Data collection and precoding constraints for communication channel information reporting based on artificial intelligence and machine learning

Abstract

The invention provides data collection and precoding constraints for communication channel information reporting based on artificial intelligence and machine learning. An apparatus may include: a receiver configured to receive a reference signal using a channel, and receive constraint information related to the channel; a processing circuit configured to determine channel information based on the reference signal, and generate a representation based on the channel information and the constraint information using a machine learning model; and a transmitter configured to transmit the representation.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to and the benefit of U.S. Provisional Patent Application Serial No. 63 / 530,016 filed on July 31, 2023 and U.S. Provisional Patent Application Serial No. 63 / 569,589 filed on March 25, 2024, both of which are incorporated herein by reference. Technical Field

[0003] The present disclosure relates generally to communication systems, and more particularly to data collection and precoding constraints for communication channel information reporting based on artificial intelligence and machine learning. Background Art

[0004] Some wireless communication systems may implement a codebook subset restriction (CBSR) feature that can reduce downlink interference by restricting user equipment (UE) to use a specific portion of the codebook for channel state information (CSI) feedback. For example, a base station may restrict a UE to report a specific precoding matrix indicator (PMI) by setting specific basis vector coefficients to zero. The problem with this approach is that in systems that use artificial intelligence and / or machine learning to generate and / or compress CSI, there may not be basis vectors for the reported precoder to determine the constrained PMI. Summary of the invention

[0005] To overcome this problem, systems and methods for providing channel-related constraint information to a UE are described herein. This can enable the UE to ensure that the precoder determined using the decoder model is orthogonal to (e.g., does not interfere with) the constrained subspace used for the channel. For example, a base station can configure a constrained subspace for the UE and / or configure a decoder output (e.g., coefficient) that can be zero at a particular index. In an embodiment where the decoder model is shared with the UE, the UE can ensure that the PMI output by the decoder is orthogonal to the constrained subspace (e.g., for CSI in the space-frequency domain) and / or ensure that the decoder output coefficient is zero at a particular index (e.g., for CSI in the angular delay domain). In an embodiment where the decoder model is not shared with the UE, the UE can ensure that the CSI input to the encoder corresponding to the decoder is orthogonal to the constrained subspace and / or ensure that the encoder input is zero at a particular index.

[0006] As another example, the constrained subspace may be used as an explicit input to train the encoder model in such a way that the output of the corresponding decoder model is within a subspace that may be orthogonal to the constrained subspace.

[0007] This approach improves upon previous approaches as it may enable systems that use artificial intelligence and / or machine learning for CSI reporting to reduce downlink interference, e.g., by implementing a constrained subspace for the PMI.

[0008] In some wireless communication systems that use artificial intelligence and / or machine learning models for CSI compression, the UE may collect data for model lifecycle management (LCM) purposes, such as model training, model inference, performance monitoring, etc. Examples of the types of data that may be collected include channel matrices, CSI feedback, target CSI, etc. Data collection at the UE may involve using UE processing resources to perform downlink channel measurements based on a CSI reference signal (CSI-RS) sent by a base station. In one method of collecting data, the base station may configure the UE with a CSI reporting configuration, which the UE may typically use to measure, calculate, and report CSI to the base station.

[0009] The problem with this approach is that different devices (e.g., UE or base station) may need to collect different types of data in different situations. Therefore, when configured with a CSI reporting configuration, the UE calculates one or more types of CSI that it may not need to collect and / or report to the base station, thereby placing an unnecessary burden on the UE's processing resources. Another problem with this approach is that the UE may not be able to and / or is not configured to support one or more CSI-RS configurations for data collection. In addition, the base station may not be able to determine the UE's capabilities and / or configuration for data collection. Another problem with this approach is that the UE may waste resources when preparing an uplink channel for sending CSI that it does not need to report. Therefore, the UE may need to operate differently when collecting data than when reporting CSI according to a normal CSI reporting configuration. Another problem with this approach is that the base station may configure the UE with a CSI reporting configuration that may not generate one or more types of data that the UE may need to collect for UE purposes (e.g., training a local model, performing reasoning with a local model, etc.). In addition, the UE may not be able to determine whether or when the base station expects it to collect data.

[0010] To overcome these issues, systems and methods are described herein to enable a UE to establish a data collection session between the UE and a base station, request collection of one or more specific types of data, and / or declare capabilities for supporting one or more data collection configurations.

[0011] In some embodiments, the UE may send a request to the base station to initiate a data collection session. The request may be accompanied by information indicating one or more types of data to be collected, one or more CSI-RS configurations to be used for data collection, etc. The base station may send an indication to the UE to start the data collection session in response to a request from the UE or at the initiation of the base station.

[0012] In some embodiments, the base station may configure the UE for data collection using a CSI reporting configuration that may be indicated as being used for data collection. In such a configuration, the UE may avoid calculating some or all of the CSI and / or reporting some or all of the CSI to the base station. Additionally or alternatively, the base station may configure the UE to use one or more CSI-RS resource sets for the purpose of data collection.

[0013] In some embodiments, the UE may declare a capability to support one or more data collection configurations. For example, the data collection configuration may include information about the purpose of the data collection, one or more types of data to be collected, and / or information about corresponding supported CSI-RS configurations that may be used to collect the data. The UE may declare data collection capabilities, for example, using one or more types of identifiers (IDs) to indicate the type of information to be collected, use cases for the collected data, supported CSI-RS configurations, etc.

[0014] This approach improves upon previous approaches because it may enable efficient initiation and / or configuration of data collection sessions, it may reduce or eliminate unnecessary computations and / or transmission of unneeded information, and / or it may ensure that the UE may collect data that the UE and / or base station may need for use with artificial intelligence and / or machine learning models.

[0015] A device may include: a receiver configured to: receive a reference signal using a channel, and receive constraint information related to the channel; a processing circuit configured to: determine channel information based on the reference signal, and generate a representation based on the channel information and the constraint information using a machine learning model; a transmitter configured to transmit the representation. The processing circuit may be configured to apply the constraint information to an input of the machine learning model. The processing circuit may be configured to: receive input configuration information for the machine learning model, and generate the representation based on the input configuration information. The machine learning model may be a generative model and the processing circuit may be configured to: receive a reconstruction model, input the representation to the reconstruction model, and apply the constraint information to an output of the reconstruction model. The reconstruction model may include a reference model. The processing circuit may be configured to train the reconstruction model. The processing circuit may be configured to train the machine learning model using the constraint information. The constraint information may indicate a constrained precoder subspace. The constraint information may indicate constrained coefficients associated with the constrained precoder subspace. The representation may include channel state information. The channel state information may include precoding information.

[0016] A device may include: a receiver, and a processing circuit configured to perform one or more operations, the operations including: using the receiver to receive a reference signal using a channel, receiving an indication for collecting data related to the channel for a machine learning model, processing the reference signal based on the indication, and collecting the data for the machine learning model based on the indication and the processing. The device may also include a transmitter, wherein the processing circuit may also be configured to perform one or more operations, the operations including using the transmitter to send a request for the indication for collecting data. The processing circuit may also be configured to perform one or more operations, the operations including using the transmitter to send a purpose of the data. The processing circuit may also be configured to perform one or more operations, the operations including using the transmitter to send a configuration of the reference signal. The processing circuit may also be configured to perform one or more operations, the operations including using the receiver to receive a report configuration for the collection. The report configuration may be suitable for the collection. The processing circuit may also be configured to perform one or more operations, the operations including using the receiver to receive a purpose of the data. The purpose of the data may include one or more of training, reasoning, or monitoring. The processing circuit may also be configured to perform one or more operations, the operations including using the receiver to receive an indication of the type of the data. The type of data may include one or more of a channel estimate, a channel information feedback, or a target channel information. The processing circuit may also be configured to perform one or more operations including receiving an identifier associated with the collection. The identifier may include one or more of a data set identifier, a use case identifier, a base station vendor identifier, or an additional condition identifier.

[0017] A device may include: a receiver, a transmitter, and a processing circuit, the processing circuit being configured to perform one or more operations, the operations including: using a receiver to receive a reference signal using a channel, processing the reference signal to generate data related to the channel for a machine learning model, and using a transmitter to send an indication of the device's ability to collect the data for the machine learning model. The indication may indicate the type of data. The type of data may include one or more of channel estimation, channel information feedback, or target channel information. The indication may indicate a configuration of the reference signal. The indication may indicate time domain information about the reference signal. The indication may indicate a first feature group of the data and a first configuration of the reference signal corresponding to the first feature group. The indication may also indicate a second feature group of the data and a second configuration of the reference signal corresponding to the second feature group. The indication may indicate the ability to collect data for a joint transmission configuration. The joint transmission configuration may include a non-coherent joint transmission configuration.

[0018] A device may include at least one receiver, and a processing circuit configured to perform one or more operations, the operations including: receiving a first reference signal using a first channel using at least one receiver, receiving a second reference signal using a second channel using at least one receiver, processing the first reference signal and the second reference signal to generate channel information for joint transmission using the first channel and the second channel, and collecting data related to the first channel and the second channel for a machine learning model based on the processing. The data may include: first channel information associated with the first channel, second channel information associated with the second channel, and a representation of the first channel information and the second channel information based on the machine learning model. The data may also include: reconstruction of the first channel information based on the representation, and reconstruction of the second channel information based on the representation. The data may include: a first channel matrix based on the first channel, a second channel matrix based on the second channel, and channel state information based on the first channel matrix and the second channel matrix. The data may also include: a first precoding matrix based on the first channel matrix, and a second precoding matrix based on the second channel matrix. The processing circuit may also be configured to perform one or more operations, the operations including training the machine learning model using a training data set. The training data set may be determined by a specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings are not necessarily drawn to scale, and throughout the drawings, similar elements of structure or function are generally represented by similar reference numerals or portions thereof for illustrative purposes. The drawings are intended only to facilitate the description of the various embodiments described herein. The drawings do not describe every aspect of the teachings disclosed herein and do not limit the scope of the claims. In order to prevent the drawings from becoming obscure, not all components, connections, etc. may be shown, and not all components have reference numerals. However, the style of component construction is apparent from the drawings. The drawings, together with the specification, illustrate example embodiments of the present disclosure and, together with the description, are used to explain the principles of the present disclosure.

[0020] Figure 1 An embodiment of a wireless communication device according to the present disclosure is shown.

[0021] Figure 2 Another embodiment of a wireless communication device according to the present disclosure is shown.

[0022] Figure 3 An embodiment of a dual-model training scheme according to the present disclosure is shown.

[0023] Figure 4 An embodiment of a system having a pair of models to provide channel information feedback according to the present disclosure is shown.

[0024] Figure 5 An example embodiment of a system for reporting downlink physical layer information according to the present disclosure is shown.

[0025] Figure 6 An example embodiment of a system for reporting uplink physical layer information according to the present disclosure is shown.

[0026] Figure 7 An example embodiment of a system for reporting downlink physical layer channel state information according to the present disclosure is shown.

[0027] Figure 8 An embodiment of a learning process of a machine learning model according to the present disclosure is shown.

[0028] Fig. 9 An example embodiment of a method for jointly training a pair of encoder and decoder models according to the present disclosure is shown.

[0029] Fig.10 An example embodiment of a method for training a model using the latest shared values ​​according to the present disclosure is shown.

[0030] Fig.11 An example embodiment of a dual-model training scheme utilizing pre-processing and post-processing according to the present disclosure is shown.

[0031] Fig.12 An embodiment of a system using a dual-model approach according to the present disclosure is shown.

[0032] Fig.13 An example embodiment of a user equipment (UE) according to the present disclosure is shown.

[0033] Fig.14 An example embodiment of a base station according to the present disclosure is shown.

[0034] Fig.15 An embodiment of a method for providing physical layer information feedback according to the present disclosure is shown.

[0035] Fig.16 An embodiment of a system having a pair of models for providing channel information feedback according to the present disclosure is shown.

[0036] Fig.17 An example embodiment of a pair of models that may be used for joint compression of channel information and channel quality information according to the present disclosure is shown.

[0037] Fig.18 An embodiment of a system having a pair of models for compressing channel information based on one or more subbands according to the present disclosure is shown.

[0038] Fig.19An embodiment of a pair of models that may be used for channel quality indicator compression across subbands according to the present disclosure is shown.

[0039] Fig. 20 An embodiment of a system having a pair of models for providing channel information compression according to the present disclosure is shown.

[0040] Fig.21 A first embodiment of a pair of models that may be used to implement a compression scheme according to the present disclosure is shown.

[0041] Fig. 22 A second embodiment of a pair of models that may be used to implement a compression scheme according to the present disclosure is shown.

[0042] Fig.23 A third embodiment of a pair of models that may be used to implement a compression scheme according to the present disclosure is shown.

[0043] Fig.24 An embodiment of a communication system using channel constraint information and artificial intelligence and / or machine learning to provide channel information feedback according to the present disclosure is shown.

[0044] Fig.25 An example embodiment of a communication system in which a decoder model is shared with a UE configured with a constrained subspace according to the present disclosure is shown.

[0045] Fig.26 An example embodiment of a communication system in which a decoder model is not shared with a UE configured with a constrained subspace according to the present disclosure is shown.

[0046] Fig. 27 An embodiment of a system according to the present disclosure using constrained subspace information as input to a model is shown.

[0047] Fig.28 An embodiment of a communication system having data collection functionality for one or more models that may be used to provide channel information feedback according to the present disclosure is shown.

[0048] Fig.29 An example embodiment of a communication system having data collection functionality for one or more models that may be used to provide channel information feedback according to the present disclosure is shown.

[0049] Fig.30 An example embodiment of a communication system having data collection functionality for one or more models that may be used to provide channel information feedback for an NCJT scheme according to the present disclosure is shown.

[0050] Fig.31 is a block diagram of an electronic device in a network environment according to an embodiment.

[0051] Fig.32 A system including a UE and a base station communicating with each other according to an embodiment is shown. DETAILED DESCRIPTION

[0052] In the following detailed description, many specific details are set forth in order to provide a thorough understanding of the present disclosure. However, those skilled in the art will appreciate that the disclosed aspects may be practiced without these specific details. In other cases, well-known methods, processes, components, and circuits are not described in detail to avoid obscuring the subject matter disclosed herein.

[0053] References throughout this specification to "one embodiment" or "embodiment" mean that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment disclosed herein. Therefore, the phrases "in one embodiment" or "in an embodiment" or "according to an embodiment" (or other phrases with similar meanings) that appear in various places throughout this specification may not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics may be combined in any suitable manner in one or more embodiments. In this regard, as used herein, the word "exemplary" means "used as an example, instance or illustration". Any embodiment described herein as "exemplary" should not be interpreted as necessarily being preferred or advantageous over other embodiments. In addition, depending on the context discussed herein, a singular term may include a corresponding plural form, and a plural term may include a corresponding singular form. Similarly, hyphenated terms (e.g., "two-dimensional", "predetermined", "pixel-specific", etc.) may occasionally be used interchangeably with corresponding non-hyphenated versions (e.g., "two-dimensional", "predetermined", "pixel-specific", etc.). This occasionally interchangeable use should not be considered inconsistent with each other.

[0054] It should also be noted that the various drawings (including component drawings) shown and discussed herein are for illustrative purposes only and are not drawn to scale. For example, the size of some elements may be exaggerated relative to other elements for clarity. In addition, where deemed appropriate, reference numerals have been repeated in the drawings to indicate corresponding and / or similar elements.

[0055] The terms used herein are only used for the purpose of describing some example embodiments and are not intended to limit the claimed subject matter. As used herein, the singular form is intended to also include the plural form, unless the context clearly indicates otherwise. It will be further understood that when used in this specification, the terms "include" and / or "comprise" specify the presence of the features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0056] It should be understood that when an element or layer is referred to as being on, "connected to" or "coupled to" another element or layer, it may be directly on, connected to or coupled to the other element or layer, or there may be intermediate elements or layers. Conversely, when an element is referred to as being "directly on," "directly connected to" or "directly coupled to" another element or layer, there are no intermediate elements or layers. The same reference numerals always represent the same elements. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0057] As used herein, the terms "first," "second," and the like are used as labels for the nouns that follow them, and do not imply any type of ordering (e.g., spatial, temporal, logical, etc.) unless explicitly defined as such. In addition, the same reference numerals may be used in two or more figures to refer to parts, components, blocks, circuits, units, or modules having the same or similar functions. However, such use is only for simplicity of illustration and ease of discussion; it does not mean that the construction or architectural details of such components or units are the same in all embodiments, or that such commonly referenced components / modules are the only way to implement some example embodiments disclosed herein.

[0058] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those commonly understood by those skilled in the art to which the subject matter belongs. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art, and will not be interpreted in an idealized or overly formal sense unless explicitly defined as such in this article.

[0059] As used herein, the term "module" refers to any combination of software, firmware, and / or hardware configured to provide the functionality described herein in conjunction with the module. For example, software may be implemented as a software package, code, and / or instruction set or instructions, and the term "hardware" as used in any embodiment described herein may include, for example, components, hardwired circuits, programmable circuits, state machine circuits, and / or firmware storing instructions executed by programmable circuits, alone or in any combination. Modules may be implemented collectively or individually as circuits that form part of a larger system, such as, but not limited to, an integrated circuit (IC), a system on a chip (SoC), a component, etc.

[0060] Overview

[0061] In some wireless communication systems, a transmitting device may rely on a receiving device to provide feedback information about channel conditions so that the transmitting device can more efficiently transmit to the receiving device through the channel. For example, in a 5G new radio (NR) system, a base station (e.g., gNodeB or gNB) may send a reference signal to a user equipment (UE) through a downlink (DL) channel. The UE may measure the reference signal to determine the channel condition for the DL channel. Subsequently, the UE may send feedback information (e.g., channel state information (CSI)) indicating the channel condition for the DL channel to the base station through an uplink (UL) channel. The base station may use the feedback information to improve the way it transmits to the UE through the DL channel (e.g., by using beamforming).

[0062] However, sending feedback information about channel conditions may consume a relatively large amount of resources as overhead. In order to reduce the amount of data used to send feedback information, some wireless communication systems may use one or more types of codebooks to enable a receiving device to send implicit channel condition feedback and / or explicit channel condition feedback to a transmitting device. For example, in a 5GNR system, a type-I codebook may be used to provide implicit CSI feedback to a gNB in ​​the form of an index, wherein the index may point to a predefined PMI selected by the UE based on the DL channel condition. The gNB may then use the PMI for beamforming in the DL channel. As another example, a type-II codebook may be used to provide explicit CSI feedback, wherein in the explicit CSI feedback, the UE may derive a PMI that can be fed back to the gNB, and the gNB may use the PMI for beamforming in the DL channel. However, the use of a type-I codebook may not provide CSI feedback with sufficient accuracy. In addition, the use of a type-II codebook may still involve the transmission of a large amount of overhead data on the UL channel.

[0063] The feedback scheme according to the present disclosure may use artificial intelligence (AI), machine learning (ML), deep learning, etc. (any one or all of artificial intelligence (AI), machine learning (ML), deep learning, etc. may be individually and / or collectively referred to as machine learning or ML) to generate a representation of physical layer information for a wireless communication system. For example, in some embodiments, the feedback scheme may use an ML model to generate a representation of feedback information for channel conditions (e.g., a representation of a channel matrix, a precoding matrix, etc.). The representation may be a compressed form, an encoded form, or a form modified in another way of the feedback information, depending on the implementation details, which may reduce the resources involved in transmitting the feedback information between devices.

[0064] Feedback schemes according to the present disclosure may also use machine learning to reconstruct physical layer information from the representation. For example, in some embodiments, the feedback scheme may use an ML model to reconstruct feedback information or an approximation of feedback information from a representation of feedback information for channel conditions. For convenience, the ML model may be referred to simply as a model.

[0065] A model that generates a representation of an input (e.g., physical layer information such as feedback information for channel conditions) may be referred to as a generative model. A model that reconstructs an input or an approximation of an input from a representation of the input may be referred to as a reconstruction model. The output of the reconstruction model may be referred to as a reconstructed input. Therefore, the reconstructed input may be an input applied to the generative model, or an approximation, estimate, prediction, etc. of the input applied to the generative model. The generative model and the corresponding reconstruction model may be collectively referred to as a pair of ML models or a pair of models. In some embodiments, the generative model may be implemented as an encoder model, and / or the reconstruction model may be implemented as a decoder model. Therefore, the encoder model and the decoder model may also be referred to as a pair of ML models or a pair of models.

[0066] For the purpose of distinguishing a model from one or more other models, any model may be referred to as a first model, a second model, model A, model B, etc., and the labels used for models are not intended to imply the type of model unless otherwise apparent from the context. For example, in the context of a pair of models, if model A refers to the generative model, model B may refer to the reconstructive model.

[0067] A node may refer to a base station, a UE, or any other device that may use one or more ML models as disclosed herein. Additional examples of nodes may include a UE-side server, a base station-side server (e.g., a gNB-side server), an eNodeB, a master node, a secondary node, and the like, whether a logical node, a physical node, or a combination thereof. For the purpose of distinguishing a node from one or more other nodes, any node may be referred to as a first node, a second node, a node A, a node B, and the like, and the labels used for the nodes are not intended to imply the type of the node unless otherwise apparent from the context. For example, in some embodiments, a first node may refer to a UE, and a second node may refer to a base station. However, in some other embodiments, a first node may refer to a first UE, and a second node may refer to a second UE configured for sidelink communication with the first UE.

[0068] In some example embodiments, the first node may encode a channel matrix, a precoding matrix, etc. using a first model (e.g., a generation model) to generate a feature vector that may be sent to the second node. The second node may decode the feature vector using a second model (e.g., a reconstruction model) to reconstruct the original information (e.g., a channel matrix, a precoding matrix, etc.) or an approximation of the original information.

[0069] According to some embodiments of the present disclosure, a dual model training scheme may be implemented, wherein the models may be trained in pairs in the dual model training scheme. For example, a reconstruction model may be used to train a generative model, and / or a generative model may be used to train a reconstruction model. In some example implementations, a pair of models may be configured to implement an autoencoder, wherein an encoder model (e.g., for a first node) may be trained using a decoder model (e.g., for a second node) in the autoencoder.

[0070] In some embodiments, a first model (e.g., a generative model) that can be used by a first node for inference can be trained using a second model (e.g., a reconstructive model) that can actually be used by a second node for inference. The training can be performed by the first node, the second node, and / or any other device, for example, by a server that can (e.g., offline) train models and transmit one or more of the trained models to one or more of the nodes for inference.

[0071] Alternatively or additionally, the first model may be trained using a second model that may provide a certain amount of matching between the first model and the second model, even if the second model is not an actual model that may be used by the second node for inference. Alternatively or additionally, the first model may be trained using a reference model for the second model. Alternatively or additionally, the first model may be trained using a second model that may be configured with values ​​for weights, hyperparameters, etc. that may be initialized to predetermined values, random values, etc.

[0072] In some embodiments, a pair of models may be trained simultaneously, sequentially (e.g., alternating between training a first model while freezing the second model, and then training the second model while freezing the first model), and / or in a similar manner using the same or different training data sets.

[0073] In some embodiments, a node may use a quantizer to convert a representation of the physical layer information into a form that can be more easily transmitted over a communication channel. For example, a quantizer may convert a real number (e.g., an integer) representation of the physical layer information into a binary bit stream that can then be applied to a polar encoder or other device for transmission over a physical uplink channel or a physical downlink channel. Similarly, a node may use a dequantizer to convert a bit stream into a representation of the physical layer information that can be used to reconstruct the physical layer information. In some embodiments, a quantizer or dequantizer may be considered to be part of an ML model. For example, a generation model may include an encoder and a corresponding quantizer, and / or a reconstruction model may include a corresponding dequantizer and a decoder.

[0074] According to some embodiments of the present disclosure, one or more frameworks for training models and / or transmitting models between nodes may be implemented. For example, in a first class of frameworks, a first node (node ​​A) may jointly train a pair of models (model A and model B). Node A may use trained model A for inference and transmit trained model B to a second node (node ​​B) that may use trained model B for inference. In a variation of the first class of frameworks, node A may transmit trained model A to node B, and node B may use trained model A to train its own model B for inference.

[0075] In the second type of framework, a reference model may be established as Model A for Node A, and then Node B may use the reference model as Model A to train Model B (e.g., assuming that Node A will use the reference model as Model A for inference). Node A may then use the reference model as Model A without further training, or Node A may continue to train the reference model for use as Model A. In some embodiments, multiple reference models may be established for Model A, and Node B may train one or more versions of Model B corresponding to one or more of the reference models for Model A. In embodiments using multiple reference models for Model A, Node B may train one or more versions of Model B based on the multiple reference models for Model A, and Node B may indicate to Node A which version of Model B it has selected to use, which version or versions of Model B provide the best performance, etc. Based on the indication from Node B, Node A may continue to utilize the reference model corresponding to Model B indicated by Node B, or Node A may select any other model to use as Model A.

[0076] In a third class of frameworks, node A may start with model A, which may be in any initial state (e.g., pre-trained (e.g., trained offline), untrained but configured with initial values, etc.). Node B may start with model B, which may also be in any initial state. In some embodiments, node A and / or node B may have models that match each other (e.g., trained together) before training their own models. One or both nodes may train their respective models for a period of time, and then one or both nodes may share the trained model values ​​and / or trained models with the other nodes. Fig.10 Example embodiments are described in more detail in Fig.10In , a first node (e.g., UE) and a second node (e.g., base station) may have a pair of models (e_0, ​​d_0), where e_0 may be an encoder model at an initial state at the UE, and d_0 may be a decoder model at an initial state at the base station. In a variation of the third class of frameworks, one or both nodes may train their respective models for one or more additional time periods, and one or both nodes may share trained model values ​​and / or trained models with other nodes, for example, at the end of each time period, at the end of alternating time periods, and so on.

[0077] In any framework disclosed herein, when a model is transmitted to or from a node, a corresponding quantizer or dequantizer may be transmitted along with the model.

[0078] In some embodiments, training data may be collected based on resource windows (e.g., windows of time resources and / or frequency resources). For example, a node may be configured to collect training data (e.g., channel estimates) for a specific frequency range (e.g., subcarriers, subbands, etc.) and a specific time range (e.g., symbols, time slots, frames, etc.). The size of the window may be determined, for example, based on the amount of training data that the node can store in memory. The collected training data may be used by one or more nodes for online training, or saved for offline training.

[0079] In some embodiments, pre-processing and / or post-processing can enable a pair of models to run more efficiently. For example, domain knowledge of one or more inputs (e.g., frequency domain knowledge) can be used to perform pre-processing operations on at least a portion of the one or more inputs to generate one or more transformed inputs. The one or more transformed inputs can be applied to a generation model to generate a representation of the one or more transformed inputs. The representation of the one or more transformed inputs can be applied to a reconstruction model, wherein the reconstruction model can generate a reconstructed transformed input (e.g., the one or more transformed inputs or an approximation thereof). Domain knowledge can also be used to perform post-processing operations (e.g., the inverse of the pre-processing operations) on the reconstructed transformed inputs to restore the original one or more inputs or an approximation thereof. Depending on the implementation details, transforming the input and / or output (e.g., based on domain knowledge) can exploit one or more correlations between elements of the one or more inputs, thereby reducing the processing burden, memory usage, power consumption, etc. of the generation model and / or reconstruction model.

[0080] In some embodiments, a node may be provided with processing time for a model. For example, if a node is configured to perform online training of a model (e.g., using a training data set provided to or collected by the node), the node may be expected to update the model within a predetermined number of symbols or other time metric.

[0081] According to some embodiments of the present disclosure, a scheme in which multiple pairs of models can be trained, deployed and / or activated for use by one or more nodes (e.g., a pair of nodes) may be implemented. For example, multiple pairs of different trained models may be activated to handle different channel environments, different matrix dimensions (e.g., for channel matrices, precoding matrices, etc.), etc. In some embodiments, a pair of models may be activated by signaling (e.g., RRC signaling, MAC-CE signaling, etc.). In some embodiments, the first node (e.g., gNB) may also indicate to the second node (e.g., UE) to switch or deactivate the current active model, for example, via RRC, MAC CE, or dynamic signaling. A pair of models may be activated to train one or more models, use one or more models for inference, etc.

