Communication method and apparatus

Abstract

The present application relates to a communication method and an apparatus. A terminal device determines channel state information based on S indexes of S vectors. Each of the S vectors is included in a first vector quantization dictionary. The first vector quantization dictionary includes N1 vectors, where N1 and S are both positive integers. The terminal device transmits the channel state information to a network device. Since the vectors included in the vector quantization dictionary usually have a relatively large dimension, quantizing the channel information using the first vector quantization dictionary is equivalent to performing a dimensional expansion on the channel information or maintaining a relatively high dimension as much as possible, so that high-precision feedback can be implemented by using a relatively low signaling overhead.

Description

Technical Field

[0002] This application relates to the field of communication technologies, and in particular, to communication methods and apparatuses.

Background Art

[0003] The 5th generation (5G) mobile communication system has high requirements for system capacity, spectral efficiency, etc. In a 5G communication system, the application of massive multiple-input multiple-output (massive-MIMO) technology plays an important role in improving the spectral efficiency of the system. By using massive-MIMO technology, a base station can provide high-quality services to more user equipments (UEs) simultaneously. An important step is that the base station precodes the downlink data of multiple UEs. Through precoding, spatial multiplexing can be implemented, interference between UEs can be reduced, the signal to interference plus noise ratio (SINR) at the receiving end can be increased, thereby improving the system throughput rate. In order to more accurately precode the downlink data of a UE, the base station can obtain the channel state information (CSI) of the downlink channel and perform precoding based on the CSI. A method for enabling the base station to effectively obtain the CSI of the downlink channel is a technical problem worthy of study.

Summary of the Invention

Means for Solving the Problems

[0004] Embodiments of this application provide a communication method and apparatus for improving CSI feedback efficiency.

[0005] According to a first aspect, a communication method is provided. This method can be executed on the terminal device side. This method can be executed using software, hardware, or a combination of software and hardware. For example, this method can be executed by a terminal device, or by a circuit system, or by a larger device including the terminal device. The circuit system can implement the functions of the terminal device. This method includes a step of determining channel state information, where the channel state information is determined based on S indices of S vectors, each of the S vectors is included in a first vector quantization dictionary, the first vector quantization dictionary includes N1 vectors, and both N1 and S are positive integers, and a step of transmitting the channel state information to a network device.

[0006] In this method, on the terminal device side, the channel state information can be determined based on the first vector quantization dictionary. For example, the channel information obtained by the terminal device through measurement can be represented by using the indices of one or more vectors included in the first vector quantization dictionary to obtain the channel state information. Since the dimension of the vector is usually relatively large, quantizing the channel information using the first vector quantization dictionary is equivalent to performing dimensional expansion or maintenance in a relatively high dimension on the channel information as much as possible, thereby improving the quantization accuracy. Therefore, in this method, high-precision feedback can be implemented using a relatively low signaling overhead.

[0007] In an optional implementation, each of the N1 vectors represents one channel feature. Since each of the N1 vectors can represent one channel feature, the corresponding channel information can be quantized based on the first vector quantization dictionary to obtain the channel state information, thereby reducing the feedback overhead and relatively guaranteeing the quantization accuracy.

[0008] In an optional implementation, the S vectors are determined based on S pieces of first channel information and a first vector quantization dictionary, and the S pieces of first channel information are the information output by the encoder network. It will be understood that the S pieces of first channel information output by the encoder network indicate the state of the downlink channel, and the S pieces of first channel information can be quantized based on the first vector quantization dictionary. Through such quantization, the feedback overhead can be reduced. In addition, in the VQ-AE network, the quantization network and the compression network are jointly designed (for example, the encoder network and the vector quantization dictionary can be obtained through joint training) so that joint optimization can be implemented between compression and quantization, thereby reducing the performance loss and improving the overall accuracy of the feedback channel state information.

[0009] Optionally, the encoder network and the decoder network cooperate so that the original channel matrix can be compressed. Specifically, instead of extracting some coefficients from the original channel matrix for feedback, the network can be used to compress the original channel matrix as a whole. In this case, the information to be discarded is reduced, thereby reducing the compression loss. In addition, in the VQ-AE network, the quantization network and the compression network are jointly designed so that joint optimization can be implemented between compression and quantization, thereby reducing the performance loss and improving the overall accuracy of the feedback channel state information.

[0010] In an optional implementation, the N1 vectors have the same length.

[0011] In an optional implementation form, the length is equal to at least one dimension in the output dimension of the encoder network. Since the lengths of the N1 vectors are equal to at least one dimension of the output dimension of the encoder network, the N1 vectors can be easily used to quantize the channel information output by the encoder network.

[0012] In an optional implementation, the input dimension of the encoder network is determined based on the input dimension of the first reference encoder network, and the output dimension of the encoder network is determined based on the output dimension of the first reference encoder network. For example, in this embodiment of the present application, one or more reference network groups may be defined by a protocol, or information about one or more reference network groups may be received from a network device. Each reference network group may include a reference encoder network and a corresponding reference decoder network. The reference encoder network and the corresponding reference decoder network may be obtained by joint training. Two communication parties may separately perform offline training on the encoder network and the decoder network that can be used in actual applications based on the reference network. For example, the encoder network currently used by a terminal device is obtained by training based on the first reference encoder network included in the first reference network (or based on the first reference encoder network and the first reference decoder network included in the first reference network). Therefore, the input dimension of the encoder network may be determined based on the input dimension of the first reference encoder network, and the output dimension of the encoder network may be determined based on the output dimension of the first reference encoder network. For example, the input dimension of the encoder network is equal to the input dimension of the first reference encoder network, and the output dimension of the encoder network is equal to the output dimension of the first reference encoder network. In this way, in order to improve the adaptability of the encoder network to the actual network environment, the encoder network used by the terminal device can meet the performance requirements, and the encoder network used by the terminal device can also meet the current network conditions.

[0013] In an optional implementation, the first reference encoder network corresponds to the first reference decoder network. For example, if the first reference encoder network belongs to the first reference network, and the first reference network includes the first reference encoder network and the first reference decoder network, the first reference encoder network corresponds to the first reference decoder network. For example, the first reference encoder network and the first reference decoder network may be obtained by joint training.

[0014] In an optional implementation, the first reference network includes the first reference encoder network and the first reference decoder network. The reference network may include a reference encoder network and a corresponding reference decoder network. For example, the reference encoder network and the reference decoder network included in the reference network may be obtained by joint training.

[0015] In an optional implementation, the first reference network further includes a vector quantization dictionary (which may also be referred to as a reference vector quantization dictionary). In this embodiment of the present application, the reference encoder network, the reference decoder network, and the vector quantization dictionary can be jointly trained and optimized. Therefore, in addition to the first reference encoder network and the first reference decoder network, the first reference network may further include a reference vector quantization dictionary. In this way, the joint optimization between compression and quantization can be better implemented. For example, the first vector quantization dictionary is a reference vector quantization dictionary, or the first vector quantization dictionary may be different from the reference vector quantization dictionary. For example, in the initial network access communication stage, the first vector quantization dictionary used by the UE is a reference vector quantization dictionary. During the communication process between the UE and the network device, the vector quantization dictionary used by the UE may be updated, and the first vector quantization dictionary is, for example, an updated vector quantization dictionary.

[0016] In an optional implementation form, the input dimension of the encoder network is M or M×2, the output dimension of the encoder network is D×S, both M and D are positive integers, and M≤N tx ×N sb ×z, S≤M, z is a positive integer greater than or equal to 1, and N tx represents the number of transmission antenna ports of the network device, and N sb represents the number of frequency domain sub-bands. N tx and N sb are both positive integers.

[0017] After obtaining the downlink channel matrix through measurement, the terminal device may preprocess the downlink channel matrix and then input the preprocessed downlink channel matrix into the encoder network. For example, in the process of preprocessing the downlink channel matrix, the terminal device may obtain an eigen subspace matrix based on the downlink channel matrix, and then perform a spatial frequency joint projection on the eigen subspace matrix to obtain a complex matrix. After obtaining the complex matrix, the terminal device may further perform a dimensionality reduction process on the complex matrix and input the processed complex matrix into the encoder network, or convert the processed complex matrix into a real matrix and then input the real matrix into the encoder network. In this case, the input dimension of the encoder network may be M or M×2. For example, the dimension of the downlink channel matrix is N tx ×N rx ×N RB and M≤N tx ×N sbIt is z. It can be understood that information compression is completed in this way. Specifically, in this step, the main feature elements are selected to approximately represent the original information. In this processing method, the matrix dimension can be further reduced to reduce the complexity of subsequent processing (such as the processing of the encoder network). The matrix dimension obtained by this preprocessing method is used as the input dimension of the subsequent neural network (for example, the input dimension of the encoder network), and further affects the dimension for reconstructing the downlink channel matrix on the network device side.

[0018] In an optional implementation form, the input dimension of the encoder network is M×N×2 or M×N, the output dimension of the encoder network is D×P×Q, S = P×Q, and M, N, D, P, and Q are all positive integers, and M≤N tx , N≤N sb , and S≤M×N, or M = N tx ×x, N = N sb ×y, and S≤(N tx ×x)×(N sb ×y), or M = N tx ×N rx , N = N sb , and S≤(N tx ×N rx )×N sb is. In this specification, N tx represents the number of transmission antenna ports of the network device, N sb represents the number of frequency domain sub-bands, N rx represents the number of reception antenna ports of the terminal device, and both x and y are greater than 1, or x is equal to 1 and y is greater than 1, or x is greater than 1 and y is equal to 1. N rx、 N tx , and N sbThey are all positive integers. For example, in the process of preprocessing the downlink channel matrix, the terminal device may obtain an eigen-subspace matrix based on the downlink channel matrix, and then perform a spatial-frequency joint projection on the eigen-subspace matrix to obtain a complex matrix. After obtaining the complex matrix, the terminal device may further perform a dimensionality reduction process on the complex matrix and input the processed complex matrix into the encoder network, or convert the processed complex matrix into a real matrix and then input the real matrix into the encoder network. In this case, the input dimension of the encoder network may be M×N×2 or M×N. In this case, for the technical effect, please refer to the description of the above optional implementation form. Alternatively, after obtaining the complex matrix, the terminal device does not need to further perform a dimensionality reduction process, but may input the complex matrix into the encoder network, or convert the complex matrix into a real matrix and then input the real matrix into the encoder network. In this case, the input dimension of the encoder network may also be M×N×2 or M×N. In this case, in order to simplify the process of determining the channel state information by the terminal device, the processing process of the terminal device is reduced. Alternatively, in the process of preprocessing the downlink channel matrix, the terminal device may first perform a dimensionality reduction process on the downlink channel matrix, and then perform a spatial-frequency joint projection on the matrix obtained by the dimensionality reduction to obtain a complex matrix. For the processing method after obtaining the complex matrix, please refer to the above description. In this case, the input dimension of the encoder network may also be M×N×2 or M×N. In this case, the terminal device reduces the process of performing the SVD operation on the downlink channel matrix to simplify the implementation.

[0019] In an optional implementation form, the input dimension of the encoder network is M×N×T×2 or M×N×T, the output dimension of the encoder network is D×P×Q×R, S = P×Q×R, and M, N, D, P, Q, R, and T are all positive integers, and M = N tx , N = N sb , T = Nrx 、S ≤ N rx ×N tx ×N sb where N tx represents the number of transmit antenna ports of the network device, N sb represents the number of frequency domain sub-bands, N rx represents the number of receive antenna ports of the terminal device. N rx、 N tx and N sb are all positive integers. In the process of preprocessing the downlink channel matrix, the terminal device may first transform the downlink channel matrix in order to obtain a two-dimensional matrix included in the matrix obtained by the transformation. Then, a spatial frequency joint projection may be performed on the two-dimensional matrix to obtain a complex matrix. For the processing method after the complex matrix is obtained, please refer to the description of the foregoing implementation form. In this case, the input dimension of the encoder network may be M×N×T×2 or M×N×T. In this case, the terminal device reduces the process of performing the SVD operation on the downlink channel matrix in order to simplify the implementation.

[0020] In an optional implementation form, the method further includes the step of transmitting information about the first vector quantization dictionary to the network device. For example, the terminal device may transmit the first vector quantization dictionary to the network device before transmitting the channel state information to the network device or (for example, in the same time unit) when transmitting the channel state information, so that the network device can obtain the downlink channel matrix by reconstruction based on the first vector quantization dictionary and the channel state information from the terminal device. In another example, the vector quantization dictionary may be updated online. For example, the terminal device may update the vector quantization dictionary so that it can better match the channel environment in which the used vector quantization dictionary is constantly changing in order to feedback the CSI more accurately.

[0021] In an optional implementation form, the step of transmitting information about the first vector quantization dictionary to the network device includes: periodically transmitting information about the first vector quantization dictionary to the network device; after the first configuration information is received from the network device, periodically transmitting information about the first vector quantization dictionary to the network device; after the first configuration information is received from the network device, transmitting information about the first vector quantization dictionary to the network device; after the first configuration information is received from the network device, transmitting information about the first vector quantization dictionary to the network device in units of the arrival time of the first delay; after the first parameter information is received from the network device, transmitting information about the first vector quantization dictionary to the network device, where the first parameter information indicates the parameter information of the decoder network; after the first parameter information is received from the network device, transmitting information about the first vector quantization dictionary to the network device in units of the arrival time of the first delay, where the first parameter information indicates the parameter information of the decoder network; after the first instruction information is received from the network device, transmitting information about the first vector quantization dictionary to the network device, where the first instruction information is used to indicate (or trigger) the transmission of the vector quantization dictionary; or after the first instruction information is received from the network device, transmitting information about the first vector quantization dictionary to the network device in units of the arrival time of the first delay, where the first instruction information is used to indicate (or trigger) the transmission of the vector quantization dictionary. For example, the reporting parameters of the vector quantization dictionary may be configured in the first configuration information. When the terminal device transmits information about the first vector quantization dictionary to the network device, there may be multiple transmission methods. Some methods are listed in this specification.The specific manner in which the terminal device transmits the first vector quantization dictionary may be negotiated by the terminal device and the network device, may be configured by the network device, or may be specified by a protocol or the like. This is relatively flexible.

[0022] Optionally, the first configuration information includes information regarding a first delay, or the information regarding the first delay is indicated using other signaling from the network device, or the first delay is agreed upon by a protocol.

[0023] In an optional implementation form, the method includes determining that the current networking method is a first networking method and determining a first vector quantization dictionary corresponding to the first networking method; determining that the current moving speed belongs to a first interval and determining a first vector quantization dictionary corresponding to the first interval; determining a handover from a second cell to a first cell and determining a first vector quantization dictionary corresponding to the first cell; determining a switch from a second reference encoder network to a first reference encoder network and determining a first vector quantization dictionary corresponding to the first reference encoder network; or receiving second indication information from the network device, where the second indication information indicates the first vector quantization dictionary.

[0024] In an optional implementation, the current networking method is the first networking method, the first vector quantization dictionary is the vector quantization dictionary corresponding to the first networking method, the moving speed of the terminal device belongs to the first interval, the first vector quantization dictionary is the vector quantization dictionary corresponding to the first interval, the terminal device is handed over from the second cell to the first cell, the first vector quantization dictionary is the vector quantization dictionary corresponding to the first cell, the reference encoder network used is switched from the second reference encoder network to the first reference encoder network, the first vector quantization dictionary is the vector quantization dictionary corresponding to the first reference encoder network, or the first vector quantization dictionary is the vector quantization dictionary indicated by the second indication information, and the second indication information is from the network device.

[0025] To use the first vector quantization dictionary, the terminal device needs to first determine the first vector quantization dictionary. The method for the terminal device to determine the first vector quantization dictionary includes, for example, an explicit method and / or an implicit method. When the explicit method is used, for example, the network device may send signaling to indicate the first vector quantization dictionary. When the implicit method is used, some of the foregoing methods may be included. In other words, a correspondence relationship may be established between the vector quantization dictionary and the corresponding information (for example, the networking method and moving speed of the terminal device), so that the terminal device can determine the first vector quantization dictionary based on the information. In order to reduce signaling overhead, the signaling indication of the network device may not be required in an implicit manner. Alternatively, the terminal device may determine the first vector quantization dictionary in a manner that combines the explicit method and the implicit method. This is not limited in this embodiment of the present application.

[0026] In an optional implementation, the method further includes a step of sending second parameter information to the network device, where the second parameter information indicates the parameters of the encoder network. In other words, the terminal device may send information regarding the parameters of the encoder network to the network device. For example, the vector quantization dictionary may alternatively be updated by the network device. It is more appropriate for the network device to have the decision-making power. In this case, optionally, the terminal device may send the parameter information of the encoder network to the network device, as a result of which the network device can update the vector quantization dictionary with reference to the parameter information of the encoder network.

[0027] In an optional implementation form, the step of transmitting the second parameter information to the network device includes the step of periodically transmitting the second parameter information to the network device; the step of periodically transmitting the second parameter information to the network device after the second configuration information is received from the network device; the step of transmitting the second parameter information to the network device after the second configuration information is received from the network device; the step of transmitting the second parameter information to the network device in units of the arrival time of the second delay after the second configuration information is received from the network device; the step of transmitting the second parameter information to the network device after the third instruction information is received from the network device, where the third instruction information is used to indicate (e.g., trigger) the transmission of the parameter information of the encoder network; or the step of transmitting the second parameter information to the network device in units of the arrival time of the second delay after the third instruction information is received from the network device, where the third instruction information indicates the transmission of the parameter information of the encoder network. For example, the reporting parameter of the second parameter information may be configured in the second configuration information. When the terminal device transmits the second parameter information to the network device, there may be multiple transmission methods. Some methods are listed in this specification. The specific method for the terminal device to transmit the second parameter information may be negotiated by the terminal device and the network device, may be configured by the network device, or may be specified by a protocol or the like. This is relatively flexible.

[0028] Optionally, the second configuration information includes information about the second delay, or the information about the second delay is indicated using other signaling from the network device, or the second delay is agreed upon by the protocol.

[0029] According to a second aspect, another communication method is provided. This method can be executed on the network device side. This method can be executed using software, hardware, or a combination of software and hardware. For example, the method can be executed by a network device, or by a larger device including the network device, or by a circuit system, or by an AI module for assisting the network device or a network element for assisting the network device. In this specification, the circuit system can implement the functions of the network device, and the AI module and the network element are independent of the network device. This is not limited. For example, the network device is an access network device such as a base station. This method includes the step of receiving channel state information, where the channel state information includes S indices of S vectors, and the step of obtaining information about a downlink channel matrix reconstructed based on the S indices and a first vector quantization dictionary, where each of the S vectors is included in the first vector quantization dictionary, the first vector quantization dictionary includes N1 vectors, and both N1 and S are positive integers.

[0030] In an optional implementation, each of the N1 vectors represents one channel feature.

[0031] In an optional implementation, the step of obtaining information about a downlink channel matrix reconstructed based on the S indices and the first vector quantization dictionary includes the step of obtaining information about a downlink channel matrix reconstructed based on the first vector quantization dictionary, the S indices, and a decoder network.

[0032] In an optional implementation form, the step of obtaining information about the reconstructed downlink channel matrix based on S indices and the first vector quantization dictionary (or, based on the first vector quantization dictionary, S indices, and the decoder network, the step of obtaining information about the reconstructed downlink channel matrix) includes the step of performing an inverse mapping on the S indices based on the first vector quantization dictionary to obtain a first matrix, and the step of obtaining information about the reconstructed downlink channel matrix based on the first matrix and the decoder network.

[0033] In an optional implementation form, the step of obtaining information about the reconstructed downlink channel matrix based on the first matrix and the decoder network includes the step of inputting the first matrix into the decoder network to obtain a second matrix, and the step of obtaining information about the reconstructed downlink channel matrix based on the second matrix.

[0034] In an optional implementation form, the reconstructed downlink channel matrix satisfies the following relationship:

Number

Number

[0035] In an optional implementation form, M = N tx , and N = N sb is, or M ≤ N tx , and N ≤ N sb is, or M = N tx ×x, and N = N sb ×y, where x and y are both greater than 1, or x is equal to 1 and y is greater than 1, or x is greater than 1 and y is equal to 1.

[0036] In an optional implementation form, the reconstructed downlink channel matrix satisfies the following relationship:

Number

Number

[0037] In an optional implementation, the reconstructed downlink channel matrix satisfies the following relationship:

Number

Number

[0038] In an optional implementation form, the reconstructed downlink channel matrix satisfies the following relationship:

Number

Number

[0039] In an optional implementation, the reconstructed downlink channel matrix satisfies the following relationship:

Number

Number

[0040] In an optional implementation form, M = N tx , and N = N sb or M ≤ N tx , and N ≤ N sb or M = N tx ×x, and N = N sb ×y, where both x and y are greater than 1.

[0041] In an optional implementation form, the reconstructed downlink channel matrix satisfies the following relationship:

Number

Number

[0042] In an optional implementation form, the reconstructed downlink channel matrix satisfies the following relationship:

Number

Number

[0043] In an optional implementation form, the reconstructed downlink channel matrix satisfies the following relationship:

Number

Number

[0044] In an optional implementation form, the method further includes the step of receiving information regarding the first vector quantization dictionary.

[0045] In an optional implementation form, the step of receiving information about the first vector quantization dictionary includes: periodically receiving information about the first vector quantization dictionary; after the first configuration information is sent, periodically receiving information about the first vector quantization dictionary; after the first configuration information is sent, receiving information about the first vector quantization dictionary; after the first configuration information is sent, receiving information about the first vector quantization dictionary in units of the arrival time of the first delay; after the first parameter information is sent, receiving information about the first vector quantization dictionary, where the first parameter information indicates the parameter information of the decoder network; after the first parameter information is sent, receiving information about the first vector quantization dictionary in units of the arrival time of the first delay, where the first parameter information indicates the parameter information of the decoder network; after the first instruction information is sent, receiving information about the first vector quantization dictionary, where the first instruction information is used to indicate (e.g., trigger) the transmission of the vector quantization dictionary; or after the first instruction information is sent, receiving information about the first vector quantization dictionary in units of the arrival time of the first delay, where the first instruction information is used to indicate (e.g., trigger) the transmission of the vector quantization dictionary.

[0046] Optionally, the first configuration information includes information about the first delay, or the information about the first delay is indicated using other signaling from the network device, or the first delay is agreed upon by the protocol.

[0047] In an optional implementation, the method includes determining that the current networking method by the terminal device is the first networking method and determining a first vector quantization dictionary corresponding to the first networking method; determining that the current moving speed of the terminal device belongs to a first interval and determining a first vector quantization dictionary corresponding to the first interval; determining that the terminal device is handed over from a second cell to a first cell and determining a first vector quantization dictionary corresponding to the first cell; determining that the terminal device is switched from a second reference encoder network to a first reference encoder network and determining a first vector quantization dictionary corresponding to the first reference encoder network; or transmitting second indication information, where the second indication information indicates information about the first vector quantization dictionary.

[0048] In an optional implementation, the current networking method is the first networking method, the first vector quantization dictionary is a vector quantization dictionary corresponding to the first networking method, the moving speed of the terminal device belongs to the first interval, the first vector quantization dictionary is a vector quantization dictionary corresponding to the first interval, the terminal device is handed over from a second cell to a first cell, the first vector quantization dictionary is a vector quantization dictionary corresponding to the first cell, the reference encoder network used by the terminal device is switched from a second reference encoder network to a first reference encoder network, the first vector quantization dictionary is a vector quantization dictionary corresponding to the first reference encoder network, or the method includes transmitting second indication information, where the second indication information indicates information about the first vector quantization dictionary.

[0049] In an optional implementation, the method further includes receiving second parameter information, where the second parameter information indicates parameters of an encoder network.

[0050] In an optional implementation, the step of receiving the second parameter information includes the step of periodically receiving the second parameter information, the step of periodically receiving the second parameter information after the second configuration information is sent, the step of receiving the second parameter information after the second configuration information is sent, the step of receiving the second parameter information in units of the arrival time of the second delay after the second configuration information is sent, the step of receiving the second parameter information after the third instruction information is sent, where the third instruction information indicates the transmission of the parameter information of the encoder network, or the step of receiving the second parameter information in units of the arrival time of the second delay after the third instruction information is sent, where the third instruction information indicates the transmission of the parameter information of the encoder network.

[0051] Optionally, the second configuration information includes information about the second delay, or the information about the second delay is indicated using other signaling from the network device, or the second delay is agreed upon by the protocol.

[0052] For the technical effects brought about by the second aspect or various optional implementations, please refer to the description of the technical effects brought about by the first aspect or the corresponding implementations.

