Method and apparatus for wireless communication

Abstract

Provided are a method and apparatus for wireless communication. The method comprises: a first device performing CSI compression on first CSI by using a first model, so as to obtain second CSI; and the first device sending the second CSI to a second device, wherein the first CSI is related to a first channel, a second model corresponding to the second device is used for determining third CSI on the basis of the second CSI, and the third CSI is determined on the basis of a first adjustment amount of the first model and / or a second adjustment amount of the second model.

Description

Methods and apparatus for wireless communication Technical Field

[0001] This application relates to the field of communication technology, and more specifically, to a method and apparatus for wireless communication. Background Technology

[0002] In communication systems, channel state information (CSI) feedback helps network devices optimize transmission parameters to improve system performance. Introducing artificial intelligence (AI) technology into the CSI feedback process can achieve more efficient CSI prediction and compression. However, improving the accuracy of inference becomes a key technical challenge when using AI for CSI compression. Summary of the Invention

[0003] This application provides a method and apparatus for wireless communication. The various aspects related to the embodiments of this application are described below.

[0004] In a first aspect, a method for wireless communication is provided, comprising: a first device performing CSI compression on a first CSI using a first model to obtain a second CSI; the first device sending the second CSI to a second device; wherein the first CSI is associated with a first channel, and a second model corresponding to the second device is used to determine a third CSI based on the second CSI, the third CSI being determined based on a first adjustment amount of the first model and / or a second adjustment amount of the second model.

[0005] In a second aspect, a method for wireless communication is provided, comprising: a second device receiving a second CSI from a first device; a second model corresponding to the second device processing the second CSI to obtain a third CSI; wherein the second CSI is determined by CSI compression of the first CSI by a first model, the first model corresponding to the first device, the first CSI being related to a first channel, and the third CSI being determined based on a first adjustment amount of the first model and / or a second adjustment amount of the second model.

[0006] Thirdly, an apparatus for wireless communication is provided, the apparatus being a first device, the apparatus comprising: a processing unit configured to perform CSI compression on a first CSI using a first model to obtain a second CSI; and a transmitting unit configured to transmit the second CSI to a second device; wherein the first CSI is associated with a first channel, and a second model corresponding to the second device is configured to determine a third CSI based on the second CSI, the third CSI being determined based on a first adjustment amount of the first model and / or a second adjustment amount of the second model.

[0007] Fourthly, an apparatus for wireless communication is provided, the apparatus being a second device, the apparatus comprising: a receiving unit for receiving a second CSI from a first device; and a processing unit for processing the second CSI using a second model to obtain a third CSI; wherein the second CSI is determined by CSI compression of the first CSI using a first model, the first model corresponding to the first device, the first CSI being related to a first channel, and the third CSI being determined based on a first adjustment amount of the first model and / or a second adjustment amount of the second model.

[0008] Fifthly, a communication device is provided, including a memory and a processor, the memory for storing a program, and the processor for calling the program in the memory to perform the method as described in the first or second aspect.

[0009] A sixth aspect provides an apparatus including a processor for calling a program from memory to perform the method as described in the first or second aspect.

[0010] A seventh aspect provides a chip including a processor for calling a program from memory, causing a device on which the chip is mounted to perform the method as described in the first or second aspect.

[0011] Eighthly, a computer-readable storage medium is provided having a program stored thereon that causes a computer to perform the method as described in the first or second aspect.

[0012] Ninth aspect, a computer program product is provided, including a program that causes a computer to perform the method as described in the first or second aspect.

[0013] In a tenth aspect, a computer program is provided that causes a computer to perform the method as described in the first or second aspect.

[0014] In the embodiments of this application, when the first device and the second device perform CSI compression based on the first model and the second model, the CSI (third CSI) related to the first channel obtained by the second device is determined according to the first adjustment amount of the first model and / or the second adjustment amount of the second model. Therefore, in CSI compression based on a bilateral model, the CSI obtained by the second device is based on adjustments made to the model, which helps to improve the inference accuracy of CSI compression. Attached Figure Description

[0015] Figure 1 is a system architecture example diagram of a wireless communication system applicable to the embodiments of this application.

[0016] Figure 2 is a schematic diagram of the network architecture applicable to the embodiments of this application.

[0017] Figures 3A and 3B are schematic diagrams of the structure of the wireless protocol stack applicable to the embodiments of this application.

[0018] Figure 4 is a schematic diagram of neurons in a neural network applicable to the embodiments of this application.

[0019] Figure 5 is a schematic diagram of the neural network applicable to the embodiments of this application.

[0020] Figure 6 is a schematic diagram of a convolutional neural network applicable to the embodiments of this application.

[0021] Figure 7 is a schematic diagram of AI-based bilateral CSI compression as an alternative to the traditional method.

[0022] Figure 8 is a schematic diagram of the AI-based bilateral CSI compression model.

[0023] Figure 9 is a flowchart illustrating the training and inference phases of a bilateral CSI compressed model.

[0024] Figure 10 is a flowchart illustrating a method for wireless communication proposed in an embodiment of this application.

[0025] Figure 11 is a schematic diagram of one possible implementation of the method shown in Figure 10.

[0026] Figure 12 is a schematic diagram of another possible implementation of the method shown in Figure 10.

[0027] Figure 13 is a schematic diagram of another possible implementation of the method shown in Figure 10.

[0028] Figure 14 is a schematic diagram of a device for wireless communication provided in an embodiment of this application.

[0029] Figure 15 is a schematic diagram of another device for wireless communication provided in an embodiment of this application.

[0030] Figure 16 is a schematic diagram of the structure of a communication device provided in an embodiment of this application. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0032] Communication system architecture

[0033] Figure 1 is a system architecture example diagram of a wireless communication system 100 to which embodiments of this application can be applied. The wireless communication system 100 may include a network device 110 and a terminal device 120. The network device 110 may be a device that communicates with the terminal device 120. The network device 110 may provide communication coverage for a specific geographical area and may communicate with the terminal device 120 located within that coverage area.

[0034] Figure 1 exemplarily illustrates a network device and multiple terminal devices, such as terminal devices 120a to 120j in Figure 1. Optionally, the wireless communication system 100 may include multiple network devices, and each network device may include other numbers of terminal devices within its coverage area; this embodiment of the application does not limit this.

[0035] Optionally, the wireless communication system 100 may also include other network entities such as a network controller and a mobility management entity, which is not limited in this embodiment.

[0036] It should be understood that the technical solutions of the embodiments of this application can be applied to various communication systems, such as: 5th-generation (5G) systems or new radio (NR) systems, long-term evolution (LTE) systems, LTE frequency division duplex (FDD) systems, LTE time division duplex (TDD) systems, advanced long-term evolution (LTE-A) systems, enhanced 5G (5G advanced) systems, etc. The technical solutions provided in this application can also be applied to future communication systems, such as 6th-generation (6G) mobile communication systems, satellite communication systems, etc.

[0037] The communication system in this application embodiment can be applied to carrier aggregation (CA) scenarios, dual connectivity (DC) scenarios, and standalone (SA) network deployment scenarios.

[0038] The terminal device in this application embodiment can also be referred to as user equipment (UE), access terminal, user unit, user station, mobile station, mobile station (MS), mobile terminal (MT), remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent, or user device. The terminal device in this application embodiment can be a device that provides voice and / or data connectivity to a user, and can be used to connect people, objects, and machines, such as a handheld device with wireless connectivity, vehicle-mounted device, etc. The terminal device in the embodiments of this application may be a mobile phone, tablet computer, laptop computer, handheld computer, camera equipment, mobile internet device (MID), wearable device, virtual reality (VR) device, augmented reality (AR) device, wireless terminal in industrial control, wireless terminal in self-driving, wireless terminal in remote medical surgery, wireless terminal in smart grid, wireless terminal in transportation safety, wireless terminal in smart city, wireless terminal in smart home, etc. Optionally, the terminal device may be used to act as a base station. For example, the terminal device may act as a scheduling entity, providing sidelink signals between UEs in vehicle-to-everything (V2X) or device-to-device (D2D) connections. For example, cellular phones and cars communicate with each other using sidelink signals. Cellular phones and smart home devices can communicate without relaying communication signals through base stations.

[0039] The network device in this application embodiment can be a device for communicating with terminal devices. This network device can also be called an access network device or a radio access network device, such as a base station (BS). In this application embodiment, the network device can refer to a radio access network (RAN) node or a next-generation RAN (NG-RAN) node (or device) that connects user equipment to a wireless network. A base station can broadly encompass, or be replaced by, various names including: NodeB, evolved NodeB (eNB), next-generation NodeB (gNB), relay station, transmitting and receiving point (TRP), transmitting point (TP), master station (MeNB), secondary station (SeNB), multi-mode radio (MSR) node, home base station, network controller, access node, wireless node, access point (AP), transmission node, transceiver node, baseband unit (BBU), remote radio unit (RRU), active antenna unit (AAU), remote radio head (RRH), central unit (CU), distributed unit (DU), positioning node, etc. A base station can be a macro base station, micro base station, relay node, donor node, or similar, or a combination thereof. A base station can also refer to a communication module, modem, or chip installed within the aforementioned equipment or apparatus. Base stations can also be mobile switching centers, devices that perform base station functions in D2D, V2X, and machine-to-machine (M2M) communications, network (NW) side devices in 6G networks, and devices that perform base station functions in future communication systems. Base stations can support networks using the same or different access technologies. The embodiments of this application do not limit the specific technologies or device forms used in the network equipment.

[0040] Base stations can be fixed or mobile. For example, a helicopter or drone can be configured to act as a mobile base station, and one or more cells can move depending on the location of the mobile base station. In other examples, a helicopter or drone can be configured as a device to communicate with another base station.

[0041] In some deployments, the network device in this application embodiment may refer to a CU or a DU, or the network device may include both a CU and a DU. The gNB may also include an AAU.

[0042] Network devices and terminal devices can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; they can also be deployed on water; and they can also be deployed in the air on airplanes, balloons, and satellites. This application does not limit the scenario in which the network devices and terminal devices are located.

[0043] It should be understood that all or part of the functions of the communication device in this application can also be implemented by software functions running on hardware, or by virtualization functions instantiated on a platform (e.g., a cloud platform).

[0044] Figure 2 illustrates a schematic diagram of a network architecture 200 according to an embodiment of this application. This network architecture 200 describes the network architecture of a 5G NR / LTE / LTE-A system, which can also be referred to as a 5G system (5GS) / evolved packet system (EPS) network architecture. The network architecture 200 includes at least one of the following: network device 110, terminal device 120, 5G core network (5GC) / evolved packet core (EPC) 210, home subscriber server (HSS) / unified data management (UDM) 220, and Internet service 230. The network device and terminal device in Figure 2 are illustrated using RAN and UE as examples, respectively.

[0045] As shown in Figure 2, network device 110 provides user plane and control plane protocol termination to terminal device 120. Network device 110 is connected to 5GC / EPC 210 via an S1 / NG interface. 5GC / EPC 210 includes a mobility management entity (MME) / authentication management field (AMF) / session management function (SMF) 211, other MMEs / AMFs / SMFs 214, a service gateway (S-GW) / user plane function (UPF) 212, and a packet data network gateway (P-GW) / UPF 213. MME / AMF / SMF 211 is the control node that handles signaling between terminal device 120 and 5GC / EPC 210. Generally, MME / AMF / SMF 211 provides bearer and connection management. All user Internet Protocol (IP) packets are transmitted through the S-GW / UPF212, which is itself connected to the P-GW / UPF213. The P-GW provides UE IP address allocation and other functions. The P-GW / UPF213 is connected to Internet service 230. Internet service 230 includes operator-compliant Internet Protocol services, specifically including the Internet, intranet, IP multimedia subsystem (IMS), and packet-switched streaming services. It is evident that network architecture 200 provides packet-switched services; however, those skilled in the art will readily understand that the various concepts presented herein can be extended to networks providing circuit-switched services or other cellular networks.

[0046] Figures 3A and 3B respectively illustrate a schematic diagram of a wireless protocol stack structure according to an embodiment of this application. Figures 3A and 3B use a 5G wireless protocol stack as an example for illustration. The 5G wireless protocol stack is divided into two planes: the user plane (UP) protocol stack and the control plane (CP) protocol stack. The user plane protocol stack is the protocol suite used for user data transmission, and the control plane protocol stack is the protocol suite used for control signaling transmission in the 5G system. The specific names of each protocol stack layer are as follows:

[0047] As shown in Figure 3A, the user plane protocol stack includes, from top to bottom, the following layers: Service Data Adaptation Protocol (SDAP) layer, Packet Data Convergence Protocol (PDCP) layer, Radio Link Control (RLC) layer, Medium Access Control (MAC) layer, and Physical (PHY) layer.

[0048] As shown in Figure 3B, the control plane protocol stack includes, from top to bottom: non-access stratum (NAS); radio resource control (RRC) layer, PDCP layer, RLC layer, MAC layer, and PHY layer.

[0049] It should be understood that the different layers in the above protocol stack have different functions, and they work together through inter-layer interaction to achieve communication between terminal devices and network devices. With the development of artificial intelligence technology, AI-assisted computing has permeated the processing implementation methods of the above protocol stack. For example, the scheduling algorithm of the MAC layer and the encoding / decoding algorithm of the PHY layer can apply artificial intelligence algorithms to improve the performance of communication algorithms.

[0050] As an example, the wireless protocol architecture in Figures 3A and 3B is applicable to the terminal device in this application, such as a UE.

[0051] As an example, the wireless protocol architecture in Figures 3A and 3B is applicable to the network devices in this application, such as gNB.

[0052] It should be understood that the interpretation of the terminology in the embodiments of this application may refer to the TS36, TS37 and TS38 series of specifications of the 3rd generation partnership project (3GPP), but may also refer to the specifications of the Institute of Electrical and Electronics Engineers (IEEE).

[0053] To facilitate understanding, some related technical knowledge involved in the embodiments of this application is first introduced. The following related technologies are optional solutions and can be arbitrarily combined with the technical solutions of the embodiments of this application, all of which fall within the protection scope of the embodiments of this application. The embodiments of this application include at least some of the following contents.

[0054] Neural Networks

[0055] AI research, exemplified by neural networks, has achieved significant results in many fields and will continue to impact people's lives and work for a long time to come. A neural network can be understood as a computational model composed of multiple interconnected neurons. In a neural network, the connection strength between nodes can be represented as the weighted values ​​corresponding to the input signals, also known as parameters. Each neuron performs a weighted summation of different input signals and outputs the result through a specific activation function. Neurons can achieve nonlinear mappings depending on the activation function.

[0056] Taking the neuron shown in Figure 4 as an example, the input of the neuron can be denoted as A, and each dimension of the input can be denoted as a. j The corresponding weighted value is denoted as w. j Where j takes values ​​of 1, 2, ..., n. The neuron's input can also be set with a bias term to adjust the output, as shown by the constant 1 in Figure 4 (corresponding to the weighting value denoted as b). The weighting value, together with the summation units (SU), enhances or weakens the input. The output of the SU can be input into the activation function f to obtain the output t.

[0057] Common neural networks include convolutional neural networks (CNN), recurrent neural networks (RNN), and deep neural networks (DNN).

[0058] The neural network applicable to the embodiments of this application is described below with reference to Figure 5. The neural network shown in Figure 5 can be divided into three categories according to the position of different layers: input layer 510, hidden layer 520, and output layer 530. Generally, the first layer is the input layer 510, the last layer is the output layer 530, and the intermediate layers between the first and last layers are hidden layers 520.

