WO2023016508A1 - 一种信道信息反馈、恢复方法及装置 - Google Patents

一种信道信息反馈、恢复方法及装置 Download PDF

Info

Publication number
WO2023016508A1
WO2023016508A1 PCT/CN2022/111628 CN2022111628W WO2023016508A1 WO 2023016508 A1 WO2023016508 A1 WO 2023016508A1 CN 2022111628 W CN2022111628 W CN 2022111628W WO 2023016508 A1 WO2023016508 A1 WO 2023016508A1
Authority
WO
WIPO (PCT)
Prior art keywords
channel state
reference signal
matrix
state matrix
dimension
Prior art date
Application number
PCT/CN2022/111628
Other languages
English (en)
French (fr)
Inventor
柴晓萌
吴艺群
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to EP22855488.7A priority Critical patent/EP4376479A1/en
Publication of WO2023016508A1 publication Critical patent/WO2023016508A1/zh
Priority to US18/437,789 priority patent/US20240187283A1/en

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0242—Channel estimation channel estimation algorithms using matrix methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station
    • H04B7/0613—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission
    • H04B7/0615—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal
    • H04B7/0619—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal using feedback from receiving side
    • H04B7/0621—Feedback content
    • H04B7/0626—Channel coefficients, e.g. channel state information [CSI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station
    • H04B7/0613—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission
    • H04B7/0615—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal
    • H04B7/0619—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal using feedback from receiving side
    • H04B7/0658—Feedback reduction
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0254—Channel estimation channel estimation algorithms using neural network algorithms
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L27/00—Modulated-carrier systems
    • H04L27/26—Systems using multi-frequency codes
    • H04L27/2601—Multicarrier modulation systems
    • H04L27/2602—Signal structure
    • H04L27/26025—Numerology, i.e. varying one or more of symbol duration, subcarrier spacing, Fourier transform size, sampling rate or down-clocking
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L5/00—Arrangements affording multiple use of the transmission path
    • H04L5/003—Arrangements for allocating sub-channels of the transmission path
    • H04L5/0048—Allocation of pilot signals, i.e. of signals known to the receiver
    • H04L5/0051—Allocation of pilot signals, i.e. of signals known to the receiver of dedicated pilots, i.e. pilots destined for a single user or terminal
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L5/00—Arrangements affording multiple use of the transmission path
    • H04L5/003—Arrangements for allocating sub-channels of the transmission path
    • H04L5/0048—Allocation of pilot signals, i.e. of signals known to the receiver
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L5/00—Arrangements affording multiple use of the transmission path
    • H04L5/003—Arrangements for allocating sub-channels of the transmission path
    • H04L5/0053—Allocation of signaling, i.e. of overhead other than pilot signals
    • H04L5/0057—Physical resource allocation for CQI

Definitions

  • the present application relates to the field of communication technologies, and in particular to a channel information feedback and recovery method and device.
  • the base station In communication systems such as long term evolution (LTE) and new air interface (new radio, NR), the base station needs to obtain downlink channel state information (channel state information, CSI), which is used to determine and schedule the downlink data channel of the terminal equipment. Resources, modulation and coding scheme (modulation and coding scheme, MCS), precoding matrix and other configurations.
  • CSI channel state information
  • MCS modulation and coding scheme
  • precoding matrix precoding matrix and other configurations.
  • TDD time division duplex
  • the base station can obtain the uplink CSI by measuring the uplink reference signal, and then infer more accurate downlink CSI, such as using the uplink CSI as the downlink CSI.
  • the uplink and downlink channels do not necessarily have reciprocity, so the downlink CSI needs to be obtained from the terminal device.
  • the terminal device obtains downlink CSI by measuring the reference signal, and feeds back the CSI to the base station.
  • CSI feedback requires a large resource overhead.
  • the terminal device performs operations such as compression and quantization on the CSI, so as to feed back the compressed and quantized CSI.
  • operations such as compression and quantization on the CSI will cause loss of accuracy of the CSI.
  • the less accurate the downlink CSI obtained by the base station is the less the scheduled downlink transmission can match the current downlink channel, and the worse the downlink transmission performance will be.
  • the present application provides a method and device for channel information feedback and recovery, which are used to improve the accuracy of channel information feedback while reducing channel information feedback overhead.
  • the present application provides a channel information feedback method, which is applicable to a scenario of AI-based channel information compression feedback.
  • the subject of execution of the method is a terminal device or a module in the terminal device, and the description here takes the terminal device as the subject of execution as an example.
  • the method includes: a terminal device receives a first reference signal from a network device; obtains a first channel state matrix according to the first reference signal; samples the first channel state matrix according to a first sparse pattern to obtain a sparse channel state matrix;
  • the sparse pattern is configured by the network device, the first sparse pattern is used to sample at least one dimension of the first channel state matrix; the first channel information is sent to the network device, and the first channel information is used to indicate the sparse channel state matrix.
  • the network device independently trains the neural network without requiring the terminal device to participate in the training of the neural network, and does not require the terminal device to feed back a large amount of complete downlink channel information.
  • the downlink channel information can be recovered through the neural network.
  • it also includes: sending a second reference signal to the network device, the second reference signal is used to train the neural network corresponding to the first sparse pattern, and the neural network is used to restore the first channel state according to the sparse channel state matrix matrix.
  • the second precoding used for sending the second reference signal is the same as the first precoding used for receiving the first reference signal.
  • the bandwidth of the first reference signal is smaller than or equal to the bandwidth of the second reference signal.
  • the subcarrier spacing corresponding to the first reference signal is greater than or equal to the subcarrier spacing corresponding to the second reference signal.
  • the solution of the embodiment of the present application can also be used when the bandwidth and subcarrier spacing of the first reference signal are inconsistent with the bandwidth and subcarrier spacing of the second channel information to be obtained.
  • the dimension of the first channel state matrix includes at least one dimension of the antenna dimension of the terminal device, the antenna dimension of the network device, and the frequency domain dimension and the time domain dimension corresponding to the first reference signal;
  • the first sparse pattern indicates at least one of the following: an index of at least one antenna in the antenna dimension of the terminal device, an index of at least one antenna in the antenna dimension of the network device, an index of at least one frequency domain unit in the frequency domain dimension, and a time Index of at least one temporal domain cell in the domain dimension.
  • the present application further provides a communication device, where the communication device implements any method provided in the first aspect above.
  • the communication device may be realized by hardware, or may be realized by executing corresponding software by hardware.
  • the hardware or software includes one or more units or modules corresponding to the above functions.
  • the communication device includes: a processor, where the processor is configured to support the communication device to perform corresponding functions in the methods shown above.
  • the communication device may also include a memory, which may be coupled to the processor, which holds program instructions and data necessary for the communication device.
  • the communication device further includes an interface circuit, which is used to support communication between the communication device and other devices.
  • the communication apparatus may be a terminal device, or a chip or a module in the terminal device.
  • the communication device includes corresponding functional modules, respectively configured to implement the steps in the above method.
  • the functions may be implemented by hardware, or may be implemented by executing corresponding software through hardware.
  • Hardware or software includes one or more modules corresponding to the above-mentioned functions.
  • the structure of the communication device includes a processing unit and a communication unit, and these units can perform corresponding functions in the above method examples.
  • these units can perform corresponding functions in the above method examples.
  • the present application provides a method for recovering channel information, which is applicable to a scenario of AI-based channel information compression feedback.