[0053] According to a third aspect, a communication device is provided. The communication device may implement the method of the first aspect. The communication device has the functions of the aforementioned terminal device. In an optional implementation form, the device may include modules that correspond one-to-one with the methods / operations / steps / actions described in the first aspect. The modules may be implemented by a hardware circuit, software, or a combination of a hardware circuit and software. In a possible implementation form, the communication device includes a baseband device and a radio device. In another optional implementation form, the communication device includes a processing unit (which may also be called a processing module) and a transceiver unit (which may also be called a transceiver module). The transceiver unit can implement a transmission function and a reception function. When the transceiver unit implements the transmission function, the transceiver unit may be called a transmission unit (or may be called a transmission module). When the transceiver unit implements the reception function, the transceiver unit may be called a reception unit (or may be called a reception module). The transmission unit and the reception unit may be the same functional module. The functional module is called a transceiver unit. The functional module can implement a transmission function and a reception function. Alternatively, the transmission unit and the reception unit may be different functional modules. The transceiver unit is a general term for these functional modules.

[0054] The processing unit is configured to determine channel state information, the channel state information is determined based on S indices of S vectors, each of the S vectors is included in a first vector quantization dictionary, the first vector quantization dictionary includes N1 vectors, and both N1 and S are positive integers. The transceiver unit (or the transmission unit) is configured to transmit the channel state information to a network device.

[0055] In another example, the communication device includes a processor and a communication interface. The processor uses the communication interface to implement the method described in the first aspect. The communication interface is used by the device to communicate with another device. For example, the communication interface may be a transceiver, a circuit, a bus, a pin, a module, or another type of communication interface. For example, the processor is configured to determine channel state information, where the channel state information is determined based on S indices of S vectors, each of the S vectors is included in a first vector quantization dictionary, the first vector quantization dictionary includes N1 vectors, both N1 and S are positive integers, and the processor transmits the channel state information to a network device using the communication interface. The apparatus may further include a memory. The memory is configured to store program instructions and data. The memory is coupled to the processor. When executing the instructions stored in the memory, the processor can implement the method described in the first aspect.

[0056] In another example, the communication device includes a processor coupled to a memory and configured to execute instructions within the memory to implement the method of the first aspect. Optionally, the communication device further includes other components, such as an antenna, an input / output module, and an interface. These components may be hardware, software, or a combination of software and hardware.

[0057] According to a fourth aspect, a communication device is provided. The communication device may implement the method according to the second aspect. The communication device has the functions of the aforementioned network device. The network device is, for example, a base station or a baseband device within a base station. In an optional implementation form, the device may include a module that corresponds one-to-one with the method / operation / step / action described in the second aspect. The module may be implemented by a hardware circuit, software, or a combination of a hardware circuit and software. In a possible implementation form, the communication device includes a baseband device and a radio device. In another optional implementation form, the communication device includes a processing unit (which may also be called a processing module) and a transceiver unit (which may also be called a transceiver module). For the implementation form of the transceiver unit, refer to the relevant description of the fourth aspect.

[0058] The transceiver unit (or receiving unit) is configured to receive channel state information. The channel state information includes S indices of S vectors.

[0059] The processing unit is configured to obtain information regarding the reconstructed downlink channel matrix based on the S indices and the first vector quantization dictionary. Each of the S vectors is included in the first vector quantization dictionary. The first vector quantization dictionary includes N1 vectors. Both N1 and S are positive integers.

[0060] In another example, the communication device includes a processor and a communication interface. The processor uses the communication interface to implement the method described in the second aspect. The communication interface is used for the device to communicate with another device. For example, the communication interface may be a transceiver, a circuit, a bus, a pin, a module, or another type of communication interface. For example, the processor uses the communication interface to receive channel state information. The channel state information includes S indices of S vectors. The processor is configured to obtain information regarding the reconstructed downlink channel matrix based on the S indices and the first vector quantization dictionary. Each of the S vectors is included in the first vector quantization dictionary. The first vector quantization dictionary includes N1 vectors. Both N1 and S are positive integers. The device may further include a memory. The memory is configured to store program instructions and data. The memory is coupled to the processor. When the instructions stored in the memory are executed, the processor can implement the method described in the second aspect.

[0061] In another example, the communication device includes a processor coupled to a memory and configured to execute instructions in the memory to implement the method of the second aspect. Optionally, the communication device further includes other components, such as an antenna, an input / output module, and an interface. These components may be hardware, software, or a combination of software and hardware.

[0062] According to a fifth aspect, a computer-readable storage medium is provided. The computer-readable storage medium is configured to store a computer program or instructions. When the computer program or instructions are executed, the method of the first aspect or the second aspect is implemented.

[0063] According to a sixth aspect, a computer program product including instructions is provided. When the computer program product is executed on a computer, the method according to the first aspect or the second aspect is implemented.

[0064] According to a seventh aspect, a chip system is provided. The chip system includes a processor and may further include a memory. The chip system is configured to implement the method according to the first aspect or the second aspect. The chip system may include a chip or may include a chip and other separate devices.

[0065] According to an eighth aspect, a communication system including the communication device according to the third aspect and the communication device according to the fourth aspect is provided.

Brief Description of the Drawings

[0066]

Figure 1

Figure 2

Figure 3

Figure 4A

Figure 4B

Figure 4C

Figure 4D

Figure 4E

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10A

Figure 10B

Figure 11

Figure 12

Figure 13

Figure 14

Embodiments for Carrying Out the Invention

[0067] In order to make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0068] The technology provided in the embodiments of this application can be applied to the communication system 10 shown in FIG. 1. The communication system 10 includes one or more communication devices 30 (e.g., terminal devices). One or more communication devices 30 are connected to one or more core network (CN) devices via one or more radio access network (RAN) devices 20 to implement communication between multiple communication devices. For example, the communication system 10 is a communication system that supports 4G (including long term evolution (LTE)) access technology, a communication system that supports 5G (also sometimes called new radio (NR)) access technology, a wireless fidelity (Wi-Fi) system, a cellular system related to the 3rd generation partnership project (3GPP), a communication system that supports the convergence of multiple wireless technologies, or a future-oriented evolved system.

[0069] Hereinafter, the terminal device and the RAN in FIG. 1 will be described in detail individually.

[0070] 1. Terminal Device A terminal device, simply called a terminal, is a device with wireless transmission and reception capabilities. The terminal device may be mobile or fixed. The terminal device may be deployed on land, including, for example, indoor devices or outdoor devices, handheld devices, or in-vehicle devices, or may be deployed on water (e.g., on a ship), or in the air (e.g., on an airplane, balloon, or satellite). The terminal device may include a mobile phone, a tablet computer (pad), a computer with wireless transmission and reception capabilities, a virtual reality (VR) terminal device, an augmented reality (AR) terminal device, a wireless terminal device for industrial control, a wireless terminal device for self-driving, a wireless terminal device for remote medical, a wireless terminal device for smart grid, a wireless terminal device for transportation safety, a wireless terminal device for smart city, or a wireless terminal device for smart home. Alternatively, the terminal device may also be a cellular phone, a cordless phone, a session initiation protocol (SIP) phone, a wireless local loop (WLL) station, a personal digital assistant (PDA), a handheld device or a computing device with wireless communication capabilities, an in-vehicle device, a wearable device, a terminal device in a future fifth generation (5G) network, a terminal device in a future evolved public land mobile network (PLMN), etc. The terminal device may sometimes be called a user equipment (UE). Optionally, the terminal device may communicate with multiple access network devices using different technologies.For example, the terminal device may communicate with an access network device that supports LTE, or may communicate with an access network device that supports 5G, or may communicate with an access network device that supports LTE and an access network device that supports 5G simultaneously. This is not limited in the embodiments of the present application.

[0071] In the embodiments of the present application, the device configured to implement the functions of the terminal device may be the terminal device itself, or a device that can support the terminal device when implementing the functions, such as a chip system, a hardware circuit, a software module, or a hardware circuit and a software module. The device may be installed in the terminal device or used in a manner consistent with the terminal device. In the technical solutions provided in the embodiments of the present application, for the purpose of explaining the technical solutions provided in the embodiments of the present application, an example in which the device for implementing the functions of the terminal device is the terminal device and the terminal device is a UE is used.

[0072] In the embodiments of the present application, the chip system may include a chip, or may include a chip and other separate devices.

[0073] 2. RAN The RAN may include one or more RAN devices, such as the RAN device 20. The interface between the RAN device and the terminal device may be a Uu interface (also called an air interface). In future communications, the names of these interfaces may remain unchanged, or may be replaced by other names. This is not limited in the present application.

[0074] A RAN device is a node or device that enables a terminal device to access a wireless network. A RAN device may also be referred to as a network device or a base station. For example, a RAN device includes, but is not limited to, a base station, a next-generation node B (gNB) in 5G, an evolved node B (eNB), a radio network controller (RNC), a node B (NB), a base station controller (BSC), a base transceiver station (BTS), a home base station (e.g., a home evolved node B or a home node B, HNB), a baseband unit (BBU), a transmitting and receiving point (TRP), a transmitting point (TP), and / or a mobile switching center. Alternatively, an access network device may be at least one of a centralized unit (CU), a distributed unit (DU), a CU control plane (CU-CP) node, a CU user plane (CU-UP) node, an integrated access and backhaul (IAB), a radio controller in a cloud radio access network (CRAN) scenario, etc. Alternatively, an access network device may be a relay station, an access point, a vehicle-mounted device, a terminal device, a wearable device, an access network device in a 5G network, an access network device in a future evolved public land mobile network (PLMN), etc.

[0075] In the embodiments of the present application, the device configured to implement the functions of an access network device may be an access network device, or a device capable of supporting an access network device when implementing functions, such as a chip system, a hardware circuit, a software module, or a hardware circuit and a software module. The device may be installed in the access network device or used in a manner consistent with the access network device. In the technical solutions provided in the embodiments of the present application, an example where the device for implementing the functions of an access network device is an access network device and the access network device is a base station is used to illustrate the technical solutions provided in the embodiments of the present application.

[0076] (1) Protocol layer structure The communication between the access network device and the terminal device follows a specific protocol layer structure. The protocol layer structure may include a control plane protocol layer structure and a user plane protocol layer structure. For example, the control plane protocol layer structure may include the functions of protocol layers such as a radio resource control (RRC) layer, a packet data convergence protocol (PDCP) layer, a radio link control (RLC) layer, a media access control (MAC) layer, and a physical layer. For example, the user plane protocol layer structure may include the functions of protocol layers such as the PDCP layer, the RLC layer, the MAC layer, and the physical layer. In one implementation, a service data adaptation protocol (SDAP) layer may further be included on top of the PDCP layer.

[0077] (2) Central unit (CU) and distributed unit (DU) The RAN device may include a CU and a DU. A plurality of DUs may be centrally controlled by one CU. For example, the interface between the CU and the DU may also be called the F1 interface. The control plane (CP) interface may be F1-C, and the user plane (UP) interface may be F1-U. The CU and the DU may be obtained by splitting based on the protocol layer of the wireless network. For example, the functions of the PDCP layer and the protocol layers above the PDCP layer are set in the CU, and the functions of the protocol layers below the PDCP layer (for example, the RLC layer and the MAC layer) are set in the DU. In another example, the functions of the protocol layers above the PDCP layer are set in the CU, and the functions of the PDCP layer and the protocol layers below the PDCP layer are set in the DU.

[0078] It will be understood that the division of the processing functions of the CU and the DU based on the protocol layer is merely an example. The division may alternatively be performed in another manner. For example, the CU or the DU may have the functions of more protocol layers by splitting. As another example, the CU or the DU may alternatively have some processing functions of the protocol layers by splitting. In one design, some functions of the RLC layer and the functions of the protocol layers above the RLC layer are set in the CU, and the remaining functions of the RLC layer and the functions of the protocol layers below the RLC layer are set in the DU. In another design, the division of the functions of the CU or the DU may alternatively be performed based on the service type or other system requirements. For example, the division may be performed based on latency. The functions that need to meet the latency time requirement are set in the DU, and the functions that do not need to meet the latency time requirement are set in the CU. In another design, the CU may alternatively have one or more functions of the core network. For example, for ease of centralized management, the CU may be located on the network side. In another design, the radio unit (RU) of the DU is remotely located. The RU has radio frequency functions.

[0079] Optionally, the DU and RU may be further divided at the physical layer (PHY). For example, the DU may implement the upper layer functions of the PHY layer, and the RU may implement the lower layer functions of the PHY layer. In the transmission procedure, the functions of the PHY layer may include cyclic redundancy check (CRC) code addition function, channel coding, rate matching, scrambling, modulation, layer mapping, precoding, resource mapping, physical antenna mapping, and / or radio frequency transmitter function. In the reception procedure, the functions of the PHY layer may include CRC, channel decoding, rate demapping, descrambling, demodulation, layer demapping, channel sounding, resource demapping, physical antenna demapping, and / or radio frequency receiver function. The upper layer functions of the PHY layer may include some functions of the PHY layer. For example, the functions are closer to the MAC layer. The lower layer functions of the PHY layer may include some other functions of the PHY layer. For example, these functions are closer to the radio frequency functions. For example, the upper layer functions of the PHY layer may include CRC code addition, channel coding, rate matching, scrambling, modulation, and layer mapping, and the lower layer functions of the PHY layer may include precoding, resource mapping, physical antenna mapping, and radio frequency transmitter function. Alternatively, the upper layer functions of the PHY layer may include CRC code addition, channel coding, rate matching, scrambling, modulation, layer mapping, and precoding, and the lower layer functions of the PHY layer may include resource mapping, physical antenna mapping, and radio frequency transmitter function.

[0080] For example, the functions of the CU may be implemented by one entity or different entities. For example, the functions of the CU may be further divided. Specifically, the control plane and the user plane are separated and implemented by using different entities, namely, a control plane CU entity (i.e., a CU-CP entity) and a user plane CU entity (i.e., a CU-UP entity), respectively. The CU-CP entity and the CU-UP entity may be coupled to the DU to jointly complete the functions of the RAN device.

[0081] It should be noted that in the foregoing architecture, the signaling generated by the CU may be transmitted to the terminal device via the DU, or the signaling generated by the terminal device may be transmitted to the CU via the DU. For example, the signaling in the RRC layer or the PDCP layer may be finally processed to obtain the signaling in the physical layer and then transmitted to the terminal device, or may be converted from the signaling received from the physical layer. In this architecture, the signaling in the RRC layer or the PDCP layer may be considered to be transmitted via the DU or via the DU and the RU.

[0082] Optionally, any one of the DU, CU, CU-CP, CU-UP, and RU may be a software module, a hardware structure, or a software module and a hardware structure. This is not limited. Different entities may exist in different forms. This is not limited. For example, the DU, CU, CU-CP, and CU-UP are software modules, and the RU is a hardware structure. These modules and the methods executed by these modules also fall within the protection scope of the embodiments of the present application.

[0083] In a possible implementation form, the RAN device includes a CU-CP, a CU-UP, a DU, and an RU. For example, the execution entity of the method in the embodiments of this application includes the DU, or includes the DU and the RU, or includes the CU-CP, the DU, and the RU. This is not limiting. The method executed by the module also falls within the protection scope of the embodiments of this application.

[0084] Note that since the network device in the embodiments of this application is mainly an access network device, unless otherwise specified, the "network device" described below may be an "access network device".

[0085] It should be understood that the number of devices in the communication system shown in FIG. 1 is only used as an example, and this embodiment of this application is not limited thereto. In actual applications, the communication system may further include more terminal devices and more RAN devices, and may further include other devices, for example, a core network device and / or a node configured to implement an artificial intelligence function.

[0086] The network architecture shown in FIG. 1 can be used in communication systems of various radio access technologies (RATs), such as a 4G communication system, or a 5G (also known as new radio (NR)) communication system, or a transitional system between an LTE communication system and a 5G communication system. The transitional system may also be referred to as a 4.5G communication system, or a future communication system, such as a 6G communication system. The network architecture and service scenarios described in the embodiments of the present application are intended to more clearly describe the technical solutions of the embodiments of the present application, and do not constitute a limitation to the technical solutions provided in the embodiments of the present application. Those skilled in the art can know that with the evolution of communication network architectures and the emergence of new service scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0087] In addition to the communication between the network device and the terminal device, the method provided in the present application can also be used for communication between other communication devices, such as the communication between a macro base station and a micro base station in a wireless backhaul link, or the communication between a first terminal device and a second terminal device in a sidelink (SL). This is not limited. In the embodiments of the present application, the communication between the network device and the terminal device is used as an example for explanation.

[0088] When transmitting data to a terminal device, the network device may perform precoding based on the CSI feedback by the terminal device. To facilitate the understanding of the embodiments of the present application, some technical terms used in the embodiments of the present application will be briefly described below.

[0089] 1. Precoding technology When the channel state is known, the network device can process the signal to be transmitted using a precoding matrix that matches the channel conditions. This technique can be used to collate the precoded signal to be transmitted with the channel in order to improve the quality of the signal received by the terminal device (e.g., SINR) and reduce the complexity of eliminating the influence between channels by the terminal device. By using precoding techniques, the transmitting device (e.g., network device) and multiple receiving devices (e.g., terminal devices) can effectively perform transmissions on the same time-frequency resources, that is, can effectively implement multi-user multiple input multiple output (MU-MIMO). Note that the related descriptions of precoding techniques are merely examples for ease of understanding and are not intended to limit the disclosed scope of the embodiments of the present application. In a specific implementation process, the transmitting device may alternatively perform precoding in another manner. For example, when channel information (e.g., but not limited to, channel matrix) cannot be known, precoding is performed using a preset precoding matrix or by weighting processing. For the sake of brevity, the specific content of precoding is not described in this specification.

[0090] 2. Dual-region compression Dual-region compression may include two-dimensional compression of spatial region compression and frequency region compression. Dual-region compression may include two steps, namely spatial-frequency joint projection and compression. Spatial-frequency joint projection is shown as separately projecting the spatial-frequency two-dimensional channel matrix H using the U1 matrix and the U2 matrix. The U1 matrix represents the spatial region basis, and the U2 matrix represents the frequency region basis. The projected equivalent coefficient matrix

Number

Number

Number

[0091] 3. CSI Feedback CSI feedback is sometimes also called a CSI report. CSI feedback is that in a wireless communication system, the receiving side (for example, a terminal device) reports information used to describe the channel attributes of a communication link to the transmitting side (for example, a network device). The CSI report includes, for example, a precoding matrix indicator (PMI), a rank indicator (RI), and a channel quality indicator (CQI). The above-listed content included in the CSI is merely an example for explanation and should not constitute any limitation to the embodiments of this application. The CSI may include one or more of the above items, or may include other information used to represent CSI other than the above-listed information. This is not limited in the embodiments of this application.

[0092] 4. Neural Network (NN) A neural network is a specific implementation form of machine learning technology. According to the universal approximation theorem, theoretically, a neural network can approximate any continuous function, and as a result, a neural network can learn any mapping. In traditional communication systems, the design of communication modules requires rich expertise. However, a deep learning communication system based on neural networks can automatically find implicit pattern structures from a large amount of dataset, establish the mapping relationship between data, and obtain better performance than traditional modeling methods.

[0093] For example, a deep neural network (DNN) is a neural network with a relatively large number of layers. Based on different network structures and usage scenarios, DNN may include a multi-layer perceptron (MLP), a convolutional neural network (CNN), a recurrent neural network (RNN), etc. The specific form of DNN is not limited in the embodiments of this application.

[0094] 5. Vector-Quantization Auto-Encoder (VQ-AE) network or simply VQ-AE The VQ-AE network may include an encoder and a corresponding decoder. For example, the encoder and / or decoder may be implemented using a neural network (such as a DNN). In this case, the encoder may sometimes be called an encoder network, and the decoder may sometimes be called a decoder network. Based on an auto-encoder (AE) network, in the VQ-AE network, a vector quantization (VQ) format is combined, and an optimization mechanism for joint training is designed. For example, in an AE network, the encoder and the corresponding decoder may be obtained by joint training, and in a VQ-AE network, the encoder, the corresponding decoder, and the corresponding VQ dictionary may be obtained by joint training. The encoder, decoder, and VQ dictionary obtained by training can be used to encode and decode information. For example, an image is encoded using VQ-AE. As shown in Figure 2, after the original image is input to the encoder of VQ-AE, the output eigenmatrix is, for example, Z e (x). In this case, the original image is the input information of the encoder, and Z e (x) is the output information or encoding result of the encoder.

[0095] In a possible implementation form, the feature distance between each D-dimensional vector in the eigenmatrix Z e (x) and each of the x preset D-dimensional vectors is calculated. The x preset D-dimensional vectors may be regarded as a vector quantization dictionary, or the vector quantization dictionary includes the x D-dimensional vectors and optionally includes another vector. As described above, the vector quantization dictionary is also obtained by training. In Figure 2, e1~e x represents the x preset D-dimensional vectors. For the D-dimensional vector in Z e (x), when the feature distance between the D-dimensional vector in the x D-dimensional vectors and the i-th D-dimensional vector is the smallest, the index matrix Z qTo obtain Z(x), the index i of the i-th D-dimensional vector among the x D-dimensional vectors is filled into the two-dimensional grid Q(z|x). The index matrix Z q (x) may be transmitted to the decoder side. In another possible implementation, the eigenmatrix Z e For each D-dimensional vector in Z(x), among the x D-dimensional vectors, if the feature distance between the D-dimensional vector and the i-th D-dimensional vector in the x D-dimensional vectors is less than the threshold, the index i of the i-th D-dimensional vector in the x D-dimensional vectors is filled into the index matrix Z q To obtain Z(x), it is filled into the two-dimensional grid Q(z|x). The index matrix Z q (x) may be transmitted to the decoder side. When the i-th D-dimensional vector is determined, all x D-dimensional vectors may be traversed, or some of the x D-dimensional vectors may be traversed. This is not limited. The specific implementation for determining the i-th D-dimensional vector from the x D-dimensional vectors is not limited in this application.

[0096] The decoder side may process the index matrix Z q (x). For example, the indices in the index matrix Z q (x) are restored to the corresponding matrix based on the vector quantization dictionary. For example, the matrix is represented as p(×|z q ), and the matrix p(×|z q ) may be used as the input information of the decoder. In this case, the decoder may reconstruct the information about the original image based on that information.

[0097] In the embodiments of the present application, unless otherwise specified, the number of nouns indicates "singular noun or plural noun", that is, "one or more". "At least one" means one or more, and "a plurality of" means two or more. "And / or" describes the relationship of association for describing associated objects, indicating that three relationships can exist. For example, A and / or B may indicate the following cases, that is, when only A exists, when both A and B exist, and when only B exists. In that case, A and B may be singular or plural. When a feature is indicated, the character " / " may indicate an "or" relationship between associated objects. For example, A / B indicates A or B. When an operation is indicated, the symbol " / " may further indicate a split operation. In addition, in the embodiments of the present application, the symbol "×" may be replaced by the symbol "*".

[0098] Ordinal numbers such as "first" and "second" in the embodiments of the present application are used to distinguish a plurality of objects and are not intended to limit the size, content, order, time series, application scenario, priority, or importance of the plurality of objects. For example, the first vector quantization dictionary and the second vector quantization dictionary may be the same vector quantization dictionary or different vector quantization dictionaries. In addition, these names do not indicate that the two vector quantization dictionaries include different contents, priorities, application scenarios, importance, etc.

[0099] In one possible implementation, the CSI feedback mechanism uses the procedure shown in FIG. 3.

[0100] S31: The base station transmits signaling, and correspondingly, the UE receives signaling from the base station.

[0101] Signaling is used to configure channel measurement information. For example, the signaling is used to notify the UE of at least one of the time information for performing channel measurement, the type of reference signal (RS) for performing channel measurement, the time domain resource of the reference signal, the frequency domain resource of the reference signal, and the reporting condition of the measured quantity.

[0102] S32: The base station transmits a reference signal to the UE, and correspondingly, the UE receives the reference signal from the base station.

[0103] The UE measures the reference signal to obtain CSI.

[0104] S33: The UE transmits CSI to the base station, and correspondingly, the base station receives the CSI from the UE.

[0105] S34: The base station transmits data to the UE based on the CSI, and correspondingly, the UE receives the data from the base station.

[0106] Data transmitted by the base station to the UE is carried on the downlink channel, for example, carried on the physical downlink shared channel (PDSCH). Optionally, CSI feedback technology may be used by the base station to transmit another downlink channel, for example, the physical downlink control channel (PDCCH), to the UE. In this embodiment of the present application, an example where CSI is used to transmit PDSCH is used for illustration. The base station may determine the amount of flow used when data is transmitted to the UE based on the RI feedback by the UE, and based on the CQI feedback by the UE, determine the modulation order used when data is transmitted to the UE, and / or the coded bit rate of the channel carrying the data, etc. In addition, the base station may determine the precoding matrix used when data is transmitted to the UE based on the PMI feedback by the UE.