[0059] The input layer 510 is used to input data, which may be, for example, a received signal received by a receiver. The hidden layer 520 is used to process the input data, for example, to decompress the received signal. The hidden layer may also be called an intermediate layer. The output layer 530 is used to output the processed output data, for example, to output the decompressed signal.

[0060] Referring to Figure 5, the neural network consists of multiple layers, each containing multiple neurons. Neurons between layers can be fully connected or partially connected. For connected neurons, the output of a neuron in one layer can serve as the input to a neuron in the next layer.

[0061] To facilitate understanding, the following uses a CNN as an example, with reference to Figure 6, to illustrate the multiple layers in a neural network. A CNN is a deep neural network with convolutional structures. As shown in Figure 6, the structure of a CNN may include an input layer 610, a convolutional layer 620, a pooling layer 630, a fully connected layer 640, and an output layer 650. The convolutional layer 620, pooling layer 630, and fully connected layer 640 are the intermediate layers of this CNN.

[0062] It should be noted that the CNN shown in Figure 6 is only an example of a convolutional neural network. In specific applications, convolutional neural networks can also exist in the form of other network models, and this application does not limit this.

[0063] CSI Feedback

[0064] CSI feedback is a crucial component of wireless communication systems. CSI feedback from terminal devices helps network equipment accurately understand the wireless channel state, thereby optimizing transmission parameters to improve system performance. The following section uses 5G NR as an example to introduce CSI feedback.

[0065] In 5G NR, two types of codebooks are defined: Type 1 and Type 2, and a set of precoding matrices is used to describe channel state information. These precoding matrices are obtained based on measurements of the Channel State Information Reference Signal (CSI-RS). Type 1 codebooks are of conventional precision and are primarily used to support single-user multiple-input multiple-output (MIMO) transmission; Type 2 codebooks are of high precision and are primarily used to support multi-user MIMO transmission to improve system spectral efficiency. Terminal devices can transmit this information to network devices through a feedback mechanism so that the network devices can perform scheduling and precoding.

[0066] As an example, all CSI codebook designs focus on feedback (FB) based on the current CSI-RS measurement. One implementation involves a network device (e.g., a gNB) periodically sending CSI-RS to an end device, which periodically provides CSI feedback (CSI-FB). The network device applies the CSI feedback to generate a pre-coding matrix indication (PMI) until the next CSI feedback is available. Furthermore, assuming the channel remains essentially unchanged over time (e.g., within milliseconds), the network device can use older CSI feedback. Another implementation involves the network device sending a cluster of CSI-RS, and the end device sending CSI feedback to the network device based on at least one CSI-RS in the cluster. The network device applies the CSI feedback to generate a time-dependent PMI until the next CSI feedback is available.

[0067] In both implementations described above, the terminal device requires frequent CSI feedback and is more suitable for low-mobility terminal devices.

[0068] As an example, a UE can measure channel quality by analyzing the reference signal (CSI-RS) transmitted by the base station.

[0069] As an example, CSI parameters can include the Channel Quality Indicator (CQI), Rank Indicator (RI), and Prefix Index (PMI). CQI is a quantized value reported by the terminal device to the network device, indicating the downlink channel quality. CQI reflects the maximum modulation and coding schemes the terminal device can receive under current channel conditions to ensure a certain bit error rate. RI represents the number of parallel data streams transmitted in a MIMO system. RI reflects the multipath propagation characteristics of the channel and the rank of the channel matrix, typically related to the spatial degrees of freedom of the channel. PMI is a feedback from the user equipment to the downlink channel state, indicating which precoding matrix the network device should select for signal transmission. The precoding matrix is ​​a linear transformation matrix used to process signals in a multi-antenna system.

[0070] It should be noted that in NR systems, when downlink (DL) / uplink (UL) channel reciprocity is good, CSI feedback based on CSI-RS is not required because the sounding reference signal (SRS) can be used for DL ​​CSI acquisition. However, when DL / UL channel reciprocity is insufficient, CSI feedback based on CSI-RS measurements is necessary. In this case, since the DL channel may be significantly different from the UL channel, DL CSI cannot be obtained solely through SRS.

[0071] Due to the complexity of wireless channels and the massive MIMO and high-frequency, high-bandwidth communication supported by 5G, CSI information typically has high dimensionality and rich detail, resulting in a very large amount of bandwidth resources required for feedback.

[0072] In 5G systems, physical layer and CSI feedback information play a crucial role in wireless communication. Relevant regulations mandate the compression of physical layer and CSI feedback information to reduce wireless resource consumption and improve system efficiency; this is known as physical layer and CSI feedback compression. Given the critical role of feedback information, effective compression is a key method for enhancing system performance.

[0073] To improve the efficiency and accuracy of CSI feedback, AI technology is introduced to drive the development and application expansion of wireless communication technology. As an example, AI / machine learning (ML) based CSI compression technology offers greater flexibility, lower distortion, and stronger adaptability. Therefore, leveraging the powerful capabilities of AI technology can achieve more efficient and lower-distortion CSI compression.

[0074] In the CSI feedback process, AI-based CSI feedback enhancement mainly focuses on the following two directions:

[0075] 1. Spatial-frequency domain CSI compression based on a bilateral AI model;

[0076] 2. Temporal CSI prediction based on a one-sided AI model.

[0077] Temporal CSI prediction based on a one-sided AI model employs a single-sided model on the terminal device side, meaning the CSI prediction model is deployed on the terminal device. This approach inputs historical CSI measurement data into the CSI prediction model and then outputs predicted CSI values ​​for future times, effectively addressing the timeliness issue of CSI feedback. In certain scenarios, predicting CSI based on AI / ML models can also improve CSI accuracy.

[0078] As an example, the overall process for CSI prediction based on AI / ML models can be divided into data preparation and preprocessing, input feature construction, AI model inference (prediction), post-processing, and performance evaluation and optimization.

[0079] Unlike time-domain CSI prediction schemes based on one-sided AI models, space-frequency domain CSI compression schemes based on two-sided AI models involve deploying a CSI generation model on the terminal device and a corresponding CSI reconstruction model on the network device. These two models can work together to complete CSI compression, feedback, and reconstruction (or rebuilding) tasks. This ensures that CSI is significantly compressed on the terminal device while guaranteeing that the network device can obtain more accurate CSI data, thereby optimizing resource scheduling and other operations.

[0080] To facilitate understanding, the AI / ML-based bilateral CSI compression method will be introduced below with reference to Figures 7 and 8.

[0081] Figure 7 illustrates an example of AI / ML-based bilateral CSI compression replacing the traditional method. As shown in Figure 7, in the traditional method, the terminal side (i.e., the terminal device side) receives the CSI-RS, performs channel estimation and equalization to obtain channel information H, and generates a precoding matrix based on H, outputting a Precoding Interface (PMI). The network side (i.e., the network device side) receives the PMI sent by the terminal side through the feedback link and obtains channel information H′ based on the acquired precoding matrix. In the AI / ML-based bilateral CSI compression method, the terminal side determines H, extracts channel features, and inputs them into the AI ​​encoder. The output of the AI ​​encoder is input into the AI ​​decoder through the feedback link, and the network side reconstructs the channel based on the output of the AI ​​decoder, thereby determining H′. Therefore, the AI ​​encoder on the terminal side can replace the traditional precoding matrix and output PMI on the terminal side, and the AI ​​decoder on the network side can replace the traditional received PMI and obtained precoding matrix on the network side.

[0082] Figure 8 shows an example of an AI / ML-based bilateral CSI model. The terminal measures the received CSI-RS to obtain H. The right singular vector obtained by performing singular value decomposition (SVD) on H can be used as model input. The input parameters have dimensions (2, N). Tx N sub ), where 2 represents the real part / imaginary part, and N Tx N represents the number of antenna ports. sub Indicates the number of sub-bands.

[0083] As shown in Figure 8, the bilateral CSI model can include a CSI production sub-model on the terminal side and a CSI reconstruction sub-model on the network side. The CSI production sub-model includes a CSI generation module and a quantization module, while the CSI reconstruction sub-model includes the output of the dequantization (also known as dequantization) module and the CSI reconstruction module.

[0084] The preceding text, with reference to Figures 7 and 8, introduced AI / ML-based bilateral CSI compression. In CSI compression, different transmit-receive unit (TxRU) mappings can be considered, and the model can be trained using a hybrid dataset. Since both CSI compression and / or CSI prediction utilize channel information measured from CSI-RS as model input / labels, the generalization performance of the following three TxRU mapping modes is adaptable to CSI compression datasets and can be extended to CSI prediction scenarios.

[0085] Training Type 1: The AI / ML model is trained on a training dataset from a scenario #A / configuration #A, and then the AI / ML model performs inference / testing on a dataset from the same scenario #A / configuration #A.

[0086] Training Type 2: The AI / ML model is trained on a training dataset from a scenario #A / Configuration #A, and then the AI / ML model performs inference / testing on a dataset different from scenario #A / Configuration #A. A scenario / configuration different from scenario #A / Configuration #A could be scenario #B / Configuration #B, or scenario #A / Configuration #B.

[0087] Training Type 3: The training dataset for training the AI / ML model is constructed by mixing datasets from multiple scenarios / configurations. For example, the training dataset may include a dataset from scenario #A / Configuration #A as well as datasets different from scenario #A / Configuration #A. The AI / ML model can then perform inference / testing on datasets from a single scenario / configuration derived from these multiple scenarios / configurations.

[0088] Optionally, the CSI generation and reconstruction parts of the model can employ an autoencoder structure. The training of this bilateral model has the following three implementation types.

[0089] Implementation Type 1: Single-sided (terminal side or network side) training, that is, training the CSI generation sub-model and the CSI reconstruction sub-model on the same side (terminal side or network side).

[0090] Implementation Type 2: Joint Training of Network and Terminal Sides. There are two implementation methods: simultaneous training and sequential training. The simultaneous training process is as follows: ① The terminal-side CSI generation sub-model generates CSI feedback and sends it to the network side; ② The network-side CSI reconstruction sub-model reconstructs the CSI based on this feedback and generates backpropagation information (such as gradients), which is then sent back to the terminal side; ③ The terminal side updates the CSI generation sub-model based on the backpropagation information from the network side, iterating repeatedly until the model converges. In the sequential training process, once the terminal-side CSI generation sub-model and the network-side CSI reconstruction sub-model are jointly trained, the structure and parameters of the network-side CSI reconstruction sub-model can be frozen. Other terminal-side CSI generation sub-models and the frozen network-side reconstruction sub-models are then trained simultaneously. However, the network-side model parameters are not further updated; they only assist the terminal-side model in calculating and transmitting forward / backward propagation information and updating the terminal-side model parameters until the model converges.

[0091] Implementation Type 3: Independent training on the network and terminal sides, meaning the CSI generation sub-model on the terminal side and the CSI reconstruction sub-model on the network side are trained independently. This training method can be further divided into two types: network-initiated and terminal-initiated. When training starts from the network side, after the network side completes its training, it shares the training set required for the CSI generation sub-model with the terminal side so that the terminal side can train the CSI generation sub-model. When training starts from the terminal side, after training is completed, the terminal side shares the training set required for the CSI reconstruction sub-model with the network side so that the network side can train the CSI reconstruction sub-model.

[0092] The following describes an embodiment of training a bilateral model with reference to Figure 9. This embodiment includes steps S1 to S3.

[0093] In step S1, the network side first trains the end-to-end model and obtains the corresponding dataset (training at NW side). This dataset may include one or more of the following information: encoder 1 (901), decoder 1 (902), target CSI (V1), CSI feedback (C1), and reconstructed target CSI.

[0094] In step S2, the terminal side trains encoder 2 (903) based on the dataset and / or model shared from the network side. This process varies depending on the option, and the dataset exchanged from the network side to the terminal side may include (V1, C1). As shown in Figure 9, the terminal-side training in step S2 can consider two schemes.

[0095] Scheme 1 (alt.1) can first train a nominal decoder, namely nominal decoder 2 (904). Among them,<C1,V1> As the model input and label, the terminal side and decoder 2 jointly train their own encoder 2. Since the network side training also uses V1 as the label, the trained nominal decoder 2 can achieve a similar model input and output mapping relationship as the network side decoder 1.

[0096] For the dataset used to train step S2, the target CSI (V') can be used. As shown in Figure 9, V' can have two sub-options.

[0097] For the network-side dataset, V1 delivered by the network side can be used as V'. The parameters sent from the network side to the terminal side correspond to encoder 1 generated by the network side. In this method, the network side first trains nominal decoder 2, and the terminal side uses the dataset of V' to jointly train nominal decoder 2 with the frozen encoder 1. However, the terminal side then uses V' to jointly train its own encoder 2 with nominal decoder 2. In this method, the network side sends the network-side dataset of V1 as V' to the terminal device.

[0098] For the terminal-side dataset, a new target CSI V2 can be used as V', which may include new data generated on the terminal side. For example, the terminal side can generate a new dataset for the terminal device that reflects a specific distribution of terminal-side additional conditions, which the network-side dataset V1 cannot represent. The terminal side can use encoder 1 to generate...<V',C2> The dataset is used as <model input, model output>. Specifically, the terminal side uses...<V',C2> The network trains its own encoder 2 using <model input, labels>. The network side passes dataset V1 as V' to the terminal device, while the terminal side uses V2 as a new dataset for V', which may include new data generated by the terminal side.

[0099] Option 2 (alt. 2) directly trains encoder 2. The terminal side uses...<C1,V1> Use the model input and labels to train your own encoder2.

[0100] In step S3, the model is deployed for inference. Encoder 2 (903) at the terminal device and decoder 1 (902) at the network device are deployed to the network for inference.

[0101] It should be understood that the above introduction to CSI feedback based on 5G is only an example. The development of 5G mobile communication systems can provide support for subsequent new communication technologies. For example, new waveforms providing coverage in the terahertz band of 6G mobile communication technology. Other new technologies include multi-antenna transmission technologies such as full-dimensional MIMO (FD-MIMO), array antennas, and large antennas; metamaterial-based lenses and antennas for improving the coverage of terahertz band signals; high-dimensional spatial multiplexing technologies using orbital angular momentum (OAM); reconfigurable intelligent surfaces (RIS); full-duplex technologies to increase the frequency efficiency of 6G mobile communication technology and improve system networks; AI-based communication technologies for system optimization by leveraging satellites and AI from the design phase and internalizing end-to-end AI support; and next-generation distributed computing technologies for implementing services at complexity levels exceeding the limits of UE operational capabilities by utilizing ultra-high-performance communication and computing resources. Therefore, CSI prediction and compression for the air interface are important research directions in 6G and subsequent communication system technologies.

[0102] The previous section introduced AI-based CSI feedback enhancement. To achieve space-frequency domain CSI compression based on a bilateral AI model and time-domain CSI prediction based on a unilateral AI model, how to collect model-related data becomes a key research issue. Furthermore, for CSI compression based on a bilateral AI model, it is also necessary to address the alignment issues between the terminal and network sides, as well as transmission problems.

[0103] As an example, CSI compression involves a two-sided model, comprising an encoder and a decoder. The encoder compresses CSI data, achieving efficient transmission by reducing its size. For example, the encoder can quantize the compressed representation to reduce its bit width, thus enabling efficient transmission. Upon receiving the compressed CSI, the network side reconstructs and estimates the original CSI data using the decoder. For example, the network device can dequantize the received quantized representation. The reconstructed CSI can be used to optimize the communication link (e.g., beamforming, power allocation) to improve performance and reliability. How quantization alignment is performed between the terminal device and the network device is a consideration.