[0082] According to some embodiments of the present disclosure, one or more formats for representation of feedback information may be implemented, wherein the representation of the feedback information may be generated by a generation model at a first node and sent to a second node for reconstruction. For example, the format for the representation of the feedback information may be established as a type of uplink control information (UCI). The format may involve one or more types of coding (e.g., polarization coding, low-density parity check (LDPC) coding, etc.), which may depend on, for example, the type of physical channel used to send the UCI.

[0083] In some embodiments, AI and / or ML may be used to improve CSI compression performance, for example by exploiting one or more correlations in the time, frequency, and / or spatial domains, and / or by defining a training data set across time, frequency, and / or space.

[0084] Some embodiments according to the present disclosure may enable a first node (e.g., a UE) to determine precoding information that may be used by a second node (e.g., a base station). The precoding information may be used, for example, to determine channel quality information for a channel that may use the precoding information. Depending on implementation details, enabling the first node to determine the precoding information used by the second node may reduce or eliminate a mismatch between the precoding information and the channel quality information that may be determined based on the precoding information. For example, a base station may share a decoder model with a UE, and the UE may use the model to determine the precoding matrix used by the base station. As another example, the UE may train a decoder (e.g., a reference decoder as described below) to reconstruct the precoding matrix used by the base station.

[0085] In some embodiments according to the present disclosure, a pair of encoder and decoder models may be trained to jointly compress channel information (e.g., channel matrix, precoding matrix, etc.) and channel quality information (e.g., CQI) that may be determined based on the channel information to reduce or eliminate the mismatch between the channel information and the channel quality information. For example, an encoder and a decoder may be trained with a training data set that may include precoding information that may match the corresponding channel quality information. Depending on implementation details, this may reduce or eliminate the mismatch between precoding information (e.g., precoding matrix) and channel quality information (e.g., CQI) that may be determined based on the precoding information.

[0086] According to some embodiments of the present disclosure, one or more machine learning models may be used to compress channel information across one or more subbands. For example, a first node may use an encoder model to generate a representation of the channel information for multiple subbands by combining (e.g., concatenating) channel information (e.g., channel quality information) for multiple subbands into a vector and compressing the vector. A second node may use a decoder to reconstruct the channel information for multiple subbands from the representation. Depending on implementation details, compressing the channel information across one or more subbands may improve performance, reduce complexity, etc.

[0087] According to some embodiments of the present disclosure, one or more decoder models may be used to implement one or more compression schemes for generating a representation of channel information and / or report channel information, which in some embodiments may include precoding information. Depending on the implementation details, such an embodiment may mimic a codebook scheme while providing improved performance and / or flexibility, reduced complexity, etc. In some example embodiments, a pair of machine learning models (e.g., an encoder and a decoder) may be configured and / or trained to generate precoding information from channel information. For example, a first node (e.g., a UE) may apply channel information (e.g., one or more reference signal measurements) to an encoder, which may generate a representation (e.g., a codeword) based on the channel information using a compression scheme. A second node (e.g., a base station) may apply the representation to a decoder, which may construct precoding information (e.g., a precoding matrix) based on the representation. In other example embodiments, a pair of machine learning models may receive any type of information (such as channel state information (e.g., channel quality information), precoding information (e.g., a precoding matrix), rank information, etc.) as input, apply one or more compression schemes, and provide any type of information that can be used to determine precoding information as output.

[0088] Some embodiments that may use one or more decoder models to implement one or more compression schemes may provide spatial compression, frequency compression, a combination of spatial and frequency compression, etc. Depending on the implementation details, the compression scheme may provide separate compression for separate subbands, combined compression for separate subbands, and / or a combination thereof.

[0089] For example, a first node (e.g., a UE) may use one or more encoders to spatially compress channel information for one or more subbands to generate one or more representations (e.g., separate representations) of the channel information for the one or more subbands. A second node (e.g., a base station) may use one or more decoders to recover the channel information for the one or more subbands (e.g., using one or more separate representations).

[0090] As another example, the first node may use one or more spatial encoders and one or more frequency encoders to provide separate spatial compression and combined frequency compression for channel information of one or more subbands to generate one or more representations (e.g., a single representation) of the channel information for the one or more subbands. The second node may use one or more spatial decoders and one or more frequency decoders to recover the channel information for the one or more subbands from the one or more representations using separate spatial decompression and combined frequency decompression.

[0091] As another example, the first node may use a joint spatial and frequency encoder to provide combined spatial and frequency compression of channel information for one or more subbands to generate a combined representation of the channel information for the one or more subbands. The second node may use a joint spatial and frequency decoder to recover the channel information for the one or more subbands from the combined representation.

[0092] Thus, in some embodiments, a decoder model may refer to one or more decoder models, an encoder model may refer to one or more encoder models, and a pair of models may refer to one or more decoder models and one or more encoder models.

[0093] Depending on implementation details, embodiments using one or more decoder models to generate precoding information and / or using other information that may be used to determine precoding information may provide improved performance and / or flexibility, reduced complexity, etc.

[0094] The present disclosure covers many inventive principles related to artificial intelligence and machine learning for the physical layer of a communication system. These principles may have independent utility and may be implemented separately, and not every embodiment will utilize every principle. In addition, the principles may also be implemented in various combinations, some of which may amplify the benefits of the individual principles in a synergistic manner.

[0095] For purposes of illustration, some embodiments may be described in the context of some specific implementation details and / or applications in a 5G NR system, such as compression, decompression, and / or transmission of channel feedback information between one or more UEs, base stations (e.g., gNBs), etc. However, the present principles are not limited to these details and / or applications, and may be applied in any other context where physical layer information may be processed and / or transmitted between wireless devices, regardless of whether any of the devices may be a base station, a UE, a peer device, etc., and regardless of whether the channel may be a UL channel, a DL channel, a peer channel, etc. In addition, the present principles may be applied to any type of wireless communication system that may process and / or exchange physical layer information, such as other types of cellular networks (e.g., 4G LTE, 6G, and / or any subsequent generation cellular networks), Bluetooth, Wi-Fi, etc.

[0096] Machine Learning Models for the Physical Layer

[0097] Figure 1 An embodiment of a wireless communication device according to the present disclosure is shown. Device 101 may include a machine learning model 103, wherein the machine learning model 103 may receive physical layer information 105 as input and generate a representation 107 of the physical layer information as output. In some embodiments, as shown by arrow 109, device 101 may send the representation 107 of the physical layer information to one or more other devices.

[0098] The representation 107 of the physical layer information may be a compressed, encoded, encrypted, mapped, or otherwise modified form of the physical layer information 105. Depending on implementation details, modification of the physical layer information 105 by the machine learning model 103 to generate the representation 107 of the physical layer information may reduce resources involved in transmitting the physical layer information 105 between devices.

[0099] The machine learning model 103 may be implemented using one or more models of any type of AI and / or ML model, including neural networks (e.g., deep neural networks), linear regression, logistic regression, decision trees, linear discriminant analysis, naive Bayes, support vector machines, learning vector quantization, etc. For example, the machine learning model 103 may be implemented using a generative model.

[0100] The physical layer information 105 may include any information related to the operation of the physical layer of the wireless communication device. For example, the physical layer information 105 may include information related to one or more physical layer channels, signals, beams, etc. (e.g., state information, precoding information, etc.). Examples of physical layer channels may include one or more of the following items: physical broadcast channel (PBCH), physical random access channel (PRACH), physical downlink control channel (PDCCH), physical downlink shared channel (PDSCH), physical uplink shared channel (PUSCH), physical uplink control channel (PUCCH), physical sidelink shared channel (PSSCH), physical sidelink control channel (PSCCH), physical sidelink feedback channel (PSFCH), etc. Examples of physical layer signals may include one or more of a primary synchronization signal (PSS), a secondary synchronization signal (SSS), a channel state information reference signal (CSI-RS), a tracking reference signal (TRS), a sounding reference signal (SRS), etc.

[0101] Figure 2 Another embodiment of a wireless communication device according to the present disclosure is shown. Device 202 may include a machine learning model 204, wherein the machine learning model 204 may receive a representation 208 of physical layer information as input and generate a reconstruction 206 of the physical layer information on which the representation 208 may be based as output. In some embodiments, as shown by arrow 210, device 202 may receive the representation 208 of physical layer information from one or more other devices.

[0102] Reconstruction 206 (which may be referred to as an input to reconstruction) may be physical layer information on which representation 208 may be based, or an approximation, estimate, prediction, etc. of physical layer information on which representation 208 may be based. Reconstruction 206 may be a decompressed, decoded, decrypted, inversely mapped, or otherwise modified form of the physical layer information on which representation 208 may be based.

[0103] The machine learning model 204 may be implemented using one or more of any type of AI and / or ML models, including neural networks (e.g., deep neural networks), linear regression, logistic regression, decision trees, linear discriminant analysis, naive Bayes, support vector machines, learning vector quantization, etc. For example, the machine learning model 204 may be implemented using a reconstruction model.

[0104] The reconstructed physical layer information 206 may include any information related to the operation of the physical layer of the wireless communication device, for example, as described above with respect to Figure 1 One or more channels, signals, etc. described in the embodiments shown in .

[0105] Although not limited to any particular use, Figure 1 and Figure 2 The wireless communication devices 101 and 202 shown in the figure can be used together to facilitate the transmission of physical layer information from between the devices. For example, in some embodiments, the device 101 can be implemented as a UE where the model 103 is implemented as a generation model, and the device 202 can be implemented as a base station where the model 204 can be implemented as a reconstruction model. In such an embodiment, the generation model 103 can generate a representation 107 by compressing the physical layer information 105 (e.g., related to a DL channel from a base station to a UE). The UE can send the representation 107 to the base station (e.g., using a UL channel). The base station can input the representation (indicated as 208) to the reconstruction model 204, where the reconstruction model 204 can generate reconstructed physical layer information 206. The base station can use the reconstructed physical layer information 206, for example, to facilitate DL transmission from the base station to the UE. Depending on the implementation details, sending the physical layer information 105 in the form of a compressed representation 107 can reduce the amount of UL resources associated with sending the physical layer information 105.

[0106] Dual model training

[0107] Figure 3 An embodiment of a dual-model training scheme according to the present disclosure is shown. Figure 3 The embodiment 300 shown in FIG. 1 may be used, for example, with Figure 1 and Figure 2 The invention may be used in conjunction with one or more of the models shown in or any other embodiments disclosed herein.

[0108] Reference Figure 3 , training data 311 may be applied to generation model 303, where generation model 303 may generate representation 307 of the training data. Reconstruction model 304 may generate reconstruction 312 of the training data based on representation 307 of the training data. In some embodiments, generation model 303 may include a quantizer to convert representation 307 into a quantized form (e.g., a bitstream) that may be transmitted over a communication channel. Similarly, in some embodiments, reconstruction model 304 may include a dequantizer, where the dequantizer may convert the quantized representation 307 (e.g., a bitstream) into a form that may be used to generate reconstructed training data 312.

[0109] The generative model 303 and the reconstruction model 304 may be trained in pairs, for example, by providing training feedback 314 to the generative model 303 and / or the reconstruction model 304 using a loss function 313. The training feedback 314 may be implemented, for example, using gradient descent, back propagation, etc. In embodiments where one or both of the generative model 303 and the reconstruction model 304 may be implemented using one or more neural networks, the training feedback 314 may update one or more values ​​of weights, hyperparameters, etc. in the generative model 303 and / or the reconstruction model 304.

[0110] In some embodiments, a loss function 313 (which may be implemented, for example, at least in part, using a reconstruction loss) may be operated to train the generation model 303 and the reconstruction model 304 to generate reconstructed training data 312 that is close to the original training data 311. This may be done, for example, by reducing or minimizing the loss output of the loss function 313.

[0111] For example, if the training data 311 is denoted as x and the reconstructed training data 312 is denoted as Then the generation model 303 can be represented by the function f(x), and the reconstruction model 304 can be represented by the function ) indicates that, and therefore, The loss function 313 can be expressed as Thus, in some embodiments, training a pair of models 303 and 304 may involve reducing or minimizing L by using training feedback 314 .

[0112] Although not limited to any particular type of representation 307 of the training data, in some embodiments, the pair of models 303 and 304 can seek to reduce the dimensionality of the representation 307 of the training data relative to the original training data 311. For example, the generative model 303 can be trained to generate feature vectors, where the feature vectors can identify or separate one or more features (e.g., latent features) of the training data, which can reduce the overhead associated with storing and / or transmitting the representation 307. The reconstruction model 304 can similarly be trained to reconstruct the original training data 311 or an approximation thereof based on the representation 307.

[0113] Once trained, the generative model 303 and / or the reconstruction model 304 can be used, for example, in Figure 1 and Figure 2 Inferences can be made in one or both of the wireless communication devices 101 and 202 shown in FIG. 1 or any other embodiment disclosed herein. Figure 3 The described dual model training scheme may be used with one or more frameworks for training models and / or transmitting models between wireless devices as disclosed herein. Figure 3 The described training may be performed anywhere, for example, at the wireless device 101, at the wireless device 202, at another location (e.g., at a server remote from both the device 101 and the device 202), or at any combination of such locations. In addition, once trained, one or both of the generated model 303 and / or the reconstructed model 304 may be transmitted to another location for inference. In some embodiments, once trained, one of the models may be discarded and the remaining model may be used, for example, as a pair with a separately trained model.

[0114] Machine learning model for channel information feedback

[0115] Figure 4 An embodiment of a system having a pair of models to provide channel information feedback according to the present disclosure is shown. Figure 4 The system 400 shown in FIG. 4 can be used to implement the devices, models, training schemes, etc. disclosed herein (including Figure 1 , Figure 2 and Figure 3 Any of those shown in ), or can use the devices, models, training schemes, etc. disclosed herein (including Figure 1 , Figure 2 and Figure 3 Any of those shown in ) is implemented.

[0116] Reference Figure 4 , the system 400 may include a first wireless device 401 and a second wireless device 402. The first wireless device 401 may be configured to receive transmissions from the second wireless device 402 over a channel 415. To improve the effectiveness (e.g., efficiency, reliability, bandwidth, etc.) of transmissions over the channel 415, the first wireless device 401 may provide feedback to the second wireless device 402 in the form of channel information 405, which may be obtained, for example, by measuring one or more signals (e.g., reference signals) sent by the second wireless device 402 over the channel 415.

[0117] The first wireless device 401 may use a first machine learning model 403 (which may be implemented as a generative model in this example) to generate a representation 407 of the channel information 405. The first wireless device 401 may send the representation 407 to the second wireless device 402, for example, using another channel, signal, etc. 416. The representation 407 may be a compressed, encoded, encrypted, mapped, or otherwise modified form of the channel information 405. Depending on implementation details, modifying the channel information 405 by the machine learning model 403 to generate the representation 407 may reduce the resources involved in sending the channel information 405 to the second wireless device 402.

[0118] The second wireless device 402 may apply a representation 407 of the channel information to a second machine learning model 404, where, in this example, the second machine learning model 404 may be implemented as a reconstruction model. The reconstruction model 404 may generate a reconstruction 406 of the channel information 405. The reconstruction 406 (which may be referred to as an input to the reconstruction) may be the channel information 405, or an approximation, estimate, prediction, etc. of the channel information 405, on which the representation 407 may be based. The reconstruction 406 may be a decompressed form, a decoded form, a decrypted form, an inversely mapped form, or a form otherwise modified by the channel information 405. The second wireless device 402 may use the channel information 405 to improve the manner in which the second wireless device 402 transmits to the first wireless device 401 over the channel 415.

[0119] Figure 4 The system 400 shown in the embodiment is not limited to any particular apparatus (e.g., UE, base station, peer device, etc.), application (e.g., 4G, 5G, 6G, Wi-Fi, Bluetooth, etc.), and / or implementation details. However, for the purpose of illustrating some of the inventive principles, some example embodiments may be described in the context of a 5G NR system, in which a UE may receive different DL signals from a gNB.

[0120] Uplink and downlink transmission

[0121] In an NR system, a UE may receive DL transmissions including various information from a gNB. For example, a UE may receive user data from a gNB in ​​a specific configuration of time and frequency resources called a physical downlink shared channel (PDSCH). The media access control (MAC) layer at the gNB may provide user data intended to be transmitted to the corresponding MAC layer at the UE side. The physical (PHY) layer of the UE may receive a physical signal received on the PDSCH and apply it as input to a PDSCH processing chain, where the output of the PDSCH processing chain may be fed as input to the MAC layer at the UE. Similarly, the UE may receive control data from a gNB using a physical downlink control channel (PDCCH). The control data may be referred to as downlink control information (DCI) and may be converted into a PDCCH signal by a PDCCH processing chain on the gNB side.

[0122] The UE may send UL signals to the gNB using the physical uplink shared channel (PUSCH) and the physical uplink control channel (PUCCH) to transmit user data and control information, respectively. The PUSCH may be used by the UE MAC layer to transmit data to the gNB. The PUCCH may be used to transmit control information, which may be referred to as uplink control information (UCI), which may be converted into a PUCCH signal by the PUCCH processing chain at the UE side.

[0123] Channel State Information

[0124] In an NR system, a UE may include a channel state information (CSI) generator, wherein the channel state information (CSI) generator may calculate a channel quality indicator (CQI), a precoding matrix indicator (PMI), a CSI reference signal resource indicator (CRI), and / or a rank indication (RI), wherein any one or all of the channel quality indicator (CQI), the precoding matrix indicator (PMI), the CSI reference signal resource indicator (CRI), and / or the rank indication (RI) may be reported to one or more gNBs serving the UE. The CQI may be associated with a modulation and coding scheme (MCS) for adaptive modulation and coding and / or frequency selective resource allocation, the PMI may be used for a channel-dependent closed-loop multiple-input multiple-output system, and the RI may correspond to the number of useful transmission layers.

[0125] In an NR system, CSI generation may be performed based on a CSI reference signal (CSI-RS) sent by a gNB. The UE may use the CSI-RS to measure downlink channel conditions and generate CSI, for example, by performing channel estimation and / or noise variance estimation based on measurements of the CSI-RS signal.

[0126] In NR systems, CSI may be reported to the serving gNB using a type-I codebook, where the type-I codebook may provide implicit CSI feedback to the gNB in ​​the form of an index that may point to a predefined PMI. Alternatively or additionally, CSI may be reported to the serving gNB using a type-II codebook, where the type-II codebook may provide explicit CSI feedback, where the UE may determine one or more principal eigenvectors or singular vectors based on the DL channel conditions. The UE may then use the principal eigenvectors or singular vectors to derive a PMI that may be fed back to the gNB, which may use the PMI for beamforming in the DL channel.

[0127] For example, in embodiments with a limited number of antenna ports and / or users, the use of a codebook may provide adequate performance. However, in systems with a larger number of antenna ports and / or users (e.g., multiple-input multiple-output (MIMO) systems), and particularly where frequency division duplexing (FDD) is used, the relatively low resolution of the type-I codebook may not provide CSI feedback with sufficient accuracy. Furthermore, the use of the type-II codebook may still involve the transmission of a large amount of overhead data on the UL channel.

[0128] Depending on implementation details, some embodiments of the machine learning based channel information feedback scheme according to the present disclosure may enable the UE to send complete CSI information to the gNB while reducing the overhead associated with the UL transmission to the gNB. In addition, the principles of the present invention are not limited to the UE sending CSI to the gNB, but can be applied to any situation where a first device can send channel information feedback to a second device (e.g., reporting channel conditions for an uplink channel from the UE to the gNB, reporting channel conditions for a sidelink channel between UEs, etc.).

[0129] Example Embodiments

[0130] Figure 5 An example embodiment of a system for reporting downlink physical layer information according to the present disclosure is shown. System 500 may include UE 501 (which may be designated as Node B) and gNB 502 (which may be designated as Node A). gNB 502 may send a transmission of DL signal 517 (e.g., a reference signal (RS) transmission) to UE 501, where UE 501 may extract measurement 518 from the transmission. UE 501 may include model 503, where model 503 may be configured, for example, as an encoder to encode measurement 518 as a feature vector related to the DL physical layer. The encoded measurement may then be quantized by quantizer 519 and sent back to gNB 502 as UL signal 520 (e.g., a bitstream). In some embodiments, the description of the model at a node may also include a quantizer and / or dequantizer description, for example, channel information (e.g., true CSI codewords) at the output of the encoder model may be mapped to a quantized value or function of the bitstream, and vice versa at a decoder model at another node. The gNB 502 may apply the received UL signal 520 to an inverse quantizer 521 to generate an equivalent feature vector, wherein the equivalent feature vector may be fed to the model 504 to extract information 522 (e.g., necessary or optional information) related to the DL physical layer.

[0131] Figure 6 An example embodiment of a system for reporting uplink physical layer information according to the present disclosure is shown. In some aspects, Figure 6 The system 600 shown in FIG. 6 may be similar to Figure 5 5 , but system 600 may be configured to report uplink physical layer information rather than downlink physical layer information.

[0132] Specifically, the system 600 may include a gNB 601 (which may be designated as a Node B) and a UE 602 (which may be designated as a Node A). The UE 602 may send a transmission of a UL signal 617 (e.g., a reference signal (RS) transmission) to the gNB 601, and the gNB 601 may extract a measurement 618 from the transmission. The gNB 601 may include a model 603, wherein the model 603 may be configured, for example, as an encoder to encode the measurement 618 into a feature vector related to the UL physical layer. The encoded measurement may then be quantized by a quantizer 619 and sent back to the UE 602 as a DL signal 620 (e.g., a bitstream). The UE 602 may apply the received DL signal 620 to an inverse quantizer 621 to generate an equivalent feature vector, wherein the equivalent feature vector may be fed to the model 604 to extract information 622 related to the UL physical layer (e.g., necessary or optional information).

[0133] Figure 7 An example embodiment of a system for reporting downlink physical layer channel state information according to the present disclosure is shown. Depending on the implementation details, Figure 7 The system 700 shown in the diagram may enable a gNB or other base station to retrieve complete CSI information (e.g., as opposed to codebook-based pointers, precoding matrix indicators, etc.) from a UE when using an ML model to compress CSI (e.g., compressing the CSI into a relatively low number of bits), thereby reducing uplink resource overhead involved in sending CSI.

[0134] The system 700 may include a UE 701 and a gNB 702. The gNB 702 may transmit a DL reference signal 717, such as a CSI-RS or a demodulation reference signal (DMRS), which may enable the UE 701 to determine CSI 718 for a DL channel 715. The UE 701 may include an ML model 703, wherein the ML model 703 may be configured as an encoder to encode the CSI 718 into a feature vector. The UE 701 may also include a quantizer 719, wherein the quantizer 719 may quantize the feature vector into a bitstream that may be sent to the gNB 702 using a UL signal 720. The gNB 702 may include an inverse quantizer 721 that may reconstruct the feature vector from the bitstream. The feature vector may then be fed to the ML model 704, wherein the ML model 704 may be configured as a decoder to reconstruct an estimate 722 of the CSI 718.

[0135] In some embodiments, the performance metric f(H,H^) may be used to evaluate the accuracy of the design, configuration and / or training of the encoder model 703, the decoder model 704, the quantizer 719 and / or the inverse quantizer 721.

[0136] For example, the performance metric f(H,H^) can be implemented as a measure of the error between channel estimates as follows:

[0137] f(H,H^)=‖HH^‖^2 / ‖H‖^2 (Equation 1) where H and H^ may represent channel estimates (e.g., CSI) at UE 701 and gNB 702, respectively. Such performance metrics may be useful, for example, for evaluating the accuracy of channel state information extracted by gNB 702.

[0138] Additionally or alternatively, the system 700 may be configured to enable the UE 701 to use the DL reference signal 717 to determine a precoding matrix based on current channel conditions. The precoding matrix may then be encoded as features by the encoder model 703, quantized by the quantizer 719, and sent to the gNB 702 using the UL signal 720. At the gNB 702, the inverse quantizer 721 may recover the eigenvectors, which may be applied to the decoder model 704 to reconstruct an estimate of the precoding matrix. For example, for a channel realization H, a suitable precoding matrix may be realized as a set S of singular vectors using a singular value decomposition (SVD) of H, where the singular value decomposition (SVD) of H may be given as H=SΣD, where Σ may be a diagonal matrix and D may be a unitary matrix. In such an embodiment, the encoder model 703, the decoder model 704, the quantizer 719 and / or the dequantizer 721 may be configured to enable the gNB 702 to extract a set (e.g., matrix) S of singular vectors and may implement performance metrics accordingly. Figure 7 The embodiment shown in reports downlink physical layer information, but other embodiments may be configured to report uplink physical layer information, sidelink physical information, etc. using similar principles according to the present disclosure.

[0139] Model development, training, and operations

[0140] Artificial intelligence (AI), machine learning (ML), deep learning, etc. (as described above, any or all of artificial intelligence (AI), machine learning (ML), deep learning, etc. may be individually and / or collectively referred to as machine learning or ML) may provide techniques for inferring one or more functions (e.g., complex functions) of data according to the present disclosure. In a machine learning process, a sample of data may be provided to an ML model, where the ML model may in turn apply one of a variety of machine learning techniques to learn how to determine the one or more functions using the provided data sample. For example, a machine learning process may cause the ML model to learn a function f(x) of a data sample input x. As described above, an ML model may also be referred to as a model.

[0141] In some embodiments, the machine learning process (also referred to as the development process) may be performed in one or more stages (also referred to as periods) such as training, validation, testing, and / or inference (also referred to as application stages). Some embodiments may omit one or more of these stages and / or include one or more additional stages. In some embodiments, all or a portion of one or more stages may be combined into one stage, and one stage may be divided into multiple stages. In addition, the order of the stages or portions thereof may be changed.

[0142] In the training phase, the model may be trained to perform one or more target tasks. The training phase may involve the use of a training data set, wherein the training data set may include i) data samples, and ii) the results of the function f(x) for the samples (e.g., each sample) in the training data set. In the training phase, one or more training techniques may enable the model to learn an approximate relationship (e.g., an approximate function) that may be expressed as a function f(x) or closely related to the function f(x).

[0143] In the validation phase, (e.g., after performing initial training) the model may be tested to assess the fitness of the trained model for one or more target tasks. If the validation results are not satisfactory, the model may undergo further training. If the validation phase provides successful results, the training phase may be considered successfully completed.

[0144] In the testing phase, the trained model may be tested to evaluate the fitness of the trained ML model for one or more target tasks. In some embodiments, the trained model may not proceed to the testing phase unless training is complete and validation provides successful results.

[0145] In the inference phase, the trained model is used to perform one or more target tasks (e.g., in real-world applications).

[0146] During the testing and / or inference phase, the model may use the learned approximate function obtained during the training phase to determine function values ​​f(x) for other data samples that may be different from those in the training phase.

[0147] In some embodiments, the success and / or execution of the machine learning process may involve the use of a sufficiently large training data set, wherein the training data set may contain sufficient information about the function f(x) and thereby enable the model to obtain an acceptably close approximation of the function f(x) through the training phase.

[0148] Figure 8An embodiment of a learning process of a machine learning model according to the present disclosure is shown. Process 800 may begin at operation 823, where a training process may be initialized. For example, the structure of the model may be determined, the values ​​of the model (e.g., neural network weights, hyperparameters, etc.) may be initialized, a training data set with a sufficient number of samples may be constructed, and so on.

[0149] In operation 824, the initialized model may be trained using the training data set to determine the configuration of the candidate trained model, for example, by updating the values ​​of neural network weights, hyperparameters, etc. using gradient descent, back propagation, etc.

[0150] In some embodiments, there may be a correlation between the construction of the training data set and the training phase. For example, the training phase may involve a relatively long duration to complete, and the duration may depend on the number of samples in the training data set. The duration may in turn depend on the type of training. For example, for full training and / or initial training, the model may be initialized, and training may be performed using a large data set consisting of many samples (e.g., samples that may not have been previously used to train the model). As another example, for partial training and / or update training, the model may be pre-trained (or partially trained), and an event (e.g., obtaining a new data sample, performance degradation of the model, a model update event, etc.) may prompt the model to be modified or adapted. In the case of partial training and / or update training, a modified data set may be used to train the model, wherein the modified data set may be different from the large training data set used for full training and / or initial training. For example, the modified training data set may be a subset of the complete data set used for initial training, a collection of new data samples that have been newly acquired, or a combination thereof.

