Method and system for machine learning based csi compression

Abstract

Methods and systems for compressing channel state information (CSI), particularly precoding vectors, using machine learning are described. At a central data center, precoding vectors acquired by user equipment (UE) are converted to coefficient vectors that represent a linear combination of one or more selected representative vectors from a set of representations, the set of representations defined by one or more representative vectors, the coefficient vectors usable to train a neural network to perform a precoding vector encoding task or a precoding vector decoding task.

Description

[0001] Cross-references to related applications

[0002] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 638,321, filed April 24, 2024, entitled “A method to support two-sided AI / ML model for CSI compression,” the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates to wireless communications, and more particularly to methods and systems for compressing channel state information (CSI) using machine learning. Background Technology

[0004] In 5G New Radio (NR), accurate Channel State Information (CSI) is crucial for enabling advanced technologies such as beamforming and spatial multiplexing. However, the massive MIMO and wideband channels in 5G NR systems generate high-dimensional CSI, which can lead to excessive overhead if reported directly without compression. To address this issue, 3GPP standardized CSI compression techniques for 5G NR in Release 15. The compression technique specified in 5G NR Release 15 is not data-driven; it relies on a predefined codebook and compression matrix derived from mathematical models and assumptions about the channel. However, with the emergence of artificial intelligence (AI) and machine learning (ML) technologies, exploring data-driven methods for CSI compression has attracted significant attention. These methods can potentially improve performance by adapting to actual channel statistics and dynamic changes.

[0005] One researched approach is to use a two-sided AI / ML model for CSI compression. In this framework, a pair of AI / ML models (typically deep neural networks) are jointly trained: an encoder model on the user equipment (UE) side and a decoder model on the gNodeB (gNB) side. The UE-side encoder model compresses the high-dimensional CSI (typically a pre-encoded matrix) into a low-dimensional representation and then feeds it back to the gNB. The gNB-side decoder model then reconstructs the original CSI based on this compressed representation. However, many challenges remain regarding inter-vendor collaboration, model complexity management, and deployment considerations.

[0006] Therefore, there is a need to provide a solution that can improve the interoperability of two-sided AI / ML models used for CSI compression between different vendors, but this is quite challenging. Summary of the Invention

[0007] In various examples, this disclosure describes methods, systems, and computer-readable media for compressing CSI using machine learning. In particular, this disclosure describes methods, systems, and computer-readable media for enabling the transmission of CSI (specifically, vectors of precoded matrices) in a compressed representation using machine learning-based encoder and decoder models. An example of this disclosure provides a centralized data center where data used to train neural networks to implement encoder and decoder models can be stored. The precoded vectors acquired by the UE can be stored at the data center using a sparse representation, which enables more efficient training of the encoder and decoder models.

[0008] The examples disclosed herein offer the following technical advantages: enabling a central data center to maintain data used for training encoder or decoder models without requiring the encoder model to depend on the architecture or complexity of the decoder model (or vice versa). This facilitates improved efficiency of the overall communication system by enabling greater flexibility and optimization of encoder and decoder models that can be developed independently. Another technical advantage is the significant savings in memory resources achieved by storing pre-encoded vector data in the data center using sparse representations.

[0009] The examples disclosed herein also offer the following technical advantages: by making the encoder model deployed at the UE better adaptable to the limited resources at the UE and the specific spatial characteristics of interest, computational resources can be used more efficiently. Simultaneously, decoder models deployed at network entities with more resources and serving a larger spatial area can be developed to be more complex. In this way, the examples disclosed herein can achieve better efficiency and / or performance on both the UE and network sides.

[0010] The examples disclosed herein can provide the following technical advantages: CSI can be transmitted for more complex scenarios than currently permitted using a fixed precoder codebook. This can help improve resource allocation and other network operations.

[0011] Some embodiments of this disclosure may include one or more of the following features, which may be used individually or together as a complete solution.

[0012] In an example of the first aspect, this disclosure describes a method comprising: receiving data including a precoded vector associated with a user equipment (UE); converting the precoded vector into a coefficient vector, the coefficient vector representing a linear combination of one or more selected representative vectors from a representation set defined by the one or more representative vectors, the coefficient vector being usable for training a neural network to perform a precoded vector encoding task or a precoded vector decoding task.

[0013] In one example of the above example of the first aspect, converting the precoded vector into the coefficient vector may include: computing candidate vectors to represent the precoded vector as a linear combination of one or more candidate representative vectors from the representation set; and determining that the candidate vectors satisfy a sparsity threshold; wherein, after determining that the candidate vectors satisfy the sparsity threshold, the candidate vectors may be the coefficient vectors, and the one or more candidate representative vectors may be the one or more selected representative vectors.

[0014] In one example of the above-described example of the first aspect, converting the precoded vector into the coefficient vector may include: determining that the representation set cannot represent the precoded vector; updating the representation set by including the precoded vector as a new representative vector in the representation set; and converting the precoded vector into the coefficient vector after the update.

[0015] In one example of the above example of the first aspect, determining that the representation set cannot represent the precoded vector may include: prior to the update, computing candidate vectors to represent the precoded vector as a linear combination of one or more candidate representative vectors from the representation set; and determining that the candidate vectors do not satisfy a sparsity threshold.

[0016] In one example of any of the above examples of the first aspect, the method may further include: tracking the usage and / or duration of existence of each representative vector in the representation set; and updating the representation set by deleting outdated representative vectors based on the tracked usage and / or duration of existence.

[0017] In one example of some of the above examples of the first aspect, the method may further include: sending a control message indicating an update to the representation set.

[0018] In one example of any of the above examples of the first aspect, the method may further include: sending a control message indicating one or more parameters to be used by the UE to generate the precoding vector; wherein the precoding vector may respond to the reception of the control message.

[0019] In one example of any of the above examples of the first aspect, the received data may further include auxiliary information related to the data acquisition performed by the UE to generate the precoding vector, and the coefficient vector corresponding to the precoding vector may be associated with the auxiliary information.

[0020] In one example of the above-described example of the first aspect, the auxiliary information may include one or more of the following: information about the frequency subband used in the data acquisition; information about the resource blocks used in the data acquisition; information about the antenna configuration used in the data acquisition; and information about the geographical location during the data acquisition process.

[0021] In one example of any of the above examples of the first aspect, the method may further include: providing the coefficient vector to the neural network for training to perform the precoded vector encoding task or the precoded vector decoding task.

[0022] In one example of any of the above examples of the first aspect, the method may further include: generating a synthetic coefficient vector based on one or more coefficient vectors obtained from the precoded vector; wherein the synthetic coefficient vector may also be used to train the neural network to perform the precoded vector encoding task or the precoded vector decoding task.

[0023] In one example of the above-described example of the first aspect, the method may further include: providing the synthesized coefficient vector to the neural network for training to perform the precoded vector encoding task or the precoded vector decoding task.

[0024] In one example of any of the above examples of the first aspect: the method can be performed at a data center of a wireless communication system including the UE; the set of one or more representative vectors can be maintained in the memory of the data center; the coefficient vectors can be stored in a dataset maintained in the memory of the data center and can be used to train the neural network to perform the precoded vector encoding task or the precoded vector decoding task.

[0025] In one example of the above-described example of the first aspect, the method may include: sending a coefficient vector from the dataset for training the neural network to perform the precoded vector decoding task, wherein the sent coefficient vector is sampled from the entire dataset.

[0026] In one example of the above-described example of the first aspect, the method may include: sending coefficient vectors from the dataset for training the neural network to perform the pre-encoding vector encoding task, wherein the sent coefficient vectors are sampled from a selected subset of the dataset, the selected subset containing coefficient vectors associated with spatial features of interest, and the sent coefficient vectors correspond to a selected subset of representations, the selected subset of representations omitting at least one representative vector from the set of representations and associated with the spatial features of interest.

[0027] In one example of the above examples of the first aspect, the method may include: sending mapping information for mapping between the subset of representations and the set of representations.

[0028] In one example of some of the above examples in the first aspect, the transmitted coefficient vector may be transmitted in a compact vector form, which is smaller than the extended vector form used to store the coefficient vector in the dataset, the compact vector form omitting elements corresponding to at least one representative vector omitted from the representation subset.

[0029] In one example of the above-described example of the first aspect, the mapping information may also be mapped between the compact vector form and the extended vector form.

[0030] In one example of the second aspect, this disclosure describes a method comprising: obtaining a selected subset of one or more representative vectors from a set of one or more representative vectors defining a representation set, the selected subset of one or more representative vectors being defined as a subset of representations of the representation set; obtaining coefficient vectors, wherein each coefficient vector represents a linear combination of one or more selected representative vectors from the representation subset, the linear combination corresponding to a precoded vector; and training a neural network to perform a precoded vector encoding task, wherein during training, each coefficient vector is converted into a corresponding precoded vector, and the corresponding precoded vector is used to train the neural network.

[0031] In one example of the above example of the second aspect, training the neural network may include: providing the corresponding precoded vector as input to the neural network and obtaining an output; and training the neural network to minimize the error between the output generated by the neural network and the coefficient vector corresponding to the precoded vector.

[0032] In one example of any of the above examples in the second aspect, the coefficient vectors may be sampled from a selected subset of data containing coefficient vectors associated with spatial features of interest, each coefficient vector in the subset of data being converted into a corresponding precoded vector using the representation subset.

[0033] In one example of the above example in the second aspect, each coefficient vector can be obtained in a compact vector form corresponding to the subset of representations, which is smaller than the extended vector form corresponding to the set of representations.

[0034] In one example of any of the above examples in the second aspect, the method may further include: deploying the trained neural network to perform the pre-encoded vector encoding task.

[0035] In one example of the above example in the second aspect, the trained neural network can be deployed at a user equipment (UE).

[0036] In one example of any of the above examples in the second aspect, the method may further include: updating the subset of representations to include new representative vectors from the set of representations or excluding outdated representative vectors from the subset of representations; obtaining another set of coefficient vectors corresponding to the updated subset of representations; and retraining the neural network using the other set of coefficient vectors to perform the pre-encoding vector encoding task.

[0037] In one example of the third aspect, this disclosure describes a method comprising: obtaining a set of one or more representative vectors defining a representation set; obtaining coefficient vectors, wherein each coefficient vector represents a linear combination of one or more selected representative vectors from the representation set, the linear combination corresponding to a precoded vector; and training a neural network to perform a precoded vector decoding task, wherein, during training, each coefficient vector is used to train the neural network.

[0038] In one example of the above example of the third aspect, training the neural network may include: providing the coefficient vector as input to the neural network and obtaining an output; training the neural network to minimize the error between the output generated by the neural network and the precoded vector corresponding to the coefficient vector.

[0039] In one example of any of the above examples in the third aspect, the method may include: receiving an update to the set of representations, wherein the neural network is retrained in response to the update.

[0040] In one example of any of the above examples in the third aspect, the method may include: receiving feedback relating to the performance of the trained neural network; and using the feedback to further refine the training of the neural network.

[0041] In one example of any of the above examples in the third aspect, the neural network can be trained using transfer learning, wherein the pre-trained neural network can be further trained using the coefficient vector.

[0042] In one example of any of the above examples in the third aspect, the method may include: deploying the trained neural network to perform the precoded vector decoding task.

[0043] In one example of the above examples in the third aspect, the method may include: receiving a coefficient vector; and decoding a precoded vector from the coefficient vector using the trained neural network.

[0044] In one example of the above example of the third aspect, the method may include: obtaining a mapping between a subset of representations and the set of representations, the subset of representations being defined by a selected subset of one or more representative vectors from the set of representations; wherein the coefficient vectors may be received in a compact vector form corresponding to the subset of representations; and the coefficient vectors may be converted from the compact vector form to an extended vector form corresponding to the set of representations using the mapping, the decoding being performed on the coefficient vectors in the extended vector form.

[0045] In one example of the fourth aspect, this disclosure describes a method comprising: receiving a coefficient vector representing a linear combination of one or more selected representative vectors from a representation set defined by the one or more representative vectors, and the linear combination corresponding to a precoded vector; and decoding the precoded vector from the coefficient vector using a trained neural network, wherein the trained neural network has been trained to perform a precoded vector decoding task.

[0046] In one example of the above examples of the fourth aspect, the method may include: obtaining a mapping between a subset of representations and the set of representations, the subset of representations being defined by a selected subset of one or more representative vectors from the set of representations; wherein the coefficient vectors may be received in a compact vector form corresponding to the subset of representations; and the coefficient vectors may be converted from the compact vector form to an extended vector form corresponding to the set of representations using the mapping, the decoding being performed on the coefficient vectors in the extended vector form.

[0047] In one example of any of the above examples in the fourth aspect, the method may include: storing the decoded precoded vector and the corresponding original coefficient vector as a data pair for further analysis / refinement of the trained neural network.

[0048] In one example of any of the above examples in the fourth aspect, the method may include: sending the decoded precoded vector for optimizing network operation and / or resource allocation.

[0049] In one example of the fifth aspect, this disclosure describes a method comprising: acquiring a precoded vector; encoding the precoded vector into a coefficient vector using a trained neural network, the coefficient vector representing a linear combination of one or more selected representative vectors from a subset of representations, the subset of representations being defined by one or more representative vectors, wherein the linear combination corresponds to the precoded vector; and transmitting the coefficient vector.

[0050] In one example of the above example in the fifth aspect, the coefficient vector can be sent as a precoding matrix indicator (PMI).

[0051] In one example of any of the above examples in the fifth aspect, the method may include: obtaining parameters of the trained neural network from a control signal.

[0052] In one example of the above examples of the fifth aspect, the method may include: updating the trained neural network based on feedback received from network entities.

[0053] In one example of any of the above examples in the fifth aspect, the precoded vector may be obtained based on one or more parameters indicated by a control message from a network entity.

[0054] In any of the above examples of the fifth aspect, the method may include sending auxiliary information associated with the precoded vector along with the coefficient vector.

[0055] In another exemplary aspect, this disclosure describes a network entity or apparatus comprising: a memory; and a processor for executing instructions stored in the memory to cause the network entity or apparatus to perform any of the above exemplary aspects of the method.

[0056] In another exemplary aspect, this disclosure describes a non-transitory computer-readable medium storing machine-executable instructions, wherein, when executed by a network entity or device, the instructions cause the network entity or device to perform any of the aforementioned exemplary aspects of the method.

[0057] In another exemplary aspect, this disclosure describes a processing module for controlling a network entity or device to cause the network entity or device to perform any of the above exemplary aspects of the method.

[0058] In any of the above examples, the network entity may be a data center. In any of the above examples, the network entity may provide services to one or more user equipment (UEs). In any of the above examples, the network entity may provide services to one or more base stations (BSs).

[0059] In another exemplary aspect, this disclosure describes a computer program that, when run on a computer, causes the computer to perform any of the above exemplary aspects of the method. Attached Figure Description

[0060] Reference will now be made to the accompanying drawings, which illustrate exemplary embodiments of the present disclosure, in which:

[0061] Figure 1 This is a simplified schematic diagram of a communication system that can be used to implement the examples of this disclosure;

[0062] Figure 2 A block diagram of an example communication system that can be used to implement the examples of this disclosure;

[0063] Figure 3 A block diagram of a device example that can be used to implement the examples of this disclosure;

[0064] Figure 4 A block diagram of an exemplary unit or module that can be used to implement the examples of this disclosure;

[0065] Figure 5 This is a schematic diagram of an exemplary method for training encoder and decoder models using a centralized dataset in some existing technical approaches;

[0066] Figure 6 is a schematic diagram illustrating the complexity difference between the ideal encoder model deployed at the UE and the ideal decoder model deployed at the network entity;

[0067] Figure 7 An example of how CSI complexity may increase due to technological advancements is shown;

[0068] Figure 8 It shows management Figure 7 The example shown illustrates the increase in the size of the central database required to increase complexity;

[0069] Figure 9An example of increased complexity in encoder and decoder models is shown, which is caused by an exemplary increase in communication that may result from technological advancements;

[0070] Figure 10 A schematic diagram illustrating how the central database provided for this disclosure example can be used to achieve interoperability;

[0071] Figure 11 A schematic diagram illustrating the data collection and registration examples provided in this disclosure;

[0072] Figure 12 This disclosure provides an example of how sparse representations can be used to store precoded vectors at the data center;

[0073] Figure 13 This disclosure provides an example of how a precoded vector can be added to a representation set at the data center.

[0074] Figure 14 This disclosure provides an example of how coefficient vectors from UEs with similar spatial characteristics can be clustered.

[0075] Figure 15 This disclosure illustrates examples of how different NW vendors or UE vendors can choose a representation set or a representation subset.

[0076] Figure 16 This illustrates an example of how the coefficient vector provided in this disclosure can be mapped from an extended vector form to a compact vector form;

[0077] Figure 17 An example of training a neural network to implement an encoder model, provided in this disclosure, is shown;

[0078] Figure 18 An example of training a neural network to implement a decoder model, provided in this disclosure, is shown;

[0079] Figure 19 This disclosure illustrates an example of how a trained neural network, as provided in this disclosure, can be used for CSI communications;

[0080] Figure 20 A block diagram of a data center example provided for this disclosure;

[0081] Figure 21 A flowchart illustrating an exemplary method for maintaining a dataset and a collection of representations, as provided in this disclosure.

[0082] Figure 22 A flowchart illustrating an exemplary method for training a neural network to perform a pre-coding vector encoding task, as provided in this disclosure.

[0083] Figure 23 A flowchart illustrating an exemplary method for training a neural network to perform a precoded vector decoding task, as provided in this disclosure.

[0084] Figure 24 A flowchart illustrating an exemplary method for decoding precoded vectors using a trained neural network, as provided in this disclosure.

[0085] Figure 25 A flowchart illustrating an exemplary method for encoding precoded vectors using a trained neural network, as provided in this disclosure.

[0086] Similar reference numerals may be used in different accompanying drawings to denote similar components. Detailed Implementation

[0087] In various examples, this disclosure describes methods, systems, and computer-readable media capable of performing machine learning-based two-sided compression of Channel State Information (CSI). "Two-sided" refers to compressing CSI using a trained encoder model on the User Equipment (UE) side and decompressing CSI using a trained decoder model on the network side. The examples disclosed herein can help improve the interoperability of encoder / decoder models developed by different vendors, specifically by providing a central data center where data can be registered and retrieved for training neural networks to implement encoder or decoder models. As disclosed herein, the data center can store data in a sparse format, which helps improve the fairness and standardization of encoder / decoder models between UE-side vendors and network-side vendors.