  • the execution body of the method is a network device or a module in the network device, and the network device is used as an execution body as an example for description here.
  • the method includes: the network device sends a first reference signal to the terminal device; receives first channel information from the terminal device, and the first channel information is used to indicate a sparse channel state matrix; At least one dimension of the channel state matrix is obtained by sampling, the first channel state matrix is determined according to the first reference signal; the sparse channel state matrix is processed by the neural network to obtain the second channel state matrix, and the second channel state matrix is for the second channel state matrix A restored value of the channel state matrix; the neural network is trained on data sampled from the first sparse pattern.
  • it also includes: receiving a second reference signal from the terminal device; obtaining a third channel state matrix according to the second reference signal; obtaining a fourth channel state matrix according to the third channel state matrix, the fourth channel state
  • the length of the matrix in the subcarrier dimension is equal to the length of the first channel state matrix in the subcarrier dimension
  • the length of the fourth channel state matrix in the receiving antenna dimension is equal to the length of the first channel state matrix in the transmitting antenna dimension
  • the fourth channel state matrix is in The length of the transmitting antenna dimension is equal to the length of the first channel state matrix in the receiving antenna dimension
  • the network device samples at least one dimension of the fourth channel state matrix according to the first sparse pattern to obtain the fifth channel state matrix
  • the neural network is obtained through multiple The fourth channel state matrix and the corresponding fifth channel state matrix are obtained through training.
  • it also includes: receiving a second reference signal from the terminal device, the second reference signal is used to train the neural network corresponding to the first sparse pattern, and the neural network is used to restore the first channel according to the sparse channel state matrix state matrix.
  • the fourth precoding used for receiving the second reference signal is the same as the third precoding used for sending the first reference signal.
  • the bandwidth of the first reference signal is smaller than or equal to the bandwidth of the second reference signal.
  • the subcarrier spacing corresponding to the first reference signal is the same as the subcarrier spacing corresponding to the second reference signal.
  • the dimension of the first channel state matrix includes at least one dimension of the antenna dimension of the terminal device, the antenna dimension of the network device, and the frequency domain dimension and the time domain dimension corresponding to the first reference signal;
  • the first sparse pattern indicates at least one of the following: an index of at least one antenna in the antenna dimension of the terminal device, an index of at least one antenna in the antenna dimension of the network device, an index of at least one frequency domain unit in the frequency domain dimension, and a time Index of at least one temporal domain cell in the domain dimension.
  • the present application further provides a communication device, and the communication device implements any method provided in the third aspect above.
  • the communication device may be realized by hardware, or may be realized by executing corresponding software by hardware.
  • the hardware or software includes one or more units or modules corresponding to the above functions.
  • the communication device includes: a processor, where the processor is configured to support the communication device to perform corresponding functions in the methods shown above.
  • the communication device may also include a memory, which may be coupled to the processor, which holds program instructions and data necessary for the communication device.
  • the communication device further includes an interface circuit, which is used to support communication between the communication device and other devices.
  • the communication apparatus may be a network device, or a chip or a module in the network device.
  • the structure of the communication device includes a processing unit and a communication unit, and these units can perform corresponding functions in the above method examples.
  • these units can perform corresponding functions in the above method examples.
  • the present application provides a channel information feedback method, which is applicable to a scenario of AI-based channel information compression feedback.
  • the subject of execution of the method is a terminal device or a module in the terminal device, and the description here takes the terminal device as the subject of execution as an example.
  • the method includes: a terminal device receives a first reference signal from a network device; performs channel estimation according to the first reference signal to obtain a first channel state matrix; determines a first channel characteristic matrix according to the first channel state matrix; Sampling the first channel characteristic matrix to obtain the first sparse matrix; the second sparse pattern is configured by the network device, and the second sparse pattern is used to sample at least one dimension of the first channel characteristic matrix; sending the first channel to the network device information, the first channel information is used to indicate the first sparse matrix.
  • the similar correlation between the uplink and downlink channels can be used without the need for the terminal device to participate in the training of the neural network.
  • Restoring the complete first channel characteristic matrix by the sparse matrix can reduce the requirement on the capability of the terminal device, and avoid the feedback overhead caused by the terminal device feeding back the complete first channel characteristic matrix.
  • the method further includes: sending a second reference signal to the network device, the second reference signal is used to train a neural network corresponding to the first sparse pattern, and the neural network is used to restore the second matrix according to the first sparse matrix.
  • the second precoding used for sending the second reference signal is the same as the first precoding used for receiving the first reference signal.
  • the bandwidth of the first reference signal is smaller than or equal to the bandwidth of the second reference signal.
  • the subcarrier spacing corresponding to the first reference signal is the same as the subcarrier spacing corresponding to the second reference signal.
  • the dimension of the first channel state matrix includes at least one dimension of the antenna dimension of the terminal device, the antenna dimension of the network device, and the frequency domain dimension and the time domain dimension corresponding to the first reference signal;
  • the second sparse pattern indicates at least one of the following: an index of at least one antenna in the antenna dimension of the terminal device, an index of at least one antenna in the antenna dimension of the network device, an index of at least one frequency domain unit in the frequency domain dimension, and a time Index of at least one temporal domain cell in the domain dimension.
  • the present application further provides a communication device, where the communication device implements any method provided in the fifth aspect above.
  • the communication device may be realized by hardware, or may be realized by executing corresponding software by hardware.
  • the hardware or software includes one or more units or modules corresponding to the above functions.
  • the communication device includes: a processor, where the processor is configured to support the communication device to perform corresponding functions in the methods shown above.
  • the communication device may also include a memory, which may be coupled to the processor, which holds program instructions and data necessary for the communication device.
  • the communication device further includes an interface circuit, which is used to support communication between the communication device and other devices.
  • the communication apparatus may be a terminal device, or a chip or a module in the terminal device.
  • the structure of the communication device includes a processing unit and a communication unit, and these units can perform corresponding functions in the above method examples.
  • these units can perform corresponding functions in the above method examples.
  • the present application provides a channel information recovery method, which is applicable to a scenario of AI-based channel information compression feedback.
  • the execution body of the method is a network device or a module in the network device, and the network device is used as an execution body as an example for description here.
  • the method includes: the network device sends a first reference signal to the terminal device; receives first channel information from the terminal device, and the first channel information is used to indicate the first sparse matrix; At least one dimension of the channel feature vector matrix is obtained by sampling, the first channel feature matrix is determined through the first channel state matrix, and the first channel state matrix is determined according to the first reference signal; the first sparse matrix is processed by a neural network , to obtain a second matrix, where the second matrix is the restored value of the first channel feature matrix; the neural network is trained by using the data sampled according to the second sparse pattern.
  • it also includes: receiving a second reference signal from the terminal device, the second reference signal is used to train the neural network corresponding to the second sparse pattern, and the neural network is used to restore the second matrix according to the first sparse matrix .
  • the fourth precoding used for receiving the second reference signal is the same as the third precoding used for sending the first reference signal.
  • the bandwidth of the first reference signal is smaller than or equal to the bandwidth of the second reference signal.
  • the subcarrier spacing corresponding to the first reference signal is the same as the subcarrier spacing corresponding to the second reference signal.
  • the dimension of the first channel state matrix includes at least one dimension of the antenna dimension of the terminal device, the antenna dimension of the network device, and the frequency domain dimension and the time domain dimension corresponding to the first reference signal;
  • the second sparse pattern indicates at least one of the following: an index of at least one antenna in the antenna dimension of the terminal device, an index of at least one antenna in the antenna dimension of the network device, an index of at least one frequency domain unit in the frequency domain dimension, and a time Index of at least one temporal domain cell in the domain dimension.