[0151] At operation 825, the trained candidate model may be validated. In some embodiments, the validation phase 825 may be performed iteratively with the training phase 824. For example, if a candidate model fails the validation phase 825, the candidate model may be returned to the training phase 824, which may generate a new candidate model. In some embodiments, different criteria (e.g., classification accuracy, minimum mean square error (MMSE), etc.) may be established to determine validation success or failure.

[0152] In some embodiments, failed candidate models may not be allowed to return to the training phase 824 (e.g., after failing a number of times exceeding a threshold, or if a performance criterion does not exceed a threshold for a particular duration or a particular number of validation steps), and the method may terminate at operation 826. However, if the performance of the candidate model using the validation data (e.g., based on the criteria used to determine success or failure) is determined to be acceptable, then the validation may be considered successful, and the trained candidate model may be passed to the testing phase at operation 827.

[0153] At operation 827, the performance of the trained model candidates that have passed the validation phase can be evaluated. The criteria used to declare successful testing and / or failure of a model during the testing phase of development can be similar to the criteria used in the validation phase. However, one or more parameters (e.g., number of steps, performance threshold, etc.) used with the criteria during the testing phase may or may not be different from the parameters used in the validation phase.

[0154] If the test is successful, the model may be designated as the final model and the process may proceed to operation 828. In some embodiments, if the model fails the testing phase, the process may return to the training phase at operation 824 for further training. However, in some embodiments, further training may not be allowed (e.g., based on criteria similar to those used during the validation phase 825) and the process may terminate at operation 826.

[0155] Model training and deployment framework

[0156] According to some embodiments of the present disclosure, one or more frameworks for training and / or deploying models may be implemented. In some embodiments of the framework disclosed herein, one or more models trained and / or developed by a node may be tested against one or more reference models for the node, for example, to assess the conformance of the model with one or more potential test cases that may be specified for a corresponding application.

[0157] In any of the embodiments of the framework disclosed herein, the quantizer function may be differentiable in some or all of the quantizer ranges (e.g., substantially in the entire range), with the derivative value being substantially zero (e.g., with probability 1). Depending on the implementation details, this may result in little or no backpropagation for updating the encoder weights being provided. Thus, in some embodiments, the quantizer function may be approximated during the training period using a differentiable function that may be referred to as f_(quantizer, approx)(x) (e.g., a reference differentiable quantizer function), while the actual quantizer function may be used during the inference period. Similarly, the dequantizer function may be approximated during the training period using a differentiable function that may be referred to as f_(dequantizer, approx)(x) (e.g., a reference differentiable dequantizer function), while the actual dequantizer function may be used during the inference period. In some embodiments, a quantizer function or dequantizer function used in conjunction with a model may be considered part of a complete description of the corresponding model and may be transmitted along with and as part of the model. Therefore, using any of the frameworks disclosed in this document, if a first node shares a trained model with a second node (for example, if node A trains a pair of models (model A and model B), the trained model B is sent to node B), the first model may also share one or both of the approximate quantizer function f_(quantizer,approx)(x) and / or the approximate dequantizer function f_(dequantizer,approx)(x) with the second node, for example via RRC signaling.

[0158] Although the framework disclosed herein is not limited to any particular application and / or implementation details, in some embodiments, and depending on the implementation details, the framework may be used to train and / or test a model that may reduce CSI feedback overhead.

[0159] Joint training framework

[0160] In some embodiments, a pair of models (e.g., model A and model B) may be jointly trained by one of the two nodes (node ​​A or node B), and the trained model for the non-training node may be transmitted to the non-training node (e.g., if the joint training is performed by node A, the trained model B may be transmitted to node B) for inference. For example, in the context of CSI compression, the base station may perform joint training of a pair of encoder and decoder models, and then transmit the encoder model to the UE. The encoder model may also be referred to as an encoder, and the decoder model may also be referred to as a decoder.

[0161] In some embodiments of the joint training framework, further training (e.g., fine-tuning) of one or both of the trained models may be performed by the following nodes (e.g., to improve or optimize one or both of the models): at which the model can be used for inference. In some embodiments, the further training can be based on online data that can be obtained by one or more of the nodes, for example during ongoing communication.

[0162] In some embodiments of the joint training framework, the training node can use the corresponding quantizer function and / or dequantizer function (e.g., approximate and / or differentiable quantizer function and / or dequantizer function) to train one or both models. The node receiving the trained model can also receive and use the corresponding quantizer function and / or dequantizer function for further training, validation, testing, inference, etc.

[0163] In some embodiments, joint training of models by nodes may produce models that can be jointly matched to the target task and thus can provide improved or optimized performance. Depending on the implementation details, this performance improvement may outweigh any communication overhead associated with transmitting the model to different nodes and / or any mismatch between models and / or nodes caused by, for example, joint training at one node that may be produced by a different manufacturer than the other nodes.

[0164] In a variation of the joint training framework, a node (e.g., a base station) may jointly train a pair of encoder models and decoder models. For example, a reference differentiable quantizer function and inverse quantizer function as described above may be used to train an encoder and decoder pair. The base station may then share the trained decoder model with the UE, for example via RRC signaling. However, the base station may or may not share the trained encoder model with the UE. If the base station shares the trained encoder model with the UE, the UE may use the trained encoder model as a reference encoder model. If the base station does not share the trained encoder model with the UE, the UE may establish a reference encoder model based on, for example, randomly initialized weights, weights that may be selected for the UE implementation, or any other basis.

[0165] The UE may then use the trained decoder model it receives from the base station to train the reference encoder model. The reference encoder model may be trained online (which may refer to training that may be performed during operation). In some embodiments, online training (which may refer to training performed using training data (e.g., channel estimates H) that may be collected during operation) may be performed on the fly. Thus, the UE may use channel estimates H that may be collected over time to train the reference encoder model. For example, at a specific point during training, the collected channel estimates may be used as a new training data set. In addition, the collected channel estimates H may also be stored for future online and / or offline training by the UE or any other device.

[0166] The UE may then use the trained encoder model for inference. The UE may also share the trained encoder model with the base station. The training process may continue as more training samples (eg, channel estimates H) are collected and used by the UE for training.

[0167] Fig. 9 An example embodiment of a method for jointly training a pair of encoder models and decoder models according to the present disclosure is shown. In operation 929, the base station may use a training data set to jointly train a pair of reference encoder models and decoder models, wherein the pair of reference encoder models and decoder models may be referred to as Enc_ref and Dec_ref. In operation 930, the base station may share a reference decoder model Dec_ref with the UE. In operation 931, the base station decides whether to share the reference encoder model with the UE. If the base station shares a reference encoder with the UE, then in operation 932, the UE may use the shared reference encoder as a reference encoder for training. If the base station does not share a reference encoder with the UE, then in operation 933, the UE may establish a reference encoder model, for example, using random weights, using weights based on UE implementation, etc. In operation 934, the UE may train the reference encoder model at a time point t_i. The time point t_i may be determined, for example, as a time at which the UE has performed and collected sufficient channel estimates since a previous time point. This can be implemented, for example, as shown in Algorithm 1, where, in Algorithm 1, for each time point t_i, the UE may have collected a new online training set S_i on the fly, where S_i may include channel estimates from t_(i-1) to t_i, and N may be the maximum number of online trainings on the UE side. In some embodiments, after completing Algorithm 1, the UE may share the trained encoder model with the base station.

[0168] Algorithm 1

[0169]

[0170] Any training and deployment framework disclosed herein can be used with any type and / or combination of devices, and with any type of models and / or physical layer information. Figure 8 In the illustrated embodiment, the base station performs initial joint training and the encoder and decoder may be trained and used using channel estimates, but in other embodiments, the joint training may be performed by the UE or any other device and the model may be trained and used using precoding matrices or any other type of physical layer information.

[0171] In some embodiments, a node such as a UE or a base station may collect new training data within a window (e.g., a well-defined time and / or frequency window). For example, the collected data may be used, for example, to construct a training data set that the node may use to train a model. The window may be configured with a start time and / or end time that may be determined, for example, by the base station. In some embodiments, the timeline used to determine the data collection window may be measured, for example, from one or more CSI-RS resources.

[0172] Alternatively or additionally, online training may be performed as follows. A first node (base station) may have a first model and a second node may have a second model, wherein the second model may form a pair with the first model. In some embodiments, the first node and the second node may operate in a connected mode (e.g., an RRC connected mode). One or both of the nodes may have obtained their models through sharing with other nodes. In this example, one of the nodes may be a base station and the other node may be a UE.

[0173] The base station may configure a predetermined online training data set for the UE, and both nodes may use the predetermined online training data set to update their respective models. When a node updates its own model, it may be assumed that another model at another node is frozen. In some embodiments, one or more online training data sets may be specified (e.g., as part of a specification and / or provided to the UE and / or base station by a third node). Once a first node updates its first model (e.g., an encoder or decoder), it may share the updated first model with a second node, and assuming that the first model is frozen, the second node may begin training its second model. The models may continue to alternate between periodic training and freezing their models, for example, until an end time is reached.

[0174] Although the embodiments disclosed above may be described in the context of a UE that can update and use an encoder, in some embodiments with online training, the UE and the base station (or any two other nodes having a pair of models, such as two UEs configured for sideband communication) may both collect new training data and use the new training data to update their own models (e.g., an encoder or a decoder) or both models (e.g., both an encoder and a decoder that may be configured, for example, as an autoencoder).

[0175] In some embodiments, the first node may share newly collected training data (e.g., channel matrices) with the second node by sending the training data as data or control information. For example, the UE may generate a binary representation of one or more channel matrices and send the representation using PUSCH or PUCCH following conventional procedures for uplink transmission (i.e., coding, modulation, etc.).

[0176] Alternatively or additionally, the UE may encode the channel matrix it has obtained using its encoder as currently trained. The UE may send the encoded channel matrix (which may be referred to as a CSI codeword) to the base station. The base station may then use its currently trained decoder to recover the channel matrix. The base station may then include the recovered channel matrix in a new training data set, which may be used for further (e.g., online) training at the base station.

[0177] In some embodiments, in addition to exchanging training data between nodes, one or more nodes may also share their most recently trained models (e.g., encoders and / or decoders) with another node. The sharing of training data and / or models may be performed at intervals (e.g., the model may be sent when it is updated), which may manage the amount of communication overhead involved in such sharing.

[0178] In a framework with online training of a model, a node may use a memory buffer to store the collected physical layer information (e.g., a CSI matrix from time t_(i-1) to time t_i). Depending on implementation details, a node may collect some or all of a new training data set before it can start using the new training data to update the model. However, if a node uses a dedicated memory buffer to store the training data set, and the interval between time t_(i-1) and time t_i exceeds a certain value, the amount of memory involved in storing the training data may exceed the available buffer size because the number of CSI-RS in the time window may become too large. In addition, even if the interval between time t_(i-1) and time t_i is generally short enough to prevent buffer overflow, a node may encounter some reference signals with relatively short periodicity (e.g., a relatively large number of reference signals (e.g., CSI-RS) may be configured in a window), and therefore, the CSI matrix collected based on the reference signal may exceed the available dedicated memory buffer.

[0179] In some embodiments, a node may declare a data buffering capacity, where the data buffering capacity may, for example, be related to the size of a training data set constructed from the collected training data. Depending on the implementation details, this may reduce or prevent the problem of exceeding the capacity of the memory buffer for new training data. For example, a node may declare or be assigned a predetermined memory buffer capacity based on: (1) a time gap (e.g., a maximum time gap) for obtaining training data and / or updating a model based on the obtained training data; (2) a maximum number of reference signals (e.g., CSI-RS) within a time window that the node expects to use to construct a training set; or (3) a short period of a reference signal (e.g., CSI-RS) used to construct a training data set. A situation in which a node may be configured with one or more reference signals and / or time windows that may violate the predetermined memory buffer capacity may be considered an error condition.

[0180] Alternatively or additionally, a default behavior may be defined when a violation of a node's predetermined memory buffer capability occurs. For example, if the configuration of the reference signal and / or time window violates the node's memory buffer capability, the node may store and / or use only a subset of the collected training data to update the model. For example, if the UE reports a maximum of N_max CSI-RS within the window, and the gNB configures a larger number of N_(CSI-RS) CSI-RS within the window, the UE may only use N_max CSI-RS out of the N_(CSI-RS) CSI-RS to update the model. How the UE selects which CSI-RS to use may be determined based on the UE implementation and / or according to one or more configured and / or fixed rules (e.g., the UE may use the nearest N_max resources out of the N_(CSI-RS) resources).

[0181] In some embodiments, the size of the buffer for the collected training data may be based on the node implementation, e.g., not involved in the specification. For example, if the UE's training data buffer overflows, the UE may stop storing newly collected data (e.g., matrices) and continue to update the model using the data in the buffer. In some embodiments, once the model is updated, the UE may flush the buffer and then start collecting new training data again.

[0182] In some embodiments, the UE may use a shared buffer to store new training data. Examples of shared buffers may include one or more buffers that are already used to store other channels, such as a PDSCH buffer, a primary CCE LLR buffer, etc. In such embodiments, since the shared buffer space may already be fully or partially occupied based on other dedicated uses, the shared buffer space may be used based on availability. In some embodiments, the buffering of the collected training data may be based on node implementation.

[0183] Training framework using reference models

[0184] In some frameworks according to the present disclosure, a reference model may be established as Model A for Node A, and then Node B may use the reference model as Model A to train Model B (e.g., assuming that Node A will use the reference model as Model A for inference). Node A may then use the reference model as Model A without further training, or Node A may continue to train the reference model for use as Model A. In some embodiments, one or more Models A may be provided and / or specified for Node A, and Node B may train one or more Models B at Node B using Model A, which may be assumed to be one or more of the reference models for Node A. For example, Node B may train a first version of Model B using Model A, which is assumed to be a reference model specified for Model A. Node B may also train a second version of Model B, where Model A is assumed to be another reference model, and so on.

[0185] The reference model may be established, for example, through specifications, signaling (eg, RRC signaling from a base station to a UE after the UE makes an RRC connection), and the like.

[0186] In some embodiments, node B may inform node A which reference models it has selected to use to train different versions of model A. In the case where there is only one reference model available for model A, no communication may be involved, as the reference model may be implicitly known. Node B may inform node A of one of a plurality of reference models that may be used to train versions of model B; the model may correspond to, for example, a reference model that provides the best performance. Alternatively or additionally, node B may inform node A of a subset of the reference models in the plurality of reference models; the subset may include a set of best performing reference models.

[0187] Regardless of any signaling from node B to node A, node A may or may not indicate to node B which reference model it has selected. Indicating a reference model may be used, for example, to establish consensus between node A and node B, while not indicating a reference model may reduce signaling overhead. In implementations utilizing multiple reference models, if the subset of best performing models includes only one reference model (e.g., only one reference model is indicated from node B to node A as the best performing reference model), node A may not provide an indication to node B, because node A's selection may be implicitly known to node B.

[0188] Once a reference model is established for node A, node A may use the reference model as model A or continue to train the reference model. Depending on implementation details, using the reference model as model A (e.g., with little or no further training or tuning) may provide a relatively high level of match (e.g., best match) between the two models because node B may train model B assuming that the reference model is for model A. If node B uses multiple reference models to train different versions of model B, node A may use a reference model corresponding to any of the trained versions of model B; this may involve establishing a consensus between node A and node B as to which trained version of model B is to be used (e.g., node B may communicate to node A which trained version of model B is to be used, or node A may inform node B which model to use).

[0189] Rather than using the reference model as Model A without further training, Node A may continue to train Model A. This may be beneficial, for example, if the reference model is not appropriate for the current network state (e.g., the wireless environment if the model is to be used for CSI compression and decompression). Therefore, having Node A further train (e.g., adjust or optimize) Model A may enable the model to match the current network state. However, when training Model B, changing Model A from the reference model assumed by Node B may result in a potential mismatch between the two models, which in turn may result in performance degradation.

[0190] In some embodiments, model A may be trained to overcome this potential mismatch. For example, to train model A, node B may send model B to node A, so the training of model A may be based on the actual model used by node B as model B.

[0191] If there are multiple trained versions of model B, node B may transmit a subset of the trained versions of model B, and node A may train multiple corresponding models A for the transmitted versions of model B. In such an embodiment, model A and model B may communicate to establish a consensus on which pair of model A and model B may be selected for use. Depending on implementation details, sharing multiple versions of model A may enable node A and / or node B to improve (e.g., optimize) performance by selecting the best pair of model A and model B among the transmitted models that may perform best. Alternatively, to reduce communication overhead, node B may transmit one of the multiple versions of model B, and node A may train model A corresponding to the transmitted version of model B.

[0192] Alternatively or additionally, if node A continues to train model A, node A may train a trial version of model B to mimic the actual model B used by node B. The degree of similarity between the trial model B and the actual model B may depend on the design and / or architecture of model B, the training data set used to train the trial version of model B, and / or the training process (e.g., initialization of weights, hyperparameters, etc.) used to train the trial version of model B. If there are multiple trained versions of model B that have been trained by node B, node A may train multiple corresponding trial versions of model B. Optionally, node A may train multiple models A using a trial version of model B corresponding to each of the available reference models for model A; this may be very useful because it may enable node A to train model A before being informed by node B which reference model or models node B has selected. In such an embodiment, model A and model B may communicate to establish a consensus on which pair of model A and model B may be selected for use.

[0193] To further reduce the mismatch between the trial model B and the actual model B, node B may share some auxiliary information with node A. Depending on the implementation details, sharing the auxiliary information may help node A train the trial model B in a way that will produce a trial model B similar to the actual model B. Examples of auxiliary information may include initialization values ​​(e.g., random seeds used by node B to train the actual model B, initial network weights, etc.), one or more optimization algorithms, one or more algorithms for feature selection, one or more algorithms for data preprocessing, information about the type of neural network (e.g., recurrent neural network (RNN), convolutional neural network (CNN), etc.), information about the structure of the model (e.g., the number of layers, the number of nodes per layer, etc.), information about the training data set, etc. Use of this information may be enforced (e.g., via a specification) or may be left to the implementation of the node.

[0194] In some embodiments, reference models for Node A and / or Node B may be specified (e.g., in a specification), e.g., for testing purposes. Such embodiments may not involve any indication of which model Node A and / or Node B uses. For example, when a gNB uses one or more reference models, a UE may be expected to meet one or more performance specifications. Depending on the implementation details, this may provide guidance on which models a node will use to achieve suitable performance for a machine learning task. In some embodiments, for example, as part of a specification, one or more performance requirements may be established for a machine learning CSI compression task.

[0195] In some embodiments of a framework utilizing a reference model, a node may use corresponding quantizer functions and / or dequantizer functions (e.g., approximate and / or differentiable quantizer functions and / or dequantizer functions) to train any model including the reference model, and any model may also use corresponding quantizer functions and / or dequantizer functions for further training, verification, testing, inference, etc.

[0196] Training framework using the latest shared values

[0197] In some frameworks according to the present disclosure, node A may start with model A, which may be in any initial state (e.g., pre-trained (e.g., offline training), untrained but configured with initial values, etc.). Node B may start with model B, which may also be in any initial state. One or both nodes may train their respective models for a period of time (which may be referred to as a training cycle or iteration), and then one or both nodes may or may not share the trained model values ​​and / or trained models with other nodes. In some embodiments, a new training data set may be provided to one or both nodes, directly or indirectly, for example, at the beginning or end of a cycle. Node A and node B may train their respective models, for example, without any model exchange, using their latest knowledge of the weights of the model at the other node.

[0198] The first node may train its model (e.g., the UE may train the encoder) using the latest weights (e.g., fed back by the second node) assuming that the model at the second node (e.g., the decoder at the base station) is frozen. The first node may train its model and update its weights, such as the maximum number of times (e.g., K_e times for the encoder), and then share the updated model weights with the second node. The same process may be implemented at the second node. Specifically, once the second node has received the updated model weights from the first node, the second node may train its model and update its weights the maximum number of times (e.g., K_d times for the decoder) assuming that the model weights of the model at the first node are frozen in the latest state of the model shared by the first node. The second node may then share its updated model weights with the first node. Therefore, the first node and / or the second node may have trained their respective models the maximum number of times, and then shared the updated model values ​​with another node (this may be referred to as a sharing cycle or iteration).

[0199] In a variation of the framework, after one or more of the nodes share model state information (e.g., weights) with another node, for example, at the end of a sharing cycle, one or two of the nodes may start another sharing cycle. For example, two nodes may train their models assuming that the value of the model on the other node is frozen as the latest value shared by the other node. At a specific point in time, or after a specific number of training cycles have been executed by the first node and / or the second node (e.g., at the end of another sharing cycle), one or two nodes may stop training and share their latest trained models with each other node. In some embodiments, at the beginning, a shared model (e.g., a fully shared model that may be initialized, for example, by offline training, handshakes, etc.) may be used for the initial values ​​of the latest shared weights.

[0200] Fig.10An example embodiment of a method for training a model using the latest shared value according to the present disclosure is shown. For illustration purposes, the following may be described in the context of a UE having an encoding model for CSI and a base station having a decoding model for CSI. Fig.10 The method is shown in , but the principle can be applied to any type of node and / or physical layer information.

[0201] Reference Fig.10 , at the beginning of the first sharing loop 1035-1, as shown in sharing point 1036-0, the encoder model may be in an initial state e_0, and the decoder model may be in an initial state d_0. The encoder model and the decoder model in the initial state (e_0, ​​d_0) may both be provided to the UE and the base station. Therefore, both the UE and the base station start with the encoder model and the decoder model in the same initial state. The UE may then perform M training loops (e.g., train its encoder M times while its decoder model remains in the initial state d_0). While the UE is performing M training loops, the base station may perform N training loops (e.g., train its encoder N times while its encoder model remains in the initial state e_0).

[0202] For example, after the first training cycle of the UE, the encoder model and decoder model of the UE may have a state (e_1, d_0), after the second training cycle of the UE, the encoder model and decoder model of the UE may have a state (e_2, d_0), and so on, until after the Mth training cycle, the encoder model and decoder model of the UE may have a state (e_M, d_0).

[0203] Similarly, after the first training cycle of the base station, the encoder model and the decoder model of the base station may have a state (e_0, ​​d_1), after the second training cycle of the base station, the encoder model and the decoder model of the base station may have a state (e_0, ​​d_2), and so on, until after the Nth training cycle, the encoder model and the decoder model of the base station may have a state (e_0, ​​d_N).

[0204] At the sharing point 1036-1 at the end of the sharing cycle 1035-1, the UE may send its trained encoder model to the base station, and the base station may send its trained decoder model to the UE. Thus, both the UE and the base station may have an encoder model and a decoder model with a state of (e_M, d_N).

[0205] In some embodiments, the UE and / or base station may stop training at this point and begin inference using its trained encoder model and decoder model. However, in some other embodiments, one or both of the UE and / or base station may begin another shared loop 1035-2. For example, the UE may then perform P training loops by training its encoder P times while its decoder model remains in state d_N, and the base station may perform Q training loops by training its decoder Q times while its encoder model remains in state e_M.

[0206] At the sharing point 1036-2 at the end of the sharing cycle 1035-2, the UE may send its trained encoder model to the base station, and the base station may send its trained decoder model to the UE. Thus, both the UE and the base station may have an encoder model and a decoder model with a state of (e_P, d_Q). The UE and / or the base station may perform any number of sharing cycles and any number of training cycles for each sharing cycle.

[0207] Fig.10 Special cases of the embodiments shown in are when M or N=0, if M>>N, or if N>>M. For example, in the case where N=0 and M>0, the base station may not update the decoder model during the sharing cycle (e.g., no training cycle may be performed). However, the UE may train its encoder M times before sharing its encoder with the base station. Similarly, in the case where M=0 and N>0, the UE may not update its encoder model, and the base station may update its decoder model N times before sharing its decoder model with the UE. Depending on the implementation details, one or more of these special cases may be beneficial in a case where, for example, one of the nodes has difficulty or is unable to obtain a training data set for online training at the node. In this case, the node that has access to (or has easier access to) the training data may continue online training, which may enable the node that continues training to provide trained models to other nodes that do not have access to the training data or have limited access to the training data.

[0208] Special cases may also be performed in an interleaved and / or alternating manner. For example, two nodes may start with one of the variables M or N equal to zero and the other variable greater than zero. Once the model with an applicable model is updated the number of times determined by the non-zero variable and shared with other nodes, the non-zero variable may take the value of zero and the other variable becomes non-zero. The process may continue with M and N alternatingly taking the value of zero. Such an interleaved training process may allow a first node (e.g., UE or gNB) to train its model (e.g., encoder or decoder) multiple times while the model at the second node is frozen. Then, after the first node shares its trained model with the second node, the second node may train its model multiple times while the model of the first node is frozen, and so on.

[0209] In some embodiments, the values ​​of the non-zero variables may affect the performance of the trained model pair. For example, if the time between sharing points is relatively large, the trained model (e_M, d_N) may have relatively poor performance, for example, if each node has been trained under the assumption of model weights on the other side that may be significantly different from the model that will be shared at the next sharing point. For example, in Fig.10 In the embodiment shown in , the base station can train its decoder assuming an encoder with weights e_0, and can later pair the trained decoder with a new encoder model e_M that may have changed significantly from e_0. Therefore, in some embodiments, sharing models at a relatively high frequency can improve the performance of the trained models.

[0210] In any framework of the framework disclosed in this document, a node can use corresponding quantizer functions and / or dequantizer functions (e.g., approximate and / or differentiable quantizer functions and / or dequantizer functions) to train any model including a reference model, and any model can also use corresponding quantizer functions and / or dequantizer functions for further training, verification, testing, inference, etc.

[0211] Using any of the frameworks disclosed herein, one or more of the nodes may transmit the collected training data and / or data sets (e.g., channel estimates, precoding matrices, etc.) to another device that can train one or more of the models. For example, the UE and / or the base station may upload the collected online training data and / or one or more models to a server (e.g., a cloud-based server), wherein the server may train one or more of the models using the uploaded training data, and download the one or more trained models to the UE and / or the base station.

[0212] Any framework in the framework disclosed herein may be modified so that a first type of node may train a model for another type of node and share the trained model with multiple instances of the second type of node. For example, a base station may train an encoder for its decoder and share the trained encoder with multiple UEs. One or more of the UEs may apply a shared encoder to compress the CSI at the UE and / or use the shared encoder for further online training. In addition, any framework in the framework disclosed herein may be implemented in a system where one of the nodes is not a base station (e.g., two UEs or other peer devices configured for sidelink communication). In such an implementation, a UE may train a decoder for its encoder and share the trained decoder with one or more other UEs, wherein the one or more other UEs may use the trained decoder for direct inference and / or use the trained decoder as a source of initial values ​​(e.g., weights) for further online training.

[0213] Model sharing mechanism

[0214] In some embodiments, the node may transmit the model, weights, etc. using any type of communication mechanism, such as one or more uplink and / or downlink channels, signals, etc. For example, when the sharing of the encoder model is triggered, the UE may use one or more MAC control elements (MAC CE) PUSCH to send the encoder model and / or weights to the gNB. Similarly, the gNB may use one or more MAC CE PDSCH to send the decoder model and / or weights to one or more UEs.

[0215] Depending on implementation details, sharing the full set of weights may be inefficient because the model may be relatively large and may consume a relatively large amount of downlink and / or uplink resources for sharing.

[0216] Some embodiments may establish one or more sets of quantized models that may be referred to as a model book. When a model is trained by a node, if sharing is requested, the node may map the model to one of the quantized models in the model book. One or more model books in the model book may be shared among nodes. Instead of sending the model, a node may send an index of the mapped model in the model book. Depending on the implementation details, this may reduce the communication resources associated with model sharing.

[0217] In some embodiments, once the set of parameters for the model is known, the final result of the training can be known with certainty. For example, given (1) a training set, (2) an initial random seed that determines the initial weights, (3) optimizer parameters (e.g., fully defined optimizer parameters), and / or a training process, the trained model at the end of a particular number of training steps (e.g., a training loop) can be uniquely determined. These parameters can be referred to as, for example, minimum description parameters. If the size of the minimum description is less than the size of the weights for the model, then nodes can share the minimum description parameters instead of the weights. Depending on the implementation details, this can reduce the communication overhead associated with sharing the model.