[0088] To aid in understanding this disclosure, first refer to Figure 1 .

[0089] refer to Figure 1 As a non-limiting illustrative example, a simplified schematic diagram of a communication system is provided. Communication system 100 includes a radio access network 120. Radio access network 120 may be a next-generation radio access network or a traditional (e.g., 5G, 4G, 3G, or 2G) radio access network. One or more communication electronic devices (EDs) 110a, 110b, 110c, 110d, 110e, 110f, 110g, 110h, 110i, 110j (generally referred to as 110) may be interconnected or connected to one or more network nodes (170a, 170b, generally referred to as 170) in radio access network 120. Core network 130 may be part of the communication system and may depend on or be independent of the radio access technology used in communication system 100. Furthermore, communication system 100 includes a public switched telephone network (PSTN) 140, the Internet 150, and other networks 160.

[0090] Figure 2 An exemplary communication system 100 is illustrated. Typically, communication system 100 enables multiple wireless or wired components to transmit data and other content. The purpose of communication system 100 may be to provide content such as voice, data, video, and / or text via broadcast, multicast, unicast, etc. Communication system 100 can operate by sharing resources (such as carrier spectrum bandwidth) among its constituent units. Communication system 100 may include terrestrial communication systems and / or non-terrestrial communication systems. Communication system 100 can provide a wide range of communication services and applications (e.g., earth monitoring, remote sensing, passive sensing and positioning, navigation and tracking, autonomous delivery and mobility, etc.). Communication system 100 can provide high availability and robustness through the joint operation of terrestrial and non-terrestrial communication systems. For example, integrating a non-terrestrial communication system (or components thereof) into a terrestrial communication system can create a heterogeneous network comprising multiple layers. Compared to traditional communication networks, heterogeneous networks can achieve better overall performance through efficient multi-link joint operation, more flexible function sharing, and faster physical layer link switching between terrestrial and non-terrestrial networks.

[0091] Terrestrial communication systems and non-terrestrial communication systems can be considered subsystems of a communication system. Figure 2 In the example shown, communication system 100 includes electronic devices (EDs) 110a, 110b, 110c, and 110d (generally referred to as ED 110), radio access networks (RANs) 120a and 120b, a non-terrestrial communication network 120c, a core network 130, a public switched telephone network (PSTN) 140, the Internet 150, and other networks 160. RANs 120a and 120b include corresponding base stations (BSs) 170a and 170b, which are generally referred to as terrestrial transmit and receive points (T-TRPs) 170a and 170b. The non-terrestrial communication network 120c includes an access node 172, which is generally referred to as a non-terrestrial transmit and receive point (NT-TRP) 172.

[0092] Alternatively or additionally, any ED 110 can be used to connect, access, or communicate with any T-TRP 170a, 170b, and NT-TRP 172, the Internet 150, the core network 130, the PSTN 140, other networks 160, or any combination thereof. In some examples, ED 110a can transmit with T-TRP 170a via terrestrial air interface 190a, via uplink and / or downlink transmissions. In some examples, ED 110a, 110b, 110c, and 110d can also communicate directly with each other via one or more sidelink air interfaces 190b. In some examples, ED 110d can transmit with NT-TRP 172 via non-terrestrial air interface 190c, via uplink and / or downlink transmissions.

[0093] Air interfaces 190a and 190b can employ similar communication technologies, such as any suitable wireless access technology. For example, communication system 100 can implement one or more channel access methods in air interfaces 190a and 190b, such as code division multiple access (CDMA), space division multiple access (SDMA), time division multiple access (TDMA), frequency division multiple access (FDMA), orthogonal FDMA (OFDMA), or single-carrier FDMA (SC-FDMA) (also known as discrete Fourier transform spread OFDMA (DFT-s-OFDMA)). Air interfaces 190a and 190b can employ other higher-dimensional signal spaces, which may involve combinations of orthogonal and / or non-orthogonal dimensions.

[0094] The non-terrestrial air interface 190c enables communication between the ED 110d and one or more NT-TRP 172s via a wireless link, or simply a link. For some examples, a link is a dedicated connection for unicast transmission, a connection for broadcast transmission, or a connection between a group of ED 110s and one or more NT-TRP 172s for multicast transmission.

[0095] RANs 120a and 120b communicate with core network 130 to provide various services, such as voice, data, and other services, to EDs 110a, 110b, and 110c. RANs 120a and 120b and / or core network 130 may communicate directly or indirectly with one or more other RANs (not shown), which may or may not be directly served by core network 130 and may or may not use the same radio access technology as RANs 120a and / or RAN 120b. Core network 130 may also serve as a gateway access between (i) RANs 120a and 120b, or EDs 110a, 110b, and 110c, or both, and (ii) other networks (such as PSTN 140, Internet 150, and other networks 160). Additionally, some or all of EDs in EDs 110a, 110b, and 110c may include functionality for communicating with different wireless networks via different radio links using different radio technologies and / or protocols. ED 110a, 110b, and 110c can communicate with service providers or exchanges (not shown) via wired communication channels and with the Internet 150, rather than via wireless communication (or via wired communication in addition to wireless communication). PSTN 140 may include a circuit-switched telephone network for providing plain old telephone service (POTS). The Internet 150 may include a computer network and / or subnet (internal network) and employ protocols such as Internet Protocol (IP), Transmission Control Protocol (TCP), and User Datagram Protocol (UDP). ED 110a, 110b, and 110c may be multimode devices capable of operating under various wireless access technologies and integrate multiple transceivers required to support such technologies.

[0096] Figure 3Another example of an ED 110 and base stations 170a, 170b, and / or 170c is shown. The ED 110 is used to connect people, things, machines, etc. The ED 110 can be widely used in various scenarios, including cellular communication, device-to-device (D2D), vehicle-to-everything (V2X), peer-to-peer (P2P), machine-to-machine (M2M), machine-type communications (MTC), Internet of Things (IoT), virtual reality (VR), augmented reality (AR), mixed reality (MR), metaverse, digital twins, industrial control, autonomous driving, telemedicine, smart grids, smart furniture, smart offices, smart wearables, smart transportation, smart cities, drones, robots, remote sensing, passive sensing, positioning, navigation and tracking, autonomous delivery and mobility, and more.

[0097] Each ED 110 represents any suitable end-user equipment for wireless operation and may include (or be referred to as): user equipment / device (UE), wireless transmit / receive unit (WTRU), mobile station, fixed or mobile subscriber unit, cellular phone, station (STA), machine-type communication (MTC) device, personal digital assistant (PDA), smartphone, laptop, computer, tablet, wireless sensor, consumer electronics, smartbook, vehicle, automobile, truck, bus, train, or IoT device, wearable device (such as watch, glasses, head-mounted device, etc.), industrial equipment, or devices that include or incorporate the above-mentioned equipment (such as communication modules, modems, or chips), etc. Next-generation ED 110 may be referred to using other terms. Base stations 170a and 170b are T-TRPs and will be referred to as T-TRP 170 below. Similarly, Figure 3As shown, NT-TRP will be referred to as NT-TRP 172 below. Each ED 110 connected to T-TRP 170 and / or NT-TRP 172 can be dynamically or semi-statically started (i.e., established, activated, or enabled), shut down (i.e., released, deactivated, or disabled), and / or configured in response to one or more of connectivity availability and connectivity necessity.

[0098] ED 110 includes a transmitter 201 and a receiver 203 coupled to one or more antennas 204. Only one antenna 204 is shown in the figure to avoid congestion. Alternatively, one, some, or all of the antennas 204 may be panels. For example, the transmitter 201 and receiver 203 may be integrated as a transceiver. The transceiver is used to modulate data or other content for transmission by at least one antenna 204 or a network interface controller (NIC). The transceiver may also be used to demodulate data or other content received through at least one antenna 204. Each transceiver includes any suitable structure for generating signals for wireless or wired transmission and / or processing signals received wirelessly or wiredly. Each antenna 204 includes any suitable structure for transmitting and / or receiving wireless or wired signals.

[0099] ED 110 includes at least one memory 208. Memory 208 stores instructions and data used, generated, or acquired by ED 110. For example, memory 208 may store software instructions or modules executed by one or more processing units (e.g., processor 210) for implementing some or all of the functions and / or embodiments described herein. Each memory 208 includes any suitable one or more volatile and / or non-volatile storage and retrieval devices. Any suitable type of memory can be used, such as random access memory (RAM), read-only memory (ROM), hard disk, optical disk, subscriber identity module (SIM) card, memory stick, secure digital (SD) memory card, and on-processor cache, etc.

[0100] ED 110 may also include one or more input / output devices (not shown) or interfaces (such as those connected to...). Figure 1(Wired interface of Internet 150 in the network). Input / output devices or interfaces can interact with users or other devices in the network. Each input / output device or interface includes any suitable structure for providing or receiving information from the user and / or for network interface communication. For example, suitable structures include speakers, microphones, keypads, keyboards, displays, touchscreens, etc.

[0101] ED 110 includes a processor 210 for performing operations, including operations related to preparing for uplink transmissions to NT-TRP 172 and / or T-TRP 170, operations related to processing downlink transmissions received from NT-TRP 172 and / or T-TRP 170, and operations related to processing sidelink transmissions transmitted to and received from another ED 110. Processing operations related to preparing for uplink transmissions may include operations such as encoding, modulation, transmit beamforming, and generating symbols for transmission. Processing operations related to processing downlink transmissions may include operations such as receive beamforming, demodulation, and decoding of received symbols. According to an embodiment, the downlink transmission may be received by receiver 203 possibly using receive beamforming, and processor 210 may extract signaling from the downlink transmission (e.g., by detecting and / or decoding signaling). For example, an example of signaling may be a reference signal transmitted by NT-TRP 172 and / or T-TRP 170. In some embodiments, processor 210 performs transmit beamforming and / or receive beamforming based on beam direction indication (e.g., beam angle information (BAI)) received from T-TRP 170. In some embodiments, processor 210 may perform operations related to network access (such as initial access) and / or downlink synchronization, such as operations related to detecting synchronization sequences, decoding, and acquiring system information. In some embodiments, processor 210 may perform channel estimation, for example, using reference signals received from NT-TRP 172 and / or T-TRP 170.

[0102] Although not shown, processor 210 may be part of transmitter 201 and / or receiver 203. Although not shown, memory 208 may constitute part of processor 210.

[0103] The processing components of processor 210, transmitter 201, and receiver 203 can be implemented by the same or different one or more processors, which execute instructions stored in memory (e.g., memory 208). Alternatively, some or all of the processing components of processor 210, transmitter 201, and receiver 203 can be implemented using dedicated circuitry, such as a programmable field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), or hardware accelerators such as graphics processing units (GPUs) or artificial intelligence (AI) accelerators.

[0104] In some implementations, T-TRP 170 can be referred to by other names, such as base station, base transceiver station (BTS), wireless base station, network node, network device, network-side device, transmit / receive node, NodeB, evolved NodeB (eNodeB or eNB), home eNodeB, next-generation NodeB (gNB), transmission point (TP), site controller, access point (AP), wireless router, relay station, ground node, ground network device, ground base station, base band unit (BBU), remote radio unit (RRU), active antenna unit (AAU), remote radio head (RRH), central unit (CU), distributed unit (DU), location node, etc. T-TRP 170 can be a macro BS, pico BS, relay node, or host node, or a combination thereof. T-TRP 170 may refer to the aforementioned equipment or to a device within the aforementioned equipment (such as a communication module, modem, or chip).

[0105] In some embodiments, the various parts of T-TRP 170 may be distributed. For example, some modules of T-TRP 170 may be located remotely from the device housing the antenna 256 for T-TRP 170 and may be coupled to the device housing the antenna 256 via a communication link (not shown), sometimes referred to as the fronthaul, such as the Common Public Radio Interface (CPRI). Therefore, in some embodiments, the term T-TRP 170 may also refer to network-side modules that perform processing operations such as ED 110 location determination, resource allocation (scheduling), message generation, and encoding / decoding, which are not necessarily part of the device housing the antenna 256 of T-TRP 170. These modules may also be coupled to other T-TRPs. In some embodiments, T-TRP 170 may actually be multiple T-TRPs that operate together to serve ED 110, for example, through cooperative multicast.

[0106] T-TRP 170 includes at least one transmitter 252 and at least one receiver 254 coupled to one or more antennas 256. Only one antenna 256 is shown in the figure to avoid congestion. Alternatively, one, some, or all of the antennas 256 may be panels. The transmitter 252 and receiver 254 may be integrated as a transceiver. T-TRP 170 also includes a processor 260 for performing operations including operations related to: preparing transmissions for downlink transmission to ED 110, processing uplink transmissions received from ED 110, preparing transmissions for backhaul transmission to NT-TRP 172, and processing transmissions received from NT-TRP 172 via backhaul. Processing operations related to preparing transmissions for downlink or backhaul transmission may include operations such as encoding, modulation, precoding (e.g., multiple-input multiple-output (MIMO) precoding), transmit beamforming, and generating symbols for transmission. Processing operations related to transmissions received in the uplink or via backhaul may include receiving beamforming, demodulating received symbols, and decoding received symbols. Processor 260 may also perform operations related to network access (e.g., initial access) and / or downlink synchronization, such as generating the contents of a synchronization signal block (SSB), generating system information, etc. In some embodiments, processor 260 also generates a beam direction indication (e.g., BAI), which can be scheduled for transmission by scheduler 253. Processor 260 performs other network-side processing operations described herein, such as determining the location of ED 110, determining the deployment location of NT-TRP 172, etc. In some embodiments, processor 260 may generate signaling to configure one or more parameters of ED 110 and / or one or more parameters of NT-TRP 172, etc. Any signaling generated by processor 260 is transmitted by transmitter 252. It should be noted that "signaling" as used herein may also be referred to as control signaling. Signaling can be transmitted in physical layer control channels such as the physical downlink control channel (PDCCH). In this case, the signaling can be called dynamic signaling. Signaling transmitted in the downlink physical layer control channel can be called downlink control information (DCI). Signaling transmitted in the uplink physical layer control channel can be called uplink control information (UCI). Signaling transmitted in the sidelink physical layer control channel can be called sidelink control information (SCI).Signaling can be included in higher-layer (e.g., above the physical layer) data packets transmitted in physical layer data channels such as the Physical Downlink Shared Channel (PDSCH). In this case, the signaling can be referred to as higher-layer signaling, static signaling, or semi-static signaling. Higher-layer signaling can also refer to radio resource control (RRC) protocol signaling or media access control-control element (MAC-CE) signaling.

[0107] Scheduler 253 may be coupled to processor 260. Scheduler 253 may be included within T-TRP 170 or may operate separately from T-TRP 170. Scheduler 253 may schedule uplink, downlink, lateral link, and / or backhaul transmissions, including issuing scheduling authorizations and / or configuring schedule-free (e.g., “configuration authorization”) resources. T-TRP 170 also includes memory 258 for storing information and data. Memory 258 stores instructions and data used, generated, or acquired by T-TRP 170. For example, memory 258 may store software instructions or modules executed by processor 260 for implementing some or all of the functions and / or embodiments described herein.

[0108] Although not shown, processor 260 may be part of transmitter 252 and / or receiver 254. Furthermore, although not shown, processor 260 may implement scheduler 253. Although not shown, memory 258 may constitute part of processor 260.

[0109] The processing components of processor 260, scheduler 253, transmitter 252, and receiver 254 can each be implemented by the same or different one or more processors, which execute instructions stored in memory (e.g., memory 258). Alternatively, some or all of the processing components of processor 260, scheduler 253, transmitter 252, and receiver 254 can be implemented using dedicated circuitry, such as a programmable FPGA, hardware accelerator (e.g., GPU or AI accelerator), or ASIC.

[0110] Although the NT-TRP 172 is shown as an example of a drone only, it can be implemented in any suitable non-terrestrial form, such as satellites and high-altitude platforms including international mobile telecommunications base stations and unmanned aerial vehicles. Furthermore, in some implementations, the NT-TRP 172 may have other names, such as a non-terrestrial node, a non-terrestrial network device, or a non-terrestrial base station. The NT-TRP 172 includes a transmitter 272 and a receiver 274 coupled to one or more antennas 280. Only one antenna 280 is shown in the figure to avoid congestion. One, some, or all of the antennas may alternatively be panels. The transmitter 272 and receiver 274 may be integrated as a transceiver. NT-TRP 172 also includes a processor 276 for performing operations related to: preparing transmissions for downlink transmission to ED 110, processing uplink transmissions received from ED 110, preparing transmissions for backhaul transmission to T-TRP 170, and processing transmissions received from T-TRP 170 via backhaul. Processing operations related to preparing transmissions for downlink or backhaul transmission may include operations such as encoding, modulation, precoding (e.g., MIMO precoding), transmit beamforming, and generating symbols for transmission. Processing operations related to processing transmissions received in the uplink or via backhaul may include operations such as receive beamforming, demodulating received symbols, and decoding received symbols. In some embodiments, processor 276 performs transmit beamforming and / or receive beamforming based on beam direction information (such as BAI) received from T-TRP 170. In some embodiments, processor 276 may generate signaling, for example, for configuring one or more parameters of ED 110. In some embodiments, the NT-TRP 172 implements physical layer processing but not higher-layer functions, such as those in the medium access control (MAC) layer or radio link control (RLC) layer. Since this is merely an example, more generally, the NT-TRP 172 may implement higher-layer functions in addition to physical layer processing.

[0111] The NT-TRP 172 also includes a memory 278 for storing information and data. Although not shown, a processor 276 may be part of the transmitter 272 and / or the receiver 274. Although not shown in the figure, the memory 278 may constitute part of the processor 276.

[0112] The processing components of processor 276, transmitter 272, and receiver 274 can each be implemented by the same or different one or more processors, which execute instructions stored in memory (e.g., memory 278). Alternatively, some or all of the processing components of processor 276, transmitter 272, and receiver 274 can be implemented using dedicated circuitry, such as a programmable FPGA, hardware accelerator (e.g., GPU or AI accelerator), or ASIC. In some embodiments, NT-TRP 172 can actually be multiple NT-TRPs that operate together to serve ED 110, for example, by using coordinated multicast transmission.