[0104] As another example, different vendors of terminal device chips may have different codebooks. Quantization alignment involves not only the terminal device and network equipment, but also different terminal device vendors and third-party servers. Therefore, how different vendors collaborate is a consideration.

[0105] As another example, how to determine the input to the AI ​​decoder to improve the accuracy of model inference is also an issue that needs to be considered.

[0106] As another example, terminal devices provide CSI feedback through CSI reports. Whether to improve upon traditional CSI reporting principles for CSI reporting in model inference is also a question that needs to be considered.

[0107] As another example, CSI compression based on bilateral AI models involves data interaction during the training and inference processes of the bilateral models. Therefore, how to collect data during the model training or inference phase is also a technical problem that needs to be solved.

[0108] As another example, in AI / ML-based CSI compression scenarios, the compressed CSI data may lead to inconsistencies between the CSI feedback and the actual channel quality, thus affecting CQI calculation and resource allocation. During CSI compression and recovery, errors may also exist between the reconstructed CSI from the network device and the target CSI, further impacting system performance due to the inconsistency between the CSI feedback and the actual channel quality.

[0109] To address the aforementioned issues, this application proposes a method for wireless communication. In this method, a first model for compressing CSI is deployed on the first device (e.g., a terminal device), and a second model for receiving and processing CSI feedback is deployed on the second device (e.g., a network device). The third CSI obtained by the second device through the second model can be determined based on a first adjustment amount of the first model and / or a second adjustment amount of the second model, thereby improving the inference accuracy of CSI compression based on a bilateral AI model.

[0110] To facilitate understanding, the method proposed in the embodiments of this application will be described in detail below with reference to Figure 10. Figure 10 is presented from the perspective of the interaction between the first device and the second device.

[0111] The first device is any of the terminal devices or terminal-side devices described above, such as a UE. In some embodiments, the first device can communicate with a network device or other terminal devices. As one embodiment, the first device can receive a reference signal sent by the network device or other terminal devices to perform channel estimation. This reference signal may be, for example, a CSI-RS, or a demodulation reference signal (DMRS), or other reference signals reflecting the channel state. As one embodiment, the first device can send an SRS to the network device.

[0112] In some embodiments, deploying the first model on the first device side can be understood as the first model being deployed on the first device, or on an auxiliary device communicating with the first device. The auxiliary device may be a non-3GPP entity corresponding to the first device, such as an OTT server.

[0113] The first model can be a model related to CSI prediction and / or CSI compression. The first device can predict and / or compress the CSI to be transmitted based on the deployed first model to improve transmission efficiency. As an example, the first model can be an AI / ML model related to CSI feedback. For instance, the first model can be used to compress a first CSI to obtain a second CSI.

[0114] As an example, the first model can be an AI / ML-based encoder, also known as an AI encoder. For instance, the first model can be an AI encoder related to CSI compression, such as encoder 901 or encoder 903 in Figure 9.

[0115] As an example, the first model can achieve CSI prediction based on a one-sided AI model, such as time-domain CSI prediction.

[0116] As an example, the first model can work in conjunction with a second model deployed on network devices or other terminal devices to achieve CSI compression based on a bilateral AI model, such as space-frequency domain CSI compression.

[0117] As an example, the first model may be a CSI generation model deployed on the first device side, such as a CSI generation sub-model.

[0118] As an example, the first model may include a CSI generation module and a quantization module.

[0119] In some embodiments, the first device can train the first model. For example, based on the training type / implementation type described above, the terminal device or OTT server can train the first model independently or jointly.

[0120] In some embodiments, the cell where the first device is located is called the first cell. The first device communicates with the network device corresponding to the first cell.

[0121] In some embodiments, the first device is one of a plurality of terminal devices. The plurality of terminal devices may be multiple types of terminal devices and / or multiple terminal-side devices. In some embodiments, the plurality of terminal devices may be multiple terminal devices of the same type to facilitate device-group-based communication between the network device and the plurality of terminal devices. In some embodiments, at least two of the plurality of terminal devices are of different types.

[0122] The second device can be any of the network devices or network-side devices described above, or it can be a terminal device other than the first device. As one embodiment, when the second device is a network device such as a base station, it can be the network device corresponding to the first cell. This network device can provide services to all terminal devices in the first cell. As another embodiment, the second device may be a UE different from the first device. The second device can perform side-channel communication with the first device.

[0123] In some embodiments, a corresponding second model may be deployed on the second device side. Deploying the second model on the second device side can be understood as deploying the second model on the second device or on an auxiliary device of the second device.

[0124] The second model can be a model related to CSI prediction and / or CSI compression. The second device can reconstruct the received CSI feedback based on the deployed second model to determine the channel state. As an example, the second model can be an AI / ML model related to the CSI feedback. For instance, the second model can be used to process the second CSI (e.g., reconstruct the CSI) to obtain a third CSI.

[0125] As an example, the second model can be an AI / ML-based decoder, also known as an AI decoder. For instance, the second model could be an AI decoder related to CSI compression, such as decoder 902 or decoder 904 in Figure 9.

[0126] As an example, the second model can work in conjunction with the first model deployed on the first device side to achieve CSI compression based on a bilateral AI model, such as space-frequency domain CSI compression.

[0127] As an example, the second model can be a CSI reconfiguration model deployed on the second device side, such as a CSI reconfiguration sub-model.

[0128] As an example, the second model may include a dequantization module and a CSI reconstruction module.

[0129] In some embodiments, the second device can train the second model. For example, the second device can train the second model independently or jointly based on the training type / implementation type described above.

[0130] As an example, leveraging the computing power of the network device, the network device can also train a first model deployed on a terminal device, and then send the trained first model to the terminal device. For instance, the network device can perform synchronous training on encoder-decoder pairs.

[0131] In some embodiments, the first model on the first device side can be the same as the second model on the second device side. For example, the CSI generation model used on the terminal device side can be the same as the actual CSI reconstruction model used on the network device side. Different input ports and output ports are set in actual use. Alternatively, the first model used on the terminal device side can be the same as the reference model provided by the network device side. Or, the first model used on the terminal device side can be the same as the proxy model developed on the terminal device side.

[0132] The process shown in Figure 10 includes steps S1010 and S1020, which are described below.

[0133] Referring to Figure 10, in step S1010, the first device performs CSI compression on the first CSI using the first model to obtain the second CSI.

[0134] The first CSI is related to the first channel and can be understood as indicating the channel state of the first channel. The first CSI can be used to capture the wireless channel characteristics between the terminal device and the network device or other terminal devices. The first channel can be any downlink channel or sidelink channel, without limitation. As an example, when the second device is a network device, the first channel is any downlink channel. As an example, when the second device is a terminal device, the first channel is any sidelink channel.

[0135] The first CSI can be the input to the first model for CSI compression. Since the first and second models are used by the second device to determine the first CSI, the first CSI can also be referred to as the target CSI or the input CSI.

[0136] The first CSI may include one or more CSIs associated with the first channel. As an example, the first CSI may include CSIs corresponding to multiple time slots. For example, the terminal device may measure the first channel in different time slots (t1, t2, t3, t4) to obtain the CSI at different time slots. As another example, the first CSI may include at least one measured result and / or at least one predicted result, and the first model may perform CSI compression based on the measured result and the predicted result.

[0137] The second CSI can be the output of the first model, used to represent the compressed CSI. Since the first model is also the CSI generation model, the second CSI can be called the generated CSI or intermediate CSI. The feedback CSI sent by the first device to the second device on the feedback link is the second CSI. The second CSI is generated based on the first CSI, therefore the second CSI is also related to the first channel.

[0138] In some embodiments, the first CSI may include one or more historical CSIs of the first channel. Historical CSI can be understood as the CSI corresponding to a time (or time slot) prior to the current time (or time slot). For example, the first CSI may include the CSI of the first channel in the current time slot and the accumulated CSI in the previous few time slots. That is, the input of the first model, or encoder, may include multiple parts, namely the CSI of the current time slot and one or more accumulated CSIs.

[0139] As an example, when the first CSI includes the current CSI and one or more historical CSIs, the second CSI is determined based on the compression of the current CSI and one or more historical CSIs according to the first model. However, the dequantized CSI of the network device is still the CSI corresponding to the current moment.

[0140] To facilitate understanding, the following example, with reference to Figure 11, illustrates the method of including multiple historical CSIs in the first CSI. Figure 11 uses the example of an encoder as the first model and a decoder as the second model, illustrating the input and output at times tn-2, tn-1, and tn, respectively.

[0141] As shown in Figure 11, the inputs at the three time points are V tn-2 V tn-1 V tn The outputs are V' tn-2 V' tn-1 V' tn For time tn-1, the inputs to both the encoder and decoder include the accumulated CSI from time tn-2, i.e., the historical CSI. For time tn, the inputs to both the encoder and decoder include the accumulated CSI from time tn-1, i.e., the historical CSI.

[0142] As shown in Figure 11, the decoder input consists of two parts: the compressed CSI of the current time slot and the accumulated CSI of the previous few time slots. In each time slot, the encoder processes CSI containing information about the current CSI and the accumulated CSI. The encoder's main task is to process the current CSI (H... t ) and past cumulative CSI (C t Feature fusion is performed to generate a compact compressed CSI (Z). t The decoder is primarily used to recover the CSI of the current time slot.

[0143] Therefore, the first device can determine the first CSI using the temporal correlation of the CSI. For the terminal device, CSI from different time slots are fed into the encoder. The encoder can compress the CSI data using spatial and temporal information.

[0144] In some embodiments, the first CSI can be determined based on the measurement result obtained after the first device measures a reference signal. This reference signal can be transmitted through a first channel. The measurement result used to determine the first CSI can be the actual measurement result (measured result) or the measurement result based on AI-based CSI prediction (predicted result).

[0145] In some embodiments, the first CSI may include one or more predicted CSIs. The one or more predicted CSIs may be the CSI of the first channel determined based on prediction results. It should be noted that when the first CSI includes one or more predicted CSIs, it may also include at least one measured CSI. The at least one measured CSI may be a CSI determined based on measurement results.

[0146] As an example, the first device may first perform CSI prediction, and then perform CSI compression based on the predicted CSI.

[0147] To facilitate understanding, the method for predicting the first CSI, including Figure 12, will be illustrated below. As shown in Figure 12, on the terminal device side, the CSI of time slot n can be obtained by predicting the CSI based on an observation window. The prediction model on the terminal device side can use the value of the observation window as a sample and predict the CSI of time slot n.

[0148] As shown in Figure 12, the terminal device (first device) compresses the CSI of time slot n using a CSI generation sub-model (first model). The network device (second device) performs reconstruction using a CSI reconstruction sub-model (second model) to recover the CSI in time slot n. For the CSI prediction part, the historical channel matrix during the observation window can be used as input to predict the channel matrix of time slot n. The input part of the first model can only contain the CSI predicted in time slot n.

[0149] In some embodiments, the first CSI may include one or more predicted CSIs, which may include the current CSI or historical CSIs. Alternatively, the first CSI may include one or more historical CSIs, which may include the measured CSI corresponding to the measured result or the predicted CSI corresponding to the predicted result.

[0150] As an example, for the first device, when CSI from different time slots is fed into the first model, the first model can use historical CSI data (from previous time slots) and / or predicted CSI from subsequent time slots to enhance the compression of the current CSI.

[0151] Referring again to Figure 10, in step S1020, the first device sends a second CSI to the second device. The second device can process the second CSI using a second model to obtain a third CSI.

[0152] The first device can send a second CSI to the second device via a CSI report. The second CSI is determined based on the first model; therefore, the CSI report sent by the first device is an AI / ML-based CSI report. This CSI report can share transmission resources with or be configured independently of traditional CSI reports.

[0153] As an example, an AI / ML-based CSI report can be a CSI report obtained by processing a first CSI using a first model.

[0154] In some embodiments, when a second CSI is sent via a CSI report, the sending of the CSI report is related to at least one of the following: determining whether to send based on an event-triggered manner; and the first priority of the CSI report.

[0155] As an example, determining whether to send a CSI report based on event triggering helps reduce the sending of invalid reports and saves signaling overhead. Event triggering is also known as asynchronous inference reporting. For the first device, CSI reporting can deviate from the traditional periodic CSI reporting method and instead adopt an asynchronous triggering mechanism, reporting inference results only when critical events occur.

[0156] In the above embodiments, the event for determining whether to send a CSI report may include referencing a traditional CSI report triggering event.

[0157] In the above embodiments, the event for determining whether to send a CSI report may further include the prediction result of the first channel. The prediction result of the first channel can be understood as the result of predicting the channel state of the first channel based on an AI model. The first device can predict the changing trend of the first channel based on the AI ​​model, thereby determining whether to trigger a CSI report for sending a second CSI.

[0158] As an example, the first priority of a CSI report is used to indicate the priority level of the CSI report. The first priority can be determined according to a first CSI priority rule. For CSI reports in inference, in addition to traditional CSI reporting principles, a first CSI priority rule can also be added. Exemplarily, the first CSI priority rule can be related to the AI / ML model; therefore, the first priority of the CSI report can be related to the first model. For example, the first priority can be determined based on the type or level of the first model.

[0159] As one implementation approach, the first priority rule can be a priority rule that considers the specific report type of AI / ML. When the CSI report generated by the first model is of a specific report type, the priority level of the first priority rule can be determined.

[0160] As another implementation, the first priority rule can be a priority rule within the bit sequence of each AI / ML-specific inference CSI report. Network devices can set priorities based on the impact of the AI / ML model on the channel state. Therefore, the first priority is determined based on the impact of the first model on the channel state of the first channel. For example, network devices can assign specific bit sequence priorities to each CSI report type. Higher-priority CSI reports are transmitted with more efficient bit sequences.

[0161] In some embodiments, CSI reports can be generated based on a multi-level mapping. The first device can divide the CSI mapping into multiple levels, thereby optimizing the CSI report through hierarchical mapping. As an example, the CSI mapping can divide the second CSI generated by the first model into multiple parts for reporting. For example, the framework of the CSI report may include Part 1 and Part 2. Part 1 and Part 2 of the CSI report may correspond to parameters or information at different levels, respectively.

[0162] As one embodiment, CSI part 1 may contain information that affects the size of CSI part 2, such as the selected rank value R and the quantization granularity. Furthermore, the size of the pre-quantized CSI payload for each layer may also be carried by CSI part 1. CSI part 2 may contain the content that generates the CSI part output, such as the mapping method for each layer and the size of the potential space for scalable dimensions.

[0163] As an example, when generating CSI reports based on multi-level mapping, the mapping method for each layer can be further optimized. For instance, each layer can use an independent quantization codebook. Alternatively, each layer can dynamically select the quantization granularity and codebook size based on the channel state. Furthermore, the mapping rules for CSI reports can be further optimized by comparing the inferred CSI with the actual CSI.

[0164] As one example, multi-level mapping can include a first-level mapping and a second-level mapping. The first-level mapping corresponds to the channel characteristics of a first channel, and the second-level mapping corresponds to the subcarriers or antennas associated with the first channel. For example, the first-level mapping can select a broad-based mapping framework based on global channel characteristics. The second-level mapping can refine the mapping for the data of each subcarrier or antenna.

[0165] In some embodiments, the content of the CSI report can be further optimized to reduce redundant information and improve adaptability in complex channel environments. As one example, the CSI report includes differential information between the current CSI and the previous CSI. That is, the CSI report is not complete CSI data, but only reports the differential information relative to the previous report. As another embodiment, the CSI report includes partial information extracted by a first model based on the first CSI. This partial information may be key channel information extracted by the first model.