[0107] In one possible implementation, in CSI feedback technology, the UE may perform compression (such as dual-domain compression) on the measured channel matrix using the sparse representation of the channel, and then quantize the coefficients extracted by the compression to obtain CSI. However, in dual-domain compression, some coefficients with higher energy need to be selected from the equivalent sparse matrix

Number

[0108] In consideration of this, a technical solution in an embodiment of the present application is provided. In an embodiment of the present application, the channel state information is determined based on a first vector quantization dictionary. For example, the channel state information may be represented using the index of one or more vectors included in the first vector quantization dictionary. Since the dimension of the vector is usually relatively large, it corresponds to dimensionality expansion of information or minimization of possible information loss, thereby improving quantization accuracy. In addition, since the feedback information is the index of the vector, the original channel matrix is compressed, and thus it may be regarded as reducing the signaling overhead. In addition, in the VQ-AE network, the quantization network and the compression network are jointly designed so that joint optimization can be implemented between compression and quantization, thereby reducing performance loss and improving the overall accuracy of the feedback channel state information.

[0109] FIG. 4A shows the architecture of a communication network in a communication system 10 according to an embodiment of the present application. Any one of the embodiments shown in FIGS. 5 to 8 provided later is applicable to the architecture. The network device included in FIG. 4A is, for example, the access network device 20 included in the communication system 10, and the terminal device included in FIG. 4A is, for example, the communication device 30 included in the communication system 10. The network device can communicate with the terminal device.

[0110] The machine learning technology in the embodiment of the present application is a specific implementation form of artificial intelligence (AI) technology. For ease of understanding, AI technology will be described below. It will be understood that the description is not intended to limit the embodiments of the present application.

[0111] AI is a technology for simulating the human brain to execute complex calculations. With the improvement of data storage and functions, AI is being increasingly applied.

[0112] Figure 4B is a schematic diagram of a first application framework of AI in a communication system. A data source is used to store training data and inference data. A model training host analyzes or trains the training data provided by the data source to obtain an AI model, and deploys the AI model to a model inference host. The model inference host uses the AI model to perform inference based on the inference data provided by the data source and obtains an inference result. The inference result is uniformly planned by an execution entity (actor) and sent to one or more execution objects (e.g., network entities) for execution.

[0113] For example, the AI model includes a decoder network within a VQ-AE network. The decoder network is deployed on the network device side. The inference result of the decoder network is used, for example, for the reconstruction of a downlink channel matrix. The AI model includes an encoder network within a VQ-AE network. The encoder network is deployed on the UE side. The inference result of the encoder network is used, for example, for the encoding of a downlink channel matrix.

[0114] Figure 4C, Figure 4D, or Figure 4E is a schematic diagram of a second application framework of AI in a communication system.

[0115] A first AI module independent of the base station receives training data. The first AI module analyzes or trains the training data to obtain an AI model. Referring to FIG. 4C, for the parameters, the first AI module may perform an inference using the corresponding AI model and inference data to obtain the parameters. Alternatively, referring to FIG. 4D, for the parameters, the first AI module may transmit information about the AI model to a second AI module arranged at the base station (or arranged in the RAN in the description), and the second AI module performs an inference using the corresponding AI model and inference data to obtain the parameters. Alternatively, referring to FIG. 4E, the AI model used by the second AI module for inference may receive the training data and be obtained by the second AI module by training the training data. The AI model includes a decoder network in the VQ-AE network. On the base station side, the inference result of the decoder network is used, for example, for reconstructing the downlink channel matrix. Optionally, the AI model includes an encoder network in the VQ-AE network, and the model information of the encoder network may be transmitted to the UE for inference performed by the UE.

[0116] In the framework from FIG. 4B to FIG. 4E, it should be noted that the AI model may sometimes be simply referred to as the model, and the AI model may be regarded as a mapping from input parameters (e.g., input matrix) to output parameters (e.g., output matrix). For example, in the case of the decoder network on the base station side, the input matrix may be a matrix determined based on the received CSI. The training data may include known input matrices, or may include known input matrices and corresponding output matrices. The training data is used to train the AI model. The training data may be data from a base station, CU, CU-CP, CU-UP, DU, RU, UE, and / or another entity, and / or data inferred using AI technology. This is not limited. The inference data includes an input matrix and is used to infer an output matrix using the model. The inference data may be data from a base station, CU, CU-CP, CU-UP, DU, RU, UE, and / or another entity. The inferred matrix is regarded as policy information and may be sent to the execution target. The inferred matrix may be sent to a base station, CU, CU-CP, CU-UP, DU, RU, UE, etc. for further processing, such as the reconstruction of the downlink channel matrix.

[0117] In this embodiment of the present application, on the network side, the decoder network in the VQ-AE network may be deployed in a network device (e.g., a base station), or may be deployed in, for example, an RU, DU, or a second AI module, or may be deployed in an AI device (e.g., a first AI module) independent of the network device. This is not limited. The inference result of the decoder network may be obtained by the network device through inference, or may be sent to the network device after the first AI module executes the inference. For the sake of simplicity, in this embodiment of the present application, an example where the decoder network is deployed in the network device is used for illustration.

[0118] In this embodiment of the present application, the encoder network within the VQ-AE network is deployed within the UE, and the UE can perform inference using the encoder network.

[0119] The method provided in the embodiments of the present application will be described below with reference to the accompanying drawings. In the accompanying drawings corresponding to the embodiments of the present application, all steps represented by dashed lines may be optional steps. In these methods, the steps or operations included are merely examples. In the embodiments of the present application, other operations or variations of various operations may be further performed. Additionally, the steps may be performed in a different order than presented in the embodiments of the present application, and it is possible that not all operations need to be performed.

[0120] In technologies where CSI feedback is performed with reference to a neural network, a relatively common neural network architecture is a dual architecture. As an example, an autoencoder is used, and compression transmission can be implemented by co-optimization of the encoder and the decoder. For example, one or more reference network groups (or referred to as one or more reference networks) can be obtained by training. Each of these reference networks may be a pair of an encoder network (or referred to as a reference encoder network) and a decoder network (or referred to as a reference decoder network). In other words, the reference network group may include a reference encoder network and a corresponding reference decoder network. These reference networks can be trained offline or online. If the training is performed offline, it may be specified by a protocol. For example, the protocol may provide parameters such as the network structure and / or weights of these reference networks (including the reference encoder network and the reference decoder network). The UE or the network device can implement these reference networks according to the protocol. Optionally, according to the protocol, the evaluated performance corresponding to these specific reference networks may be provided based on an agreed dataset.

[0121] For example, the encoding is performed on the UE side. In actual applications, the UE selects a reference decoder network within one or more reference network groups based on factors such as complexity or performance, and may train one or more encoder networks in a targeted manner (offline or online) that can be used in actual applications. In the case of a reference decoder network, the encoder network trained by the UE in a targeted manner may be referred to as an encoder network corresponding to the reference decoder network. Correspondingly, the network device may also select a reference encoder network within one or more reference network groups and train one or more decoder networks in a targeted manner (offline or online) that can be used in actual applications. In the case of a reference encoder network, the decoder network trained by the network device in a targeted manner may be referred to as a decoder network corresponding to the reference encoder network. The protocol may not limit whether the actually used encoder network or decoder network is a reference network pre-defined by the protocol, but may limit the evaluated performance of the actually used encoder network and the corresponding reference decoder network so as to meet the indicators specified by the protocol, and may further limit the evaluated performance of the actually used decoder network and the corresponding reference encoder network so as to meet the indicators specified by the protocol. The protocol may limit that the actually deployed encoder network has the same input and output dimensions as the corresponding reference encoder network, and / or may limit that the decoder network has the same input and output dimensions as the corresponding reference decoder network.

[0122] One embodiment of the present application is based on a VQ-AE network. The difference from an autoencoder is that the reference network group (or reference network) in the VQ-AE network can be a pair of a vector quantization encoder network (sometimes simply called a reference encoder network) and a vector quantization decoder network (sometimes simply called a reference decoder network), and a corresponding vector quantization dictionary (set). The vector quantization dictionary (set) is sometimes called a reference vector quantization dictionary (set). The vector quantization encoder network, the vector quantization decoder network, and the vector quantization dictionary (set) can be jointly trained and optimized. Similar to the above steps, one or more reference network groups can be trained offline or online. When an offline training method is used, the protocol may define one or more reference network groups and optionally define a method and indicator for evaluating the performance of the reference network.

[0123] Optionally, the vector quantization dictionary A defined by the reference network group does not necessarily have to be a universal set of the vector quantization dictionary B defined by the protocol. For example, the protocol may define a vector quantization dictionary B that contains more vectors, and the vector quantization dictionary A corresponding to the group of reference encoder networks / reference decoder networks may only be a subset of the vector quantization dictionary B. The "definition" in this specification can be a related relationship. For example, the protocol defines a reference vector quantization dictionary set. The set contains a plurality of vector quantization dictionaries. The reference vector quantization dictionary A included in the reference network group is, for example, one of the reference vector quantization dictionaries in the reference vector quantization dictionary set. For example, in the reference vector quantization dictionary set, each reference vector quantization dictionary corresponds to an index. The reference vector quantization dictionary included in the reference network group may correspond to an index in the reference vector quantization dictionary set.

[0124] Optionally, before step S501 in one embodiment of the present application starts, it may be assumed that two communication parties train an encoder network and a decoder network online or offline respectively for an actual application (e.g., meeting performance indicators) based on one or more reference network groups specified by a protocol. Since the vector quantization dictionary needs to be known to the two communication parties, at the initial stage of network access communication, the vector quantization dictionary used by the two communication parties may be predefined in the corresponding reference network or determined in another way, for example. With the support of subsequent signaling mechanisms, one communication party may exchange the vector quantization dictionary trained offline by the communication parties with the other communication party and actually use the vector quantization dictionary.

[0125] FIG. 5 is a flowchart of a first communication method according to an embodiment of the present application.

[0126] S501: The network device transmits configuration information to the UE, and correspondingly, the UE receives the configuration information from the network device.

[0127] For example, the configuration information is called configuration information 1 or third configuration information. The third configuration information may be used to configure parameters used during CSI measurement and / or CSI reporting. Therefore, the configuration information may also be called CSI measurement configuration information.

[0128] The third configuration information can be sent by the network device to the UE. The third configuration information may indicate, for example, a CSI resource configuration (CSI resource setting) and / or a CSI reporting configuration (reporting amount). For example, the third configuration information can be used to indicate at least one of the type of reference signal used for measurement (e.g., channel state information reference signal (CSI-RS) or a signal in the synchronization signal and physical broadcast channel block (SSB)), the time-domain resource and / or frequency-domain resource of the reference signal (e.g., the index of the reference signal, the port number of the reference signal, and / or the period of the reference signal), the CSI reporting format, the time-domain resource and / or frequency-domain resource used for CSI reporting, the trigger condition for CSI reporting, and / or the period of CSI reporting.

[0129] In this embodiment of the present application, the UE determines CSI using a vector quantization dictionary in the VQ-AE network. The vector quantization dictionary in the VQ-AE network is similar to a public key or a public codebook and needs to be known in advance to both communication parties (e.g., the network device and the UE). Therefore, the third configuration information may also include information about the vector quantization dictionary. Examples are used below to explain the definition method of the vector quantization dictionary.

[0130] 1. The protocol predefines a vector quantization dictionary. The vector quantization dictionary may be predefined by a protocol. Specifically, the protocol may define one or more vector quantization dictionaries, and may also define the format of each vector quantization dictionary, specific parameters, etc. In this case, two communication parties may obtain all information regarding the vector quantization dictionary according to the protocol. The information includes the specific values of each vector included in the vector quantization dictionary (or weights, i.e., may sometimes be called the weights of the vector quantization dictionary including the values of some or all elements within the vector quantization dictionary). Each vector quantization dictionary includes one or more vectors.

[0131] The weights of the vector quantization dictionary may be updated together with the encoder network / decoder network. Therefore, in an offline training scenario, the weights of the vector quantization dictionary can also be determined in an offline training manner. However, offline training usually focuses on the generalization performance of the model. Therefore, the vector quantization dictionary obtained by offline training may generally be considered applicable to all users in a given communication scenario. Generally, until the UE is notified that it is necessary to switch or update the vector quantization dictionary, it is appropriate for the device to preload the general-purpose vector quantization dictionary obtained by offline training and use it as the first vector quantization dictionary when the UE first accesses the network. Generally, a predefined vector quantization dictionary with generality and a pair of a reference encoder network and a reference decoder network may form a default reference network group. Optionally, the default reference network group may further include another parameter. This is not limited. The default reference network group may be the reference network used by the UE when the UE first accesses the network, when the UE enters the fallback mode, and / or when the network device does not indicate the UE using signaling.

[0132] 2. The protocol predefines the format of the vector quantization dictionary. The protocol may define one or more vector quantization dictionaries, and may define the format of each vector quantization dictionary. For example, it may define the length of the vectors included in the vector quantization dictionary and / or the number of vectors. Optionally, the weights of the vectors may not be defined, or some elements of the vectors may be defined, or the weights of some vectors in the dictionary may be defined. The undefined weights of the vectors can be filled or updated via another mechanism.

[0133] When a solution that uses a VQ-AE network (e.g., simply called a VQ-AE-CSI solution) to feedback CSI is actually applied, for some UEs located at relatively fixed positions within the cell, the UEs may not require generalization of the network. Therefore, parameter filling or updating can be performed on the vector quantization dictionary based on the real-time channel environment. This method can be described as a step of determining dictionary weights that conform to the channel environment of the small area where these UEs are located in some online training methods. In this way, the CSI reconstruction accuracy can be further improved. The online training (or called online update) method is described in another embodiment below.

[0134] 3. The network device transmits the vector quantization dictionary. The network device may send one or more vector quantization dictionaries in a format and / or specific parameters (e.g., weight parameters) to the UE, and the UE uses the vector quantization dictionary from the network device as the vector quantization dictionary in the VQ-AE-CSI solution. When the network device sends multiple vector quantization dictionaries to the UE, the network device may further indicate to the UE the vector quantization dictionary for a specific use, or the UE may determine the vector quantization dictionary for a specific use based on other factors (e.g., application scenario). This may be the case in a scenario where multiple small cells are deployed. Each small cell may have limited coverage. In the case of limited coverage, the small cell obtains the target vector quantization dictionary through offline training or online training. The vector quantization dictionary can be used by the UE accessing the small cell. In this case, the network device may determine the format of the possible vector quantization dictionary, weight parameters, etc., and distribute the vector quantization dictionary to the UE.

[0135] The protocol may pre-define multiple vector quantization dictionaries. The UE may determine the vector quantization dictionary in an implicit manner, or an explicit manner, or a combination of an implicit manner and an explicit manner. Hereinafter, examples are used for illustration.

[0136] 1. Explicit manner For example, the explicit manner includes the UE determining the vector quantization dictionary for a specific use based on the signaling (e.g., indication signaling or switching signaling) sent by the network device.

[0137] In this embodiment of the present application, the signaling sent by the network device to the UE is, for example, radio resource control (RRC) signaling, media access control (MAC) control element (CE), or downlink control information (DCI).

[0138] Referring to this embodiment of the present application, the signaling includes, for example, third configuration information or other signaling. For example, the third configuration information further includes the format and / or weight parameters of the first vector quantization dictionary (corresponding to the aforementioned method in which the network device transmits the vector quantization dictionary), or further includes an index (the protocol corresponds to the aforementioned method of predefining the vector quantization dictionary, and for example, the third configuration information indicates the index of the first vector quantization dictionary from a plurality of vector quantization dictionaries predefined by the protocol). The first vector quantization dictionary is, for example, a dictionary used by the VQ-AE network. The network device may use the third configuration information to send the first vector quantization dictionary or the index of the first vector quantization dictionary to the UE. This is a relatively simple dictionary distribution method. In a scenario where real-time update is not required, the first vector quantization dictionary can be distributed using the third configuration information (usually RRC signaling).

[0139] When the signaling is MAC-CE or DCI, the signaling may indicate the index or index set of the first vector quantization dictionary. In this design, the signaling overhead can be reduced.

[0140] When the notification is executed in an explicit manner, the network device may send one signaling or multiple signalings to the UE. Regarding the method of sending multiple signalings, the notification method of multiple signalings may be a combination of several aforementioned signaling notification methods. For example, in the combination method, the network device uses Signaling 1 (e.g., RRC signaling or MAC-CE) to send an index set of a group of vector quantization dictionaries, and the network device further indicates the index of the first vector quantization dictionary using DCI. The first vector quantization dictionary is included in the group of vector quantization dictionaries, and the UE can determine the first vector quantization dictionary based on the two signalings. The format of the group of vector quantization dictionaries, specific parameters, etc. may be included in the third configuration information. In other words, information such as the format and parameters of the group of vector quantization dictionaries is sent using the third configuration information, and then Signaling 1 and DCI are used to indicate the specific vector quantization dictionary to be used. Alternatively, the formats, specific parameters, and / or indexes of various vector quantization dictionaries within the group of vector quantization dictionaries may be alternatively specified by the protocol. In another example, in another combination method, the network device uses RRC signaling to send a group of vector quantization dictionaries (including the format, weight, and / or index of each group of vector quantization dictionaries), and uses MAC-CE or DCI to indicate the index of the vector quantization dictionary within the group of vector quantization dictionaries, and the UE can determine the vector quantization dictionary based on the two signalings. Alternatively, there is another combination method. This is not limited in this embodiment of the present application.

[0141] The aforementioned explicit indication method is merely an example of the solution in this embodiment of the present application and is not a limitation to the solution in this embodiment of the present application. Any method of determining the vector quantization dictionary in an explicit manner falls within the protection scope of this embodiment of the present application.

[0142] 2. Implicit manner In an alternative implementation of the implicit manner, in the protocol, a vector quantization dictionary (e.g., the first vector quantization dictionary) may be predefined. The first vector quantization dictionary is common to all channels of the UE. In this case, the switching of the vector quantization dictionary is not relevant. In this case, the network device may not need to transmit the first vector quantization dictionary, and the UE may determine the first vector quantization dictionary according to the protocol, and the network device may also determine the first vector quantization dictionary according to the protocol. This is a relatively simple implementation of the protocol. In this embodiment of the present application, the first vector quantization dictionary may be defined together with the (default) reference network. In other words, the reference network may include a reference vector quantization dictionary, a decoder network, and an encoder network.

[0143] In another alternative implementation of the implicit manner, the protocol may define a plurality of vector quantization dictionaries. For example, these vector quantization dictionaries have a correspondence with the configuration of one or more other system parameters. Therefore, when one or more system parameters change, the vector quantization dictionary corresponding to the system parameters is also switched accordingly. Hereinafter, examples are used for illustration.

[0144] For example, there is a correspondence between a networking scenario (or what is called a networking method) and each of a plurality of vector quantization dictionaries. Different networking methods may coincide with different vector quantization dictionaries or may coincide with the same vector quantization dictionary. In this case, for example, if the UE determines that the current networking method is the first networking method, the UE may determine a vector quantization dictionary corresponding to the first networking method (for example, the first vector quantization dictionary). Therefore, the network device may not need to transmit the first vector quantization dictionary. The network device may also determine the first vector quantization dictionary according to the protocol. For example, different networking methods include a line of sight (LoS) path and a non line of sight (nLoS) path, and two networking methods may respectively correspond to vector quantization dictionaries. In another example, different networking methods include an urban microcell (Umi) and an urban macrocell (Uma), and Umi and Uma may respectively correspond to their own vector quantization dictionaries. In this implementation form, the vector quantization dictionary is bound to the networking method. Therefore, the switching of the vector quantization dictionary is in an implicit manner. Specifically, when the communication scenario (or networking method) is determined, the vector quantization dictionary is also determined. Conversely, when the communication scenario is switched, the vector quantization dictionary may be switched accordingly.

[0145] As another example, the plurality of vector quantization dictionaries may correspond to different moving speeds. For example, the moving speed of the UE may be divided into a plurality of intervals, and different intervals respectively correspond to corresponding vector quantization dictionaries. The division granularity is not limited in this embodiment of the present application. Different intervals may correspond to different vector quantization dictionaries or the same vector quantization dictionary. In this case, for example, if the UE determines that the current moving speed belongs to the first interval, the UE may determine the vector quantization dictionary corresponding to the first interval (for example, the first vector quantization dictionary). Therefore, the network device may not need to transmit the first vector quantization dictionary. In this implementation, the vector quantization dictionary is bound to the moving speed of the UE. Therefore, the switching of the vector quantization dictionary is in an implicit manner. Specifically, when the moving speed of the UE is determined, the vector quantization dictionary is also determined. Correspondingly, when the moving speed of the UE changes, the corresponding vector quantization dictionary can also be switched.

[0146] In another example, the plurality of vector quantization dictionaries may correspond to different antenna configurations and / or bandwidths. Here, the bandwidth may be, for example, the bandwidth of the carrier wave used for communication between two communication parties, or the maximum bandwidth supported by the UE, or the bandwidth of the UE's active bandwidth part (BWP). This is not limited. When the antenna configuration and / or bandwidth changes, the input and output dimensions of the encoder network and the decoder network may change accordingly. In this embodiment of the present application, the vector quantization dictionary may be associated with the encoder network and / or the decoder network. In particular, when the vector quantization dictionary and the encoder network and / or the decoder network are jointly trained and optimized, if the encoder network and / or the decoder network change, the vector quantization dictionary may also change accordingly. Therefore, after determining the antenna configuration and / or bandwidth, the UE determines the first vector quantization dictionary. In this implementation, a correspondence relationship is established between the vector quantization dictionary and the antenna configuration and / or bandwidth. Therefore, the switching of the vector quantization dictionary is in an implicit manner.

[0147] Some of the aforementioned implicit methods may be applied separately, or any combination of a plurality of the aforementioned implicit methods may be applied. For example, in the way of applying a combination, the protocol may define a plurality of vector quantization dictionaries, and these vector quantization dictionaries may correspond to different reference networks and moving speeds. For example, there may be a plurality of combinations of moving speed and reference network, and different combinations may correspond to the corresponding vector quantization dictionaries. For example, when the moving speed of the UE belongs to the first interval and the current reference network is the first reference network, this case corresponds to the vector quantization dictionary 1, or when the moving speed of the UE belongs to the first interval and the current reference network is the second reference network, this case corresponds to the vector quantization dictionary 2.

[0148] The protocol can define the corresponding vector quantization dictionary, encoder network, and decoder network using the form of the reference network. Considering this, the implicit instructions in some of the foregoing implementation forms can be uniformly given by the index of the reference network. In other words, the UE can determine the index of the reference network based on the current networking method, moving speed, antenna configuration, or system bandwidth, and then determine the first vector quantization dictionary based on the index.

[0149] 3. Combination of explicit and implicit methods In one implementation form, the protocol can define one or more sets of vector quantization dictionaries (for example, the protocol can pre-define a plurality of vector quantization dictionaries, and some or all of the vector quantization dictionaries may belong to one or more sets of vector quantization dictionaries), and each set of vector quantization dictionaries includes one or more vector quantization dictionaries. The UE can determine the set of vector quantization dictionaries in an explicit manner and determine the vector quantization dictionary to be used within the set of vector quantization dictionaries in an implicit manner. Alternatively, the UE can determine the set of vector quantization dictionaries to be used in an implicit manner and determine the vector quantization dictionary to be used within the set of vector quantization dictionaries in an explicit manner.

[0150] For example, the network device may send signaling to the UE, and the signaling indicates the vector quantization dictionary set 1 within a plurality of vector quantization dictionary sets defined by the protocol. In other words, the UE may determine the vector quantization dictionary set in an explicit manner. The vector quantization dictionary set 1 includes a plurality of vector quantization dictionaries. For example, the vector quantization dictionaries included in the vector quantization dictionary set 1 have a correspondence with the moving speed of the UE, and the correspondence is preconfigured in the UE. In this case, the UE may determine the current moving speed of the UE and determine that the vector quantization dictionary corresponding to the interval in which the current moving speed of the UE is located is the vector quantization dictionary to be used. It can be found that the UE determines the vector quantization dictionary within the vector quantization dictionary set in an implicit manner.

[0151] In another example, there is a correspondence between the networking method and each of a plurality of vector quantization dictionary sets defined by the protocol, and the correspondence is preconfigured in the UE. In this case, the UE may determine the current networking method of the UE and the network device and determine that the vector quantization dictionary set corresponding to the networking method is the vector quantization dictionary set to be used. For example, the vector quantization dictionary set is the vector quantization dictionary set 1. In other words, the UE may determine the vector quantization dictionary set in an implicit manner. The network device may send signaling to the UE. The signaling indicates the vector quantization dictionary A within the vector quantization dictionary set 1. In this case, the UE determines to use the vector quantization dictionary A. In other words, the UE may determine the vector quantization dictionary in an explicit manner.

[0152] Of course, the above are merely examples. The implementation forms combining the explicit manner and the implicit manner are not limited thereto.

[0153] S502: The UE prepares a neural network (NN)-CSI network, and the network device prepares an NN-CSI network. The neural network here may be, for example, the aforementioned VQ-AE network or another neural network.

[0154] Step S502 is the behavior of the device implementation form. The specific implementation form of step S502 is not limited. For example, when the network device sends the third configuration information to the UE, the third configuration information may include information bits for indicating to the UE to enable the artificial intelligence (AI)-based CSI feedback procedure or for indicating to the UE to enable the VQ-AE network-based CSI feedback procedure. In this case, the network device immediately prepares an NN-CSI network (for example, an NN decoder network). After receiving the third configuration information, the UE prepares an NN-CSI network (for example, an NN encoder network).