[0113] T-TRP 170, NT-TRP 172 and / or ED 110 may include other components, but for clarity these components are omitted.

[0114] according to Figure 4 One or more steps of the exemplary methods provided herein can be performed by the corresponding units or modules. Figure 4 Units or modules in the device are shown, such as in ED 110, T-TRP 170, or NT-TRP 172. For example, signals may be transmitted by a transmitting unit or transmitting module. Signals may be received by a receiving unit or receiving module. Signals may be processed by a processing unit or processing module. Other steps may be performed by an artificial intelligence (AI) module or a machine learning (ML) module. The corresponding units or modules may be implemented by hardware, one or more components or devices executing software, or a combination thereof. For example, one or more of these units or modules may be circuits such as integrated circuits. Examples of integrated circuits include programmable FPGAs, GPUs, or ASICs. For example, one or more of these units or modules may be logic, such as a logical function executed by a circuit, a portion of an integrated circuit, or software instructions executed by a processor. It should be understood that if these modules are implemented, for example, by a processor using software, these modules can be retrieved by the processor, wholly or partially, individually or collectively, for processing, in one or more instances, and these modules themselves may include instructions for further deployment and instantiation.

[0115] Further details regarding ED 110, T-TRP 170, and NT-TRP 172 are known to those skilled in the art. Therefore, these details are omitted herein.

[0116] The CSI compression framework in 5G NR is based on the principle that CSI is compressed before the UE feeds back to the terrestrial transmit and receive point (T-TRP) (e.g., gNB). In various examples, this disclosure refers to the T-TRP using either the gNB or the BS as an example; it should be understood that this is merely exemplary and not intended to be limiting. This compression is achieved by leveraging dimensionality reduction techniques that utilize the inherent spatial and temporal correlations in the radio channel. Specifically, the compression scheme operates on a precoding matrix indication (PMI), which represents a recommended precoding vector or precoding matrix, or beamforming matrix, for different transport layers and subbands (which can be multiple consecutive resource blocks).

[0117] 5G NR Release 15 specifies two main compression modes for CSI feedback: Type I and Type II. Type I compression is a codebook-based method, where the UE selects a precoding matrix from a standardized codebook and feeds back the corresponding PMI index. Type II compression is a more flexible approach, allowing the UE to provide both wideband PMI and subband differential PMI sets, thus enabling finer-grained CSI reporting.

[0118] In addition, the 3GPP standard supports the enhanced Type II (eType II) compression mode, which introduces additional compression capabilities. eType II builds upon Type II, adding two enhancement techniques: spatial compression and frequency compression. Spatial compression utilizes correlation across the angular domain, while frequency compression utilizes correlation across adjacent subbands.

[0119] The compression techniques specified in 5G NR Release 15 are not data-driven; that is, they rely on predefined codebooks and compression matrices derived from mathematical models and assumptions about the channel. However, with the emergence of AI and ML technologies, there is growing interest in exploring data-driven approaches for CSI compression that can potentially improve performance by adapting to real-world channel statistics and dynamic characteristics.

[0120] While the non-data-driven CSI compression techniques standardized in 5G NR Release 15 provide a baseline, there is growing interest in exploring AI and ML-based data-driven approaches to CSI compression. These AI / ML-based techniques have the potential to improve compression performance by adapting to real-world channel statistics and dynamic characteristics, rather than relying on predefined codebooks and assumptions.

[0121] One approach that has been investigated is to use a two-sided AI / ML model for CSI compression. In this framework, a pair of AI / ML models (typically deep neural networks) are jointly trained: an encoder model on the UE side and a decoder model on the gNB side. The encoder model on the UE side compresses the high-dimensional CSI (typically a pre-encoded matrix) into a low-dimensional representation, which is then fed back to the gNB. The decoder model on the gNB side then reconstructs the original CSI based on this compressed representation.

[0122] The advantage of the two-sided model lies in its ability to support end-to-end optimization of the compression and reconstruction processes. By jointly training the encoder and decoder models on representative channel data, the two-sided model can learn the most effective compression strategies and adapt to the specific channel characteristics of the deployment scenario.

[0123] During the training phase, the dual-side models are trained on a large dataset implemented through the contact channel, typically obtained through simulation or field measurements. The encoder and decoder models are iteratively updated to minimize a loss function that quantifies the reconstruction error between the original CSI and the reconstructed CSI. After training, the encoder model is deployed on the UE side, while the decoder model is deployed on the gNB side.

[0124] A key aspect of two-sided models is the need for collaboration between the UE and NW vendors during training. Since the encoder and decoder models are interdependent, the respective vendors of these models need to coordinate to ensure compatibility and interoperability. Various approaches, such as sequential training, joint training, or model exchange, have been proposed to facilitate this collaboration.

[0125] Two-sided AI / ML models for CSI compression have shown promising results in terms of compression performance and adaptability to real-world channel conditions. However, several challenges remain, including inter-vendor collaboration, model complexity management, and deployment considerations. This research aims to address these challenges and lay the foundation for the practical implementation of AI / ML-based CSI compression technology in future wireless communication systems.

[0126] We will now discuss some previously proposed methods for CSI compression techniques based on AI / ML.

[0127] To improve the interoperability of two-sided AI / ML models for CSI compression developed by different vendors, a distributed training scheme based on a standardized dataset is proposed. This approach introduces a central repository, a database accessible to all participating vendors, containing a large amount of channel instance data. This data typically comes from simulations or real-world measurements, covering the statistical characteristics and dynamic behavior of actual wireless channels, thus providing a common foundation for model training.

[0128] A central database serves as a unifying element, enabling UE and NW vendors to train their respective encoder and decoder models using the same underlying data distribution. This shared foundation theoretically lays the groundwork for seamless interoperability, ensuring that trained models can effectively communicate and collaborate regardless of their specific design or implementation.

[0129] Figure 5 An example of this concept is illustrated, describing how both the NW (New Wireless) vendor and the UE (User Equipment) vendor access a standardized dataset of the precoding matrix (V) from a central database. Each vendor then independently trains its autoencoder (AE) model using the shared dataset, which consists of an encoder model and a decoder model. Once training is complete, the NW vendor deploys the decoder portion of its model (often called the CSI generator or CSI decompressor), while the UE vendor deploys the encoder portion (often called the CSI compressor), thus achieving end-to-end CSI compression and reconstruction.

[0130] like Figure 5 As shown, the precoding matrix V is stored in the central database 502 and is accessible to both the NW (Network Wi-Fi) vendor and the UE (User Equipment) vendor, used for training on the NW side and the UE side respectively. For example, each matrix V can have dimension N. BS × r, where N BS Let r represent the number of antennas at BS, and r represent the rank produced by the singular value decomposition (SVD) used to generate V. In the NW-side vendor training, the AE model (denoted as ENC) is generated by the encoder model (denoted as ENC). NW 504) and decoder model (denoted as DEC) NW 506) Composition) by minimizing a certain loss function (e.g., min(|v'–v|) 2 Similarly, in the UE-side vendor training, another AE model (represented as ENC) is trained by the encoder model (represented as ENC). UE 508) and decoder model (denoted as DEC) UE 510) Composition) by minimizing a certain loss function (e.g., min(|v'–v|) 2 The UE-side encoder (ENC) is used for training. After training, the trained UE-side encoder (ENC) is used for further training. UE 508 during inference and the trained NW-side decoder DEC NW Used in conjunction with 506.

[0131] The advantage of this approach lies in its simplicity and the potential to minimize direct collaboration between vendors, without requiring the sharing of trained models among vendors (as trained models are considered highly proprietary). Vendors can bypass complex coordination and data exchange processes, instead using a central database as a trusted source of training data, thus simplifying the development process and promoting interoperability.

[0132] However, relying solely on standardized datasets has limitations. The inherent diversity of model architecture, parameter (or meta-parameter) selection, training methods, and even the randomness of AI / ML training algorithms can lead to differences in the learned latent representations (used as PMI mappings between the encoder and decoder). This means that even when trained on the same dataset, models from different vendors may have different internal interpretations of the information reflected in the distribution of the PMI mapping layers, which can hinder seamless communication and collaboration during inference.

[0133] In summary, while centralized training datasets provide a good starting point for achieving interoperability, additional mechanisms are needed to ensure consistent and reliable performance across a diverse ecosystem of vendors and devices.

[0134] Sequential training methods offer a potential solution to the interoperability challenges inherent in two-sided AI / ML models used for CSI compression. By setting a predetermined order for model development, it is possible to ensure consistency between the encoder and decoder in the latent representation (PMI mapping layer), i.e., the internal data structure learned by the model.

[0135] As outlined, this process is typically initiated by the NW vendor. The NW vendor trains its AE model using precoding matrices from a central database, thereby creating a baseline latent representation. Subsequently, they share both the input data and the corresponding latent layer outputs, essentially providing the UE vendor with a "key" to unlock compatible coding.

[0136] This approach simplifies the collaboration process and ensures compatibility between the final deployed models. However, it also introduces potential drawbacks that need to be considered. Primarily, it places NW vendors in a dominant position, as their initial model training determines the upper limit of the UE vendor's encoder performance. This could stifle innovation and limit the UE vendor's ability to differentiate its products based on superior compression performance.

[0137] Furthermore, the inherent differences in design philosophy between UE vendors and NW vendors exacerbate this problem. UE vendors, constrained by the power and size of user equipment, typically prioritize smaller, more efficient models, even at the cost of some performance. The efficiency of these models is closely related to the sparsity of CSI data, which in turn is affected by the degree of data localization. In contrast, NW vendors, with less stringent hardware constraints, generally tend to favor larger models with greater generalization capabilities to cover a wider range of scenarios and configurations.

[0138] Figure 6A and Figure 6B This illustrates an example of this difference between NW vendor priorities and UE vendor priorities.

[0139] In this example, BS 602 can communicate with user equipment (UE) 604 through multiple different channels, which may include relays (e.g., via building 606). BS 602 may have a large number of antennas (e.g., N). BS = 512), while UE604 can have a fewer number of antennas (e.g., 16). Furthermore, BS 602 can focus on the channel conditions of all UEs in its service area 608, while UE 604 can focus only on its own service area 610. This means that although the trained NW-side DEC NW 506 might be more complex, but ideally, the trained UE-side ENC... UE The 508 is preferably significantly smaller in size to avoid placing a greater burden on the UE 604 (which has more limited computing power, battery capacity, and memory resources compared to the BS 602), such as Figure 6B As shown.

[0140] This difference in priority can lead to an imbalance in the sequential training process. When the NW vendor decides on the potential representation, the UE vendor's encoder model may have to learn irrelevant information, adding unnecessary complexity and potentially reducing efficiency. This ultimately puts the UE vendor at a disadvantage and limits its ability to optimize models for specific use cases or device capabilities.

[0141] Generally, it can be expected that the complexity and / or volume of communication in wireless communication networks will increase with technological advancements. For example, the technological advancements from 5G to 6G are discussed below, but this disclosure is not intended to be limited to 6G embodiments or any other particular generation of technology. In 6G very large-scale MIMO systems, the size of the wireless interface has been significantly expanded compared to its predecessor, 5G. This increase is reflected not only in the massive scale of the infrastructure but also in the complexity of the information it processes.

[0142] Figure 7A direct comparison is provided, highlighting the differences. While 5G Release 19 massive MIMO systems typically operate with a bandwidth of 50 MHz, using 64 antennas at the gNB and 8 antennas at the UE, 6G envisions a much larger scale. With bandwidth reaching 400 MHz, the gNB is equipped with 512 antennas, and the UE with 16 antennas, resulting in an exponential increase in the amount of CSI that needs to be processed and switched within the network.

[0143] This increase in scale has a ripple effect on all aspects of the system. The size of the channel matrix (representing the propagation characteristics between the transmitter and receiver) increases dramatically. This, in turn, leads to more complex precoding matrices (V), which are crucial for beamforming and spatial multiplexing techniques. Therefore, the dimensionality of CSI explodes, creating a data deluge that challenges traditional compression and feedback methods.

[0144] Furthermore, the increased number of antennas and bandwidth leads to a greater number of MIMO subchannels. This means richer spatial and frequency domain resources, offering greater potential for spatial multiplexing and higher data rates. However, this also means that AI / ML models used for CSI compression need to be more advanced and capable of handling this additional complexity. This results in larger models, increased training data requirements, and higher computational demands.

[0145] The most direct consequence of the expanded 6G air interface is a dramatic surge in storage demands for central databases. Compared to 5G, 6G's data footprint could increase by 100 times, or even more. Let's analyze these contributing factors to understand the scale of this data explosion.

[0146] First, the bandwidth increased from 50 MHz in 5G to 400 MHz in 6G, resulting in an eightfold increase in the amount of frequency domain information that needs to be stored. Second, the number of antennas at the base station surged from 64 to 512, leading to an eightfold increase in spatial domain data. Considering these two factors, the data volume has increased 64 times compared to 5G.

[0147] In addition, factors such as the number of users, the number of cells in the network, and the frequency of data collection also affect overall storage requirements. Furthermore, decisions regarding data retention strategies and the need for historical data used for model training and analysis will further influence storage footprint.

[0148] Taking these factors into account, it is obvious that maintaining such a large-scale central database is no easy task.

[0149] Figure 8 This demonstrates the potential challenges of managing this data deluge, highlighting the need for robust and scalable infrastructure, efficient data management strategies, and advanced data security measures.

[0150] like Figure 8 As shown, the precoding matrix V in 5G R19 may be relatively small, while the precoding matrix V in 6G is expected to be significantly larger. Therefore, the central database in 6G (Release 20) is expected to be at least 100 times larger than the central database in 5G R19.

[0151] Beyond data storage challenges, the increase in scale to 6G will inevitably require a significant increase in the size and complexity of AI / ML models themselves. As the amount of CSI that needs to be compressed grows, the demand for more advanced models capable of handling this increased dimensionality and complexity is also growing.

[0152] Figure 9 This paper describes this challenge, illustrating how expanded air interfaces and a greater number of MIMO subchannels drive the growth of AI / ML models. Models must be able to effectively capture the complex spatial and frequency correlations within the channels, requiring more parameters and more sophisticated architectures. This directly leads to an increased demand for training data, as larger models require more diverse and representative samples to learn effectively and avoid overfitting.

[0153] like Figure 9 As shown, due to the increased size of the precoding matrix V in 6G, the encoder and decoder models in 6G are expected to be significantly larger and more complex than those in 5G.

[0154] The impact of larger AI / ML models is multifaceted. Significantly longer training times necessitate more powerful computing infrastructure and may lead to bottlenecks in model development. The computational resources required for model inference also increase, potentially straining the capabilities of UE and gNB. Furthermore, the energy consumption associated with training and running (inference) these models becomes a critical issue, impacting the overall energy efficiency and sustainability of the system.

[0155] These two key factors—the exponential growth in data storage needs and the demand for increasingly complex AI / ML models—are converging to severely hinder the practical implementation of CSI compression methods based on a central database in 6G. The sheer scale of data and the computational demands of large models are pushing current technological capabilities to their limits, raising concerns about the sustainability and efficiency of this centralized approach.

[0156] The pressure on network resources, potential bottlenecks in data transmission, and the energy consumption associated with managing and processing such massive amounts of information all indicate the need for fundamental change. While the centralized model was effective in smaller-scale systems in previous generations, it is insufficient to meet the demands of 6G.

[0157] The move towards 6G will bring about a major transformation in antenna design. 6G no longer confines itself to the traditional assumption of uniform planar arrays, paving the way for more diverse and irregular antenna configurations. This shift presents both challenges and opportunities, especially for CSI compression technology.

[0158] When dealing with non-traditional antenna arrays, the limitations of traditional CSI compression methods become apparent. Methods that rely on predefined codebooks or structured assumptions about the channel struggle to effectively capture the unique characteristics and spatial correlations present in irregular antenna layouts.

[0159] However, this challenge also provides ample room for innovation. One potential area of ​​exploration is the CSI-RS antenna port quiescent scenario. By selectively activating specific antenna ports, whether for energy conservation or due to hardware limitations, irregular antenna patterns can be created. This deviation from uniformity allows for the development of more adaptive and efficient CSI compression techniques that can take advantage of the inherent sparsity and unique spatial correlations present in such configurations.

[0160] Against this backdrop, eigenvector-based CSI compression has emerged as a promising candidate. By focusing on the principal eigenvectors of the channel matrix that represent the most salient spatial features, compression ratio and reconstruction accuracy can be improved compared to traditional methods. Furthermore, by combining AI / ML techniques, models capable of learning and adapting to the specific characteristics of different irregular antenna layouts can be developed, thereby further improving performance and flexibility.

[0161] The potential advantages extend beyond simple port-quiet scenarios. Exploring more complex, unconventional antenna configurations, such as cylindrical or spherical arrays, can yield even greater gains. These configurations offer unique beamforming capabilities and spatial diversity, enabling more efficient use of the radio spectrum and potentially significantly improving CSI compression and overall system performance.

[0162] When antenna elements are arranged in a non-uniform manner, this deviation from uniformity poses a challenge to traditional codebook generation methods, especially those that rely on the principle of discrete Fourier transform (DFT).

[0163] Release 18 of 5G, as part of the 3GPP 5G network specification, introduced Type II and eType-II codebooks. These codebooks utilize PMIs to describe channel characteristics in multi-antenna systems, which facilitates efficient CSI feedback. However, like many DFT-based methods, these codebooks fundamentally assume that antenna elements are evenly spaced within the array. This simplifies the codebook generation process and supports efficient mapping between precoded vectors and their corresponding PMIs via simple DFT.

[0164] However, when dealing with irregular antenna arrays with non-uniform spacing between elements, the assumptions of DFT-based methods no longer hold. The intrinsic relationship between the physical antenna layout and the spatial characteristics of the signal becomes more complex, rendering traditional DFT-based codebook generation techniques inapplicable.

[0165] Although eigenvector channels are generally still present in the Fourier transform, directly applying DFT-based methods to generate codebooks presents problems when dealing with a wider signal space that goes beyond the limitations of traditional DFT assumptions. Irregular antenna arrangements disrupt the direct mapping between spatial features and frequency domain representations, necessitating new codebook generation methods.