[0218] In some embodiments, one or more values ​​of the model (e.g., weights of a CSI encoding and / or decision model at a node) may be arranged in a vector W (e.g., a vector of weight elements). A dedicated compression autoencoder model (e.g., an encoder and decoder model pair) may be trained to compress W using an encoder at one node and a decoder at another node. If sharing of a CSI model is triggered and / or requested, the node may construct a vector W of the CSI model, encode it using a model compression encoder, and send the encoded vector to another node. The other node may use a model compression decoder to recover the weight vector W. Depending on the implementation details, this may reduce the communication overhead associated with sharing the model.

[0219] Online training processing time

[0220] In embodiments where a node may perform online training of a model, a resource grant (e.g., a grant of processing time, processing resources, etc.) may be provided to the node to perform training. Such a grant may be provided to train using an online training data set that may be collected by the node (e.g., a channel estimate based on measurements performed by the node) or an online training data set that may be RRC configured (or reconfigured) or MAC-CE activated. The resource processing time grant may ensure that the node may have sufficient time to update the model using the online training data set before the node is expected to have completed the update (e.g., to share the updated model with another node). However, in some embodiments, a processing time grant may be provided to the node regardless of whether the node is expected to share the trained model after processing.

[0221] For example, in embodiments where the UE may collect an online training data set by computing a channel estimate, the UE may be provided an amount of time determined by N_(AIML,upadte) symbols from the end of the last symbol of the latest CSI-RS for the online training set (e.g., to update the encoder model). If the UE is configured to report updated models to another node (e.g., a gNB), the UE may not be expected to report the model to the gNB earlier than N_(AIML,report) symbols from the last symbol of the latest CSI-RS in the training set.

[0222] As another example, in an embodiment where the UE may perform online training of the encoder using an online training data set for which the UE is RRC configured (or reconfigured) or MAC-CE activated, the UE may not be expected to update and / or report its encoder earlier than N symbols from the latest symbol in which the corresponding RRC (re)configuration was completed or a MAC-CE activation command had been received.

[0223] Preprocessing based on domain knowledge

[0224] For compression purposes, a machine learning encoder may receive an input signal and generate a set of output features that may be sufficient for a decoder to use to reconstruct the input signal. In the case of maximum compression, the output features may be expected to be independent of each other, otherwise they may be further compressed.

[0225] Although a pair of machine learning models may be able to generate a feature vector from an input and reconstruct the input from the feature vector, in some embodiments according to the present disclosure, one or more pre-processing and / or post-processing operations may be performed on the input to the generated model and / or the output from the reconstructed model. Depending on the implementation details, this may provide one or more potential benefits, such as reducing the processing burden and / or memory usage of one or both of the models, improving the accuracy and / or efficiency of one or both of the models, etc.

[0226] In some embodiments, pre-processing and / or post-processing may be based on domain knowledge of the input signal. In some embodiments, pre-processing and / or post-processing may provide auxiliary information from the domain knowledge to the encoder, which, depending on the implementation details, may reduce the processing burden on the encoder. For example, if the vector to be compressed by the encoder can be characterized as a low-pass signal with relatively small changes, a discrete Fourier transform (DFT) and / or inverse DFT (IDFT) may be performed to analyze the frequency domain representation of the vector. If the DC component of the DFT vector is larger (e.g., significantly larger) than the other components, this may indicate that the signal has a low change, and therefore the signal may be pre-processed before being compressed by the encoder (and post-processed after being decompressed by the decoder) to reduce the burden on the encoder / decoder pair.

[0227] In some embodiments, performing a transform and / or inverse transform such as a DFT and / or IDFT can provide the machine learning model with a clearer understanding of the level of correlation between elements of the input vector. For example, in some embodiments (e.g., using any of the frameworks disclosed herein), the CSI matrix may be input to a preprocessor, wherein the preprocessor may, for example, apply a transform (e.g., DFT / IDFT, discrete cosine transform (DCT) / inverse DCT (IDCT), etc.) to all or part of the input on different CSI-RS ports. The transformed signal may then be input to the encoder and compressed. On the decoder side, the output of the decoder may be applied to an inverse operator of the preprocessor transform (the inverse operator of the preprocessor transform may, for example, be implemented using a postprocessor) to generate a reconstructed input signal.

[0228] Fig.11 An example embodiment of a dual model training scheme utilizing pre-processing and post-processing according to the present disclosure is shown. In some aspects, Fig.11 The embodiment 1100 shown in FIG. 1 may be similar to Figure 3 , and similar components may be identified by reference designators ending with the same number. Fig.11The embodiment shown in 1104 may include a pre-processor 1137 and a post-processor 1138. The pre-processor 1137 may apply any type of transformation to the training data 1111 before applying it to the generation model 1103. Similarly, the post-processor 1138 may apply any type of inverse transformation (e.g., the inverse of the transformation applied by the pre-processor 1137) to the output of the reconstruction model 1104 to generate the final reconstructed training data 1112.

[0229] In some embodiments, as shown by solid lines 1139 and 1140, a loss function 1113 for training model 1103 and model 1104 may be defined between the input of generation model 1103 and the output of reconstruction model 1104. However, in some embodiments, as shown by dashed lines 1141 and 1142, a loss function 1113 may be defined between the input of preprocessor 1137 and the output of postprocessor 1138. Fig.11 As shown in FIG. 1 , models 1103 and 1104 are trained and can be used for inference.

[0230] Although the principles related to pre-processing and / or post-processing are not limited to any particular implementation details, for the purpose of illustrating the principles of the present invention, an example embodiment of a scheme for pre-processing and post-processing a CSI matrix based on domain knowledge may be implemented as follows. In the case of a channel matrix of size N_rx×N_tx, for each pair of RX and TX antennas (i, j), the channel elements corresponding to the pairs of all resource elements (REs) used within the time and frequency window may be concatenated to obtain a combined matrix H_(i, j) of size M×N, where M and N may be the number of subcarriers and orthogonal frequency division multiplexing (OFDM) symbols of the CSI-RS in the window. In some embodiments, it may be assumed that the matrix is ​​a complex number. In an example embodiment of the pre-processing scheme, H_(i, j) may be transformed, for example, using a DFT matrix. If U_freq and U_time are M×M and N×N DFT matrices, respectively, the matrix H_(i, j) may be transformed into X_(i, j) as follows:

[0231] X_(i,j)=U_freq^*H_(i,j)U_time (Equation 2)

[0232] X_(i,j) can be called the Delay-Doppler Representation (DDR) of H_(i,j). The matrix H can be reconstructed from the DDR as follows i,j :

[0233] H_(i,j)=U_freq X_(i,j)U_time^* (Equation 3)

[0234] In some embodiments, the use of the DDR transform may produce a sparse matrix X, which in turn may reduce learning and inference complexity.

[0235] In some embodiments, the use of pre-processing and / or post-processing transforms may enable the original training set to be transformed to the corresponding DDR matrix. In such embodiments, CSI compression may then compress the transformed training set. Thus, pre-processing (e.g., DDR transformation) may be performed on the UE side, while post-processing (e.g., inverse of DDR to recover H) may be performed on the gNB side.

[0236] In some embodiments, the transformed matrix (e.g., Fig.11 As shown in , a loss function is defined between the transformed matrix input to the encoder and the transformed matrix output of the decoder).

[0237] In some embodiments, the matrix H may be constructed based on the union of the individual CSI matrices in the time domain and / or frequency domain for each spatial channel (e.g., for each transmit antenna (port) and each receive antenna (port) pair). One or more models may be trained and tested for each spatial channel. In some embodiments, H may be constructed based on the channel matrices of the REs, for example, where each matrix may have a size of N_r×N_t, where N_r and N_t may be the number of receive antennas at the UE and the number of transmit antennas at the gNB, respectively.

[0238] CSI Matrix Formulation

[0239] In some embodiments, the CSI information of a RE or a group of REs that can be compressed by the UE may be referred to as a CSI matrix. From the analysis of multiple-input multiple-output (MIMO) channels, the capability distribution may be a Gaussian distribution with possibly different power allocations across transmit antennas. If the channel matrix is ​​decomposed into H_(r×t)=UΣV^H, the capability distribution can be obtained by first setting x~=Vx, where x is an iid Gaussian random vector with zero mean and unit variance, and then multiplying each element of the following term by the power allocation given by the water filling algorithm:

[0240]

[0241] The power allocation for the i-th channel may be P_i, i=1,…,t, where P_i may be obtained from the singular value decomposition of the channel matrix H. Thus, the information used at the gNB (e.g., the complete information required at the gNB) may be both the right singular value matrix V and the singular values ​​themselves. Thus, in some embodiments, the CSI matrix may be formulated (e.g., defined) in any one or more of the following terms: (a) the CSI matrix may be formulated as the channel matrix H; (b) the CSI matrix may be formulated as the concatenation of the singular values ​​and V; and / or (c) the CSI matrix may be formulated as the matrix U.

[0242] In some embodiments, the UE may be configured to report any of the above CSI matrices via RRC (re) configuration, MAC-CE command, or dynamically via DCI. In embodiments where the CSI matrix is ​​formulated according to (b) (e.g., concatenation of singular values ​​and V), the UE may also be configured to report only singular values.

[0243] For the purpose of model training, when the UE is configured to report a specific CSI matrix, the training set and / or loss function may be formulated based on the applicable CSI matrix. For example, when the UE is configured to report V, the training set may include a V matrix obtained from the estimated channel matrix, and the loss may be formulated based on the V matrix input to the encoder and the reconstructed V matrix at the output of the decoder.

[0244] Node Capacity

[0245] Implementation of any of the frameworks disclosed herein may involve the use of resources (such as memory resources, processing resources, and / or communication resources), for example, to store new training data, share models between nodes, and train and / or apply a particular type of neural network architecture, such as a CNN or RNN for the model. Different nodes, such as UEs, may have different capabilities for implementing neural networks. For example, a UE may or may be able to support CNNs but not RNNs. In some embodiments, a UE may report its ability to support a particular type of neural network architecture (e.g., a network type such as CNN, RNN, etc.), and / or any other aspect reflecting its limitations and capabilities for applying an encoder model.

[0246] In some embodiments, a node (e.g., a UE) may report its capabilities and / or limitations using a list that may include any number of the following items: (a) one or more network types, e.g., CNN, RNN, a specific type of RNN, a gated recurrent unit (GRU), a long short-term memory (LSTM), a transformer, etc.; (b) one or more aspects related to the size of the model, e.g., the number of layers, the number of input and / or output channels of a CNN, the number of hidden states of an RNN, etc.; and / or (c) any other type of structural limitation.

[0247] Depending on its reported capabilities, a node (e.g., UE) may not be expected to train or test an encoder model that violates any of the node's reported constraints and / or that requires capabilities beyond the node's declared capabilities. In some embodiments, this may be ensured regardless of the applicable framework and / or location of the training / inference of the model. For example, if the framework is implemented so that the gNB can pre-train encoders and decoders and share the encoders and / or decoders with the UE, the UE may not expect the encoder model to violate its capabilities. As another example, if one or more pairs of encoders and decoders are trained offline and specified in the applicable specification (e.g., NR specification), the UE may not expect the applicable model to violate its capabilities. In some embodiments, the UE may report capabilities by signaling to activate one or more models, and may declare which specific encoder / decoder pairs, or individual encoders or decoders, it can support. The gNB may then indicate to the UE which encoder / decoder pairs are applicable to the UE. The indication may be provided, for example, by system information, RRC configuration, dynamic signaling in DCI, etc.

[0248] Adjustment through online training

[0249] In some frameworks, a node such as a UE may be expected to train its encoder model or both encoder and decoder models online (e.g., by collecting new training data (e.g., samples) on the fly or based on offline provisioning and updating of one or more models). If the node only updates the encoder model, then since the loss function may also depend on the decoder weights, the encoder model adjustment and / or optimization may also depend on the decoder weights and / or model. In such an implementation, the node may declare the ability to handle the limitations of the decoder model even if the encoder may be used on the gNB side. One or more such restrictions may be applied as follows: (1) one or more online training features and / or fine-tuning may be declared by a node as capabilities; (2) a node that reports the ability to support online training may further report restrictions on supported structures for encoder models, wherein the restrictions may be, for example, any of the restrictions mentioned above as (a) to (c); (3) a node that reports the ability to support online training may further report restrictions on supported structures for decoder models, wherein the restrictions may be, for example, any of the restrictions mentioned above as (a) to (c); and / or (4) if multiple models including encoder and decoder pairs are specified in the specification, the node may declare capabilities indicating which encoder and decoder pairs or which individual encoders and / or decoders the node can support.

[0250] Multiple pairs of models

[0251] In some embodiments, multiple pairs of models (e.g., encoder / decoder pairs) may be trained and / or deployed for operation (e.g., simultaneous operation) on two nodes (e.g., a UE and a gNB). The multiple pairs may differ from each other in terms of a) both encoder and decoder, b) only encoder, or c) only decoder. In some embodiments, multiple pairs of models may be configured (e.g., optimized) to handle different situations that may be specified to handle different channel environments, which in turn may result in different distributions of training data sets and / or test data sets.

[0252] Multiple pairs of models may be used, for example, to accommodate different dimensions in the training data. For example, the dimensions of the CSI matrix may be determined based on the number of CSI-RS ports. In some embodiments, if the UE is to report a first CSI matrix H_1 and a second matrix H_2 of different dimensions, a single encoder and decoder pair may be used to process matrices of different sizes. In such an embodiment, the encoder and decoder may be trained as follows. The matrix input to the encoder may be reformatted to have a fixed size by appending zeros in a configuration that may be agreed upon between the UE and the gNB. Thus, the training set may initially include matrices of different sizes, wherein the matrices of different sizes may be modified by appending zeros as described above to convert the matrices to a fixed matrix size. The communication mechanism may be implemented to enable the gNB and UE to share the same understanding of the size of the CSI matrix requested, for example, in the UE's report. Depending on the implementation details, the matrix resizing technique may be applicable to any matrix size.

[0253] Alternatively or additionally, multiple pairs of models (e.g., encoder / decoder pairs) may be trained, where different model pairs may be configured to handle different CSI matrix sizes.

[0254] In some embodiments, and depending on implementation details, multiple pair models may be implemented without increasing complexity. For example, in the case of multiple pair models, if the CSI report includes a CSI matrix corresponding to a specific number of CSI-RS ports, the reference time for calculating the CSI report may be less in the case of multiple pairs than in the case of a single pair. In addition, if each RRC configuration or MAC-CE activation includes a CSI report for a specific case corresponding to a specific pair, the UE may load the applicable model into the modem while keeping one or more of the other models in the UE controller. Depending on implementation details, this may reduce modem internal memory usage. In embodiments with multiple pairs, different pairs may be classified according to any of the following configurations. (a) Each pair model may be configured to handle a specific CSI matrix size. For example, a pair model may receive a CSI matrix estimated based on a CSI-RS with a specific number of ports and also associated with a specific number of receive antennas of the UE. The UE may report its number of receive antennas to the gNB in ​​one report or report its number of receive antennas to the gNB separately for different numbers of CSI-RS ports. (b) Each pair of models can be configured to process different distributions of the training dataset and / or the test dataset. (c) Each pair of models can be configured to process different channel environments for the training dataset and / or the test dataset.

[0255] Training set association and model pair configuration

[0256] In embodiments where a model pair may be configured to handle different cases, a node (e.g., a UE or a base station) may be configured with different training data sets (e.g., different training data sets for specific cases or pairs of models (e.g., encoder / decoder pairs)). Thus, the UE and / or base station may have different training data sets, e.g., one for each of the different pairs. Once a trigger occurs for a node (e.g., a UE or a gNB), the node may also be signaled as to which pair of models should be trained. For example, in the case of online training, the gNB may indicate to the UE to start training a specific pair of models. If online training is performed on the fly by collecting a new data set, an association may be provided, for example, between the CSI-RS and the encoder / decoder pair, e.g., via the number of CSI-RS ports.

[0257] Once multiple pairs of models have been trained and are ready to be deployed in the inference period, a node (e.g., a UE) may need to know which pair to use to encode the channel matrix. For example, each pair may be associated with a specific dimension for the CSI matrix for encoding. The dimension may be referred to as the input dimension of the encoder model. In some embodiments, the UE may determine the model pair used to encode the CSI matrix as follows. The CSI-RS may be implicitly or explicitly associated with a pair of models. The UE uses the model pair associated with the CSI-RS to encode the CSI matrix. In the case of implicit association, the CSI-RS may be mapped to a specific pair based on the number of CSI-RS ports and / or the number of receive antennas at the UE. Therefore, if the dimension of the CSI matrix obtained from the CSI RS is equal to the input dimension of a pair, the CSI-RS may be mapped to the pair. If multiple pairs have the same qualified input dimension, the reference pair may be selected, for example, based on a rule that may be established between the UE and the gNB. In the case of explicit association, the CSI-RS reporting the CSI matrix may be configured via RRC or dynamically indicated in the DCI (e.g., using a pair of indices).

[0258] In any of the above implementations, if the UE is signaled to report the CSI matrix via a pair of models with different input dimensions, the UE may append zeros to match the size of the matrix to the input dimensions. However, the UE may not expect to be signaled to report the CSI matrix using a pair of models with input dimensions that are smaller than the dimensions of the CSI matrix.

[0259] Compression with reduced model size

[0260] In some embodiments, a pair of models may be configured as an autoencoder to compress the CSI matrix and / or exploit redundancy and / or correlation between CSI matrix elements. If the CSI matrix is ​​reported for each RE, the correlation may only be the spatial correlation between different paths between different pairs of transmit antennas (e.g., CSI-RS ports) and receive antennas. Depending on the implementation details, the amount of such correlation may be limited, and thus the autoencoder may not be able to adequately compress the CSI matrix.

[0261] In some embodiments, the compression capability of the autoencoder may be related to the redundancy and / or correlation (which may be referred to as spatial correlation) between the elements of the CSI matrix. Since the wireless channels may also be correlated in the time domain and / or frequency domain, there may also be time and / or frequency correlation. Therefore, the estimated channels for multiple OFDM symbols and / or multiple resource elements (REs), resource blocks (RBs), or subbands may be input as a single training sample. For example, a channel matrix corresponding to multiple REs may be specified as an input to the autoencoder. In one such method according to the present disclosure, the UE may be configured via RRC using such a configuration that may specify a time and / or frequency resource bundle (bundling) for forming a training data set.

[0262] Depending on the implementation details, the compression performance of the autoencoder can be improved by compressing CSI for multiple REs across different frequency and / or time resources. Therefore, a combined CSI matrix of multiple REs can be input in a time and frequency window. The combined CSI matrix can then be obtained by cascading the individual CSI matrices of the REs in the window. Depending on the implementation details, due to the time and frequency flatness of the channel, the combined CSI matrix may be more likely to have significant correlations between its elements. Therefore, if the model takes the combined CSI matrix as input, the model can compress the combined CSI matrix to a higher degree than multiple models acting on each per RE matrix. In some embodiments, the UE can be configured using a time and / or frequency window and one or more configurations that can indicate which REs the UE can use to determine the combined CSI matrix. Such a configuration can be used in both the training period and / or the test period to obtain the combined matrix.

[0263] Input size is reduced via subsets of the CSI matrix

[0264] In some embodiments, the autoencoder may encode the CSI matrices of different REs in a specific time and frequency window. If the channel is such that the correlation between the elements of the channel matrix does not exist or is not strong in a specific domain (e.g., time or frequency), the set of elements in the union of the CSI matrices may be divided into subsets with relatively strong intra-subset element correlation and relatively weak inter-subset element correlation. For example, if the autoencoder is to compress four CSI matrices of four REs on the same OFDM symbol, the matrices may be represented as follows:

[0265]

[0266] If the correlation in the frequency domain is strong and there is little or no correlation in the spatial domain (i.e., between elements of a matrix), the autoencoder can be configured to compress vectors of length 4, and the autoencoder can be applied four times to the following subsets: subset 1 (a_1, b_1, c_1, d_1); subset 2 (a_2, b_2, c_2, d_2); subset 3 (a_3, b_3, c_3, d_3); and subset 4 (a_4, b_4, c_4, d_4).

[0267] The CSI matrix can then be reconstructed at the decoder by, for example, reconstructing the four vectors using the same decoder. As described above, subsets may be selected so that they can exploit one or more correlations in one or more domains. To further illustrate, in the example above, if there is a correlation between elements in the spatial domain, the subset selection described above prevents the network from exploiting the correlation to further compress the CSI matrix. Conversely, the following subset selections may allow the exploitation of correlations in both the frequency domain and the spatial domain: subset 1 (a_1, a_2, b_1, b_2); subset 2 (c_1, c_2, d_1, d_2); subset 3 (a_3, a_4, b_3, b_4); and subset 4 (c_3, c_4, d_3, d_4).

[0268] In some embodiments, the following framework may be used for a reduced model size with N_features input dimensions based on this approach. (1) The UE may be configured to report CSI matrices for M REs that may be on the same or different OFDM symbols and within a time and / or frequency window. Each CSI matrix may have N elements. (2) The UE may partition the M×N elements into (M×N) / N_features subsets. A common rule for subset selection may be established between the UE and the gNB. (3) An autoencoder (e.g., a single autoencoder) may be used to compress and recover the N_feature elements in each subset. In the example subsets above, M=N=4 and N_features=4.

[0269] Input size reduction via resource element selection

[0270] The size of the encoder network can be reduced by reducing the size of the combined input matrix. In some embodiments, the size of the combined matrix can be reduced by (a) removing specific elements of each per-RE matrix (for example, if there are two REs in a window of CSI matrices H_1 and H_2 with the same dimensions). The combined matrix can be constructed to have the same dimensions as H_1 or H_2, but by selectively selecting (i, j) elements from H_1 or H_2. Alternatively or additionally, the size of the combined matrix can be reduced by (b) constructing a matrix that excludes the CSI matrix for specific REs in the window.

[0271] These examples are shown in Table 1, which shows a window with two REs and two CSI matrices. Using method (a), the combined matrix can be constructed as shown in Table 1, while using method (b), the combined matrix can be constructed by selecting one of the two matrices.

[0272] Table 1

[0273]

[0274]

[0275] Number of CSI-RS ports and training set

[0276] The channel matrix estimated from a CSI-RS having N_port ports and reception via N_r receive antennas at the UE may have a dimension of N_r×M_port. An autoencoder may be used to remove redundancy from the matrix. In some embodiments, if the training set includes matrices of different dimensions (e.g., corresponding to CSI-RS with different numbers of ports), it may be more difficult to identify redundant forms and / or remove redundant forms. Therefore, in some embodiments, the training set for any of the methods disclosed herein may include only or primarily matrices of the same dimension and / or associated with the same number of CSI-RS ports. Therefore, the UE may not be expected to be configured with a training set or a CSI report and measurement configuration (config), which results in the training set having dimensions different from the dimensions of the training data set matrix.

[0277] UCI Format

[0278] Using any of the frameworks disclosed herein, the output of a generation model (e.g., an ML encoder) may be considered a type of UCI (which may be referred to as, for example, artificial intelligence machine learning (AIML) CSI). In some embodiments, the AIML CSI may be obtained from a CSI report and a measurement configuration with associated CSI-RS resources and reporting settings. In some embodiments, the AIML CSI may be sent to the gNB via PUCCH or PUSCH (e.g., following Rel-15 behavior). Thus, the format of the representation of feedback information for physical layer information may be established as a type of uplink UCI. The format may involve one or more types of coding (e.g., polarization coding, low-density parity check (LDPC) coding, etc.), which may depend on, for example, the type of physical channel used to send the UCI. In some embodiments, the type of uplink UCI (e.g., AIML CSI) sent using PUCCH may use polarization coding, while sending using PUSCH may use LDPC coding. In addition, before encoding, the CSI may be quantized. Thus, the AIML CSI may be quantized into a bit stream (0 and 1) and input to a polarization encoder or an LDPC encoder.

[0279] Adaptability to different network providers

[0280] When a UE connects to a network, it may not know which network vendor created the network it is connected to. Since different vendors may utilize different training techniques and / or network architectures for machine learning models, the availability of this information may affect the training model on the UE side. Therefore, in some embodiments, a network indication or AI / MI index may be provided to the UE via system information (e.g., via one of the SIBs). The UE can then use this information to adapt its training to a specific network vendor configuration.

[0281] ML Model Lifecycle Management

[0282] In some ML applications, the performance of the ML model may degrade over time and may not perform adequately for the duration of the application it was trained for. Therefore, the ML model may be frequently updated to adapt to temporal changes that may occur in its operating environment, e.g., statistical changes in the wireless channel in the case of CSI compression.

[0283] According to some embodiments of the present disclosure, a management framework may be provided to enable efficient and / or timely updating of one or more ML models with acceptable overhead. To facilitate such a framework, one embodiment may implement model monitoring in which nodes can keep track of the performance of ML models. In some embodiments, this may involve model monitoring in which nodes can keep track of the performance of ML models. Model monitoring can be based on one or more performance metrics as follows. (1) Task-based metrics can be used to (e.g., directly) evaluate the performance of tasks performed by ML models. For example, these metrics may include accuracy, mean square error (MSE) performance, etc. (2) System-based metrics can be used to track the overall performance of the system, such as correct decoding of transmissions, or other system-level key performance indicators (KPIs) that can provide less direct measurements of the performance of ML models used by nodes in the system.

[0284] The management framework may initiate an ML model update process when the performance of the ML model is deemed unacceptable based on the negotiated and / or configured metrics. For example, (1) when the ML model performance is unacceptable based on one or more negotiated and / or configured metrics; and / or (2) if the performance of the ML model is unacceptable for a particular duration greater than a threshold time, the performance may be deemed unacceptable.

[0285] The threshold for determining unacceptable performance may be implemented as a configured and / or specified parameter. Duration may be measured i) cumulatively (e.g., any duration of unacceptable performance may be added to a global counter and the global counter value may be compared to the threshold), or ii) continuously (e.g., only consecutive durations of unacceptable performance greater than the threshold may be considered).

[0286] When the performance of the ML model is deemed unacceptable, the management framework may trigger an update process that can be implemented in one of the following ways. (1) The management framework may require, for example, Figure 8 The management framework may require a complete training process as described. In this case, the ML model may be retrained from scratch, or the ML model may be retrained starting from the current ML model. Training in this case may use the entire training data set with or without additional data samples that may have been recently acquired. (2) The management framework may require partial training, where the ML model may be retrained starting from the current ML model and possibly using new data samples that may have been recently acquired.

[0287] Performance metrics during training and testing

[0288] In order to evaluate the performance of different models for the CSI compression task, some embodiments of the present disclosure may focus on aspects of CSI compression. In such an embodiment, different models may be compared based on their respective abilities to compress the CSI matrix and recover the CSI matrix so that the recovered matrix is ​​as close to the true CSI matrix as possible. The determination of the proximity may be related to the operation of the gNB utilizing the CSI matrix. For example, if the gNB calculates the SVD of the channel matrix as H_(r×t)=UΣV^H and uses the right singular vector V to determine the precoder, the proximity may be determined between V at the input of the encoder and the recovered V at the output of the decoder.

[0289] In some embodiments, a proximity metric between two matrices may be implemented element by element, and an average may be taken over some or all elements to provide a single loss value. Alternatively or additionally, having one or a few erroneous elements in a matrix may be as harmful as having many erroneous elements. In this case, a loss function may be determined matrix by matrix (e.g., the maximum value of the element-by-element error over all elements of the matrix).

[0290] In some embodiments, the performance of the CSI encoder and decoder models can also be evaluated in conjunction with other blocks of the system. For example, if the block error rate (BLER) is used as a system performance metric, the comparison between different CSI models can be based on their resulting BLER. Other system KPIs (such as throughput, resource utilization, etc.) can also be used for this purpose.

[0291] In embodiments where BLER is the metric of interest, configuring the gNB to use the information provided by the CSI matrix may impact system performance. For example, assuming that the CSI model is perfect in the sense that the channel matrix sent by the UE is fully recovered at the gNB and the channel matrix indicates a rank 1 channel, decoding may fail if the gNB schedules a rank 2 PDSCH. Therefore, in order to establish a connection between the compression capabilities of the CSI model and system performance, assumptions about the operation of the gNB may be used. In some embodiments, the gNB that processes the function f_gNB may be defined as taking the output of the decoder (e.g., H^) and providing an estimate of the resulting BLER as BLER = f_gNB(H^) (or BLER = f_gNB(H,H^)). The loss function during training may then be defined taking into account both aspects of CSI compression and gNB operation. For example, the loss may be defined as a weighted sum of two terms as follows:

[0292] loss(H,H^)=α.loss_CSI(H,H^)+β.BLER (Equation 6)

[0293] Among them, α and β are hyperparameters used for training.