  • the present application further provides a communication device, where the communication device implements any method provided in the seventh aspect above.
  • the communication device may be realized by hardware, or may be realized by executing corresponding software by hardware.
  • the hardware or software includes one or more units or modules corresponding to the above functions.
  • the communication device includes: a processor, where the processor is configured to support the communication device to perform corresponding functions in the methods shown above.
  • the communication device may also include a memory, which may be coupled to the processor, which holds program instructions and data necessary for the communication device.
  • the communication device further includes an interface circuit, which is used to support communication between the communication device and other devices.
  • the communication apparatus may be a network device, or a chip or a module in the network device.
  • the structure of the communication device includes a processing unit and a communication unit, and these units can perform corresponding functions in the above method examples.
  • these units can perform corresponding functions in the above method examples.
  • a communication device including a processor and a memory, and the memory stores computer programs or instructions; the processor is used to execute the computer programs or instructions stored in the memory, so as to realize any possible method in the implementation.
  • a communication device including a processor memory, in which computer programs or instructions are stored; the processor is used to execute the computer programs or instructions stored in the memory, to achieve any possible implementation in the aforementioned third aspect methods in methods.
  • a communication device including a processor and a memory, where computer programs or instructions are stored in the memory; the processor is used to execute the computer programs or instructions stored in the memory, so as to realize any possibility in the fifth aspect above. method in the implementation of .
  • a communication device including a processor memory, in which computer programs or instructions are stored; the processor is used to execute the computer programs or instructions stored in the memory, to implement any possible method in the implementation.
  • a communication device including a processor and an interface circuit, and optionally, a memory in which computer programs or instructions are stored; the interface circuit is used to receive information from other communication devices other than the communication device. and transmit the signal to the processor or send the signal from the processor to other communication devices other than the communication device, and the processor is used to execute the computer program or instructions stored in the memory to realize the aforementioned first aspect or The method in any possible implementation manner in the fifth aspect.
  • a communication device including a processor and an interface circuit, and optionally, a memory in which computer programs or instructions are stored; the interface circuit is used to receive information from other communication devices other than the communication device. and transmit the signal to the processor or send the signal from the processor to other communication devices other than the communication device, and the processor is used to execute the computer program or instructions stored in the memory to realize the aforementioned third aspect or The method in any possible implementation manner in the seventh aspect.
  • a computer-readable storage medium is provided, and a computer program or instruction is stored in the computer-readable storage medium, and when the computer program or instruction is run on a computer, the computer is enabled to implement the aforementioned first A method in any possible implementation of the third aspect, the fifth aspect, or the seventh aspect.
  • a computer program product storing computer-readable instructions, and when the computer-readable instructions are run on a computer, the computer is made to implement the aforementioned first aspect or the third aspect or the fifth aspect Or the method in any possible implementation manner in the seventh aspect.
  • a chip in a seventeenth aspect, includes a processor, and may also include a memory for executing computer programs or instructions stored in the memory, so that the chip system realizes the aforementioned first aspect or the third aspect or the fifth aspect Aspect or a method in any possible implementation of the seventh aspect.
  • a communication device including a processor and an interface circuit, and the interface circuit is used to receive signals from other communication devices other than the communication device and transmit them to the processor or transmit signals from the processor Sending to other communication devices other than the communication device, the processor is used to execute computer programs or instructions to implement the method in any possible implementation manner of the aforementioned first aspect or third aspect or fifth aspect or seventh aspect.
  • a communication device including a module for implementing the method in any possible implementation manner of the foregoing first aspect.
  • a communication device including a module for implementing the method in any possible implementation manner of the aforementioned third aspect.
  • a communication device including a module for implementing the method in any possible implementation manner of the foregoing fifth aspect.
  • a communication device including a module for implementing the method in any possible implementation manner of the foregoing seventh aspect.
  • a twenty-third aspect provides a communication system, where the system includes the apparatus (such as terminal equipment) described in the second aspect and the apparatus (such as network equipment) described in the fourth aspect.
  • a twenty-fourth aspect provides a communication system, where the system includes the apparatus (such as a terminal device) described in the sixth aspect and the apparatus (such as a network device) described in the eighth aspect.
  • FIG. 1 is a schematic diagram of a network architecture provided by an embodiment of the present application.
  • Fig. 2 is a schematic diagram of the layer relationship of a neural network provided by the embodiment of the present application.
  • FIG. 3 is a schematic diagram of an AI network structure applicable to the embodiment of the present application.
  • FIG. 4 is a schematic flow chart of a neural network training method provided in an embodiment of the present application.
  • FIG. 5 is a schematic structural diagram of an AI model provided by an embodiment of the present application.
  • FIG. 6 is a schematic flowchart of a method for channel information feedback and restoration provided in an embodiment of the present application.
  • FIG. 7 is a schematic flow chart of a neural network training method provided in an embodiment of the present application.
  • FIG. 8 is a schematic structural diagram of an AI model provided by an embodiment of the present application.
  • FIG. 9 is a schematic flowchart of a method for channel information feedback and restoration provided in an embodiment of the present application.
  • FIG. 10 is a schematic structural diagram of a communication device provided by an embodiment of the present application.
  • FIG. 11 is a schematic structural diagram of a communication device provided by an embodiment of the present application.
  • the terminal device can be a device with wireless transceiver function or a chip that can be set in any device, and can also be called user equipment (user equipment, UE), access terminal, user unit, user station , mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, wireless communication device, user agent, or user device.
  • the terminal device in the embodiment of the present application may be a mobile phone, a tablet computer (Pad), a computer with a wireless transceiver function, a virtual reality (virtual reality, VR) terminal, an augmented reality (augmented reality, AR) terminal, an industrial Wireless terminals in industrial control, wireless terminals in self driving, etc.
  • VR virtual reality
  • AR augmented reality
  • the network device may be a wireless access device of various standards, for example, it may be a next-generation base station (next Generation node B, gNB) in the NR system, or it may be an evolved Node B (evolved Node B, eNB), radio network controller (radio network controller, RNC) or node B (Node B, NB), base station controller (base station controller, BSC), base transceiver station (base transceiver station, BTS), home base station (such as , home evolved NodeB, or home Node B, HNB), baseband unit (baseband unit, BBU), access point (access point, AP) in wireless fidelity (wireless fidelity, WIFI) system, wireless relay node, wireless Backhaul node, transmission point (transmission and reception point, TRP or transmission point, TP), etc., can also be gNB or transmission point in 5G (NR) system, one or a group of base stations in 5G system (including multiple Antenna Panel)
  • NR next-
  • the base station in the 5G system can also be called a transmission reception point (transmission reception point, TRP) or a next-generation node B (generation Node B, gNB or gNodeB).
  • the base station in this embodiment of the present application may be an integrated base station, or may be a base station including a centralized unit (centralized unit, CU) and a distributed unit (distributed unit, DU).
  • a base station including CU and DU may also be referred to as a base station with separated CU and DU, for example, the base station includes gNB-CU and gNB-DU.
  • CU can also be separated into CU control plane (CU control plane, CU-CP) and CU user plane (CU user plane, CU-CP), such as the base station includes gNB-CU-CP, gNB-CU-UP and gNB -DU.