[0166] This challenge becomes even more apparent when considering scenarios where antenna elements can be dynamically activated or deactivated, such as in energy-saving technologies or when dealing with faulty hardware. Dynamic changes in antenna array configuration further complicate the task of generating a codebook that accurately reflects the dynamic spatial characteristics of the system.

[0167] Therefore, exploring alternative methods for codebook generation is crucial to fully unlocking the potential of irregular antenna arrays in 6G. These methods need to take into account the non-uniform spacing of antenna elements and adapt to dynamic array configurations, thus laying the foundation for efficient CSI compression and feedback in future wireless communication systems.

[0168] To facilitate seamless data exchange between vendors and promote interoperability of AI / ML models used for CSI compression, a central database can be introduced. This centralized repository (such as a 3GPP file server or a dedicated core network entity) serves as a platform where vendors can share and access standardized datasets used for model training.

[0169] The process begins with NW and UE vendors registering their respective datasets in a central database. These datasets are carefully curated and prepared according to predefined standards, covering a wide range of channel implementations and scenarios to reflect the diverse situations encountered in real-world deployments. By making these datasets readily available, the central database creates a collaborative environment where vendors can leverage a common data pool to train their models.

[0170] UE vendors can then access the relevant datasets from the central database to train their encoder models. Similarly, NW vendors can access the same datasets to train their decoder models. This shared access to training data ensures that the resulting models are compatible and can interoperate effectively, regardless of the specific vendor or implementation.

[0171] Now for reference Figure 10 The example illustrates data exchange using a central database. In this example, the process includes a data acquisition step 1002, where the acquired data (e.g., in the form of a channel matrix H, containing channel estimates) can be processed using down-rank SVD to obtain a precoding matrix V. Then, in the next step 1004, the precoding matrix V is registered to the central database 1006. At some point later in step 1008, the NW vendor and / or UE vendor can retrieve V from the central database 1006.

[0172] To ensure smooth and consistent operation, the registration and retrieval process for the central database is standardized. This includes defining common data formats, metadata specifications, and access protocols. By adhering to these standards, vendors can seamlessly exchange data and ensure compatibility across different platforms and implementations.

[0173] Security and access control mechanisms are a critical component of central database infrastructure. Robust authentication and authorization protocols are implemented to ensure that only authorized vendors can access and contribute data. Furthermore, data encryption and integrity checks are employed to protect the confidentiality and integrity of stored information.

[0174] The registration process should include a well-defined set of steps, including data format validation, metadata submission, and quality checks. These measures ensure that the submitted datasets conform to established standards and are suitable for model training. Furthermore, access control mechanisms are implemented to ensure that only authorized vendors can register datasets and to grant appropriate permissions for data access and modification.

[0175] Similarly, the acquisition process follows standardized protocols to ensure consistency and ease of use. Vendors can query the database using specific criteria (such as scenario type, frequency range, or antenna configuration) to identify relevant datasets that meet their specific needs. The acquired datasets are then transmitted in a standardized format for seamless integration into the vendor's training pipeline.

[0176] By implementing these standardized processes, the central database establishes a reliable and efficient platform for data exchange, promotes interoperability, and accelerates the development of robust and high-performance AI / ML models for CSI compression.

[0177] As a critical hub for data exchange and collaboration, the central database requires efficient and secure communication channels. Vendors have multiple options for interacting with this central repository, each meeting different needs and scenarios.

[0178] In the core network, application programming interfaces (APIs) have emerged as a powerful solution. Standardized APIs provide vendors with a structured, well-defined interface for registering, retrieving, and querying datasets. This approach facilitates efficient data transfer, enhances security, and simplifies integration with existing systems. Alternatively, established technologies such as the Secure File Transfer Protocol (FTP) can be used to upload and download datasets. While simpler to implement, FTP lacks the flexibility and efficiency of APIs for complex data interactions. For vendors seeking more granular control and manipulation of data, one option is to interact directly with the database using standardized query languages ​​such as SQL. However, this requires expertise in database administration, which, if not carefully managed, can introduce potential security vulnerabilities.

[0179] In the wireless domain, the air interface provides a direct communication path between user equipment (UE), base station (gNB), and central database. Dedicated control channels can be used for low-latency data exchange, ensuring dedicated resources for this critical function. However, this approach requires the allocation of additional radio resources and raises concerns about potential interference.

[0180] Existing user plane data channels provide an alternative for transmitting smaller datasets or model updates (not CSI encoder or decoder models, but models that help register, store, and retrieve data). Utilizing existing infrastructure eliminates the need for additional resource allocation, but may be unsuitable for large datasets due to bandwidth limitations. In scenarios where device-to-device communication is available, sidelink channels offer an efficient way to exchange local data, thus reducing the burden on the core network. However, the range and reliability of sidelink channels are limited compared to cellular connections.

[0181] Regardless of the communication method chosen, establishing standardized signaling mechanisms and protocols is paramount. This includes defining consistent message formats, implementing robust authentication and authorization protocols, using data encryption, and ensuring data integrity through checksums or other mechanisms. The chosen solution should prioritize scalability to accommodate the growing demands of 6G, while maintaining security, efficiency, and flexibility to adapt to different use cases and deployment scenarios.

[0182] To illustrate the process of collecting and registering data in the central database, the journey of a single data sample is traced from its initial capture at the UE to its eventual storage in the central repository. This detailed breakdown provides a clear understanding of the relevant steps, which can be easily extended to the collection, registration, and acquisition of large numbers of data samples.

[0183] Now for reference Figure 11 This illustrates an example of how data collected at the UE can be sent to a central database in the network, as provided in this disclosure. Figure 11 The data acquisition at the UE is shown as 1102. For example, using a resource block (RB) or subband (f k On the CSI-RS, the UE performs channel estimation to obtain the channel matrix H(f) k Then, the precoded matrix V is obtained using reduced-rank SVD.

[0184] Given that real-time processing is not strictly required during data acquisition, a dedicated role can be assigned to the UE to collect the raw precoded data or matrix (V). This involves dedicating a portion of the UE's resources to this task, thereby ensuring efficient and centralized data acquisition. The precoded data or precoded matrix (V) is processed in subbands (f k The unit is f. To simplify the discussion below, f is omitted. k The methods discussed below can be readily applied to various sub-bands, resource blocks, or resource units.

[0185] The process begins with the UE performing channel estimation on the downlink CSI-RS (i.e., the reference signal (pilot) in the downlink from the gNB to the UE). This step is necessary because it allows the UE to estimate the current state and characteristics of the radio channel. Subsequently, the UE performs a decomposition of the estimated channel to identify the optimal precoder, i.e., a beamforming vector that maximizes signal quality and minimizes interference. Once determined, this precoded data is typically transmitted back to the network via the uplink data channel in the form of a precoding matrix (number of BS antennas × number of MIMO layers).

[0186] To provide context and ensure the usability of the acquired data, the UE also collects and reports auxiliary information. This includes details such as the specific subband or resource block (RB) used during data acquisition and the geographic location of the data acquisition. These additional parameters provide valuable insights into the environmental conditions and system configuration for data capture, thereby enhancing the overall value and applicability of the dataset. This auxiliary information can be sent to the network along with the precoding matrix V.

[0187] Optionally, to optimize resource utilization and reduce transmission overhead, the precoded data can be compressed before being sent to the network. Depending on the specific requirements and capabilities of the system, various compression techniques can be employed, such as quantization-based or dimensionality-reduction-based compression techniques.

[0188] This comprehensive data acquisition process ensures that the central database is populated with a rich and information-rich dataset, providing a solid foundation for training robust and high-performance AI / ML models for CSI compression. This is an exemplary and general design for protocols and signaling between the UE, base station (gNB), and network, focusing on the physical and MAC layers over the air interface.

[0189] Below is an exemplary and general design of protocols and signaling between the UE, base station (gNB), and network, focusing on the physical and MAC layers over the air interface.

[0190] First, let's describe an example data acquisition signaling:

[0191] Triggering data collection can include:

[0192] The network can initiate data collection by sending a dedicated RRC message (e.g., AI_DataCollectionRequest) to the UE. This message can include the following parameters:

[0193] o Target CSI-RS resource: Specifies the CSI-RS resources that the UE should use for channel estimation.

[0194] o Data collection duration: Defines the duration or number of time instances used for data collection.

[0195] o Reporting mode: Indicates whether the UE should report data periodically, upon completion, or based on specific events (e.g., handover, channel condition change).

[0196] o Data compression options: Specifies whether and how the UE should compress the acquired precoded data.

[0197] The data collected and reported may include:

[0198] The UE can report the collected data using existing uplink channels (e.g., PUCCH, PUSCH) or potentially via a dedicated control channel. The report may include:

[0199] o Precoding matrix (V): Raw or compressed precoded data acquired from CSI-RS measurements.

[0200] o Auxiliary information: Additional parameters such as RB allocation, geographic location, and timestamp.

[0201] o Data quality indicator: An optional metric that indicates the quality or reliability of the collected data.

[0202] The following describes an exemplary model update signaling. In this context, the term "model" may refer to a set of representations maintained in a central database (discussed further below) rather than the encoder / decoder model. This model (also referred to as the representation set) includes parameters that can be used to ensure consistency in data registration and management across network entities. These parameters may include a representation set ID (e.g., an identifier that helps ensure the correct version of the representation set is used across different network entities), the weights and biases of the neural network (e.g., parameters learned by the trained neural network to implement the encoder / decoder model), hyperparameters used during training (e.g., parameters including the learning rate, batch size, and regularization strength used during neural network training), regularization terms (e.g., terms that help ensure the trained encoder / decoder model operates efficiently while avoiding overfitting), and quantization parameters (e.g., parameters that help ensure the quantization process at the UE and the dequantization process at the network entity are consistent, which may help maintain the integrity of compressed information). These parameters can be transmitted to the UE to maintain synchronization with the central data center.

[0203] Model update indicators (also known as indicators representing set updates) may include:

[0204] • The network can notify the UE of model updates (also known as representation set updates) by sending a dedicated RRC message (e.g., AI_ModelUpdate). This message may include:

[0205] o Model ID (also known as Representation Set ID): A unique identifier for a new model (also known as a new representation set).

[0206] o Model size (also known as representation set size): The size of the model parameters (also known as representation set parameters) to be transmitted.

[0207] o Model transmission method (also known as representation set transmission method): Specifies the method used to transmit model parameters (also known as representation set parameters) (e.g., RRC signaling, user plane data channel).

[0208] Model parameter transfer (also known as representation set parameter transfer methods) can include:

[0209] The network can transmit model parameters (also known as representation set parameters) to the UE in the following ways:

[0210] oRRC signaling: This method is suitable for smaller models (also known as smaller representation sets) and provides reliable transmission through a dedicated control channel.

[0211] o User plane data channel: This approach is suitable for larger models (also known as larger representation sets) and utilizes existing data channels, but may have higher latency and potential impact on user services.

[0212] Model parameters (also known as representation set parameters) can be sent in segments and have appropriate error detection and correction mechanisms to ensure reliable transmission.

[0213] Storing raw precoded data (V) directly in a central database can be inefficient due to its high dimensionality and potential redundancy. To optimize storage and retrieval, using sparse representation becomes crucial. This involves representing the data using a minimal set of basic components, thereby discarding redundant or unimportant information.

[0214] When processing a precoded matrix (V) obtained through a specific decomposition technique such as Singular Value Decomposition (SVD), the resulting column vectors exhibit orthogonality. This means that each user's data columns are linearly independent and cannot be further simplified or compressed by representing these data columns as linear combinations of other vectors. Essentially, they represent the most basic and indivisible components of the data from a reporting user.

[0215] However, while individual precoding matrices lack inherent sparsity, a different form of sparsity emerges when considering data collected under similar wireless environmental conditions. Precoding matrices collected in such environments tend to exhibit strong correlations and similarities. This means that when new data is introduced into a central database, it is likely to be highly correlated with existing data, or highly approximated using sparse representations of data already existing in the database.

[0216] This observation opens a new avenue for efficient data storage. These correlations can be leveraged and represented using coefficients relative to the data basis (also called the representation set) formed by existing data or its sparse representation, instead of storing each new precoding matrix independently. This effectively reduces storage requirements, as only the coefficients and basis vectors (also called representation vectors) constituting the representation set need to be stored, rather than the entire high-dimensional precoding matrix. In some examples, this disclosure may refer to the representation set as the data basis. It should be understood that the term "basis" as used herein is not necessarily a basis in a strict mathematical sense.

[0217] Given the independent nature of each column in the precoding matrix (V), they can be considered as distinct data units (v). i(i=1, 2... MIMO layer number) for storage and management in a central database. Each column vector (representing a unique precoding vector or beamforming vector) is stored as a separate entry for fine-grained access and manipulation of individual data points. This organization facilitates efficient data processing and analysis, enabling the identification of specific precoders or the acquisition of subsets of data based on specific features or conditions.

[0218] For each newly arriving data record, representing a single column beamforming vector or precoder, the central database performs a critical evaluation to determine its potential for sparse representation. The goal is to determine whether this new precoder can be efficiently represented as a sparse combination of existing data in the database. This evaluation process is divided into two distinct scenarios:

[0219] Scenario 1: Sparse Representation Feasible: In this scenario, the new precoder can be approximately, but accurately, represented as a sparse combination of existing data points or basis vectors (also known as representative vectors) in the database. In some examples, this disclosure may refer to the representative vectors as basis vectors. It should be understood that the term "basis vector" as used herein is not necessarily a basis vector in a strictly mathematical sense. This implies that the new precoder has significant similarity and correlation with previously stored information. Therefore, the database incorporates this new data in its sparse form, storing only the correlation coefficients and their corresponding basis vectors (also known as their corresponding representative vectors), thereby minimizing storage requirements and optimizing efficiency.

[0220] Scenario 2: Sparse Representation Not Feasible: If the database cannot identify a suitable sparse combination to represent the new precoder, indicating that it has unique characteristics that distinguish it from existing data, then a different approach is recommended. In this case, the database can choose to store the new precoder in its original form, or it can employ alternative dimensionality reduction techniques to create new basis vectors (also known as new representative vectors), which can be used to represent the new data along with other similar precoders that may be acquired in the future.

[0221] Regardless of the scenario, the process of storing precoded data in a central database begins with computing its sparse representation. This involves utilizing a “data base” (also known as a representation set) derived from the database itself.

[0222] Scenario 1: When represented as v i When a precoded vector or beamforming vector can be represented in a sparse format, it means that the precoded vector or beamforming vector can be accurately reconstructed using a linear combination of several column vectors (also called representative vectors) selected from an established public "data base" (also called the representation set) denoted as Ψ. This data base (also called the representation set) comprises a set of representative precoded vectors that embody the fundamental characteristics of the data distribution in the central database.

[0223] The key to this sparse representation lies in the coefficient vector s. i The vector is primarily composed of zeros, with only a few non-zero elements strategically placed. The indices of these non-zero elements directly correspond to the contributions of v in the data base (also known as the representation set). i A specific column vector (also called a representative vector) is constructed. The values ​​of these non-zero elements represent the weights assigned to each basis vector (also called a representative vector), indicating the relative contribution of these basis vectors to the formation of the original preencoder, such as... Figure 12 As shown.

[0224] like Figure 12 As shown in the example, the precoding matrix V can be represented as a set of precoding vectors v i (Where i is the index of the MIMO layer, from 1 to r). Given a precoding vector v i It can be represented or approximated as a linear combination of representative vectors stored in a central database, denoted as Φ[1] to Φ[K]. The representative vectors together can be considered to constitute a representation set, denoted as Ψ. The coefficient vectors s i Most are zero, with only a few non-zero elements (in Figure 12 (Represented by dark bars) corresponds to the selected representative vector. A coefficient vector s can be used. i The linear combination of coefficients contained in the non-zero elements is used to select a representative vector to approximately recover the precoded vector v. i This means that each precoded vector v i (And therefore the precoded matrix V) can be stored in a central database using a sparse representation.

[0225] This sparse representation offers significant advantages in terms of storage efficiency. It eliminates the need to store the entire high-dimensional precoded vector v. i Instead, only a much smaller coefficient vector s needs to be stored. i And the associated data base (also known as the representation set) Ψ. This method significantly reduces the storage footprint while retaining the ability to accurately reconstruct the original precoder when needed.

[0226] Scenario 2: When the new precoding vector v i When a sparse representation cannot be achieved using the existing data base (also known as the representation set) Ψ, it indicates that the vector possesses unique characteristics that distinguish it from the current set of basis vectors (also known as representation vectors). Instead of discarding this valuable information, it can be integrated into the data base (also known as the representation set) itself to enrich the data base's representational capabilities and enhance its ability to capture the diversity of pre-encoded data.

[0227] One direct method is to add v iThis expands the data base (also known as the representation set) as a new column vector (also known as a new representative vector). This effectively increases the dimension of the base (also known as the representation set), thus providing a new dimension to represent data with the same dimensions as v. i A preencoder for similar features. For this new preencoder, the corresponding sparse representation vector s i (Also known as the coefficient vector, not to be confused with the representative vector discussed above) will be a vector consisting of 0s and ending with a single "1", representing v. i It is directly represented by the newly added basis vectors (also called the newly added representative vectors), such as Figure 13 As shown.

[0228] like Figure 13 As shown in the example, the precoding matrix V can also be represented as a set of precoding vectors v i In this example, it was found that the representative vectors Φ[1] to Φ[K] constituting the set Ψ cannot represent a given precoded vector v. i In other words, the precoding vector v i A linear combination of representative vectors cannot be used to approximate (to achieve an acceptable level of accuracy). Therefore, the precoded vector v can be... i The new representative vector (denoted as Φ[K+1]) is added to the representation set Ψ. Therefore, the representation set Ψ can be updated. After the update, the precoded vector v... i This can now be represented as a linear combination of vectors, for example, using a coefficient vector s. i The coefficient vector has only one non-zero element (in Figure 13 (represented by dark bars), corresponding to the newly added representative vector Φ[K+1] as the selected representative vector.