[0294] Reliability aspects of the uplink channel

[0295] In some embodiments, it may be assumed that the output of the encoder (which may also be referred to as CSI codewords) is available error-free at the decoder side. Therefore, the CSI codewords may be sent via PUCCH or PUSCH on an uplink channel with infinite reliability so that PUCCH and / or PUSCH decoding does not fail. However, in some cases, during the inference period, for example, when PUSCH / PUCCH decoding fails, CSI codewords with one or more errors may be transmitted to the gNB (decoder). In this case, a noisy version of the CSI codeword may be available at the decoder. The impact of imperfections in the uplink channel during the training period may be modeled as follows.

[0296] For each training example input to the encoder, the CSI codeword at the output of the encoder can be denoted as x. Taking into account the imperfections of the uplink channel, the input y to the decoder can be modeled as:

[0297] y = x + ω (Equation 7)

[0298] Wherein, ω is the additive noise that can model the residual error after decoding of the uplink channel. The additive noise can be generated as follows during the training period.

[0299] In method 1, ω can be modeled as a Gaussian random vector with zero mean and variance σ^2. The variance can be indicated to the UE via RRC configuration or left to the UE implementation. In method 2, the channel between x and y can be modeled by performing PUCCH and / or PUSCH decoding on each training example, obtaining a residual error vector ω, and then assuming that this vector is added to x to obtain y.

[0300] Federated Learning

[0301] In the case of federated learning (FL), a global model at the server may be learned by individual learning at multiple nodes connected to the server and sharing the learned model with the server. The server may then perform one or more operations on the received model to obtain the final model. Such an arrangement may be motivated, for example, by privacy aspects of the nodes and / or a requirement not to share their data with the server.

[0302] In the case of CSI compression use case, the server may be considered as a gNB and different UEs connected to the gNB may be considered as model update nodes. Different UEs may have different training sets, where the training sets have the same or different distributions. If the distribution is the same, each UE may update the model using its own training set and share the model with the gNB. The gNB may then perform one or more operations, such as averaging the models to obtain a final model. The gNB may share the obtained final model with the UEs that share their models. The final model may be expected to be better than the individual received models because the final model is trained based on the union of all training sets on all participant UEs. Therefore, FL may be used to improve CSI compression performance. In the case where different distributions are available at different UEs, FL may help capture distributions that have not been seen by a particular UE through models shared by UEs that have seen the distribution. In any case, FL may be used to obtain a model that takes into account the different environments observed by the UEs.

[0303] In the case of the FL framework according to the present disclosure, the gNB may configure a group of UEs in a FL group. UEs in the same FL group may be configured to have the same encoder and / or decoder (e.g., autoencoder or AE) architecture. Therefore, their encoders and decoders may differ only in the actual trained weights, but have the same configuration in terms of the number of layers, number of units, activation functions, and other parameters defining the network structure.

[0304] For UEs in a group, the size of the inputs to the encoder model and the decoder model may be the same or similar. The inputs to the encoder may also have the same or similar meaning for the UEs. For example, the input to the encoder of a UE (e.g., all UEs) may be a channel matrix or a singular value matrix V. The gNB may instruct the UEs to update their models and share the updates with the gNB via RRC, DCI, or MAC CE commands. In some embodiments, not all UEs in a group participate in the update process at the same time. The gNB may send information about training, hyperparameters, and / or other aspects of the FL via a group common (GC) DCI, where UEs in the same FL group may have a specific portion of the DCI configured via RRC.

[0305] Additional Embodiments

[0306] Fig.12 An embodiment of a system using a dual model approach according to the present disclosure is shown. The system can be described in the context of one or more models in the test model. Fig.12 , but the same or similar embodiments can also be used to use any model in the model disclosed in this article (for example, after training, using Figure 3 The generated model 303 and / or reconstructed model 304 shown in are used for verification, inference, etc.

[0307] Reference Fig.12 , system 1200 may include a first node (Node 1) having a generative model 1203 and a second node (Node B) having a reconstruction model 1204. Test data 1211 may be applied to the generative model 1203, which may generate a representation 1207 of the test data. The reconstruction model 1204 may generate a reconstruction 1212 of the test data based on the representation 1207 of the test data. In some embodiments, the generative model 1203 may include a quantizer to convert the representation 1207 into a quantized form (e.g., a bitstream) that can be transmitted over a communication channel. Similarly, in some embodiments, the reconstruction model 1204 may include a dequantizer, wherein the dequantizer may convert the quantized representation 1207 (e.g., a bitstream) into a form that can be used to generate the reconstructed test data 1212.

[0308] The generative model 1203 and the reconstruction model 1204 may be obtained in any manner, including using any of the frameworks described herein. For example, using a joint training framework, the generative model 1203 and the reconstruction model 1204 may be trained as a pair at node A, and node A may send the reconstruction model 304 to node B. Other embodiments may use a training framework using a reference model, a training framework using the latest shared values, or any other framework and / or technique to obtain and / or train the generative model 1203 and the reconstruction model 1204.

[0309] Fig.13 An example embodiment of a user equipment (UE) according to the present disclosure is shown. Fig.13 The embodiment 1300 shown in 1300 may include a radio transceiver 1302 and a controller 1304, which may control the operation of the transceiver 1302 and / or any other components in the UE 1300. The UE 1300 may be used, for example, to implement any of the functions described in the present disclosure, including: determining channel information based on one or more reference signals from a base station, generating a representation of the channel information based on the conditions of the channel using a machine learning model, sending a representation of the channel information, such as collecting training data during a window, such as performing pre-processing and / or post-processing on a CSI matrix, deploying and / or activating one or more pairs of ML models, etc.

[0310] The transceiver 1302 may transmit one or more signals to / receive one or more signals from a base station and may include an interface unit for such transmission / reception. For example, the transceiver 1302 may receive one or more signals from a base station and / or may transmit a representation of channel information to a base station on a UL channel.

[0311] The controller 1304 may include, for example, one or more processors 1306 and a memory 1308, wherein the memory 1308 may store instructions for one or more processors 1306 to execute code to implement any of the functions described in the present disclosure. For example, the controller 1304 may be configured to implement one or more machine learning models as disclosed herein, and determine channel information based on one or more reference signals from a base station, generate a representation of the channel information based on the conditions of the channel using the machine learning model, send a representation of the channel information, such as collecting training data during a window, such as performing pre-processing and / or post-processing for a CSI matrix, deploying and / or activating one or more pairs of ML models, etc.

[0312] Additionally or additionally, the controller 1304 may be configured to implement any functionality disclosed herein related to using constraint information for a channel, including, for example, providing a constrained subspace for a channel to the UE and / or enabling the UE to ensure that a precoder determined using a reconstruction model is orthogonal to the constrained subspace. For example, the controller 1304 may be configured to implement respectively Fig.24 , Fig.25 and / or Fig.26 Shown and / or referred to in Fig.24 , Fig.25 and / or Fig.26 Some or all of the CSI determination logic 2452, the precoding determination logic 2552, and / or the precoding determination logic 2652 described.

[0313] Additionally or alternatively, the controller 1304 may be configured to implement any of the functions disclosed herein related to data collection for use with artificial intelligence and / or machine learning models, including, for example, establishing a data collection session and / or declaring data collection capabilities. The controller 1304 may be configured, for example, to implement Fig.28 shown in and / or about Fig.28 The data collection logic 2866 and / or the data collection logic 2867 and / or the Fig.29 shown in and / or about Fig.29 Some or all of the data collection logic 2966 and / or data collection logic 2967 described.

[0314] Additionally or alternatively, the controller 1304 may be configured to implement any functionality disclosed herein related to data collection for non-coherent joint transmission, including, for example, reporting data collection capabilities and / or data set descriptions. The controller 1304 may be configured to implement, for example, Fig.30 shown in and / or about Fig.30 Some or all of the data collection logic 3066, 3067A, and / or 3067B described.

[0315] Fig.14 An example embodiment of a base station according to the present disclosure is shown. Fig.14 The embodiment 1400 shown in the figure may include a radio transceiver 1402 and a controller 1404, wherein the controller 1404 may control the operation of the transceiver 1402 and / or any other components in the base station 1400. The base station 1400 may be used, for example, to implement any of the functions described in the present disclosure, including: sending one or more reference signals to the UE on a DL channel, reconstructing a representation of channel information, such as performing pre-processing and / or post-processing for a CSI matrix, deploying and / or activating one or more pairs of ML models, etc.

[0316] The transceiver 1402 may send / receive one or more signals to / from the user equipment and may include an interface unit for such sending / receiving. For example, the transceiver 1402 may send one or more reference signals to the UE on a DL channel and / or receive precoding information from the UE on a UL channel.

[0317] The controller 1404 may include, for example, one or more processors 1406 and a memory 1408, wherein the memory 1408 may store instructions for one or more processors 1406 to execute code to implement any base station functions described in the present disclosure. For example, the controller 1404 may be used to implement one or more machine learning models as disclosed herein, and send one or more reference signals to the UE on a DL channel, reconstruct a representation of channel information, such as performing pre-processing and / or post-processing for a CSI matrix, deploying and / or activating one or more pairs of ML models, etc.

[0318] Additionally or alternatively, controller 1404 may be configured to implement any functionality disclosed herein related to using constraint information for a channel, including, for example, providing a constrained subspace for the channel to the UE and / or enabling the UE to ensure that a precoder determined using a reconstruction model is orthogonal to the constrained subspace.

[0319] Additionally or alternatively, controller 1404 may be configured to implement any functionality related to data collection for use with artificial intelligence and / or machine learning models disclosed herein, including, for example, establishing a data collection session and / or declaring data collection capabilities.

[0320] Additionally or additionally, the controller 1404 may be configured to implement any functionality disclosed herein related to data collection for non-coherent joint transmission, including, for example, reporting data collection capabilities and / or data set descriptions.

[0321] exist Fig.13 and Fig.14In the embodiment shown in , transceiver 1302 and transceiver 1402 can be implemented using various components (such as amplifiers, filters, modulators and / or demodulators, A / D and / or DA converters, antennas, switches, phase shifters, detectors, couplers, conductors, transmission lines, etc.) to receive and / or transmit RF signals. Controller 1304 and / or controller 1404 can be implemented using hardware, software, and / or any combination thereof. For example, all or part of the hardware implementation may include combinational logic, sequential logic, timers, counters, registers, gate arrays, amplifiers, synthesizers, multiplexers, modulators, demodulators, filters, vector processors, complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), systems on chips (SOCs), system-in-package (SIPs), multi-chip modules, state machines, data converters such as ADCs and DACs, etc. All or part of the software implementation may include one or more processor cores, memories, programs and / or data storage devices, etc., wherein the one or more processor cores, memories, programs and / or data storage devices, etc. may be located locally and / or remotely and may be programmed to execute instructions to perform one or more functions of the controller. Some embodiments may include one or more processors (such as microcontrollers), CPUs (such as complex instruction set computer (CISC) processors (such as x86 processors) and / or reduced instruction set computer (RISC) processors (such as ARM processors), graphics processing units (GPUs), neural processing units (NPUs), tensor processing units (TPUs), etc. that execute instructions stored in any type of memory.

[0322] Fig.15 An embodiment of a method for providing physical layer information feedback according to the present disclosure is shown. The method may start at operation 1502. At operation 1504, the method may determine physical layer information for the wireless device at the wireless device. At operation 1506, the method may generate a representation of the physical layer information using a machine learning model. At operation 1508, the method may send the representation of the physical layer information from the wireless device. The method may end at operation 1510.

[0323] exist Fig.15 In the embodiments shown in and any embodiments disclosed herein, the components and / or operations shown are exemplary only. Some embodiments may involve various additional components and / or operations not shown, and some embodiments may omit some components and / or operations. In addition, in some embodiments, the arrangement of components and / or the temporal order of operations may be changed. Although some components may be shown as separate components, in some embodiments, some components shown separately may be integrated into a single component, and / or some components shown as a single component may be implemented using multiple components.

[0324] Precoding information does not match channel information

[0325] In an NR system, a UE may measure downlink channel conditions based on a reference signal (e.g., CSI-RS or DMRS) sent by a gNB. The UE may use the channel measurement to determine (e.g., calculate) channel information that the UE may report to the gNB, such as a precoding matrix and / or a channel quality indicator (CQI). The UE may calculate a precoding matrix that may result in a good (e.g., best available) equivalent channel for downlink transmission if the gNB applies the reported precoding matrix to subsequent transmissions. The CQI that the UE may calculate based on the precoding matrix may indicate the channel quality that may be expected if the gNB uses the reported precoding matrix. As an example, the reported CQI may be used to select a modulation order and / or code rate, etc., for subsequent transmissions where the gNB uses the reported precoding matrix.

[0326] As described above, the CQI may be calculated based on a precoding matrix. Thus, accurately determining channel quality information (e.g., CQI) may involve or require the UE's knowledge or assumptions about the precoding information (e.g., precoding matrix) applied by the gNB. In NR systems where a codebook may be used to determine the precoding matrix (e.g., using the RI and / or PMI reported by the UE), the UE may have knowledge of the precoding matrix applied by the gNB to the CSI-RS.

[0327] However, in systems that use a machine learning architecture to report channel information, the UE may not have knowledge or assumptions about the precoding information applied by the gNB. For example, in some embodiments, a pair of models (e.g., an encoder at the UE and a decoder at the gNB) may be trained such that when the channel matrix is ​​applied as input to the encoder at the UE, the decoder at the gNB can directly output the precoding matrix. Such embodiments may be used, for example, Figure 7 , where the channel state information 718 may be implemented as a channel matrix determined by measuring the reference signal 717 and applied to the machine learning encoder 703 at the UE 701. In such an embodiment, reconstruction of the channel state information 722 may be implemented as a precoding matrix that may be obtained (e.g., directly) as an output of the machine learning decoder 704 at the gNB 702.

[0328] For example, you can use Figure 3Such an embodiment can be trained using the configuration shown in , where the training data 311 can include data pairs, where each data pair can include a channel matrix and a corresponding precoding matrix calculated by the UE based on the channel matrix. During training, the channel matrix can be applied as input to the generation model 303, and the corresponding precoding matrix can be used as a training target (e.g., for the loss function 313), so that the reconstruction model 304 generates output data 312 as a precoding matrix.

[0329] However, in such an embodiment, the UE may not have access to the trained decoder used by the gNB. Figure 7 , the encoder 703 and the decoder 704 may be trained by the network, which may transmit only the encoder 703 to the UE 701 and only the decoder 702 to the gNB 702. As another example, the gNB 702 may train the encoder 703 and the decoder 704 and transmit only the encoder 703 to the UE 701. As a result, the UE may not be able to determine the precoding matrix that the UE may report to the gNB (e.g., in the form of the output of the encoder 703) and the gNB 702 may reconstruct using the decoder 704. (Although the gNB 702 may reconstruct the precoding matrix, in some systems it may not be necessary to use the reconstructed precoding matrix or any precoding matrix previously reported by the UE 701). As a result, the UE 701 may not have access to the precoding matrix for calculating the channel quality information. Depending on the implementation details, this may result in a mismatch between the channel quality information reported by the UE to the gNB and the precoding information. In some embodiments, as used in the context, a mismatch may refer to a situation where the channel quality information reported by the wireless device may not be sufficiently based on the corresponding precoding information reported to the wireless device.

[0330] In some embodiments according to the present disclosure, a first wireless device (e.g., UE) may determine precoding information (e.g., precoding matrix) used by a second wireless device (e.g., gNB) to enable the first wireless device to determine channel quality information (e.g., CQI) based on the precoding information. Depending on implementation details, this may reduce or eliminate a mismatch between the channel quality information reported by the first wireless device to the second wireless device and the precoding information.

[0331] Fig.16 An embodiment of a system having a pair of models for providing channel information feedback according to the present disclosure is shown. Fig.16 The system 1600 shown in FIG. 1 may be used to implement any of the devices, models, training schemes, etc. disclosed herein, or may be implemented with any of the devices, models, training schemes, etc. disclosed herein. The system 1600 may include a Figure 4 and / or Figure 7One or more elements (eg, components, operations, etc.) of those in the embodiments shown in , where similar elements may be indicated by reference numerals ending with and / or containing the same numbers, letters, etc.

[0332] exist Fig.16 In the system 1600 shown in FIG. 1 , a first wireless device (e.g., UE) 1601 may receive a signal (e.g., a reference signal) 1617 from a second wireless device (e.g., a base station) 1602 via a channel 1615, for example, to enable the first wireless device 1601 to determine a channel condition of the channel 1615. The second wireless device 1602 may apply precoding information (e.g., a precoding matrix) 1645 to subsequent transmissions (e.g., PDCCH, PDSCH, etc.).

[0333] The first wireless device 1601 may include precoding determination logic 1643. Additionally or alternatively, the second wireless device 1602 may include sharing logic 1644. The precoding determination logic 1643 and / or the sharing logic 1644 may individually and / or collaboratively implement one or more schemes to enable the first wireless device 1601 to determine (e.g., calculate) precoding information 1645 (or a reconstruction thereof), which in turn may enable the first wireless device 1601 to determine channel quality information 1659. Depending on implementation details, enabling the first wireless device 1601 to determine the precoding information 1645 applied by the second wireless device 1602 may help reduce or eliminate mismatches between the precoding information 1645 (or a reconstruction thereof) and the channel quality information 1659 reported by the first wireless device 1601 to the second wireless device 1602.

[0334] In some example embodiments, the first wireless device 1601 may use a reconstruction model (e.g., a decoder model) to generate a reconstruction 1646 of the precoding information 1645. For example, the sharing logic 1644 at the second wireless device 1602 may transmit the reconstruction model 1604 (e.g., information regarding model size, specifications, weights, etc.) to the first wireless device 1601, where the transmitted model may be indicated as the shared model 1604A.

[0335] Such an embodiment may be implemented, for example, with a generation model 1603 implemented by an encoder and a reconstruction model 1604 implemented by a decoder. The second wireless device 1602 may train the encoder 1603 and the decoder 1604 (e.g., in an autoencoder configuration). The shared logic 1644 at the second wireless device 1602 may transmit both the encoder 1603 and the decoder 1604A to the first wireless device 1601, for example, using an over-the-air (OTA) transmission. The first wireless device 1601 may use the encoder 1603 to generate a representation 1607 based on channel information 1605, which may include, for example, channel measurements (e.g., a channel matrix). The second wireless device 1602 may use the decoder 1604 to generate precoding information (e.g., a precoding matrix) 1645 from the representation 1607. However, because the first wireless device 1601 may also receive the decoder 1604A, it may use the decoder 1604A to generate reconstructed precoding information (e.g., a precoding matrix) 1646 that may be the same or similar to the precoding information 1645. The first wireless device 1601 may then perform calculations 1660 using the reconstructed precoding information 1646 to determine channel quality information 1659 .

[0336] The first wireless device 1601 may send the channel quality information 1659 to the second wireless device 1602 in any suitable manner. For example, the first wireless device 1601 may combine the channel quality information 1659 with the representation 1607 generated by the generation model 1603 (e.g., by appending the channel quality information 1659 to the representation 1607), and send the combined channel quality information 1659 and representation 1607 to the second wireless device 1602 using another channel (e.g., an uplink channel), a signal, etc. 1616. In some embodiments, the second wireless device 1602 may remove the channel quality information 1659 from the combined information before applying the representation 1607 to the reconstruction model 1604.

[0337] As another example, the first wireless device 1601 may receive the reconstructed model 1604A from a server (e.g., using OTA transmission) on a wireless network on which the first wireless device 1601 and the second wireless device 1602 may operate. In such an embodiment, the reconstructed model 1604A may be identified and / or registered with the network, and the network may use the model identifier to activate the reconstructed model 1604A. If the reconstructed model 1604A is shared by the second wireless device 1602 with the first wireless device 1601, the second wireless device 1602 may share the reconstructed model 1604A corresponding to the activated generated model 1603.

[0338] In some additional example embodiments, the first wireless device 1601 may use a locally trained reconstruction model 1647 to generate a reconstruction 1646 of the precoding information 1645. For example, the first wireless device 1601 may train a reference model (e.g., a reference decoder) 1647 that may match the reconstruction model 1604 used by the second wireless device 1602. The first wireless device 1601 may then use the locally trained model 1647 to generate (e.g., reconstruct) the precoding information 1646, which may then be used to determine (e.g., calculate) channel quality information (e.g., CQI) 1659. One or more of these operations may be controlled, supported, etc., individually and / or collaboratively by the precoding determination logic 1643.

[0339] In some embodiments, the first wireless device 1601 may implement the reference model as a reference decoder model (which may be the same or different from the decoder model used by the second wireless device 1602) that the first wireless device 1601 may refine, for example, by fine-tuning. Alternatively or additionally, the first wireless device 1601 may implement the reference model using its own encoder model, which the first wireless device 1601 may use to train a decoder model, which may be different from the decoder model used by the second wireless device 1602, but which the first wireless device 1601 may still use to determine a precoding matrix, which in turn may be used to determine channel quality information to be reported to the second wireless device 1602.

[0340] In some other example embodiments where the first wireless device 1601 (e.g., UE) may report channel information 1605 (e.g., channel matrix) to the second wireless device 1602, a training data set including precoding information (e.g., one or more precoding matrices) 1646 calculated by the first wireless device 1601 may be used. Thus, the first wireless device 1601 may know the corresponding precoding information 1645 obtained at the second wireless device 1602. However, since the number of columns of the precoding matrix may correspond to the rank reported using the RI, the second wireless device 1602 may not know the size of the precoding matrix and, in turn, the corresponding UCI payload size. In such an embodiment, the first wireless device 1601 may report enough matrices (e.g., a full N_T×N_T matrix) to enable the second wireless device 1602 to determine the precoding information (e.g., precoding matrix) 1646. For example, the second wireless device 1602 may then use a certain number of the first columns of the matrix as the precoding matrix, where the number of the first columns may be determined by (e.g., equal to) the rank as reported by the RI. In some embodiments where the first wireless device 1601 is implemented by a UE, the UE may use an existing algorithm to calculate a precoding matrix and then calculate a CQI based on the precoding matrix. One or more of these operations may be controlled, supported, etc., by the precoding determination logic 1643, individually and / or collaboratively.

[0341] According to some additional embodiments of the present disclosure, the mismatch between precoding information and channel quality information can be reduced or eliminated by jointly processing (e.g., compressing) channel information (e.g., channel information matrix, such as channel matrix (H), precoding matrix (P), etc.) and channel quality information (e.g., CQI) that can be determined based on the channel information. For example, in some embodiments, the generation model and the reconstruction model (e.g., in an autoencoder configuration) can be trained to jointly compress and / or decompress the channel information and the channel quality information that can be determined based on the channel information.

[0342] Fig.17 An example embodiment of a pair of models that may be used for joint compression of channel information and channel quality information according to the present disclosure is shown. Fig.17 The embodiments shown in can be used, for example, to implement Figure 4 The generated model 403 and / or the reconstructed model 404 shown in FIG. 4 , and Figure 7 The encoder 703 and / or decoder 704 shown in .

[0343] refer to Fig.17 , the generation model can be implemented using the machine learning encoder 1703, and the reconstruction model can be implemented using the machine learning decoder 1704. iand CQI i , where i=1, ..., N, where N may indicate the number of subbands) may be applied as input to encoder 1703, which may generate a joint representation 1707 of the input. A reconstructed precoding matrix and CQI (respectively denoted as RecPrecoder , ... i and RecCQI i , where i=1,…,N) is the output of decoder 1704.

[0344] The encoder 1703 and the decoder 1704 may be trained using training data, which may include target channel information (e.g., target CQI) for one or more (e.g., each) subband i, which may be or may be based on a precoding matrix for the subband and a CQI calculated by the UE based on the corresponding precoding matrix. The UE may calculate the precoding matrix and / or the corresponding CQI for the subband using one or more existing algorithms for precoding matrix and / or CQI calculation. Once the pair of models is trained (e.g., when used for inference), the reconstructed precoding output RecPrecoder i Can match the corresponding reconstructed CQI output RecCQI i , for example, because the precoding information and the corresponding CQI may match in the training data used to train the pair of models 1703 and 1704. Alternatively or additionally, the training data used to train the pair of models 1703 and 1704 may be generated by any other source (e.g., network, base station, etc.) that may provide training data, which may include sufficient matching between precoding information (e.g., precoding matrix) and channel quality information (e.g., CQI). One or more of these operations may be individually and / or collaboratively controlled, supported, etc. by joint processing logic that may be located at the UE, gNB, network, or any other location or combination thereof.

[0345] The training of the encoder 1703 and the decoder 1704 may be performed at the UE, at the base station, at the network (e.g., a server), or at any other location, and the trained encoder 1703 and decoder 1704 may be transmitted to the location where they will be used for inference. For example, the encoder 1703 may be transmitted to the UE (if it has not been trained at the UE), and the decoder 1704 may be transmitted to the base station (if it has not been trained at the base station).

[0346] Although the present invention may be described in the context of one or more precoding matrices and / or CQIs as input and / or output Fig.17, but any type of channel information, precoding information, etc. may be used as input and / or output. For example, in some embodiments, for each subband, in addition to or as an alternative to the precoding matrix, a channel matrix may be applied as an input to the encoder 1703. In such embodiments, for each subband, in addition to or as an alternative to the reconstructed precoding matrix, a reconstruction of the channel matrix may be reconstructed as an output by the decoder 1704. In some embodiments where the channel matrix may be applied as an alternative to the precoding matrix to the input of the encoder 1703, for each subband, the UE may calculate the precoding matrix as an intermediate result, from which a corresponding CQI that may be applied to the encoder 1703 is calculated.

[0347] Furthermore, although it may be described in the context of multiple subbands Fig.17 The embodiments shown in , however, these principles can be applied to train a pair of models to jointly compress channel information and precoding information for multiple and / or individual frequency bands, sub-bands, etc.

[0348] In some embodiments, in addition to potentially providing a match between channel information components (such as CQI) and precoding information, joint compression (e.g., for joint reporting) may provide one or more other potential benefits, depending on implementation details. For example, using different ML models (e.g., encoders and / or decoders) to report each channel information quantity (e.g., RI, CQI, PMI, etc.) may increase the time, complexity, etc. involved in training and / or reasoning for different models. However, according to embodiments of the present disclosure, joint compression (and / or joint reporting) may be used for values ​​such as RI, CQI, PMI, channel matrix, precoding matrix, etc., which may reduce the time, complexity, etc. involved in training and / or reasoning, depending on implementation details. Some embodiments may increase the size of the model used for joint compression and / or reporting based on, for example, different distributions of channel information quantities to improve compression performance. In some embodiments, a wireless device (e.g., UE) may be configured with different encoders and / or decoders, where each encoder, decoder, model pair, etc. may be associated with an identifier (e.g., a model ID) and mapped to a specific combination of channel information quantities. In some embodiments, more than one model may be mapped to the same set of quantities, where different encoders, decoders, model pairs, etc. (identified by different model IDs) may be used to handle different channel environments, configurations, etc.

[0349] CQI compression across subbands

[0350] In NR systems, the UE may report CQI using compression with a differential scheme that may be based on CQIs of different subbands that deviate around a wideband value. For example, instead of using 4×N_sb bits (where N_sb may indicate the number of subbands) to report CQI, the UE may report a wideband CQI value and a differential value of the CQI using 2 bits per subband. However, such a reporting scheme may not provide sufficient compression in some implementations, particularly in systems where frequency division duplexing (FDD) may be used and / or where there may be an increased antenna size with a relatively large amount of feedback reporting.

[0351] A scheme for compressing and / or reporting channel information using machine learning according to the present disclosure may compress a combination of channel information for a group of subbands. Such an embodiment may be useful, for example, for compressing and / or reporting quantities such as CQI, where a UE may be configured to report a subband CQI for one or more subbands (e.g., each or a subset) for channel information reporting. Depending on implementation details, a scheme for compressing and / or reporting channel information on a subband basis using machine learning according to the present disclosure may improve performance and / or flexibility, reduce complexity, etc., using one or more correlations (e.g., in the time domain, frequency domain, and / or spatial domain).

