaul, which includes the interface 16 from network interface 50 of TRP 10 to DU 42, and the F1 interface 46 from DU 42 to CU 40. Core network 20 is connected to CU 40 via the N2 (also known as NG-C) interface for carrying control data and via the N3 (also known as NG-U) interface for carrying user data.

[0029] Core network 20 may include core network functions, such as location management function (LMF) for managing the location of wireless communication devices in the wireless communication network. In addition, core network 20 may include one or more network functions for managing, training, and storing artificial intelligence / machine learning (AI / ML) models.

[0030] Although references have been made to 4G / LTE and 5G NR above, it should be understood that this disclosure applies to subsequent wireless communication technologies, including 6G. In the case of 6G, a base station (as an example of network infrastructure equipment) may also be referred to as a 6G NB (6G Node B), a 6G RAN node, etc. In 6G, the core network 20 can be one or more network functions.

[0031] Artificial Intelligence (AI)

[0032] In existing technologies (also known as "conventional technologies"), a UE or gNB can determine its location based on positioning measurements performed by the gNB or UE. Positioning measurements can be performed by the UE on downlink signals such as a Positioning Reference Signal (PRS), or by the gNB on uplink signals such as a Sounding Reference Signal (SRS). Positioning measurements can include one or more of time of arrival (TOA), angle of arrival (AOA), angle of departure (AOD), round-trip time (RTT), reference signal received power (RSRP), and any other measurements used in determining the UE's location. A UE can determine its location, for example, by receiving multiple downlink signals (each from a different gNB), measuring the time of arrival and / or angle of the downlink signals, and determining the UE's location based on these measurements (e.g., multilateral positioning).

[0033] However, the accuracy of location determination using existing technologies is limited. To address the technical challenges in determining the precise location / position estimation of communication devices, the use of artificial intelligence / machine learning (AI / ML) models for positioning was first explored in 3GPP standard version 18, and further developed in subsequent versions.

[0034] AI / ML-based positioning often outperforms other positioning technologies because AI / ML models are configured to collect and process large amounts of positioning measurements. By performing model training (based on input data), AI / ML models learn and build an understanding of the environment associated with the positioning measurements. Subsequently, AI / ML models can be deployed, for example at a user equipment (UE) or a base station (e.g., a gNB), to generate outputs such as improved (i.e., more accurate) positioning measurements or location estimates. This process / operation is also known as AI / ML model inference. For example, output positioning measurements can be improved in the sense that they are more likely to lead to more accurate location estimates. For example, a UE can receive a DL-PRS from a gNB and perform positioning measurements on it. However, the UE may not know that the DL-PRS is reflected by an object before reaching the UE. AI / ML models can be configured to correct for such reflections. AI / ML models are able to generate outputs with refined LOS / NLOS component identifiers. For more details on how AI / ML models can be used to generate more accurate positioning measurements or generate location estimates, see reference [2], the contents of which are incorporated herein by reference in their entirety.

[0035] As an example, AI / ML models can utilize one or more of supervised learning, generative AI, autoencoders, and reinforcement learning.

[0036] Supervised learning

[0037] AI / ML models can implement supervised machine learning models.

[0038] Supervised learning models are trained using labeled training data to learn a function that maps inputs (typically provided as feature vectors) to outputs (i.e., labels). Labeled training data consists of pairs of inputs and corresponding output labels. Output labels are typically provided by the operator to indicate the expected output for each input. The supervised learning model processes the training data to generate an inference function that can be used to map new (i.e., unseen) inputs to labels.

[0039] Input data (during training and / or inference) can include various types of data, such as numerical values, images, videos, text, or audio. The raw input data can be preprocessed to obtain appropriate feature vectors for use as input to the model; for example, features can be extracted from image or text input to obtain the corresponding feature vectors. It should be understood that the type of input data and the preprocessing techniques used (if necessary) can be chosen based on the specific task the supervised learning model is intended for.

[0040] Once ready, the labeled training dataset is used to train the supervised learning model. During training, the model adjusts its internal parameters (e.g., weights) to optimize (e.g., minimize) the error function, which aims to minimize the difference between the model's predicted output and the labels provided as part of the training data. In some cases, the error function may include a regularization penalty to reduce overfitting of the model to the training dataset.

[0041] Supervised learning models can use one or more machine learning algorithms to learn the mapping between their inputs and outputs. Suitable exemplary learning algorithms include linear regression, logistic regression, artificial neural networks, decision trees, support vector machines (SVM), random forests, and the K-nearest neighbors algorithm.

[0042] For localization AI / ML models, the input to the supervised learning model may include localization measurements (e.g., measurements of one or more PRS or one or more SRS, such as time of arrival and angle of arrival), and the output of the supervised learning model may include location estimates of the communication device or improved localization measurements.

[0043] Once trained, a supervised learning model can be used for inference, i.e., to predict outputs for previously unseen input data. Supervised learning models can perform classification and / or regression tasks. In classification tasks, the supervised learning model predicts discrete class labels for input data and / or assigns input data to predetermined classes. In regression tasks, the supervised learning model predicts labels as continuous values.

[0044] For localization AI / ML models, supervised learning models are used to predict outputs that include improved localization measurements or location estimates based on inputs that include localization measurements.

[0045] In some cases, the amount of labeled data available for model training may be limited (e.g., due to the high cost or difficulty in implementing data labeling). In such cases, supervised learning models can be extended to further utilize unlabeled data and / or generate labeled data.

[0046] If unlabeled data is required, the training data can include both labeled and unlabeled training data, and semi-supervised learning can be used to learn the mapping between the model input and output. For example, graph-based methods such as Laplacian regularization can be used to extend the SVM algorithm to Laplacian SVM to perform semi-supervised learning on partially labeled training data.

[0047] To generate labeled data, an active learning model can be used, where the model actively queries information sources (such as users or operators) to add labels to data points in the desired output. Typically, labels are requested only for a subset of the training dataset, thus reducing the amount of labels required compared to fully supervised learning. The model can choose the examples for which labels are requested—for instance, it can request labels for data points that are most likely to change the current model or that are most likely to reduce the model's generalization error. A semi-supervised learning algorithm can then be used to train the model on a portion of the labeled dataset.

[0048] Generative AI

[0049] AI / ML models can enable generative artificial intelligence (AI).

[0050] Generative AI systems learn patterns and structures in their input training data in order to subsequently generate new output data that exhibits similar characteristics to the training data. For localization AI / ML models, the input training data may include localization measurements, and the output training data may include location estimates or improved localization measurements from communication devices.

[0051] Generative AI systems can generate output data based on input prompts. Prompts can include different types of data, such as images, videos, text, or audio. Prompts can have the same or different data types as the model's training and / or output data.

[0052] Generative AI / ML models can include generative models that are trained to learn the probability distribution of input training data and generate new output data based on that learned distribution. For example, given a set of data instances / observable variables (X) and a set of labels / target variables (Y) in a training dataset, the generative model can learn the joint probability distribution p(X,Y) of the data instances and labels, and / or the probability distribution p(X) of the data instances (e.g., when no labels are available).

[0053] Suitable exemplary generative models for learning the probability distribution of input training data include variational autoencoders (VAEs), Transformer-based models, diffusion models (e.g., denoising diffusion probabilistic models (DDPM)), reinforcement learning (RL), and generative adversarial networks (GANs). The choice of generative model can depend on the specific task performed by the generative AI / ML.

[0054] Generative models can include one or more artificial neural networks. For example, a variational autoencoder (VAE) can include a pair of neural networks that act as an encoder and decoder, respectively, for encoding training data into a reduced-dimensional representation (i.e., a latent space representation) and recovering the original data from that representation. A generative adversarial network (GAN) can include a first “generator” neural network and a second “discriminator” neural network, wherein the first “generator” neural network generates new data, and the second “discriminator” neural network learns to distinguish generated data from real data. The one or more component neural networks of a generative model can be trained jointly or individually.

[0055] During training, the generative model can adjust its internal parameters (e.g., neural network weights) to optimize (e.g., minimize) the loss / error function, aiming to minimize the difference between the generated output data and the desired output data. It should be understood that the specific loss function and the algorithm used to optimize it can vary depending on the nature of the generative model and its intended application. For example, the mean squared error loss function can be used for image generation tasks, and the cross-entropy loss function can be used for text generation tasks. These loss functions can be optimized using various existing optimization algorithms (e.g., gradient descent).