[0229] This process of incorporating unique precoded vectors into the data base (also known as the representation set) ensures that the central database continuously evolves and adapts to the ever-changing environment of wireless channels and antenna configurations. It supports efficient representations of a wider range of precoders, thereby enhancing the flexibility and effectiveness of AI / ML models used for CSI compression.

[0230] To ensure the continued relevance and efficiency of the central database, a dynamic mechanism can be implemented to track the usage and duration of each basis vector (also known as a representative vector) in the data base (also known as the representation set). This involves associating a timestamp with each column vector (also known as a representative vector) in the data base (also known as the representation set), marking when that column vector was introduced or last updated.

[0231] By monitoring these timestamps, the database can identify basis vectors (also known as representative vectors) that have not been used by newly arriving data for a considerable period. If a particular basis vector (also known as a representative vector) remains unused for an extended period, it indicates that the data it represents may no longer be relevant to the current channel conditions or deployment scenario. In such cases, the database can automatically discard these outdated basis vectors (also known as obsolete representative vectors) to ensure that the data base (also known as the representation set) remains compact and reflects the evolving characteristics of the wireless environment.

[0232] This dynamic evolution of the data base (also known as the representation set) reflects the ever-changing nature of wireless communication networks. As network configuration, user equipment capabilities, and channel conditions evolve over time, the data base (also known as the representation set) should also adapt to maintain its validity and relevance. By removing outdated information and incorporating new representative data, the central database ensures its continued value as the foundation for training accurate and efficient AI / ML models for CSI compression.

[0233] The challenge of efficiently representing precoded data in a central database can be concisely formulated as a mathematical optimization problem. The goal is to find the optimal coefficient vector s. i The goal is to minimize the combination of reconstruction error and sparsity. This optimization problem can be formulated as:

[0234] Find vector s i (f k ), such that: α min (|| v i (f k ) – Ψ(f k )s i (f k )|| 2 ) +β min (|s i (f k )|)

[0235] The first term α×min (|| v i – Ψs i || 2 The reconstruction error is measured using the data basis (also known as the representation set) Ψ and the coefficient vector s. i Quantize the original pre-encoding vector v i The difference between the sparse representation and its reconstructed vector. Minimizing this term ensures that the sparse representation accurately captures the essential information contained in the original precoder.

[0236] The second term β×min (|s i |) By minimizing the coefficient vector s iThe number of non-zero elements in Ψ is used to promote sparsity. This leads to the use of the set of minimal basis vectors (also known as representative vectors) from Ψ to represent the precoder, thereby reducing storage requirements and improving efficiency.

[0237] The parameters α and β control the relative importance of these two objectives. Adjusting these parameters allows for fine-tuning the balance between reconstruction accuracy and sparsity based on the specific needs and constraints of the system.

[0238] A powerful tool for solving this optimization problem is the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm. LASSO effectively identifies the most relevant basis vectors (also known as representative vectors) from Ψ while minimizing the reconstruction error. Other suitable techniques can be used to solve this optimization problem.

[0239] To determine whether a sparse representation is feasible for a given precoding vector, a threshold is established. If the coefficient vector |s i If the obtained L1-norm is below this threshold, it indicates that the precoder can be accurately represented using a sparse combination of existing basis vectors (also known as representative vectors) (Scenario 1, e.g., ...). Figure 12 As shown). Conversely, if |s i If the threshold is exceeded, it indicates that the precoder has unique characteristics that necessitate its inclusion in the data base (also known as the representation set) (Scenario 2, for example, ...). Figure 13 (As shown).

[0240] Recent data analysis of real-world 5G deployments has revealed a key characteristic of wireless channels: their inherent sparsity and spatial consistency. This means that despite their complexity, high-dimensional MIMO signals typically exhibit a considerable degree of redundancy and predictability, especially within localized regions. Channels in adjacent locations often share similar propagation characteristics, reflecting a common underlying physical environment and scattering properties.

[0241] This observation aligns with theoretical expectations and has significant implications for AI / ML-based CSI compression. By leveraging localized data, the model can effectively utilize the spatial consistency of the channel, thereby improving compression ratio and reconstruction accuracy. This is because a model trained on localized data can better capture the nuances and correlations specific to that region, resulting in more efficient (i.e., sparser) representations and better generalization within that particular region.

[0242] As we move towards 6G, with its higher-dimensional channels and more complex antenna configurations, sparsity and spatial consistency become more pronounced. This presents a valuable opportunity to further optimize CSI compression techniques by incorporating localization learning and data-driven approaches.

[0243] Within the s vector space representing the sparse representation of the precoded data, this spatial consistency manifests as a distinct distribution. Precoded vectors from nearby locations tend to cluster together, reflecting their shared features and potential correlations. This spatial distribution can be effectively captured and utilized by AI / ML models trained on localized datasets, leading to more efficient and accurate CSI compression. Figure 14 A general diagram is shown.

[0244] Figure 14 An exemplary BS 1402 communicating with multiple UEs is shown. It should be noted that a first group of UEs 1404A can be centralized in one location and served by a first vendor (e.g., UE vendor-A), while a second group of UEs 1404B can be centralized in another location and served by a second vendor (e.g., UE vendor-B). Each UE vendor can be a corresponding network entity and can have [specific features / equipment / etc.]. Figure 4 The module shown is similar to the module shown. Figure 14 Further examples of coefficient vectors in the simplified s-domain (or sparse domain) 1406 are shown. The coefficient vectors corresponding to the first set of UEs 1404A (i.e., s...) i The vectors are not shaded, while the coefficient vectors corresponding to the second group of UE 1404B are shaded. The spatial correlation of each group of UE 1404A and 1404B can make their corresponding precoding vectors correlated, which can be reflected by the clustering of their corresponding coefficient vectors in the s domain 1406.

[0245] The scalability property of the "data base" (also known as the representation set) Ψ is that, given s i Preserving the spatial relationships and topological properties of precoded vectors within the sparse domain of vector representation offers a key advantage. This capability is crucial, or at least useful, for maintaining spatial consistency and separability, both critical aspects for both NW and UE vendors when retrieving data from a central database.

[0246] Spatial consistency: By incorporating new precoded vectors into the data base (also known as the representation set), spatial correlations between nearby locations can be captured and preserved. This allows the model to effectively learn and utilize these relationships, thereby improving prediction accuracy and generalizing better within a specific region.

[0247] Spatial separability: Data bases (also known as representation sets) also help to separate precoders from different locations or environments. This allows vendors to selectively acquire data relevant to their specific needs, avoiding unnecessary processing of irrelevant or out-of-distribution data. For example, a UE vendor might be interested in training models specifically for urban environments, while an NW vendor might need data for rural deployments. Data bases enable the efficient acquisition of these specific subsets of data, thereby optimizing training efficiency and model performance.

[0248] Benefits for NW and UE vendors: This ability to preserve spatial relationships within a sparse domain offers several benefits to both NW and UE vendors:

[0249] Improved model performance: By utilizing spatially consistent data, manufacturers can train models that are better adapted to specific environments and exhibit higher accuracy in predicting precoders of nearby locations.

[0250] Reduced complexity: The ability to selectively acquire relevant data based on spatial location reduces the amount of data that needs to be processed and stored, thereby reducing computational complexity and improving efficiency.

[0251] Enhanced generalization: Models trained on spatially diverse data from a central database can generalize better to new, unseen locations, thereby improving the overall robustness and adaptability of these models.

[0252] Therefore, scalable data bases (also known as representation sets) play a key role in preserving spatial information within sparse domains, enabling efficient and effective use of pre-coded data to train AI / ML models for CSI compression.

[0253] While data-driven approaches offer the flexibility of scalable bases (also known as representation sets), fixed-base-dependent methods such as the Discrete Fourier Transform (DFT) and the Cosine Discrete Transform (CDT) face limitations when dealing with irregular antenna arrays and varying channel conditions. These fixed bases are predetermined and lack the ability to adapt or extend to incorporate unique features of new information or data.

[0254] This lack of flexibility becomes a significant drawback when dealing with irregular antenna configurations, where the spatial relationships between antenna elements deviate from the uniform antenna array assumptions upon which DFT and CDT are based. Consequently, these transformations struggle to accurately capture the complex spatial correlations and nuances present in such scenarios, resulting in suboptimal performance in tasks such as beamforming and CSI compression.

[0255] This limitation is particularly evident in the challenges faced by the DFT-based Type II codebooks introduced in 5G NR Release 18. While these codebooks provide efficient CSI compression for uniform antenna arrays, they struggle with irregular antenna layouts, leading to performance degradation and reduced accuracy.

[0256] Fixed bases lack scalability, limiting their ability to adapt to the evolving environment of wireless communication systems. As 6G adopts more diverse antenna configurations and explores the potential of non-traditional arrays, the need for data-driven approaches using scalable bases becomes increasingly critical to achieve optimal performance and flexibility in CSI compression and other mission-critical tasks.

[0257] The flexibility of the central database disclosed in this article supports different data acquisition strategies to meet the specific needs and priorities of both network vendors and UE vendors. While NW vendors may choose a more comprehensive approach, extracting a larger data base (also known as a representation set) to cover a wider range of scenarios and configurations, UE vendors typically prioritize the advantages of localized data, focusing on a smaller, more specific subset of the data base (also known as a representation set).

[0258] Figure 15 The example illustrates this concept. In this example, BS 1402 communicates with a first group of UEs 1404A (served by UE vendor A, the first vendor) and a second group of UEs 1404B (served by UE vendor B, the second vendor), as previously described. UE vendor A and UE vendor B each extract different sub-data bases (also referred to as subsets of representation set 1502) (e.g., UE vendor A extracts representation subset 1504A, and UE vendor B extracts representation subset 1504B) to reflect their specific regions of interest within the overall area. This localization approach enables these vendors to leverage the spatial consistency and correlation present within their respective areas, thereby building more efficient and accurate models for their targeted deployments.

[0259] In contrast, BS 1402 addresses channel conditions over a broader area 1408. Therefore, the NW vendor (serving BS 1402) acquires the entire data base (also known as the representation set 1502), which includes data relevant to both UE vendor-A and UE vendor-B, as well as additional data from other areas or scenarios. This broader perspective enables network vendors to develop more generalized and adaptive models, thus handling a wider range of channel conditions and UE configurations.

[0260] Localized data not only supports customized model training but also provides opportunities for further optimization of sparse representations, particularly for UE vendors. By focusing on specific regions or subsets of the data base, UE vendors can achieve more compact and efficient pre-encoder representations within their domain of interest.

[0261] Figure 16 An example is shown. For example, in Figure 16 In the process, for UE vendor A, by extracting relevant parts from the global s vector space 1406 (also known as the global s domain 1406), a coefficient vector s is derived. UE_A The localized sparse vector subspace is extracted. This extraction process involves selecting elements in s that correspond to representative vectors in the subdata basis (also known as representation subset 1504A) used by UE vendor A. The resulting subspace retains the essential information needed to represent the precoder within a specific region of interest while discarding irrelevant components, resulting in a more compact and efficient representation. Consider the coefficient vector s corresponding to the selection of representation subset 1504A. UE_A (i.e., s) UE_A It only has non-zero elements corresponding to the representative vectors contained in the subset 1504A. UE_A This can be stored in the central data center using an extended vector form 1602, which includes elements corresponding to all representative vectors contained in the complete representation set 1502. To utilize resources more efficiently, s UE_A This can be represented using a more compact vector form 1604A, which corresponds only to the representative vectors contained in the representation subset 1504A. This means that, compared to the representation of the same coefficient vectors at the central data center, UE vendor A can represent the coefficient vectors within the region of interest using a more compact vector form.

[0262] To ensure accurate data acquisition and reconstruction, UE vendors documented the data used to generate the more compact 1604A. UE_A The specific extraction method is as follows. This information (denoted as Mapping_A), along with the sub-data base (also known as representation subset 1504A) itself, enables UE vendors to effectively utilize sparse representations to train their models and accurately reconstruct the precoder within their target regions. As will be discussed further below, this stored mapping information can be used to convert the coefficient vector from compact vector form 1604A (corresponding to representation subset 1504A) back to extended vector form 1602 (corresponding to the complete representation set 1502). Similarly, UE vendor B can use another mapping (denoted as Mapping_B) to convert the coefficient vector s UE_B The coefficient vector s maps from the extended vector form 1602 to the more compact vector form 1604B. UE_B It is selected as the representation subset 1504B used by UE vendor B. It should be noted that since the size of the representation subset 1504A used by UE vendor A and the representation subset 1504B used by UE vendor B can be different, the sizes of the compact vector forms 1604A and 1604B can also be different.

[0263] Figure 17 An example of training a UE-side encoder (i.e., training a neural network to perform pre-encoder encoding tasks, which will be deployed at the UE) is shown. This training can be performed by any network entity, such as at the UE vendor. Figure 17 In the context of CSI compression using a central database and sparse representation, the training process of the UE-side encoder (performed by UE vendor A in this example) can be viewed as an optimization problem within a minimum mean square error (MMSE) framework. The goal is to find the optimal solution when provided with synthesized precoded data (v) and the corresponding sparse representation (s). UE_A When minimizing the original sparse representation (s) UE_A ) and the output of the encoder network (s UE_A The difference between =ENC(v; β)).

[0264] The following is Figure 17 A detailed breakdown of the process shown in the example:

[0265] 1. Obtaining the data base and sparse representation: UE vendor A first obtains the relevant "data base" (also known as the representation subset) (Ψ) from the central database. UE-A 1504A, as mentioned above, is a subset of Ψ; for convenience, (f) can be omitted. k (symbols) and a set of data samples (s) in sparse representation (or compact vector form 1604A). UE_A As mentioned earlier, is a subset of s; in some examples, s UE_A It can be simplified to s A These constitute the training encoder model 1702A (which can be represented as ENC). UE_A The foundation of ). This will be about the transition from s to s UE_A and Ψ to Ψ UE-A The mapping information is recorded as mapping A. The mapping information can be in the form of a transformation matrix, for example, it can be recorded at a central data center.

[0266] 2. Synthesize precoded data: Use the obtained representation subset (Ψ) UE-A ) and coefficient vector (s UE_A The UE vendor synthesizes the corresponding precoding vector (v) from the sparse representation of the data. This involves multiplying the data basis matrix (also known as the representation subset matrix) with the sparse representation vector to reconstruct the original precoding data: v = Ψ UE-A s UE-A

[0267] 3. Training the encoder network: Synthesized data pairs consist of the pre-encoder (v) and its corresponding coefficient vector (s).UE_A The sparse representation of is used to train the encoder network (s' UE_A =ENC(v; β), where s' UE_A This represents the encoder's output, to distinguish it from the true value s. UE_A The training process aims to optimize the encoder's parameters (β) to minimize the original sparse representation (s) given the synthesized pre-encoded data and its sparse representation. UE_A The mean square error between the encoder output and the target signal. MMSE optimization framework: The training objective can be mathematically expressed as: min (|s UE_A – ENC(Ψ UE-A s UE_A ; β)| 2 Or more simply min(|s) UE_A – s' UE_A | 2 (It should be noted that this is equivalent to min (|s') UE_A – s UE_A | 2 This equation reflects the goal of minimizing the difference between the original sparse representation and the encoded sparse representation. The specific method used to achieve this optimization is determined by the UE vendor. Deep neural networks are a common choice, allowing the model to learn the optimal parameters (β) through a data-driven approach. However, other optimization techniques or machine learning algorithms can also be explored.

[0268] This framework allows UE vendors to flexibly design and optimize their encoder models based on their specific requirements and priorities. These UE vendors can choose the model architecture, training algorithms, and optimization techniques that best suit their needs, while still ensuring compatibility with the standardized data base and sparse representation format used in the central database through recorded mapping information.

[0269] Figure 18 An example of training an NW side encoder (i.e., training a neural network to perform a pre-encoder decoding task, which will be deployed at network nodes, such as BS or TRP) is shown. This training can be performed by any network entity, such as at an NW vendor (which can have...) Figure 4 The module shown is similar to the module that executes the code. Unlike the UE-side encoder, the network-side decoder operates in slightly different scenarios, such as... Figure 18 As shown. Although the data base used on the network side (also known as representation set 1502) (Ψ) NW It is mainly composed of representative vectors that are not necessarily strictly orthogonal, but its scalability introduces a certain degree of compression leeway.

[0270] Therefore, the network-side decoder 1802 (v=DEC) NWThe role of (s;θ) is not limited to reconstruction. It also acts as a compressor to further refine the sparse representation received from the UE and recover the original precoded data with minimal information loss.

[0271] During training, the raw pre-encoded data is synthesized from coefficient vectors s sampled from a dataset maintained in a central data center. For example, as... Figure 18 As shown, the coefficient vector (in the original extended vector form 1602) can be obtained, and then it is combined with the representation set Ψ. NW Multiply to obtain the precoded vector v. The optimization objective of the decoder network is to minimize the mean square error between its output (which can be represented as v') and the original precoded data. This can be mathematically expressed as:

[0272] min(|DEC NW (s;θ) – v| 2 )

[0273] Where s represents the coefficient vector (which will be the sparse representation received from the UE during inference; and the sampled training data during training), θ represents the decoder parameters, and v is the original precoding vector.

[0274] Similar to the UE-side encoder, the specific method for training the decoder network is determined by the NW vendor. Neural networks offer a powerful and flexible approach, enabling the model to learn optimal parameters (θ) through a data-driven process. However, other optimization techniques or machine learning algorithms can also be explored based on the NW vendor's specific requirements and preferences.

[0275] Figure 19 Examples of how trained encoder and decoder models can be used in inference are shown. During training (e.g., as combined above) Figure 17 and Figure 18 After that, the trained encoder model can be deployed to UE 110 (e.g., an encoder model trained by UE vendor A (denoted as ENC)). UE_A 1702A) can be deployed to UE_A 110A served by UE vendor A; encoder model (denoted as ENC) trained by UE vendor B UE_B 1702B) can be deployed to UE_B 110B served by UE vendor B, while the trained decoder model (e.g., denoted as DEC) NW 1802) can then be deployed to network node 170, such as BS or T-TRP. Mapping information (e.g., Mapping_A and Mapping_B) can be provided to network entities from the central data center (e.g., via control signaling or other suitable signaling).