[0166] As an example, for differential CSI reports, the second device maintains historical CSI estimates, while the first device only sends differential data.

[0167] As an example, the first model can introduce channel semantic compression based on the channel semantic report. Through channel semantic compression, the CSI report can report only key semantic information of the channel instead of the raw data. For example, the first model can transmit only the extracted semantic features to improve transmission efficiency while retaining key channel information. Semantic features include, for example, channel quality and direction information.

[0168] In some embodiments, if a CQI is configured in the CSI report, the CSI report may include the CQI. The CQI is determined according to one of the following methods: calculating the CQI based on the target CSI and actual channel measurements; calculating the CQI based on the target CSI and potential adjustments; calculating the CQI based on a conventional codebook; or calculating the CQI based on the output of the CSI reconstruction portion of the actual channel estimation.

[0169] As an example, CQI may not be calculated based on the output of the CSI reconstruction portion of the actual channel estimate.

[0170] As an example, when the CSI reconstruction model is available on the first device side, the CQI is determined based on the output of the CSI reconstruction model. In this scenario, the first device side can perform reconstruction model inference with potential adjustments based on the CSI reconstruction model. For example, the first device can use the precoding reference signal corresponding to the reconstructed precoder to derive the CQI. It should be noted that the CSI reconstruction part model on the first device side may differ from the actual CSI reconstruction part model (second model) used on the network side.

[0171] As an example, the first device can use a two-stage method to calculate the CQI.

[0172] The second model is used by the second device to determine the third CSI based on the second CSI. The third CSI is also related to the first channel. The third CSI is the CSI of the first channel obtained after the second model reconstructs the CSI based on the CSI feedback; therefore, the third CSI can be called the reconstructed CSI or the output CSI.

[0173] The third CSI can be the output of the second model, enabling the second device to determine the CSI of the first channel. In some embodiments, the second device can determine the channel state of the first channel based on the third CSI, and can also transmit a precoded reference signal based on the third CSI.

[0174] The third CSI can be determined based on the first adjustment amount of the first model and / or the second adjustment amount of the second model. In some embodiments, when determining the reconstructed third CSI based on jointly trained or independently trained bilateral models (the first model and the second model), adjustments can be made based on the adjustment amount during the model's inference process, or the model's inference results can be adjusted based on the adjustment amount, to improve the accuracy of the third CSI. It should be understood that the process of adjusting the first model and / or the second model based on the first adjustment amount and / or the second adjustment amount is also known as CSI adjustment.

[0175] In some embodiments, the third CSI may be determined based solely on a first adjustment amount of the first model, or solely on a second adjustment amount of the second model. That is, the first and second adjustment amounts may be presented separately to reduce the complexity of the adjustment.

[0176] In some embodiments, the third CSI can be determined based on a first adjustment amount of the first model and a second adjustment amount of the second model. That is, the first and second adjustment amounts can be presented together to more accurately infer the process of CSI generation and reconstruction.

[0177] In some embodiments, the first adjustment amount and / or the second adjustment amount may be 0. In response to the satisfaction of a first condition, the first adjustment amount is 0, and / or the second adjustment amount is 0. Thus, the first condition can be used to determine whether to trigger an adjustment of the first model and / or the second model.

[0178] As an example, the first condition is related to the performance of the first model and / or the second model. For instance, when the accuracy of the first model and / or the second model meets the threshold requirement, the first adjustment amount and / or the second adjustment amount can be set to 0, that is, no adjustment is made.

[0179] As an example, the first condition is related to the operating state of the first model and / or the second model. For example, the first model and / or the second model can be set with an upper limit on the number of adjustments, and after a certain number of adjustments, the first adjustment amount and / or the second adjustment amount can be set to 0.

[0180] As one embodiment, the first condition is related to the motion state of the first device. For example, for a first device in low-speed motion or at a standstill, the first adjustment amount and / or the second adjustment amount can be set to 0. Alternatively, for a first device in high-speed motion, the first adjustment amount and / or the second adjustment amount can be related to the motion parameters of the first device.

[0181] As one example, the first condition is related to the current measured or predicted results of the first device. For instance, the first device can determine whether to trigger CSI adjustment based on the current channel state. Alternatively, the first device can determine whether to trigger CSI adjustment based on the CSI prediction result. As mentioned above, the CSI prediction result can be determined based on an AI model.

[0182] In some embodiments, the first adjustment amount and / or the second adjustment amount may be determined based on the mismatch between the third CSI and the first CSI. For example, the difference between the first CSI and the third CSI may be used to determine the first adjustment amount and / or the second adjustment amount.

[0183] As one embodiment, the first adjustment amount can be determined based on the output of the first model and / or the output of the second model. The output of the first model includes, but is not limited to, the second CSI. The output of the second model includes, but is not limited to, the third CSI. As one implementation, since the first model is deployed on the first device side, the first or second device can evaluate the inference performance of the first model based on its output and determine the first adjustment amount. As another implementation, the second device can also evaluate the inference performance of the first model based on its output and determine the first adjustment amount. Once the second device determines the first adjustment amount, it can send it to the first device via an instruction message.

[0184] In the above embodiments, the first device can receive instruction information from the second device, the instruction information including a first adjustment amount.

[0185] As one example, the second adjustment amount can be determined based on the output of the second model. Since the second model is deployed on the second device side, the second device can evaluate the inference performance of the second model based on its output and determine the second adjustment amount.

[0186] In some embodiments, the first adjustment amount and / or the second adjustment amount are related to one or more of the following: intermediate evaluation parameters corresponding to CSI compression; auxiliary information from the second device. That is, the adjustment of the CSI compression model can be triggered by the first device itself, by the first and second devices jointly, or by the second device alone.

[0187] As an example, the first adjustment amount can be independently determined by the first device based on intermediate evaluation parameters. These intermediate evaluation parameters are used for AI / ML-based CSI compression. For CSI compression, the intermediate evaluation parameters can also be referred to as intermediate metrics or key performance indicators (KPIs) for CSI compression.

[0188] As an example, the first adjustment amount can be determined based on auxiliary information from the second device. This auxiliary information can be used to compensate for the feature vector difference between the third CSI and the first CSI. In AI / ML-based CSI compression, the first device does not have a model for CSI reconstruction except for joint training on the first device side; therefore, the first device is unaware of the output channel matrix / feature vector on the second device side. The first device also cannot reconstruct the output channel matrix / feature vector on the second device side. When the first device uses the original feature vector for CSI calculation, this differs from the feature vector recovered by the second device. This misalignment between the original and recovered feature vectors will lead to CSI inaccuracies between the second and first devices. To report more accurate CSI, the first device needs to compensate for the CSI calculated using the original feature vector. Since the first device may not be able to recover the information of the third CSI, the CSI can be compensated based on the auxiliary information indicated by the second device. That is, to improve the accuracy of CSI reporting, the first device needs to compensate for the CSI calculated based on the original feature vector according to the auxiliary information provided by the second device, or compensate for the CQI calculated through the CSI.

[0189] As one implementation, auxiliary information may include historical values ​​or historical compensation information of the third CSI. For example, the second device indicates to the first device the previous third CSI for compensation. The first device can calculate compensation based on the historical values ​​of the third CSI. Alternatively, the network device may provide historical compensation information, and the first device performs CSI compensation according to the network device's instructions.

[0190] In the above embodiments, the first device can determine a first adjustment amount based on local information and auxiliary information from the second device. The second device can send this auxiliary information to the first device to facilitate the first device in determining the first adjustment amount. For example, when the first device receives instruction information from the second device, the instruction information may include the auxiliary information from the second device.

[0191] As one embodiment, the second adjustment amount can be determined based on the first CSI and the third CSI. The second device can determine the second adjustment amount of the second model based on the difference between the output of the second model and the input of the first model.

[0192] As one embodiment, the second device can indicate auxiliary information to the first device in various ways. These various ways correspond to different types of information. For example, the auxiliary information from the second device is used to determine one or more of the following: the historical CSI offset value corresponding to the first channel; the weighting factor of the second CSI; and the compensation reference value corresponding to the first channel. The compensation reference value includes a global compensation value and / or a local compensation value.

[0193] In some embodiments, the first adjustment amount and / or the second adjustment amount are related to one or more of the following: intermediate evaluation parameters of CSI compression; historical CSI offset values ​​corresponding to the first channel; weighting factors of the second CSI; and compensation reference values ​​corresponding to the first channel.

[0194] For the sake of brevity, the following section will take the first adjustment as an example to introduce the model adjustment based on intermediate evaluation parameters and auxiliary information.

[0195] In some embodiments, for CSI compression based on a bilateral AI model, the model can be adjusted by directly estimating intermediate evaluation parameters. In some scenarios, adjusting the CSI compression model using intermediate evaluation parameters allows for direct prediction of the adjusted CSI value without reconstructing the target CSI. In other scenarios, adjusting the model using intermediate evaluation parameters eliminates the need to send reference signals based on the output of the second model. In still other scenarios, the second device can also estimate monitoring outputs other than intermediate evaluation parameters without reconstructing the target CSI.

[0196] As an example, the intermediate evaluation parameter is the squared generalized cosine similarity (SGCS). The SGCS value measures the correlation between the second CSI and the first CSI. SGCS can be an effective intermediate KPI for guiding CSI adjustments. For the first model, the SGCS value can reduce the error or bias between the second and first CSIs, thereby improving the accuracy of the second CSI in the CSI report. For example, the first adjustment amount can be determined based on the intermediate metric SGCS. The input to the SGCS estimator can be the intermediate output of the CSI compression model, i.e., the output of the first model.

[0197] For example, the value of SGCS can be In one case,<V,C> represents the inner product between the first and second CSIs (measuring similarity); ||V|| and ||C|| are the norms of the first and second CSIs, respectively (measuring channel energy); SGCS ∈ [0,1], when SGCS is close to 1, it indicates that the second CSI and the first CSI are infinitely close (or highly close); SGCS is close to 0, it indicates that the error between the second CSI and the first CSI is large (or irrelevant). In another case,<V,C> SGCS represents the inner product between the first CSI and the third CSI, where |V| and |C| are the norms of the first CSI and the third CSI, respectively. When SGCS is close to 1, it means that the third CSI and the first CSI are infinitely close. When SGCS is close to 0, it means that the error between the third CSI and the first CSI is relatively large.

[0198] In the above embodiments, the first device can obtain an accurate SGCS value using a simple SGCS estimator, or it can develop another estimator to determine the SGCS value. Exemplarily, the SGCS estimator can efficiently predict the SGCS value and use it for real-time CSI adjustment. Exemplarily, by setting an intermediate encoder and implementing a two-stage method to directly predict the adjusted CSI value, the efficiency and robustness of the system can be improved. This two-stage method can be divided into Stage 1 and Stage 2. Stage 1 performs SGCS estimation, predicting the SGCS value using the SGCS estimator. Stage 2 performs CSI adjustment, adjusting the second CSI and / or the third CSI based on the predicted SGCS value.

[0199] In the above embodiments, the first device can correct the second CSI reported by the first device based on the SGCS value, making it closer to the actual channel quality. That is, the SGCS can be adjusted to adjust the second CSI based on an adjustment formula. For example, the adjusted value of the second CSI is: CSI′=α(SGCS)*CSI+(1-α(SGCS))*ΔCSI;

[0200] Wherein, CSI is the initial value before the second CSI adjustment (the original value after compression), α(SGCS) = SGCSγ, γ is the adjustment parameter, and ΔCSI is the measured result of the first channel or determined according to the instructions of the network device.

[0201] In the above adjustment formula, the first adjustment can be the difference between CSI′ and CSI. α(SGCS) is the CSI adjustment weight related to SGCS, and γ is used to control the influence of SGCS on CSI adjustment. When the SGCS value is close to 1, i.e., α(SGCS)≈1, the influence of CSI adjustment is small. When the SGCS value is close to 0, α(SGCS)≈0, indicating that the CSI deviation is large, and therefore a larger correction is needed.

[0202] In some embodiments, auxiliary information can be used to determine the historical CSI offset value corresponding to the first channel, i.e., the first adjustment amount. The second device can feed back the previous CSI error offset, i.e., the historical CSI offset value. The historical CSI offset value can be determined based on the historical values ​​of the first CSI and the third CSI. The historical values ​​of the first CSI and the third CSI can also be used to determine the historical CSI feature vector difference.

[0203] For example, the historical CSI offset value can be expressed as Δ = Avg(CSI) NW -CSI UE ), of which CSI NW It is the historical value of the third CSI calculated by the second device. UE It is the historical value of the first CSI calculated by the first device.

[0204] As an example, the historical CSI offset value can also represent the difference in historical CSI feature vectors. The second device can indicate the previously reconstructed CSI feature vector, and the first device can calculate the difference between the current original feature vector and the historically reconstructed feature vector, using this as the source of CSI error estimation. When the historical CSI feature vector difference is Δ, the adjusted second CSI is CSI′ = CSI - Δ. Optionally, Δ can be an error determined by the second device based on historical channel conditions. Optionally, Δ can also be used to adjust the weighting factor of the second CSI.

[0205] As one example, the historical CSI offset value is used for multiple terminal devices, including the first device. For instance, the network device may predefine a fixed historical CSI offset value and indicate this uniform offset value to all terminal devices via signal broadcasting.

[0206] In some embodiments, auxiliary information can be used to determine the weighting factor of the second CSI, i.e., the first adjustment amount. For example, in, The weighting factor is used as a feedback weighting factor to correct the second CSI. Specifically, the second device can calculate the weighting factor based on a channel prediction model and broadcast it to the terminal device.

[0207] As one example, the weighting factor can be dynamically updated based on the channel state, or it can be a constant set by the network side. For example, the second device can dynamically update the weighting factor based on the statistical characteristics of the channel state H of the first channel. Statistical characteristics include, for example, the signal-to-noise ratio (SNR) or path loss (PL).

[0208] As an example, the weighting factor can be determined based on the correlation between the third CSI and the first CSI. The correlation between the third CSI and the first CSI can be the ratio between historical values ​​of the third CSI and historical values ​​of the first CSI. For example,

[0209] As an example, the weighting factor can be determined based on the channel quality of the first channel at different times. The channel quality of the first channel can be the channel quality estimated by the second device or network device. For example, the weighting factor of the second CSI at time t. for:

[0210] Where α is the smoothing coefficient. The weighting factor at time t-1 SNR(t) and SNR(t-1) are the signal-to-noise ratios of the first channel estimated by the network device at times t and t-1. g(·) can be a dynamic update function.

[0211] In the above formula, the weighting factor is determined based on the factor time window averaging method to reduce instantaneous fluctuations.

[0212] In some embodiments, auxiliary information can be used to determine a compensation reference value corresponding to the first channel, i.e., a first adjustment amount. The compensation reference value can be used by the first device to compensate for CSI, i.e., to adjust CSI.

[0213] As one embodiment, the first device can perform CSI compensation based on a distributed compensation mechanism. Based on the distributed compensation mechanism, the first device can perform collaborative compensation by combining partial feedback from the first device, without relying on instructions from the network side.

[0214] As an example, the first device can determine a compensation reference value based on a global compensation value and a local compensation value from the network side. For example, the network device can send a global compensation reference value (i.e., the global compensation value). The first device can determine a local compensation reference value (i.e., the local compensation value) based on local real-time channel measurement results, thereby fine-tuning the compensation reference value. For example, the adjusted second CSI is:

[0215] CSI′=CSI-(ΔCSI NW +ΔCSI UE );

[0216] Wherein, ΔCSI NW It is the global compensation value, ΔCSI UE It is the local compensation value derived by the first device based on channel measurements.