[0352] Fig.18 An embodiment of a system having a pair of models for compressing channel information based on one or more subbands according to the present disclosure is shown. Fig.18 The system 1800 shown in can be used to implement any of the devices, models, training schemes, etc. disclosed herein, or can be implemented with any of the devices, models, training schemes, etc. disclosed herein.

[0353] System 1800 may include a Figure 4 and / or Fig.16 One or more elements (e.g., components, operations, etc.) of those elements in the embodiments shown in , where similar elements may be indicated by reference numerals ending with and / or containing the same numbers, letters, etc. However, in Fig.18 In the illustrated system 1800, the first wireless device 1801 may include subband compression logic 1848. Additionally or alternatively, the second wireless device 1802 may include subband decompression logic 1849. The subband compression logic 1848 and / or the subband decompression logic 1849 may individually and / or collaboratively implement one or more schemes to enable the first wireless device 1801 to compress a combination of channel information for a group (e.g., one or more) of subbands.

[0354] For example, at a first wireless device 1801 (e.g., UE), the subband compression logic 1848 may concatenate the values ​​of the channel information 1805 for the plurality of subbands into a vector, which may be compressed by the generation model 1803 to generate a representation 1807 of the values ​​of the channel information 1805 for the plurality of subbands. At a second wireless device 1802 (e.g., base station), the reconstruction model 1804 may recover the vector or an approximation of the vector, and the subband decompression logic 1849 may divide the vector or an approximation of the vector into respective subbands to generate reconstructed channel information 1806 or an approximation thereof for the plurality of subbands.

[0355] Fig.19 An embodiment of a pair of models that may be used for CQI compression across subbands according to the present disclosure is shown. Fig.19 The embodiments shown in can be used, for example, to implement Fig.18 The generation model 1803 and / or reconstruction model 1804 shown in . For illustrative purposes, when used with the system 1900, the first wireless device 1801 and the second wireless device 1802 can be implemented as a UE and a gNB, respectively.

[0356] refer to Fig.19 , a machine learning encoder 1903 may be used to implement a generation model, and a machine learning decoder 1904 may be used to implement a reconstruction model. CQI vector (CQI 1 ,CQI 2 ,…,CQI N ) may include a CQI value for one or more (eg, each) of subbands 1, 2, ..., N (where N may indicate the number of subbands). The CQI values ​​CQI for the subbands may be concatenated, for example. 1 、CQI 2 , …, CQI N To form the CQI vector (CQI 1 ,CQI 2 ,…,CQI N ). The CQI vector may be applied to encoder 1903 so that each value CQI 1 、CQI 2 , …, CQI N can be a separate input to encoder 1903, such as Fig.19 Alternatively or additionally, the entire CQI vector may be applied as a single input to the encoder 1903. The encoder 1903 may use the CQI vector and / or the individual CQI values ​​to generate a representation 1907 of the value of the channel information for the plurality of subbands.

[0357] Decoder 1904 may receive representation 1907 and generate reconstructed CQI values ​​RCQI for subbands 1, 2, ..., N 1RCQI 2 ,…,RCQI N As individual outputs and / or as reconstruction vectors (RCQI 1 ,RCQI 2 ,…,RCQI N ), reconstruction vector (RCQI 1 ,RCQI 2 ,…,RCQI N ) may be partitioned (eg, by subband decompression logic) to provide individual outputs for one or more subbands.

[0358] In some embodiments, the CQI value CQI 1 、CQI 2 , …, CQI N One or more of may be provided as an input (e.g., a discrete input) to a CQI table (e.g., a CQI index, which may be a scalar value, e.g., between 0 and 15). For example, if the CQI table has sixteen rows, then the CQI value CQI i (where i = 1, 2, ..., N) can be a 4-bit value. Optionally or additionally, the CQI value CQI 1 、CQI 2 , …, CQI N One or more of may be provided in the form of a continuous value (e.g., an arbitrary real value) of the coding rate and / or the modulation order. In embodiments using one or more continuous values ​​as CQI values, for example, a modulation and coding scheme (MCS) table with a relatively high resolution may be used to accommodate a finer granularity indication by the gNB.

[0359] According to the implementation details, as mentioned above Fig.18 and Fig.19 The described scheme for compressing and / or reporting channel information (e.g., CQI) on a subband basis may exploit one or more correlations in the input data (e.g., between elements of a CQI vector) to improve performance and / or flexibility, reduce complexity, etc. For example, the encoder 1903 may be able to compress CQI vectors having multiple elements with the same or similar values ​​to a greater extent than CQI vectors having different values.

[0360] In some embodiments, one or more of the encoder 1903, the decoder 1904, the generation model 1803, and / or the reconstruction model 1804 may be implemented using a machine learning model that may be adapted to work with inputs having different correlation properties. Thus, for example, a model for compressing CQI values ​​across subbands may be specialized (e.g., as opposed to a general ML model) to improve or optimize compression performance based on, for example, the range of input values, vector and / or matrix dimensions, weights, etc.

[0361] Compression of channel information

[0362] In an NR system, the UE may compress the precoding matrix in the spatial domain and / or frequency domain using a codebook. For example, the UE may report the RI and the PMI associated with the RI (e.g., to the gNB). The rank indicated by the RI and the PMI may be used together to determine the precoding matrix. For example, the PMI may be selected from a set of support matrices that may be referred to as a codebook. In some embodiments, the codebook may be implemented as a mapping table between a set of i indices and a precoding matrix. For example, the UE may use an index (i_1, 1, i_1, 2, i_1, 3, i_2) with a type 1 single-panel codebook that may uniquely define a precoding matrix for a given rank. However, the codebook compression scheme may not provide sufficient compression in some embodiments, especially in systems where frequency division duplexing (FDD) may be used and / or where there may be an increased antenna size with a relatively large amount of feedback reporting.

[0363] According to the scheme for compressing and / or reporting channel information using machine learning disclosed herein, one or more decoder models can be used to generate precoding information and / or other information that can be used to, for example, determine precoding information. In some embodiments, such a scheme can imitate a codebook scheme and / or implement a hierarchical compression mechanism. In some embodiments, such a scheme can implement one or more encoders and / or decoders with a specific structure for compression in the time domain, frequency domain, and / or spatial domain. According to implementation details, such a scheme can improve performance and / or flexibility, reduce complexity, etc.

[0364] Fig. 20 An embodiment of a system having a pair of models for providing channel information compression according to the present disclosure is shown. Fig. 20 The system 2000 shown in FIG. 1 can be used to implement any of the devices, models, training schemes, etc. disclosed herein, or can be implemented with any of the devices, models, training schemes, etc. disclosed herein. The system 2000 may include a Figure 4 , Fig.16 and / or Fig.18 One or more elements (eg, components, operations, etc.) of those in the embodiments shown in , where similar elements may be indicated by reference numerals ending with and / or containing the same numbers, letters, etc.

[0365] exist Fig. 20In the system 2000 shown in FIG. 1 , a first wireless device (e.g., UE) 2001 may receive a signal (e.g., a reference signal) 2017 from a second wireless device (e.g., a base station) 2002 via a channel 2015. The second wireless device 2002 may apply precoding information (e.g., a precoding matrix) 2045 to the signal 2017. The first wireless device 2001 may include compression scheme logic 2050. Additionally or alternatively, the second wireless device 2002 may include compression scheme logic 2051. The compression scheme logic 2050 and / or 2051 may individually and / or collaboratively implement one or more schemes to enable the first wireless device 2001 and the second wireless device 2002 to implement one or more compression schemes for generating a representation 2016 of the channel information 2005.

[0366] Fig. 20 The embodiments shown in are not limited to any particular form of channel information 2005 applied to the generation model 2003 and / or any particular form of reconstructed channel information 2006 generated by the reconstruction model 2004. For example, in some example embodiments, the generation model 2003 and the reconstruction model 2004 may be implemented with an encoder and a decoder, respectively (e.g., configured as an autoencoder), and the encoder and the decoder may be configured and / or trained to generate precoding information based on channel information (e.g., a channel matrix) using one or more compression schemes. In such an embodiment, a first wireless device 2001, which may be implemented as a UE, may perform one or more CSI-RS measurements to obtain channel information such as one or more channel matrices, eigenvectors, etc. The UE may apply the channel information to the encoder to generate a codeword that the UE may send to a second wireless device 2002, which may be implemented as a base station (e.g., a gNB). The base station may apply the codeword to the decoder to construct a precoding matrix. In such an embodiment, the codeword may mimic one or more codebook indexes, and / or the decoder may mimic the codebook, but depending on the implementation details, the scheme using one or more decoders as described above may improve performance and / or flexibility, reduce complexity, etc.

[0367] However, in other embodiments, the channel information 2005 and / or the reconstruction of the channel information 2006 may be implemented in any other form. For example, instead of generating a precoding matrix as an output of the generation model 2003 based on the channel information applied as input to the generation model 2003, the generation model 2003 and the reconstruction model 2004 may be trained so that the reconstruction model 2004 may reconstruct the channel information applied as input to the generation model 2003 or an approximation thereof. The second wireless device 2002 may then determine (e.g., calculate) the precoding matrix using the reconstructed channel information. In other embodiments, the channel information 2005 and / or the reconstruction of the channel information 2006 may be implemented using CQI, RI, and / or any other type of information.

[0368] In embodiments where an encoder at the UE and a decoder at the base station can be configured and / or trained to generate precoding information such as a precoding matrix based on channel information such as a channel matrix, the output of the decoder (and post-processing, if any) can be a precoding matrix. In some such embodiments, the UE may also be configured with a decoder that is trained to provide an output as a precoding matrix. The decoder at the UE may be obtained, for example, from a base station that may share a decoder model including weights, input and / or output dimensions, etc. In some embodiments, the UE and the base station may have a common understanding of the decoder output type (e.g., how precoding information may be constructed based on the decoder output).

[0369] In embodiments where a post-processing scheme may be applied after the decoder, the pre-processing scheme may be shared with the UE. One or more implicit and / or explicit indications may be provided to the UE of how to construct precoding information (e.g., a precoding matrix) from the output of the decoder (and any post-processing, if any). For example, if the output dimension of the decoder is N_t*ν, where ν is the rank indicated by the RI, the precoder may be constructed by reconstructing the output vector into a N_t×v matrix on a column-by-column and / or row-by-row basis.

[0370] The following describes the Fig. 20 Some embodiments of the training and / or testing scheme of the system shown in . For illustration purposes, embodiments of the training and / or testing scheme may be described in the context of a system in which the first wireless device 2001 and the second wireless device 2002 may be implemented with a UE and a gNB, respectively, and the generation model 2003 and the reconstruction model 2004 may be implemented with an encoder and a decoder, respectively. However, the principles are not limited to these or any other implementation details.

[0371] In a first embodiment of a training and / or testing scheme, the encoder and decoder may be jointly designed, configured, trained, etc. based on a training data set (e.g., by a UE vendor and a gNB vendor, respectively), which may include a set (X, Y), where X may refer to one or more CSI-RS measurements, such as, for example, a channel matrix of a CSI-RS set acquired in a time and frequency window, and Y may refer to a precoding matrix that may be calculated, for example, by the UE. The decoder may be shared with the UE (e.g., transmitted to the UE) for use during inference, for example, to generate a precoding matrix that the UE may use for CSI-RS measurements. During inference, the UE may apply the encoder in a manner similar to that used during training.

[0372] In a second embodiment of the training and / or testing scheme, the gNB may train the decoder using a training data set, which may or may not be available from the UE. The gNB may share the decoder with the UE, including one or more weights, input and / or output dimensions, post-processing, etc. In some embodiments, the training of the encoder and one or more input types and / or dimensions may be determined by the UE based on, for example, the UE's implementation. When training the encoder, the UE may seek to minimize the loss function between the output P_dec of the decoder and the desired target precoder P_target that may be calculated by the UE. In some embodiments, the gNB may configure the UE with one or more time and / or frequency windows and the CSI-RS used in the windows, which may be used as training data to train the encoder. The UE may then use the trained encoder for inference.

[0373] In a third embodiment of the training and / or testing scheme, the gNB may train the decoder based on a training dataset, which may or may not be obtained from the UE, and the gNB may share the decoder with the UE, including one or more weights, input and / or output dimensions, post-processing, etc. In some embodiments, the training of the encoder and one or more input types and / or dimensions may be determined by the UE based on, for example, the UE's implementation. However, in this type of embodiment, the gNB may also share the training dataset used to train the encoder with the UE, including one or more inputs and / or labels. The UE may then use the shared training dataset to train the encoder for use during inference.

[0374] In any of the three embodiments of the above training and / or testing schemes, because the precoding matrix dimension may depend on the reported rank, the UE may apply one or more different models for one or more different reported ranks. For example, if the UE reports a first rank as RI, the UE may use a first encoder model trained for a first decoder, and if the UE reports a second rank as RI, the UE may use a second encoder model trained for a second decoder.

[0375] In any of the three embodiments of the training and / or testing schemes described above, depending on the implementation details, the decoder can be characterized as performing a role similar to that of the NR codebook, however, the machine learning decoder model has the potential to improve performance and / or flexibility, reduce complexity, etc. compared to using a codebook.

[0376] Fig.21 A first embodiment of a pair of models that may be used to implement a compression scheme according to the present disclosure is shown. Fig.21 The embodiment 2100 illustrated in FIG. 2 can be used, for example, to implement Fig. 20 The generated model 2003 and the reconstructed model 2004 are shown in FIG.

[0377] refer to Fig.21 , embodiment 2100 may include a generation model 2103 and a reconstruction model 2104 that may be configured to operate, for example, as an autoencoder. The generation model 2103 (which may be located, for example, at a UE) may include one or more spatial encoders 2153-1 to 2153-N, which may be arranged to receive one or more spatial encoder inputs SEI-1 to SEI-N corresponding to channel information for subbands 1, 2, ..., N, respectively, where N may refer to the number of subbands. The one or more spatial encoders 2153-1 to 2153-N may generate one or more spatial compressed outputs SEO1 to SEO-N based on the corresponding encoder inputs SEI-1 to SEI-N, respectively.

[0378] The compressed outputs SEO-1 to SEO-N may be transmitted to a reconstruction model 2104 (which may be located, for example, at a base station), where the compressed outputs SEO-1 to SEO-N may be applied as subband decoder inputs SDI-1 to SDI-N, respectively, to one or more spatial decoders 2154-1 to 2154-N, respectively. The one or more spatial decoders 2154-1 to 2154-N may spatially decompress the decoder inputs SDI-1 to SDI-N, respectively, to generate one or more decompressed outputs SDO-1 to SDO-N, respectively.

[0379] Although not limited to any particular implementation details, in some embodiments, the encoder input SEI-1 to SEI-N can be implemented as a channel matrix H_1^i (where i=1, 2, ..., N) for the corresponding subband, and the spatial encoder output SEO-1 to SEO-N can be implemented as a representation v_1^i of the precoding matrix P_1^i for the corresponding subband. Then, the spatial decoder 2154-1 to 2154-N can generate a decompressed output SDO-1 to SDO-N from the representation v_1^i as a reconstructed precoding matrix. Therefore, the reporting of precoding information can be performed independently in different subbands.

[0380] In some embodiments, for example, in embodiments where the spatial encoders 2153-1 to 2153-N may be implemented with more than one set of hardware (e.g., a separate processor, circuit, etc. for each encoder), the spatial encoders 2153-1 to 2153-N may perform spatial compression operations in parallel. In some other embodiments, for example, in embodiments where the spatial encoders 2153-1 to 2153-N may be implemented with less than N hardware instances (e.g., a single separate processor, circuit, etc. for all N encoders), one or more of the spatial encoders 2153-1 to 2153-N may perform spatial compression operations sequentially. Similarly, the spatial decoders 2154-1 to 2154-N may perform spatial decompression operations in parallel or sequentially, depending on, for example, the number of hardware sets used to implement the decoders.

[0381] Fig. 22 A second embodiment of a pair of models that may be used to implement a compression scheme according to the present disclosure is shown. Fig. 22 The embodiment 2200 shown in FIG. 2 can be used, for example, to implement Fig. 20 The generation model 2003 and the reconstruction model 2004 shown in FIG. Embodiment 2200 may include Fig.21 One or more elements (eg, components, operations, etc.) that are similar to the elements in the illustrated embodiments may be indicated by reference numerals that end with and / or include the same numbers, letters, etc.

[0382] refer to Fig. 22, embodiment 2200 may include a generation model 2203 and a reconstruction model 2204 that may be configured to operate, for example, as an autoencoder. The generation model 2203 (which may be located, for example, at a UE) may include one or more spatial encoders 2253-1 to 2253-N, which may be arranged to receive one or more spatial encoder inputs SEI-1 to SEI-N corresponding to channel information for subbands 1, 2, ..., N, respectively, where N may refer to the number of subbands. The one or more spatial encoders 2253-1 to 2253-N may generate one or more spatial compressed outputs SEO1 to SEO-N based on the corresponding encoder inputs SEI-1 to SEI-N, respectively.

[0383] The generation model 2203 may further comprise a frequency encoder 2255 which may be arranged to receive one or more spatially compressed outputs SEO-1 to SEO-N and generate a spatially and frequency compressed representation FEO of the subband encoder inputs SEI-1 to SEI-N.

[0384] The compressed representation FEO may be transmitted to a reconstruction model 2204 (which may be located, for example, at a base station), where the compressed representation FEO may be applied as an input to a frequency decoder 2256, which may generate one or more spatially decompressed outputs SDI-1 to SDIN corresponding to subbands 1, 2, ..., N, respectively, which may be decompressed in the frequency domain but still compressed in the spatial domain. The N outputs of the frequency decoder 2256 may be applied as inputs SDI-1 to SDI-N, respectively, to one or more spatial decoders 2254-1 to 2254-N. The one or more spatial decoders 2254-1 to 2254-N may spatially decompress the decoder inputs SDI-1 to SDI-N, respectively, to generate one or more decompressed outputs SDO-1 to SDO-N, respectively, which may be decompressed in both the frequency domain and the spatial domain. Thus, spatial compression and decompression of precoding information may be performed independently in different subbands, while frequency compression and decompression may be performed simultaneously across subbands.

[0385] Although not limited to any particular implementation details, in some embodiments, the input SEI-1 to SEI-N of the spatial encoders 2253-1 to 2253-N can be implemented as a channel matrix H_1^i (where i=1, 2, ..., N) for the corresponding subband, and the spatial encoder output SEO-1 to SEO-N can be implemented as a spatially compressed representation v_1^i of the precoding matrix P_1^i for the corresponding subband. The frequency encoder 2255 can compress the representation v_1^i in the frequency domain to generate an output FEO as a representation vector v, which can be compressed in both the frequency domain and the spatial domain. The frequency decoder 2256 can decompress the representation vector v in the frequency domain to reconstruct the spatially compressed representation v_1^i, which can be applied as one or more inputs SDI-1 to SDIN to one or more spatial decoders 2154-1 to 2154-N, and then the spatial decoders 2154-1 to 2154-N can generate decompressed outputs SDO-1 to SDO-N as the reconstructed precoding matrix based on the representation v_1^i.

[0386] Therefore, in Fig. 22 In the illustrated embodiment, the generation model 2203 may operate in two stages. In the first stage, the precoding information may be compressed independently at each subband via spatial domain compression. In this stage, N encoder models may be used, for example, one encoder model for each subband. The models may be the same or different, depending, for example, on various parameters associated with different subbands within the channel. In the second stage, one coding model may be used to jointly compress the spatially compressed precoding information of the subbands in the frequency domain. In some embodiments, the outputs SEO-1 to SEO-N of one or more spatial encoders 2253-1 to 2253-N may represent a precoding matrix for each subband without using inter-subband correlation. The frequency decoder 2256 may then frequency compress the spatially compressed precoding matrices for each subband and compress them into a vector v, which may include information for representing all precoding matrices. Thus, Fig. 22 The embodiments shown in may provide frequency domain decompression and a per-subband decoder model that may reconstruct a precoding matrix for each subband.

[0387] In some embodiments, for example, in embodiments where the spatial encoders 2253-1 to 2253-N and the frequency encoder 2255 may be implemented with more than one set of hardware (e.g., a separate processor, circuit, etc. for each encoder), the spatial encoders 2253-1 to 2253-N and / or the frequency encoder 2255 may perform compression operations in parallel. In some other embodiments, for example, in embodiments where the spatial encoders 2253-1 to 2253-N and / or the frequency encoder 2255 may be implemented with less than N+1 hardware instances (e.g., a single separate processor, circuit, etc. for all N+1 encoders), one or more of the spatial encoders 2253-1 to 2253-N and / or the frequency encoder 2255 may perform compression operations sequentially. Similarly, the spatial decoders 2254-1 to 2254-N and / or the frequency decoder 2256 may perform spatial decompression operations in parallel or sequentially, depending on, for example, the number of hardware sets used to implement the decoders.

[0388] Fig.23 A third embodiment of a pair of models that may be used to implement a compression scheme according to the present disclosure is shown. Fig.23 The embodiment 2300 shown in FIG. 2300 may be used, for example, to implement Fig. 20 The generation model 2003 and the reconstruction model 2004 shown in FIG. Embodiment 2300 may include Fig.21 and / or Fig. 22 One or more elements (eg, components, operations, etc.) similar to the elements in the embodiments shown in , where similar elements may be indicated by reference numerals ending with and / or containing the same numbers, letters, etc.

[0389] refer to Fig.23 , embodiment 2300 may include a generation model 2303 and a reconstruction model 2304 that may be configured to operate, for example, as an autoencoder. The generation model 2303 (which may be located, for example, at a UE) may include a joint spatial and frequency encoder 2357 that may receive one or more encoder inputs SFEI-1 to SFEI-N corresponding to channel information for subbands 1, 2, ..., N, respectively, where N may refer to the number of subbands. The joint spatial and frequency encoder 2357 may generate an output SFEO that may be a spatially and frequency compressed representation of the encoder inputs SFEI-1 to SFEI-N.

[0390] The spatially and frequency compressed representation SFEO may be transmitted to a reconstruction model 2304 (which may be located, for example, at a base station), where the spatially and frequency compressed representation SFEO may be applied as input to a joint spatial and frequency decoder 2358, which may generate one or more decoder outputs SFDO-1 to SFDO-N, which may be decompressed in both the spatial and frequency domains, corresponding to subbands 1, 2, ..., N, respectively. Thus, spatial and frequency compression may be performed simultaneously across subbands, and spatial and frequency decompression may be performed simultaneously across subbands.

[0391] Although not limited to any particular implementation details, in some embodiments, the input SFEI-1 to SFEI-N may be implemented as a channel matrix H_1^i for the corresponding subband (where i=1, 2, ..., N), and the output SFEO may be implemented as a representation vector v, which may be a spatially and frequency-compressed representation of the N precoding matrices corresponding to the N subbands. The joint spatial and frequency decoder 2358 may decompress the representation vector v in the spatial domain and the frequency domain to recover the reconstructed precoding matrix P^_1^i. Therefore, in some embodiments, Fig.23 In the embodiment 2300 shown in , the joint spatial and frequency encoder 2357 can be characterized as jointly computing the precoding matrices P_1^N and compressing them into a vector v, and the joint spatial and frequency decoder 2358 can be characterized as restoring the matrices to P^_1^N.

[0392] In some embodiments, spatial correlation may refer to correlation across one or more transmit (Tx) antenna ports. Any of the embodiments disclosed above may implement spatial compression individually (e.g., per layer) or jointly across multiple (e.g., all) layers of a channel. For example, in some embodiments, a precoding matrix may be applied as input to an encoder. For example, a 4×3 precoding matrix may have three layers (e.g., columns), each of which has a precoding vector of length 4 (e.g., four rows). In some embodiments, spatial compression may be implemented jointly across multiple layers (e.g., all layers) by applying the entire matrix to the encoder. Alternatively or additionally, spatial compression may be implemented layer by layer, for example, applying the layers (e.g., columns) of the matrix to the encoder one layer at a time (e.g., applying a vector (column) having four elements (rows) at a time). The number of layers may be indicated, for example, by RI.

[0393] In NR systems, one or more of the following constraints may be imposed on channel information reporting. The UE may calculate CSI parameters (where reported) assuming the following dependencies between CQI parameters (where reported): LI may be calculated conditioned on reported CQI, PMI, RI, and CRI; CQI may be calculated conditioned on reported PMI, RI, and CRI; PMI may be calculated conditioned on reported RI and CRI; and / or RI may be calculated conditioned on reported CRI.

[0394] In some embodiments, one or more of the ML-based compression and / or reporting schemes disclosed herein may implement one or more constraints, which may be similar to the constraints described above, depending on implementation details. Thus, in some embodiments, if a UE reports a CQI, the CQI for the subband (e.g., each subband) may be calculated based on precoding information (e.g., a precoding matrix) reported for one or more subbands. For example, in the case of Fig.21 In the case of the described embodiments, the CQI for the subband may be calculated based on the precoding matrix P^_i, which may involve sharing the decoder with the UE. In embodiments where the decoder may not be shared with the UE, the UE may calculate the CQI based on information for each subband #i. As another example, in the case of Fig. 22 and Fig.23 In the case of the described embodiments, if the decoder is shared with the UE, the UE may calculate the CQI for subband #i based on the subband PMI i or the decoder output P^_i. In some embodiments, for the purpose of specifying this condition, the precoding information (e.g., PMI) of each subband may refer to the subband encoder output (or input) or the corresponding decoder output, which may be assumed to represent the subband precoding matrix according to implementation details.

[0395] In an NR system, the bit width of the amount of channel information reported may depend on one or more RRC parameters. In some embodiments, the bit width for reporting precoding information (e.g., PMI) may additionally depend on the rank reported by RI. Within an architecture for reporting channel information using machine learning according to the present disclosure, if RI, CQI, and encoder codewords (e.g., the output of an encoder representing channel information) are all sent in the same PUCCH, one or more mechanisms may be implemented to ensure the same understanding of the UCI payload size between the UE and the base station. Because the bit width for precoding information (e.g., PMI) may depend on RI, in some embodiments, precoding information may be reported separately in a PUCCH other than the PUCCH carrying RI. In some other embodiments, the UCI payload may be appended with one or more elements (e.g., zeros), where the number of elements may depend on the indicated rank. For example, if Nmax is the maximum UCI payload size over all possible supported ranks, then for a given rank v to be reported with a payload size N(v), the UE may max-v Zeros are appended to the UCI payload.

[0396] Processing time

[0397] In some embodiments (e.g., as part of life cycle management (LCM)), the base station may instruct the UE to update the currently active model (e.g., fine-tune a new data set), switch to a new model, activate a new model, disable a model, etc. When updating the model, if the UE is instructed to update the encoder model based on an online training set (which may be collected via channel estimation, RRC configuration, MAC-CE activation, etc.), the UE may use (e.g., require) a minimal amount of time to update the model. The UE may use (e.g., require) a minimal amount of time regardless of whether the UE will share its updated model with the base station.

[0398] In some embodiments, if the online training set is collected by the UE via online channel estimation, it may not be expected that the UE updates the model before an amount of time has expired, which may be expressed as, for example, a number of symbols from the end of the last symbol of the latest CSI-RS used for the online training set (which may be referred to as, for example, N_(AIML,upadte) symbols). In some embodiments, if the UE is configured to report an updated model to the gNB, it may not be expected that the UE reports the model to the gNB earlier than an amount of time, which may be expressed as, for example, a number of symbols from the end of the last symbol of the latest CSI-RS used in the training set (which may be referred to as, for example, N_(AIML,report) symbols).

[0399] In some embodiments, if a training set is RRC configured to the UE, it may not be expected that the UE updates or updates and reports the encoder earlier than a certain amount of time, which may be expressed as, for example, a number of symbols from the last symbol that may have triggered the update command (which may be referred to as, for example, N symbols).

[0400] In some embodiments, when the UE is instructed to switch to a new model, a processing time similar to the processing time provided for updating the model may be provided to the UE. For example, if the UE receives a command to switch to a new model (e.g., via DCI, MAC CE, RRC, etc.), it may not be expected that the UE switches to the model (which may involve activating the model) before a time period, which may be represented as a number of symbols (which may be referred to as, for example, N_(AIML, switch) symbols) from the end of the last symbol at which the switch command may be delivered to the UE (e.g., the last symbol of the PDCCH).