  • CU control plane CU control plane, CU-CP
  • CU user plane CU user plane
  • the base station includes gNB-CU-CP, gNB-CU-UP and gNB -DU.
  • the device for realizing the function of the network device may be a network device; it may also be a device capable of supporting the network device to realize the function, such as a chip system.
  • the device can be installed in the network equipment or matched with the network equipment.
  • the technical solutions provided by the embodiments of the present application are described by taking the apparatus for realizing the functions of the network equipment as network equipment and taking the network equipment as a base station as an example.
  • a special artificial intelligence (artificial intelligence, AI) AI network element or module may also be introduced into the network. If an AI network element is introduced, it corresponds to an independent network element; if an AI module is introduced, it can be located inside a certain network element, and the corresponding network element can be gNB, UE, etc.
  • AI artificial intelligence
  • FIG. 1 is a schematic diagram of an architecture of a communication system provided by an applicable embodiment of the present application.
  • the communication system includes a network device and a terminal device. Terminal devices can access network devices and communicate with network devices.
  • FIG. 1 is only a schematic diagram, and the embodiment of the present application does not limit the number of network devices and terminal devices included in the communication system.
  • the communication system may also include a node for realizing the AI function, the node may communicate with the network device, and the node may also be located in the network device and be a module in the network device.
  • the method for the terminal device to feed back downlink CSI is as follows: the network device sends the downlink reference signal to the terminal device, the terminal device performs channel estimation according to the downlink reference signal, and selects a pre-defined codebook that best matches the channel according to the channel estimation result. encoding, and feed back the selected precoding information to the network device, where the precoding information is a precoding matrix indicator (precoding matrix indicator, PMI).
  • the terminal device can also indicate the modulation and coding mode supported by the current channel judged by the terminal device through feedback of channel quality indicator (CQI), and indicate the downlink channel suggested by the terminal device through feedback of rank indicator (RI).
  • CQI channel quality indicator
  • RI rank indicator
  • the number of layers transmitted, etc., feedback information such as PMI, CQI, and RI can be used to represent the downlink CSI.
  • the complete CSI is usually very large.
  • operations such as compression and quantization are usually performed on the CSI, which will cause loss of CSI accuracy.
  • CSI feedback requires a trade-off between feedback overhead and feedback accuracy.
  • neural network neural network
  • NN neural network
  • a neural network is a specific implementation of machine learning. According to the general approximation theorem, the neural network can theoretically approximate any continuous function, so that the neural network has the ability to learn any mapping. Therefore, neural networks can accurately and abstractly model complex high-dimensional problems.
  • the idea of a neural network is derived from the neuronal structure of brain tissue. Each neuron performs a weighted sum operation on its input values, and passes the weighted sum result through an activation function to generate an output.
  • a neural network generally includes a multi-layer structure, and each layer may include one or more neurons. Increasing the depth and/or width of a neural network can improve the expressive power of the neural network, providing more powerful information extraction and abstract modeling capabilities for complex systems.
  • the depth of the neural network may refer to the number of layers included in the neural network, and the number of neurons included in each layer may be referred to as the width of the layer.
  • Figure 2 it is a schematic diagram of the layer relationship of the neural network.
  • a neural network includes an input layer and an output layer. The input layer of the neural network processes the input received by neurons, and then passes the result to the output layer, and the output layer obtains the output result of the neural network.
  • a neural network in another implementation, includes an input layer, a hidden layer, and an output layer.
  • the input layer of the neural network processes the input received by neurons, and then passes the result to the hidden layer in the middle, and the hidden layer then passes the calculation result to the output layer or the adjacent hidden layer, and finally the output layer obtains the result of the neural network. Output the result.
  • a neural network may include one or more sequentially connected hidden layers without limitation.
  • a loss function can be defined. The loss function describes the gap or difference between the output value of the neural network and the ideal target value, and the application does not limit the specific form of the loss function.
  • the training process of the neural network is to adjust the parameters of the neural network, such as the number of layers of the neural network, the width, the weight of the neuron, and/or the parameters in the activation function of the neuron, etc., so that the value of the loss function is less than the threshold threshold value Or the process of meeting the target requirements.
  • FIG. 3 it is a schematic diagram of an AI network structure applicable to the embodiment of the present application.
  • the AI network structure shown in Figure 3 is a neural network based on autoencoders (AE).
  • the encoder (encoder) and decoder (decoder) are each a neural network.
  • the encoder part of the self-encoder can be deployed on the terminal device side, and the decoder can be deployed on the network device side, or the decoder can be independent of the network device.
  • the network device can send a downlink reference signal to the terminal device.
  • the terminal equipment performs channel estimation on the downlink reference signal to obtain channel state information.
  • the encoder in the terminal device compresses the channel state information, and the terminal device sends the compressed channel state information to the decoder.
  • the decoder is used to restore the compressed channel information, and the self-encoder is trained through enough channel samples, so that the difference between the channel state information output by the decoder and the channel information input by the encoder is small enough.
  • the neural network is a mathematical model that imitates the behavior characteristics of animal neural networks and performs distributed parallel information processing, and is a special form of AI model.
  • the network architecture and business scenarios described in the embodiments of the present application are for more clearly illustrating the technical solutions of the embodiments of the present application, and do not constitute limitations on the technical solutions provided by the embodiments of the present application.
  • the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.
  • the embodiments of this application relate to model training, model deployment and application of neural networks.
  • the model training of the neural network requires data collection, and the collected sample data is used for model training.
  • the subject of data collection can be an AI network element or AI module, or a specialized data collection network element or module.
  • the network device uses the sparse uplink channel state matrix and the complete uplink channel state matrix to train the neural network model, and the collected data is ⁇ sparse uplink channel state matrix, uplink channel state matrix ⁇ , where the sparse uplink channel state matrix It can be used as a training sample, and the uplink channel state matrix can be used as a sample label.
  • the complete uplink channel state matrix is referred to as the uplink channel state matrix for short in this application.
  • the training sample refers to the data input to the neural network when training the neural network
  • the sample label refers to the expected output value when the neural network inputs the sample, which can be understood as the actual value corresponding to the training sample.
  • the purpose of the network is to input a training sample into the neural network, and its output is as close as possible to the sample label.
  • the specific method of data collection is that the terminal device sends an uplink reference signal, such as a sounding reference signal (sounding reference signal, SRS), and the network device performs channel estimation according to the received uplink reference signal, and obtains the uplink channel state matrix , and then according to the predetermined sparse pattern, some elements are extracted from the uplink channel state matrix to obtain a sparse uplink channel state matrix.
  • the sparse uplink channel state matrix and its corresponding uplink channel state matrix can be used as training data.
  • multiple training data need to be collected. The number of samples collected can be determined according to the actual situation, which is not limited in this application.
  • the downlink reference signal sent by the network device to the terminal device is called the first reference signal
  • the uplink reference signal sent by the terminal device to the network device is called the second reference signal.
  • the downlink channel state matrix determined by the reference signal is called the first channel state matrix
  • the sparse downlink channel matrix corresponding to the downlink channel state matrix is called the sparse channel state matrix
  • the sparse channel state matrix is processed through the neural network, and the obtained matrix is called is the second channel state matrix
  • the uplink channel state matrix determined by the network device according to the second reference signal is called the third channel state matrix.