[0276] UE_A 110A and UE_B 110B can be similar to ED 110 as previously described, including Figure 4 The units and modules shown. For simplicity, Figure 19 These units and modules are omitted. The trained encoder model ENC UE_A 1702A can be provided to UE_A 110A by UE vendor A after training, as previously combined. Figure 17 As described above. In some examples, the encoder model ENC... UE_A 1702A can be a submodule of any previously described unit or module of UE_A 110A. For example, the encoder model ENC. UE_A 1702A can be implemented in the logic of the processing module of UE_A 110A. Typically, the encoder model is ENC. UE_A The 1702A can be implemented as software, hardware, or a combination of software and hardware (e.g., implemented as an integrated circuit and / or as software instructions executed by a processor, etc.).

[0277] UE_A 110A can receive a reference signal through its receiving module. UE_A 110A uses this reference signal to calculate the channel matrix H, which can then be decomposed into a precoding matrix V. Each vector in the precoding matrix V is called a precoding vector v. A UE_A 110A can generate precoding vectors v A And use the trained encoder model ENC UE_A The 1702A will precode the vector v A Encoded as a coefficient vector (which can be represented as s) UE_A Or simpler s A In this case, the coefficient vector is in compressed vector form 1604A. The coefficient vector can be used as a precoding matrix indicator (denoted as PMI). A The PMI A (For example, the coefficient vector s in compressed vector form 1604A) UE_A Appropriate signaling can be sent to network node 170.

[0278] Network node 170 can be similar to the previous reference Figure 4 The T-TRP 170 is described above, and may include the units and modules discussed in conjunction with this diagram. At network node 170, a receiving module can receive PMI. A (That is, the coefficient vector s in compressed vector form 1604A) UE_ANetwork node 170 has stored mapping information (e.g., Mapping_A) in memory, which is used to transform or map the coefficient vector from compressed vector form 1604A to extended vector form 1602. This enables the use of the trained decoder model DEC. NW 1802 decodes the coefficient vector in extended vector form 1602 to recover the precoded vector v' A The trained decoder model DEC NW 1802 can be provided to network node 170 by the NW vendor after training, as previously combined. Figure 18 As described above. In some examples, the decoder model DEC NW 1802 can be trained by network node 170 itself. In some examples, the decoder model DEC... NW 1802 can be a submodule of any previously described unit or module of network node 170. For example, the decoder model DEC. NW 1802 can be implemented in the logic of the processing module of network node 170. Typically, the decoder model is DEC. NW 1802 can be implemented as software, hardware, or a combination of software and hardware (e.g., implemented as an integrated circuit and / or as software instructions executed by a processor, etc.).

[0279] Similar to the operation of UE_A 110A, UE_B 110B can generate a precoding vector v. B And using the trained encoder model ENC UE_B 1702B will precode vector v B Encoded into a precoded matrix indicator (represented as PMI) B In this case, the precoding matrix indicator is a coefficient vector s in compressed vector form of 1604B. UE_B The precoding matrix indicates that appropriate signaling can be sent to network node 170. At network node 170, mapping information (e.g., Mapping_B) is used to transform or map the coefficient vector from compressed vector form 1604B to extended vector form 1602, which in turn is used by the trained decoder model DEC. NW 1802 Decoding to recover the precoded vector v' B .

[0280] The proposed method for CSI compression, utilizing sparse representations and representation sets (also known as data bases, though not intended to be limited to the strict mathematical meaning of "base"), provides a solid foundation for achieving interoperability between models developed by different vendors. This is achieved by mapping PMIs (e.g., coefficient vectors in compressed vector form) from encoders of different UE vendors back to the same extended vector form (because from s... UE_A Mapping from s - A and from s UE_B The mapping to s-B is recorded in detail. This extended vector form uses a standardized format to encode key information about the pre-encoder, establishing a universal language that transcends specific model architectures or training methods.

[0281] This approach enables UE vendors to leverage the advantages of localized data. By focusing on a specific subset of the data base relevant to their target deployment areas, these UE vendors can develop smaller, more efficient encoder models. This makes PMI mapping simpler, reduces overhead, and improves communication efficiency.

[0282] Furthermore, the proposed framework imposes no restrictions on the specific models or training methods used by either UE or NW vendors. This provides room for flexibility and innovation, enabling vendors to explore and implement AI / ML technologies best suited to their needs and expertise. This is because the S domain is standardized.

[0283] To facilitate seamless integration and operation, a certain level of over-the-air signaling is introduced. These signaling mechanisms enable the UE and the network to exchange information, such as data base, sparse representation, and model updates.

[0284] Based on the previous discussion, the following are the protocol and signaling mechanism designs for the physical layer and MAC layer, designed to facilitate the integration and operation of AI / ML-based CSI compression with a central database and sparse representation:

[0285] System information and UE capability signaling may include:

[0286] • Data base indication: The network broadcasts information about available data bases (e.g., Ψ) via system information messages. NW This information may include:

[0287] o Data base ID: A unique identifier for the data base.

[0288] o Basis vector description: parameters such as dimension, quantization method, and possible subsets representing basis vectors.

[0289] • UE Capability Reporting: The UE reports the AI / ML functions and capabilities it supports, including:

[0290] o Supported Data Base ID: Indicates the data base that the UE can process.

[0291] o Model capability: Report the complexity and performance characteristics of the supported encoder models.

[0292] o Local data availability: (optionally) Indicates the availability of locally acquired data for model fine-tuning.

[0293] Data collection and registration may include:

[0294] • Triggering data collection: The network sends a dedicated RRC message (e.g., AI_DataCollectionRequest) to the UE to specify CSI-RS resources, data collection duration, and reporting mode.

[0295] • UE Data Acquisition: The UE performs channel estimation on the specified CSI-RS resources and extracts the precoding matrix (V). Then, it uses the data basis to compute the sparse representation (S) and reports S and auxiliary information to the network.

[0296] Model updates and synchronization can include the following. In this context, "model" refers to the collection of representations maintained in a central database, not the encoder / decoder model:

[0297] • Model update indication (also known as representation set update indication): The network notifies the UE of a new model update via an RRC message (e.g., AI_ModelUpdate), which includes the model ID (also known as representation set ID) and the transmission method.

[0298] • Model parameter transmission (also known as representation set parameter transmission): Depending on the selected method, the network sends model parameters (also known as representation set parameters) to the UE via RRC signaling or user plane data channel to ensure reliable transmission using segmentation and error correction mechanisms.

[0299] • Synchronization: Optionally, the network can provide synchronization signals (e.g., via RRC or MAC control elements) to ensure that the UE and gNB use the same data base and model version (also known as representation set version).

[0300] Physical layer enhancements may include:

[0301] • CSI-RS Configuration: The network can dynamically configure CSI-RS resources based on feedback from the UE and gNB, thereby ensuring sufficient measurements for accurate channel estimation and data acquisition. For example, this feedback mechanism can include information such as channel conditions, interference levels, and signal quality metrics reported by the UE. By dynamically adjusting the CSI-RS configuration, the network can optimize resource allocation and improve overall system performance.

[0302] • Uplink Channel Selection: The UE and gNB can dynamically select a suitable uplink channel for data reporting based on payload size and channel conditions. Feedback from the UE regarding current channel quality, signal-to-noise ratio (SNR), and other relevant metrics can be used to make informed decisions about the most suitable channel for data transmission. This dynamic selection process helps ensure efficient data transmission with minimal errors, thus adapting in real time to constantly changing network conditions.

[0303] • Hybrid Automatic Repeat Request (HARQ): Implements a HARQ mechanism to ensure reliable transmission of CSI data and model updates. HARQ combines error detection, correction, and retransmission protocols to enhance data reliability. In this scenario, feedback from the receiving end (e.g., UE or gNB) can be used to indicate whether packets have been correctly received or whether retransmission is necessary. This feedback loop supports continuous monitoring and correction of transmission errors, maintaining the integrity and accuracy of CSI data and model updates. The feedback mechanism can include a continuous loop where the UE periodically reports updated CSI to the gNB, including parameters such as link quality, interference, and signal strength. The gNB analyzes this feedback to dynamically adjust the CSI-RS configuration and optimize resource allocation. This helps ensure the network adapts to real-time conditions and maintains optimal performance. Uplink channel selection can benefit from a real-time feedback system where the UE sends periodic updates on channel conditions. This data allows the gNB to determine the optimal uplink channel for data reporting in a timely manner. This process can include selecting a channel that can efficiently handle the current data payload to reduce latency and improve data throughput. HARQ mechanisms typically rely heavily on feedback for error correction. When a data packet is received, the receiver checks for errors. If the packet is error-free, it sends an acknowledgment (ACK); if an error is detected, it sends a negative acknowledgment (NACK). The sender uses this feedback to continue sending the next data packet or retransmit the erroneous packet. This process helps ensure high data reliability and efficient utilization of network resources. By utilizing UE feedback, the network can continuously optimize CSI-RS configuration, uplink channel selection, and the HARQ process. This proactive approach helps mitigate the impact of adverse conditions such as interference and poor signal quality, which can contribute to improved overall network efficiency and robustness.

[0304] MAC layer enhancements may include:

[0305] • Resource allocation: The MAC layer scheduler should prioritize data acquisition and model update transmission while ensuring fairness and efficiency for other services.

[0306] • Scheduling algorithm: An adaptive scheduling algorithm can be used to dynamically allocate resources based on real-time channel conditions and service requirements, thereby ensuring the timeliness and efficiency of data exchange.

[0307] • Data aggregation: gNB can aggregate data reports from multiple UEs to reduce overhead and improve efficiency, and then send these data reports to a central database.

[0308] Security enhancements may include:

[0309] • Secure boot and secure storage: Implement secure boot mechanisms and secure storage schemes on UE and gNB to protect the integrity of AI / ML models and prevent unauthorized modifications.

[0310] • Authentication and Authorization: Utilizes robust authentication and authorization protocols to ensure that only trusted entities can access and provide data to the central database and receive model updates.

[0311] • Secure communication channels: All communications between the UE, gNB, and central database employ encryption and integrity protection mechanisms.

[0312] This protocol and signaling design provides a comprehensive framework for integrating AI / ML-based CSI compression with a central database and sparse representation into existing NR frameworks. This framework ensures interoperability, efficiency, and security, while providing vendors with the flexibility to innovate and optimize their implementations.

[0313] Now for reference Figure 20 .

[0314] Figure 20 An exemplary apparatus is shown, which can be an implementation of a data center 2000 (also referred to as a database) as previously described. The data center 2000 can be a wireless communication system (e.g., Figure 1 The communication system 100) includes a central data center 2000. Data center 2000 can be, in addition to... Figure 1 The network entity of the wireless communication system of the devices shown (e.g., other than ED 110, T-TRP 170, or NT-TRP 172); for example, data center 2000 may be a server or computing system at core network 130 and may be able to communicate with any of ED 110, T-TRP 170, and / or NT-TRP 172. Data center 2000 may include [devices / systems]. Figure 4 The module shown is similar to other modules, but for simplicity, Figure 20Not all units or modules of the data center 2000 are shown in the diagram. The data center 2000 may include a receiving unit or module 2002 (also referred to as receiving module 2002 for simplicity), a transmitting unit or module 2004 (also referred to as transmitting module 2004 for simplicity), and a processing unit or module 2006 (also referred to as processing module 2006 for simplicity), the functions of which have been previously combined. Figure 4 The discussion took place. Data center 2000 may also include memory units for storing training dataset 2008 and the same or additional memory units for storing representation set 2010.

[0315] The receiving module 2002, transmitting module 2004, and / or processing module 2006 can be implemented as software, hardware, or a combination of software and hardware. For example, the receiving module 2002, transmitting module 2004, and / or processing module 2006 can be circuits such as integrated circuits. Examples of integrated circuits include programmable FPGAs, GPUs, or ASICs. For example, the receiving module 2002, transmitting module 2004, and / or processing module 2006 can be implemented at least partially as logic, such as logic functions performed by circuits, logic functions performed by a portion of an integrated circuit, or logic functions performed by software instructions executed by a processor. The training dataset 2008 and the representation set 2010 can be stored in one or more physical memory units of the data center 2000, such as RAM, ROM, hard disk, optical disk, and any other suitable one or more volatile and / or non-volatile storage and retrieval devices.

[0316] Data center 2000 can receive communications via receiving module 2002, such as precoding vectors from ED 110 (which is a UE) and requests for training data sampled from training dataset 2008. Data center 2000 can send communications via sending module 2004, such as update information about representation set 2010 and requested data samples. Processing module 2006 can perform operations to implement various functions of data center 2000 disclosed herein, such as adding data to training dataset 2008 and updating representation set 2010.

[0317] Figure 21 This is a flowchart of an exemplary method 2100 for maintaining a training dataset and a set of representations. Method 2100 can be used in the data center of a wireless communication system (e.g., Figure 20 Method 2100 can be implemented at a data center (2000) or by any suitable network entity. Method 2100 can be executed by a functional unit, such as a processor or any other suitable unit. For simplicity, method 2100 will be described in the operational context of the data center, but this is not intended to be limiting.

[0318] At the start of method 2100, the UE may broadcast or report the availability of its collected data. For simplicity, the following description refers to a single UE, but it should be understood that this is not intended to be limiting.

[0319] Optionally, in operation 2102, the data center's memory 2020 maintains a defined representation set (e.g., Figure 20 A set of one or more representative vectors (representation set 2010). A dataset (e.g., Figure 20 The training dataset (2008) can also be maintained in the data center's storage (2020).

[0320] In operation 2104, data is received, including a precoded vector associated with the UE. In some examples, the data acquisition performed by the UE may be initiated by a data center or other network entity. For example, a dedicated RRC message may be sent to the UE to trigger data acquisition, and data may be received from the UE in response to a control message. The control message may indicate relevant parameters, such as indications of the CSR-RS resources the UE should use for channel estimation, indications of the duration or number of time instances for which the UE should acquire data, indications of how the UE should report the data (e.g., periodically, upon completion, based on specific events such as handover, changes in channel conditions, etc.), and / or any indications of compression the UE should use to compress the acquired data, etc.

[0321] For example, data from the UE can be received via an existing uplink channel (e.g., PUCCH or PUSCH) or via a dedicated control channel. The data received from the UE may also include ancillary information related to the data acquisition performed by the UE to generate associated precoding vectors. For example, the ancillary information may include information about the frequency subbands used in the data acquisition, information about the resource blocks used in the data acquisition, information about the antenna configuration used in the data acquisition, and / or information about the geographical location during the data acquisition process, etc. The data received from the UE may also include indicators indicating the quality or reliability of the acquired data.

[0322] In operation 2106, the precoded vector is converted into a coefficient vector. The coefficient vector represents a linear combination of one or more selected representative vectors from a representation set defined by the one or more representative vectors. Therefore, the coefficient vector can be a sparse representation of the precoded vector based on the representation set. The coefficient vector can be used to train a neural network to perform a precoded vector encoding task or a precoded vector decoding task, as described above.

[0323] Converting a precoded vector to a coefficient vector may include checking whether the precoded vector can be sparsely represented using the current set of representations. This may include computing candidate vectors to represent the precoded vector as a linear combination of one or more candidate representative vectors from the current set of representations. The candidate vectors are then compared to a sparsity threshold. For example, the magnitude of the candidate vectors (e.g., ,in, The candidate vector is compared with a predefined threshold. If the magnitude is less than or equal to the threshold, the sparsity threshold may be met; if the magnitude is greater than the threshold, the candidate vector may not meet the sparsity threshold. If the candidate vector meets the sparsity threshold, it means that the precoded vector can be sparsely represented as a coefficient vector using the current set of representations. Therefore, a candidate vector that meets the sparsity threshold is a coefficient vector representing the precoded vector.

[0324] If a candidate vector fails to meet the sparsity threshold, it means that the precoded vector cannot be sparsely represented as a coefficient vector using the current representation set. The representation set can be updated by including the precoded vector as a new representative vector. Then, after updating the representation set, the precoded vector can be converted into a coefficient vector. For example, if the precoded vector is added as a new k+1 representative vector to the updated representation set, the precoded vector can be represented as a sparse coefficient vector, where the only non-zero element is the (k+1)th element.

[0325] It should be noted that after updating the representation set (e.g., by adding precoded vectors as new representative vectors or through other operations disclosed herein), the data center may send a control message (e.g., an RRC message) to notify the UE and other network entities of the update. The model update message may include a unique identifier for the updated representation set, as well as information about the representative vectors, such as dimensions, quantization methods, and other parameters.

[0326] In operation 2108, the data center may store the coefficient vector (generated in operation 2106) in a dataset (e.g., training dataset 2008). The coefficient vector may be included in the dataset, which can be used to train a neural network to perform a pre-encoded vector encoding task or a pre-encoded vector decoding task. For example, the coefficient vector may be used to perform the above combination... Figure 17 or Figure 18 This discussion concerns a portion of the training data. If the data received from the UE in operation 2104 includes auxiliary data, this auxiliary data can be stored in association with the corresponding coefficient vector in the dataset. This auxiliary data can be used to sample data from the dataset for spatial features of interest. For example, the coefficient vector can be stored along with auxiliary data about a specific geographic location associated with the data acquisition. The coefficient vector can then be included in the sampled data for that specific geographic location of interest.

[0327] In some examples, a data center can generate one or more synthetic coefficient vectors based on one or more coefficient vectors derived from real-world precoded vectors. For instance, real-world coefficient vectors can be clustered, and synthetic coefficient vectors can be generated based on these clusters. Synthetic coefficient vectors generated in this way can help augment the real-world data in the dataset, and can be used to train neural networks just like real-world coefficient vectors. When sampled data is provided to NW or UE vendors for training neural networks, the data can include both synthetic and real-world coefficient vectors.

[0328] In operation 2110, the data center can provide coefficient vectors sampled from the dataset for training a neural network to perform a pre-encoding vector encoding task (e.g., Figure 17 (as shown in the training) or pre-encoded vector decoding task (e.g., Figure 18 (As shown in the training example). As previously described, the data center can sample coefficient vectors from the entire dataset or from a selected subset of the dataset. For example, if an NW vendor requests data for training a decoder model designed for network entities serving a large geographic area (e.g., T-TRP 170), the coefficient vectors provided by the data center for training the decoder model can be sampled from the entire dataset. In another example, if a UE vendor requests data for training an encoder model designed for a UE in a specific spatial region of interest, the coefficient vectors provided by the data center for training the encoder model can be sampled from a selected subset of the dataset containing coefficient vectors associated with the spatial region of interest (e.g., based on auxiliary information associated with each coefficient vector). The subset of data may correspond to a representation subset that includes only representative vectors representing the subset of data, omitting at least one representative vector not used to represent the subset of data.