[0217] Optionally, the first device can determine the local compensation value based on factors such as SNR difference. The first device can also perform event-triggered compensation. As mentioned earlier, the first device can determine whether the current channel state triggers a compensation update. Combining the local compensation value, the first device can calculate the compensation value based on a threshold and use the threshold as a parameter for triggering the event. For example, ΔCSI... UE It can be:

[0218] Where β represents the channel change trigger threshold, k represents the channel quality coefficient, and SNR is... UE The first device measures the channel quality, SN R. NW,avg SNR represents the average channel quality estimated by the second device. UE-pre This indicates the historical channel quality of the first device.

[0219] The model adjustment method for CSI compression has been introduced above with reference to Figures 10 to 12. For ease of understanding, the adjustment method is illustrated below with reference to Figure 13. Compared with Figure 12, in Figure 13, the CSI generation sub-model on the terminal device side will be adjusted according to adjustment amount 1 (first adjustment amount), and the CSI reconstruction sub-model on the network device side will be adjusted according to adjustment amount 2 (second adjustment amount).

[0220] As shown in Figure 13, adjustment amount 1 can be determined based on the output of the CSI-generated partial sub-model (first model) and / or the CSI-reconstructed partial sub-model (second model). Adjustment amount 2 can be determined based on the output of the CSI-reconstructed partial sub-model (second model).

[0221] The preceding text, with reference to Figures 10 to 13, introduced the bilateral AI models (the first model and the second model) in CSI compression. As mentioned earlier, data collection is also a crucial consideration for CSI compression. In some embodiments, the data collection method in CSI compression can refer to that in CSI prediction. In other embodiments, data collection in CSI compression needs to consider more factors to support the interaction between the bilateral models.

[0222] The first data related to the first model and / or the second model can refer to any data collected during the data collection process for training and inference of the first model and / or the second model. That is, the first data is any data collected during the data collection phase that supports CSI compression.

[0223] In some embodiments, the first data may include corresponding conditional information to enable the first and second models to determine the necessary information related to CSI compression. The conditional information includes, for example, the type of CSI to be compressed, the original channel or precoding matrix, and the CSI configuration. The CSI configuration includes, for example, the number of antenna ports, the number of subbands, and the rank.

[0224] As an example, condition information is used to describe configuration information directly related to CSI data compression, specifying the data structure and characteristics. Fields in the condition information may include one or more of the following: CSI type, number of antenna ports, number of subbands, rank, frequency bandwidth, and time resource configuration. Table 1 provides an example of condition information fields.

[0225] Table 1

[0226] In some embodiments, the first data may include first data information to describe contextual information related to a cell, site, scene, or device for localization model adaptation and personalization optimization. The first data information is also referred to as additional conditional information or added conditional information. For model localization, the first data information may include cell / site / scene-related information such as cell / site-scene identifier (ID), indoor / outdoor indication, line-of-sight (LoS) / non-line-of-sight (NLoS) flags, and UE ID.

[0227] As an example, the first data information may include one or more of the following: the identifier (ID) of the cell corresponding to the first data; the network device scenario identifier corresponding to the first data; the environment type corresponding to the first data; whether the channel corresponding to the first data is line-of-sight or non-line-of-sight; and the identifier of the terminal device corresponding to the first data. The fields of the first data information may be as shown in Table 2.

[0228] Table 2

[0229] As one example, the first data information can be used by the network device to classify and store the collected data. After collecting the data, the network device can classify and store it according to conditional information and / or the first data information, and use the data to train a localized model. For example, the network device can classify the data into specific environmental data according to the "cell ID" and "site-scene ID" and store the data accordingly. As another example, the network device can associate the "LoS / NLoS flag" with CSI measurement results to analyze channel characteristics under different propagation conditions. Furthermore, the network device can use "environment type" and "scene ID" to improve the model's adaptability under specific conditions.

[0230] In some embodiments, the first data may include corresponding condition information and first data information. That is, any data in the data collection phase that supports CSI compression needs to be accompanied by "condition information" and "attached condition information".

[0231] As one embodiment, the first device can perform CSI measurements based on condition information. During data recording, first data information is appended to facilitate the network side in distinguishing data sources for different scenarios. The first device can then feed back the CSI data, along with the condition information and the first data information, to the network device via the uplink channel.

[0232] In some embodiments, condition information and / or first data information can be indicated via higher-level signaling so that the terminal device is aware of the context of data collection. For example, configuration signaling for the first data information can be indicated via the RRC layer and the MAC layer.

[0233] As mentioned earlier, a key technical issue in CSI compression based on a bilateral AI model is how to achieve quantization alignment between the first and second devices. In other words, the CSI generation model on the first device side and the CSI reconstruction model on the second device side need to be aligned in quantization. To address this, quantization alignment between the first and second devices can be achieved through model pairing or standardized quantization schemes. As an example, the quantization scheme between the first and second devices can include scalar quantization (SQ) and vector quantization (VQ). To ensure alignment between quantization performed by the first model and dequantization performed by the second model, quantization and dequantization must be performed based on the same codebase.

[0234] In some embodiments, CSI compression may be based on SQ and / or VQ based on a first codebook. The first codebook may also be referred to as the first parameter or dictionary. For quantization alignment using a standardized quantization scheme, the first codebook may include information related to configuring / reporting / updating quantization.

[0235] As an example, the first codebook can be an SQ codebook or a VQ codebook.

[0236] As one embodiment, the first codebook is a pre-configured codebook or a codebook defined by the protocol. For example, when the first codebook is an SQ codebook, the first codebook can be a pre-configured fixed codebook or a fixed codebook defined by the protocol.

[0237] As one embodiment, the first codebook can be transmitted together with either the first or second model. Transmitting the first codebook together with either the first or second model can be understood as transmitting the first codebook along with the model when the first and second devices are exchanging data. For example, if the first codebook is a VQ codebook, since it's difficult to specify a fixed codebook when training the VQ codebook alongside the AI / ML model, alignment can be ensured by having the model training side indicate the VQ codebook to the other side. The transmission of the first codebook may include the transmission of the codebook format and codebook size.

[0238] In the above embodiments, the training entity / side can send the first codebook along with the model to another party, such as the terminal device side sending it to the network device side. As an example, the entity / side performing the first step of training on the model (such as the network device side) can send the first codebook to the counterpart side performing the second step of training (such as the terminal device side), while the counterpart side only trains the model but keeps the codebook unchanged.

[0239] In some embodiments, when the first codebook is a VQ codebook, it may include multiple sub-codebooks. These sub-codebooks are multiple VQ codebooks with short VQ vectors. Taking an encoder as an example, the encoder's direct output (i.e., the latent space) has a relatively large dimension, making it unlikely to generate VQ codebooks with extremely long VQ vectors / extremely large codebook sizes that directly map the entire latent space. To address this issue, the encoder's long latent space can be segmented into multiple segments, each of which can be mapped to each VQ codebook with short VQ vectors. This method segments the latent space to form multiple segments. Since each segment has a low dimension, a separate short VQ codebook can be trained, making VQ more efficient and easier to implement.

[0240] As an example, the latent space can be segmented using a uniform partitioning method. After partitioning, each segment can be independently generated or mapped into a vector-quantized codebook. That is, each segment corresponds to each sub-codebook.

[0241] As an example, the size of the first codebook is determined based on the sizes of multiple sub-codebooks. Each sub-codebook corresponds to a segment, and the size of any sub-codebook is positively correlated with the dimension of that segment. Assuming the dimension of that segment is N, that segment can generate 2... b A subcodebook (VQ codebook) for each codeword, where b is the codeword length. b is related to N.

[0242] As an example, if multiple segments have similar distribution characteristics, multiple segments can share the same sub-codebook, which helps to reduce the total size of the first codebook.

[0243] In some embodiments, for scalar quantization (SQ), the total payload of SQ is related to the model's output and / or latent dimensions. The latent dimensions may include the dimensions of the latent space. As one embodiment, the size of the total payload of SQ is linearly related to the size of the model's output. As another embodiment, the size of the total payload of SQ is linearly related to the number of bits per latent dimension.

[0244] For example, the total payload size of SQ is positively correlated with the model output size and / or the number of bits per potential dimension.

[0245] For example, suppose the number of bits in each potential dimension is N. bit The output size is z dim Then the total effective payload size of SQ = N bit *z dim .

[0246] In some embodiments, for vector quantization (VQ), the total payload of VQ is related to at least one of the model's output, the piecewise parameters of the latent space, and the latent dimensions. As an example, the total payload of VQ is linearly related to the model's output size. As an example, the total payload of VQ is linearly related to the number of bits per latent dimension. As an example, the total payload of VQ is linearly related to the number of pieces in the latent space. As an example, the total payload of VQ is linearly related to the piecewise size of the latent space. As an example, the total payload of VQ is linearly related to the number of bits per piece.

[0247] For example, the total payload size of VQ is positively correlated with the model output size and / or the number of bits per latent dimension. Similarly, the total payload size of VQ is positively correlated with the number of segments in the latent space and / or the number of bits per segment.

[0248] For example, let the segment size be N. seg The number of bits in each potential dimension is still N. bit Number of bits per segment = N bit *N seg Number of segments = z dim / N seg Then the total payload size of VQ = number of segments * number of bits per segment.

[0249] In some embodiments, the AI / ML model may remain unchanged to account for changes in channel state, but the first codebook will be updated and needs to be indicated to the other party to adapt to the channel changes. Updating the first codebook may involve reconfiguring the quantization granularity of the SQ codebook, meaning the quantization granularity of the SQ codebook is related to the channel state. Updating the first codebook may also involve re-indicating the VQ codebook.

[0250] As one embodiment, in response to a change in the state of the first channel, the first device may adjust the quantization granularity of the SQ codebook; and / or, in response to a change in the state of the first channel, the first device may adjust or update the VQ codebook.

[0251] As one example, the first device can dynamically adjust the VQ codebook based on a feedback mechanism, thereby adapting to channel changes while maintaining model stability. For instance, the first device can dynamically adjust the VQ codebook when reporting codebook performance to the second device.

[0252] In some embodiments, quantization alignment involves not only the alignment between the first and second devices, but also the alignment between third-party servers or different vendors. For example, collaboration between different terminal device or network equipment vendors. A third-party server is, for example, an OTT server.

[0253] As one example, quantization alignment can have two options: a standardized quantization codebook and a proprietary codebook with codebook exchange, as described in the method for determining the first codebook. Since standardized signaling can be used to exchange intermediate datasets or decoders, the transmission of a proprietary codebook is supported. For example, the network side can exchange the proprietary codebook with the terminal device side. When the terminal device performs encoder training, it can freeze the decoder and the quantization codebook. Compared to fixed quantization methods, optimal quantization performance can be obtained when the quantization method itself is trained with the model. Alternatively, quantization can be considered as part of the model design. The transmission of the quantization codebook can be implemented together with encoder model / parameter sharing.

[0254] In some embodiments, the encoder-decoder pair in CSI compression can be viewed as quantization of a continuous space, where the encoder divides the continuous space into multiple intervals, and the decoder provides a representative value for each interval. Ideal quantization (i.e., segmentation and mapping to representative values) can be optimized based on the data distribution. This data distribution corresponds to the initial training of datasets B and A, based on data from the network device side. If the actual data distribution does not match the quantization design, it can lead to a degraded model or system performance.

[0255] As an example, network device-side data (dataset A) may or may not include terminal device-side data (dataset B). If dataset A includes dataset B, it means that the network device has collected data from the terminal device in a timely manner. Depending on additional conditions on the network device side, dataset A and dataset B are aligned in this scenario. However, due to time-varying channel characteristics and additional conditions on the terminal device side, dataset A and dataset B may become mismatched. These additional conditions on the terminal device side include the SVD phase normalization method, terminal device antenna virtualization, antenna imbalance, antenna spacing / layout, etc.

[0256] As an example, additional conditions on the terminal device side can lead to dataset mismatches. This issue can be addressed by timely collection of network device-side data for new terminal device vendors, provided it demonstrates good generalization performance across various terminal device-side conditions. However, timely network device-side data collection is not a scalable solution because new chipsets or new terminal devices will enter the market more frequently. Furthermore, after data collection on the network device side, the network device may update its decoder and exchange the new dataset with existing and new terminal devices for a new round of offline engineering.

[0257] In the above embodiments, the dataset for retraining the first model on the terminal device side may have two sources. One is shared data from the network device side, which needs to include at least additional information about the target CSI; the other is data collection on the terminal device side. Parameter exchange between the first or second model can be performed offline or via an air interface. If parameters are exchanged offline, both parameter exchange and subsequent engineering can be performed offline. However, whether parameters are exchanged offline or via an air interface, it can lead to varying degrees of complexity in inter-vendor collaboration. For example, if an offline method is used, the offline parameter exchange from the network device side to the terminal device side involves the complexity of collaboration between different vendors. Specifically, a terminal device vendor may need to collaborate with multiple network device vendors, thus increasing the complexity of offline collaboration, and vice versa. In contrast, if parameter exchange is performed via an air interface, procedures and signaling can be specified, which can alleviate the complexity of inter-vendor collaboration to some extent.

[0258] To address the aforementioned issues, different vendors can design encoders or decoders based on a standardized reference model architecture. Once the model architecture is standardized, the complexity of collaboration between vendors can be alleviated to some extent. In other words, the first model can be designed based on a common reference model to align with vendors of different terminal devices.

[0259] To address the aforementioned issues, parameter exchange can involve only a subset of parameters. For example, in the first model, only the parameters of the CSI generation component are exchanged from the network device side to the terminal device side. Therefore, the terminal device-side / terminal device-side OTT server only needs to retrain the CSI generation component.

[0260] To address the complexity of inter-vendor collaboration in bilateral CSI compression, the following section provides an exemplary solution to this problem, using the training of the encoder or decoder on the terminal device side or network device side as an example, combined with several embodiments.

[0261] Example 1: Encoder training on the terminal device side

[0262] As shown in Figure 9, there are two alternative schemes for encoder retraining: one is to first train the nominal decoder on the terminal device side, and then retrain the actual encoder relative to the nominal decoder; the other is to directly develop and train the actual encoder on the terminal device side.

[0263] During encoder training, one or more of the following additional information can be shared from the network device side to the end device side: performance objectives, dataset, and information related to the collected dataset. This additional information can be used on the end device side to retrain / redevelop the actual encoder. If performance objectives are shared, they can be incorporated into encoder retraining as performance guidelines. Whether information about the dataset or related to the collected dataset can be considered additional information depends on whether the end device side can train a high-performance encoder without exchanging the dataset with the network device side. In particular, if additional conditions on the end device side cause a data distribution mismatch between dataset A and dataset B, the additional information needs to include the dataset or related data.

[0264] As an example, the dataset in the additional information can be dataset A or a dataset related to dataset A. Dataset A is, for example, {target CSI (first CSI), CSI feedback (second CSI)} on the network device side. Optionally, dataset A can be a dedicated dataset for data collection during the model training phase, or it can be a traditional historical dataset.

[0265] As an example, the additional information may not include dataset B. Dataset B is, for example, the {target CSI} collected on the terminal device side. For encoder training on the terminal device side, the network device does not need to provide dataset B as additional information.