[0155] The "preparation" described in this specification can be understood as loading the network model (into the memory) or switching to the neural network module (from the conventional codebook calculation module).

[0156] S503: The network device sends a reference signal to the UE, and correspondingly, the UE receives the reference signal from the network device.

[0157] The network device may send a reference signal to the UE based on the reference signal format indicated by the third configuration information. The reference signal includes, for example, CSI-RS and / or SSB.

[0158] S504: The UE acquires the downlink channel matrix.

[0159] The UE may measure reference signals to obtain the downlink channel matrix. The downlink channel matrix may also be referred to as the downlink MIMO channel. For example, the dimension of the downlink channel matrix is N tx ×N rx ×N RB where N tx represents the number of transmit antenna ports of the reference signal transmitting device. The transmit antenna ports may include transmit antenna ports with different polarization directions. In this specification, N tx is a positive integer, N rx represents the number of receive antenna ports of the UE, and the value of N rx is a positive integer. The receive antenna ports may include receive antenna ports with different polarization directions. In this specification, N RB represents the number of frequency domain resource blocks (RBs), and the value of N RB is a positive integer. N RB may represent the number of RBs included in the measurement bandwidth. The measurement bandwidth may be, for example, the system bandwidth, the sub-band bandwidth, the maximum bandwidth supported by the UE, the bandwidth of the UE's BWP, etc. This is not limited. For example, when the carrier bandwidth is 10 MHz and the sub-carrier spacing is 15 kHz, N RB is 52. In another example, when the carrier bandwidth is 20 MHz and the sub-carrier spacing is 15 kHz, N RB is 104. In another example, when the carrier bandwidth is 10 MHz and the sub-carrier spacing is 30 kHz, N RB is 26, etc. For example, in this embodiment of the present application, the network device transmits reference signals, and N tx represents the number of transmit antenna ports of the network device. However, when the technical solution provided in this embodiment of the present application is applied to another network architecture, the reference signal transmitting device may change, and N tx represents the number of transmit antenna ports of the transmitting device.

[0160] S505: The UE preprocesses the downlink channel matrix.

[0161] The UE can preprocess the downlink channel matrix in multiple ways. Hereinafter, examples are used for explanation.

[0162] 1. The first method In the first method, the preprocessing process of the UE may include multiple steps. The steps are described below.

[0163] Step 1: Determine the physical meaning of H, in other words, determine the eigen-subspace matrix H.

[0164] Optionally, the UE performs singular value decomposition (SVD) on the downlink channel matrix obtained in step S504 based on the MIMO channel (N tx ×N rx ) of each subband to obtain the eigen-subspace matrix H. The dimension of the eigen-subspace matrix H is, for example, N tx ×N sb . This corresponds to performing dimensionality reduction processing on the downlink channel matrix via SVD. N sb is the number of frequency-domain subbands, for example, N sb =N RB / a, where a represents the frequency-domain subband granularity or subband bandwidth, that is, the number of RBs included in each subband. The common frequency-domain subband granularity is 2 RBs or 4 RBs. When 4 RBs are used as an example, N sb =N RB / 4. It should be understood that the frequency-domain subband granularity can be any value. For example, when the granularity is 1 RB, N sb =N RB .

[0165] Step 2: Determine the spatial domain basis and the frequency domain basis.

[0166] Optionally, the spatial domain basis and the frequency domain basis may be generated according to the discrete Fourier transform (DFT) equation. When an orthogonal DFT basis is generated, the dimension of the basis satisfies the A×A format.

[0167] The specific form of the DFT basis is not limited in this application. For example, refer to the corresponding description in the 3GPP standard protocol TS38.214, or refer to another possible form.

[0168] For example, the eigen subspace matrix H is a matrix of dimension N tx ×N sb When the spatial domain basis is generated, A = N tx When the frequency domain basis is generated, A = N sb Therefore, two groups of orthogonal DFT bases can be generated according to the DFT equation, namely the spatial domain basis

Number

Number

Number

Number

[0169] Alternatively, optionally, a plurality of orthogonal spatial domain bases {G1, G2, G3...} and a plurality of orthogonal frequency domain bases {F1, F2, F3...} may be generated in an oversampling manner, where G i and F jThe group is selected as the spatial domain basis and the frequency domain basis in this embodiment of the present application. For example, a group with a relatively accurate projection direction may be selected. For example, each group of the spatial domain basis and the corresponding frequency domain basis is processed by the following spatial frequency joint projection method. Based on the result of the spatial frequency joint projection, a group with a relatively accurate projection direction can be determined. The group is determined as the spatial domain basis and the frequency domain basis in this embodiment of the present application.

[0170] Step 3: Sparse representation C of the eigen-subspace matrix H of the downlink channel matrix complex to determine.

[0171] Optionally, the UE may perform a spatial frequency joint projection on H to determine the sparse representation C of H. For the method of performing a spatial frequency joint projection on H, refer to the following formula. complex For the method of performing a spatial frequency joint projection on H, refer to the following formula.

Equation

[0172] G H is the Hermitian matrix of G, also called the self-conjugate matrix, and can be obtained by performing conjugate transpose on the matrix G. G i and F j When a group of G and F is selected from multiple groups of the spatial domain basis and the frequency domain basis, G in Equation 1 H is

Equation

[0173] The complex matrix C obtained according to Equation 1 complex is the sparse representation of the eigen-subspace matrix H of the downlink channel matrix, and the dimension of C complex is the same as the dimension of the eigen-subspace matrix H, for example, N tx ×N sb matches.

[0174] Step 4: C complex Determine whether to further compress it.

[0175] Optionally, the dimension of C complex remains unchanged. In other words, Step 4 may not be executed, and Step 5 may be executed after Step 3 is completed.

[0176] Optionally, the UE may further compress C complex There are multiple compression methods described in detail in Step 4 of the second method. Details are not described here.

[0177] Step 5: Convert the dimension of the matrix according to the input requirements of the encoder network.

[0178] The UE may perform dimension conversion on the coefficient matrix C according to the input requirements of the encoder network.

[0179] Generally, neural networks process real numbers. When Step 4 is executed, in this step, after the complex matrix C complex is further compressed, the matrix obtained in Step 4 usually needs to be converted into a real matrix. The real matrix is the input matrix of the encoder network. Alternatively, when Step 4 is not executed, in this step, usually, the complex matrix C complex obtained in Step 3 needs to be converted into a real matrix. The real matrix is the input matrix of the encoder network. The method for converting a complex matrix into a real matrix is generally to concatenate in a new dimension after obtaining the matrix formed by the real part of the complex matrix and the matrix formed by the imaginary part of the complex matrix respectively.

[0180] Regardless of whether compression is performed in Step 4, the complex matrix before conversion in Step 5 is

Number

[0181] When a complex matrix is a one - dimensional vector, for example

Number

Number

Number

[0182] A method for converting a complex matrix with a higher dimension into a real matrix can be inferred by analogy.

[0183] In this embodiment of the present application, for example, for the complex matrix

Number

Number

[0184] After the preprocessing of the above five steps, in the first method, the downlink channel matrix (complex matrix) with dimensions N tx ×N rx ×N RB is converted into a coefficient matrix (real matrix) with dimensions M×N×2. M = N tx 、and N = N sb . The above preprocessing steps are just examples. This is not limited.

[0185] 2. The second method In the second method, a method for further compressing the complex matrix C complex in step 4 is mainly described, and contents such as the influence of further compression on the matrix dimensions before and after preprocessing are explained. For other steps, please refer to the first method.

[0186] Step 1: Determine the physical meaning of H.

[0187] Optionally, refer to step 1 of the first method. The eigen-subspace matrix H is obtained. The dimension of H is N tx ×N sb .

[0188] Step 2: Determine the spatial domain basis and the frequency domain basis.

[0189] Optionally, refer to step 2 of the first method. A pair of orthogonal DFT spatial frequency bases are obtained:

Number

Number

[0190] Step 3: Determine the sparse representation C of H complex to be determined.

[0191] Optionally, referring to Step 3 of the first method, a spatial frequency joint projection is performed according to Equation 1.

[0192] Step 4: Determine whether to further compress C complex to be determined.

[0193] Optionally, in the second method, the UE may further compress the sparsity of the complex matrix C complex by using the coefficient energy. The compression process may be performed in multiple ways. Hereinafter, an example is used for explanation.

[0194] For example, the method is Method 1. Method 1 is to convert the complex matrix C complex ∈ C U×V into a one-dimensional vector C col ∈ C (U×V)×1 , sort the coefficients included in the one-dimensional vector based on the energy values, and select the first M coefficients having the maximum energy, where M ≤ U × V. In this way, a coefficient matrix is obtained. The coefficient matrix is, for example, represented as C topM_complex ∈ C M×1 .

[0195] For example, another method is Method 2. Method 2 is to select the first K rows (K ≤ U) having the maximum energy in the spatial domain dimension of C complex ∈ C U×V , and select the first L columns (L ≤ V) having the maximum energy in the frequency domain dimension of C complex ∈ C U×V . The coefficient matrix composed of K rows and L columns is CtopK×L_complex ∈ C K×L is represented as. Matrix C topK×L_complex ∈ C K×L is converted into a one-dimensional vector C topK×L_col ∈ C (K×L)×1 The coefficients included in the one-dimensional vector are sorted based on the energy values. The first M coefficients (M ≤ K × L) with the largest energy are selected. In this way, a coefficient matrix is obtained. The coefficient matrix is, for example, C topM_complex ∈ C M×1 is represented as. M, L, and K are positive integers.

[0196] For example, yet another method is Method 3. Method 3 selects the first M rows (M ≤ U) having the maximum energy in the spatial domain dimension of C complex ∈ C U×V , and selects the first N columns (N ≤ V) having the maximum energy in the frequency domain dimension of C complex ∈ C U×V . The coefficient matrix composed of M rows and N columns is C topM×N_complex ∈ C M×N is represented as. M and N are positive integers.

[0197] In the three methods mentioned above for further compressing C included in this step, some coefficients are selected according to rules based on the original size N complex of C complex , and dimensionality reduction is performed to obtain C tx × N sb or C topM_complex ∈ C M×1 or C topM×N_complex ∈ C M×N . In this step, information compression is completed. Therefore, in any one of the three methods, U = N tx , and V = N sbThat is to say, in this step, the original information can be approximately represented by selecting the main features. The dimension of the matrix obtained by the processing in this preprocessing method is usually smaller than the dimension of the matrix obtained by the first method, and is used as the input dimension of the subsequent neural network (for example, the input dimension of the encoder network), thereby reducing the complexity of the neural network.

[0198] C complex It should be understood that when the original size of C changes, only the values of U and V need to be replaced.

[0199] Step 5: Determine the dimension of the input matrix of the encoder network.

[0200] In the second method, the complex matrix output in step 4 is a one-dimensional matrix

Number

Number

[0201] Based on Method 1 or Method 2 of Step 4, the real matrix corresponding to the complex matrix

Number

[0202] Based on Method 3 of Step 4, the real matrix corresponding to the complex matrix

Number

[0203] After the preprocessing of the above five steps, the downlink channel matrix (complex matrix) with dimensions N tx ×N rx ×N RB is converted into a coefficient matrix (real matrix) with dimensions M×2 or M×N×2. Of course, the above preprocessing steps are just examples. This is not limited.

[0204] 3. The third method The third method mainly describes contents such as a method for improving the spatial resolution by using an oversampled spatial frequency basis, and the influence of this method on the matrix dimensions before and after preprocessing.

[0205] Step 1: Determine the physical meaning of H.

[0206] Optionally, refer to step 1 of the first method. The eigen-subspace matrix H is obtained. The dimension of H is N tx ×N sb is.

[0207] Step 2: Determine the spatial domain basis and the frequency domain basis.

[0208] Optionally, when an oversampled DFT basis is generated, the dimension of the basis satisfies the form of A×B, and A < B. Generally, B = A×o, where o is the oversampling coefficient.

[0209] For example, H is an eigen-subspace matrix with dimensions N tx ×N sb is. When the spatial domain basis is generated, A = N tx , and B = N tx ×x. When the frequency domain basis is generated, A = N sb , and B = N sbIt is x × y. In this specification, x and y are oversampling coefficients in the spatial domain and the frequency domain, respectively. Both x and y are greater than 1, or x is equal to 1 and y is greater than 1, or x is greater than 1 and y is equal to 1, and x and y may or may not be equal.

[0210] The eigen subspace matrix H is assumed to be a complex matrix of dimension N tx × N sb In this case, the UE can generate a plurality of spatial domain bases {G1, G2, G3...} and a plurality of frequency domain bases {F1, F2, F3...} using different oversampling rates. According to the above rules, the UE concatenates the plurality of spatial domain bases column by column to obtain the spatial domain base set

Number

Number

Number

Number

[0211] Compared with the first method and the second method, the column vector dimensions of the spatial domain base set and the frequency domain base set still match the corresponding dimensions of H, but the number of column vectors in the base set is increased. This helps to improve the DFT beam direction resolution.

[0212] Step 3: Determine the sparse representation C of H complex

[0213] ​ For the method of performing spatial frequency joint projection on H, refer to the following equation:

Number

[0214]

Number

Number

[0215] The complex matrix C obtained according to Equation 2 complex is the sparse representation of the eigen-subspace matrix H of the downlink channel matrix, and the dimension of C complex is different from the dimension of the eigen-subspace matrix H. The eigen-subspace matrix H is a complex matrix with dimension N tx ×N sb . The dimension of the complex matrix C complex is (N tx ×x)×(N sb ×y). In the third method, it can be found that the dimension of the coefficient matrix can be increased after performing spatial frequency joint projection according to Equation 2. Therefore, the complex matrix obtained by the third method may be regarded as the dimension expansion process being performed on the complex matrix in the first method. From the perspective of sparse projection, the resolution of the original information can be increased, and the accuracy of the original information reconstructed by the network device can be improved.

[0216] Step 4: Determine whether to further compress C complex .

[0217] Optionally, refer to Step 4 of the first method. C complex remains unchanged.

[0218] Alternatively, optionally, refer to any method of step 4 of the second method. C complex is further compressed. For reference, in three exemplary methods of step 4 of the second method, U is N tx replaced by ×x, and V is N sb replaced by ×y.

[0219] Step 5: Determine the dimension of the input matrix of the encoder network.

[0220] Optionally, for example,

Number

Number

[0221] If the complex matrix has another dimension, referring to the example of step 5 of the second method, the corresponding real matrix can also be obtained correspondingly. Details are not described again here.

[0222] After the preprocessing of the five steps described above, if C complex is not further compressed in step 4, the downlink channel matrix (complex matrix) with dimension N tx ×N rx ×N RB is converted into a coefficient matrix (real matrix) with dimension M × N × 2, where M = N tx ×x, and N = N sb ×y. If C complex is further compressed in step 4, the dimension is N tx ×N rx ×N RBThe downlink channel matrix (complex matrix) to be transformed is converted into a coefficient matrix (real matrix) with dimensions of M×2 or M×N×2. For the value ranges of M and N, refer to the relevant descriptions of Method 1 to Method 3 in Step 4 of the second method on the condition that U is replaced by N tx ×x, and V is replaced by N sb ×y. For the relevant descriptions of Method 1 to Method 3 in Step 4 of the second method, refer to the relevant descriptions.

[0223] 4. The Fourth Method, also known as the Full Channel Feedback + 2D Convolutional Network Method In the fourth method, CSI feedback is performed based on the original downlink channel, and DFT joint projection is performed for dimension N tx ×N rx ×N.

[0224] Step 1: Determine the physical meaning of H.

[0225] Optionally, the downlink channel matrix H is obtained using Step S504, and the dimension is N tx ×N rx ×N RB ×N. Different from the eigen-subspace matrix H of the previous method, it can be found that H in this specification represents the downlink channel instead of the eigen-subspace (that is, in the fourth method, SVD is not performed on the downlink channel matrix).

[0226] Step 2: Determine the spatial domain basis and the frequency domain basis.

[0227] Optionally, two orthogonal DFT bases are generated. When the spatial domain basis is generated, A = N tx ×N rx ×N. When the frequency domain basis is generated, A = N sb ×N. Therefore, the two generated bases are the spatial domain basis

Number

Number

[0228] Optionally, G and F may be a group of bases having the most accurate projection directions selected from a plurality of groups of orthogonal oversampling spatial frequency bases (for specific methods, see the foregoing description).

[0229] Optionally, two orthogonal DFT bases are generated. When the spatial domain base is generated, A = N tx ×N rx , B = N tx ×N rx ×x. When the frequency domain base is generated, A = N sb , and B = N sb ×y. In this specification, x and y are oversampling coefficients in the spatial domain and the frequency domain, respectively, and both x and y are greater than 1, or x is equal to 1 and y is greater than 1, or x is greater than 1 and y is equal to 1, and x and y may or may not be equal.

Number

Number

[0230] Step 3: Sparse representation C of Hcomplex Determine.

[0231] The UE performs a spatial-frequency joint projection on the matrix H. When G and F are orthogonal bases, refer to Equation 1 for the spatial-frequency joint projection method. In this specification, [Number] is obtained.

[0232] [Number] represents a 2D matrix. The content in the parentheses is regarded as the dimension. For example, (N tx ×N rx ) represents one dimension, and N sb represents the other dimension. If there are similar cases for other equations, the understanding is the same.

[0233] Optionally, when G and F are oversampled DFT bases, referring to Equation 1, G in Equation 1 H is [Number] may be replaced by, and F is [Number] may be replaced by. For the spatial-frequency joint projection formula, refer to the following formula: [Number]

[0234] Step 4: Determine whether to further compress C complex .

[0235] Optionally, if C complex is not further compressed, this step is skipped.

[0236] Alternatively, optionally, refer to the corresponding steps in a second method for further compression to be performed on C complex For example, U is replaced by N tx ×N rx and V is replaced by N sb or U is replaced by N tx ×N rx ×x and V is replaced by N sb ×y.

[0237] Step 5: Determine the dimension of the input matrix of the encoder network.

[0238] If C complex is not further compressed in Step 4, the real matrix C real ∈R M×N×2 can be obtained, where M = N tx ×N rx and N = N sb . The real matrix is the input matrix of the encoder network.

[0239] If C complex is further compressed in Step 4, the real matrix C real ∈R M×2 or C ∈ R M×N×2 can be obtained. The range of the values of M and N and the relationship with U and / or V remain unchanged as the relative relationship described in Step 4 of the second method. For the absolute values, U is replaced by N tx ×N rx and V is replaced by N sb or U is replaced by N tx ×N rx ×x and V is replaced by N sb ×y. The real matrix is the input matrix of the encoder network.

[0240] After the preprocessing of the five steps described above, if C complex is not further compressed in Step 4, the dimension is N tx ×N rx ×N RBThe downlink channel matrix (complex matrix) that is obtained is converted into a coefficient matrix (real matrix) of dimension M×N×2, where M = N tx ×N rx and N = N sb is obtained. If C complex is further compressed in step 4, the downlink channel matrix (complex matrix) of dimension N tx ×N rx ×N RB is converted into a coefficient matrix (real matrix) of dimension M×2 or M×N×2. The range of values of M and N and the relationship with U and / or V remain unchanged as the relative relationship explained in step 4 of the second method. For absolute values, U is replaced by N tx ×N rx , V is replaced by N sb , or U is replaced by N tx ×N rx ×x, and V is replaced by N sb ×y.

[0241] In the fourth method, the channel dimension N rx is reserved, and joint projection is performed in N tx . In this way, the original channel information (without SVD compression) can be fed back using CSI, and the 2D convolutional network can be used in cooperation to reduce the complexity of the operation.

[0242] 5. The fifth method, also called the full-channel feedback + 3-dimensional (dimension, D) convolutional network method In the fifth method, CSI feedback is performed based on the original downlink channel, and DFT projection is performed separately in dimension N tx and N rx .

[0243] Step 1: Determine the physical meaning of H.

[0244] Optionally, the downlink channel matrix H is obtained using step S504, and the dimension is N tx ×Nrx ×N RB Unlike the eigen subspace matrix H of the first method, the second method, or the third method, it can be found that H in this specification represents the downlink channel instead of the eigen subspace (i.e., SVD is not performed on the downlink channel matrix).

[0245] Step 2: Determine the spatial domain basis and the frequency domain basis.

[0246] Optionally, two orthogonal DFT bases are generated. When the spatial domain basis is generated, A = N tx When the frequency domain basis is generated, A = N sb Therefore, the two generated bases are the spatial domain basis

Number

Number

[0247] Optionally, G and F may be the group with the most accurate projection direction selected from a plurality of groups of orthogonal oversampled spatial frequency bases.

[0248] Alternatively, optionally, two oversampled DFT bases are generated. When the spatial domain basis is generated, A = N tx and B = N tx ×x. When the frequency domain basis is generated, A = N sb and B = N sb ×y. x and y are the oversampling coefficients of the spatial domain and the frequency domain respectively, and both x and y are greater than 1, or x is equal to 1 and y is greater than 1, or x is greater than 1 and y is equal to 1, and x and y may or may not be equal.

Number

Number

[0249] Step 3: Determine the sparse representation C of H complex to be determined.

[0250] The UE performs a spatial-frequency joint projection on H. When G and F are orthogonal bases, for the spatial-frequency joint projection method, refer to the following equation:

Number

[0251] When G and F are oversampled DFT bases, G in Equation 4 H is

Number

Number

Number

[0252] Step 4: Determine whether to further compress C complex to be determined.

[0253] The difference from the above-mentioned multiple methods is that C complex is a three-dimensional matrix.

[0254] Optionally, C complex is not further compressed. In this case, Step 4 may not be executed, and Step 5 is executed after Step 3 is completed.​

[0255] Optionally, C complex is further compressed.

[0256] For example, the compression method is Method 1. Method 1 is to convert the complex matrix

Number

Number

Number

[0257] For example, another compression method is Method 2. Method 2 is

Number

Number

Number

Number

[0258] For example, yet another compression method is Method 3. Method 3 [Number] selects the first M rows (M ≦ U) having the maximum energy in the spatial domain dimension of [Number] and selects the first N columns (N ≦ V) having the maximum energy in the frequency domain dimension of [Number] is represented as.

[0259] Step 5: Determine the dimension of the input matrix of the encoder network.

[0260] If Step 4 is not executed, in Step 5, the complex matrix C complex can be converted into a real matrix [Number] where M = N tx , and N = N sb .

[0261] If C complex is further compressed in Step 4, the compression matrix [Number] is a real matrix [Number] may be converted or the compression matrix

Number

Number

[0262] After the preprocessing of the above five steps, if step 4 is not executed, the downlink channel matrix (complex matrix) with dimension N tx ×N rx ×N RB is converted into a coefficient matrix (real matrix) with dimension N rx ×M×N×2, where M = N tx , and N = N sb . If C complex is further compressed in step 4, the downlink channel matrix (complex matrix) with dimension N tx ×N rx ×N RB is converted into a coefficient matrix (real matrix) with dimension N rx ×M×2 or N rx ×M×N×2. The range of the values of M and N and the relationship with U and / or V remain unchanged as the relative relationship described in step 4. For the absolute values, U is replaced by N tx , V is replaced by N sb , or U is replaced by N tx ×x, and V is replaced by N sb ×y.

[0263] In the fifth method, the channel dimension N rx is reserved, but the spatial frequency projection is not performed. In this way, the original channel information (without SVD compression) can be fed back using CSI, and a 3D convolutional network can be used in cooperation to explore the coefficient correlation of dimension N rx , thereby obtaining a better compression effect.

[0264] In some of the foregoing methods, it is assumed that the encoder network is a real number network, and the input of the encoder network can have multiple possibilities such as a two-dimensional tensor, a three-dimensional tensor, and a four-dimensional tensor. The channel dimension of the tensor is usually equal to 2, that is, it represents the real part and the imaginary part. Another dimension represents the shape of the original complex matrix, such as a one-dimensional vector or a two-dimensional / three-dimensional matrix. Based on the above method, the input of the encoder network can be one of the following three cases, namely, 2×M, 2×M×N, and 2×M×N×T. Correspondingly, the output dimension of the encoder network may also be one of the three cases of D×S, D×P×Q, and D×P×Q×R, where D represents the channel dimension of the tensor in the encoder network, and P, Q, R, etc. represent the dimensions of the output matrix. Generally, the number of dimensions of the input tensor and the output tensor of the neural network is the same. For example, when the dimension of the input tensor is a two-dimensional tensor, the dimension of the output tensor is also a two-dimensional tensor. Therefore, the input / output combination corresponding one-to-one to the foregoing two-dimensional / three-dimensional / four-dimensional dimensions is a preferred combination of the encoder network. Optionally, the encoder network may also meet the requirements of the input and output dimensions according to the actual application scenario. In this case, the combination of the input and output dimensions of the encoder network may be any combination of the foregoing nine combinations (that is, when the input of the encoder network has the foregoing three cases and the output of the encoder network has the foregoing three cases, if the consistency between the input dimension and the output dimension is not considered, the three inputs and the three outputs may be randomly combined, that is, 3×3). It should be understood that the order of the first dimension, the second dimension, and the third dimension can be correspondingly changed.