  • the dimensions of the uplink state matrix include at least one dimension in ⁇ the transmitting antenna of the terminal device, the receiving antenna of the network device, the frequency domain, and the time domain ⁇ . If the length of a certain dimension is 1, then it is considered that this dimension does not exist, where the granularity in the frequency domain dimension is a frequency domain unit, such as a subcarrier or resource block (resource block, RB), and the granularity in the time domain dimension is a time domain unit, such as an orthogonal frequency domain Division multiplexing (orthogonal frequency division multiplexing, OFDM) symbols or time slots, etc.
  • the dimension of the downlink state matrix includes at least one dimension in ⁇ receiving antenna of the terminal device, transmitting antenna of the network device, frequency domain, time domain ⁇ .
  • the dimension of the uplink channel state matrix is the same as the dimension of the resources corresponding to the uplink reference signal, but the uplink channel state matrix is a matrix determined through channel estimation, and is not exactly the same as the resource actually sending the uplink reference signal in each dimension. That is to say, the resources corresponding to the uplink channel state matrix and the uplink reference signal may have the same or different specific value ranges in each dimension.
  • the resources corresponding to the uplink reference signal in the frequency domain dimension range from 0 to 6RB , but the uplink channel state matrix can range from 0 to 10 RB in the frequency domain dimension; for another example, the resource corresponding to the uplink reference signal includes two antennas in the transmit antenna dimension of the terminal device, but the uplink channel state matrix is in the transmit antenna of the terminal device The number of antennas specifically included in the dimension is not limited.
  • the dimension of the resource corresponding to the uplink reference signal includes dimensions such as an antenna for sending the uplink reference signal, an antenna for receiving the uplink reference signal, time domain, and frequency domain.
  • the dimension of the downlink channel state matrix is the same as the dimension of the resource corresponding to the downlink reference signal, but the specific value range of each dimension may be the same or different.
  • the dimension of the resource corresponding to the downlink reference signal includes dimensions such as an antenna for sending the downlink reference signal, an antenna for receiving the downlink reference signal, time domain, and frequency domain.
  • both the transmitting antenna and the receiving antenna are collectively referred to as antennas.
  • FIG. 4 it is a schematic diagram of a neural network model training process provided by the embodiment of the present application. This process involves the interaction between terminal devices, network devices, and AI entities.
  • the AI entity may be an independent network element, or a module in other devices, for example, a module in a network device.
  • the terminal device sends a second reference signal to the network device.
  • the terminal device may send multiple second reference signals to the network device, and the specific number is not limited.
  • the second reference signal is an uplink reference signal, such as an SRS, and the specific type is not limited.
  • the second reference signal may be used to train the neural network corresponding to the first sparse pattern.
  • the network device receives the second reference signal from the terminal device, and performs channel estimation according to the second reference signal, to obtain a third channel state matrix.
  • the third channel state matrix is also the uplink channel state matrix.
  • the network device can obtain the second channel state information by performing channel estimation on the second reference signal.
  • the second channel state information can indicate the channel response of the uplink channel between the network device and the terminal device.
  • the second channel state information It can be a multi-dimensional matrix, that is, the uplink channel state matrix.
  • the uplink channel state matrix corresponding to the channel response of the uplink channel may include ⁇ transmitting antenna of the terminal device, receiving antenna of the network device, frequency domain, time domain ⁇ At least one dimension, or include at least one dimension in ⁇ antenna port of the terminal device, antenna port of the network device, frequency domain, time domain ⁇ , or other dimensions.
  • the antenna port of the terminal device is an antenna port for sending the second reference signal
  • the antenna port of the network device is an antenna port for receiving the second reference signal. Therefore, in the embodiment of the present application, the uplink channel state matrix, that is, the third channel state matrix may be a multidimensional matrix.
  • the method for the network device to perform channel estimation according to the second reference signal to obtain the third channel state matrix is not limited, and may be a traditional channel estimation algorithm, for example, the minimum mean square error estimation algorithm (minimum mean square error estimation, MMSE), or a channel estimation algorithm based on a neural network.
  • MMSE minimum mean square error estimation
  • the network device determines a fifth channel state matrix according to the third channel state matrix.
  • dimension conversion may be performed on the third channel state matrix to obtain a fourth channel state matrix.
  • dimension conversion refers to operating the third channel state matrix, so that the third channel state matrix matches the dimensions of the downlink channel state matrix and the length of each dimension.
  • the fourth channel state matrix corresponds to the dimensions of the downlink state matrix, that is, the antenna dimension of the terminal device in the fourth channel state matrix corresponds to the antenna dimension of the terminal device in the downlink state matrix, and the antenna dimension of the network device corresponds to that in the downlink state matrix.
  • the antenna dimensions of network devices need to correspond, the frequency domain dimension corresponds to the frequency domain dimension, and the time domain dimension corresponds to the time domain dimension.
  • the dimension order of the fourth channel state matrix is the antenna of the terminal equipment, the antenna of the network equipment, the frequency domain, and the time domain; if the dimension order of the downlink state matrix is the antenna of the terminal equipment, the antenna of the network equipment, and the frequency domain , time domain; then the dimensions of the fourth channel state matrix and the downlink state matrix are corresponding; if the dimension order of the downlink state matrix is other cases, then the dimensions of the fourth channel state matrix and the downlink state matrix are not corresponding.
  • the lengths of the fourth channel state matrix and the downlink state matrix in each dimension are the same, that is, the length of the fourth channel state matrix in the antenna dimension of the terminal device is equal to the downlink
  • the length of the channel state matrix in the antenna dimension of the terminal device, the length of the fourth channel state matrix in the antenna dimension of the network device is equal to the length of the downlink channel state matrix in the antenna dimension of the network device, and the length of the fourth channel state matrix in the frequency
  • the length of the domain dimension is equal to the length of the downlink channel state matrix in the frequency domain dimension
  • the length of the fourth channel state matrix in the time domain dimension is equal to the length of the downlink channel state matrix in the time domain dimension.
  • the network device may extract some elements from the fourth channel state matrix according to the first sparse pattern to obtain a sparse uplink channel state matrix (also referred to as the fifth channel state matrix).
  • the sparse uplink channel state matrix includes elements from the fourth channel state matrix Extracted partial elements. That is to say, the first sparse pattern can be used to extract some elements from the fourth channel state matrix, so as to obtain the fifth channel state matrix.
  • the first sparse pattern is predetermined by the network device. Since the embodiment of the present application uses the third channel state matrix to train the neural network and applies the neural network to restore the downlink channel state matrix, the terminal device also performs the channel state matrix feedback stage.
  • the first thinning pattern is used, which is the same as the first thinning pattern used by the network device.
  • the first sparse pattern may indicate at least one dimension of the uplink channel state matrix (the third channel state matrix) or the downlink channel state matrix (the matrix obtained by the terminal device through channel estimation based on the downlink reference signal) The index of the extracted element.
  • the first sparse pattern may indicate the index of at least one antenna in the antenna dimension of the network device, the index of at least one antenna in the antenna dimension of the terminal device At least one of the index of at least one antenna, the index of at least one frequency domain unit in the frequency domain dimension corresponding to the second reference signal, and the index of at least one time domain unit in the time domain dimension corresponding to the second reference signal .
  • the number of time domain units corresponding to the second reference signal is 1, so the fourth channel state matrix is a 3-dimensional matrix of 4 ⁇ 64 ⁇ 120.