[0266] As an example, additional information may include dataset B. Optionally, dataset B may be a dedicated dataset for data collection during the model training phase, or it may be a traditional historical dataset.

[0267] In Example 1, the network device trains a decoder on dataset A, while the terminal device trains an encoder based on dataset A, dataset B, or both. The choice of dataset for encoder training depends on inter-vendor collaboration, training options, and the selection of the training target CSI.

[0268] Example 2: Encoder-decoder training on the network device side

[0269] The network device can train an encoder-decoder on dataset A and transmit the encoder parameters to the terminal device. Upon receiving the encoder parameters from the network device, the terminal device can directly perform local inference. The terminal device and network device can exchange parameters via the over-the-air interface for model transmission / delivery.

[0270] As an example, the encoder trained on the network device side can be a universal encoder to facilitate use by different terminal devices from different vendors. The feasibility of using a universal encoder across different terminal devices needs to consider performance and terminal device capabilities.

[0271] Regarding encoder performance, different end devices may be located in different scenarios, such as LosS / NLoS and indoor / outdoor environments. Therefore, the design and training of a universal encoder need to take into account the performance impact of different scenarios.

[0272] Regarding the capabilities of terminal devices, different terminal devices may have different capabilities in terms of AI / ML model complexity. If the complexity of a universal encoder exceeds the capabilities of one or more terminal devices, then such a universal encoder is not feasible.

[0273] As one example, a network device can train multiple encoders for different terminals. The feasibility of this embodiment also needs to consider both performance and the capabilities of the terminal devices.

[0274] Regarding encoder performance, it is feasible to train multiple encoders using different datasets from different end devices. For example, a network device can obtain the same encoder structure with different parameters and then exchange it with different end devices.

[0275] Regarding the capabilities of terminal devices, different terminal devices may require different encoder structures with varying model complexities. However, this will lead to more standardization work on model structures.

[0276] Example 3: Training a specified encoder or decoder

[0277] When a terminal device trains its actual encoder while a network device trains its actual decoder, there may be compatibility issues between the actual encoder on the terminal device side and the decoder on the network device side, especially when the data distribution on the terminal device side does not match the data distribution on the network device side.

[0278] In some embodiments, to meet minimum requirements, the terminal device can directly deploy a specified encoder, which simplifies the operation on the terminal device side. Furthermore, because the encoder is specified, different terminal devices and terminal device vendors will have the same encoder, thus allowing the network device side to train a compatible network device-side decoder for all terminal devices.

[0279] In some embodiments, network devices can directly deploy designated decoders, while different terminal devices / terminal device vendors / chipset vendors can train compatible encoders on the terminal device-side server.

[0280] In CSI compression based on a bilateral AI model, the encoder is used for CSI generation, and the decoder is used for CSI reconstruction (or rebuilding). A fully specified encoder or decoder completes the designated CSI generation or CSI reconstruction portion. By fully specifying the CSI generation (generation) portion, performance evaluation of the model's training / fine-tuning can be based on simulated and / or real-world field data when developing / training a compatible CSI reconstruction (generation) portion on the network device (end device) side. This simulation can simulate data distribution mismatches between synthetic and field data, for example, by simulating data distribution mismatches between network device-side field data and end device-side field data.

[0281] Example 4: Reference encoder or reference decoder

[0282] The encoder on the terminal device side and / or the decoder on the network device side can be trained based on a reference model. The reference model can be either a reference encoder or a reference decoder. For example, the encoder on the terminal device side can be trained based on a reference decoder. Similarly, the decoder on the network device side can be trained based on a reference encoder.

[0283] As an example, the first model on the first device side is trained based on a reference model. This reference model is a standardized reference model.

[0284] The reference model can be trained on a general-purpose dataset. As an example, this general-purpose dataset is an offline dataset. An offline general-purpose dataset is essential for the reference model; otherwise, it cannot be trained with standardized parameters. However, obtaining an offline general-purpose dataset is a critical issue because mismatches between the dataset distribution used to train the reference model and the data distributions on the terminal device and network device sides can impact performance.

[0285] As an example, the reference model can be trained on an offline general dataset (referred to as dataset closure). To enable training of the terminal device-side encoder, the following additional information can be specified: target CSI (first CSI), reference decoder, cumulative information of historical CSIs, SGCS target value, error feedback information, etc. For example, the training of the first model can be based on one or more of the following additional information: first CSI, historical CSIs, reference decoder, SGCS target value, and error feedback information related to the first adjustment and / or the second adjustment.

[0286] As one embodiment, the encoder on the first device can be one of multiple encoders. These multiple encoders can be used on multiple terminal devices, including the first device. The multiple encoders can be trained based on a reference encoder and / or a reference decoder. The reference encoder and / or reference decoder can be trained based on a general dataset. As described above, this general dataset can be an offline general dataset.

[0287] As one embodiment, the decoder on the second device can be one of multiple decoders. These multiple decoders can be used by multiple network devices or multiple terminal devices communicating with the first device. The multiple decoders can be trained based on a reference decoder and / or a reference encoder.

[0288] In the above embodiments, the additional information can enable encoder training, verification, and testing on the terminal device side. Correspondingly, the additional information including the reference encoder can enable decoder training, verification, and testing on the network device side. For example, the actual CSI generation part corresponding to the first model can be trained on the terminal device side based on the specified reference encoder / decoder. Similarly, the actual CSI reconstruction part corresponding to the second model can be trained on the network device side according to the specified reference encoder / decoder.

[0289] In the above embodiments, after separate training on the terminal device side, the terminal device only knows the performance of (the actual CSI generation part + the specified reference decoder). Similarly, after separate training on the network device side, the network device only knows the performance of (the specified reference encoder + the actual CSI reconstruction part). Neither the terminal device nor the network device knows whether (the actual CSI generation part + the actual CSI reconstruction part) performs well after separate training. Therefore, a verification procedure should be performed after training and before actual deployment to confirm the performance of (the actual CSI generation part + the actual CSI reconstruction part). This verification process can refer to the model monitoring mechanisms on the network device side and the terminal device side.

[0290] For network device-side verification, the terminal device reports the target CSI label to the network device, which then verifies the performance (actual CSI generation part + actual CSI reconstruction part) by calculating the SGCS of the target label and the output of the actual CSI reconstruction part. If the SGCS meets the performance target, the actual performance can be approved. In this case, the additional information from the terminal device to the network device includes at least the target label.

[0291] For network device-side verification, the network device sends the output of the actual CSI reconstruction portion to the terminal device. The terminal device can then verify the performance of (actual CSI generation portion + actual CSI reconstruction portion) by calculating the SGCS of the target label and the output of the actual CSI reconstruction portion. Furthermore, performance targets should be indicated from the network device to the terminal device for verification. In this case, the additional information from the network device to the terminal device includes at least the output of the actual CSI reconstruction portion and the performance targets.

[0292] As an example, the encoder or decoder model structure can be designed based on the structure of a standardized reference model. For scalability studies of the encoder or decoder model structure, choices based on token and / or feature dimensions can be introduced. For instance, the encoder or decoder model structure can use subbands as the token dimension and transmit (Tx) ports as the feature dimension, with the number of tokens varying with the number of subbands. Alternatively, the encoder or decoder model structure can use Tx ports as the token dimension and subbands as the feature dimension, with the number of tokens varying with the number of Tx ports. Yet another example is that the encoder or decoder model structure can use fixed-size sub-blocks (e.g., n_Tx_ports*m_subbands) of Tx ports and subband matrices as tokens, representing the input as a token sequence. Here, the number of tokens varies with the number of Tx ports and the number of subbands.

[0293] In some implementations, to achieve scalability across the feature dimension, each feature size may include a specific embedding layer and / or a generic embedding layer with padding. This padding may correspond to zero padding or other padding value techniques.

[0294] In some implementations, to achieve scalability at the token dimension, embedding can be done at locations specific to each token index, enabling load-configuration-based scalability. For example, token dimension scaling can be implemented based on one or more of the following: a specific output linear layer for each load configuration, truncation / masking of the output linear layer output, and changes to quantization parameters.

[0295] As one possible implementation, the first model is based on a reference model with a first extension. The first extension is implemented based on the token dimension and / or feature dimension. The token dimension corresponds to the number of sub-bands and / or the number of transmit ports. The feature dimension corresponds to the number of transmit ports and / or the number of sub-bands. For example, the encoder on the first device side is based on the reference encoder with a first extension.

[0296] In some implementations, for scalable model architectures (e.g., encoder or decoder model architectures), model architectures with fixed hyperparameters can be supported. For example, assuming a general model architecture, a single parameter set or different parameter sets can be trained on different {number of Tx ports, CSI feedback payload size, bandwidth}. In this scenario, it can also be reported whether a single parameter set or different parameter sets were used on different {number of Tx ports, CSI feedback payload size, bandwidth}. A single parameter set is, for example, a single parameter set spanning different payload sizes and bandwidths; different parameter sets are, for example, different parameter sets spanning different numbers of Tx ports.

[0297] In some implementations, scalable model architectures can also support model structures with different hyperparameters. For example, the hyperparameters associated with the input and output of the model architecture are selected as the best correspondence for each specific {number of Tx ports, CSI feedback payload size, bandwidth}, regardless of scalability. Furthermore, for scalable model architectures, different parameter sets can be trained with different numbers of Tx ports, CSI feedback payload sizes, and bandwidths.

[0298] In some implementations, for each choice of scalable CSI compression model architecture, training can be performed for each {number of Tx ports, CSI feedback payload size, bandwidth}, and the associated gain can be reported. For example, for each {number of Tx ports, CSI feedback payload size, bandwidth}, the average gain (%) relative to the non-AI / ML standard, along with the loss (%) of the average gain, can be reported in SGCS. Specifically, the SGCS gain (%) for each {number of Tx ports, CSI feedback payload size, bandwidth} is first calculated, and then the SGCS gain values ​​(%) are averaged across multiple {number of Tx ports, CSI feedback payload size, bandwidth}.

[0299] As one possible implementation, the first model is a second extension based on the reference model. The second extension is trained using a single parameter set and / or multiple parameter sets based on the structural parameters of the reference model. The structural parameters of the reference model may correspond to at least one of the following: the number of Tx ports, the size of the CSI feedback payload, and the bandwidth.

[0300] Based on the above embodiments, datasets, parameters, or reference models can be exchanged using over-the-air signaling or offline signaling. When training on the terminal device side is offline, the training dataset, parameters, or reference model must be sent to the training server on the terminal device side. As one embodiment, the serving cell can configure the terminal device to provide information for transmitting / transferring training datasets, parameters, or reference models to the terminal device side. For example, the serving cell can provide this configuration information through operation administration and maintenance (OAM) or network function (NF) to transmit / transfer training datasets, parameters, or reference models.

[0301] As one example, the training dataset, parameters, or reference model can be transferred / transmitted from a network entity (e.g., RAN / OAM / NF) to a training server area (within or outside the 3GPP network) on the terminal device side. This training server is, for example, an OTT server.

[0302] In some embodiments, an important piece of information needed when collecting data on the terminal device side is pairing information, which helps in labeling data, training models, and configuring post-trained inference. As one example, each training server can serve as an intermediate dataset, and network devices can share a pairing ID for each intermediate dataset or decoder.

[0303] As one example, a pairing ID can be used to distinguish different datasets used in network device-side decoder training and reference encoder training. Different pairing IDs can also be associated with different models. For example, an end device can use the indicated pairing ID to request a data collection reference signal for data collection on the end device side. The network device side then sends a data collection reference signal accordingly. These reference signals can be transmitted in the same manner as the network device-side data used for data collection and associated with the same pairing ID.

[0304] As one example, the terminal device can decide to develop multiple generic models on the indicated pairing ID. The pairing ID can serve as a bridge for model identification and inference configuration.

[0305] As an example, when there are multiple standardized reference models, each reference model can be assigned a pairing ID. The network device can train a decoder for each specified encoder through network device-side data collection. The terminal device can use the pairing ID to request the collection of reference signals, and the reference signals sent by the network device can be the same as those used for network device-side data collection.

[0306] As an example, when training is performed only on the network device side, the encoder model / parameters can be directly transmitted to the terminal device for inference. Pairing IDs can also be used in the transmission of encoder model / parameters. For example, after the terminal device receives the model / parameters for the first time, it can store them in its memory. The next time the network device wants to transmit the model / parameters to the terminal device, the network device can first query the terminal device by indicating the pairing ID whether the model / parameters need to be transmitted. If the terminal device already has the model / parameters in its memory, then transmission is unnecessary. Encoder / decoder compatibility can be achieved through model / parameter sharing or intermediate datasets generated from the training model or a specified model; therefore, using pairing IDs to maintain consistency between training and inference is more convenient and efficient. Therefore, this information needs to be provided in reference signal requests, reference signal configurations, data collection, and inference configurations.

[0307] The method embodiments of this application have been described in detail above with reference to Figures 4 to 13. The apparatus embodiments of this application will now be described in detail with reference to Figures 14 to 16. It should be understood that the descriptions of the apparatus embodiments correspond to the descriptions of the method embodiments; therefore, any parts not described in detail can be referred to the preceding method embodiments.

[0308] Figure 14 is a schematic block diagram of a wireless communication device according to an embodiment of this application. The device 1400 can be any of the terminal devices described above. The device 1400 shown in Figure 14 includes a processing unit 1410 and a transmitting unit 1420.

[0309] The processing unit 1410 can be used to perform CSI compression on the first CSI using the first model to obtain the second CSI.

[0310] The transmitting unit 1420 can be used to transmit the second CSI to the second device; wherein the first CSI is related to the first channel, and the second model corresponding to the second device is used to determine the third CSI based on the second CSI, wherein the third CSI is determined based on the first adjustment amount of the first model and / or the second adjustment amount of the second model.

[0311] Optionally, the first CSI includes one or more predicted CSIs, and / or the first CSI includes one or more historical CSIs of the first channel.

[0312] Optionally, the first adjustment amount is determined based on the output of the first model and / or the output of the second model, and the second adjustment amount is determined based on the output of the second model.

[0313] Optionally, the first adjustment amount and / or the second adjustment amount are related to one or more of the following: intermediate evaluation parameters corresponding to the CSI compression; auxiliary information of the second device; wherein the auxiliary information is used to compensate for the feature vector difference between the third CSI and the first CSI.

[0314] Optionally, the intermediate evaluation parameter is SGCS, which is used to adjust the second CSI. The adjusted value of the second CSI is: CSI′=α(SGCS)*CSI+(1-α(SGCS))*ΔCSI;

[0315] Wherein, CSI is the initial value before the second CSI adjustment, α(SGCS) = SGCSγ, γ is the adjustment parameter, and ΔCSI is the measured result of the first channel or determined according to the instructions of the network device.

[0316] Optionally, the auxiliary information is used to determine one or more of the following: the historical CSI offset value corresponding to the first channel; the weighting factor of the second CSI; and the compensation reference value corresponding to the first channel.

[0317] Optionally, the historical CSI offset value is determined based on the historical value of the first CSI and the historical value of the third CSI, and the historical CSI offset value is used for multiple terminal devices including the first device.

[0318] Optionally, the weighting factor is determined based on the correlation between the third CSI and the first CSI, or the weighting factor is determined based on the channel quality of the first channel at different times.

[0319] Optionally, the weighting factor of the second CSI at time t for:

[0320] Where α is the smoothing coefficient. The weighting factor at time t-1 SNR(t) and SNR(t-1) are the signal-to-noise ratios of the first channel estimated by the second device at time t and time t-1.