[0265] Among the five methods described above, the first method is a preprocessing process based on an example that uses an eigen-subspace matrix and orthogonal DFT bases: G and F (or orthogonal bases generated by an oversampling method), and it is a relatively basic process. In the second method, based on the first method, the sparsely projected coefficient matrix C complex is further compressed. In the third method, based on the first method, the oversampled DFT bases:

Number

Number

[0266] In particular, all of the examples of the five aforementioned methods are based on the DFT basis. It should be understood that these are merely examples and do not impose limitations. Optionally, another method may be used to generate the spatial frequency projection basis. For example, the eigenvectors obtained via SVD may be executed on the channel (covariance matrix) and used as the projection basis. This basis is usually referred to as the Eigen basis. Generally, the dimension of the Eigen basis is the same as that of the orthogonal DFT basis, i.e., A×A. However, sampling cannot be performed in the Eigen basis. In the five aforementioned methods, the orthogonal DFT basis can be replaced by the Eigen basis in any manner not related to oversampling. The dimension of the Eigen basis coincides with the dimension of the orthogonal DFT basis in the five aforementioned methods. The method of replacing the orthogonal DFT basis with the Eigen basis is also included within the scope of protection of the embodiments of this application.

[0267] S506: The UE obtains the output matrix of the encoder network based on the real matrix and the encoder network. For example, the UE inputs the real matrix into the encoder network to obtain the output matrix of the encoder network.

[0268] For example, the output matrix of the encoder network is called the first output matrix. For example, the reference network currently used by the UE and the network device is the first reference network, and the first reference network includes a first reference encoder network and a first reference decoder network. Optionally, the UE may perform encoding using the first reference encoder network. Alternatively, optionally, the protocol may not limit whether the encoder network actually used by the UE is the reference encoder network, and may limit that the evaluated performance of the actually used encoder network and the corresponding reference decoder network needs to meet the indicators specified by the protocol. Therefore, the characteristics of the encoder network currently used by the UE can be determined based on the characteristics of the first reference encoder network. For example, the input dimension of the encoder network currently used by the UE is determined based on the input dimension of the first reference encoder network. For example, the input dimension of the encoder network currently used by the UE is equal to the input dimension of the first reference encoder network. Similarly, for example, the output dimension of the encoder network currently used by the UE is also determined based on the output dimension of the first reference encoder network. For example, the output dimension of the encoder network currently used by the UE is equal to the output dimension of the first reference encoder network.

[0269] For example, the UE obtains a real matrix with dimensions M×N×2 through a preprocessing process. The input dimension of the encoder network (in other words, the dimension of the input matrix of the encoder network) is M×N×2. The real matrix is the input matrix of the encoder network. The UE uses the real matrix as the input of the encoder network and, through inference, the output la∈R of the encoder network D×SObtain it, and la represents the first output matrix. The first output matrix is a real matrix with dimensions D×P×Q. It will be understood that the output dimension of the encoder network is D×P×Q and S = P×Q. The encoder network can output S vectors, each of the S vectors represents one pixel, the dimension of each of the S vectors is D, and D may be regarded as representing the number of channels of the tensor of the output layer of a neural network (e.g., the encoder network). M≦N tx If it is the case that sb , and S≦M×N. For example, generally, S≪M×N. In other words, the original information (downlink channel matrix) obtained by the UE through measurement may be compressed to a lower dimension by using the encoder network, and ≪ indicates "much smaller than". Alternatively, M = N tx ×x, and N = N sb ×y, then S≦(N tx ×x)×(N sb ×y). Alternatively, M = N tx ×N rx , and N = N sb , then S≦(N tx ×N rx )×N sb .

[0270] In another example, the UE obtains a real matrix with dimensions M×2 through a preprocessing process. The input dimension of the encoder network (in other words, the dimension of the input matrix of the encoder network) is M×2. The real matrix is the input matrix of the encoder network. The UE uses the real matrix as the input of the encoder network and obtains the output la∈R D×S of the encoder network by inference, and la represents the first output matrix. The first output matrix is a real matrix with dimensions D×S. In this specification, S≦N tx ×N sb ×z, where z is a positive integer greater than or equal to 1. For example, in the aforementioned second method, z = 1 is used as an example. In addition, S≦M.

[0271] In yet another example, the UE obtains a real matrix with dimensions M×N×T×2 via a preprocessing process. The input dimension of the encoder network (in other words, the dimension of the input matrix of the encoder network) is M×N×T×2. The real matrix is the input matrix of the encoder network. The UE uses the real matrix as the input to the encoder network and, through inference, obtains the output la∈R of the encoder network D×S where la represents a first output matrix. The first output matrix is a real matrix with dimensions D×P×Q×R. For example, S = P×Q×R. M = N rx , N = N tx , and T = N sb In the case where, then S ≤ N rx ×N tx ×N sb .

[0272] It can be found that the dimensions of the matrix can change after the encoder network

[0273] S507: The UE determines the CSI

[0274] In this step, the UE uses the first vector quantization dictionary. For the method of determining the first vector quantization dictionary, refer to step S501. For example, the first vector quantization dictionary includes N1 vectors. The lengths of the N1 vectors may be the same. The length may be equal to at least one dimension in the output dimension of the encoder network. Therefore, the first vector quantization dictionary can be used by the receiving end of the CSI to restore (or reconstruct) the downlink channel matrix. For example, the UE can perform quantization processing on the first output matrix based on the first vector quantization dictionary to obtain the CSI. As a result, the receiving end of the CSI can restore the downlink channel matrix based on the CSI. The UE can directly perform quantization processing on the first output matrix based on the first vector quantization dictionary. Alternatively, the UE can perform corresponding processing on the first output matrix and then perform quantization processing on the processed first output matrix based on the first vector quantization dictionary. In this embodiment of the present application, it is used as an example that the UE directly performs quantization processing on the first output matrix based on the first vector quantization dictionary.

[0275] Performing quantization processing on the first output matrix based on the first vector quantization dictionary corresponds to replacing (or representing) the channel features included in the first output matrix using some or all of the N1 vectors included in the first vector quantization dictionary. Therefore, from this perspective, each of the N1 vectors included in the first vector quantization dictionary may be regarded as representing one channel feature or one type of channel information. The channel features included in the matrix output by the encoder network can also be understood as channel features represented by many vectors. Therefore, the vectors in the first vector quantization dictionary can be used to approximately represent the vectors in the first matrix. In other words, the first output matrix is quantized based on the first vector quantization dictionary.

[0276] The dimension of the first vector quantization dictionary is, for example, D×E. N1 vectors are, for example, all the vectors included in the first vector quantization dictionary. In this case, N1 is equal to E. Alternatively, the N1 vectors are, for example, some of the vectors included in the first vector quantization dictionary. In this case, N1 is less than E. In this specification, E is a positive integer. The output dimension of the encoder network is la∈R D×S is. It can be found that one dimension of the first vector quantization dictionary (for example, the dimension is also the length of the N1 vectors) and one dimension of the output dimension of the encoder network are both D. Here, an example where the length of the N1 vectors included in the first vector quantization dictionary is equal to one dimension of the output dimension of the encoder network is used. In some other implementations, the length of the N1 vectors included in the first vector quantization dictionary may also be equal to a plurality of dimensions in the output dimension of the encoder network.

[0277] The specific implementation for performing quantization processing on the first output matrix based on the first vector quantization dictionary is not limited in this embodiment of the present application. For example,

Number

Number

[0278]

Number

Number

[0279] Dic j,d represents the d-th element of the j-th vector in the first vector quantization dictionary, and la i,d is la iIt represents the d-th element of the vector. By using Equation 6, the continuous floating-point channel representation can be approximately represented by using a vector represented by discrete indices. The error generated by the processing of this step may be regarded as quantization error.

[0280] An index matrix with a dimension of 1*S can be used as PMI. Alternatively, the UE may further convert the index matrix into a binary bit stream. The binary bit stream can be used as PMI. This is not limited.

[0281] S508: The UE transmits CSI to the network device, and correspondingly, the network device receives CSI from the UE.

[0282] The UE can determine CSI based on information such as CQI, RI, and the obtained PMI, and transmit the CSI to the network device. For example, if the UE receives the third configuration information in step S501 and the third configuration information indicates a CSI reporting configuration and / or a CSI resource configuration, etc., the UE can transmit the CSI to the network device based on the third configuration information. Since the determination of PMI is related to the first vector quantization dictionary, CSI is determined based on the first vector quantization dictionary. More specifically, CSI may be regarded as being determined based on the index of one or more vectors included in the first vector quantization dictionary.

[0283] S509: The network device can obtain information regarding the reconstructed downlink channel matrix based on S indices and the first vector quantization dictionary.

[0284] For example, the information regarding the reconstructed downlink channel matrix is either the reconstructed downlink channel matrix, or the information regarding the reconstructed downlink channel matrix is not the reconstructed downlink channel matrix, but the reconstructed downlink channel matrix can be obtained, or the transmission parameters of the downlink channel can be determined based on the information.

[0285] The S indices are the S indices of the S vectors of the first vector quantization dictionary included in the CSI (or PMI). For example, the reference network currently used by the UE and the network device is the first reference network, and the first reference network includes a first reference encoder network and a first reference decoder network. The protocol may not limit whether the decoder network actually used by the network device is the reference decoder network, but may limit that the evaluated performance of the actually used decoder network and the corresponding reference encoder network needs to meet the indicators specified by the protocol. Therefore, the characteristics of the decoder network currently used by the network device can be determined based on the characteristics of the first reference decoder network. For example, the input dimension of the decoder network currently used by the network device is determined based on the input dimension of the first reference decoder network. For example, the input dimension of the decoder network currently used by the network device is equal to the input dimension of the first reference decoder network. Similarly, for example, the output dimension of the decoder network currently used by the network device is also determined based on the output dimension of the first reference decoder network. For example, the output dimension of the decoder network currently used by the network device is equal to the output dimension of the first reference decoder network.

[0286] Step S509 may include a plurality of steps. An explanation is provided below.

[0287] S5091: The network device performs an inverse mapping for S indices based on the first vector quantization dictionary to obtain a first matrix. It is understood that the network device performs an inverse mapping for the PMI included in the CSI based on the first vector quantization dictionary to obtain the first matrix.

[0288] When the PMI is an index matrix, the network device may find S column vectors corresponding to the S indices included in the index matrix from the first vector quantization dictionary and obtain a matrix including the S column vectors. The matrix is, for example, called the first matrix. The first matrix is a real matrix. The dimension is D×S. When the PMI is a binary bit stream, the network device first parses the binary bit stream to obtain an index matrix, and then obtains the first matrix based on the first vector quantization dictionary.

[0289] S5092: The network device obtains information about the reconstructed downlink channel matrix based on the first matrix and the decoder network.

[0290] For example, the network device may input the first matrix into the decoder network. The matrix output by the decoder network is called the second matrix. Then, the network device obtains information about the reconstructed downlink channel matrix based on the second matrix.

[0291] For example, the dimension of the first matrix is D×S. The network device inputs the first matrix into the decoder network, and the decoder network may obtain the second matrix by inference. When the input dimension of the encoder network is M×N×2, the second matrix is

Number

[0292] Based on different output dimensions of the decoder network, the way for the network device to obtain information about the downlink channel matrix reconstructed based on the second matrix may also be different. Hereinafter, examples are used for illustration.

[0293] 1. The output dimension of the decoder network is M×N×2. In this specification, it is assumed that the dimension of the second matrix is M×N×2. When M×N of the dimension of the second matrix is the same as the dimension of the downlink channel feature matrix H' to be reconstructed, the corresponding complex matrix [Mathematics] can be obtained based on the second matrix by directly referring to the following method of converting a real matrix into a complex matrix. When the real matrix is a two-dimensional matrix, for example, C real ∈R M×2 when it is, the complex matrix corresponding to the real matrix is usually, for example, [Mathematics] It is a one-dimensional vector represented as. Alternatively, when the real matrix is a 3D matrix, for example, C real ∈R M×N×2 in the case of, the complex matrix corresponding to the real matrix is usually, for example

Number

Number

[0294] The method of converting a real matrix of higher dimension into a complex matrix can be obtained by analogy.

[0295] Alternatively, when the dimension M×N of the second matrix is different from the dimension of the downlink channel characteristic matrix H' to be reconstructed (for example, in step S505, further compression is performed with reference to method 3 of step 4 of the second method), optionally, the network device converts the second matrix into a complex matrix, and based on the M×N position information obtained from the CSI, fills one by one the elements of the complex matrix of the all-zero matrix having the same dimension as the downlink channel characteristic matrix H' according to the corresponding positions of the elements. This is

Number

[0296] The network device performs an inverse transformation on the complex matrix

Number

[0297] To distinguish from the aforementioned eigen-subspace matrix, the aforementioned eigen-subspace matrix (for example, the eigen-subspace matrix H in step S505) may be referred to as the first eigen-subspace matrix, and the eigen-subspace matrix H' in this specification may be referred to as the second eigen-subspace matrix. For example, the network device may use the following method to obtain the second eigen-subspace matrix for the complex matrix

Number

Number

[0298] When the output dimension of the decoder network is M×N×2, step S509 may be expressed using a uniform reconstruction formula. On the condition that the information about the reconstructed downlink channel matrix satisfies the first relationship, the network device may be regarded as being able to obtain the information about the reconstructed downlink channel matrix in any way. The steps (for example, steps S5091 and S5092) executed by the network device in step S509 are merely examples and are not intended to limit the behavior of the network device. The first relationship includes the following:

Number

[0299] In Equation 8, {ind i} i=1…S represents the S indices of the S vectors included in the first vector quantization dictionary, that is, the S indices of the S vectors corresponding to the PMI. The function q(x) indicates mapping the indices of the S vectors to obtain a first matrix with dimensions D×S based on the first vector quantization dictionary. The function f dec(x) indicates reconstructing the first matrix using a decoder network to obtain a second matrix with dimensions M×N×2. Both the first matrix and the second matrix are real matrices. The function C(x) converts the second matrix into a complex matrix whose dimensions are M×N (for example, the complex matrix described in step S509).

Number

Number

Number

[0300] Note that in all equations in the embodiments of this application, the count always starts from 1 and ends with a specific character. For example, in Equation 8, k is an integer from 1 to M, that is, k = 1, 2,..., M. Alternatively, in some or all of the equations in the embodiments of this application, the count may start from 0 and end with (character - 1). For example, k = 1, 2,..., M in Equation 8 may be replaced by k = 0, 1,...,(M - 1).

[0301] {U 1,k}} k=1…M represents a set of spatial domain basis vectors. The set of spatial domain basis vectors is, for example, the spatial domain basis used when the UE performs the spatial frequency joint projection described above. U 1,k represents the k-th vector in the set of spatial domain basis vectors. The length of the vectors included in the set of spatial domain basis vectors is N tx (the physical meaning of the corresponding downlink channel matrix H is the eigen subspace matrix obtained through SVD) or N tx ×N rx(The physical meaning of the corresponding downlink channel matrix H is the original channel). {U 2,l} l=1…N represents a frequency-domain basis vector set. The frequency-domain basis vector set is, for example, the above-mentioned frequency-domain basis used when the UE performs spatial-frequency joint projection. For example, the network device may use the DFT formula to obtain the spatial-domain basis vector set and the frequency-domain basis vector set. U 2,l represents the l-th vector in the frequency-domain basis vector set. The length of the vector included in the frequency-domain basis vector set is N sb . When the frequency-domain granularity is one RB, N sb = N rb .

[0302] The values of M and N depend on the preprocessing method in step S505. In the embodiments of the present application, the relationship satisfied by the reconstructed downlink channel matrix (for example, Equation 8) restricts the preprocessing method performed by the UE. For example, when the ranges of the values of M and N are determined, the relationship satisfied by the reconstructed downlink channel matrix may be regarded as known to the UE, and the UE can determine how to perform the preprocessing. Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are orthogonal DFT basis vector sets, and C complex is not further compressed. In this case, the values of M and N are respectively equal to the lengths of the vectors included in the corresponding basis vector sets. Specifically, M = N tx (The physical meaning corresponding to the downlink channel matrix H is the eigen-subspace matrix obtained through SVD) or M = N tx ×N rx (The physical meaning corresponding to the downlink channel matrix H is the original channel), and N = N sb .

[0303] Optionally, {U 1,k} k=1…M and {U2,l} l=1…N is an orthogonal DFT basis vector set, and C complex is compressed with reference to Method 3 of Step 4 (of the second method). In this case, the values of M and N are each smaller than the length of the vectors included in the corresponding basis vector set. Specifically, M < N tx (The physical meaning corresponding to the downlink channel matrix H is the eigen subspace matrix obtained via SVD) or M < N tx × N rx (The physical meaning corresponding to the downlink channel matrix H is the original channel), and N < N sb is true.

[0304] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are oversampled DFT basis vector sets, and C complex is not further compressed. In this case, the values of M and N are each equal to the product of the length of the vectors included in the corresponding basis vector set and the oversampling coefficient. Specifically, M = N tx × x (The physical meaning corresponding to the downlink channel matrix H is the eigen subspace matrix obtained via SVD) or M = N tx × N rx × x (The physical meaning corresponding to the downlink channel matrix H is the original channel), and N = N sb × y. In this specification, x and y are the spatial frequency oversampling coefficients, respectively.

[0305] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are oversampled DFT basis vector sets, and C complex is compressed with reference to Method 3 of Step 4 (of the second method). In this case, the values of M and N are each smaller than the product of the length of the vectors included in the corresponding basis vector set and the oversampling coefficient. Specifically, M < N tx×x (The physical meaning corresponding to the downlink channel matrix H is the eigen-subspace matrix obtained via SVD) or M < N tx ×N rx ×x (The physical meaning corresponding to the downlink channel matrix H is the original channel), and N < N sb ×y. In particular, to achieve the purpose of information compression, generally, in this case, M < N tx or M < N tx ×N rx and N < N sb hold.

[0306] The reconstructed channels obtained using Equations 7 and 8 are equivalent. Equation 7 represents the inverse transform of the spatial frequency projection. The dimensions of S, F, and

Number

Number

[0307] Equation 8 restores the channel information included in PMI to several D-dimensional vectors using the function q(x), and the function f decUnderstand that a general process is represented by using (x) to transform these D-dimensional vectors into M×N×2 real matrices, calculating the complex matrix form of the real matrices, obtaining M×N weighted complex coefficients, multiplying the M spatial domain basis vectors and the N frequency domain basis vectors separately by the weighted complex coefficients, and finally performing addition to obtain the reconstructed downlink channel characteristic matrix H'. The above M×N values are used as examples, not limitations. In other words, the specific values of M×N are different from the above examples due to different preprocessing methods. On the condition that the end-to-end reconstruction method satisfies the first relationship described in Equation 8, this shall fall within the protection scope of the embodiments of the present application.

[0308] Equation 8 can further impose restrictions on the preprocessing procedure in step S505. C complex When C is compressed, the selected coefficients are 2D matrices that separately correspond to the spatial domain dimension and the frequency domain dimension (similar to Method 3).

[0309] 2. The output dimension of the decoder network is M×2. Assume that the dimension of the second matrix is M×2. When M of the dimension of the second matrix is the same as the dimension of the downlink channel characteristic matrix H' to be reconstructed, the second matrix is the corresponding complex matrix

Number

[0310] Alternatively, when the dimension M of the second matrix is different from the dimension of the downlink channel feature matrix H' to be reconstructed (for example, the UE performs further compression with reference to Method 1 or Method 2 of Step 4 of the second method in Step S505), the network device obtains information regarding the downlink channel matrix reconstructed based on the second matrix. In one implementation, the network device converts the second matrix into a complex matrix whose dimension is M, and then, based on the positions of M complex coefficients in the original N tx ×N sb 2D plane, satisfies the M complex coefficients at the corresponding positions. The other positions in the original N tx ×N sb 2D plane are set to 0. In this way, the complex matrix

Number

[0311] The network device performs an inverse transformation on the complex matrix

Number

Number

[0312] When the output dimension of the decoder network is M×2, step S509 may be expressed using a uniform reconstruction formula. Under the condition that the information regarding the reconstructed downlink channel matrix satisfies a second relationship, the network device may be regarded as being able to obtain the information regarding the reconstructed downlink channel matrix in any manner. The steps (for example, steps S5091 and S5092) executed by the network device in step S509 are merely examples and are not intended to limit the behavior of the network device. The second relationship includes the following:

Number

[0313] In Equation 9, {ind i} i=1…S represents the S indices of the S vectors included in the first vector quantization dictionary, that is, the S indices of the S vectors corresponding to the PMI. The function q(x) indicates mapping the indices of the S vectors to obtain a first matrix with dimensions D×S based on the first vector quantization dictionary. The function f dec (x) indicates reconstructing the first matrix using the decoder network to obtain a second matrix with dimensions M×2. Both the first matrix and the second matrix are real matrices. The function C(x) indicates converting the second matrix into a complex matrix (for example, the complex matrix described in step S509

Number

Number

Number

[0314] {U 1,(j,k)} j=1…M represents a spatial domain basis vector set. The spatial domain basis vector set is, for example, the spatial domain basis used when the UE performs the above-described spatial frequency joint projection. The length of the vectors included in the spatial domain basis vector set is N tx (the physical meaning of the corresponding downlink channel matrix H is the eigen-subspace matrix obtained via SVD) or N tx ×N rx (the physical meaning of the corresponding downlink channel matrix H is the original channel).

Number

[0315] The value of M depends on the preprocessing method in step S505. Optionally, {U 1,(j,k)} j=1…M and [Number] is an orthogonal DFT basis set, and C complex is not further compressed. In this case, the value of M is the dimension product of the C complex matrix. In this specification, M = N tx × N sb (The physical meaning corresponding to the downlink channel matrix H is the eigen subspace matrix obtained through SVD) or M = N tx × N rx × N sb (The physical meaning corresponding to the downlink channel matrix H is the original channel).

[0316] Optionally, {U 1,(j,k)} j=1…M and [Number] is an orthogonal DFT basis set, and C complex is compressed with reference to Method 1 or 2 in Step 4 of (the second method). In this case, the value of M is the Ccomplex It is less than the dimensional product of the matrix. In this specification, M < N tx × N sb (The physical meaning corresponding to the downlink channel matrix H is the eigen-subspace matrix obtained through SVD) or M < N tx × N rx × N sb (The physical meaning corresponding to the downlink channel matrix H is the original channel).

[0317] Optionally, {U 1,(j,k)} j=1…M and

Number

[0318] Optionally, {U 1,(j,k)} j=1…M and

Number

[0319] The reconstructed channels obtained using Equation 7 and Equation 9 are equivalent. Equation 7 represents the inverse transform of the spatial frequency projection. S, F, and

Number

Number

[0320] Equation 9 restores the channel information contained in PMI to several D-dimensional vectors using the function q(x), and the function f dec (x) is used to convert these D-dimensional vectors into M × 2 real matrices, the complex matrix form of the real matrices is calculated, M weight complex coefficients are obtained, the weight complex coefficients are multiplied separately by M groups of spatial frequency basis vectors, and finally addition is performed to obtain the reconstructed downlink channel characteristic matrix H'. It should be understood that the above value of M is used as an example without limitation. In other words, the specific value of M is different from the above example due to different preprocessing methods. Provided that the end-to-end reconstruction method satisfies the second relationship described in Equation 9, this shall fall within the protection scope of the embodiments of the present application.

[0321] Generally, H is a two-dimensional matrix plane (spatial domain × frequency domain). Equation 9 may further impose the following restrictions on the preprocessing procedure of step S505, and the selected coefficients need to be arranged according to a one-dimensional vector. For example, H is converted into a one-dimensional vector (without compression), or the candidate coefficient matrix is converted into a one-dimensional vector, and then M coefficients with the maximum energy are selected (similar to compression method 1 or 2).

[0322] 3. The output dimension of the decoder network is M×N×T×2. In this specification, it is assumed that the dimension of the second matrix is M×N×T×2. In this case, it is generally the fifth method, and T = N rx is.