[0056] Once training is complete, the generative model can generate new output data based on input prompts. These prompts can be provided by the user or by the appropriate device (e.g., using an Application Programming Interface (API)). Therefore, generative AI / ML models allow for the generation of new output data based solely on prompts, without requiring detailed instructions.

[0057] Self-encoder

[0058] AI / ML models can be self-encoded.

[0059] An autoencoder is an unsupervised machine learning model that uses one or more artificial neural networks to learn an efficient representation of unlabeled input data. Autoencoders can be used to encode various types of data, such as images, videos, text, audio, or location measurements.

[0060] An autoencoder may include an encoder neural network that encodes input data into a reduced-dimensional representation (also known as a "latent space"), and a decoder neural network designed to recreate the input data from the encoded reduced-dimensional representation. The latent space typically has a lower dimension than the input data; therefore, the latent space generated by the encoder usually provides a more efficient, compressed representation of the input data that requires less memory storage than the original input data.

[0061] An encoder neural network can include one or more layers that transform input data into a reduced-dimensional representation. The encoder neural network receives input data, and the last layer of the encoder neural network outputs a reduced-dimensional representation of the input data, namely the latent space (also known as the "bottleneck layer").

[0062] A decoder neural network consists of one or more layers that transform data from a latent space into output data with the same dimensions as the data input to the encoder. The decoder aims to reconstruct the data initially input to the encoder neural network from the latent space representation of the data.

[0063] Encoder and / or decoder neural networks typically include multiple hidden layers. For example, an encoder may include multiple hidden layers that progressively extract further reduced representations of the input data. Using deeper neural networks (i.e., with a greater number of hidden layers) for the encoder and / or decoder can improve the performance of the autoencoder and, in some cases, reduce the amount of training data required.

[0064] Encoder and decoder neural networks are typically trained jointly. During training, the autoencoder can adjust its internal parameters (e.g., the weights and biases of the encoder and decoder neural networks) to optimize (e.g., minimize) a loss / error function aimed at minimizing the difference between the data input to the encoder and the data reconstructed from the output generated by the decoder. It should be understood that the specific loss function and the algorithm used to optimize it can vary depending on the nature of the autoencoder model and its intended application. In the example, a mean squared error loss function optimized by gradient descent can be used. In some cases, sparse autoencoders can be used to promote sparsity of the latent space representation (relative to the input) and prevent the autoencoder from learning an identity function—for example, a sparse autoencoder can be implemented by modifying the loss function to include a sparsity regularization penalty.

[0065] In some cases, an autoencoder can be a variational autoencoder (VAE). A VAE is a specific type of autoencoder in which the training process imposes a probabilistic model on the encoded representation (the training process penalizes deviations from this probabilistic model). VAEs can be used in generative artificial intelligence applications to generate new output data that exhibits similar features to the input encoded data by sampling from the learned latent space.

[0066] For location AI / ML models, input data may include location measurements, and output data may include location estimates or improved location measurements from communication devices.

[0067] reinforcement learning

[0068] AI / ML models can implement reinforcement learning (RL).

[0069] Reinforcement learning is a type of machine learning that aims to train an artificial intelligence agent to take actions in an environment to maximize cumulative rewards. During reinforcement learning, the agent interacts with the environment and learns from the consequences of its actions, allowing the agent to progressively improve its decision-making.

[0070] RL models typically include an action-reward feedback loop. The feedback loop includes: environment, state, agent, policy, action, and reward. The environment is the system in which the agent interacts and operates—for example, the environment could be the virtual environment of a video game. The state represents the current conditions in the environment. The agent receives the state as input and takes actions that may affect the environment and change its state. The agent takes actions based on its policy, which is a mapping from the environment state to the agent's actions. The policy can be deterministic or stochastic. The reward represents the environment's feedback to the agent's actions. The reward provides an indication (usually in numerical form) of the desirability of the agent's actions. Rewards can include positive signals for rewarding desirable behavior and / or negative signals for punishing undesirable behavior.

[0071] Through multiple iterations of the action-reward feedback loop, the agent aims to maximize its total cumulative reward, thus learning how to take the best action in the environment. Therefore, reinforcement learning allows the agent to learn the optimal policy that maximizes cumulative reward. A value function can be used to estimate the cumulative reward, which estimates the expected return obtainable from a given state, or a given state and action. Using cumulative reward in reinforcement learning allows the agent to consider the long-term effects of its policy.

[0072] Reinforcement learning algorithms can be used to optimize the agent's policy and value function through multiple iterations of the action-reward feedback loop. This learning algorithm can be based on an environment model (e.g., Markov Decision Process (MDP)) or a model-free architecture. Suitable exemplary model-free reinforcement learning algorithms include Q-learning, SARSA (State-Action-Reward-State-Action), Deep Q-Network (DQN), or Deep Deterministic Policy Gradient (DDPG).

[0073] It should be understood that agents typically explore and exploit their operating environment. During exploration, agents usually take random actions to gather information about the environment and identify potentially desirable actions (i.e., actions that maximize cumulative rewards). During exploitation, agents take actions that are expected to maximize rewards (e.g., actions chosen based on the agent's latest policy). Various techniques can be used to control the ratio of exploratory to exploitative actions taken by the agent—for example, a predetermined probability of taking an exploratory action in a given feedback loop iteration can be set (and optionally reduced over time to allow the agent to shift more towards exploitation over time, thereby maximizing cumulative rewards despite diminishing returns from further exploration).

[0074] In some cases, RL models can be configured to learn from user-provided feedback. Utilizing user feedback in this way allows the agent to improve its action choices and better align with user preferences. For example, reinforcement learning from human feedback (RLHF) can be used. RLHF involves training a reward model based on user feedback and using that model to determine the reward in the aforementioned reinforcement learning process. User feedback can be received in various forms depending on the specific reinforcement learning problem being solved—for example, feedback could be received as a user's ranking of instances of agent actions. Therefore, RLHF allows for the incorporation of user feedback into the reinforcement learning process. RLHF methods can be advantageous in situations where users are better able to evaluate the quality of the machine learning model's output than the algorithm itself (e.g., for generative AI RL models).

[0075] For location AI / ML models, input data may include location measurements, and output data may include location estimates or improved location measurements from communication devices.

[0076] Although various types of AI / ML models have been described above in conjunction with positioning, the types of AI / ML described above can also be used for beam management or CSI tasks, or in fact for other tasks known to those skilled in the art.

[0077] For beam management AI / ML models, input data may include one or both of the L1-RSRP and its associated beam / resource ID. Output data may include one or more of the following: predicted beam, narrow beam, improved beam prediction, or refined RSRP.

[0078] For CSI AI / ML models (such as CSI measurement and reporting, CSI compression, or CSI prediction), input data may include one or more of the following: signal-to-interference-plus-noise ratio (SINR) estimate, modulation and coding (MCS) index, channel quality indicator (CQI) index, or precoding matrix index (PMI). Output data may include one or more of the following: refined or predicted MCS, CQI index, PMI index, or RI index. In some examples of CSI compression AI / ML models, the input data may be a complete CSI compression report, and the output data may be a compressed CSI report.

[0079] AI / ML model lifecycle management (LCM) architecture

[0080] 3GPP has defined a general AI / ML framework for the NR air interface to support various RF applications, including positioning applications, as well as beam management and CSI feedback enhancement applications. The framework aims to cover a common architecture that addresses the entire AI model lifecycle, namely lifecycle management (LCM), including data collection, model training, etc. The LCM for AI / ML in the NR air interface is described below. This description also applies to 6G.

[0081] Examples of lifecycle management (LCM) architectures for artificial intelligence models include... Figure 4 As illustrated, the Figure 4 This is copied from reference [2], the content of which is incorporated into this paper in its entirety through citation. Figure 4 As shown, the LCM architecture includes data collection function 402, inference function 404, management function 406, model training function 408, and model storage function 410.

[0082] Data collection function 402 is configured to provide training data to model training function 408, monitoring data to management function 406, and inference data to inference function 404. Data collection function 402 is a process / function by which network nodes, management entities, or UEs collect data for the purposes of AI / ML model training, data analysis, and inference.