[0329] In some examples, the data center may send coefficient vectors in a compact vector form (e.g., sampled from a subset of the dataset), which is smaller than the original extended vector form used to store the coefficient vectors in the dataset. For example, if the coefficient vectors are sampled from a subset of the dataset corresponding to a subset of representations that omits a particular representative vector, the compact vector form can be obtained by omitting the elements of the coefficient vectors that correspond to the omitted representative vectors (since the elements corresponding to the omitted representative vectors should all be zero).

[0330] Mapping information used for mapping (or transforming) from compact vector form to extended vector form can be stored. This mapping information is also used for mapping (or transforming) between subsets and sets of representations. For example, the mapping information can be a transformation matrix. Different UE vendors can use different subsets of data to train their encoder models, resulting in different mapping information. As mentioned above, network entities can also use this mapping information to map compact vector form to extended vector form before decoding precoded vectors. For example, the mapping information can be sent to network entities (e.g., in control messages).

[0331] The data center can perform operations to ensure that the representation set is up-to-date and contains information for representing the dynamic, real-time characteristics of the wireless channel. For example, the data center can update the representation set by including unique precoding vectors that cannot be sparsely represented using the current representation set as new representative vectors in the updated representation set. In another example, the data center can track the usage and / or duration of existence of each representative vector in the representation set. Each representative vector can be associated with a timestamp (or other time indicator) indicating when the representative vector was added to the representation set or when it was last used to represent a precoding vector. Based on this tracked usage and / or duration of existence, outdated representative vectors (e.g., those whose timestamps are earlier than a preset time threshold, such as 1 day) can be removed from the representation set.

[0332] After updating the representation set, the data center can send control messages to notify the UE and other network entities of the updated representation set. For example, the control message may include the updated identifier of the updated representation set and / or information about the representative vectors, such as the number of representative vectors, the dimensions of the representative vectors, the indices of added or deleted representative vectors, etc.

[0333] Figure 22 This is a flowchart of an exemplary method 2200 for training a neural network to perform a pre-coding vector encoding task. Method 2200 can be implemented at a network entity (e.g., a UE vendor) or other suitable network entity where the trained neural network can be deployed at the UE. Method 2200 can be executed by a functional unit, such as a processor or any other suitable unit.

[0334] In operation 2202, the UE vendor obtains a selected subset of one or more representative vectors selected from a representation set (e.g., a set defined by all representative vectors maintained at a data center). The selected representation subset is smaller than the representation set and may include only one or more representative vectors associated with a feature of interest (e.g., a spatial feature of interest, such as a specific geographic region). The representation subset may contain representative vectors that, together, are sufficient to sparsely represent the precoded vectors associated with the feature of interest (e.g., capable of sparsely representing precoded vectors generated by UEs within a specific geographic region).

[0335] For example, a UE vendor can send a request to a data center for one or more representative vectors associated with one or more features of interest. The data center can select one or more representative vectors associated with one or more features of interest from a set of representations (e.g., based on auxiliary data stored in association with each representative vector) and send the one or more selected representative vectors to the UE vendor in a control message or using other signaling. A subset of representations defined by the one or more selected representative vectors can be assigned a unique identifier.

[0336] In operation 2204, the UE vendor obtains coefficient vectors. Each coefficient vector represents a linear combination of one or more selected representative vectors from a subset of representations (obtained in operation 2202).

[0337] For example, a UE vendor can send a request to a data center for coefficient vectors associated with one or more features of interest. The data center can sample coefficient vectors from a subset of data sets associated with the one or more features of interest (e.g., based on auxiliary data stored in association with each coefficient vector) and send the sampled coefficient vectors to the UE vendor in a control message or using other signaling. Each sampled coefficient vector can be converted into a corresponding precoded vector using only the representative vectors contained in the representation subset. The sampled coefficient vectors can be sent to the UE vendor in a compact vector form corresponding to the representation subset (and smaller than the extended vector form corresponding to the representation set), as described above.

[0338] In operation 2206, the UE vendor uses the acquired coefficient vectors to train a neural network to perform a precoding vector encoding task. To train the neural network, the coefficient vectors are transformed into corresponding precoding vectors using a subset of representations (e.g., by multiplying the coefficient vectors by the subset of representations to obtain the corresponding precoding vectors). The precoding vectors and their corresponding coefficient vectors are then used as data pairs to train the neural network. For example, the precoding vectors are fed as input to the neural network, and the output generated by the neural network is obtained. The neural network can then be trained using appropriate machine learning techniques (e.g., for minimizing MMSE) to minimize the error between the output generated by the neural network and the coefficient vector corresponding to the precoding vector. In some examples, transfer learning methods can be used to train the neural network.

[0339] Transfer learning involves taking an existing pre-trained model that has been trained on different but related tasks and further training that pre-trained model using data specific to the target task. For example, a pre-trained model that can be fine-tuned using the acquired coefficient vectors to perform a pre-coding vector encoding task could be a general encoder / decoder neural network that has been trained on large datasets for similar purposes, such as image recognition, natural language processing, or even an initial version of pre-coding vector encoding.

[0340] Further training using the coefficient vectors allows the pre-trained model to adapt to the specific nuances of the pre-encoding vector encoding task. This approach leverages the knowledge the model has already acquired during its pre-training phase, which helps to accelerate the training process and / or improve model performance. In contrast to de novo training, where the neural network starts with randomly initialized parameter values ​​(or weights) and gradually learns the task, transfer learning starts with a pre-trained model and fine-tunes it using data specific to the desired task (in this case, the pre-encoding vector encoding task).

[0341] The difference between further training using transfer learning and the MMSE method is that it refines an already functional model rather than building it from scratch. While the MMSE method focuses on minimizing the mean squared error between the output and the target coefficient vector, transfer learning involves using the parameters of a pre-trained model as a starting point and tuning those parameters in a new, more specific task context to minimize the MMSE. Because the model is initially trained on a broader dataset, this approach can achieve faster convergence and potentially better performance.

[0342] In Operation 2208, for example, a trained neural network can be deployed to a UE served by a UE vendor. The UE vendor can deploy the trained neural network to the UE using appropriate signaling. For example, control messages (e.g., dedicated RRC messages) can contain information about the trained neural network, such as an identifier of the encoder model implemented by the trained neural network, an indication of the encoder model size (e.g., the number of model parameters), and / or the method of deploying the trained neural network to the UE (e.g., via control messages, via data channels, etc.). When the neural network is relatively small (e.g., with few parameters), the parameters of the trained neural network can be deployed to the UE via control messages (e.g., using RRC signaling); when the neural network is large (e.g., with many parameters), deployment via data channels (e.g., user plane data channels) may be more suitable. In some examples, the parameters of the trained neural network can be sent in segments, with appropriate error detection and / or correction mechanisms to help ensure a more reliable deployment. Furthermore, synchronization signals (e.g., from a data center, from the UE vendor, or from another network entity) can be used to ensure that the encoder model deployed at the UE is consistent and up-to-date with the decoder model deployed at the network entity. This synchronization enables efficient communication and data processing, and helps ensure seamless integration and functionality within wireless communication systems.

[0343] In operation 2210, the UE vendor may update the representation subset. For example, the UE vendor updates the representation subset to include new representative vectors (e.g., by obtaining new representative vectors corresponding to new features of interest from a data center, such as new spatial features of the UE served by the UE vendor) and / or exclude outdated representative vectors (e.g., representative vectors that have not been used for at least a predetermined period of time). Then, method 2200 may return to operation 2204. The UE vendor may obtain new data corresponding to the updated representation subset (e.g., a new set of coefficient vectors representing linear combinations of the representative vectors in the updated representation subset), and may again perform operation 2206 to retrain the neural network using the new set of coefficient vectors to perform the pre-encoding vector encoding task. The UE vendor may then deploy the retrained neural network to the UE as described above.

[0344] Figure 23 This is a flowchart of an exemplary method 2300 for training a neural network to perform a precoded vector decoding task. Method 2300 can be implemented at a network entity (e.g., an NW vendor) or at other suitable network entities where the trained neural network can be deployed at network nodes (e.g., at a TRP). Method 2300 can be executed by a functional unit, such as a processor or any other suitable unit.

[0345] In operation 2302, the NW vendor obtains a set of one or more representative vectors that define a representation set. For example, the NW vendor may obtain a complete representation set maintained at a data center. The NW vendor may send a request for the representation set to the data center, and the data center may, for example, send the representation set to the NW vendor using control messages.

[0346] In Operation 2304, the NW vendor acquires coefficient vectors. Each coefficient vector represents a linear combination of one or more selected representative vectors from a representation set, and each coefficient vector corresponds to a precoded vector. For example, the NW vendor may send a request to the data center for coefficient vectors sampled from the entire dataset maintained by the data center. The data center may sample coefficient vectors from the dataset and send the sampled coefficient vectors to the NW vendor in a control message or using other signaling. Each sampled coefficient vector can be converted into a corresponding precoded vector using representative vectors contained in the representation set. The sampled coefficient vectors can be acquired by the NW vendor in extended vector form.

[0347] In Operation 2306, the NW vendor uses the acquired coefficient vectors to train a neural network to perform a precoded vector decoding task. To train the neural network, the coefficient vectors are transformed into corresponding precoded vectors using a set of representations (e.g., by multiplying the coefficient vectors by a subset of representations to obtain the corresponding precoded vectors). The precoded vectors and their corresponding coefficient vectors are then data pairs used to train the neural network. For example, the coefficient vectors are fed as input to the neural network, and the output generated by the neural network is obtained. The neural network can then be trained using appropriate machine learning techniques (e.g., for minimizing MMSE) to minimize the error between the output generated by the neural network and the precoded vectors corresponding to the coefficient vectors. In some examples, transfer learning methods can be used to train the neural network, where a pre-trained model is further trained using the coefficient vectors.

[0348] This transfer learning can be similar to the process described in previous combination operation 2206. For example, the acquired coefficient vector can be used to fine-tune a pre-trained model that has been pre-trained on a relevant task (e.g., image recognition, natural language processing, or an initial version of pre-coded vector decoding).

[0349] In Operation 2308, for example, a trained neural network can be deployed to a network entity (e.g., a TRP) served by an NW vendor. The NW vendor can deploy the trained neural network to the network entity using appropriate signaling. For example, control messages (e.g., dedicated RRC messages) can contain information about the trained neural network, such as an identifier of the decoder model implemented by the trained neural network, an indication of the decoder model size (e.g., the number of model parameters), and / or the method of deploying the trained neural network to the network entity (e.g., via control messages, via data channels, etc.). When the neural network is relatively small (e.g., with a small number of parameters), the parameters of the trained neural network can be deployed to the network entity via control messages (e.g., using RRC signaling); when the neural network is large (e.g., with a large number of parameters), deployment via data channels (e.g., user plane data channels) may be more suitable. In some examples, the parameters of the trained neural network can be sent in segments, with appropriate error detection and / or correction mechanisms employed to help ensure a more reliable deployment. Furthermore, synchronization signals (e.g., from a data center, an NW vendor, or another network entity) can be used to ensure that the decoder model deployed at the network entity is consistent and up-to-date with the encoder model deployed at the UE. This consistency enables efficient communication and data processing, helping to ensure seamless integration and functionality within the wireless communication system.

[0350] In some examples, the trained neural network can be trained at the same network entity where training was performed. In such cases, deploying the trained neural network may simply involve storing the trained parameters in memory. After deployment, a coefficient vector can be received (e.g., from the UE), and the trained neural network can be used to decode the coefficient vector to recover the pre-coded vector (e.g., using method 2400, discussed further below).

[0351] In operation 2310, the updated representation set can be received. For example, after updating the representation set at the data center (e.g., by adding new representative vectors or deleting outdated representative vectors), the data center can send an update message to the NW vendor containing information about the updated representation set (e.g., including updated identifiers, identifiers of deleted representative vectors, and / or identifiers of added representative vectors). Then, method 2300 can return to operation 2304, in which a new coefficient vector corresponding to the updated representation set is obtained. Then, operation 2306 can be performed again to retrain the neural network using the new set of coefficient vectors to perform the pre-encoded vector decoding task. The NW vendor can then deploy the retrained neural network as described above.

[0352] Figure 24This is a flowchart of an exemplary method 2400 for decoding a precoded vector using a trained neural network. Method 2400 can be implemented at a network entity (e.g., a TRP). Method 2400 can be executed by a functional unit, such as a processor or any other suitable unit. For example, a trained neural network may have been trained using the method 2300 described above, and that trained neural network may have been deployed to a network entity.

[0353] In operation 2402, a coefficient vector is received (e.g., from the UE). The coefficient vector corresponds to a precoding vector and can be received as a PMI. As described above, the coefficient vector can be a sparse representation of the precoding vector, wherein the coefficient vector represents a linear combination of one or more selected representative vectors from a representation set defined by one or more representative vectors.

[0354] In some examples, the coefficient vector may be received in a compact vector form, which corresponds to a subset of representations smaller than the representation set. To decode the precoded vector from the coefficient vector, it may be necessary to convert the coefficient vector to an extended vector form corresponding to the representation set. In operation 2404, a mapping can be used to convert the coefficient vector from the compact vector form to the extended vector form. For example, the mapping could be a transformation matrix that transforms between the compact and extended vector forms. This mapping can be obtained in advance (e.g., obtained from a data center via control signaling). An identifier associated with the mapping (or with the subset of representations associated with the mapping) can be provided by or associated with the UE to enable the network entity to apply the appropriate mapping to convert the coefficient vector to the extended vector form. Decoding can then be performed on the coefficient vector in extended vector form.

[0355] In operation 2406, a trained neural network is used to decode the precoded vector from the coefficient vector.

[0356] The decoded precoded vector can then be used for various operations. For example, in operation 2408, the decoded precoded vector can be provided to optimize network operations and resource allocation, and so on.

[0357] In some examples, the decoded pre-encoded vector can be stored together with the corresponding original coefficient vector. This can form data pairs that can be used for further training and / or refinement of the neural network. For example, data pairs formed in this way can be analyzed (e.g., using machine learning or statistical techniques) to identify patterns or differences between the decoded and original data, which can provide insights that guide the neural network architecture or parameter tuning. Fine-tuning the neural network using these data pairs allows for further training or refinement, contributing to improved accuracy and robustness. The stored data pairs can enable the neural network to be trained in real-world scenarios, potentially helping to enhance the model's performance over time by updating it with new data. This iterative process can help ensure that the neural network remains effective and reliable, adapting to changing conditions and improving its predictive capabilities.

[0358] In some examples, information about the accuracy of the decoded precoded vectors can be provided as feedback to further refine the training of the neural network. Feedback may include metrics such as error rate, signal-to-noise ratio (SNR), bit error rate (BER), and / or any discrepancy between the expected and actual performance of the decoded precoded vectors. This information can be collected from the network entity and sent back to the data center or the entity responsible for training the neural network (e.g., an NW vendor).

[0359] Feedback can be used to identify specific aspects of a neural network's performance where improvement is possible. For example, if feedback indicates a high error rate for certain types of channels or conditions, the training process can be adjusted to focus more on these scenarios (e.g., a network vendor might request more training data from data centers specific to these scenarios). The neural network can then be retrained using additional data reflecting these conditions or by modifying the training algorithm to better handle the identified problems.

[0360] Furthermore, feedback can be used to perform incremental updates to the neural network. Instead of retraining the entire model from scratch, the neural network can undergo a fine-tuning phase, where only certain layers or parameters are adjusted based on feedback (e.g., by fixing the hidden layers and focusing training only on the final output layer). This process ensures that the model stays in sync with the latest running data and can adapt to new patterns or changes in the network environment.

[0361] By repeatedly incorporating feedback from network entities after deploying the neural network, the trained neural network can be fine-tuned to reflect the latest real-world conditions. This can help the neural network achieve higher accuracy and reliability, ultimately improving the overall efficiency and performance of wireless communication systems.

[0362] Figure 25This is a flowchart of an exemplary method 2500 for encoding pre-encoded vectors using a trained neural network. Method 2500 can be implemented by a processor that executes instructions at an ED (e.g., UE). For example, the trained neural network may have been trained using the method 2200 described above and then deployed to the UE.

[0363] The trained neural network may have been deployed to the UE by the UE through control signals and / or by acquiring the parameters of the trained neural network through data channels, as described above.

[0364] In operation 2502, the UE acquires the precoding vector. The UE can acquire the precoding vector based on one or more parameters indicated by a control message from a network entity. For example, the control message sent to the UE (e.g., an RRC message) can specify CSI-RS resources, data acquisition duration, and reporting mode, etc. The UE can then perform channel estimation on the specified CSI-RS resources and extract the precoding matrix (whose vector is the precoding vector).

[0365] In operation 2504, a trained neural network is used to encode the precoded vector into a coefficient vector. The coefficient vector is a sparse representation of the precoded vector. Each coefficient vector represents a linear combination of one or more selected representative vectors from a subset of representations defined by one or more representative vectors.

[0366] In operation 2506, the coefficient vector is sent. For example, the coefficient vector can be sent as a PMI. In some examples, auxiliary information (e.g., related to data acquisition of the precoded vector, as described above) can be sent along with the coefficient vector.

[0367] The examples disclosed also provide protocol and signaling designs for the physical and MAC layers. For example, at the data center, information about the representation set can be broadcast via system information messages, such as a unique identifier for the representation set and parameters of the representative vectors (e.g., dimensions, quantization method, etc.). The data center can also provide network entities with requested representation sets or subsets of representations (which can be identified using unique identifiers different from the identifier of the representation set), and / or requested data sampled from a dataset of coefficient vectors. Optionally, the data center can provide synchronization signals (e.g., via RRC or MAC control elements) to help ensure that the UE and network entities use the same representation set version and encoder / decoder model version.