[0401] In some embodiments, a similar processing time may be provided to the UE for activating a model. For example, if the UE receives a command to activate and / or deactivate a model (e.g., via DCI, MAC CE, RRC, etc.), it may not be expected that the UE activates and / or deactivates the model before a time period, which may be represented as a number of symbols (which may be referred to as, for example, N_(AIML,activate) and / or N_(AIML,deactivate) symbols) from the end of the last symbol (e.g., the last symbol of the PDCCH) at which the activation and / or deactivation command may be transmitted to the UE.

[0402] In some embodiments, one or more of the above processing times may be described in time units based on a subcarrier spacing (SCS) parameter set of the cell transmitting the command and / or the cell for which the model may be tested (e.g., the cell on which the corresponding CSI report may be sent). In some embodiments, the SCS of the cell on which the model is active may also be considered. For example, a number of symbols (such as those mentioned in the above method) used to represent the amount of time may be described based on the minimum SCS among the SCS values.

[0403] In some embodiments, the UE may declare (e.g., based on a query of a base station) one or more capabilities for processing time, such as PDSCH and / or PUSCH processing time capabilities. Additionally or alternatively, the UE may also declare capabilities for any processing time disclosed herein. For example, in some embodiments, the UE may declare the minimum amount of time or minimum number of symbols for N_(AIML,upadte), N_(AIML,switch), and N_(AIML,activate) / N_(AIML,deactivate) that it can support (e.g., can use or need). For example, the UE may declare to the base station that it has a capability of 20 symbols for model switching (e.g., minimum processing time), which may indicate that the UE may not be expected to switch to the new model before 20 symbols after the end symbol of the switching command.

[0404] Using a separate model for reporting

[0405] Within the AI / ML channel information reporting architecture, if the UE reports a channel matrix, the gNB can infer the channel quality including rank, eigenvectors and / or eigenvalues, and therefore the SVD precoding matrix. However, within the AI / ML CSI reporting architecture, since the performance may depend on the UE downlink signal processing, including the UE capabilities regarding the supported ranks, the UE-specific precoding information (e.g., precoding matrix) calculation algorithm, the UE may also report one or more other values, such as RI and / or CQI. For purposes of illustration, some embodiments of an AI / ML (which may be referred to as ML) based reporting scheme may be described in the context of a system that performs AI / ML CSI reporting by reporting CSI matrices (e.g., channel matrices, precoding matrices, etc.) via an autoencoder, however, the principles are not limited to the use of an autoencoder.

[0406] In some embodiments, in addition to CSI matrix reporting, the UE may also report RI and CQI via existing reporting or AI / ML architecture. Under the AI / ML architecture, the UE may be configured with subbands to report the CSI matrix for each subband or subset of subbands according to a bitmap. Some examples of reporting RI, PMI, and / or CQI are described below.

[0407] In the NR CSI report, the UE may report only one RI for all subbands specified in the CSI reporting configuration. The same approach may also be used for some embodiments of the AI / MLCSI reporting architecture. In some additional embodiments, the one or more RI values ​​corresponding to one or more (e.g., each) subbands may be reported by concatenating the one or more RI values ​​corresponding to a vector and compressing the vector, and depending on the implementation details, the compression may be similar to the CSI matrix compression. Since the relevant properties of the RI vector may be different from those of the CSI matrix, an ML model (e.g., a network) that may be different from the ML model used for CSI matrix compression may be used to compress the RI vector. For example, the ML model used for RI reporting may have input and / or output dimensions, weights, etc. that may be customized for compressed RI vectors.

[0408] The AI / ML architecture based CSI matrix reporting may be performed according to any of the schemes disclosed herein. In some embodiments, the UE may be configured to report a CSI matrix for each subband or a subset of subbands according to a CSI reporting configuration.

[0409] CQI reporting based on the AI / ML architecture may be performed according to any of the schemes disclosed herein.

[0410] In some embodiments, if the UE is configured to report one or more of the three quantities discussed above (e.g., RI, CSI, and / or CQI) separately in different PUCCHs, in order to reduce the input size of the second model (CSI matrix reporting model), the UE may report different quantities (e.g., three quantities) at different stages via PUCCH using time domain multiplexing (TDM). For example, the UE may report RI in the first PUCCH. The UE may report a CSI matrix in the second PUCCH with a dimension that may depend on the reported RI, for example, because the gNB may only be able to decode the second PUCCH after successfully decoding the first PUCCH. Depending on the implementation details, decoding of the first PUCCH and the third PUCCH may be performed at any time, but the second PUCCH may be decoded only after decoding the first PUCCH. One or more of these aspects may be based on the assumption that PUCCH decoding may involve a decoder model running an autoencoder at the gNB.

[0411] Lifecycle management based on model ID

[0412] According to some communication systems disclosed herein, an LCM scheme may be implemented based on a model ID. In some embodiments of a model ID-based LCM, one or more UE capabilities may be based on multiple models and / or functions. In some model ID-based LCM schemes, models may be identified and / or registered to the network. However, it may be redundant for the UE to identify two different models that are logically identical or similar and correspond to the same or similar physical models. Therefore, some embodiments may assume that each logical model identified by the base station has a different physical model. Alternatively, a logical model may correspond to multiple different physical models (for example, the UE may have a low-complexity model and a high-complexity model for a given artificial intelligence and / or machine learning feature and / or function). Once the model is identified, the UE may report to the base station which models it supports. For example, due to the nature of model identification and capability reporting, some embodiments may support relatively more models in a function-based LCM system and / or a model ID-based LCM.

[0413] However, one aspect that may be problematic is that there may be multiple physical models for one identified logical model. In some embodiments, this type of situation may be handled in the UE capability report, as the UE may be aware of the situation and may adjust the capability report to accommodate the situation.

[0414] In some embodiments, the UE capability report may include one or more details that enable the UE to report to the base station which models among the identified models the UE may support. In a reporting (e.g., signaling) scheme regarding such details, one part of the report (e.g., a first part) may include the UE vendor ID. Additionally or alternatively, another part of the report (e.g., a second part) may include one or more indexes of supported models identified by the UE vendor. For example, a bitmap of length N may be used (e.g., by the vendor) to identify N models for a base station, such that each of the N elements of the bitmap may correspond to an identified model (e.g., in ascending or descending order of the model ID). From the perspective of the UE and / or base station, once the vendor is identified, a local logical ID from 0 to N-1 may be assigned to the identified model. The UE may then use the bitmap to declare support for the identified model.

[0415] In addition to indicating one or more supported model IDs to the base station, the UE may also declare (e.g., as a capability report) the maximum number of simultaneously activated models that it can support (e.g., N active,max ). Then, the UE can expect that the base station will not activate more than N active,max Model.

[0416] In some embodiments, to accommodate the additional complexity that may result from running multiple physical models against the same identified logical model, the UE may also report a scalar value α for each supported model ID i i >1, where α can be identified by the network during model identification. The UE may not expect that more than N base stations can be activated simultaneously. active,max model, where the scalar value α can be considered i to count the number of activated models. Therefore, if the base station activation has the set In the model with the ID, the number of activated models can be calculated as:

[0417] Number of activated models = ∑ i∈S α i (Equation 8)

[0418] Among them, the UE can expect that the number is no greater than N active,max .

[0419] Some additional aspects of the present disclosure relate to meta information and / or coexistence with other RRC configurations. When a model is identified, a type of information referred to as meta information may be revealed to the network. In different embodiments, the meta information may include different types of information. For example, in some embodiments, the meta information may include some (e.g., all) information that may be used by the model for reasoning operations. In such an embodiment, the base station may activate a specific model that may be similar or equivalent to performing some (e.g., all) RRC configurations that may be used by the model for reasoning operations. When the base station activates the model ID, it may activate a specific CSI report configuration. In some NR implementations, the CSI report configuration may include information about subband configurations, codebook types, etc., and all of this information may be included in the meta information. Such an approach may reduce RRC overhead, but may reduce the flexibility of the base station because, for example, (1) some (e.g., each) functions and / or RRC configurations may be completed via model identification, and / or (2) unlike some NR implementations, the base station may not configure the meta information.

[0420] Some additional aspects of the present disclosure relate to model transmission and / or delivery. In some communication systems, model delivery (e.g., downloading a model from a server) may not be part of the specification. However, the specification may include model transmission (e.g., transmitting the model to a UE covered by the specification using signaling provided by the specification). One or more of the following techniques for transmitting a model to a UE may be implemented according to some embodiments of the present disclosure.

[0421] Some embodiments may transmit one or more weights of a known model structure. For example, to transmit the model to the UE, a structure of the model may be established between the UE and the base station (e.g., by specifying a pool of structures such as standard structures), and each structure may have an ID. The UE may declare its support for certain structures as UE capability signaling. To transmit the model, the base station may indicate the structure ID of the specified structure to the UE, and then send N weights using a downlink channel (e.g., PDSCH). The mapping between vectorized weights and network structures may be specified in the specification. The base station may also indicate the model ID when transmitting the network weights to the UE.

[0422] Some embodiments may transmit the model as a model ID. For example, a UE vendor and a base station vendor may develop multiple models offline and store them on a server and / or in a registry. The base station may then indicate the applicable model ID to the UE. The UE may then retrieve the model from the server using the model ID.

[0423] Some embodiments may implement model transmission with an indication of a dataset ID and / or an additional ID (e.g., an additional condition ID). For example, the base station may indicate the dataset ID or the additional condition ID to the UE. The UE may know which model to use based on the mapping between the trained model and the dataset ID in the training phase. Therefore, in some embodiments, one or more IDs intended for different purposes may be used as the model ID.

[0424] Any of the techniques discussed above related to model ID based LCM may be applied in the context of CSI compression, CSI prediction, etc.

[0425] Some additional aspects of the present disclosure relate to a model identification scheme for CSI feedback in a system using non-coherent joint transmission (NCJT). For example, in the case of NCJT CSI reporting for a UE that receives transmissions from multiple transmission reception points (TRPs), the UE may report CSI for different TRP (e.g., base station) transmission assumptions. The UE may be configured to report X CSIs (where X=0, 1, 2) associated with a single TRP measurement assumption and one CSI associated with the NCJT measurement assumption. These different assumptions may generate different precoding matrices to be reported. Since the UE may report different precoding matrices for different transmission schemes, the AIML encoder / decoder pair (e.g., autoencoder) may be different for different assumptions. One or more of the following techniques for NCJT CSI reporting may be implemented in accordance with some embodiments of the present disclosure.

[0426] Regarding model ID, if the NCJT CSI report corresponds to M different hypotheses, the corresponding model ID may include M individual model IDs (e.g., pairing IDs) for M different autoencoders. Therefore, the model ID can be represented in the form of an M-tuple (ID#1, ..., ID#M). In addition or alternatively, it can be assumed that the UE can use the same model for any sTRP CSI. For example, if the UE reports CSI for three hypotheses: sTRP#1, sTRP#2, and NCJT transmission, the ID can take the form of a pair (ID of the model for sTRP, ID of the model for NCJT).

[0427] In some embodiments, LCM for a model for NCJT CSI feedback may be implemented in a manner similar to that for a single TRP report, with the following differences: (1) for model activation, switching, deactivation, and / or other model ID-based LCMs, the base station may indicate the model in the form of an NCJT (e.g., a tuple); and / or (2) the CSI report configuration for NCJT may include an IE indicating the model ID in the form of a tuple.

[0428] CSI payload size based on number of models and / or features

[0429] Using a relatively small number of encoder models (e.g., a single model or only a few models) to report different numbers of subbands and / or different CSI payload sizes (e.g., the output of the encoder) can help keep hardware complexity relatively low at the UE. Thus, a single CSI generation model can be implemented with a relatively large number of output features to accommodate the maximum number of expected features that can be used. However, in some cases, a relatively small number of subbands may be reported, and therefore, the actual number M of features sent may be less than N, resulting in a compression rate that may be unnecessarily low.

[0430] In this case, according to embodiments of the present disclosure, a common understanding of which M features are sent and which NM features are punctured may be provided to the UE and the base station. This may enable the base station to correctly apply the received features to the N corresponding input features of the CSI reconstruction model. The generation model at the UE and the reconstruction model at the base station may be trained so that the punctured features are set to fixed values ​​(e.g., zero) at the reconstruction model.

[0431] During inference operation, the base station may indicate to the UE (e.g., statically and / or dynamically, via RRC, DCI, MAC CE signaling, etc.) which feature set the UE may send. In an exemplary embodiment (which may be relatively efficient), an order may be established between feature sets (e.g., M features), and the base station and / or UE may indicate the number of features to send (e.g., N).

[0432] In some embodiments, once the model pair is activated, the base station may provide information about the CSI payload to the UE (eg, via RRC, DCI, MACCE signaling, etc.). The payload information may be provided, for example, using an index.

[0433] One type of CSI payload information may include, for example, a CSI payload size expressed in terms of the number of features. For example, if a generation mode (e.g., an encoder) generates a CSI output codeword of length N, the base station may indicate an actual CSI payload of size M, where M may be less than, equal to, or greater than N. In such an embodiment, the UE may only send M features from the encoder output (e.g., after appropriate quantization).

[0434] Another type of CSI payload information may include a set of transmitted features. For example, an index may determine which subset of M features from a set of N features the UE may transmit. In some embodiments, the index may correspond to a bitmap of length N, where the first binary value (e.g., 1 or 0) may indicate that the feature is transmitted, and the second binary value (e.g., 0 or 1) may indicate that the feature is not transmitted. In some embodiments, the UE may transmit M features in ascending order of the indexes of the M features. Additionally or alternatively, the starting index and length (number) of the transmitted features may be used to indicate the set of transmitted features. Depending on the implementation details, this may simplify the implementation. The CSI payload information may indicate, for example, a starting index of 10 and a length of 20 among a CSI codeword having N=30 features. In this case, where the CSI codeword may include a feature set (x_1, x_2, ..., x_10, x_11, ..., x_29, x_30), the UE may transmit (x_10, ..., x_29) after quantizing the transmitted features.

[0435] Another type of CSI payload information may include one or more of quantization method, granularity, etc., which may indicate to the UE how to quantize the M transmitted CSI codeword characteristics.

[0436] Codebook Subset Constraint

[0437] As described above, the UE may provide CSI feedback to the base station using a codebook to indicate the PMI selected by the UE based on the DL channel conditions measured in response to the reference signal. The base station may then use the PMI for beamforming in the DL channel. However, in some cases, the use of a specific PMI may result in beamforming that may cause interference (e.g., with a neighboring cell). Therefore, some communication systems may implement a codebook subset restriction (CBSR) feature that may restrict the UE to use a specific portion of the codebook for channel estimation and / or feedback. For example, a base station may restrict the UE to report a specific PMI by setting a specific basis vector coefficient to zero (e.g., using a bitmap). Therefore, the UE may select a beam and / or coefficient in an unconstrained set, thereby ensuring that the UE can only report a PMI that is orthogonal to the constrained PMI (e.g., does not interfere with the constrained PMI). However, in a system that uses artificial intelligence and / or machine learning to generate and / or compress CSI, there may be no basis vectors for the precoder to be reported. Therefore, the UE may not be able to determine the constrained PMI.

[0438] Constrained Subspace Information for CSI Feedback Using AI / ML

[0439] In a wireless communication system according to the present disclosure, constraint information related to a channel may be provided to a UE. This may enable the UE to ensure that a precoder or other CSI feedback determined using artificial intelligence and / or machine learning may result in a beamforming that is orthogonal (e.g., non-interfering) to a constrained subspace for the channel.

[0440] For example, in a system where a linear transform (e.g., an angular delay transform) is used to determine a precoder from the output of a decoder model, the base station can configure the UE with a constrained subspace and / or decoder output value (e.g., a coefficient that can be zero at a specific index). In an embodiment where the decoder model is shared with the UE, the UE can generate CSI that causes the decoder output to be zero at the specified index. In an embodiment where the decoder model is not shared with the UE, information can be provided to the UE to ensure that the encoder input and the decoder output are of the same type and / or format. The UE can then apply a subspace constraint (e.g., a specific index corresponding to a zero coefficient) to the encoder input, which may result in the decoder output having a coefficient that is zero at the corresponding index.

[0441] As another example, in a system that uses the output of a decoder model as CSI (e.g., the decoder model generates a precoder matrix as output), the base station can configure the UE with a constrained subspace. In an embodiment where the decoder model is shared with the UE, the UE can ensure that the decoder output is orthogonal to the constrained subspace. In an embodiment where the decoder model is not shared with the UE, the UE can ensure that the encoder input is orthogonal to the constrained subspace, and thus can ensure that the decoder output is sufficiently orthogonal to the constrained subspace.

[0442] As another example, the system can use information about the constrained subspace as an input to the encoder model (e.g., an explicit input). For example, the encoder model can be trained using the constrained subspace as an input to the model in a manner that causes the output of the corresponding decoder model to be within a subspace that is orthogonal to the constrained subspace.

[0443] For convenience, a precoder, PMI, coefficient, index, matrix, vector, subcarrier, subband, subspace, subchannel and / or any other element may be referred to and / or characterized as being orthogonal if they directly or indirectly contribute to orthogonal transmission. Orthogonality may be determined, for example, in the time domain, the spatial domain, the frequency domain and / or the code domain.

[0444] Fig.24 An embodiment of a communication system using channel constraint information and artificial intelligence and / or machine learning to provide channel information feedback according to the present disclosure is shown. Fig.24The system 2400 shown in the figure can be used to implement any of the devices, models, training schemes, etc. disclosed herein, or can be implemented together with any of the devices, models, training schemes, etc. disclosed herein. The system 2400 may include one or more elements (e.g., components, operations, etc.), which may be similar to Figure 4 , Figure 7 and / or those elements in the embodiments shown in other figures, where similar elements may be indicated by reference numerals ending with and / or containing the same numbers, letters, etc.

[0445] exist Fig.24 In the system 2400 shown in FIG. 2 , a first wireless device (e.g., UE) 2401 may receive a reference signal 2417 from a second wireless device (e.g., base station) 2402 via a channel 2415, for example, to enable the first wireless device 2401 to determine a channel condition of the channel 2415. The second wireless device 2402 may apply precoding information (e.g., precoding matrix) 2445 to some transmissions such as PDCCH, PDSCH, etc. and some reference signals such as CSI-RS.

[0446] The first wireless device 2401 may perform channel measurements based on the reference signal 2417 and the channel conditions, e.g., generate channel information (e.g., channel estimates, channel matrices, precoding information, etc.) 2405, which it may send to the second wireless device 2402, in the form of a representation 2407, which it may generate using the generation model 2403. The first wireless device 2401 may send the representation 2407 to the second wireless device 2402, e.g., using another channel 2416 (e.g., an uplink channel), a signal, etc.

[0447] The first wireless device 2401 may also receive constraint information 2461 related to the channel 2415 from the second wireless device 2402. The constraint information 2461 may enable the first wireless device 2401 to ensure that the channel information 2405 generated by it can comply with one or more channel constraint requirements indicated by the constraint information 2461. For example, the constraint information 2461 may include information about a constrained subspace for precoding information. The CSI determination logic 2452 at the first wireless device 2401 may implement one or more schemes as described in more detail below to ensure that the precoding information generated by the first wireless device 2401 is orthogonal to the constrained subspace. Thus, the CSI determination logic 2452 can use the constraint information 2461 to interact with the channel information 2405, the generation model 2403, and / or the representation 2407 to ensure that the representation 2407 is reconstructed 2406 (e.g., precoding information 2445, such as a precoding matrix) after being applied to the reconstruction model 2404 at the second wireless device 2402, and the second wireless device 2402 can use the reconstruction 2406 to send a transmission over the channel 2415 that is orthogonal to the constrained subspace indicated by the constraint information 2461.

[0448] In some embodiments, the generation model 2403 and the reconstruction model 2404 may be trained (e.g., in an autoencoder configuration) to jointly compress and / or decompress channel information using any training scheme (including one or more joint training and / or deployment frameworks disclosed herein).

[0449] Fig.25 An example embodiment of a communication system in which a decoder model is shared with a UE configured with a constrained subspace according to the present disclosure is shown. Fig.25 The system 2500 shown in FIG. 2 can be used to implement, for example, Fig.24 , where similar elements may be indicated by reference numerals ending with and / or containing the same numbers, letters, etc. For example, referring to Fig.25 For illustrative purposes, a first communication device may be implemented using UE 2501, a second communication device may be implemented using base station 2502, constraint information may be implemented using constrained subspace 2561, CSI determination logic may be implemented using precoding determination logic 2552, channel information may be implemented using precoding information 2505, a generation model may be implemented using encoder model 2503 (which may also be referred to as an encoder), and / or a reconstruction model may be implemented using decoder model 2504 (which may also be referred to as a decoder).

[0450] In various embodiments, the constrained subspace 2561 can be represented, for example, by a matrix, a column space of a matrix, a vector, an indication of one or more decoder outputs that can be set to a particular value (e.g., zero) (e.g., using one or more indices), etc. and / or one or more of multiple and / or combinations of the above.

[0451] In various embodiments, the precoding information 2505 can be represented, for example, by a precoder (e.g., a precoding matrix, vector, etc.), a set of coefficients from which the final (e.g., reconstructed) precoding information 2506 can be calculated (e.g., based on an angular delay transform and / or any other linear transform), etc., and / or one or more of multiple and / or combinations of the above.

[0452] UE 2501 may be configured with a constrained subspace 2561, for example, via RRC, DCI, MAC CE signaling, etc.

[0453] The encoder model 2503 and decoder model 2504 can be trained (e.g., in an autoencoder configuration) to jointly compress and / or decompress the precoded information 2505 using any training scheme (including one or more joint training and / or deployment frameworks disclosed herein).

[0454] The sharing logic 2544 at the base station 2502 can share the decoder model 2504 by transmitting information about the model size, dimensions, weights, etc. to the UE 2501, and the UE 2501 can use the transmitted information to implement the shared model 2504A. The sharing logic 2544 can transmit the information about the decoder model 2504 using wireless transmission, for example.

[0455] Because UE 2501 has access to a shared version 2504A of the decoder model 2504 used by base station 2502, UE 2501 can observe a version 2506A of the reconstructed precoding information 2506 that can be generated at base station 2502. In such an embodiment, and depending on implementation details, few constraints, if any, may be imposed on the inputs that UE 2501 may apply to encoder model 2503. For example, in some embodiments, as long as the inputs applied to encoder model 2503 cause the reconstructed precoding information 2506A generated at the shared decoder model 2504A to be orthogonal to constrained subspace 2561, the reconstructed precoding information 2506 generated by the decoder model 2504 at base station 2502 should also be orthogonal to constrained subspace 2561.

[0456] Thus, in some embodiments, the precoding determination logic 2552 is free to implement any scheme for determining the precoding information 2505, and the resulting representation 2507 generated by the encoder model 2503 and sent to the base station 2502 (e.g., using an uplink channel, signal, etc. 2516) should cause the reconstructed precoding information 2546 at the base station 2502 to be orthogonal to the constrained subspace 2561.

[0457] Fig.26 An example embodiment of a communication system in which a decoder model is not shared with a UE configured with a constrained subspace according to the present disclosure is shown. Fig.26 The system 2600 shown in FIG. 2 can be used to implement, for example, Fig.24 , where similar elements may be indicated by reference numerals ending with and / or including the same numbers, letters, etc. For example, for illustrative purposes, a reference numeral similar to Fig.25 UE 2601, base station 2602, constrained subspace 2661, precoding determination logic 2652, precoding information 2605, encoder model 2603, and / or decoder model 2604 are implemented in the manner of those in the embodiments shown in . In addition, the encoder model 2603 and the decoder model 2604 can be trained (for example, in an autoencoder configuration) to jointly compress and / or decompress the precoding information 2605 using any training scheme (including one or more joint training and / or deployment frameworks disclosed herein).

[0458] However, in Fig.26 In the system 2600 shown in , since the decoder model 2604 may not be shared with the UE 2601, the precoding determination logic 2652 may not be able to determine (e.g., directly using a shared model) what kind of reconstruction precoding information 2606 the decoder model 2604 at the base station 2602 can generate in response to one or more specific inputs applied to the encoder model 2603.

[0459] Fig.26 The system 2600 shown in can implement a subspace constraint scheme, where one or more inputs 2663 of the encoder model 2603 can be coordinated with one or more outputs 2664 of the decoder model 2604. The scheme can then develop joint training of the encoder model 2603 and the decoder model 2604 (e.g., in an autoencoder configuration) to ensure that the output 2664 of the decoder model 2604 is orthogonal to the constrained subspace 2661.

[0460] For example, the base station 2602 may send information about the configuration of the output 2664 of the decoder 2604 to the UE 2601. The precoding determination logic 2652 at the UE 2601 may use this information to apply an input 2663 having the same or similar configuration as the output 2664 of the decoder 2604 to the encoder 2603. If the precoding determination logic 2652 selects an input 2663 of the encoder 2603 that is orthogonal to the constrained subspace 2661, joint training with the encoder model 2603 and the decoder model 2604 may ensure that the output 2664 of the decoder 2604 is also orthogonal to the constrained subspace 2661. Depending on the implementation details, this may help overcome the lack of a shared decoder at the UE 2601.

[0461] In some embodiments, if the output 2664 of the decoder 2604 is within an acceptable error amount of the constrained subspace 2661, within an acceptable error amount of the input 2663 of the encoder 2603 which is orthogonal to the constrained subspace 2661, etc., then the output 2664 of the decoder 2604 can be considered to be orthogonal to the constrained subspace 2661.

[0462] In some embodiments, information about the configuration of the output 2664 of the decoder 2604 can be sent to the UE 2601 via RRC, DCI, MAC CE signaling, etc.

[0463] In some example embodiments described below, the type of CSI may refer to the nature (e.g., one or more characteristics) of CSI information used as input to a generation model (e.g., an encoder) and / or output from a reconstruction model (e.g., a decoder). For example, the CSI may be implemented as a precoder type CSI, a channel matrix type CSI, etc. In addition, the format may refer to the dimensions in the quantity and / or matrix, the order of the elements, etc. For example, if the input to an encoder and the output from the corresponding decoder have the same format, it may be indicated that they are matrices of the same dimension, where the elements (i, j) may represent the same entity, for example, both are the second feature of the latent space, or both are the second elements of the precoder matrix for the same subband.

[0464] Report CSI in the angular delay domain or using a linear transform

[0465] In some embodiments, the output of the decoder may be represents the target CSI precoder (which can be based on the output of the decoder can be calculated by Represented by, and the constrained subspace can be represented by the matrix S. Precoder Can be Therefore, the target CSI precoder orthogonal to the constrained subspace S is generated. It may involve determining one or more inputs to be applied to an encoder-decoder pair (e.g., an autoencoder) that may cause the decoder to generate a precoder that results in orthogonal to the constrained subspace S. (For example, ) Furthermore, as shown below, in some embodiments, in order to generate Precoder Decoder Output The specific coefficient of must be zero.

[0466] In some embodiments, the output of the decoder may be represented as one or more coefficients of a set of basis vectors (e.g., as opposed to a decoder that may generate the target CSI directly in the form of a precoder or channel matrix). The following analysis may be applied to such embodiments to generate a precoder that results in orthogonal to the constrained subspace S. (For example, )

[0467] For each layer, the spatial-frequency domain CSI matrix, channel or precoder matrix Angular delay transformation Can be defined as:

[0468] H=F c H a F d (Equation 9)

[0469] Among them, F c YesN tx ×N tx DFT matrix, and F d YesN sb ×N sb DFT matrix. This can be done by removing The appropriate zero rows and columns of H M×N Rewrite, where F c It is M×N tx DFT submatrix, and F d YesN sb ×N DFT submatrices, where M <N tx and N <N sb The following analysis can be performed without loss of generality by assuming

[0470] As mentioned above, the decoder output can be obtained by Therefore, It can represent the set of coefficients from which the final output CSI can be calculated (for example, it can be obtained from Calculate the precoder matrix The calculation may be based on, for example, a linear transformation with certain basis vectors.

[0471] The CSI matrix can be a precoder matrix It can be obtained from the output of the decoder, for example, via an angular delay transform like:

[0472]

[0473] or any other linear transformation (where F c and F d It can be obtained by using a basis matrix (e.g., a DFT matrix). Assume Need and by matrix Given that the constrained subspaces are orthogonal, the constraints can be applied directly to the output of the decoder, for example by setting certain coefficients to certain values ​​(eg, zero).