[0083] Inference function 404 is configured to provide inference output to management function 406, receive management instructions from management function 406, and receive models from model storage function 410. Inference function 404 is a process / function that uses a trained AI / ML model to produce a set of outputs based on a set of inputs.

[0084] The management function is configured to provide performance feedback and / or retraining requests to the model training function 408, and to provide model delivery requests to the model storage unit 410.

[0085] The model training function 408 is configured to provide updated models to the model storage unit 410. The model training function 408 is a process / function of training an AI / ML model in a data-driven manner (e.g., by learning input / output relationships) and obtaining a trained AI / ML model for inference.

[0086] like Figure 4 As shown, management function 406 represents the core of the LCM architecture. The function of management function 406 is to monitor model performance at different entities (such as the UE or gNB) and request updated models. Model management includes two processes—model switching and model updating.

[0087] Existing positioning processes in wireless communication networks (such as 5G NR networks) involve communication between the UE, gNB, and LMF. AI / ML models can be deployed on the UE, gNB, or LMF. The AI / ML model used for positioning can be a "direct AI / ML positioning model" or an "AI / ML-assisted positioning model." Both direct AI / ML positioning models and AI / ML-assisted positioning models use positioning measurements (e.g., measurements based on received PRS or SRS) as input data. However, the output data of a direct AI / ML positioning model is an estimate of the UE's location (such as coordinates), while the output data of an AI / ML-assisted positioning model is a more accurate positioning measurement or a new set of positioning measurements. Figure 5 Examples of existing localization techniques using direct and assisted AI / ML localization models are shown in the figure.

[0088] Figure 5 Segment (A) illustrates an example of UE-based positioning using a direct AI / ML model on the UE side. In this example, one or more gNBs send a PRS to the UE. The UE performs positioning measurements on the PRS and uses a direct AI / ML model to generate an estimate of the UE's location based on the positioning measurements. The UE forwards the location estimate to the LMF.

[0089] Figure 5 Section (B) illustrates an example of UE-assisted / LMF-based positioning using a UE-side AI / ML-assisted positioning model. In this example, one or more gNBs send a PRS to the UE. The UE performs positioning measurements on the PRS and generates improved positioning measurements of the PRS using the AI / ML-assisted positioning model. The UE forwards the improved measurements of the PRS to the LMF. The LMF generates a location estimate for the UE based on the improved measurements of the PRS.

[0090] Figure 5 Section (C) illustrates an example of LMF-based positioning using a direct AI / ML model on the LMF side. In this example, one or more gNBs send a PRS to the UE. The UE performs positioning measurements on the PRS and sends the positioning measurements from the PRS to the LMF. The LMF then uses a direct AI / ML model based on the positioning measurements from the PRS to generate a location estimate for the UE.

[0091] Figure 5 Segment (D) illustrates an example of NG-RAN node-assisted localization using a gNB-side AI / ML-assisted localization model. In this example, the UE sends an SRS to one or more gNBs. The gNB performs localization measurements on the DRS and uses the AI / ML-assisted localization model to generate further improved localization measurements based on the SRS. The gNB then sends the improved localization measurements of the SRS to the LMF. The LMF generates a location estimate for the UE based on the improved localization measurements of the SRS.

[0092] Figure 5 Segment (E) illustrates NG-RAN node-assisted positioning using an LMF-side direct AI / ML model. In this example, the UE sends an SRS to one or more gNBs. The gNB performs positioning measurements on the SRS and sends these measurements to the LMF. Based on the received positioning measurements from the SRS, the LMF uses a direct AI / ML model to generate a location estimate for the UE.

[0093] AI / ML localization models can be deployed for specific regions, while different AI / ML localization models can be deployed for different regions. In this way, AI / ML localization models can be adapted to be effective in the regions where they are deployed. For example, a particular AI / ML model can be trained under specific environmental conditions, as these conditions will be similar to those where the AI / ML model is deployed. Having multiple AI / ML models, each adapted to different regions, also facilitates model switching and selection, thereby enabling high-precision localization in diverse regions.

[0094] However, environmental conditions in the area where AI / ML positioning models are deployed can change over time. For example, the position (e.g., layout or configuration) of a communication device relative to passive environmental objects (such as tables or walls) in the area may change over time. Since passive environmental objects can reflect / scatter signals transmitted by the communication device, changes in the position of the communication device relative to passive environmental objects mean that different signals will be reflected / scattered in different ways over time. This can affect channel conditions and propagation characteristics in the area, such as line-of-sight (LOS). Consequently, AI / ML models adapted to a particular area may become outdated over time. If an AI / ML model becomes outdated, the accuracy of positioning measurements or positioning estimates generated by the AI / ML model decreases to unacceptable levels. The performance degradation of AI / ML models over time may be due to data drift and / or model drift.

[0095] Data drift occurs when the distribution of input data (e.g., location measurements) fed into an AI / ML model changes over time. This happens, for example, when the distribution of location measurements obtained by a UE deviates from the distribution of location measurements used to train the AI / ML model. For instance, when training an AI / ML model, it might be assumed that UEs are uniformly distributed throughout a room. However, in practice, or in reality over a given time period, UEs tend to remain in sub-regions of the room (e.g., most UEs are likely closer to the center of the room). This deviation between the assumed UE statistics (the distribution of location and positioning measurements) during training and the actual UE statistics can impair overall performance.

[0096] Model drift occurs when the relationship between the input and output data of an AI / ML model changes. This relationship may change due to variations in environmental conditions. For example, a room may have different reflectors, such as tables and chairs. The positions of these reflectors may not always be fixed. Consider a scenario where the arrangement of tables and chairs in the room changes, but the user interface (UE) still uses the same AI / ML model for localization; this leads to a model drift problem and impaired localization performance.

[0097] Therefore, the performance degradation of AI / ML problems over time represents a technical issue.

[0098] Therefore, improved methods, communication devices, infrastructure equipment, and information processing servers are needed to address at least some of the aforementioned problems.

[0099] Figure 6AA partially schematic, partial message flow diagram of a wireless communication system according to an exemplary embodiment is shown. The wireless communication system includes a communication device 601 (e.g., UE 14), an infrastructure device 602 of a radio access network (e.g., gNB), and an information processing server 603 (e.g., an LMF server, a dedicated AI server, or another server configured to communicate with an LMF server). In some embodiments, the information processing server 603 is a server in the core network, and the infrastructure device 602 forms part of the radio access network through which the communication device can access the core network.

[0100] Communication device 601 includes transceiver 601.1 (or transceiver circuitry) and controller 601.2 (or controller circuitry). Infrastructure device 602 includes transceiver 602.1 (or transceiver circuitry) and controller 602.2 (or controller circuitry). Information processing server 603 includes transceiver 603.1 (or transceiver circuitry) and controller 603 (or controller circuitry). Transceivers 601.1, 602.1, and 603.1 are configured to transmit and receive signals. Transceivers 601.1, 602.1, and 603.1 (or transceiver circuitry) may each include an independent transmitter or receiver (or independent transmitter and receiver circuitry), or transceivers 601.1, 602.1, and 603.1 (or transceiver circuitry) may each include means (or circuitry) configured to perform both transmitting and receiving. Each of the controllers 601.2, 602.2, and 603.2 can be, for example, a microprocessor, a CPU, or a dedicated chipset (system-on-a-chip, SoC).

[0101] Communication device 601 is configured to communicate with infrastructure device 602 via a radio access interface using wireless communication signals. Infrastructure device 602 is configured to communicate with information processing server 603 using signals that can be transmitted wirelessly, via a wired connection, or a combination of both. Communication device 601 is configured to communicate with information processing server 603 via infrastructure device 602 through a radio access network.

[0102] like Figure 6A As shown, the controller 603.2 of the information processing server 603 is configured to control the transceiver 603.1 to send trigger configuration information 604 to the communication device 601 via the infrastructure equipment 602 of the radio access network. Therefore, the controller 601.2 of the communication device 601 is configured to control the transceiver 601.1 of the communication device 601 to receive the trigger configuration information.