[0323] When the dimension M×N×T of the second matrix is the same as the dimension of the downlink channel feature matrix H' to be reconstructed, the second matrix is directly converted into a complex matrix

Number

[0324] Alternatively, when the dimension M×N×T of the second matrix is different from the dimension of the downlink channel feature matrix H' to be reconstructed (for example, in step S505, further compression is performed with reference to method 3 of step 4 of the second method), optionally, the network device converts the second matrix into a complex matrix and fills the elements of the complex matrix of the all-zero matrix having the same dimension as the downlink channel feature matrix H' one by one based on the M×N position information obtained from the CSI

Number

[0325] Alternatively, if the dimensions M×N×T of the second matrix are different from those of the downlink channel feature matrix H' to be reconstructed (for example, in step S505, further compression is performed with reference to method 1 or 2 in step 4 of the second method), optionally, the network device converts the second matrix into a complex matrix and fills the elements of the complex sequence of the all-zero matrix having the same dimensions as the downlink channel feature matrix H' one by one based on the M position information obtained from the CSI,

Number

[0326] The network device performs an inverse transformation on the complex matrix

Number

Number

Number

[0327] Comparing with Equation 7, in the relationship expressed in Equation 10, the dimension T of M×N×T is not involved in the inverse transformation. This is equivalent to T times the inverse transformation performed in the M×N dual-region plane.

[0328] When the output dimension of the decoder network is M×N×T, step S509 may be expressed using a uniform reconstruction formula. Under the condition that the information about the reconstructed downlink channel matrix satisfies a third relationship, the network device may be regarded as being able to obtain the information about the reconstructed downlink channel matrix in any manner. The steps (for example, steps S5091 and S5092) executed by the network device in step S509 are merely examples and are not intended to limit the behavior of the network device. The third relationship includes the following:

Number

[0329] In Equation 11, q(ind i ) represents the S indices of the S vectors included in the first vector quantization dictionary, that is, the S indices of the S vectors corresponding to the PMI. The function q(x) indicates mapping the indices of the S vectors to a first matrix with dimensions D×S based on the first vector quantization dictionary. The function f dec (x) indicates reconstructing the first matrix using the decoder network to obtain a second matrix with dimensions M×N×T×2. In this specification, T = N rx . Both the first matrix and the second matrix are real matrices. The function C(x) indicates converting the second matrix into a complex matrix with dimensions M×N×T (for example, the complex matrix described in step S509

Number

Number

Number

[0330] {U 1,k} k=1…M represents a set of spatial domain basis vectors. The set of spatial domain basis vectors is, for example, the spatial domain basis used when the UE performs the spatial frequency joint projection described above. The length of the vectors included in the set of spatial domain basis vectors is N tx . {U 2,l} l=1…N represents a set of frequency domain basis vectors. The set of frequency domain basis vectors is, for example, the frequency domain basis described above used when the UE performs the spatial frequency joint projection. For example, the network device may use the DFT formula to obtain the set of spatial domain basis vectors and the set of frequency domain basis vectors. The length of the vectors included in the set of frequency domain basis vectors is N sb . When the frequency domain granularity is one RB, N sb = N rb .

Number

[0331] The values of M and N depend on the preprocessing method in step S505. Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are sets of orthogonal DFT basis vectors, and C complex is not further compressed. In this case, the values of M and N are respectively equal to the lengths of the vectors included in the corresponding basis vector sets. In this specification, M = N tx , and N = N sb .

[0332] Optionally, {U1,k} k=1…M and {U 2,l} l=1…N is an orthogonal DFT basis vector set, and C complex is compressed with reference to Method 3 of Step 4 (of the second method). In this case, the values of M and N are each smaller than the length of the vectors included in the corresponding basis vector set. In this specification, M < N tx , and N < N sb is.

[0333] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N is an oversampled DFT basis vector set, and C complex is not further compressed. In this case, the values of M and N are each equal to the product of the length of the vectors included in the corresponding basis vector set and the oversampling coefficient. In this specification, M = N tx ×x, and N = N sb ×y, where x and y are the spatial frequency oversampling coefficients respectively. For the range of values of x and y, refer to the foregoing content.

[0334] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N is an oversampled DFT basis vector set, and C complex is compressed with reference to Method 3 of Step 4 (of the second method). In this case, the values of M and N are each smaller than the product of the length of the vectors included in the corresponding basis vector set and the oversampling coefficient. In this specification, M < N tx ×x, N < N sb ×y. In particular, in order to achieve the purpose of information compression, in this case, generally, M < N tx and N < N sb is.

[0335] The reconstructed channels obtained using Equation 10 and Equation 11 are equivalent. Equation 10 represents the inverse transform of the spatial frequency projection. S, F, and

Number

Number

[0336] Equation 11 restores the channel information included in PMI to several D-dimensional vectors using the function q(x), and the function f dec (x) is used to convert these D-dimensional vectors into M×N×T×2 real matrices, the complex matrix form of the real matrices is calculated, M×N×T weighted complex coefficients are obtained, M×N weighted complex coefficients are taken element by element based on the dimension T, M spatial domain basis vectors and N frequency domain basis vectors are separately multiplied by the weighted complex coefficients, and finally addition is performed to obtain the reconstructed downlink channel feature matrix H'. It should be understood that the above M×N values are used as examples without limitation. In other words, the specific values of M×N are different from the above examples due to different preprocessing methods. Under the condition that the end-to-end reconstruction method satisfies the third relationship described in Equation 11, this shall fall within the protection scope of the embodiments of the present application.

[0337] In particular, Equation 11 further imposes the following restrictions on the preprocessing procedure of step S505. When C complex is compressed, the selected coefficients need to be 2D matrices corresponding separately to the spatial domain dimension and the frequency domain dimension (similar to Method 3).

[0338] 4. The output dimension of the decoder network is M×T×2. In this specification, it is assumed that the dimension of the second matrix is M×T×2. In this case, it is generally the fifth method, and T = N rx is the case.

[0339] For example, if the dimension M×T×2 of the second matrix is different from the dimension of the downlink channel feature matrix H' to be reconstructed (for example, in step S505, further compression is performed with reference to method 1 or method 2 of step 4 of the second method), optionally, the network device converts the second matrix into a complex matrix whose dimension is T×M. The T×M complex matrices include T complex vectors whose dimension is M. For each complex vector whose dimension is M, the network device satisfies the M complex coefficients at the corresponding positions based on the positions of the M complex coefficients in the original two-dimensional spatial frequency plane reported using the CSI. The other positions are set to 0. In this way, the complex matrix

Number

Number

[0340] When the output dimension of the decoder network is M×T×2, step S509 may be expressed using a uniform reconstruction formula. Under the condition that the information about the reconstructed downlink channel matrix satisfies the fourth relationship, the network device may be regarded as being able to obtain the information about the reconstructed downlink channel matrix in any manner. The steps (for example, steps S5091 and S5092) executed by the network device in step S509 are merely examples and are not intended to limit the behavior of the network device. The fourth relationship includes the following:

Number

[0341] In Equation 12, q(ind i ) represents the S indices of the S vectors included in the first vector quantization dictionary, that is, the S indices of the S vectors corresponding to the PMI. The function q(x) indicates mapping the indices of the S vectors to the first matrix with dimensions D×S based on the first vector quantization dictionary. The function f dec (x) indicates reconstructing the first matrix using a decoder network to obtain the second matrix with dimensions M×T×2. In this specification, T = N rx . Both the first matrix and the second matrix are real matrices. The function C(x) indicates converting the second matrix into a complex matrix with dimensions M×T (for example, the complex matrix described in step S509

Number

Number

Number

[0342] {U 1,(j,k)}} j=1…M represents a set of spatial domain basis vectors. The set of spatial domain basis vectors is, for example, the spatial domain basis used when the UE performs the spatial frequency joint projection described above. The length of the vectors included in the set of spatial domain basis vectors is Ntx is as follows.

Number

Number

[0343] The value of M depends on the preprocessing method in step S505. Optionally, {U 1,(j,k)} j=1…M and

Number

[0344] Optionally, {U 1,(j,k)} j=1…M and

Number

[0345] Optionally, {U 1,(j,k)} j=1…M and

Number

[0346] Optionally, {U 1,(j,k)} j=1…M and

Number

[0347] The reconstructed channels obtained using Equation 10 and Equation 12 are equivalent. Equation 10 shows the inverse transform of the spatial frequency projection. S, F, and

Number

Number

[0348] Equation 12 restores the channel information contained in the PMI to several D-dimensional vectors using the function q(x), and the function f dec (x) is used to convert these D-dimensional vectors into M×T×2 real matrices, calculate the complex matrix form of the real matrices, obtain M×T weighted complex coefficients, take M weighted complex coefficients for each element based on the dimension T, multiply the M groups of spatial frequency basis vectors by the weighted complex coefficients separately, and finally perform addition to obtain the reconstructed downlink channel feature matrix H'. It should be understood that this represents a general process. The above value of M is used as an example rather than a limitation. In other words, the specific value of M is different from the above example due to different preprocessing methods. As long as the end-to-end reconstruction method satisfies the fourth relationship described in Equation 12, this shall fall within the protection scope of the embodiments of the present application.

[0349] Generally, H is a two-dimensional matrix plane (spatial domain × frequency domain). Equation 12 may further impose the following restrictions on the preprocessing procedure in step S505, and the selected coefficients need to be arranged according to a one-dimensional vector. For example, H is converted into a one-dimensional vector (without compression), or the candidate coefficient matrix is converted into a one-dimensional vector, and then M coefficients with the maximum energy are selected (similar to compression method 1 or 2).

[0350] In this embodiment of the present application, steps S501 to S506 are optional steps. In addition, steps S5091 and S5092 included in step S509 are also optional steps.

[0351] In addition, it should be noted that this embodiment of the present application may be applied to the network architecture and network devices of the UE, or may be applied to the network architecture, network devices, and AI modules of the UE. The architecture may include two AI modules. One AI module may be disposed inside the network device and may implement functions such as real-time inference. The other AI module may be disposed outside the network device. The AI module may implement model training, non-real-time function inference, etc. When this embodiment of the present application is applied to the network architecture, network devices, and AI modules of the UE, for example, the decoder network may be trained offline, the NN-CSI network may be prepared by the network device in step S502, and the first matrix may be input to the decoder network in step S509 to obtain the second matrix and other processes may be executed by the AI module.

[0352] In this embodiment of the present application, the CSI can be determined based on the first vector quantization dictionary. For example, the PMI may be represented using the index of one or more vectors included in the first vector quantization dictionary. In this case, the information obtained by measurement may be quantized using the index included in the first vector quantization dictionary. The UE may transmit the quantized information to reduce signaling overhead. For example, since the dimension of the quantization vector is usually relatively large, it is equivalent to information or dimension expansion of maintenance at a relatively high dimension as much as possible, thereby improving the quantization accuracy. Additionally, optionally, in cooperation with the encoder network and the decoder network, the original channel matrix may be compressed. Specifically, instead of extracting some coefficients from the original channel matrix for feedback, the original channel matrix may be compressed as a whole using the network. In this case, the information to be discarded is reduced, thereby reducing the compression loss. Additionally, in this embodiment of the present application, a VQ-AE network is used. In the VQ-AE network, the quantization network and the compression network are jointly designed so that joint optimization can be implemented between compression and quantization, reducing the performance loss and improving the overall accuracy of the feedback channel state information. Additionally, in this embodiment of the present application, to adjust the complexity of CSI feedback, the input and output dimensions of the neural network (e.g., the input and output dimensions of the encoder and / or the input and output dimensions of the decoder) may be restricted, thereby emphasizing the advantage of using the neural network to compress the downlink channel matrix.

[0353] In the embodiment shown in FIG. 5, a conversion between a complex matrix and a real matrix is performed. In this case, the UE needs to convert the complex matrix into a real matrix, and the network device needs to convert the real matrix into a complex matrix. Hereinafter, a second communication method provided in an embodiment of the present application will be described. According to this method, the matrix conversion process can be reduced, thereby simplifying the processing complexity. FIG. 6 is a flowchart of this method.

[0354] S601: The network device transmits configuration information to the UE, and correspondingly, the UE receives the configuration information from the network device. For example, the configuration information is called configuration information 1 or the third configuration information, and the third configuration information can be used to configure CSI measurement.

[0355] For further details of step S601, please refer to step S501 of the embodiment shown in FIG. 5.

[0356] S602: The UE prepares the NN-CSI network, and the network device prepares the NN-CSI network.

[0357] For further details of step S602, please refer to step S502 of the embodiment shown in FIG. 5.

[0358] S603: The network device transmits a reference signal to the UE, and correspondingly, the UE receives the reference signal from the network device.

[0359] The network device may transmit a reference signal to the UE based on the third configuration information.

[0360] S604: The UE acquires the downlink channel matrix.

[0361] For further details of step S604, please refer to step S504 of the embodiment shown in FIG. 5.

[0362] S605: The UE preprocesses the downlink channel matrix.

[0363] For example, for the method by which the UE acquires the complex matrix C complex please refer to any one of the first to fifth methods described in step S505 of the embodiment shown in FIG. 5. In step S505 of the embodiment shown in FIG. 5, in the first to fifth methods, the complex matrix C complexneeds to be converted into a real matrix. However, in this embodiment of the present application, regardless of the specific method used by the UE to obtain the complex matrix C complex , the UE does not need to convert the complex matrix C complex into a real matrix. In other words, the preprocessing result in this embodiment of the present application is that the complex matrix C complex is obtained. The complex matrix C complex can be obtained by any one of the five methods in step S505 in the embodiment shown in FIG. 5.

[0364] S606: The UE obtains the output matrix of the encoder network based on the complex matrix and the encoder network. For example, the UE inputs the complex matrix into the encoder network to obtain the output matrix of the encoder network. For example, the output matrix of the encoder network is called the second output matrix. For example, the reference network currently used by the UE and the network device is the first reference network, and the first reference network includes a first reference encoder network and a first reference decoder network. The protocol cannot limit whether the actually used encoder network is the reference encoder network, but can limit that the evaluated performance of the actually used encoder network and the corresponding reference decoder network needs to meet the indicators specified by the protocol. Therefore, the characteristics of the encoder network currently used by the UE can be determined based on the characteristics of the first reference encoder network. For example, the input dimension of the encoder network currently used by the UE is determined based on the input dimension of the first reference encoder network. For example, the input dimension of the encoder network currently used by the UE is equal to the input dimension of the first reference encoder network. Similarly, for example, the output dimension of the encoder network currently used by the UE is also determined based on the output dimension of the first reference encoder network. For example, the output dimension of the encoder network currently used by the UE is equal to the output dimension of the first reference encoder network.

[0365] For example, the UE obtains a complex matrix with dimensions M×N through a preprocessing process. The input dimension of the encoder network (in other words, the dimension of the input matrix of the encoder network) is M×N. The complex matrix is the input matrix of the encoder network. The UE uses the complex matrix C∈C M×N as the input to the encoder network and, through inference, obtains the output la∈R of the encoder network D×SObtain it, and la represents the first output matrix. The first output matrix is a real matrix whose dimension is D×P×Q, or the first output matrix is a complex matrix whose dimension is D×P×Q. For example, S is regarded as P×Q. Therefore, it will also be understood that the output dimension of the encoder network is D×S. In this embodiment of the present application, for example, M≦N tx , N≦N sb , and S≦M×N, or M = N tx ×x, N = N sb ×y, and S≦(N tx ×x)×(N sb ×y). In this specification, both x and y are greater than 1, or x is equal to 1 and y is greater than 1, or x is greater than 1 and y is equal to 1.

[0366] In another example, the UE obtains a complex matrix with dimension M through a preprocessing process. The input dimension of the encoder network (in other words, the dimension of the input matrix of the encoder network) is M. The complex matrix is the input matrix of the encoder network. The UE uses the complex matrix C∈C M as the input of the encoder network, and obtains the output la∈R of the encoder network by inference D×S , and la represents the first output matrix. The first output matrix is a real matrix whose dimension is D×S and contains S real vectors with length D. Alternatively, the first output matrix is a complex matrix whose dimension is D×S and contains S complex vectors with length D. It will be understood that the output dimension of the encoder network is D×S. In this embodiment of the present application, for example, M≦N tx ×N sb ×z, and S≦M.

[0367] In yet another example, the UE obtains a complex matrix with dimension M×N×T through a preprocessing process. The input dimension of the encoder network (in other words, the dimension of the input matrix of the encoder network) is M×N×T. The complex matrix is the input matrix of the encoder network. The UE uses the complex matrix C∈C M×N×Tis used as the input of the encoder network, and through inference, the output la ∈ R of the encoder network is obtained. la represents the first output matrix. The first output matrix is a real matrix with dimensions D × P × Q × R, or the first output matrix is a complex matrix with dimensions D × P × Q × R. For example, it is considered that S = P × Q × R. Therefore, it can also be understood that the output dimension of the encoder network is D × S. In this embodiment of the present application, for example, M = N D×S is obtained, and N = N rx is obtained, and T = N tx is obtained, and S ≤ N sb × N rx × N tx × N sb is satisfied.

[0368] S607: The UE determines the CSI. For details such as the description of the first vector quantization dictionary, refer to step S507 of the embodiment shown in FIG. 5.

[0369] For example, the UE may perform quantization processing on the second output matrix based on the first vector quantization dictionary to obtain the CSI. As a result, the downlink channel matrix can be restored based on the CSI. The UE may directly perform quantization processing on the second output matrix based on the first vector quantization dictionary. Alternatively, the UE may perform corresponding processing on the second output matrix and then perform quantization processing on the processed second output matrix based on the first vector quantization dictionary. In this embodiment of the present application, an example is used where the UE directly performs quantization processing on the second output matrix based on the first vector quantization dictionary.

[0370] The dimension of the first vector quantization dictionary is, for example, D × E. The output dimension of the encoder network is la ∈ R D×S is satisfied.

Number

Number

Number

[0371] Alternatively, when the first vector quantization dictionary is a complex matrix and the first output matrix is a complex matrix,

Number

Number

[0372] Dic j represents the j-th vector in the first vector quantization dictionary,

Number

[0373] By using Equation 13, the continuous floating-point channel representation can be approximately represented by using a vector represented by discrete indexes. The error generated by the processing of this step may be regarded as quantization error.

[0374] For example, when the first output matrix is a complex matrix, Equation 13 is used, or when the first output matrix is a real matrix, Equation 6 is used. Of course, this is just an example and not a specific limitation. For example, even if the first output matrix is a complex matrix, Equation 6 may be used (for example, Equation 6 may be correspondingly modified based on the complex matrix), or even if the first output matrix is a real matrix, Equation 13 may be used (for example, Equation 13 may be correspondingly modified based on the real matrix).

[0375] The index matrix can be used as PMI. Alternatively, the UE may further convert the index matrix into a binary bit stream. The binary bit stream can be used as PMI.

[0376] S608: The UE transmits CSI to the network device, and correspondingly, the network device receives CSI from the UE.

[0377] For further details of step S608, refer to step S508 of the embodiment shown in FIG. 5.

[0378] S609: The network device can obtain information regarding the reconstructed downlink channel matrix based on S indexes and the first vector quantization dictionary.

[0379] For example, the information regarding the reconstructed downlink channel matrix is either the reconstructed downlink channel matrix, or the information regarding the reconstructed downlink channel matrix is not the reconstructed downlink channel matrix, but the reconstructed downlink channel matrix can be obtained, or the transmission parameters of the downlink channel can be determined based on the information.

[0380] The S indices are the S indices of the S vectors of the first vector quantization dictionary included in the CSI (or PMI). For example, the reference network currently used by the UE and the network device is the first reference network, and the first reference network includes a first reference encoder network and a first reference decoder network. The protocol may not be able to limit whether the actually used decoder network is the reference decoder network pre-defined by the protocol, but may limit that the evaluated performance of the actually used decoder network and the corresponding reference encoder network needs to meet the indicators specified by the protocol. Therefore, the characteristics of the decoder network currently used by the network device can be determined based on the characteristics of the first reference decoder network. For example, the input dimension of the decoder network currently used by the network device is determined based on the input dimension of the first reference decoder network. For example, the input dimension of the decoder network currently used by the network device is equal to the input dimension of the first reference decoder network. Similarly, for example, the output dimension of the decoder network currently used by the network device is also determined based on the output dimension of the first reference decoder network. For example, the output dimension of the decoder network currently used by the network device is equal to the output dimension of the first reference decoder network.

[0381] Step S609 may include a plurality of steps. An explanation is provided below.

[0382] S6091: The network device performs an inverse mapping on the CSI based on the first vector quantization dictionary to obtain a first matrix. It is understood that the network device performs an inverse mapping on the PMI included in the CSI based on the first vector quantization dictionary to obtain the first matrix. The first matrix is a complex matrix.

[0383] For further details of step S6091, refer to step S5091 of the embodiment shown in FIG. 5.

[0384] S6092: The network device obtains information about the reconstructed downlink channel matrix based on the first matrix and the decoder network.

[0385] For example, the network device may input the first matrix into the decoder network. The matrix output by the decoder network is called the second matrix. Then, the network device obtains information about the reconstructed downlink channel matrix based on the second matrix.

[0386] For example, the dimension of the first matrix is D×S. The network device inputs the first matrix into the decoder network, and the decoder network may obtain the second matrix by inference. If the input dimension of the encoder network is M×N, the dimension of the second matrix is M×N. In other words, the first matrix can be reconstructed by using the decoder network to obtain a second matrix with a dimension of M×N. The second matrix is also a complex matrix. It should be understood that the input dimension of the decoder network is D×P×Q, the output dimension is M×N, and S = P×Q. For details such as the values of M, N, and S, refer to the description of step S606.

[0387] Alternatively, when the input dimension of the encoder network is M, the dimension of the second matrix is M. In other words, the first matrix can be reconstructed using the decoder network to obtain the second matrix whose dimension is M. The second matrix is also a complex matrix. It will be understood that the input dimension of the decoder network is D×S and the output dimension is M. For details such as the values of M and S, refer to the description of step S606.

[0388] Alternatively, when the input dimension of the encoder network is M×N×T, the dimension of the second matrix is M×N×T. In other words, the first matrix can be reconstructed by using the decoder network to obtain the second matrix whose dimension is M×N×T. The second matrix is also a complex matrix. It will be understood that the input dimension of the decoder network is D×P×Q×R and the output dimension is M×N×T. For details such as the values of M, N, T, and S, refer to the description of step S606.

[0389] Based on different output dimensions of the decoder network, the way for the network device to obtain information about the downlink channel matrix reconstructed based on the second matrix may also be different. Hereinafter, examples are used for illustration.

[0390] 1. The output dimension of the decoder network is M×N. In this specification, it is assumed that the dimension of the second matrix is M×N×2. When the M×N of the dimension of the second matrix is the same as the dimension of the downlink channel feature matrix H' to be reconstructed, the corresponding complex matrix

Number

Number

Number

Number

[0391] A method for converting a real matrix of higher dimension into a complex matrix can be obtained by analogy.

[0392] Alternatively, if the dimension M×N of the second matrix is different from the dimension of the downlink channel characteristic matrix H' to be reconstructed (for example, in step S505, further compression is performed with reference to method 3 of step 4 of the second method), optionally, referring to Equation 7, the network device converts the second matrix into a complex matrix and, based on the M×N position information obtained from the CSI, fills one by one the elements of the complex matrix of the all - zero matrix having the same dimension as the downlink channel characteristic matrix H' at the corresponding positions. This corresponds to

Number

[0393] The network device performs an inverse transformation on the complex matrix

Number

[0394] To distinguish from the aforementioned eigen-subspace matrix, the aforementioned eigen-subspace matrix (e.g., the eigen-subspace matrix H in step S505) may be referred to as the first eigen-subspace matrix, and the eigen-subspace matrix H' in this specification may be referred to as the second eigen-subspace matrix. When the network device is a complex matrix

Number

[0395] When the output dimension of the decoder network is M×N×2, step S509 may be expressed using a uniform reconstruction formula. On the condition that the information regarding the reconstructed downlink channel matrix satisfies the fifth relationship, the network device may be regarded as being able to obtain the information regarding the reconstructed downlink channel matrix in any manner. The steps (e.g., steps S5091 and S5092) performed by the network device in step S509 are merely examples and are not intended to limit the behavior of the network device. The fifth relationship includes the following:

Number

[0396] In Equation 14, {ind i} i=1…S represents the S indices of the S vectors included in the first vector quantization dictionary, that is, the S indices of the S vectors corresponding to the PMI. The function q(x) indicates mapping the indices of the S vectors to a first matrix with dimensions D×S based on the first vector quantization dictionary. The function f dec (x) indicates reconstructing the first matrix using the decoder network to obtain a second matrix with dimensions M×N×2. Both the first matrix and the second matrix are real matrices.

Number

[0397] {U 1,k} k=1…M represents a set of spatial domain basis vectors. The set of spatial domain basis vectors is, for example, the spatial domain basis used when the UE performs the spatial frequency joint projection described above. The length of the vectors included in the set of spatial domain basis vectors is N tx (the physical meaning of the corresponding downlink channel matrix H is the eigen subspace matrix obtained through SVD) or N tx ×N rx (the physical meaning of the corresponding downlink channel matrix H is the original channel). {U 2,l} l=1…N represents a set of frequency domain basis vectors. The set of frequency domain basis vectors is, for example, the frequency domain basis described above used when the UE performs the spatial frequency joint projection. For example, the network device may use the DFT formula to obtain the set of spatial domain basis vectors and the set of frequency domain basis vectors. The length of the vectors included in the set of frequency domain basis vectors is N sb . When the frequency domain granularity is one RB, N sb = N rb .