[0474] As mentioned above, the precoder Can be Substituting into Eq. 10 To get Multiply And note that S T F c may be a binary matrix with more columns than rows (which may be called a fat matrix), outputting Certain coefficients of may be set (e.g., forced) to zero to satisfy the orthogonality constraint. Thus, the base station may indicate to the UE one or more indices corresponding to the coefficients of the output of the decoder, where the coefficients of the output of the decoder should be zero to ensure that the precoder Orthogonality to the constrained subspace.

[0475] In some example embodiments where CSI may be reported in the angular delay domain or using a linear transform, the decoder used by the base station may not be available to the UE. Fig.26The subspace constraint scheme shown and / or described can implement such an example embodiment. In such an embodiment, the base station 2602 can configure the UE 2601 with a constrained subspace 2661, which is implemented with a constrained subspace vector set or a precoder vector set (e.g., a DFT beam) in the form of a matrix S. Since the base station 2602 can know the relationship between the output of the decoder 2604 and the precoder matrix, the base station 2602 can also configure the UE 2601 with an index set corresponding to the coefficients for which the output of the decoder should be zero.

[0476] One or more of the following techniques may then be used to implement a zero condition at the output of the decoder 2604. (1) The UE 2601 may be relied upon to report a CSI codeword (e.g., a representation 2607 output by the encoder 2603) that causes the decoder output to satisfy the zero condition. Thus, the UE 2601 may select any input type available with its implementation as long as it satisfies the decoder output requirements. (2) Information about the configuration of the output 2664 of the decoder 2604 may be provided to the UE. The precoding determination logic 2652 at the UE 2601 may use this information to ensure that it uses an input 2663 for the encoder 2603 that has the same configuration (e.g., type, format, etc.) as the output 2664 of the decoder 2604. Thus, the output of the decoder 2604 may have the same dimensions as the input 2663 of the encoder 2603. The UE 2601 may force the value of the encoder input at the indicated / configured index corresponding to the coefficient for which the output of the decoder should be zero to be zero. Since the decoder output can be a reconstruction (e.g., reconstruction 2606) of the encoder input (e.g., precoded information 2605) with a certain reconstruction error, it can generally be expected that the zero condition applied to the input of encoder 2603 is replicated at the output of decoder 2604, up to the certain error.

[0477] In some other example embodiments where CSI may be reported in the angular delay domain or using a linear transform, the decoder used by the base station may be shared with the UE. Fig.25 The subspace constraint scheme shown and / or described implements such an embodiment. In such an embodiment, the base station 2502 can configure the UE 2501 with a constrained subspace 2561, which is implemented with a constrained subspace vector set or a precoder vector set (e.g., DFT beam) in the form of a matrix S. Since the base station 2502 can know the output of the decoder 2504 With the precoder matrix , so the base station 2502 can also configure the UE 2501 with an index set corresponding to the coefficient at which the output of the decoder should be zero.

[0478] One or more of the following techniques may then be used to implement a zero condition at the output of the decoder 2504. (1) The UE 2501 may be relied upon to report CSI codewords that cause the decoder output to satisfy the zero condition (e.g., representation 2507 output by the encoder 2503). However, because the UE 2501 has access to a shared decoder model 2504A, there may be no constraints on the configuration of the inputs applied to the encoder 2503. Thus, the UE 2501 may select any input available with its implementation as long as the input causes the output generated at the decoder 2504 (or reference decoder) to satisfy the zero condition at the desired index. (2) The UE 2501 may train its encoder 2503 so that the decoder output has zero values ​​for the coefficients at the indicated indexes. In such a training framework, the encoder 2503 at the UE 2501 may be trained using information about the constrained subspace (e.g., the subspace itself and / or one or more constrained coefficients, e.g., a binary vector having zeros at the decoder output and / or the constrained index of the constrained subspace).

[0479] In some alternative embodiments, it may not be desirable for the UE to ensure orthogonality of the CSI it may report to the base station with the constrained subspace. For example, the UE may report the CSI in a normal manner by sending an encoded representation of the reported CSI as a codeword that causes a decoder at the base station to generate a reconstructed output When the base station outputs Construct the precoder matrix When the base station may force one or more coefficients of the decoder output to zero to enforce orthogonality with the constrained subspace (the original decoder output at those indices may or may not be zero). Therefore, it is not expected that the UE reports a CSI codeword at the decoder output that results in a zero value at the required index. In such an embodiment, the base station may or may not indicate the index to the UE.

[0480] Reporting CSI in the space-frequency domain

[0481] In the case where CSI feedback (eg, precoder matrix ), it may not be feasible or convenient to describe the output encoder via a vector space. In such embodiments, the following analysis can implement a scheme for ensuring that the reported precoder matrix is ​​orthogonal to the constrained subspace S (e.g. ).

[0482] The output of the decoder can be expressed as where e can represent the error vector (e.g., the magnitude of the error vector). The constrained subspace can be represented by the matrix Column space representation of .

[0483] In some embodiments of the orthogonality scheme, the UE may project one or more precoder vectors into a space orthogonal to the constrained subspace such that w T The one or more precoder vectors are derived using S=0. The following equation can be used to confirm that this property remains unchanged at the output:

[0484]

[0485] In some embodiments, it can be expected that an encoder-decoder pair (e.g., an autoencoder) performs compression and / or decompression with a relatively small mean square error (MSE). Therefore, this property can also remain unchanged at the output, which may have an error as follows:

[0486] Where i = 1, ..., Z (Equation 12)

[0487] The MSE can be reduced (e.g., minimized) so that the right side can be less than a given ∈:

[0488]

[0489] Assume that the output precoder vector is normalized for power (e.g., or N tx ), when the reconstructed precoder vector and the constrained subspace basis vectors s i When (i=1,…,Z) are orthogonal, the inner product can be reduced (eg, minimized).

[0490] Therefore, some embodiments may generate CSI feedback (e.g., precoder matrix φ) that is orthogonal to the constrained subspace by ensuring orthogonality at the input of the encoder and allowing some relatively small error at the output of the corresponding decoder. ).

[0491] In the case where CSI can be reported in the space-frequency domain (e.g., as a precoder matrix ), the decoder used by the base station may not be available to the UE. Fig.26 The subspace constraint scheme shown and / or described implements such an embodiment. In such an embodiment, the base station 2602 can configure the UE 2601 with a constrained subspace 2661, which is implemented in the form of a matrix S with a constrained subspace vector set or a precoder vector set (e.g., a DFT beam).

[0492] One or more of the following techniques may then be used to ensure orthogonality of the reported precoder matrix to the constrained subspace 2661. (1) The UE 2601 may be relied upon to report the output of the decoder 2604 to satisfy the zero condition (e.g., ) CSI codeword (e.g., representation 2607 output by encoder 2603). Therefore, UE 2601 can select any input type available with its implementation as long as it meets the decoder output requirements. (2) Information indicating that the input of encoder 2603 and the output of decoder 2604 are configured as precoder vectors can be provided to the UE. UE 2601 can be relied upon to determine the input of encoder 2603 so that the output of the encoder (e.g., precoder matrix ) is orthogonal to the constrained subspace. In some embodiments, the UE 2601 may ensure that the residual orthogonality (or reconstruction) error at the output of the decoder is less than a certain value. For example, the UE may report CSI feedback such that the orthogonalization error

[0493] In the case where CSI can be reported in the space-frequency domain (e.g., as a precoder matrix ) In some other example embodiments, the decoder used by the base station can be shared with the UE. Fig.25 The subspace constraint scheme shown and / or described implements such an embodiment. In such an embodiment, the base station 2502 can configure the UE 2501 with a constrained subspace 2561, which is implemented with a constrained subspace vector set or a precoder vector set (e.g., DFT beam) in the form of a matrix S. Because the UE 2501 can access the shared decoder model 2504A, there may be no constraints on the configuration of the input applied to the encoder 2503. Therefore, the UE 2501 can select any input available with its implementation as long as the input causes the output generated at the decoder 2504 (e.g., precoder matrix ) is orthogonal to the constrained subspace. In some embodiments, UE 2501 can be relied upon to report CSI feedback so that the orthogonalization error

[0494] In the above about Fig.24 , Fig.25 and / or Fig.26In the described embodiments, one or more models may be trained without using information about the constrained subspace. Thus, the autoencoder may be trained without any assumptions about the constrained subspace. In such a framework, one or more models may be trained using examples of training data of different possibilities that may implicitly include subset constraints, but information about the constrained subspace may be omitted from the explicit training input. Nevertheless, in the inference phase, constraints may be ensured based on consistency between the training phase and the inference phase regarding how the UE determines encoder inputs for a given configured subset constraint.

[0495] Training and inference with constrained subspace information

[0496] In some embodiments, the constrained subspace information can be used as an input (e.g., explicit training input and / or explicit inference input) to a generative model (e.g., an encoder of an autoencoder pair of the model).

[0497] Fig. 27 An embodiment of a system according to the present disclosure using constrained subspace information as input to a model is shown. Fig. 27 The system 2700 shown in the figure can be used to implement any of the embodiments disclosed herein for generating and / or compressing CSI using artificial intelligence and / or machine learning or can be implemented with the embodiments, including those shown and / or described with respect to other figures, in which similar elements can be indicated by reference numerals ending with and / or containing the same numbers, letters, etc.

[0498] refer to Fig. 27 , the system 2700 may include a first node (node ​​A) having a generative model 2703 and a second node (node ​​B) having a reconstruction model 2704. The system 2700 is illustrated in an inference operation phase, where input data 2762 applied to the generative model (e.g., encoder) 2703 may include channel information (e.g., precoding information) 2705 and constraint information (e.g., constrained subspace information) 2761. However, in a training operation phase, the input data 2762 may include training data instead of the channel information 2705.

[0499] The generative model 2703 may generate a representation 2707 of the input data 2762, and the reconstruction model 2704 may use the representation 2707 to generate a reconstruction (e.g., precoding information) 2706 of the channel information 2705. Because the generative model 2703 and the reconstruction model 2704 may be trained using the constraint information 2761, the reconstruction 2706 of the channel information may be outside the constrained subspace indicated by the constraint information 2761. In some embodiments, the scheme shown in the diagram 2700 may be referred to and / or characterized as constraint-aware.

[0500] The constraint information 2761 may include, for example, the constrained subspace itself, one or more constraints on the coefficients of the basis vector space (e.g., indicating the indices of coefficients that should be zero), etc. In some embodiments, the channel information (e.g., precoding information) 2705 and / or reconstruction 2706 may be similar to those quantities in a system implemented without constraint-aware operation, but because the constraint information 2761 is included in the input data 2762 used for training and / or inference, the benefit is to ensure that the reconstruction 2706 (e.g., the reported precoder) is outside the constrained subspace.

[0501] In some embodiments, the generation model 2703 may include a quantizer for converting the representation 2707 into a quantized form (e.g., a bitstream) that can be sent via a communication channel. Similarly, in some embodiments, the reconstruction model 2704 may include a dequantizer that can convert the quantized representation 2707 (e.g., a bitstream) into a form that can be used to generate reconstruction training data 2712.

[0502] The generative model 2703 and the reconstruction model 2704 may be obtained in any manner, including using any of the frameworks described herein. For example, using a joint training framework, the generative model 2703 and the reconstruction model 2704 may be trained as a pair at node B, and node B may send the reconstruction model 2704 to node A. Other embodiments may obtain and / or train the generative model 2703 and the reconstruction model 2704 using a training framework with a reference model, a training framework with the latest shared values, or any other framework and / or technique.

[0503] In some embodiments, the loss may be calculated using labels that may include constraint information at the output of the reconstruction model 2704. Thus, it may be ensured that the labels (which may be referred to as ground truth labels) are in the constrained subspace.

[0504] In some embodiments, the loss function may be implemented with a penalty that may be introduced into the loss function when the output of the reconstruction model 2704 is not orthogonal to the constrained subspace determined by the constraint information 2761. In this embodiment, Fig. 27 The scheme described in may report acceptable or best CSI codewords (eg, output of an encoder and / or PMI index) within an allowed subspace (eg, outside a constrained subspace) for a given reconstruction model 2704 .

[0505] Additionally or alternatively, the loss function may be implemented as a conventional loss function that ensures the closeness of the output of the reconstructed model 2704 to the reference true generative model 2703. Since the closeness (e.g., in the case of MSE) may reduce the norm of the orthogonality error, it may cause the output to be within the allowed subspace (e.g., it may satisfy one or more orthogonality conditions (e.g., )).

[0506] In some embodiments, during the inference phase, the base station may configure the UE with constraint information (e.g., constrained subspace) 2761, which may be used (e.g., directly) as input to the generative model 2703. Then, using the constraint information 2761 for training and / or inference operations may enable the reconstruction (e.g., precoding information) 2706 to be within the allowed subspace.

[0507] In an embodiment where the reconstruction model 2704 is shared with the UE, the base station may configure the UE with a constrained subspace vector set or a precoder vector set (e.g., a DFT beam) in the form of a matrix S, and the type of output CSI may be indicated as a precoder vector. In this embodiment, the UE may be relied upon to report CSI codewords (e.g., representation 2707 output by the generation model 2703) such that the output of the reconstruction model 2704 is orthogonal to the constrained subspace. In some such embodiments, the UE 2701 may ensure that the residual orthogonality (or reconstruction) error at the output of the decoder is less than a specific value. For example, the UE may report CSI feedback such that the orthogonalization error

[0508] In an embodiment where the reconstruction model 2704 is not shared with the UE, the base station may configure the UE with a constrained subspace vector set or a precoder vector set (e.g., a DFT beam) in the form of a matrix S, and may then use one or more of the following techniques to ensure that the reported CSI (e.g., precoder) may be orthogonal to the constrained subspace. (1) The input configuration of the generation model 2703 and the output configuration of the reconstruction model 2704 may be indicated (or assumed by the UE in the absence of an indication received by the network) as a precoder. The UE may report the CSI codeword (e.g., a representation 2707 output by the generation model 2703) such that the precoder output by the reconstruction model 2704 is orthogonal to the constrained subspace. (2) The UE may assume that the input configuration of the generation model 2703 is the same or similar to the output configuration of the reconstruction model 2704 (e.g., precoder). The UE may determine that the input to the generation model 2703 is orthogonal to the constrained subspace, and therefore the output of the reconstruction model 2704 will also be orthogonal. (3) The UE may train a reference model to be used as the generation model 2703. The UE may then report the CSI codewords such that the corresponding reference used to reconstruct the model 2704 also satisfies the orthogonality condition.

[0509] Data collection for AIML models for CSI reporting

[0510] In some wireless communication systems using artificial intelligence models and / or machine learning models, the UE may collect data for different purposes (e.g., use cases), such as model training, model inference, model performance monitoring, etc. Examples of the type of data collected for one or more of these purposes may include channel information (e.g., a channel matrix that may be input to an encoder at the UE), CSI feedback (e.g., an output of an encoder model at the UE, which may be sent to a base station and used as an input to a decoder model at the base station), target CSI (e.g., a precoder or channel matrix that may be output from a decoder model at the base station), one or more gradients (e.g., which may be used for backpropagation using type 2 training), and the like.

[0511] Data collection at the UE may involve using UE processing resources to perform downlink channel measurements (e.g., channel estimation) based on CSI-RS signals sent by the base station. In one method of collecting data, the base station may configure the UE with a CSI reporting configuration, which the UE may generally use to measure, calculate, and report CSI to the base station.

[0512] However, in different situations, different devices (e.g., UE or base station) may need to collect different types of data. Therefore, when configured with a CSI reporting configuration, the UE may calculate one or more types of CSI that it may not need to collect and / or report to the base station, thereby placing an unnecessary burden on the processing resources of the UE. Another potential problem is that the UE may not be able to and / or is not configured to support one or more CSI-RS configurations for data collection. In addition, the base station may not be able to determine the UE's capabilities and / or configuration for data collection. Another potential problem is that the UE may waste resources when preparing an uplink channel for sending CSI that it does not need to report. Therefore, the UE may need to operate differently when collecting data than when reporting CSI according to a normal CSI reporting configuration. Another potential problem is that the base station may configure the UE with a CSI reporting configuration that may not generate one or more types of data that the UE may need to collect for the UE's purpose (e.g., training a local model, performing reasoning with a local model, etc.). In addition, the UE may not be able to determine whether or when the base station expects it to collect data.

[0513] To overcome these potential problems, systems and methods according to the present disclosure may enable a UE to establish a data collection session between the UE and a base station to request collection of one or more specific types of data and / or indicate capabilities for supporting one or more data collection configurations.

[0514] Fig.28 An embodiment of a communication system having data collection functionality for one or more models that can be used to provide channel information feedback according to the present disclosure is shown. System 2800 may include one or more elements (e.g., components, operations, etc.), which may be similar to Figure 4 , Figure 7 Elements in the embodiments shown in the and / or other figures, where similar elements may be indicated by reference numerals ending with and / or containing the same numbers, letters, etc.

[0515] exist Fig.28 In the system 2800 shown in FIG. 2 , a first wireless device (e.g., UE) 2801 may receive a reference signal 2817 from a second wireless device (e.g., base station) 2802 via a channel 2815, for example, to enable the first wireless device 2801 to determine a channel condition of the channel 2815. The second wireless device 2802 may apply precoding information (e.g., a precoding matrix) 2845 to some transmissions (such as PDCCH, PDSCH, etc.) and some reference signals (such as CSI-RS).

[0516] The first wireless device 2801 may perform channel measurements based on the reference signal 2817 and the channel conditions, e.g., to generate channel information (e.g., channel estimates, channel matrices, precoding information, etc.) 2805, which it may send to the second wireless device 2802, in the form of a representation 2807, which it may generate using the generation model 2803. The first wireless device 2801 may send the representation 2807 to the second wireless device 2802, e.g., using another channel (e.g., an uplink channel), a signal 2816, etc. The second wireless device 2802 may include a reconstruction model 2804, which may generate a reconstruction 2806 of the channel information 2805 using the reconstruction model 2804.

[0517] In some embodiments, the generation model 2803 and the reconstruction model 2804 may be trained (e.g., in an autoencoder configuration) to jointly compress and / or decompress the channel information 2805 using any training scheme (including one or more joint training and / or deployment frameworks disclosed herein).

[0518] The first wireless device 2801 and the second wireless device 2802 may include data collection logic 2866 and 2867, respectively, which may individually and / or collectively implement some or all of the data collection functions disclosed herein. For example, the data collection logic 2866 and / or 2867 may enable the first wireless device 2801 to establish a data collection session between the first wireless device and the second wireless device 2802. For example, the first wireless device 2801 may send a request to the second wireless device 2802 to start a data collection session. The request may be accompanied by information indicating one or more types of data to be collected, one or more reference signal configurations to be used for data collection, and the like. The second wireless device 2802 may send an indication to the first wireless device 2801 to start a data collection session in response to a request from the first wireless device 2801 or at the initiation of the second wireless device 2802.

[0519] As another example, the data collection logic 2866 and / or 2867 may enable the first wireless device 2801 to declare one or more data collection capabilities 2868 to the second wireless device 2802. In some embodiments, the first wireless device 2801 may declare capabilities to support one or more data collection configurations, which may include information about the purpose of the data collection, one or more types of data to be collected, and / or information about corresponding supported reference signal configurations that may be used to collect data.

[0520] Fig.29 An example embodiment of a communication system having data collection functionality for one or more models that may be used to provide channel information feedback according to the present disclosure is shown. Fig.29 The system 2900 shown in FIG. 2 may include one or more elements (eg, components, operations, etc.) that may be similar to Figure 4 , Figure 7 , Fig.28 and / or those elements in the embodiments shown in other figures, where similar elements may be indicated by reference numerals ending with and / or containing the same numbers, letters, etc. Fig.29 The system 2900 shown in the figure can be used, for example, to implement the data collection scheme for CSI compression in the space-frequency (SF) domain and / or the time-space-frequency (TSF) domain as described below, as well as other data collection schemes according to the present disclosure.

[0521] For the purpose of illustrating some example embodiments of data collection for CSI compression, Fig.29The system 2900 shown in is shown to have some example implementation details. Specifically, the first communication device can be implemented by UE 2901, the second communication device can be implemented by base station 2902, the reference signal can be implemented by CSI-RS signal 2917, the channel information can be implemented by channel matrix 2905, the generation model can be implemented by encoder model 2903 (which can also be called encoder), and the representation of channel information can be implemented by CSI feedback 2907. The reconstruction model can be implemented by decoder model 2904 (which can also be called decoder), and the reconstruction of channel information can be implemented by target CSI 2906 (for example, precoder matrix, one or more coefficients of precoder matrix that can be used to determine by combination of basis vectors, etc.). However, Fig.29 The principles of system 2900 described in are not limited to these or any other implementation details.

[0522] Some or all of the data collection functions disclosed herein may be implemented, at least in part, individually and / or collectively, by data collection logic 2966 and 2967 located at the UE 2901 and the base station 2902, respectively. In some embodiments, the UE 2901 may determine the target CSI 2906 (e.g., the output of the decoder model 2904 at the base station 2902) by using, for example, a decoder model, a locally trained decoder model, a shared copy of a reference decoder model, etc.

[0523] In some embodiments, UE 2901 may send a request 2969 to initiate a data collection session to base station 2902. Request 2969 may be accompanied by information indicating one or more types of data to be collected, one or more CSI-RS configurations to be used for data collection, etc. Request 2969 may be sent to base station 2902, for example, using a dedicated scheduling request (SR).

[0524] The base station 2902 may send an indication (e.g., command, trigger, signal, activation, etc.) 2970 to the UE 2901 for starting a data collection session in response to a request 2969 from the UE 2901 or at the initiation of the base station 2902. For example, if the UE 2901 has previously declared a capability to support one or more specific data collection operations (e.g., a capability to collect one or more specific types of data using one or more specific CSI-RS configurations), the base station 2902 may initiate a data collection session. The base station 2902 may send an indication 2970 to the UE 2901 for starting a data collection session using, for example, RRC, MAC CE, DCI, etc.

[0525] In some embodiments, the base station 2902 may configure the UE 2901 for data collection using a CSI reporting configuration that may be indicated as being used for data collection. When configured with such a CSI reporting configuration, the UE 2901 may avoid reporting some or all of the CSI to the base station 2902. Depending on implementation details, the UE 2901 may also avoid performing one or more calculations for CSI feedback 2907, target CSI 2906 (e.g., precoding information), etc. The CSI reporting configuration may be indicated as being used for data collection (e.g., a data collection use case), for example, using an information element (IE). Additionally or alternatively, the base station 2902 may configure the UE 2901 to collect data using one or more CSI-RS resource sets (e.g., when the UE 2901 is not expected to provide a CSI report). In such a configuration, the one or more CSI-RS resource sets may be coordinated (e.g., matched) with the purpose of the collected data, the type of data to be collected, etc.

[0526] In some embodiments, the UE 2901 may declare a capability 2968 for supporting one or more data collection configurations. For example, a data collection configuration (which may also be referred to as a feature group) may include information about the purpose of data collection (e.g., CSI compression (for training, inference, and / or monitoring), CSI prediction, etc.), one or more types of data to be collected (e.g., channel matrix 2905, CSI feedback 2907, and / or target CSI 2906), and / or information about corresponding supported CSI-RS configurations (e.g., CSI-RS subbands, number of CSI-RS ports, supported SCs, time domain window information for CSI compression, etc.) that may be used to collect data.

[0527] In some embodiments, various identifiers (IDs) may be used to indicate the type of information to be collected, use cases for the collected data, and the like. For example, each type of collected data (e.g., channel matrix 2905, CSI feedback 2907, target CSI 2906, etc.) may be indicated by a corresponding data type ID number. Thus, a data collection configuration (e.g., sent as part of a declaration of data collection capabilities 2968) may include a set of numbers indicating the type of data that the UE 2901 may be able to collect. As another example, each use case (e.g., purpose, scenario, etc.) of the collected data may be indicated by a corresponding use case ID. Examples of use cases that may have corresponding use case IDs may include CSI compression (for training, reasoning, and / or monitoring), CSI prediction, and the like. For example, when the base station 2902 initiates a data collection session, it may indicate to the UE 2901 which use case the data collection is for by indicating the use case ID. Additional examples of identifiers may include a vendor ID (e.g., used to indicate the vendor of base station 2902 to UE 2901, which may implement a vendor-level agreement regarding data collection configuration), a data set ID (e.g., used to enable different data sets of collected data based on different models, different distributions of channel matrices across subbands and time, different beam codebooks, different antenna configurations, etc.), a condition ID (e.g., used to indicate one or more additional conditions for data collection, such as CSI-RS resources, sets, reports, etc.).

[0528] According to the implementation details, Fig.28 and / or Fig.29 One or more of the features shown and / or described may enable efficient initiation and / or configuration of data collection sessions, may reduce or eliminate unnecessary computations and / or transmission of unneeded information, and / or may ensure that UE 2901 may collect data that the UE and / or base station 2902 may need for use with artificial intelligence and / or machine learning models.

[0529] Establishing a data collection session

[0530] Some aspects of the present disclosure provide techniques for establishing a data collection session be...

Claims

1. A device comprising: The receiver is configured as: using the channel to receive a reference signal; and receiving constraint information associated with the channel; The processing circuit is configured to: determining channel information based on the reference signal; and generating, using a machine learning model, a representation based on the channel information and the constraint information; and A transmitter is configured to transmit the representation.

2. The device according to claim 1, wherein: The processing circuit is configured to apply the constraint information to an input of the machine learning model.

3. The device of claim 2, wherein: The processing circuit is configured to: receiving input configuration information for the machine learning model; and The representation is generated based on the input configuration information.

4. The device of claim 1, wherein: The machine learning model is a generative model, and the processing circuit is configured to: inputting the representation into a reconstruction model; and The constraint information is applied to an output of the reconstructed model.

5. The device of claim 4, wherein: The reconstructed model includes a reference model.

6. The device of claim 1, wherein: The processing circuit is configured to train the machine learning model using the constraint information.

7. The device of claim 1, wherein: The constraint information indicates a constrained precoder subspace.

8. A device comprising: Receiver; as well as The processing circuit is configured to perform one or more operations, the operations comprising: Using the receiver, receiving a reference signal using a channel; receiving an instruction to collect data related to the channel for use in a machine learning model; processing the reference signal based on the indication; and Based on the indication and the processing, the data for the machine learning model is collected.

9. The apparatus of claim 8, further comprising a transmitter, and wherein: The processing circuit is further configured to perform one or more operations including sending, using the transmitter, a request for the indication to collect data.

10. The device of claim 9, wherein: The processing circuit is further configured to perform one or more operations including receiving, using the receiver, a reporting configuration for the collection.

11. The device of claim 10, wherein: The report configuration is adapted for the collection.

12. The apparatus of claim 8, wherein: The processing circuit is further configured to perform one or more operations including receiving, using the receiver, an indication of a type of the data.

13. The apparatus of claim 12, wherein: The type of the data includes one or more of channel estimation, channel information feedback, or target channel information.

14. A device comprising: Receiver; Transmitter; as well as The processing circuit is configured to perform one or more operations, the operations comprising: Using the receiver, receiving a reference signal using a channel; processing the reference signal to generate data related to the channel for use in a machine learning model; and Using the transmitter, an indication of the device's ability to collect the data for the machine learning model is sent.

15. The apparatus of claim 14, wherein: The indication is used to indicate the type of the data.

16. The device of claim 15, wherein: The indication is used to indicate the configuration of the reference signal.

17. The apparatus of claim 15, wherein: The indication is used to indicate a first characteristic group of the data and a first configuration of the reference signal corresponding to the first characteristic group.

18. The apparatus of claim 14, wherein: The indication is used to indicate the capability of collecting data for a joint transmission configuration.

19. The apparatus of claim 14, wherein: The receiver is at least one of one or more receivers, the reference signal is a first reference signal, the channel is a first channel, and the data associated with the first channel is first data; as well as The processing circuit is further configured to perform one or more operations, the one or more operations comprising: receiving, using at least one of the one or more receivers, a second reference signal using a second channel; processing the second reference signal based on the indication; and Based on the indication and the processing of the second reference signal, second data for the machine learning model is collected.

20. The apparatus of claim 19, wherein: The first data includes first channel information associated with the first channel; The second data includes second channel information associated with the second channel; as well as The processing circuit is further configured to perform one or more operations including collecting representations of the first channel information and the second channel information based on the machine learning model.

Images