[0103] The trigger configuration information is used to trigger an update of the artificial intelligence (AI) model used by the communication device 601 to perform a task. This task can be, for example, a positioning task for locating the communication device 601, a beam management task, a channel state information (CSI) measurement and reporting task, a CSI compression task, or a CSI prediction task. Alternatively or additionally, the task can be another task performed by the AI / ML model.

[0104] In some implementations, the AI ​​model may be a machine learning model. Machine learning AI models will be referred to herein as "AI / ML models." While "AI / ML models" will be mentioned in conjunction with exemplary embodiments, this disclosure is not limited thereto, and AI models that do not implement machine learning may be used.

[0105] The trigger configuration information includes one or more trigger conditions for triggering the communication device 601 to send a request to update the AI ​​model. Each of the one or more trigger conditions may be based on one or more performance metrics that indicate the suitability of the AI ​​model to perform the task.

[0106] The triggering condition may include one or more thresholds, which, if met, trigger the communication device 601 to send a request to update the AI ​​model. For example, the threshold may be a confidence level of positioning accuracy (such as a percentage confidence level). The calculated performance metric can provide an indication of the positioning accuracy achieved by the communication device 601. In another example, the threshold may be a statistic representing the maximum acceptable deviation in positioning measurements over a period of time.

[0107] The controller 601.2 of the communication device 601 is configured to control the communication device 601 to calculate one or more performance metrics for each of one or more trigger conditions that can be configured by trigger configuration information.

[0108] The controller 601.2 of the communication device 601 is configured to control the communication device 601 to determine that one or more trigger conditions have been met based on one or more performance metrics calculated for each of one or more trigger conditions that can be configured by trigger configuration information.

[0109] In response to determining that one or more triggering conditions have been met, the controller 601.2 of the communication device 601 is configured to control the transceiver 601.1 of the communication device 601 to send a request 607 to update the AI ​​model to the information processing server 603 via the infrastructure device 602 of the radio access network. Alternatively, from the perspective of the communication device 601 (e.g., the UE), the destination of the request sent by the communication device 601 may be the infrastructure device 602. In this case, the infrastructure device 602 may transmit the request or information corresponding to the request to the information processing server 603.

[0110] According to an exemplary embodiment, after the information processing server 603 receives a request to update the AI ​​model, the controller 603.2 of the information processing server 603 can control the information processing server 604 to update the AI ​​model. Updating the AI ​​model may include updating one or more parameters of the AI ​​model and / or updating the structure of the AI ​​model. Updating the AI ​​model may include retraining the AI ​​model and / or fine-tuning the AI ​​model. For example, the AI ​​model information processing server 603 may obtain updated training data and use the updated training data to retrain the AI ​​model. The information processing server 603 may determine how to obtain the updated training data. For example, the information processing server 603 may send a request to other communication devices (such as PRUs) within the area where the AI / ML model is applicable to collect updated training data (such as new positioning measurements). The communication devices may report the updated training data to the information processing server 603 for use by the information processing server 603 in updating the AI / ML model.

[0111] According to an exemplary embodiment, the controller 603.2 of the information processing server 603 can control the transceiver 603.1 of the information processing server 603 to send an instruction for an updated AI model to the communication device 601. The communication device 601 can then use the updated AI model to perform tasks (e.g., positioning tasks, beam management tasks, CSI tasks, or other tasks performed by the AI / ML model).

[0112] Therefore, exemplary implementations can trigger a communication device to send a request to update the AI ​​model based on one or more performance metrics that indicate the suitability of the AI ​​model for performing a task. Thus, exemplary implementations can update the AI ​​model before it becomes unsuitable for performing the task due to degradation over time (e.g., due to model drift or data drift). Therefore, the performance of the AI ​​model in the wireless communication network is improved.

[0113] Figure 6B A partial schematic, partial message flow illustration of a wireless communication system according to an exemplary embodiment is shown. The wireless communication system includes infrastructure equipment 602 of a radio access network (e.g., a gNB) and an information processing server 603 (such as an LMF server, a dedicated AI server, or another server configured to communicate with an LMF server). Reference has been made. Figure 6A The configuration of infrastructure device 602 and information processing server 603 is described, but for the sake of brevity, it will not be repeated here.

[0114] like Figure 6BAs shown, the controller 603.2 of the information processing server 603 is configured to control the transceiver 603.1 to send 608 trigger configuration information to the infrastructure device 602 of the radio access network. Therefore, the controller 602.2 of the infrastructure device is configured to control the transceiver 602.1 of the infrastructure device 602 to receive the trigger configuration information.

[0115] The trigger configuration information is used to trigger the update of the artificial intelligence (AI) model used by infrastructure device 602 to perform a task. This task can be, for example, a positioning task for locating communication device 601, a beam management task, or a channel state information (CSI) task.

[0116] The trigger configuration information includes one or more trigger conditions for triggering infrastructure device 602 to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics that indicate the suitability of the AI ​​model for performing the task.

[0117] The controller 602.2 of the infrastructure device 602 is configured to control the infrastructure device 602 to calculate one or more performance metrics for each of one or more trigger conditions that can be configured by trigger configuration information.

[0118] The controller 602.2 of the infrastructure device is configured to control the infrastructure device 602 to determine that one or more trigger conditions have been met based on one or more performance metrics calculated for each of one or more trigger conditions that can be configured by trigger configuration information.

[0119] In response to determining that one or more triggering conditions have been met, the controller 602.2 of the infrastructure device 602 is configured to control the transceiver 602.1 of the infrastructure device 602 to send a request 611 to the information processing server 603 to update the AI ​​model.

[0120] According to an exemplary embodiment, after the information processing server 604 receives a request to update the AI ​​model, the controller 603.2 of the information processing server 604 can control the information processing server 604 to update the AI ​​model. Updating the AI ​​model may include updating one or more parameters of the AI ​​model and / or updating the structure of the AI ​​model. Updating the AI ​​model may also include retraining the AI ​​model and / or fine-tuning the AI ​​model. For example, the AI ​​model information processing server can obtain updated training data and use the updated training data to retrain the AI ​​model.

[0121] According to an exemplary embodiment, the controller 603.2 of the information processing server 604 can control the transceiver 603.1 of the information processing server 604 to send instructions on an updated AI model to the communication device 601. The infrastructure device 602 can then use the updated AI model to perform tasks (e.g., positioning tasks, beam management tasks, CSI tasks, or another task performed by the AI / ML model).

[0122] Therefore, exemplary implementations can trigger infrastructure devices in a radio access network to send requests to update the AI ​​model based on one or more performance metrics that indicate the suitability of the AI ​​model for performing a task. Thus, exemplary implementations can update the AI ​​model before it becomes unsuitable for performing the task due to degradation over time (e.g., due to model drift or data drift). Consequently, the performance of the AI ​​model in the wireless communication network is improved.

[0123] The following exemplary embodiments will refer to the UE, gNB, and LMF server for illustrative purposes only. Unless otherwise stated, the exemplary embodiments described with reference to the UE are generally applicable to communication devices, the exemplary embodiments described with reference to the gNB are generally applicable to infrastructure equipment of radio access networks, and the exemplary embodiments described with reference to the LMF server are generally applicable to information processing servers.

[0124] The exemplary implementation is applicable to scenarios where direct or auxiliary AI / ML models are deployed on the UE or gNB. Example scenarios to which the exemplary implementation is applicable include: - UE-based localization using direct AI / ML models on the UE side (e.g., Figure 5 Section A) - UE-assisted / LMF-based positioning using UE-side AI / ML-assisted positioning models (e.g., Figure 5 Section B) - NG-RAN node-assisted localization using gNB-side AI / ML-assisted localization models (e.g.) Figure 5 (Section D) According to an exemplary embodiment, a method is provided for updating an AI / ML model performing a positioning task in a radio node (such as a communication device of a UE, or an infrastructure device of a radio access network such as a gNB). As described above, an information processing server sends trigger configuration information to trigger the update of the artificial intelligence (AI) model used by the radio node to perform the task. In some embodiments, the trigger configuration information includes an indication in a format for calculating one or more performance metrics for each of one or more trigger conditions. In some embodiments, the one or more trigger conditions are based on comparing the performance metrics of the trigger conditions with a threshold. In such embodiments, the threshold adopts the same format as indicated in the trigger configuration information for calculating the performance metrics.