  • the first sparse pattern may be ⁇ Indexes of antennas of terminal equipment: 1,3; Indexes of antennas of network equipment: 1,5,11,...,61; Indexes of frequency domain units: 1,21,...,101 ⁇ , wherein, the antenna index of the terminal device indicates the antenna with the index 1 and 3; the antenna index of the network device indicates the antenna with the index 1,5,11,...,2n-1,...,61, and n is Positive integer; the index of the frequency domain unit, indicating the frequency domain unit whose index is 1, 21,...,20m+1,...,101, m is an integer greater than or equal to 0.
  • the first sparse pattern can be used to extract elements corresponding to antennas with indexes 1 and 3 in the fourth channel state matrix, and extract elements corresponding to antennas with indexes 1, 5, 11, ..., 61 in the fourth channel state matrix.
  • element, and extract the elements corresponding to the subcarriers with indexes 1, 21,...,101 in the fourth channel state matrix, and the elements extracted from the fourth channel state matrix form the fifth channel state matrix, that is, the sparse uplink channel state matrix , at this time the sparse uplink channel state matrix is a 3-dimensional matrix of 2 ⁇ 13 ⁇ 6.
  • S404 The network device sends the fourth channel state matrix and the fifth channel state matrix to the AI entity.
  • the network device may send the third channel state matrix and the fifth channel state matrix.
  • the AI entity receives the fourth channel state matrix and the fifth channel state matrix, and trains the neural network according to the fourth channel state matrix and the fifth channel state matrix.
  • the fifth channel state matrix can be used as a training sample in the training data, that is, the data input to the neural network when training the neural network;
  • the fourth channel state matrix can be used as the sample label in the training data, that is, the neural network expects
  • the obtained output value can be understood as the real value corresponding to the training sample.
  • the purpose of training the neural network is to input a fourth channel state matrix into the neural network, and its output is as close as possible to the fifth channel state matrix.
  • the AI entity and the network device are independent of each other as an example, and the AI entity may also be a module of the network device. If the AI entity is a module of the network device, that is, the AI entity is a part of the network device, then the network device may not send the fourth channel state matrix and the fifth channel state matrix. The fourth channel state matrix and the fifth channel state matrix may be internally delivered to the AI entity in the network device.
  • a fourth channel state matrix and its corresponding fifth channel state matrix may be used as a piece of training data.
  • AI entities can obtain multiple training data, and the specific number of training data is not limited.
  • a third channel state matrix and its corresponding fifth channel state matrix can be used as a training data, and the AI entity converts the third channel state matrix into a fourth channel state matrix, where The training data acquired by the AI entity is the third channel state matrix and the fifth channel state matrix.
  • the network device can obtain multiple training data for training the neural network, and the multiple training data can form a data set.
  • different network deployment environments can use different training data sets, for example, a factory environment uses one data set, an office environment uses another data set, or the training data corresponding to each terminal device A data set is formed, or the training data corresponding to each sparse pattern forms a data set, and different AI models are trained according to different data sets.
  • the collected training data is first clustered, similar training data forms a new data set, and the new data set is used for model training.
  • the collected data set may include an uplink channel state matrix corresponding to the second reference signal sent by the terminal device at different geographic locations.
  • the AI entity after the AI entity obtains the training data, it can select an appropriate AI model and use the training data to train the AI model.
  • the specific structure of the AI model is not limited.
  • FIG. 5 it is a schematic diagram of an AI model provided in an embodiment of the present application.
  • the input of the neural network in this AI model is a sparse uplink channel state matrix
  • the output is the restored uplink channel state matrix
  • the neural network can be seen as the arrive
  • the mapping function of The purpose of training the neural network is to output
  • a loss function needs to be defined. The selection of the loss function is related to the objective of the task. Commonly used loss functions include mean square error, cross entropy, etc.
  • the mean square error between the state matrix and the complete uplink channel state matrix, namely N is a matrix The number of elements to include.
  • the embodiment of the present application does not limit the specific process of training the AI model.
  • the stochastic gradient descent method may be used for training, and the iterative algorithm may also be used for training to obtain optimal neural network parameters.
  • the AI model can also be updated when new training data is obtained.
  • the update of the AI model can be periodic, for example, the AI entity obtains a new data set every certain period of time, and the AI model is updated based on the data set; it can also be event-triggered, when the output error exceeds a threshold, then The AI model is updated based on the latest dataset.
  • AI model can also consider the trade-off between complexity and performance. Taking the neural network as an example, when there is enough training data, the more layers and neurons in the neural network, the higher the complexity of the AI model and the better the performance. For a certain amount of training data, there may be too many parameters of the AI model, resulting in overfitting, that is, the model performs well on the training set, but poorly on the test set. Therefore, it is necessary to consider the selection of AI models in combination with actual application scenarios. For network devices with strong computing capabilities, such as macro stations, AI models with more parameters can be used; for network devices with weak computing capabilities, such as small cells and micro cells, AI models with fewer parameters can be used.
  • the network device can train multiple neural networks for different sparse patterns. For example, the network device pre-determines 3 sparse patterns, and then trains the neural network for each sparse pattern to obtain 3 neural networks, namely There is a corresponding relationship between the neural network and the sparse pattern, and the network device can choose to use one of the neural networks to restore the downlink CSI according to the actual situation.
  • the network device obtains the third channel state matrix by performing channel estimation on the received second reference signal, and converts the third channel state matrix into the fourth channel state matrix. Then extract elements of the fourth channel state matrix in at least one dimension through the first sparse pattern to obtain the fifth channel state matrix.
  • the fifth channel state matrix is composed of elements extracted from the fourth channel state matrix, which is equivalent to passing The first sparse pattern compresses the fourth channel state matrix, and the fifth channel state matrix can be regarded as a compressed matrix of the fourth channel state matrix.
  • the neural network is trained through the fourth channel state matrix and the fifth channel state matrix, so that the difference between the channel state matrix obtained after the neural network restores the fifth channel state matrix and the fourth channel state matrix small enough, for example, the mean square error between them is smaller than a preset value, that is, the recovered channel state matrix is close to the fourth channel state matrix.
  • the neural network can be used in the downlink channel to restore the downlink channel state matrix.
  • the terminal device may perform channel estimation on the received downlink reference signal to obtain a downlink channel state matrix.
  • the terminal device uses the first sparse pattern to extract elements of the downlink channel state matrix in at least one dimension to obtain a sparse downlink channel state matrix.
  • the sparse downlink channel state matrix can be regarded as a compressed downlink channel state matrix.
  • the terminal device sends the sparse downlink channel state matrix to the network device, and after the network device obtains the sparse downlink channel state matrix, the trained neural network can be used to restore the sparse downlink channel state matrix to obtain the restored downlink channel state matrix.
  • the difference between the restored downlink channel state matrix and the downlink channel state matrix estimated by the terminal equipment is small enough, for example, the mean square error between them is smaller than the preset value, that is, the restored uplink channel state matrix is close to the one estimated by the terminal equipment Channel state matrix, described in detail below.
  • FIG. 6 it is a schematic flowchart of a channel recovery method provided in the embodiment of the present application.
  • the method includes:
  • S601 The network device sends first information to the terminal device, where the first information is used to indicate a first sparse pattern.
  • S602 The terminal device receives first information from the network device.
  • a network device can send a message to a terminal device through a radio resource control (radio resource control, RRC) message or a medium access control (medium access control, MAC) control element (control element, CE) or downlink control information (DCI)
  • RRC radio resource control
  • MAC medium access control
  • CE control element
  • DCI downlink control information
  • the network device can configure one sparse pattern to the terminal device, or configure multiple sparse patterns. Activate a sparse pattern in , and the end device uses the activated sparse pattern.