[0321] Optionally, the device 1400 further includes a receiving unit, which can be used to receive indication information from the second device; wherein the indication information is used to indicate the auxiliary information and / or the first adjustment amount.

[0322] Optionally, in response to the first condition being met, the first adjustment amount is 0, and / or the second adjustment amount is 0.

[0323] Optionally, the second CSI is sent via a CSI report, the sending of which is related to at least one of the following:

[0324] Whether to send is determined based on an event-triggered method, wherein the event includes the prediction result of the first channel;

[0325] The first priority of the CSI report is related to the first model.

[0326] Optionally, the CSI report is generated based on a multi-level mapping method, which includes a first-level mapping and a second-level mapping. The first-level mapping corresponds to the channel characteristics of the first channel, and the second-level mapping corresponds to the subcarriers or antennas associated with the first channel.

[0327] Optionally, the CSI report includes differential information between the current CSI and the previous CSI, or the CSI report includes partial information extracted by the first model based on the first CSI.

[0328] Optionally, the first data related to the first model and / or the second model includes first data information. The first data information is used by the network device to classify and store the collected data. The first data information includes one or more of the following: the identifier of the cell corresponding to the first data; the network device scenario identifier corresponding to the first data; the environment type corresponding to the first data; the identifier of whether the channel corresponding to the first data is line-of-sight or non-line-of-sight; and the identifier of the terminal device corresponding to the first data.

[0329] Optionally, the CSI compression is based on scalar quantization and / or vector quantization using a first codebook, where the first codebook is a pre-configured or protocol-defined codebook, or the first codebook is transmitted together with the first model or the second model.

[0330] Optionally, the first codebook includes a scalar quantization codebook and / or a vector quantization codebook, and the processing unit 1410 is further configured to adjust the quantization granularity of the scalar quantization codebook in response to a state change of the first channel; and / or, the processing unit 1410 is further configured to adjust or update the vector quantization codebook in response to a state change of the first channel.

[0331] Optionally, when the first codebook is a vector quantization codebook, the first codebook includes multiple sub-codebooks, any one of the multiple sub-codebooks corresponds to a segment of the output of the first model, and the size of any one sub-codebook is positively correlated with the dimension of the segment.

[0332] Optionally, the first model is an AI / ML-based encoder, and the second model is an AI / ML-based decoder.

[0333] Optionally, the encoder is one of a plurality of encoders used in a plurality of terminal devices, the plurality of encoders being trained based on a reference encoder and / or a reference decoder, the reference encoder and / or the reference decoder being trained based on a common dataset.

[0334] Figure 15 is a schematic block diagram of another device for wireless communication according to an embodiment of this application. The device 1500 can be any of the network devices described above. The device 1500 shown in Figure 15 includes a receiving unit 1510 and a processing unit 1520.

[0335] The receiving unit 1510 can be used to receive a second CSI from the first device.

[0336] The processing unit 1520 can be used to process the second CSI through a second model to obtain a third CSI; wherein the second CSI is determined by CSI compression of the first CSI through a first model, the first model corresponds to the first device, the first CSI is related to the first channel, and the third CSI is determined according to a first adjustment amount of the first model and / or a second adjustment amount of the second model.

[0337] Optionally, the first CSI includes one or more predicted CSIs, and / or the first CSI includes one or more historical CSIs of the first channel.

[0338] Optionally, the first adjustment amount is determined based on the output of the first model and / or the output of the second model, and the second adjustment amount is determined based on the output of the second model.

[0339] Optionally, the first adjustment amount and / or the second adjustment amount are related to one or more of the following: intermediate evaluation parameters corresponding to the CSI compression; auxiliary information of the second device; wherein the auxiliary information is used to compensate for the feature vector difference between the third CSI and the first CSI.

[0340] Optionally, the intermediate evaluation parameter is SGCS, which is used to adjust the second CSI. The adjusted value of the second CSI is: CSI′=α(SGCS)*CSI+(1-α(SGCS))*ΔCSI;

[0341] Wherein, CSI is the initial value before the second CSI adjustment, α(SGCS) = SGCSγ, γ is the adjustment parameter, and ΔCSI is the measured result of the first channel or determined according to the instructions of the network device.

[0342] Optionally, the auxiliary information is used to determine one or more of the following: the historical CSI offset value corresponding to the first channel; the weighting factor of the second CSI; and the compensation reference value corresponding to the first channel.

[0343] Optionally, the historical CSI offset value is determined based on the historical value of the first CSI and the historical value of the third CSI, and the historical CSI offset value is used for multiple terminal devices including the first device.

[0344] Optionally, the weighting factor is determined based on the correlation between the third CSI and the first CSI, or the weighting factor is determined based on the channel quality of the first channel at different times.

[0345] Optionally, the weighting factor of the second CSI at time t for:

[0346] Where α is the smoothing coefficient. The weighting factor at time t-1 SNR(t) and SNR(t-1) are the signal-to-noise ratios of the first channel estimated by the second device at time t and time t-1.

[0347] Optionally, the device 1500 further includes a transmitting unit, which can be used to transmit indication information to the first device; wherein the indication information is used to indicate the auxiliary information and / or the first adjustment amount.

[0348] Optionally, in response to the first condition being met, the first adjustment amount is 0, and / or the second adjustment amount is 0.

[0349] Optionally, the second CSI is transmitted via a CSI report, the transmission of which is related to at least one of the following: determining whether to transmit based on an event-triggered method, the event including the prediction result of the first channel; and a first priority of the CSI report, the first priority being related to the first model.

[0350] Optionally, the CSI report is generated based on a multi-level mapping method, which includes a first-level mapping and a second-level mapping. The first-level mapping corresponds to the channel characteristics of the first channel, and the second-level mapping corresponds to the subcarriers or antennas associated with the first channel.

[0351] Optionally, the CSI report includes differential information between the current CSI and the previous CSI, or the CSI report includes partial information extracted by the first model based on the first CSI.

[0352] Optionally, the first data related to the first model and / or the second model includes first data information. The first data information is used by the network device to classify and store the collected data. The first data information includes one or more of the following: the identifier of the cell corresponding to the first data; the network device scenario identifier corresponding to the first data; the environment type corresponding to the first data; the identifier of whether the channel corresponding to the first data is line-of-sight or non-line-of-sight; and the identifier of the terminal device corresponding to the first data.

[0353] Optionally, the CSI compression is based on scalar quantization and / or vector quantization using a first codebook, where the first codebook is a pre-configured or protocol-defined codebook, or the first codebook is transmitted together with the first model or the second model.

[0354] Optionally, the first codebook includes a scalar quantization codebook and / or a vector quantization codebook, wherein the quantization granularity of the scalar quantization codebook is adjusted in response to a change in the state of the first channel; and / or, the vector quantization codebook is adjusted or updated in response to a change in the state of the first channel.

[0355] Optionally, when the first codebook is a vector quantization codebook, the first codebook includes multiple sub-codebooks, any one of the multiple sub-codebooks corresponds to a segment of the output of the first model, and the size of any one sub-codebook is positively correlated with the dimension of the segment.

[0356] Optionally, the first model is an AI / ML-based encoder, and the second model is an AI / ML-based decoder.

[0357] Optionally, the encoder is one of a plurality of encoders used in a plurality of terminal devices, the plurality of encoders being trained based on a reference encoder and / or a reference decoder, the reference encoder and / or the reference decoder being trained based on a common dataset.

[0358] Figure 16 is a schematic diagram of the structure of a communication device according to an embodiment of this application. The dashed lines in Figure 16 indicate that the unit or module is optional. This device 1600 can be used to implement the methods described in the above method embodiments. Device 1600 can be a chip, a terminal device, or a network device.

[0359] Apparatus 1600 may include one or more processors 1610. The processor 1610 may support apparatus 1600 in implementing the methods described in the preceding method embodiments. The processor 1610 may be a general-purpose processor or a special-purpose processor. For example, the processor may be a central processing unit (CPU). Alternatively, the processor may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0360] The apparatus 1600 may further include one or more memories 1620. The memories 1620 store a program that can be executed by the processor 1610, causing the processor 1610 to perform the methods described in the preceding method embodiments. The memories 1620 may be independent of the processor 1610 or integrated within the processor 1610.

[0361] The device 1600 may also include a transceiver 1630. The processor 1610 can communicate with other devices or chips via the transceiver 1630. For example, the processor 1610 can send and receive data with other devices or chips via the transceiver 1630.

[0362] This application also provides a computer-readable storage medium for storing a program. This computer-readable storage medium can be applied to a terminal device or network device provided in this application, and the program causes a computer to execute the methods performed by the terminal device or network device in various embodiments of this application.

[0363] The computer-readable storage medium can be any available medium that a computer can read, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital video discs, DVDs), or semiconductor media (e.g., solid-state disks, SSDs), etc.

[0364] This application also provides a computer program product. The computer program product includes a program. This computer program product can be applied to a terminal device or network device provided in this application embodiment, and the program causes a computer to execute the methods performed by the terminal device or network device in the various embodiments of this application.

[0365] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0366] This application also provides a computer program. This computer program can be applied to a terminal device or network device provided in this application, and the computer program causes the computer to execute the methods performed by the terminal or network device in various embodiments of this application.

[0367] In this application, the terms "system" and "network" are used interchangeably. Furthermore, the terminology used in this application is only for explaining specific embodiments of the application and is not intended to limit the application. The terms "first," "second," "third," and "fourth," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. In addition, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0368] In the embodiments of this application, the term "instruction" can be a direct instruction, an indirect instruction, or an indication of a relationship. For example, A instructing B can mean that A directly instructs B, such as B being able to obtain information through A; it can also mean that A indirectly instructs B, such as A instructing C, so B can obtain information through C; or it can mean that there is a relationship between A and B.

[0369] In the embodiments of this application, the term "correspondence" may indicate a direct or indirect correspondence between two things, or an association between two things, or a relationship such as instruction and being instructed, configuration and being configured.

[0370] In the embodiments of this application, "predefined" or "preconfigured" can be implemented by pre-storing corresponding codes, tables, or other means that can be used to indicate relevant information in the device (e.g., including terminal devices and network devices). This application does not limit the specific implementation method. For example, predefined can refer to what is defined in the protocol.

[0371] In the embodiments of this application, the term "protocol" may refer to standard protocols in the field of communications, such as LTE protocols, NR protocols, and related protocols applied in future communication systems. This application does not limit the scope of these protocols.

[0372] In the embodiments of this application, determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information.

[0373] In the embodiments of this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0374] In the embodiments of this application, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0375] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0376] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0377] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0378] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