[0125] In some implementations, one or more trigger conditions include a condition regarding the expiration time of the AI / ML model. For example, the information processing server may determine, based on prior experience, that the AI / ML model is no longer suitable for performing a task (e.g., a location task) after its expiration time. This could be due to model drift or data drift over time. In such implementations, the calculation of performance metrics may include calculating the time the AI / ML model has been used, and the trigger condition is that the time the AI / ML model has been used is greater than or equal to the expiration time. In cases where the information processing server expects the performance of the AI / ML model to degrade over time, the usage time of the AI / ML model serves as an indication of its suitability for performing the task.

[0126] In some implementations, the trigger configuration information includes an indication of the computation trigger time for the AI / ML model. In such implementations, in response to a communication device or infrastructure device determining that the computation trigger time has arrived, one or more performance metrics are calculated for each of one or more trigger conditions. In such implementations, power consumption can be saved because the communication device or infrastructure device only needs to calculate the performance metrics when the computation time arrives.

[0127] In some implementations, once a radio node determines that a triggering condition has been met, it can send an update request immediately, after a predefined interval, or periodically. In some implementations, the triggering configuration information includes an indication of the time interval at which a request to update the AI / ML model should be sent. This time interval may be referred to as the AI / ML update request computation time, AI / ML update request processing time, etc.

[0128] In some implementations, trigger configuration information may include instructions for instructing one or more radio nodes to calculate one or more performance metrics. For example, trigger configuration information may include instructions to a first radio node that the first radio node should calculate a performance metric indicating data drift. Trigger configuration information sent from an information processing server to another radio node may include instructions for calculating a performance metric indicating model drift. In such implementations, the second radio node may be a Positioning Reference Unit (PRU).

[0129] In some implementations, one or more performance metrics for one or more trigger conditions include performance metrics indicating data drift of the input data of the AI / ML model. Data drift may be caused by UE behavior. For example, UEs may all tend to stay in the center of a room rather than be evenly distributed throughout the room. Performance metrics indicating data drift can measure changes in the statistical properties of the model input (e.g., positioning measurements). Correcting data drift does not require a truth (GT) label. Since mobile UEs may not be able to accurately obtain their location from other sources, correcting for data drift is easier than correcting for model drift (which requires a GT label). Therefore, using one or more performance metrics indicating data drift for one or more trigger conditions is particularly advantageous for mobile UEs. Examples of performance metrics indicating data drift include: i. Variations in the mean, variance, bias, or other relevant statistics of the committed information rate (CIR), time of arrival (TOA), and reference signal received power (RSRP) captured at different time periods.

[0130] ii. Statistical distance metrics, such as the Kolmogorov-Smirnov (KS) statistic, are used to quantify the differences in the distribution of location measurements over different time periods.

[0131] To calculate a performance metric indicating data drift, a radio node can perform multiple location measurements over multiple time periods. For example, a UE can perform 100 location measurements in a day and calculate a reference statistic S1 based on these measurements. On each subsequent day, the UE can perform another 100 location measurements and measure a statistic Sn, where n is the date the location measurements were performed. The performance metric can be the difference between the statistics calculated from the location measurements each day, such as |Sn-S1|. In one example, the trigger condition could be that the UE sends a request to update the AI / ML model when |Sn-S1| is first determined to exceed a threshold X. For example, if on day 4, the statistic S4 satisfies |S4-S1|>X, the UE sends a request to update the AI / ML model. In another example, the statistic |Sn-S1| can be averaged over a period of time, and the UE sends a request to update the AI / ML model when the average exceeds a threshold X.

[0132] In some implementations, one or more performance metrics for one or more trigger conditions include performance metrics indicating model drift of the AI / ML model's input data. Typically, model drift is caused by environmental changes in the area where the AI / ML model is applicable. Performance metrics indicating model drift can measure changes in the relationship between positioning measurements and AI / ML model output data, such as improved positioning measurements or location estimates. Correction of model drift requires GT tags. Therefore, for fixed UEs with fixed, known locations (e.g., Positioning Reference Units (PRUs) or UEs capable of obtaining accurate location estimates from other sources), correction of model drift is easier.

[0133] Examples of performance metrics that indicate model drift include: i. A statistic indicating the difference between the model output (prediction) and the GT label (in other words, localization accuracy).

[0134] ii. Statistics indicating whether there is a sudden drop or fluctuation in positioning accuracy.

[0135] iii. A / B testing metrics. A / B metrics can be one or more of the following: localization accuracy, F1 score, precision, and recall of the AI / ML model.

[0136] For direct AI / ML localization models, both the model output and the ground truth (GT) label are location estimates for the UE (such as the UE's coordinates). For assisted AI / ML localization models, the model output and the GT label can be accurate localization measurements (such as LOS / NLOS indications, TOA, or RSRP).

[0137] When multiple performance metrics need to be calculated, radio nodes can calculate the metrics simultaneously or calculate each metric separately at different times. Furthermore, the calculation of performance metrics can be performed by a specific radio node. For example, a performance metric indicating model drift can be calculated by a UE (such as a PRU).

[0138] In some implementations, the request to update the AI / ML model includes an identifier of the AI / ML model currently used by the communication device. For example, the ID could be a global ID of the AI / ML model. In some implementations, the request to update the AI / ML model includes an indication of one or more performance metrics calculated for a met trigger condition. In some implementations, where the met trigger condition is based on multiple performance metrics calculated by a radio node, the update request may include indications of multiple performance metrics calculated by the node. The indication may be a calculated value of the performance metric. Alternatively or additionally, the indication may be an indication of a calculated performance metric or one or more indices associated with a calculated performance metric. In such implementations, the indication of each performance metric may be sent simultaneously or at different times.

[0139] In response to receiving an update request, the information processing server updates the AI / ML model. For example, the information processing server (e.g., its model training unit) can initiate a data collection event and retrain or fine-tune the AI / ML model based on the collected data. The updated model can be distributed to the radio nodes and the model storage unit of the information processing server to replace outdated AI / ML models.

[0140] Figure 7 This is a flowchart illustrating a model management method according to an exemplary implementation. The method is performed by a radio node (such as a UE or gNB). The method begins at step S702.

[0141] In step S704, the radio node receives trigger configuration information (e.g., an LMF server) from the information processing server. Figure 7 In the example shown, trigger configuration information is used to update the AI / ML model used by the radio node to perform the positioning task. The trigger configuration information includes one or more trigger conditions that trigger the communication device to send a request to update the AI / ML model. Each of the one or more trigger conditions is based on one or more performance metrics that indicate the suitability of the AI / ML model for performing the positioning task.

[0142] In step S706, the radio node calculates one or more performance metrics for each of the one or more trigger conditions. Figure 7In the example shown, the radio node calculates a performance metric for the triggering condition based on channel and location measurements. For example, the performance metric could be a statistic indicating the deviation between channel and location measurements over time.

[0143] In step S708, the radio node compares the performance metric with a second threshold (Threshold2). The second threshold represents a threshold deviation of the channel and positioning measurements that is higher than the threshold deviation of the channel and positioning measurements represented by a first threshold (Threshold1). If the radio node determines that the performance metric is greater than the second threshold, the method proceeds to step S710. In step S170, the radio node uses conventional positioning techniques (i.e., positioning techniques that do not use AI / ML models) to determine the UE's location. The high deviation indicated by the performance metric indicates that conventional techniques will provide a more accurate location estimate compared to techniques using AI / ML models, because the high deviation indicates that the currently used AI / ML model is not working effectively.

[0144] If the radio node determines that the performance metric is less than or equal to the second threshold, the method proceeds to step S712. In step S712, the radio node compares the performance metric with the first threshold. If the radio node determines that the performance metric is greater than the first threshold, the method proceeds to step S714. In other words, the radio node determines that the triggering condition for sending an update request to the information processing server has been met.

[0145] In step S714, the radio node sends a request to the information processing server to update the AI / ML model. Figure 7 In the example shown, the radio node also sends the calculated performance metrics to the information processing server.