  • the first sparse pattern may indicate an index of an extracted element in at least one dimension of the downlink channel state matrix
  • the downlink channel state matrix is a matrix obtained by the terminal device performing channel estimation according to the downlink reference signal. That is to say, the first sparse pattern can be used to sample at least one dimension of the downlink channel state matrix.
  • the downlink reference signal is described as the first reference signal
  • the dimension of the downlink channel state matrix includes the dimension of the antenna of the terminal device (corresponding to the antenna receiving the first reference signal), the antenna of the network device (corresponding to the antenna for sending the first reference signal) antenna) dimension and the frequency domain dimension corresponding to the first reference signal
  • the first sparse pattern may indicate the index of at least one antenna in the antenna dimension of the terminal device, the index of at least one antenna in the antenna dimension of the network device, and the first At least one item of indexes of at least one frequency domain unit in the frequency domain dimension corresponding to the reference signal.
  • the antenna of the network device is used to send the first reference signal
  • the antenna of the terminal device is used to receive the first reference signal
  • the frequency domain unit corresponding to the first reference signal is a frequency domain unit carrying the first reference signal.
  • the first sparse pattern indicates the antennas with indexes 1 and 3 among the receiving antennas of the terminal device, and indicates the antennas with indexes 1, 5, 11, ..., 2n-1, ..., 61 among the antennas that the network device sends the first reference signal Antennas, and frequency domain units with indexes 1, 21, ..., 20m+1, ..., 101, where n is a positive integer, and m is an integer greater than or equal to 0.
  • the sparse downlink channel state matrix is 2 ⁇ 13 ⁇ 6 3D matrix.
  • the dimensions of the uplink state matrix and the downlink state matrix may not be exactly the same, the dimensions of the uplink state matrix include ⁇ transmitting antenna of the terminal device, receiving antenna of the network device, frequency domain, time domain ⁇ , downlink
  • the dimensions of the state matrix include ⁇ the receiving antenna of the terminal device, the transmitting antenna of the network device, the frequency domain, and the time domain ⁇ . Therefore, it is necessary to ensure that the receiving configuration of the terminal device matches the sending configuration, and that the receiving configuration and the sending configuration of the network device match.
  • the receiving configuration includes at least one of receiving antenna, receiving right, and receiving bandwidth.
  • the receiving right can also be called receiving beam or receiving precoding or receiving spatial filter.
  • the sending configuration includes at least one of sending antenna, sending right, and sending bandwidth.
  • the transmission right can also be called a transmission beam or a transmission precoding or a transmission spatial filter, wherein the configuration of the receiving antenna and the transmitting antenna can be established by establishing a first association between the transmitting antenna of the terminal device and the receiving antenna of the terminal device, and /or the second association relationship between the transmitting antenna of the network device and the receiving antenna of the network device is realized.
  • the first association relationship and the second association relationship can be predefined by the protocol, configured by the network device, or preset by the factory. Definition implementation, for example, the first association relationship is determined by the terminal device according to its own internal configuration.
  • the terminal's transmitting antenna 1 corresponds to receiving antenna 1
  • transmitting antenna 3 corresponds to receiving antenna 2.
  • the terminal device can report the first association relationship to the network device.
  • the second association relationship is notified by the network device to the terminal device.
  • the matching of the receiving right and the sending right means that the terminal device and the network device use the same weight when receiving and sending the reference signal.
  • Receiving bandwidth and sending bandwidth refer to the fact that terminal devices and network devices use the same bandwidth when receiving and sending reference signals, or make the bandwidth of the uplink channel greater than or equal to the bandwidth of the downlink channel.
  • the downlink channel state matrix is called the first channel state matrix
  • the sparse downlink channel matrix corresponding to the downlink channel state matrix is called the sparse channel state matrix
  • the network device sends the first reference signal to the terminal device, and correspondingly, the terminal device receives the first reference signal from the network device.
  • the first reference signal may be a downlink reference signal, and the network device may send multiple first reference signals to the terminal device, which is not limited in this embodiment of the present application.
  • the receiving beam used by the terminal device when receiving the first reference signal is the same as the transmitting beam used when sending the second reference signal, that is, the spatial domain filter used by the terminal device when receiving the first reference signal
  • the filter is the same as the spatial domain filter used for sending the second reference signal.
  • the second precoding used by the terminal device to send the second reference signal is the same as the first precoding used by the terminal device to receive the first reference signal.
  • the receiving beam used by the network device to receive the second reference signal is the same as the transmit beam used when sending the first reference signal, that is, the third precoding used by the network device to send the first reference signal is the same as the third precoding used by the network device to receive the second reference signal.
  • the fourth precoding used by the signals is the same.
  • the second reference signal received by the network device can be expressed as in is the fourth precoding used when the network device receives the second reference signal
  • the second precoding used when sending the second reference signal for the terminal equipment if and is the identity matrix, it means that precoding is not used
  • H ul is the uplink channel matrix
  • s ul is the second reference signal
  • n is noise
  • the uplink channel state matrix estimated by the network equipment according to the second reference signal is That is, the uplink channel state matrix used when training the neural network is an equivalent channel matrix including precoding.
  • the first reference signal received by the terminal device may be expressed as in is the first precoding used when the terminal device receives the first reference signal, are respectively the third precoding used when the network equipment sends the first reference signal, H dl is the downlink channel matrix, s dl is the first reference signal, and the first channel state matrix estimated by the terminal equipment according to the first reference signal (ie, the uplink channel state matrix) is There is a similar correlation between the uplink channel matrix H ul and the downlink channel matrix H dl , when and hour, and There is also a correlation between them, that is, the second precoding is the same as the first precoding, and the third precoding is the same as the fourth precoding, that is, the terminal device transmits the transmission beam of the second reference signal and receives the second reference signal.
  • the method for the terminal device to perform channel estimation is not limited in the present application, and may be a traditional channel estimation algorithm, for example, the MMSE algorithm, or a channel estimation algorithm based on a neural network.
  • the terminal device determines a sparse channel state matrix according to the first reference signal and the first sparse pattern.
  • the terminal device may obtain the first channel state matrix according to the first reference signal. Specifically, the terminal device may perform channel estimation according to the first reference signal to obtain the first channel state matrix.
  • the first channel state information may indicate the channel response of the downlink channel between the network device and the terminal device.
  • the first channel state information may be a multidimensional matrix.
  • the dimensions of the first channel state matrix include the antenna dimension of the terminal device, the network device At least one dimension in the antenna dimension of the antenna dimension and the subcarrier dimension corresponding to the first reference signal.
  • the terminal device may obtain a sparse channel state matrix according to the first sparse pattern and the first channel state matrix, for example, extract the first channel state matrix according to the first sparse pattern to obtain a sparse channel state matrix, or obtain the sparse channel state matrix according to the first sparse pattern.
  • a channel state matrix is extracted, and elements not included in the first channel state matrix are estimated to obtain a sparse channel state matrix.
  • the sparse channel state matrix is equivalent to a compressed matrix of the first channel state matrix.
  • the network device does not need to send the first information to the terminal device, that is, the network device does not need to send the first sparse pattern to the terminal device, then the terminal device may consider that obtaining the first channel state matrix according to the first reference signal is also Sparse channel state matrix. For example, when the network device sends the first reference signal, the transmit antenna and the frequency domain resources used have been consistent with the predetermined first sparse pattern, then the terminal device estimates the The channel is the sparse channel state matrix.