[0368] The UE can report the AI / ML capabilities it supports, such as the representative vectors it can process (e.g., representative vectors that form a subset of the representations used to train the encoder model deployed at the UE). The UE can implement multiple trained encoder models. The UE can report the complexity and performance characteristics of the encoder models it supports. Optionally, the UE can indicate the availability of locally acquired data (e.g., locally acquired pre-encoded vectors) for fine-tuning the model.

[0369] Network entities (such as data centers or other schedulers) can trigger data acquisition at the UE via dedicated RRC messages or other control messages. For example, control messages may indicate the resources used to perform channel estimation (e.g., CSI-RS resources), the data acquisition duration, and the reporting mode.

[0370] Network entities (such as UE vendors) can transmit parameters of the trained neural network implementing the encoder model to the UE. These parameters can be transmitted via control signaling (e.g., RRC signaling) or via a data channel, depending on the number of parameters to be transmitted. Various error detection and correction mechanisms can be used to ensure reliable transmission of model parameters. Optionally, synchronization signals can be provided to ensure the UE uses the latest model.

[0371] Various feedback mechanisms can be implemented to help ensure satisfactory performance of the trained model. For example, the UE and TRP deploying the trained encoder / decoder model can provide pre-coded vector and coefficient vector data pairs that can be used for further model training. The performance of the trained model can also be reported. Each UE and TRP can dynamically select an appropriate channel for data reporting, for example, based on payload size and / or channel conditions. A HARQ mechanism can also be implemented to improve reliability.

[0372] Data reported from multiple UEs can be grouped together at the TRP to transmit data to the data center more efficiently and / or reduce overhead.

[0373] Based on feedback from the UE and TRP, resources used for channel estimation and data acquisition (e.g., CSI-RS resources) can be dynamically configured to help ensure sufficient up-to-date measurements.

[0374] The MAC layer scheduler can prioritize data acquisition and model update transmissions while balancing this communication with other services. For example, various adaptive scheduling algorithms can be used for dynamic resource allocation based on real-time channel conditions and / or service requirements.

[0375] Various security mechanisms can be implemented. For example, secure boot mechanisms and secure storage schemes can be implemented on the UE and TRP (and other network entities) to help prevent unauthorized modifications to the trained encoder and decoder models. Authentication and / or authorization protocols can be used to ensure that only trusted entities can access and provide data to the data center. Authentication and / or authorization protocols can also help ensure that models and model updates are only communicated to trusted entities. All communication between the UE, TRP, network entities, and data center can be protected with appropriate encryption and integrity mechanisms.

[0376] This disclosure covers a variety of embodiments, including not only method embodiments but also other embodiments, such as apparatus embodiments and embodiments related to non-transitory computer-readable storage media. Embodiments may be incorporated individually or in combination with the features disclosed herein.

[0377] While this disclosure references illustrative embodiments, it is not intended to be construed in a limiting sense. Those skilled in the art will recognize, upon referring to this description, various modifications and combinations of the illustrative embodiments, as well as other embodiments of this disclosure.

[0378] Additionally or alternatively, features disclosed herein in the context of any particular embodiment may be implemented in other embodiments. For example, method embodiments may be implemented in apparatus, system, and / or computer program product embodiments. Furthermore, although embodiments are described primarily in the context of methods and apparatus, other implementations are contemplated as instructions stored in one or more non-transitory computer-readable media. Such media may store programs or instructions to perform any of the various methods consistent with this disclosure.

[0379] The terms "system" and "network" are used interchangeably in the embodiments of this application. "At least one" means one or more, and "multiple" means two or more. The term "and / or" describes the association relationship between related objects, indicating that three relationships may exist. For example, A and / or B can indicate the following three cases: only A exists, both A and B exist, and only B exists, where A and B can be singular or plural. The character " / " generally indicates an "OR" relationship between related objects. "At least one of the following items" or similar expressions indicate any combination of these items, including a single item or any combination of multiple items. For example, "at least one of A, B, or C" includes A, B, C, A and B, A and C, B and C, or A, B, and C. "At least one of A, B, and C" can also be understood to include A, B, C, A and B, A and C, B and C, or A, B, and C. In addition, unless otherwise stated, the ordinal numbers such as "first" and "second" in the embodiments of this application are used to distinguish multiple objects and are not used to limit the order, time sequence, priority or importance of multiple objects.

[0380] It should be understood that examples of this disclosure may be embodied in methods, apparatus, non-transitory computer-readable media, processing modules, chipsets, system-on-a-chip, or computer programs, etc. Apparatus may include a transmitting module for performing the transmitting steps described above and a receiving module for performing the receiving steps described above. Apparatus may include processing modules, processors, or processing units for controlling or causing the apparatus to perform the examples disclosed herein.

[0381] Although this disclosure describes methods and processes by way of steps performed in a certain order, one or more steps in the methods and processes may be omitted or modified as appropriate. Where appropriate, one or more steps may be performed in an order other than that described.

[0382] Although this disclosure describes at least part of the methodological aspects, those skilled in the art will understand that this disclosure also relates to various components, whether hardware components, software, or any combination thereof, for performing at least some aspects and features of the described methods. Therefore, the technical solutions of this disclosure can be embodied in the form of a software product. Suitable software products may be stored in pre-recorded storage devices or other similar non-volatile or non-transitory computer-readable media, including DVDs, CD-ROMs, USB flash drives, removable hard drives, or other storage media. The software product includes instructions tangibly stored thereon that enable a processing device (e.g., a personal computer, server, or network device) to perform the method examples disclosed herein. Machine-executable instructions may be in the form of code sequences, configuration information, or other data that, when executed, cause a machine (e.g., a processor or other processing device) to perform the steps in the methods provided in the examples of this disclosure.

[0383] This disclosure may be embodied in other specific forms without departing from the subject matter of the claims. The exemplary embodiments described are illustrative in all respects and not restrictive. Features selected from one or more of the foregoing embodiments may be combined to create alternative embodiments not explicitly described, and features suitable for such combinations will be understood within the scope of this disclosure.

[0384] Furthermore, all values ​​and sub-ranges within the scope of the disclosure are also disclosed. Additionally, although the systems, devices, and processes disclosed and illustrated herein may include a specific number of elements / components, these systems, devices, and components may be modified to include additional or fewer such elements / components. For example, while any element / component disclosed may be referenced as a single quantity, embodiments disclosed herein may be modified to include multiple such elements / components. The subject matter described herein is intended to cover and include all suitable technical modifications.

Claims

1. A method, characterized in that, include: Receive data, the data including a precoded vector associated with a user equipment (UE); The precoded vector is converted into a coefficient vector, which represents a linear combination of one or more selected representative vectors from a representation set defined by one or more representative vectors. The coefficient vector can be used to train a neural network to perform a precoded vector encoding task or a precoded vector decoding task.

2. The method according to claim 1, characterized in that, Converting the precoded vector into the coefficient vector includes: Compute candidate vectors to represent the precoded vector as a linear combination of one or more candidate representative vectors from the representation set; and The candidate vectors are determined to satisfy the sparsity threshold. Wherein, after determining that the candidate vector satisfies the sparsity threshold, the candidate vector is the coefficient vector, and the one or more candidate representative vectors are the one or more selected representative vectors.

3. The method according to claim 1, characterized in that, Converting the precoded vector into the coefficient vector includes: It is determined that the set of representations cannot represent the precoded vector; The representation set is updated by including the precoded vector as a new representative vector in the representation set; and After the update, the precoded vector is converted into the coefficient vector.

4. The method according to claim 3, characterized in that, Determining that the representation set cannot represent the precoding vector includes: Prior to the update, candidate vectors are computed to represent the precoded vector as a linear combination of one or more candidate representative vectors from the representation set; and The candidate vector is determined to not satisfy the sparsity threshold.

5. The method according to any one of claims 1 to 4, characterized in that, Also includes: Track the usage and / or duration of existence of each representative vector in the representation set; as well as The representation set is updated by removing outdated representative vectors based on the usage of the tracking and / or the duration of existence.

6. The method according to any one of claims 3 to 5, characterized in that, Also includes: Send a control message that indicates an update to the representation set.

7. The method according to any one of claims 1 to 6, characterized in that, Also includes: Send a control message indicating one or more parameters to be used by the UE to generate the precoding vector; The precoding vector is received in response to the control message.

8. The method according to any one of claims 1 to 7, characterized in that, The received data also includes auxiliary information, which is related to the data acquisition performed by the UE to generate the precoding vector, and the coefficient vector corresponding to the precoding vector is associated with the auxiliary information.

9. The method according to claim 8, characterized in that, The auxiliary information includes one or more of the following: Information regarding the frequency sub-bands used in the data acquisition; Information regarding the resource blocks used in the data acquisition; Information regarding the antenna configuration used in the data acquisition; as well as Information regarding the geographical location during the data collection period.

10. The method according to any one of claims 1 to 9, characterized in that, Also includes: The coefficient vector is provided to the neural network for training to perform the pre-encoded vector encoding task or the pre-encoded vector decoding task.

11. The method according to any one of claims 1 to 10, characterized in that, Also includes: A synthetic coefficient vector is generated based on one or more coefficient vectors obtained from the precoded vector; wherein the synthetic coefficient vector can also be used to train the neural network to perform the precoded vector encoding task or the precoded vector decoding task.

12. The method according to claim 11, characterized in that, Also includes: The synthesized coefficient vector is provided to the neural network for training to perform the pre-encoded vector encoding task or the pre-encoded vector decoding task.

13. The method according to any one of claims 1 to 12, characterized in that: The method is performed at a data center of a wireless communication system including the UE; The set of one or more representative vectors is maintained in the memory of the data center; as well as The coefficient vectors are stored in a dataset, which is maintained in the memory of the data center and can be used to train the neural network to perform the pre-encoded vector encoding task or the pre-encoded vector decoding task.

14. The method according to claim 13, characterized in that, Also includes: The coefficient vectors from the dataset are sent to train the neural network to perform the precoded vector decoding task, wherein the sent coefficient vectors are sampled from the entire dataset.

15. The method according to claim 13, characterized in that, Also includes: The coefficient vectors in the dataset are sent to train the neural network to perform the pre-encoding vector encoding task, wherein the sent coefficient vectors are sampled from a selected subset of the dataset, the selected subset containing coefficient vectors associated with the spatial features of interest, and the sent coefficient vectors correspond to a selected subset of representations, which omits at least one representative vector from the set of representations and is associated with the spatial features of interest.

16. The method according to claim 15, characterized in that, Also includes: Send mapping information for mapping between the subset of representations and the set of representations.

17. The method according to claim 15 or 16, characterized in that, The transmitted coefficient vector is transmitted in a compact vector form, which is smaller than the extended vector form used to store the coefficient vector in the dataset. The compact vector form omits elements corresponding to at least one representative vector omitted from the representation subset.

18. The method according to claim 17, which is dependent on claim 16, characterized in that, The mapping information also maps between the compact vector form and the extended vector form.

19. A method, characterized in that, include: Obtain a selected subset of one or more representative vectors from a set of one or more representative vectors that define a representation set, wherein the selected subset of one or more representative vectors is defined to be smaller than the representation subset of the representation set; Obtain coefficient vectors, wherein each coefficient vector represents a linear combination of one or more selected representative vectors from the subset of representations, the linear combination corresponding to a precoded vector; and A neural network is trained to perform a pre-encoding vector encoding task, wherein during training, each coefficient vector is converted into a corresponding pre-encoding vector, and the corresponding pre-encoding vector is used to train the neural network.

20. The method according to claim 19, characterized in that, Training the neural network includes: The corresponding precoded vector is provided as input to the neural network and the output is obtained; and The neural network is trained to minimize the error between the output generated by the neural network and the coefficient vector corresponding to the precoded vector.

21. The method according to claim 19 or 20, characterized in that, The coefficient vectors are sampled from a selected subset of data containing coefficient vectors associated with spatial features of interest. Each coefficient vector in the subset of data can be converted into a corresponding precoded vector using the representation subset.

22. The method according to claim 21, characterized in that, Each coefficient vector is obtained in a compact vector form corresponding to the subset of representations, which is smaller than the extended vector form corresponding to the set of representations.

23. The method according to any one of claims 19 to 22, characterized in that, Also includes: The trained neural network is deployed to perform the pre-encoded vector encoding task.

24. The method according to claim 23, characterized in that, The trained neural network is deployed at the user equipment (UE).

25. The method according to any one of claims 19 to 24, characterized in that, Also includes: Update the subset of representations to include new representative vectors from the set of representations or exclude outdated representative vectors from the subset of representations; Obtain another set of coefficient vectors corresponding to the updated subset of representations; and The neural network is retrained using the other set of coefficient vectors to perform the pre-encoding vector encoding task.

26. A method, characterized in that, include: Get a set of one or more representative vectors that represent a set; Obtain coefficient vectors, wherein each coefficient vector represents a linear combination of one or more selected representative vectors from the representation set, the linear combination corresponding to a pre-encoded vector; and A neural network is trained to perform a precoded vector decoding task, wherein, during training, each coefficient vector is used to train the neural network.

27. The method according to claim 26, characterized in that, Training the neural network includes: The coefficient vector is provided as input to the neural network and the output is obtained; and The neural network is trained to minimize the error between the output generated by the neural network and the precoded vector corresponding to the coefficient vector.

28. The method according to claim 26 or 27, characterized in that, Also includes: The neural network receives an update to the set of representations, wherein the neural network is retrained in response to the update.

29. The method according to any one of claims 26 to 28, characterized in that, Also includes: Receive feedback related to the performance of the trained neural network; as well as The feedback is used to further refine the training of the neural network.

30. The method according to any one of claims 26 to 29, characterized in that: The neural network is trained using transfer learning, wherein the pre-trained neural network is further trained using the coefficient vector.

31. The method according to any one of claims 26 to 30, characterized in that, Also includes: Deploy the trained neural network to perform the precoded vector decoding task.

32. The method according to claim 31, characterized in that, Also includes: Receiver coefficient vector; and The trained neural network is used to decode the precoded vector from the coefficient vector.

33. The method according to claim 32, characterized in that, Also includes: Obtain the mapping between a subset of representations and the set of representations, wherein the subset of representations is defined by a selected subset of one or more representative vectors from the set of representations; in: The coefficient vector is received in a compact vector form corresponding to the subset of representations; and The coefficient vector is converted from the compact vector form to an extended vector form corresponding to the representation set using the mapping, and the decoding is performed on the coefficient vector in the extended vector form.

34. A method, characterized in that, include: The received coefficient vector represents a linear combination of one or more selected representative vectors from a representation set defined by one or more representative vectors, and the linear combination corresponds to a precoded vector; as well as The precoded vector is decoded from the coefficient vector using a trained neural network, wherein the trained neural network has been trained to perform the precoded vector decoding task.

35. The method according to claim 34, characterized in that, Also includes: Obtain the mapping between a subset of representations and the set of representations, wherein the subset of representations is defined by a selected subset of one or more representative vectors from the set of representations; in: The coefficient vector is received in a compact vector form corresponding to the subset of representations; and The coefficient vector is converted from the compact vector form to an extended vector form corresponding to the representation set using the mapping, and the decoding is performed on the coefficient vector in the extended vector form.

36. The method according to claim 34 or 35, characterized in that, Also includes: The decoded precoded vector and the corresponding original coefficient vector are stored as data pairs for further analysis / refinement of the trained neural network.

37. The method according to any one of claims 34 to 36, characterized in that, Also includes: The decoded precoded vector is sent to optimize network operations and / or resource allocation.

38. A method, characterized in that, include: Obtain the precoded vector; The precoded vector is encoded into a coefficient vector using a trained neural network. The coefficient vector represents a linear combination of one or more selected representative vectors from a subset of representations, which is defined by one or more representative vectors, wherein the linear combination corresponds to the precoded vector. as well as Send the coefficient vector.

39. The method according to claim 38, characterized in that, The coefficient vector is transmitted as a precoding matrix indication (PMI).

40. The method according to claim 38 or 39, characterized in that, Also includes: The parameters of the trained neural network are obtained from the control signal.

41. The method according to any one of claims 38 to 40, characterized in that, Also includes: The trained neural network is updated based on feedback received from network entities.

42. The method according to any one of claims 38 to 41, characterized in that: The precoding vector is obtained based on one or more parameters indicated by control messages from network entities.

43. The method according to any one of claims 38 to 42, characterized in that, Also includes: Auxiliary information related to the precoded vector is sent along with the coefficient vector.

44. A network entity, characterized in that, include: Memory; A processor for executing instructions stored in the memory to cause the network entity to perform the method according to any one of claims 1 to 37.

45. The network entity according to claim 44, characterized in that, The network entity is a data center.

46. ​​The network entity according to claim 44, characterized in that, The network entity provides services to one or more user equipment (UEs).

47. The network entity according to claim 44, characterized in that, The network entity provides services to one or more base stations (BS).

48. An apparatus, characterized in that, include: Memory; A processor for executing instructions stored in the memory to cause the device to perform the method according to any one of claims 38 to 43.

49. A non-transitory computer-readable medium storing machine-executable instructions, characterized in that, When executed by a network entity, the instructions cause the network entity to perform the method according to any one of claims 1 to 37.

50. The non-transitory computer-readable medium according to claim 49, characterized in that, The network entity is a data center.

51. The non-transitory computer-readable medium according to claim 49, characterized in that, The network entity provides services to one or more user equipment (UEs).

52. The non-transitory computer-readable medium according to claim 49, characterized in that, The network entity provides services to one or more base stations (BS).

53. A non-transitory computer-readable medium storing machine-executable instructions, characterized in that, When executed by the device, the instructions cause the device to perform the method according to any one of claims 38 to 43.

54. A processing module, characterized in that, Configured to control a network entity to cause the network entity to perform the method according to any one of claims 1 to 37.

55. The processing module according to claim 54, characterized in that, The network entity is a data center.

56. The processing module according to claim 54, characterized in that, The network entity provides services to one or more user equipment (UEs).

57. The processing module according to claim 54, characterized in that, The network entity provides services to one or more base stations (BS).

58. A processing module, characterized in that, Configured as a control device to cause the device to perform the method according to any one of claims 38 to 43.

59. A computer program, characterized in that, When the computer program is run on a computer, it causes the computer to perform the method according to any one of claims 1 to 43.

Images