[0146] If, in step S712, the radio node determines that the performance metric is less than or equal to a first threshold, the method proceeds to step S716, where the method terminates. In other words, the radio node determines that the AI / ML model is currently sufficiently suitable for performing the localization task and therefore does not require updating.

[0147] Figure 8 This is a signaling diagram illustrating model management according to an exemplary implementation. Figure 8 In the example, the AI / ML model is the UE-side model.

[0148] In step S802, the information processing server sends the AI / ML model for performing the positioning task to the UE via the gNB. In other words, the gNB transparently transmits the received AI / ML model to the UE. In step S804, the server sends the trigger configuration information to the UE via the gNB. In other words, the gNB transparently transmits the received trigger configuration information to the UE. Figure 8In the example shown, the trigger configuration information is used to trigger an update of the AI / ML model sent to the UE in step S802. The trigger configuration information includes one or more trigger conditions for triggering the UE to send a request to update the AI / ML model. Each of the one or more trigger conditions is based on one or more performance metrics that indicate the suitability of the AI / ML model to perform a positioning task.

[0149] In step S806, the gNB sends one or more downlink location reference signals (e.g., DL-PRS) to the UE. In some implementations, one or more other gNBs also send one or more DL-PRS to the UE. In step S808, the UE performs positioning measurements on the received DL-PRS. Positioning measurements may include, for example, TOA measurements and / or AOA measurements. The UE calculates a performance metric based on the positioning measurements (e.g., a statistic indicating the deviation of the positioning measurements over time). In one example, the UE may acquire one hundred positioning measurements in a day and calculate a reference distribution represented by statistic S1 based on the measurements. On the next day, the UE may perform another one hundred positioning measurements and calculate a distribution represented by statistic S2. The UE can then calculate the performance metric based on the difference between S2 and S1. Since the performance metric is based on the deviation of the input data of the AI / ML model, the performance metric indicates data drift.

[0150] In step S810, the UE determines whether the triggering condition has been met based on the calculated performance metric. If the triggering condition has been met (e.g., the performance metric exceeds the threshold indicated in the triggering condition), the UE then sends a request to update the AI / ML model to the information processing server via the gNB in ​​step S812.

[0151] In step S814, the information processing server updates the AI / ML model by, for example, retraining or fine-tuning the model.

[0152] In step S816, the information processing server sends the updated AI / ML model to the UE via the gNB.

[0153] In step S818, the UE deploys the updated AI / ML model and uses the model to perform future positioning tasks.

[0154] Figure 9 This is a signaling diagram illustrating model management according to an exemplary implementation. Figure 9 In the example, the AI / ML model is the UE-side model. Figure 9In this system, the information processing server includes a model management unit, a model training unit, and a model storage unit. These units may be located in the same or different devices. One or more of these units may be located within an LMF server.

[0155] In step S902, the UE sends a request for an AI / ML model for the positioning task to the model management unit.

[0156] In step S904, the model management forwards the AI / ML model request to the model storage unit.

[0157] In step S906, the model storage unit sends the requested AI / ML model to the UE via the gNB. In other words, the gNB transparently transmits the received AI / ML model to the UE.

[0158] In step S908, the model management unit sends the trigger configuration information to the UE via the gNB. In other words, the gNB transparently transmits the received trigger configuration information to the UE. Figure 9 In the example shown, the trigger configuration information is used to update the AI / ML model sent to the UE in step S906. The trigger configuration information includes one or more trigger conditions for triggering the UE to send a request to update the AI / ML model. Each of the one or more trigger conditions is based on one or more performance metrics that indicate the suitability of the AI / ML model to perform a positioning task.

[0159] In step S910, the gNB sends one or more downlink location reference signals (e.g., DL-PRS) to the UE. In some implementations, one or more other gNBs also send one or more DL-PRS to the UE. In step S912, the UE performs positioning measurements on the received DL-PRS. The UE calculates a performance metric based on the positioning measurements (e.g., a statistic indicating the deviation of the positioning measurements over time). Since the performance metric is based on the deviation of the input data of the AI / ML model, it indicates data drift.

[0160] In step S912, the UE determines whether the triggering condition has been met based on the calculated performance metric. If the triggering condition has been met (e.g., the performance metric exceeds the threshold indicated in the triggering condition), the UE then sends a request to update the AI / ML model to the model management unit via the gNB in ​​step S916.

[0161] In step S918, the model management unit forwards the update request to the model training unit.

[0162] In step S920, the model training unit updates the AI / ML model (e.g., by retraining and / or fine-tuning the AI / ML model).

[0163] In step S922, the model training unit sends the updated AI / ML model to the UE via the gNB.

[0164] In step S924, the model training unit sends the updated AI / ML model to the model storage unit for storage.

[0165] In step S926, the UE deploys the updated AI / ML model and uses the model to perform future positioning tasks.

[0166] In step S928, the model storage unit updates the AI / ML model in the storage unit to the updated AI / ML model. For example, the model storage unit may delete the previously stored AI / ML model and store the updated AI / ML model.

[0167] Although Figure 8 and Figure 9 The UE-side AI / ML model is described, in which the UE calculates performance metrics based on positioning measurements from DL-PRS. However, the exemplary implementation is also applicable to cases where a gNB-side AI / ML model exists, in which the gNB calculates performance metrics based on positioning measurements from uplink signals (such as SRS).

[0168] In some implementations, the information processing server can determine whether one or more trigger conditions for updating the AI / ML model are met based on calculated performance metrics received from radio nodes (such as communication devices or infrastructure equipment of a radio access network). In such implementations, the radio nodes periodically send indications of one or more performance metrics calculated by the radio nodes to the information processing server. Indications of one or more performance metrics can be sent periodically without receiving any triggers or trigger conditions for sending the one or more performance metrics to the radio nodes. For example, the radio nodes may not receive instructions to send performance metrics, nor may they receive trigger conditions instructing them to send performance metrics when conditions are met. In implementations where the information processing server determines whether one or more trigger conditions for updating the AI / ML model are met, the information processing server does not send any indications to the radio nodes. Figure 6A and Figure 6B The description of the trigger configuration information.

[0169] Those skilled in the art should understand that adjustments can be made based on the implementation of this technology. Figures 6A to 9 The method shown. For example, such methods may include additional intermediate steps, or the steps may be performed in any logical order. Although it has been passed Figures 6A to 9 The exemplary systems and methods shown illustrate implementations of the present technology, but it will be apparent to those skilled in the art that they are equally applicable to other systems besides those described herein.

[0170] Those skilled in the art will further recognize that, based on the various arrangements and implementations discussed in the foregoing paragraphs, such infrastructure equipment and / or communication devices as defined herein can be further defined. Those skilled in the art will further understand that such infrastructure equipment and communication devices as defined and described in this disclosure can form part of a communication system other than the communication system defined in this disclosure.

[0171] The following numbered paragraphs provide further exemplary aspects and features of this technology: Paragraph 1. A method of operating a communication device, the method comprising: Trigger configuration information is received from an information processing server via a radio access network. This trigger configuration information is used to trigger an update of the artificial intelligence (AI) model used by the communication device to perform a task. The trigger configuration information includes one or more trigger conditions for triggering the communication device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics that indicate the suitability of the AI ​​model for performing the task. One or more performance metrics are calculated for each of the one or more trigger conditions. Based on one or more performance metrics calculated for each of the one or more trigger conditions, it is determined that one or more trigger conditions have been met, and this is taken as a response. A request to update the AI ​​model is sent to the information processing server via a radio access network.

[0172] Paragraph 2. According to the method described in paragraph 1, the information processing server includes one or more of the following: Location Management Function (LMF) server, AI servers, and The server configured to communicate with the LMF server.

[0173] Paragraph 3. The method according to paragraph 1 or paragraph 2, wherein one or more performance metrics for one or more triggering conditions include performance metrics that indicate data drift of the input data of the AI ​​model.

[0174] Paragraph 4. The method according to any one of paragraphs 1 to 3, wherein one or more performance metrics for one or more of the triggering conditions include performance metrics that indicate model drift of the AI ​​model.

[0175] Paragraph 5. The method according to any one of paragraphs 1 to 4, wherein the trigger configuration information includes an indication in a format for calculating one or more performance metrics for each of one or more trigger conditions.