[0398] The values of M and N depend on the preprocessing method in step S605. For example, the specific method used by the UE to perform the preprocessing may be specified by the protocol, or the network device may send signaling to the UE to indicate the preprocessing method. Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are sets of orthogonal DFT basis vectors, and Ccomplex is not further compressed. In this case, the values of M and N are respectively equal to the lengths of the vectors included in the corresponding base vector sets. Specifically, M = N tx (The physical meaning corresponding to the downlink channel matrix H is the eigen subspace matrix obtained through SVD) or M = N tx ×N rx (The physical meaning corresponding to the downlink channel matrix H is the original channel) and N = N sb is.

[0399] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are orthogonal DFT base vector sets, and C complex is compressed with reference to Method 3 of Step 4 (in the second method). In this case, the values of M and N are respectively smaller than the lengths of the vectors included in the corresponding base vector sets. Specifically, M < N tx (The physical meaning corresponding to the downlink channel matrix H is the eigen subspace matrix obtained through SVD) or M < N tx ×N rx (The physical meaning corresponding to the downlink channel matrix H is the original channel) and N < N sb is.

[0400] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are oversampled DFT base vector sets, and C complex is not further compressed. In this case, the values of M and N are respectively equal to the products of the lengths of the vectors included in the corresponding base vector sets and the oversampling coefficient. Specifically, M = N tx ×x (The physical meaning corresponding to the downlink channel matrix H is the eigen subspace matrix obtained through SVD) or M = N tx ×N rx ×x (The physical meaning corresponding to the downlink channel matrix H is the original channel) and N = Nsb It is x × y. In this specification, x and y are respectively spatial frequency oversampling coefficients.

[0401] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are oversampled DFT basis vector sets, and C complex is compressed with reference to Method 3 of Step 4 (of the second method). In this case, the values of M and N are each smaller than the product of the length of the vectors included in the corresponding basis vector set and the oversampling coefficient. Specifically, M < N tx × x (the physical meaning corresponding to the downlink channel matrix H is the eigen-subspace matrix obtained via SVD) or M < N tx × N rx × x (the physical meaning corresponding to the downlink channel matrix H is the original channel), and N < N sb × y. In particular, in order to achieve the purpose of information compression, in this case, generally, M < N tx or M < N tx × N rx , and N < N sb .

[0402] The reconstructed channels obtained using Equation 7 and Equation 14 are equivalent. Equation 7 represents the inverse transform of spatial frequency projection. S, F, and

Number

Number

[0403] Equation 14 restores the channel information included in the PMI into several D-dimensional vectors using the function q(x), and uses the function f dec (x) to convert these D-dimensional vectors into M×N×2 real matrices, calculates the complex matrix form of the real matrices, obtains M×N weighted complex coefficients, multiplies the M spatial domain basis vectors and the N frequency domain basis vectors by the weighted complex coefficients separately, and finally performs addition to obtain the reconstructed downlink channel characteristic matrix H'. It should be understood that the above M×N values are used as examples rather than limitations. In other words, the specific values of M×N are different from the above examples due to different preprocessing methods. If the end-to-end reconstruction method satisfies the fifth relationship described in Equation 14, this shall fall within the protection scope of the embodiments of the present application.

[0404] In particular, Equation 14 further imposes the following restrictions on the preprocessing procedure of step S505, C complex When C is compressed, the selected coefficients need to be 2D matrices corresponding separately to the spatial domain dimension and the frequency domain dimension (similar to Method 3).

[0405] 2. The output dimension of the decoder network is M. Assume that the dimension of the second matrix is M×2. When M of the dimension of the second matrix is the same as the dimension of the downlink channel characteristic matrix H' to be reconstructed, the second matrix is the corresponding complex matrix

Number

[0406] Alternatively, when M of the dimension of the second matrix is different from the dimension of the downlink channel feature matrix H' to be reconstructed (for example, in step S505, further compression is performed with reference to method 1 or method 2 of step 4 of the second method), the network device obtains information about the downlink channel matrix reconstructed based on the second matrix. In one implementation form, the network device converts the second matrix into a complex matrix whose dimension is M, and then, based on the positions of the M complex coefficients in the original N tx ×N sb 2D plane, satisfies the M complex coefficients at the corresponding positions. The other positions in the original N tx ×N sb 2D plane are set to 0. In this way, the complex matrix

Number

[0407] The network device performs an inverse transformation on the complex matrix

Number

Number

[0408] When the output dimension of the decoder network is M×2, step S509 may be expressed using a uniform reconstruction formula. Under the condition that the information regarding the reconstructed downlink channel matrix satisfies the sixth relationship, the network device may be regarded as being able to obtain the information regarding the reconstructed downlink channel matrix in any manner. The steps (for example, steps S5091 and S5092) performed by the network device in step S509 are merely examples and are not intended to limit the behavior of the network device. The sixth relationship includes the following:

Number

[0409] In Equation 15, {ind i} i=1…S represents the S indices of the S vectors included in the first vector quantization dictionary, that is, the S indices of the S vectors corresponding to the PMI. The function q(x) indicates mapping the indices of the S vectors to the first matrix with dimensions D×S based on the first vector quantization dictionary. The function f dec (x) indicates reconstructing the first matrix using the decoder network to obtain the second matrix with dimensions M×2. Both the first matrix and the second matrix are real matrices.

Number

Number

[0410] {U 1,(j,k)} j=1…Mrepresents a spatial domain basis vector set. The spatial domain basis vector set is, for example, the spatial domain basis used when the UE performs the spatial frequency joint projection described above. The length of the vectors included in the spatial domain basis vector set is N tx (The physical meaning of the corresponding downlink channel matrix H is the eigen-subspace matrix obtained through SVD) or N tx ×N rx (The physical meaning of the corresponding downlink channel matrix H is the original channel). [Number] represents a frequency domain basis vector set. The frequency domain basis vector set is, for example, the above-described frequency domain basis used when the UE performs spatial frequency joint projection. For example, the network device may use the DFT formula to obtain the spatial domain basis vector set and the frequency domain basis vector set. The length of the vectors included in the frequency domain basis vector set is N sb . When the frequency domain granularity is one RB, N sb =N rb . For the description of (j,k) and (j,l), refer to the embodiment shown in FIG. 5.

[0411] The value of M depends on the preprocessing method in step S605. Optionally, {U 1,(j,k)} j=1…M and [Number] is an orthogonal DFT basis set, and C complex is not further compressed. In this case, the value of M is the dimension product of the C complex matrix. In this specification, M = N tx ×N sb (The physical meaning corresponding to the downlink channel matrix H is the eigen-subspace matrix obtained through SVD) or M = N tx ×N rx ×N sb(The physical meaning corresponding to the downlink channel matrix H is the original channel).

[0412] Optionally, {U 1,(j,k)} j=1…M and

Number

[0413] Optionally, {U 1,(j,k)} j=1…M and

Number

[0414] Optionally, {U 1,(j,k)} j=1…M and

Number

[0415] The reconstructed channels obtained using Equation 7 and Equation 15 are equivalent. Equation 7 represents the inverse transform of the spatial frequency projection. S, F, and

Number

Number

[0416] Equation 15 restores the channel information included in PMI to several D-dimensional vectors using the function q(x), and the function f decUnderstand that a general process is represented by using (x) to transform these D-dimensional vectors into M×2 real matrices, calculating the complex matrix form of the real matrices, obtaining M complex weight coefficients, multiplying the complex weight coefficients separately in M groups of spatial frequency basis vectors, and finally performing addition to obtain the reconstructed downlink channel feature matrix H'. The value of M above is used as an example without limitation. In other words, the specific value of M is different from the above example for different preprocessing methods. On the condition that the end-to-end reconstruction method satisfies the sixth relationship described in Equation 15, this is considered to fall within the protection scope of the embodiments of the present application.

[0417] In particular, generally, H is a two-dimensional matrix plane (spatial domain × frequency domain). Equation 15 may further impose the following restrictions on the preprocessing procedure of step S505, and the selected coefficients need to be arranged according to a one-dimensional vector. For example, H is transformed into a one-dimensional vector (without compression), or the candidate coefficient matrix is transformed into a one-dimensional vector, and then M coefficients with the maximum energy are selected (similar to compression method 1 or 2).

[0418] 3. The output dimension of the decoder network is M×N×T×2. In this specification, it is assumed that the dimension of the second matrix is M×N×T×2. In this case, it is generally only the fifth method, and T = N rx is.

[0419] When the dimension of the second matrix M×N×T is the same as the dimension of the downlink channel feature matrix H' to be reconstructed, the second matrix is the corresponding complex matrix

Number

[0420] Alternatively, if the dimensions M×N×T of the second matrix are different from those of the downlink channel feature matrix H' to be reconstructed (for example, in step S505, further compression is performed with reference to method 3 of step 4 of the second method), optionally, the network device converts the second matrix into a complex matrix, and based on the M×N position information obtained from the CSI, fills the elements of the complex matrix of the all-zero matrix having the same dimensions as the downlink channel feature matrix H' one by one,

Number

[0421] Alternatively, if the dimensions M×N×T of the second matrix are different from those of the downlink channel feature matrix H' to be reconstructed (for example, in step S505, further compression is performed with reference to method 1 or 2 of step 4 of the second method), optionally, the network device converts the second matrix into a complex matrix, and based on the M position information obtained from the CSI, fills the elements of the complex sequence of the all-zero matrix having the same dimensions as the downlink channel feature matrix H' one by one,

Number

[0422] The network device performs an inverse transformation on the complex matrix

Number

Number

[0423] When the output dimension of the decoder network is M×N×T, Step S509 may be expressed using a uniform reconstruction formula. Under the condition that the information regarding the reconstructed downlink channel matrix satisfies the seventh relationship, the network device may be regarded as being able to obtain the information regarding the reconstructed downlink channel matrix in any manner. The steps (for example, Steps S5091 and S5092) performed by the network device in Step S509 are merely examples and are not intended to limit the behavior of the network device. The seventh relationship includes the following:

Number

[0424] In Equation 16, q(ind i ) represents the S indices of the S vectors included in the first vector quantization dictionary, that is, the S indices of the S vectors corresponding to the PMI. The function q(x) indicates mapping the indices of the S vectors to a first matrix with dimensions D×S based on the first vector quantization dictionary. The function f dec (x) indicates reconstructing the first matrix using the decoder network to obtain a second matrix with dimensions M×N×T×2. In this specification, T = N rx . Both the first matrix and the second matrix are real matrices. (x)| t=1…T indicates that, in this specification, to obtain the reconstructed downlink channel matrix, T matrices with dimensions M×N are processed (or reconstructed) based on the second matrix according to Equation 16.

Number

Number

[0425] {U 1,k} k=1…M represents a set of spatial domain basis vectors. The set of spatial domain basis vectors is, for example, the spatial domain basis used when the UE performs the spatial frequency joint projection described above. The length of the vectors included in the set of spatial domain basis vectors is N tx . {U 2,l} l=1…N represents a set of frequency domain basis vectors. The set of frequency domain basis vectors is, for example, the frequency domain basis described above used when the UE performs the spatial frequency joint projection. For example, the network device may use the DFT formula to obtain the set of spatial domain basis vectors and the set of frequency domain basis vectors. The length of the vectors included in the set of frequency domain basis vectors is N sb . When the frequency domain granularity is one RB, N sb = N rb .

[0426] The values of M and N depend on the preprocessing method in step S605. Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are orthogonal DFT basis vector sets, and C complex is not further compressed. In this case, the values of M and N are respectively equal to the lengths of the vectors included in the corresponding basis vector sets. In this specification, M = N tx , and N = N sb .

[0427] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are orthogonal DFT basis vector sets, and C complexis compressed with reference to Method 3 of Step 4 (of the second method). In this case, the values of M and N are each smaller than the lengths of the vectors included in the corresponding base vector sets. In this specification, M < N tx , and N < N sb .

[0428] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are oversampled DFT base vector sets, and C complex is not further compressed. In this case, the values of M and N are each equal to the product of the length of the vector included in the corresponding base vector set and the oversampling coefficient. In this specification, M = N tx ×x, and N = N sb ×y, where x and y are spatial frequency oversampling coefficients, respectively.

[0429] Optionally, {U 1,k} k=1…M and {U 2,l} l=1…N are oversampled DFT base vector sets, and C complex is compressed with reference to Method 3 of Step 4 (of the second method). In this case, the values of M and N are each smaller than the product of the length of the vector included in the corresponding base vector set and the oversampling coefficient. In this specification, M < N tx ×x, N < N sb ×y. In particular, in order to achieve the purpose of information compression, in this case, generally, M < N tx and N < N sb .

[0430] The reconstructed channels obtained using Equation 10 and Equation 16 are equivalent. Equation 10 shows the inverse transform of the spatial frequency projection. S, F, and

Number

[0431] Equation 16 restores the channel information included in PMI to several D-dimensional vectors using the function q(x), and the function f dec (x) is used to convert these D-dimensional vectors into M×N×T×2 real matrices, calculate the complex matrix form of the real matrices, obtain M×N×T weight complex coefficients, take M×N weight complex coefficients for each element based on dimension T, multiply the M spatial domain basis vectors and N frequency domain basis vectors separately by the weight complex coefficients, and finally perform addition to obtain the reconstructed downlink channel feature matrix H'. It should be understood that the above M×N values are used as examples rather than limitations. In other words, the specific values of M×N are different from the above examples due to different preprocessing methods. On the condition that the end-to-end reconstruction method satisfies the seventh relationship described in Equation 16, this shall fall within the protection scope of the embodiments of the present application.

[0432] In particular, Equation 16 further imposes the following restrictions on the preprocessing procedure of step S505, where C complex When is compressed, the selected coefficients need to be 2D matrices corresponding separately to the spatial domain dimension and the frequency domain dimension (similar to Method 3).

[0433] 4. The output dimension of the decoder network is M×T×2. In this specification, it is assumed that the dimension of the second matrix is M×T×2. In this case, it is generally only the fifth method, where T = N rx is the case.

[0434] For example, if the dimension M×T×2 of the second matrix is different from the dimension of the downlink channel feature matrix H' to be reconstructed (for example, in step S505, further compression is performed with reference to method 1 or method 2 in step 4 of the second method), optionally, the network device converts the second matrix into a complex matrix whose dimension is T×M. The T×M complex matrices include T complex vectors whose dimension is M. For each complex vector whose dimension is M, the network device satisfies the M complex coefficients at the corresponding positions based on the positions of the M complex coefficients in the original two-dimensional spatial frequency plane reported using CSI. Other positions are set to 0. In this way, the complex matrix

Number

Number

[0435] When the output dimension of the decoder network is M×T×2, step S509 may be expressed using a uniform reconstruction formula. On the condition that the information about the reconstructed downlink channel matrix satisfies the eighth relationship, the network device may be regarded as obtaining the information about the reconstructed downlink channel matrix in any manner. The steps (for example, steps S5091 and S5092) executed by the network device in step S509 are merely examples and are not intended to limit the behavior of the network device. The eighth relationship includes the following:

Number

[0436] In Equation 17, q(ind i) represents the S indices of the S vectors included in the first vector quantization dictionary, that is, the S indices of the S vectors corresponding to the PMI. The function q(x) indicates mapping the indices of the S vectors to a first matrix with dimensions D×S based on the first vector quantization dictionary. The function f dec (x) indicates reconstructing the first matrix using a decoder network to obtain a second matrix with dimensions M×T×2. In this specification, T = N rx is the case. Both the first matrix and the second matrix are real matrices. The function C(x) indicates converting the second matrix into a complex matrix (for example, the complex matrix described in step S509) with dimensions M×T

Number

Number

Number

[0437] {U 1,(j,k)}} j=1…M represents a spatial domain basis vector set. The spatial domain basis vector set is, for example, the spatial domain basis used when the UE performs the spatial frequency joint projection described above. The length of the vectors included in the spatial domain basis vector set is N tx .

Number

[0438] The value of M depends on the preprocessing method in step S605. Optionally, {U 1,(j,k)} j=1…M and

Number

[0439] Optionally, {U 1,(j,k)} j=1…M and

Number

[0440] Optionally, {U 1,(j,k)} j=1…M and

Number

[0441] Optionally, {U 1,(j,k)} j=1…M and

Number

[0442] The reconstructed channels obtained using Equation 10 and Equation 17 are equivalent. Equation 10 represents the inverse transform of the spatial frequency projection. The dimensions of S, F, and

Number

Number

[0443] Equation 17 restores the channel information included in the PMI into several D-dimensional vectors using the function q(x), and uses the function f dec (x) to convert these D-dimensional vectors into M×T×2 real matrices, calculate the complex matrix form of the real matrices, obtain M×T weighted complex coefficients, take M weighted complex coefficients for each element based on dimension T, multiply the M groups of spatial frequency basis vectors by the weighted complex coefficients separately, and finally perform addition to obtain the reconstructed downlink channel feature matrix H'. It should be understood that the above value of M is used as an example rather than a limitation. In other words, the specific value of M is different from the above example for different preprocessing methods. On the condition that the end-to-end reconstruction method satisfies the eighth relationship described in Equation 17, this shall fall within the protection scope of the embodiments of the present application.

[0444] In particular, generally, H is a two-dimensional matrix plane (spatial domain × frequency domain). Equation 17 may further impose the following restrictions on the preprocessing procedure of step S505. The selected coefficients need to be arranged according to a one-dimensional vector. For example, H is converted into a one-dimensional vector (without compression), or the candidate coefficient matrix is converted into a one-dimensional vector, and then M coefficients with the maximum energy are selected (similar to compression method 1 or 2).

[0445] In this embodiment of the present application, steps S601 to S606 are optional steps. In addition, steps S6091 and S6092 included in step S609 are also optional steps.

[0446] In addition, it should be noted that this embodiment of the present application may be applied to the network architecture and network devices of the UE, or may be applied to the network architecture, network devices, and AI nodes of the UE. When this embodiment of the present application is applied to the network architecture, network devices, and AI nodes of the UE, for example, the decoder network is trained offline, the NN-CSI network is prepared by the network device in step S602, and the first matrix is input to the decoder network in step S609 to obtain the second matrix, etc. The process may be executed by the AI node.

[0447] In this embodiment of the present application, the CSI can be determined based on a first vector quantization dictionary. For example, the PMI may be represented using the indices of one or more vectors included in the first vector quantization dictionary. In this case, the information obtained by measurement may be quantized using the indices included in the first vector quantization dictionary. Since the dimension of the vector is usually relatively large, it is equivalent to dimension expansion of information, thereby improving the quantization accuracy. In addition, the original channel matrix can be compressed by the cooperation between the encoder network and the decoder network, thereby reducing the compression loss. In addition, in this embodiment of the present application, a VQ-AE network is used. In the VQ-AE network, the quantization network and the compression network are jointly designed so that joint optimization can be implemented between compression and quantization, reducing the performance loss and improving the overall accuracy of the feedback channel state information. In addition, in this embodiment of the present application, in order to adjust the complexity of CSI feedback, the input and output dimensions of the neural network (for example, the input and output dimensions of the encoder and / or the input and output dimensions of the decoder) can be restricted, thereby emphasizing the advantage of using the neural network to compress the downlink channel matrix. In addition, in this embodiment of the present application, since there is no need to perform conversion between complex matrices and real matrices, the CSI feedback procedure can be reduced, thereby improving the feedback efficiency and simplifying the implementation complexity. In addition, by using a complex number network, the physical meaning of complex number signals can be considered in aspects such as the convolution method and loss function design. This is more suitable for the processing of communication signals, thereby helping to retain the original phase information of communication complex number signals.

[0448] The embodiment shown in FIG. 5 and the embodiment shown in FIG. 6 are both related to the vector quantization dictionary. Regarding the method for determining the vector quantization dictionary by the UE, the embodiment shown in FIG. 5 or the embodiment shown in FIG. 6 may be described, or the vector quantization dictionary may be updated online so that the vector quantization dictionary can conform to the actual situation of the network. Alternatively, the UE may determine the vector quantization dictionary in the manner described in the embodiment shown in FIG. 5 or the embodiment shown in FIG. 6, and the vector quantization dictionary may be updated online (in other words, the vector quantization dictionary may also be determined in the manner described in this embodiment of the present application). FIG. 7 is a flowchart of a third communication method according to an embodiment of the present application. The online update of the vector quantization dictionary can be implemented using this method.

[0449] S701: The network device transmits the first configuration information to the UE, and correspondingly, the UE receives the first configuration information from the network device.

[0450] The first configuration information may be used to configure the online update of the vector quantization dictionary, or the first configuration information may be used to configure the reporting parameters of the vector quantization dictionary. For example, the first configuration information may include one or more of the information regarding the first delay, the format information of the vector quantization dictionary, or the periodic configuration information for reporting the vector quantization dictionary.

[0451] In this embodiment of the present application, the UE may update the vector quantization dictionary, and the information regarding the first delay may be used to indicate the UE to wait for the first delay before transmitting the vector quantization dictionary to the network device.

[0452] The format information of the vector quantization dictionary includes, for example, one or more of the dimension information of the vector quantization dictionary, the type information of the vector quantization dictionary, or the index of the vector quantization dictionary. The dimension information of the vector quantization dictionary may indicate the dimension of the vector quantization dictionary. For example, the dimension of the vector quantization dictionary is D×E. The index of the vector quantization dictionary indicates the index of the vector quantization dictionary configured for the UE from among a plurality of vector quantization dictionaries. Each of the plurality of vector quantization dictionaries corresponds to one unique index.

[0453] The periodic configuration information for reporting the vector quantization dictionary may be used to configure the periodic reporting of the vector quantization dictionary. For example, the periodic configuration information may include a period R1. For example, the network device transmits first parameter information according to the period R1 to trigger the UE to transmit the vector quantization dictionary. The UE receives the first parameter information and transmits the vector quantization dictionary to the network device. This is regarded as a process in which the UE transmits the vector quantization dictionary to the network device according to the period R1. Alternatively, the periodic configuration information may include a period R1. For example, the network device transmits first scheduling information according to the period R1. The first scheduling information may indicate uplink resources. In this way, the UE is triggered to transmit the vector quantization dictionary. The UE receives the first scheduling information and transmits the vector quantization dictionary to the network device. This is regarded as a process in which the UE transmits the vector quantization dictionary to the network device according to the period R1.

[0454] Alternatively, the periodic configuration information for reporting the vector quantization dictionary may be used to configure a semi-persistent report of the vector quantization dictionary. For example, the periodic configuration information may be used to configure the resources used to transmit the vector quantization dictionary and the period R1. When a semi-persistent report of the vector quantization dictionary is configured, the network device may use signaling to activate the mechanism of the semi-persistent report of the vector quantization dictionary. The UE receives the signaling. Before receiving the de-activation command, the UE may transmit the vector quantization dictionary to the network device according to the period R1, and the network device does not need to schedule the UE every time. The signaling used for activation and / or the signaling used for de-activation may be, for example, MAC CE or other signaling.

[0455] Alternatively, the periodic configuration information for reporting the vector quantization dictionary may be used to configure an aperiodic report of the vector quantization dictionary. When an aperiodic report of the vector quantization dictionary is configured, the way for the UE to transmit the vector quantization dictionary to the network device waits for a trigger executed by the network device. For example, the network device transmits first indication information (e.g., scheduling information, or trigger information specifically used to trigger the UE) to the UE to trigger the UE to transmit the vector quantization dictionary. After receiving the first indication information, the UE may transmit the vector quantization dictionary to the network device. In other words, when the network device executes the trigger, the UE transmits the vector quantization dictionary, or when the network device does not execute the trigger, the UE does not transmit the vector quantization dictionary. In this way, the network device can trigger the UE to transmit the vector quantization dictionary as needed, thereby reducing the process of receiving redundant information.

[0456] In this embodiment of the present application, the periodic configuration information for reporting the vector quantization dictionary can be used to constitute one or more of the periodic reporting of the vector quantization dictionary, the semi-persistent reporting of the vector quantization dictionary, or the aperiodic reporting of the vector quantization dictionary. In other words, the UE may send the vector quantization dictionary to the network device in one of the methods, or may send the vector quantization dictionary to the network device in any combination of the methods. For example, when the network device sends the first configuration information at a specific time, it is for reporting the vector quantization dictionary, and the periodic configuration information included in the first configuration information is used to constitute the periodic reporting of the vector quantization dictionary. In this case, the UE periodically sends the vector quantization dictionary to the network device. Then, the network device newly sends the first configuration information, and the periodic configuration information included in the first configuration information for reporting the vector quantization dictionary is used to newly constitute the aperiodic reporting of the vector quantization dictionary. In this case, the UE regards the newly sent first configuration information as an implicit instruction, uses the implicit instruction as instruction information for ending the periodic reporting, and then may report the vector quantization dictionary according to the aperiodic reporting mechanism after receiving the corresponding first instruction information. Alternatively, when the network device newly sends the first configuration information and the periodic configuration in the configuration information is changed to another method, for example, semi-persistent, the UE waits for an activation command according to the new configuration before performing the periodic reporting.