A method for wireless communication, comprising: include: The first device performs CSI compression on the first channel state information (CSI) using a first model to obtain the second CSI; The first device sends the second CSI to the second device; Wherein, the first CSI is related to the first channel, and the second model corresponding to the second device is used to determine the third CSI based on the second CSI. The third CSI is determined based on the first adjustment amount of the first model and / or the second adjustment amount of the second model. The method of claim 1, wherein The first CSI includes one or more predicted CSIs, and / or the first CSI includes one or more historical CSIs of the first channel. The method according to claim 1 or 2, characterized in that The first adjustment amount is determined based on the output of the first model and / or the output of the second model, and the second adjustment amount is determined based on the output of the second model. The method according to any one of claims 1-3, characterized in that The first adjustment amount and / or the second adjustment amount are related to one or more of the following: The intermediate evaluation parameters corresponding to the CSI compression; Auxiliary information of the second device; The auxiliary information is used to compensate for the feature vector differences between the third CSI and the first CSI. The method according to claim 4, characterized in that The intermediate evaluation parameter is the squared generalized cosine similarity (SGCS), which is used to adjust the second CSI. The adjusted value of the second CSI is: CSI′=α(SGCS)*CSI+(1-α(SGCS))*ΔCSI; Wherein, CSI is the initial value before the second CSI adjustment, α(SGCS) = SGCSγ, γ is the adjustment parameter, and ΔCSI is the measured result of the first channel or determined according to the instructions of the network device. The method according to claim 4, characterized in that The auxiliary information is used to determine one or more of the following: The historical CSI offset value corresponding to the first channel; The weighting factor of the second CSI; The compensation reference value corresponding to the first channel. The method according to claim 6, characterized in that The historical CSI offset value is determined based on the historical value of the first CSI and the historical value of the third CSI, and the historical CSI offset value is used for multiple terminal devices including the first device. The method according to claim 6, characterized in that The weighting factor is determined based on the correlation between the third CSI and the first CSI, or the weighting factor is determined based on the channel quality of the first channel at different times. The method according to claim 8, characterized in that, a weighting factor of the second CSI at time t For: wherein a is a smoothing coefficient, a weighting factor for time t-1, SNR(t) and SNR(t-1) are the signal-to-noise ratios of the first channel estimated by the second device at time t and time t-1. The method according to any one of claims 4-9, characterized in that The method further includes: The first device receives instruction information from the second device; The indication information is used to indicate the auxiliary information and / or the first adjustment amount. The method according to any one of claims 1-10, characterized in that In response to the fulfillment of the first condition, the first adjustment amount is 0, and / or the second adjustment amount is 0. The method according to any one of claims 1 to 11, characterized in that The second CSI is sent via a CSI report, the sending of which is related to at least one of the following: Whether to send is determined based on an event-triggered method, wherein the event includes the prediction result of the first channel; The first priority of the CSI report is related to the first model. The method of claim 12, wherein The CSI report is generated based on a multi-level mapping method, which includes a first-level mapping and a second-level mapping. The first-level mapping corresponds to the channel characteristics of the first channel, and the second-level mapping corresponds to the subcarriers or antennas associated with the first channel. The method according to claim 12 or 13 is characterized in that, The CSI report includes differential information between the current CSI and the previous CSI, or the CSI report includes partial information extracted by the first model based on the first CSI. The method according to any one of claims 1-14 is characterized in that, The first data related to the first model and / or the second model includes first data information, which is used by the network device to classify and store the collected data. The first data information includes one or more of the following: The first data corresponds to the identifier of the cell; The network device scenario identifier corresponding to the first data; The environment type corresponding to the first data; The channel corresponding to the first data is identified as either line-of-sight or non-line-of-sight; The first data corresponds to the identifier of the terminal device. The method according to any one of claims 1-15 is characterized in that, The CSI compression is based on scalar quantization and / or vector quantization using a first codebook, which is a pre-configured or protocol-defined codebook, or the first codebook is transmitted together with the first model or the second model. The method according to claim 16, characterized in that, The first codebook includes a scalar quantization codebook and / or a vector quantization codebook, and the method further includes: In response to a change in the state of the first channel, the first device adjusts the quantization granularity of the scalar quantization codebook; and / or, In response to a change in the state of the first channel, the first device adjusts or updates the vector quantization codebook. The method according to claim 16 or 17, characterized in that, When the first codebook is a vector quantization codebook, the first codebook includes multiple sub-codebooks, each of the multiple sub-codebooks corresponds to a segment of the output of the first model, and the size of each sub-codebook is positively correlated with the dimension of the segment. The method according to any one of claims 1-18, characterized in that, The first model is an encoder based on artificial intelligence (AI) / machine learning (ML), and the second model is a decoder based on AI / ML. The method according to claim 19, characterized in that, The encoder is one of a plurality of encoders used in a plurality of terminal devices. The plurality of encoders are trained based on a reference encoder and / or a reference decoder, which are trained based on a common dataset. A method for wireless communication, characterized in that, include: The second device receives the second channel status information (CSI) from the first device. The second model corresponding to the second device processes the second CSI to obtain the third CSI; The second CSI is determined by CSI compression of the first CSI using a first model, the first model corresponding to the first device, the first CSI being related to the first channel, and the third CSI being determined based on a first adjustment amount of the first model and / or a second adjustment amount of the second model. The method according to claim 21, characterized in that, The first CSI includes one or more predicted CSIs, and / or the first CSI includes one or more historical CSIs of the first channel. The method according to claim 21 or 22 is characterized in that, The first adjustment amount is determined based on the output of the first model and / or the output of the second model, and the second adjustment amount is determined based on the output of the second model. The method according to any one of claims 21-23 is characterized in that, The first adjustment amount and / or the second adjustment amount are related to one or more of the following: The intermediate evaluation parameters corresponding to the CSI compression; Auxiliary information of the second device; The auxiliary information is used to compensate for the feature vector differences between the third CSI and the first CSI. The method according to claim 24, characterized in that, The intermediate evaluation parameter is the squared generalized cosine similarity (SGCS), which is used to adjust the second CSI. The adjusted value of the second CSI is: CSI′=α(SGCS)*CSI+(1-α(SGCS))*ΔCSI; Wherein, CSI is the initial value before the second CSI adjustment, α(SGCS) = SGCSγ, γ is the adjustment parameter, and ΔCSI is the measured result of the first channel or determined according to the instructions of the network device. The method according to claim 24, characterized in that, The auxiliary information is used to determine one or more of the following: The historical CSI offset value corresponding to the first channel; The weighting factor of the second CSI; The compensation reference value corresponding to the first channel. The method according to claim 26, characterized in that, The historical CSI offset value is determined based on the historical value of the first CSI and the historical value of the third CSI, and the historical CSI offset value is used for multiple terminal devices including the first device. The method according to claim 26, characterized in that, The weighting factor is determined based on the correlation between the third CSI and the first CSI, or the weighting factor is determined based on the channel quality of the first channel at different times. The method according to claim 28, characterized in that, The weighting factor of the second CSI at time t for: Where α is the smoothing coefficient. The weighting factor at time t-1 SNR(t) and SNR(t-1) are the signal-to-noise ratios of the first channel estimated by the second device at time t and time t-1. The method according to any one of claims 26-29 is characterized in that, The method further includes: The second device sends an instruction message to the first device; The indication information is used to indicate the auxiliary information and / or the first adjustment amount. The method according to any one of claims 21-30 is characterized in that, In response to the fulfillment of the first condition, the first adjustment amount is 0, and / or the second adjustment amount is 0. The method according to any one of claims 21-31 is characterized in that, The second CSI is sent via a CSI report, the sending of which is related to at least one of the following: Whether to send is determined based on an event-triggered method, wherein the event includes the prediction result of the first channel; The first priority of the CSI report is related to the first model. The method according to claim 32, characterized in that, The CSI report is generated based on a multi-level mapping method, which includes a first-level mapping and a second-level mapping. The first-level mapping corresponds to the channel characteristics of the first channel, and the second-level mapping corresponds to the subcarriers or antennas associated with the first channel. The method according to claim 32 or 33 is characterized in that, The CSI report includes differential information between the current CSI and the previous CSI, or the CSI report includes partial information extracted by the first model based on the first CSI. The method according to any one of claims 21-34 is characterized in that, The first data related to the first model and / or the second model includes first data information, which is used by the network device to classify and store the collected data. The first data information includes one or more of the following: The first data corresponds to the identifier of the cell; The network device scenario identifier corresponding to the first data; The environment type corresponding to the first data; The channel corresponding to the first data is identified as either line-of-sight or non-line-of-sight; The first data corresponds to the identifier of the terminal device. The method according to any one of claims 21-35 is characterized in that, The CSI compression is based on scalar quantization and / or vector quantization using a first codebook, which is a pre-configured or protocol-defined codebook, or the first codebook is transmitted together with the first model or the second model. The method according to claim 36, characterized in that, The first codebook includes a scalar quantization codebook and / or a vector quantization codebook, and the quantization granularity of the scalar quantization codebook is adjusted in response to a state change of the first channel; And / or, in response to a change in the state of the first channel, the vector quantization codebook is adjusted or updated. The method according to claim 36 or 37 is characterized in that, When the first codebook is a vector quantization codebook, the first codebook includes multiple sub-codebooks, each of the multiple sub-codebooks corresponds to a segment of the output of the first model, and the size of each sub-codebook is positively correlated with the dimension of the segment. The method according to any one of claims 21-38 is characterized in that, The first model is an encoder based on artificial intelligence (AI) / machine learning (ML), and the second model is a decoder based on AI / ML. The method according to claim 39, characterized in that, The encoder is one of a plurality of encoders used in a plurality of terminal devices. The plurality of encoders are trained based on a reference encoder and / or a reference decoder, which are trained based on a common dataset. A device for wireless communication, characterized in that, The device is a first device, the device comprising: The processing unit is used to perform CSI compression on the first channel state information (CSI) using a first model to obtain a second CSI. The transmitting unit is used to transmit the second CSI to the second device; Wherein, the first CSI is related to the first channel, and the second model corresponding to the second device is used to determine the third CSI based on the second CSI. The third CSI is determined based on the first adjustment amount of the first model and / or the second adjustment amount of the second model. The apparatus according to claim 41 is characterized in that, The first CSI includes one or more predicted CSIs, and / or the first CSI includes one or more historical CSIs of the first channel. The apparatus according to claim 41 or 42 is characterized in that, The first adjustment amount is determined based on the output of the first model and / or the output of the second model, and the second adjustment amount is determined based on the output of the second model. The apparatus according to any one of claims 41-43 is characterized in that, The first adjustment amount and / or the second adjustment amount are related to one or more of the following: The intermediate evaluation parameters corresponding to the CSI compression; Auxiliary information of the second device; The auxiliary information is used to compensate for the feature vector differences between the third CSI and the first CSI. The apparatus according to claim 44 is characterized in that, The intermediate evaluation parameter is the squared generalized cosine similarity (SGCS), which is used to adjust the second CSI. The adjusted value of the second CSI is: CSI′=α(SGCS)*CSI+(1-α(SGCS))*ΔCSI; Wherein, CSI is the initial value before the second CSI adjustment, α(SGCS) = SGCSγ, γ is the adjustment parameter, and ΔCSI is the measured result of the first channel or determined according to the instructions of the network device. The apparatus according to claim 44 is characterized in that, The auxiliary information is used to determine one or more of the following: The historical CSI offset value corresponding to the first channel; The weighting factor of the second CSI; The compensation reference value corresponding to the first channel. The apparatus according to claim 46 is characterized in that, The historical CSI offset value is determined based on the historical value of the first CSI and the historical value of the third CSI, and the historical CSI offset value is used for multiple terminal devices including the first device. The apparatus according to claim 46 is characterized in that, The weighting factor is determined based on the correlation between the third CSI and the first CSI, or the weighting factor is determined based on the channel quality of the first channel at different times. The apparatus according to claim 48 is characterized in that, The weighting factor of the second CSI at time t for: Where α is the smoothing coefficient. The weighting factor at time t-1 SNR(t) and SNR(t-1) are the signal-to-noise ratios of the first channel estimated by the second device at time t and time t-1. The apparatus according to any one of claims 44-49 is characterized in that, The device further includes: The receiving unit is used to receive indication information from the second device; The indication information is used to indicate the auxiliary information and / or the first adjustment amount. The apparatus according to any one of claims 41-50 is characterized in that, In response to the fulfillment of the first condition, the first adjustment amount is 0, and / or the second adjustment amount is 0. The apparatus according to any one of claims 41-51 is characterized in that, The second CSI is sent via a CSI report, the sending of which is related to at least one of the following: Whether to send is determined based on an event-triggered method, wherein the event includes the prediction result of the first channel; The first priority of the CSI report is related to the first model. The apparatus according to claim 52 is characterized in that, The CSI report is generated based on a multi-level mapping method, which includes a first-level mapping and a second-level mapping. The first-level mapping corresponds to the channel characteristics of the first channel, and the second-level mapping corresponds to the subcarriers or antennas associated with the first channel. The apparatus according to claim 52 or 53 is characterized in that, The CSI report includes differential information between the current CSI and the previous CSI, or the CSI report includes partial information extracted by the first model based on the first CSI. The apparatus according to any one of claims 41-54 is characterized in that, The first data related to the first model and / or the second model includes first data information, which is used by the network device to classify and store the collected data. The first data information includes one or more of the following: The first data corresponds to the identifier of the cell; The network device scenario identifier corresponding to the first data; The environment type corresponding to the first data; The channel corresponding to the first data is identified as either line-of-sight or non-line-of-sight; The first data corresponds to the identifier of the terminal device. The apparatus according to any one of claims 41-55 is characterized in that, The CSI compression is based on scalar quantization and / or vector quantization using a first codebook, which is a pre-configured or protocol-defined codebook, or the first codebook is transmitted together with the first model or the second model. The apparatus according to claim 56 is characterized in that, The first codebook includes a scalar quantization codebook and / or a vector quantization codebook, and the processing unit is further configured to: In response to a change in the state of the first channel, adjust the quantization granularity of the scalar quantization codebook; and / or, In response to a change in the state of the first channel, the vector quantization codebook is adjusted or updated. The apparatus according to claim 56 or 57 is characterized in that, When the first codebook is a vector quantization codebook, the first codebook includes multiple sub-codebooks, each of the multiple sub-codebooks corresponds to a segment of the output of the first model, and the size of each sub-codebook is positively correlated with the dimension of the segment. The apparatus according to any one of claims 41-58 is characterized in that, The first model is an encoder based on artificial intelligence (AI) / machine learning (ML), and the second model is a decoder based on AI / ML. The apparatus according to claim 59 is characterized in that, The encoder is one of a plurality of encoders used in a plurality of terminal devices. The plurality of encoders are trained based on a reference encoder and / or a reference decoder, which are trained based on a common dataset. A device for wireless communication, characterized in that, The device is a second device, and the device includes: The receiving unit is used to receive second channel state information (CSI) from the first device; A processing unit is used to process the second CSI using a second model to obtain a third CSI; The second CSI is determined by CSI compression of the first CSI using a first model, the first model corresponding to the first device, the first CSI being related to the first channel, and the third CSI being determined based on a first adjustment amount of the first model and / or a second adjustment amount of the second model. The apparatus according to claim 61, characterized in that, The first CSI includes one or more predicted CSIs, and / or the first CSI includes one or more historical CSIs of the first channel. The apparatus according to claim 61 or 62 is characterized in that, The first adjustment amount is determined based on the output of the first model and / or the output of the second model, and the second adjustment amount is determined based on the output of the second model. The apparatus according to any one of claims 61-63 is characterized in that, The first adjustment amount and / or the second adjustment amount are related to one or more of the following: The intermediate evaluation parameters corresponding to the CSI compression; Auxiliary information of the second device; The auxiliary information is used to compensate for the feature vector differences between the third CSI and the first CSI. The apparatus according to claim 64, characterized in that, The intermediate evaluation parameter is the squared generalized cosine similarity (SGCS), which is used to adjust the second CSI. The adjusted value of the second CSI is: CSI′=α(SGCS)*CSI+(1-α(SGCS))*ΔCSI; Wherein, CSI is the initial value before the second CSI adjustment, α(SGCS) = SGCSγ, γ is the adjustment parameter, and ΔCSI is the measured result of the first channel or determined according to the instructions of the network device. The apparatus according to claim 64, characterized in that, The auxiliary information is used to determine one or more of the following: The historical CSI offset value corresponding to the first channel; The weighting factor of the second CSI; The compensation reference value corresponding to the first channel. The apparatus according to claim 66 is characterized in that, The historical CSI offset value is determined based on the historical value of the first CSI and the historical value of the third CSI, and the historical CSI offset value is used for multiple terminal devices including the first device. The apparatus according to claim 66 is characterized in that, The weighting factor is determined based on the correlation between the third CSI and the first CSI, or the weighting factor is determined based on the channel quality of the first channel at different times. The apparatus according to claim 68, characterized in that, The weighting factor of the second CSI at time t for: Where α is the smoothing coefficient. The weighting factor at time t-1 SNR(t) and SNR(t-1) are the signal-to-noise ratios of the first channel estimated by the second device at time t and time t-1. The apparatus according to any one of claims 66-69, characterized in that, The device further includes: The sending unit is used for the first device to send indication information; The indication information is used to indicate the auxiliary information and / or the first adjustment amount. The apparatus according to any one of claims 61-70 is characterized in that, In response to the fulfillment of the first condition, the first adjustment amount is 0, and / or the second adjustment amount is 0. The apparatus according to any one of claims 61-71 is characterized in that, The second CSI is sent via a CSI report, the sending of which is related to at least one of the following: Whether to send is determined based on an event-triggered method, wherein the event includes the prediction result of the first channel; The first priority of the CSI report is related to the first model. The apparatus according to claim 72 is characterized in that, The CSI report is generated based on a multi-level mapping method, which includes a first-level mapping and a second-level mapping. The first-level mapping corresponds to the channel characteristics of the first channel, and the second-level mapping corresponds to the subcarriers or antennas associated with the first channel. The apparatus according to claim 72 or 73 is characterized in that, The CSI report includes differential information between the current CSI and the previous CSI, or the CSI report includes partial information extracted by the first model based on the first CSI. The apparatus according to any one of claims 61-74 is characterized in that, The first data related to the first model and / or the second model includes first data information, which is used by the network device to classify and store the collected data. The first data information includes one or more of the following: The first data corresponds to the identifier of the cell; The network device scenario identifier corresponding to the first data; The environment type corresponding to the first data; The channel corresponding to the first data is identified as either line-of-sight or non-line-of-sight; The first data corresponds to the identifier of the terminal device. The apparatus according to any one of claims 61-75 is characterized in that, The CSI compression is based on scalar quantization and / or vector quantization using a first codebook, which is a pre-configured or protocol-defined codebook, or the first codebook is transmitted together with the first model or the second model. The apparatus according to claim 76 is characterized in that, The first codebook includes a scalar quantization codebook and / or a vector quantization codebook, and the quantization granularity of the scalar quantization codebook is adjusted in response to a state change of the first channel; And / or, in response to a change in the state of the first channel, the vector quantization codebook is adjusted or updated. The apparatus according to claim 76 or 77 is characterized in that, When the first codebook is a vector quantization codebook, the first codebook includes multiple sub-codebooks, each of the multiple sub-codebooks corresponds to a segment of the output of the first model, and the size of each sub-codebook is positively correlated with the dimension of the segment. The apparatus according to any one of claims 61-78 is characterized in that, The first model is an encoder based on artificial intelligence (AI) / machine learning (ML), and the second model is a decoder based on AI / ML. The apparatus according to claim 79 is characterized in that, The encoder is one of a plurality of encoders used in a plurality of terminal devices. The plurality of encoders are trained based on a reference encoder and / or a reference decoder, which are trained based on a common dataset. A communication device, characterized in that, It includes a memory and a processor, the memory being used to store a program, and the processor being used to invoke the program in the memory to perform the method as described in any one of claims 1-40. An apparatus characterized in that, Includes a processor for calling a program from memory to perform the method as described in any one of claims 1-40. A chip characterized in that, Includes a processor for calling a program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1-40. A computer-readable storage medium, characterized in that, It contains a program that causes a computer to perform the method as described in any one of claims 1-40. A computer program product, characterized in that, Includes a program that causes a computer to perform the method as described in any one of claims 1-40. A computer program, characterized in that, The computer program causes the computer to perform the method as described in any one of claims 1-40.

Images