[0176] Paragraph 6. According to the method described in paragraph 5, one or more of the triggering conditions are based on comparing the performance metric of the triggering condition with a threshold.

[0177] Paragraph 7. According to the method described in paragraph 6, wherein the indication of the threshold of one or more trigger conditions is included in the trigger configuration information.

[0178] Paragraph 8. The method described in paragraph 6 or paragraph 7, wherein the threshold is in the same format as that used to calculate the performance metric.

[0179] Paragraph 9. The method according to any one of paragraphs 1 to 8, wherein one or more of the triggering conditions include conditions relating to the expiration time of the AI ​​model.

[0180] Paragraph 10. The method according to any one of paragraphs 1 to 9, wherein the triggering configuration information includes an indication of the time interval at which a request to update the AI ​​model should be sent.

[0181] Paragraph 11. The method according to any one of paragraphs 1 to 10, wherein the request to update the AI ​​model includes an identifier of the AI ​​model currently used by the communication device.

[0182] Paragraph 12. The method according to any one of paragraphs 1 to 11, wherein the request to update the AI ​​model includes an indication of one or more performance metrics calculated for a met trigger condition.

[0183] Paragraph 13. The method according to paragraph 12, wherein the satisfied triggering conditions are based on multiple performance metrics calculated by the communication device, and the method includes: Instructions for multiple performance indicators calculated by the communication device are sent to an information processing server via a radio access network, wherein each of the multiple performance indicators calculated by the communication device is sent at a different time.

[0184] Paragraph 14. The method according to any one of paragraphs 1 to 13 comprises: The AI ​​model is received from the information processing server via a radio access network.

[0185] Paragraph 15. The method according to any one of paragraphs 1 to 14, wherein the task performed by the AI ​​model is a positioning task for determining the location of a communication device, or a beam management task, or a channel state information (CSI) measurement and reporting task, a CSI prediction task, or a CSI compression task.

[0186] Paragraph 16. According to the method described in paragraph 15, the task performed by the AI ​​model is a localization task, and the input data of the AI ​​model includes localization signal measurements, and the AI ​​model is configured to generate a more accurate estimate of the localization signal measurements based on the input data, or to generate an estimate of the location of a communication device based on the input data.

[0187] Paragraph 17. The method according to any one of paragraphs 1 to 16, wherein calculating one or more performance metrics for each of one or more triggering conditions comprises: Use AI models to calculate one or more performance metrics.

[0188] Paragraph 18. A method for operating infrastructure equipment of a radio access network, the method comprising: The system receives trigger configuration information from the information processing server. This trigger configuration information is used to trigger updates to the artificial intelligence (AI) model used by the infrastructure device to perform tasks. The trigger configuration information includes one or more trigger conditions for triggering the infrastructure device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics indicating the suitability of the AI ​​model for performing tasks. A request to update the AI ​​model is sent to the information processing server, wherein, in response to calculating one or more performance metrics for one or more trigger conditions, and determining that one or more trigger conditions have been met based on the one or more performance metrics calculated for one or more trigger conditions, the request to update the AI ​​model is sent.

[0189] Paragraph 19. A method for operating an information processing server, the method comprising: The system sends configuration information to infrastructure equipment in a radio access network or via infrastructure equipment to a communication device to trigger an update of an artificial intelligence (AI) model used by the infrastructure equipment or communication device to perform a task. The trigger configuration information includes one or more trigger conditions for triggering the infrastructure equipment or communication device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics indicating the suitability of the AI ​​model for performing the task. Receive requests to update AI models from infrastructure equipment or from communication devices via infrastructure equipment.

[0190] Paragraph 20. According to the method described in paragraph 19, the information processing server includes one or more of the following: Location Management Server (LMF) AI server The server is configured to communicate with LMF.

[0191] Paragraph 21. According to the method described in paragraph 20, the information processing server includes: The model management unit is configured to receive requests to update the AI ​​model from infrastructure devices. The model training unit is configured to receive requests to update the AI ​​model from the model management unit and update the AI ​​model based on the requests. The model storage unit is configured to receive and store updated AI models from the model training unit.

[0192] Paragraph 22. A method of operating a communication device, the method comprising: The calculation indicates one or more performance metrics that assess the suitability of the artificial intelligence (AI) model used by the communication device to perform the task, and The system periodically sends one or more calculated performance metrics to the information processing server via a radio access network.

[0193] Paragraph 23. A method for operating infrastructure equipment of a radio access network, the method comprising: The information processing server periodically sends one or more performance metrics calculated by the infrastructure device, which indicate the suitability of the artificial intelligence (AI) model used by the infrastructure device to perform the task.

[0194] Paragraph 24. A method for operating an information processing server, the method comprising: The system periodically receives indications from infrastructure equipment of a radio access network or from a communication device via infrastructure equipment, which are calculated by the infrastructure equipment or communication device. These indications indicate the suitability of the artificial intelligence (AI) model used by the infrastructure equipment or communication device to perform the task. Based on the received performance metrics, it is determined that one or more triggering conditions for updating the AI ​​model have been met; Update AI models, and Instructions to send updated AI models to infrastructure equipment or via infrastructure equipment to communication devices.

[0195] Paragraph 25. A communication device comprising: The transceiver circuit is configured to transmit and receive signals. The controller circuit is configured to cooperate with the transceiver circuit to: Trigger configuration information is received from an information processing server via a radio access network. This trigger configuration information is used to trigger an update of the artificial intelligence (AI) model used by the communication device to perform the task. The trigger configuration information includes one or more trigger conditions for triggering the communication device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics that indicate the suitability of the AI ​​model for performing the task. Calculate one or more performance metrics for each of the one or more trigger conditions. Based on one or more performance metrics calculated for each of the one or more trigger conditions, it is determined that one or more trigger conditions have been met, and this is taken as a response. A request to update the AI ​​model is sent to the information processing server via a radio access network.

[0196] Paragraph 26. An infrastructure device for a radio access network, the infrastructure device comprising: The transceiver circuit is configured to transmit and receive signals. The controller circuit is configured to cooperate with the transceiver circuit to: The system receives trigger configuration information from the information processing server. This trigger configuration information is used to trigger an update of the artificial intelligence (AI) model used by the infrastructure device to perform the task. The trigger configuration information includes one or more trigger conditions for triggering the infrastructure device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics that indicate the suitability of the AI ​​model for performing the task. A request to update the AI ​​model is sent to the information processing server, wherein, in response to calculating one or more performance metrics for one or more trigger conditions, and determining that one or more trigger conditions have been met based on the one or more performance metrics calculated for one or more trigger conditions, the request to update the AI ​​model is sent.

[0197] Paragraph 27. An information processing server, the information processing server comprising: The transceiver circuit is configured to transmit and receive signals. The controller circuit is configured to cooperate with the transceiver circuit to: The system sends configuration information to infrastructure equipment in a radio access network or via infrastructure equipment to a communication device to trigger an update of an artificial intelligence (AI) model used by the infrastructure equipment or communication device to perform a task. The trigger configuration information includes one or more trigger conditions for triggering the infrastructure equipment or communication device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics indicating the suitability of the AI ​​model for performing the task. Receive requests to update AI models from infrastructure equipment or from communication devices via infrastructure equipment.

[0198] It should be understood that, for clarity, the above description has referenced different functional units, circuit systems, and / or processors in describing the implementation. However, it will be apparent that any suitable allocation of functions among the different functional units, circuit systems, and / or processors can be used without departing from the implementation.

[0199] The described embodiments can be implemented in any suitable form, including hardware, software, firmware, or any combination thereof. The described embodiments can optionally be implemented, at least in part, as computer software operating on one or more data processors and / or digital signal processors. Elements and components of any embodiment can be implemented physically, functionally, and logically in any suitable manner. In practice, functionality can be implemented in a single unit, in multiple units, or as part of other functional units. Therefore, the disclosed embodiments can be implemented in a single unit or can be physically and functionally distributed among different units, circuit systems, and / or processors.

[0200] Although this disclosure has been described in conjunction with some embodiments, it is not intended to limit it to the specific forms set forth herein. Furthermore, while features may appear to be described in conjunction with particular embodiments, those skilled in the art will recognize that the different features of the described embodiments can be combined in any manner suitable for implementing the technology.