[0457] This embodiment of the present application can be applied in combination with the embodiment shown in FIG. 5 or the embodiment shown in FIG. 6. In this case, the first configuration information and the third configuration information may be the same configuration information or different configuration information. When the first configuration information and the third configuration information are different configuration information and the network device sends the first configuration information and the third configuration information to the UE, the first configuration information may be sent after the third configuration information, or the first configuration information and the third configuration information may be sent simultaneously.

[0458] Step S701 is an optional step. In other words, the network device may not send the first configuration information to the UE.

[0459] S702: The network device determines the first parameter information.

[0460] The first parameter information is, for example, the parameter information of the decoder network maintained by the network device. The decoder network includes, for example, the decoder network currently used by the network device and the UE, or all the decoder networks maintained by the network device.

[0461] For example, the network device reconstructs the downlink channel matrix based on the CSI sent by the UE. For example, the UE performs one or more measurements, and the downlink channel matrix obtained through the measurements is {H1, H2, …, H T}. After the UE obtains the downlink channel matrix through measurement each time, the UE may send the CSI to the network device. For example, the downlink channel matrix reconstructed by the network device based on the CSI from the UE is

Number

[0462] For processes such as the process by which the UE determines CSI and the process by which the network device reconstructs the downlink channel matrix, the solution provided in the embodiment shown in FIG. 5 or the embodiment shown in FIG. 6 may be used, or another solution may be used.

[0463] S703: The network device transmits first parameter information to the UE, and correspondingly, the UE receives the first parameter information from the network device.

[0464] For example, the first parameter information may include one or more items in Table 1.

[0465] [Table 1]

[0466] In addition, the first parameter information may further include one or more other parameters of the decoder network.

[0467] Both steps S702 and S703 are optional steps. In other words, the network device does not have to determine the first parameter information and correspondingly does not have to transmit the first parameter information. Alternatively, even if the network device determines the first parameter information, the network device does not necessarily have to transmit the first parameter information.

[0468] S704: The UE determines a second vector quantization dictionary. The second vector quantization dictionary is, for example, the first vector quantization dictionary, or the second vector quantization dictionary is, for example, a vector quantization dictionary obtained by updating the first vector quantization dictionary.

[0469] For example, the UE may update the first vector quantization dictionary to obtain the second vector quantization dictionary. The UE may update the first vector quantization dictionary based on one or more of the following items, namely, the first parameter information, the downlink channel matrices {H1, H2, …, H T}, and the history data (e.g., the historical downlink channel matrices obtained through measurements).

[0470] Alternatively, the UE may determine the first vector quantization dictionary that is currently used or should be used by the UE without updating the vector quantization dictionary used by the UE.

[0471] S705: The UE transmits the second vector quantization dictionary to the network device, and correspondingly, the network device receives the second vector quantization dictionary from the UE. It will be understood that the UE transmits the parameters of the second vector quantization dictionary to the network device, and the network device receives the parameters of the second vector quantization dictionary from the UE. When the second vector quantization dictionary is the updated first vector quantization dictionary and the embodiment shown in FIG. 8 is combined with the embodiment shown in FIG. 5, step S705 may be executed after the execution of the embodiment shown in FIG. 5. When the embodiment shown in FIG. 8 is combined with the embodiment shown in FIG. 5, step S705 may be executed after the execution of the embodiment shown in FIG. 6. Alternatively, when the second vector quantization dictionary is the first vector quantization dictionary and the embodiment shown in FIG. 8 is combined with the embodiment shown in FIG. 5, step S705 may be executed before step S509 of the embodiment shown in FIG. 5. When the embodiment shown in FIG. 8 is combined with the embodiment shown in FIG. 5, step S705 may be executed before step S609 in the embodiment shown in FIG. 6.

[0472] The UE may transmit the second vector quantization dictionary to the network device in multiple implementation forms. Hereinafter, examples are used for illustration.

[0473] In a first optional implementation, the UE may periodically transmit a second vector quantization dictionary to the network device. The period is configured, for example, using first configuration information and is, for example, period R1. In this case, after receiving the first configuration information, the UE may transmit the second vector quantization dictionary to the network device according to period R1. For example, if the first configuration information includes periodic configuration information for reporting the vector quantization dictionary and the periodic configuration information is used to configure an aperiodic report of the vector quantization dictionary, this solution may be applicable. Alternatively, the period for the UE to transmit the second vector quantization dictionary is specified by a protocol, for example. For example, the network device does not need to transmit the first configuration information. In this case, the UE transmits the second vector quantization dictionary to the network device according to the period. In this embodiment of the present application, the fact that the UE periodically transmits the second vector quantization dictionary to the network device means that the UE transmits the second vector quantization dictionary to the network device in each period, or the UE does not need to transmit the second vector quantization dictionary to the network device in each period, but may transmit the second vector quantization dictionary to the network device in one or more periods, and the one or more periods may be continuous or discontinuous in the time domain, which should be noted.

[0474] In a second optional implementation, when the UE receives the first configuration information from the network device, or the UE receives the first configuration information and the first parameter information from the network device, or the UE receives the first configuration information, the first parameter information, and the first indication information from the network device, the UE may transmit the second vector quantization dictionary to the network device after receiving the first configuration information. For example, if the first configuration information includes periodic configuration information for reporting the vector quantization dictionary and the periodic configuration information is used to configure an aperiodic report of the vector quantization dictionary, this solution may be applicable.

[0475] In a third optional implementation form, the UE receives the first configuration information from the network device, or the UE receives the first configuration information and the first parameter information from the network device, or the UE receives the first configuration information, the first parameter information, and the first indication information from the network device. When the first configuration information includes information regarding the first delay, the UE may transmit the second vector quantization dictionary to the network device in arrival time units of the first delay after the first configuration information is received. For example, if the time unit when the UE receives the first configuration information is T0 and the first delay is represented using T delay1 , the UE may transmit the second vector quantization dictionary to the network device at the time unit of T0 + T delay1 . In this case, the network device also receives the second vector quantization dictionary from the UE in arrival time units of the first delay after the first configuration information is transmitted. For example, the network device may reserve uplink resources for the UE at the time unit of T0 + T delay1 , the UE may transmit the second vector quantization dictionary on the uplink resources, and the network device may also detect and receive the second vector quantization dictionary on the uplink resources. However, when the first configuration information does not include information regarding the first delay, after receiving the first configuration information, the UE may transmit the second vector quantization dictionary to the network device without waiting. For example, when the first configuration information includes periodic configuration information for reporting the vector quantization dictionary and the periodic configuration information is used to configure the aperiodic reporting of the vector quantization dictionary, this solution may be applied.

[0476] In a fourth optional implementation form, when the UE receives the first configuration information and the first parameter information from the network device, or when the UE receives the first configuration information, the first parameter information, and the first indication information from the network device, the UE may transmit the second vector quantization dictionary to the network device after the first parameter information is received. For example, when the first configuration information includes periodic configuration information for reporting the vector quantization dictionary and the periodic configuration information is used to constitute an aperiodic report of the vector quantization dictionary, this solution may be applied.

[0477] In a fifth optional implementation form, when the UE receives the first configuration information and the first parameter information from the network device, or when the UE receives the first configuration information, the first parameter information, and the first indication information from the network device, and the first configuration information includes information regarding the first delay, the UE may transmit the second vector quantization dictionary to the network device in arrival time units of the first delay after the first parameter information is received. For example, if the time unit when the UE receives the first parameter information is T1 and the first delay is represented by T delay1 , the UE may transmit the second vector quantization dictionary to the network device in time units of T1 + T delay1 . In this case, the network device also receives the second vector quantization dictionary from the UE in arrival time units of the first delay after the first parameter information is transmitted. For example, the network device receives it at T1 + T delay1The uplink resources for the UE may be reserved in time units, the UE may transmit a second vector quantization dictionary on the uplink resources, and the network device may also detect and receive the second vector quantization dictionary on the uplink resources. If the first configuration information does not include information regarding the first delay, after receiving the first parameter information, the UE may transmit the second vector quantization dictionary to the network device without waiting. For example, if the first configuration information includes periodic configuration information for reporting the vector quantization dictionary, and the periodic configuration information is used to configure an aperiodic report of the vector quantization dictionary, a semi-persistent report of the vector quantization dictionary, or an aperiodic report of the vector quantization dictionary, this solution may be applicable.

[0478] In a sixth alternative implementation, if the UE receives the first configuration information, the first parameter information, and the first indication information from the network device, the UE may transmit the second vector quantization dictionary to the network device after receiving the first indication information. For example, if the first configuration information includes periodic configuration information for reporting the vector quantization dictionary, and the periodic configuration information is used to configure an aperiodic report of the vector quantization dictionary, this solution may be applicable.

[0479] In a seventh alternative implementation, if the UE receives the first configuration information, the first parameter information, and the first indication information from the network device, and the first configuration information includes information regarding the first delay, the UE may transmit the second vector quantization dictionary to the network device at the arrival time of the first delay after the first indication information is received. For example, if the time unit when the UE receives the first indication information is T2, and the first delay is represented by T delay1 when used, the UE may transmit the second vector quantization dictionary to the network device in time units of T2+T delay1 In this case, the network device may also receive the second vector quantization dictionary from the UE at the arrival time unit of the first delay after the first indication information is transmitted. For example, the network device may receive it at T2+T delay1The uplink resources for the UE may be reserved in time units, the UE may transmit a second vector quantization dictionary on the uplink resources, and the network device may also detect and receive the second vector quantization dictionary on the uplink resources. If the first configuration information does not include information regarding the first delay, after receiving the first indication information, the UE may transmit the second vector quantization dictionary to the network device without waiting. For example, this solution may be applicable if the first configuration information includes periodic configuration information for reporting the vector quantization dictionary and the periodic configuration information is used to configure an aperiodic report of the vector quantization dictionary.

[0480] The preparation time may be reserved for the UE by indicating the first delay such that the UE has a corresponding time for determining the second vector quantization dictionary.

[0481] Optionally, the first delay may be specified by the protocol. For example, the first delay may be applied to all UEs, or the first delay may be set separately for different network devices or different cells. Alternatively, the first delay may be determined by the network device. For example, the network device may determine the same first delay for all UEs covered by the network device, or the network device may determine the first delay separately for different cells provided by the network device, or the network device may determine the first delay separately for different UEs. For example, the UE may send capability information to the network device. The capability information may indicate information regarding the third delay. The third delay may be determined based on the processing capability of the UE. For example, the third delay is the shortest processing time required by the U...

Claims

1. A communication method, comprising: determining channel state information, wherein the channel state information is determined based on S indices of S vectors, the S vectors are determined based on information output by an encoder network, each of the S vectors is included in a first vector quantization dictionary, and the first vector quantization dictionary includes N 1 vectors, and both N 1 and S are positive integers; transmitting the channel state information to a network device; and a method comprising the above steps.

2. The method according to claim 1, wherein the S vectors are determined based on S pieces of first channel information and the first vector quantization dictionary, and the S pieces of first channel information are information output by an encoder network. The method according to claim 1.

3. The method according to claim 2, wherein the N 1 vectors have the same length, and the length is equal to at least one dimension in the output dimension of the encoder network.

4. The method according to claim 2 or 3, wherein the input dimension of the encoder network is determined based on the input dimension of a first reference encoder network, and the output dimension of the encoder network is determined based on the output dimension of the first reference encoder network.

5. The method according to claim 4, wherein the first reference encoder network corresponds to a first reference decoder network.

6. The method according to claim 5, wherein a first reference network includes the first reference encoder network and the first reference decoder network, and the first reference network further includes a reference vector quantization dictionary.

7. The input dimension of the encoder network is M or M×2, the output dimension of the encoder network is D×S, both M and D are positive integers, M≤N tx ×N sb ×z, S≤M, z is a positive integer greater than or equal to 1, N tx represents the number of transmission antenna ports of the network device, N sb represents the number of frequency domain sub-bands, The method according to claim 4.

8. The input dimension of the encoder network is M×N×2 or M×N, the output dimension of the encoder network is D×P×Q, S = P×Q, and M, N, D, P, and Q are all positive integers, M≤N tx , N≤N sb , S≤M×N, or M = N tx ×x, N = N sb ×y, S≤(N tx ×x)×(N sb ×y), or M = N tx ×N rx , N = N sb , S≤(N tx ×N rx )×N sb where N tx represents the number of transmission antenna ports of the network device, N sb represents the number of frequency domain sub-bands, N rx represents the number of receiving antenna ports of the terminal device, and both x and y are greater than 1, The method according to claim 4.

9. The input dimension of the encoder network is M×N×T×2 or M×N×T, the output dimension of the encoder network is D×P×Q×R, S = P×Q×R, and M, N, D, P, Q, R, and T are all positive integers, M = N tx , N = N sb , T = N rx , and S ≤ N rx ×N tx ×N sb and N tx represents the number of transmission antenna ports of the network device, and N sb represents the number of frequency domain sub - bands, and N rx represents the number of reception antenna ports of the terminal device The method according to claim 4

10. The method further includes a step of transmitting information about the first vector quantization dictionary to the network device The method according to any one of claims 1 to 3

11. The step of transmitting information about the first vector quantization dictionary to the network device includes a step of periodically transmitting the information about the first vector quantization dictionary to the network device a step of periodically transmitting the information about the first vector quantization dictionary to the network device after the first configuration information is received from the network device a step of transmitting the information about the first vector quantization dictionary to the network device after the first configuration information is received from the network device a step of transmitting the information about the first vector quantization dictionary to the network device in units of the arrival time of the first delay after the first configuration information is received from the network device, where the first configuration information includes information about the first delay a step of transmitting the information about the first vector quantization dictionary to the network device after the first parameter information is received from the network device, where the first parameter information indicates parameter information of a decoder network After the first parameter information is received from the network device, transmitting the information regarding the first vector quantization dictionary to the network device in arrival time units of a first delay, where the first parameter information indicates parameter information of a decoder network, step After the first instruction information is received from the network device, transmitting the information regarding the first vector quantization dictionary to the network device, where the first instruction information is used to trigger transmission of the vector quantization dictionary, step, or After the first instruction information is received from the network device, transmitting the information regarding the first vector quantization dictionary to the network device in arrival time units of a first delay, where the first instruction information is used to trigger transmission of the vector quantization dictionary, step The method according to claim 10, comprising

12. The current networking method is a first networking method, and the first vector quantization dictionary is a vector quantization dictionary corresponding to the first networking method The moving speed of the terminal device belongs to a first interval, and the first vector quantization dictionary is a vector quantization dictionary corresponding to the first interval The terminal device is handed over from a second cell to a first cell, and the first vector quantization dictionary is a vector quantization dictionary corresponding to the first cell The reference encoder network used is switched from a second reference encoder network to a first reference encoder network, and the first vector quantization dictionary is a vector quantization dictionary corresponding to the first reference encoder network, or The first vector quantization dictionary is a vector quantization dictionary indicated by second instruction information, and the second instruction information is from the network device The method according to any one of claims 1 to 3

13. the method further comprising: transmitting second parameter information to the network device, the second parameter information indicating parameters of the encoder network; The method according to claim 2 or 3.

14. The step of transmitting the second parameter information to the network device is periodically transmitting the second parameter information to the network device; periodically transmitting the second parameter information to the network device after second configuration information is received from the network device; transmitting the second parameter information to the network device after second configuration information is received from the network device; transmitting the second parameter information to the network device in arrival time units of a second delay after second configuration information is received from the network device, the second configuration information including information regarding the second delay; transmitting the second parameter information to the network device after third indication information is received from the network device, the third indication information being used to trigger transmission of parameter information of the encoder network; or transmitting the second parameter information to the network device in arrival time units of a second delay after third indication information is received from the network device, the third indication information being used to trigger transmission of parameter information of the encoder network; The method according to claim 13.

15. A method for receiving channel state information, comprising: A step of receiving channel state information, wherein the channel state information includes S indexes of S vectors, and the S vectors are determined based on information output by an encoder network; A step of obtaining information on a downlink channel matrix reconstructed based on the S indexes and a first vector quantization dictionary, wherein each of the S vectors is included in the first vector quantization dictionary, and the first vector quantization dictionary includes N 1 vectors, and both N 1 and S are positive integers; A method comprising the above steps. **Claim 16** The step of obtaining information on a downlink channel matrix reconstructed based on the S indexes and a first vector quantization dictionary includes: A step of performing an inverse mapping on the S indexes based on the first vector quantization dictionary to obtain a first matrix; A step of obtaining information on the reconstructed downlink channel matrix based on the first matrix and a decoder network; The method according to claim 15, comprising the above steps. **Claim 17** The step of obtaining information on the reconstructed downlink channel matrix based on the first matrix and a decoder network includes: A step of inputting the first matrix into the decoder network to obtain a second matrix; A step of obtaining the information on the reconstructed downlink channel matrix based on the second matrix; The method according to claim 16, comprising the above steps. **Claim 18** The reconstructed downlink channel matrix satisfies: 【Equation 1】 and satisfies: {ind i} i＝1…S represents the S indexes of the S vectors in the first vector quantization dictionary, and the function q(x) indicates mapping the S indexes to a first matrix with dimensions D×S based on the first vector quantization dictionary, and the function f dec (x) indicates reconstructing the first matrix using a decoder network to obtain a second matrix with dimensions M×N×2, both the first matrix and the second matrix are real matrices, and the function C(x) indicates converting the second matrix into a complex matrix with dimensions M×N, {U 1，k} k＝1…M represents a spatial domain basis vector set, and U 1，k represents the k-th vector in the spatial domain basis vector set, and the length of the vectors included in the spatial domain basis vector set is N tx and {U 2，l} l＝1…N represents a frequency domain basis vector set, and U 2，l represents the l-th vector in the frequency domain basis vector set, and the length of the vectors included in the frequency domain basis vector set is N sb and N tx represents the number of transmit antenna ports of the network device, and N sb represents the number of frequency domain sub-bands, 【Equation 2】 represents the conjugate transpose matrix of U 2，l , and D, S, M, and N are all positive integers, The method according to any one of claims 15 to 17.

19. M = N tx , and N = N sb and M ≤ N tx , and N ≤ N sb and M = N tx ×x, and N = N sb ×y, where both x and y are greater than 1, The method according to claim 18.

20. the reconstructed downlink channel matrix satisfies 【Equation 3】 and {ind i} i＝1…S represents the S indices of the S vectors in the first vector quantization dictionary, the function q(x) indicates mapping the S indices to a first matrix of dimension D×S based on the first vector quantization dictionary, the function f dec (x) indicates reconstructing the first matrix using a decoder network to obtain a second matrix of dimension M×2, both the first matrix and the second matrix are real matrices, the function C(x) indicates converting the second matrix to a complex matrix of dimension M, where M≦N tx ×N sb and 、 {U 1，（j，k）} j＝1…M represents a spatial domain basis vector set, U 1，（j，k） represents the j-th vector in the spatial domain basis vector set, the length of the vectors included in the spatial domain basis vector set is N tx and, {U 2，（j，l）} j＝1…M represents a frequency domain basis vector set, U 2，（j，l） represents the j-th vector in the frequency domain basis vector set, the length of the vectors included in the frequency domain basis vector set is N sb and, N tx represents the number of transmit antenna ports of the network device, N sb represents the number of frequency domain sub-bands, 【Equation 4】 is the conjugate transpose matrix of U 2，（j，l） , and D, S, and M are all positive integers. The method according to any one of claims 15 to 17.

21. The reconstructed downlink channel matrix satisfies 【Equation 5】 satisfies, q(ind i ) represents the S indices of the S vectors in the first vector quantization dictionary, and the function q(x) indicates mapping the S indices to a first matrix with dimensions D×S based on the first vector quantization dictionary. The function f dec (x) indicates reconstructing the first matrix using a decoder network to obtain a second matrix with dimensions M×N×T×2, where T = N rx and 、 N rx represents the number of receiving antenna ports of the terminal device. Both the first matrix and the second matrix are real matrices. The function C(x) indicates converting the second matrix to a complex matrix with dimensions M×N×T, and (x)| t＝1…T indicates obtaining the reconstructed downlink channel matrix by separately processing T matrices with dimensions M×N based on the second matrix, where {U 1，k} k＝1…M represents a spatial domain basis vector set, and U 1，k represents the k-th vector in the spatial domain basis vector set. The length of the vectors included in the spatial domain basis vector set is N tx and {U 2，l} l＝1…N represents a frequency domain basis vector set, and U 2，l represents the l-th vector in the frequency domain basis vector set. The length of the vectors included in the frequency domain basis vector set is N sb and N tx represents the number of transmitting antenna ports of the network device, and N sb represents the number of frequency domain sub-bands. 【Equation 6】 is the conjugate transpose vector of U 2，l . D, S, M, N, and T are all positive integers. The method according to any one of claims 15 to 17.

22. The reconstructed downlink channel matrix satisfies 【Equation 7】 where q(ind i ) represents the S indices of the S vectors in the first vector quantization dictionary, the function q(x) indicates mapping the S indices to a first matrix with dimensions D×S based on the first vector quantization dictionary, and the function f dec (x) indicates reconstructing the first matrix using a decoder network to obtain a second matrix with dimensions M×T×2, where T = N rx and 、 both the first matrix and the second matrix are real matrices, the function C(x) indicates converting the second matrix to a complex matrix with dimensions M×T, and (x)| t＝1…T indicates obtaining the reconstructed downlink channel matrix by separately processing T matrices with dimension M based on the second matrix, {U 1，（j，k）} j＝1…M represents a spatial domain basis vector set, U 1，（j，k） represents the j-th vector in the spatial domain basis vector set, and the length of the vectors included in the spatial domain basis vector set is N tx and 【Equation 8】 represents a frequency domain basis vector set, U 2，（j，l） represents the j-th vector in the frequency domain basis vector set, and the length of the vectors included in the frequency domain basis vector set is N sb and N tx represents the number of transmit antenna ports of the network device, and N sb represents the number of frequency domain subbands. 【Equation 9】 is the conjugate transpose vector of U 2（j，l） , and D, S, M, and T are all positive integers. The method according to any one of claims 15 to 17.

23. The method further comprises receiving information regarding the first vector quantization dictionary The method according to any one of claims 15 to 17.

24. The step of receiving information regarding the first vector quantization dictionary comprises periodically receiving the information regarding the first vector quantization dictionary periodically receiving the information regarding the first vector quantization dictionary after the first configuration information is transmitted receiving the information regarding the first vector quantization dictionary after the first configuration information is transmitted receiving the information regarding the first vector quantization dictionary in arrival time units of a first delay after the first configuration information is transmitted, wherein the first configuration information includes information regarding the first delay receiving the information regarding the first vector quantization dictionary after the first parameter information is transmitted, wherein the first parameter information indicates parameter information of a decoder network receiving the information regarding the first vector quantization dictionary in arrival time units of a first delay after the first parameter information is transmitted, wherein the first parameter information indicates parameter information of the decoder network receiving the information regarding the first vector quantization dictionary after the first instruction information is transmitted, wherein the first instruction information is used to trigger transmission of the vector quantization dictionary, or receiving the information regarding the first vector quantization dictionary in arrival time units of a first delay after the first instruction information is transmitted, wherein the first instruction information is used to trigger transmission of the vector quantization dictionary The method according to claim 23.

25. The current networking method is the first networking method, and the first vector quantization dictionary is a vector quantization dictionary corresponding to the first networking method. The moving speed of the terminal device belongs to a first interval, and the first vector quantization dictionary is a vector quantization dictionary corresponding to the first interval. The terminal device is handed over from a second cell to a first cell, and the first vector quantization dictionary is a vector quantization dictionary corresponding to the first cell. The reference encoder network used by the terminal device is switched from a second reference encoder network to a first reference encoder network, and the first vector quantization dictionary is a vector quantization dictionary corresponding to the first reference encoder network, or The method further includes a step of transmitting second indication information, where the second indication information indicates information related to the first vector quantization dictionary. The method according to any one of claims 15 to 17.

26. The method is a step of receiving second parameter information, where the second parameter information indicates parameters of the encoder network The method according to any one of claims 15 to 17, further including this step.

27. The step of receiving the second parameter information is a step of periodically receiving the second parameter information, a step of periodically receiving the second parameter information after the second configuration information is transmitted, a step of receiving the second parameter information after the second configuration information is transmitted, a step of receiving the second parameter information in arrival time units of a second delay after the second configuration information is transmitted, where the second configuration information includes information related to the second delay. After the third instruction information is sent, receiving the second parameter information, wherein the third instruction information indicates sending parameter information of the encoder network, or After the third instruction information is sent, receiving the second parameter information in units of the arrival time of a second delay, wherein the third instruction information indicates sending parameter information of the encoder network The method according to claim 26, comprising: **Claim 28** A communication device comprising a module configured to execute the method according to any one of claims 1 to 3. **Claim 29** A communication device comprising a module configured to execute the method according to any one of claims 15 to 17.

Images