[0201] References

[0202] [1] Holma H. ​​and Toskala A, “LTE for UMTS OFDMA and SC-FDMA basedradio access”, John Wiley and Sons, 2009.

[0203] [2] Technical Report (TR) 38.843, v18.0.0, December 2023, 3rdGeneration Partnership Project (3GPP).

Claims

1. A method of operating a communication device, the method comprising: The communication device receives trigger configuration information from an information processing server via a radio access network. This trigger configuration information is used to trigger an update of the artificial intelligence (AI) model used by the communication device to perform the task. The trigger configuration information includes one or more trigger conditions for triggering the communication device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics indicating the suitability of the AI ​​model for performing the task. Calculate the one or more performance metrics for each of the one or more triggering conditions. Based on one or more performance metrics calculated for each of the one or more trigger conditions, it is determined that one or more of the trigger conditions have been met, and in response, The request to update the AI ​​model is sent to the information processing server via the radio access network.

2. The method according to claim 1, wherein, The information processing server includes one or more of the following: Location Management Function (LMF) server, AI servers, and A server configured to communicate with the LMF server.

3. The method according to claim 1, wherein, The performance metrics for one or more of the triggering conditions include performance metrics that indicate data drift of the input data of the AI ​​model.

4. The method according to claim 1, wherein, The one or more performance metrics for one or more of the triggering conditions include performance metrics that indicate model drift of the AI ​​model.

5. The method according to claim 1, wherein, The trigger configuration information includes an indication of a format for calculating the one or more performance metrics for each of the one or more trigger conditions.

6. The method according to claim 5, wherein, One or more of the triggering conditions are based on comparing the performance metrics of the triggering condition with a threshold.

7. The method according to claim 6, wherein, The indication of the threshold of one or more of the trigger conditions is included in the trigger configuration information.

8. The method according to claim 6, wherein, The threshold is in the same format as that used to calculate the performance metric.

9. The method according to claim 1, wherein, The request to update the AI ​​model includes an identifier of the AI ​​model currently being used by the communication device.

10. The method according to claim 1, wherein, The request to update the AI ​​model includes an indication of one or more performance metrics calculated for the already met triggering conditions.

11. The method according to claim 10, wherein, The already satisfied triggering conditions are based on multiple performance metrics calculated by the communication device, and the method includes: Indicators of the plurality of performance metrics calculated by the communication device are sent to the information processing server via the radio access network, wherein each of the plurality of performance metrics calculated by the communication device is sent at a different time.

12. The method according to claim 1, comprising: The updated AI model is received from the information processing server via the radio access network.

13. The method according to claim 1, wherein, The task performed by the AI ​​model is a positioning task to determine the location of the communication device, or a beam management task, or a channel state information (CSI) measurement and reporting task, a CSI prediction task, or a CSI compression task.

14. The method according to claim 13, wherein, The task performed by the AI ​​model is a localization task, and the input data of the AI ​​model includes localization signal measurements. The AI ​​model is configured to generate a more accurate estimate of the localization signal measurements based on the input data, or to generate an estimate of the location of the communication device based on the input data.

15. The method according to claim 1, wherein, Calculating the one or more performance metrics for each of the one or more triggering conditions includes: The AI ​​model is used to calculate one or more performance metrics.

16. A method of operating infrastructure equipment for a radio access network, the method comprising: The system receives trigger configuration information from an information processing server. This trigger configuration information is used to trigger an update of the artificial intelligence (AI) model used by the infrastructure device to perform the task. The trigger configuration information includes one or more trigger conditions for triggering the infrastructure device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics indicative of the suitability of the AI ​​model for performing the task. Sending the request to update the AI ​​model to the information processing server, wherein, in response to calculating the one or more performance metrics for one or more of the triggering conditions, and determining that one or more of the triggering conditions have been met based on the one or more performance metrics calculated for the one or more triggering conditions, the request to update the AI ​​model is sent.

17. A method for operating an information processing server, the method comprising: The system sends configuration information to an infrastructure device of a radio access network, or via the infrastructure device to a communication device, for triggering an update of an artificial intelligence (AI) model used by the infrastructure device or the communication device to perform a task. The configuration information includes one or more triggering conditions for triggering the infrastructure device or the communication device to send a request to update the AI ​​model. Each of the one or more triggering conditions is based on one or more performance metrics indicating the suitability of the AI ​​model to perform the task. Receive a request to update the AI ​​model from the infrastructure device or via the infrastructure device from the communication device.

18. The method according to claim 17, wherein, The information processing server includes one or more of the following: Location Management Server (LMF) AI server A server configured to communicate with the LMF.

19. The method according to claim 18, wherein, The information processing server includes: The model management unit is configured to receive the request to update the AI ​​model from the infrastructure device. The model training unit is configured to receive the request to update the AI ​​model from the model management unit and update the AI ​​model based on the request. The model storage unit is configured to receive updated AI models from the model training unit and store the updated AI models.

20. A method of operating a communication device, the method comprising: The calculation includes one or more performance metrics indicating the suitability of the artificial intelligence (AI) model used by the communication device to perform the task, and The system periodically sends one or more calculated performance metrics to the information processing server via a radio access network.

21. A method for operating infrastructure equipment of a radio access network, the method comprising: The information processing server periodically sends indications of one or more performance metrics calculated by the infrastructure device, the one or more performance metrics indicating the suitability of the artificial intelligence (AI) model used by the infrastructure device to perform the task.

22. A method for operating an information processing server, the method comprising: The system periodically receives indications from infrastructure equipment of a radio access network, or from a communication device via said infrastructure equipment, of one or more performance metrics calculated by said infrastructure equipment or said communication device, said one or more performance metrics indicating the suitability of the artificial intelligence (AI) model used by said infrastructure equipment or said communication device to perform the task. Based on the received performance metrics, it is determined that one or more triggering conditions for updating the AI ​​model have been met. Update the AI ​​model, and Instructions to send updated AI models to the infrastructure equipment or via the infrastructure equipment to the communication device.

23. A communication device, the communication device comprising: The transceiver circuit is configured to transmit and receive signals. The controller circuit is configured to cooperate with the transceiver circuit to: The communication device receives trigger configuration information from an information processing server via a radio access network. This trigger configuration information is used to trigger an update of the artificial intelligence (AI) model used by the communication device to perform the task. The trigger configuration information includes one or more trigger conditions for triggering the communication device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics indicating the suitability of the AI ​​model for performing the task. Calculate the one or more performance metrics for each of the one or more triggering conditions. Based on one or more performance metrics calculated for each of the one or more trigger conditions, it is determined that one or more of the trigger conditions have been met, and in response, The request to update the AI ​​model is sent to the information processing server via the radio access network.

24. An infrastructure device for a radio access network, the infrastructure device comprising: The transceiver circuit is configured to transmit and receive signals. The controller circuit is configured to cooperate with the transceiver circuit to: The system receives trigger configuration information from an information processing server. This trigger configuration information is used to trigger an update of the artificial intelligence (AI) model used by the infrastructure device to perform the task. The trigger configuration information includes one or more trigger conditions for triggering the infrastructure device to send a request to update the AI ​​model. Each of the one or more trigger conditions is based on one or more performance metrics indicative of the suitability of the AI ​​model for performing the task. Sending a request to update the AI ​​model to the information processing server, wherein, in response to calculating the one or more performance metrics for one or more of the triggering conditions, and determining that one or more of the triggering conditions have been met based on the one or more performance metrics calculated for the one or more triggering conditions, the request to update the AI ​​model is sent.

25. An information processing server, the information processing server comprising: The transceiver circuit is configured to transmit and receive signals. The controller circuit is configured to cooperate with the transceiver circuit to: The system sends configuration information to an infrastructure device of a radio access network, or via the infrastructure device to a communication device, for triggering an update of an artificial intelligence (AI) model used by the infrastructure device or the communication device to perform a task. The configuration information includes one or more triggering conditions for triggering the infrastructure device or the communication device to send a request to update the AI ​​model. Each of the one or more triggering conditions is based on one or more performance metrics indicating the suitability of the AI ​​model to perform the task. Receive a request to update the AI ​​model from the infrastructure equipment or from the communication device via the infrastructure equipment.

Images