  • the terminal device may use the antenna corresponding to the antenna index of the terminal device indicated by the first sparse pattern to receive the first reference signal, and/or correspond to the antenna index of the network device indicated by the first sparse pattern
  • the channel on the frequency domain resource corresponding to the antenna and the frequency domain index is estimated, and the sparse channel state matrix is obtained directly.
  • the frequency domain unit in the frequency domain is a subcarrier, assuming that the first sparse pattern indicates the antennas with indexes 1 and 3 among the antennas of the terminal device, and indicates that the indexes of the antennas of the network device are 1, 5, 11,...,2n Antennas of -1,...,61, and subcarriers with indexes 1, 21,...,20m+1,...,101, where n is a positive integer, and m is an integer greater than or equal to 0.
  • the terminal device when receiving the first reference signal, uses the antennas with indexes 1 and 3, and estimates the subcarriers with indexes 1, 21,...,101 and the subcarriers with indexes 1, 5, 11,..., 2n in the network device- 1,...,61 antennas corresponding to the channel, the terminal device can obtain a sparse channel state matrix with a dimension of 2 ⁇ 13 ⁇ 6.
  • the length of the first channel state matrix in the subcarrier dimension is equal to the length of the fourth channel state matrix in the subcarrier dimension
  • the length of the first channel state matrix in the receiving antenna dimension is equal to the length of the fourth channel state matrix in the transmitting antenna.
  • the length of the dimension, the length of the first channel state matrix in the dimension of the transmitting antenna is equal to the length of the dimension of the fourth channel state matrix in the dimension of the receiving antenna.
  • the terminal device sends first channel information to the network device, and correspondingly, the network device receives the first channel information from the terminal device.
  • the first channel information is used to indicate a sparse channel state matrix.
  • the first channel information may be a set of values of elements in the sparse channel state matrix, or may be in other forms, which is not limited in this embodiment of the present application.
  • the network device processes the sparse channel state matrix indicated by the first channel information through the neural network to obtain a second channel state matrix.
  • the second channel state matrix is the restored value of the first channel state matrix
  • the neural network is trained by the data sampled according to the first sparse pattern, specifically refer to the description in the flow shown in FIG. 4 above.
  • the network device may determine parameters such as precoding and resources used when sending data to the terminal device according to the restored second channel state matrix.
  • the network device needs to apply the neural network to restore the first channel state matrix, so if the training of the neural network is completed on the network device or the AI module located in the network device, it is not necessary to configure the trained neural network in the network device; If the training of the neural network is completed on an independent AI network element, it is necessary to configure the trained neural network into the network device.
  • the specific configuration process is not limited in this embodiment of the present application, and will not be repeated here.
  • the network device can only train the neural network through the uplink channel state matrix without requiring the terminal device to participate in the neural network training, and restore the complete downlink channel through the neural network and the sparse downlink channel state matrix fed back by the terminal device
  • the state matrix can reduce the requirement on the capability of the terminal equipment, and avoid the feedback overhead caused by the terminal equipment feeding back the complete downlink channel state matrix.
  • the embodiment of the present application also provides a method, which can be used for the terminal device to feed back the channel feature matrix, which will be described in detail below.
  • the downlink reference signal sent by the network device to the terminal device is called the first reference signal
  • the uplink reference signal sent by the terminal device to the network device is called the second reference signal
  • the downlink channel state matrix determined according to the first reference signal is called the first matrix
  • the channel characteristic matrix determined according to the downlink channel state matrix is called the first channel characteristic matrix
  • the sparse downlink channel characteristic matrix corresponding to the first channel characteristic matrix It is called the first sparse matrix
  • the matrix obtained by processing the first sparse matrix through the neural network is called the second matrix
  • the uplink channel state matrix determined by the network equipment according to the second reference signal is called the third matrix, and will be based on
  • the channel characteristic matrix determined by the uplink channel state matrix is called the second channel characteristic matrix
  • the matrix obtained by extracting elements from the second channel characteristic matrix is called the second
  • the dimension of the uplink state matrix includes at least one dimension in ⁇ the transmitting antenna of the terminal equipment, the receiving antenna of the network equipment, the frequency domain, the time domain ⁇ , if the length of a certain dimension is 1, it can also be considered that this dimension is not Existence, wherein the granularity of frequency domain can be subcarrier or RB, and the granularity of time domain can be OFDM symbol or time slot, etc.
  • the dimension of the downlink state matrix includes at least one dimension in ⁇ receiving antenna of the terminal device, transmitting antenna of the network device, frequency domain, time domain ⁇ .
  • FIG. 7 it is a schematic diagram of a neural network model training process provided by the embodiment of the present application. This process involves the interaction between terminal devices, network devices, and AI entities.
  • the AI entity may be an independent network element, or may be located in a network device.
  • the specific method of data collection is that the terminal device sends a second reference signal, such as SRS, and the network device performs channel estimation according to the received second reference signal to obtain a third matrix. Obtaining the second channel characteristic matrix according to the third matrix, and then extracting some elements from the second channel characteristic matrix according to the predetermined second sparse pattern to obtain the second sparse matrix.
  • the second sparse matrix and the second channel characteristic matrix can be used as one training data, and multiple training data need to be collected in the data collection stage, and the number of collected samples can be determined according to the actual situation, which is not limited in this application.
  • the terminal device sends a second reference signal to the network device, and correspondingly, the network device receives the second reference signal from the terminal device.
  • the terminal device may send multiple second reference signals to the network device, and the specific number is not limited.
  • the second reference signal is an uplink reference signal, such as an SRS, and the specific type is not limited.
  • the network device performs channel estimation according to the second reference signal, and obtains uplink channel state information.
  • the network device performs channel estimation on each second reference signal to obtain uplink channel state information
  • the uplink channel state information may indicate the channel response of the uplink channel between the network device and the terminal device
  • the dimension of the uplink channel state information may include ⁇ terminal device
  • the antenna of the network device refers to the receiving antenna for receiving the second reference signal
  • the subcarrier refers to the subcarrier corresponding to the second reference signal.
  • the third matrix is 64 ⁇ 120 2D matrix of .
  • the network device determines a second sparse matrix according to the uplink channel state information.
  • the network device may determine the second airspace covariance matrix according to the uplink channel state information.
  • the second spatial covariance matrix can be calculated by an uplink channel, for example, the second spatial covariance matrix can satisfy the following form:
  • R i represents the spatial covariance matrix corresponding to each terminal antenna, frequency domain unit, and time domain unit
  • H i represents the channel vector corresponding to each terminal antenna, frequency domain unit, and time domain unit.
  • the dimension of the obtained second airspace covariance matrix is a*c*network equipment antenna*network equipment antenna , and take the average of dimensions a, b, and c, then the dimension of the obtained second airspace covariance matrix is network device antenna*network device antenna.
  • the second spatial covariance matrix may also be obtained by calculating multiple uplink channel state matrices, for example, by averaging the spatial covariance matrices of multiple uplink channel state matrices within a period of time, which is not limited in this embodiment of the present application.
  • the network device may perform eigenvalue decomposition (eigen value decomposition, EVD) on the second airspace covariance matrix to obtain the second channel characteristic matrix, or may directly perform singular value decomposition (singular value decomposition, SVD) on the uplink channel state matrix , to obtain the second channel characteristic matrix.
  • eigenvalue decomposition eigen value decomposition, EVD
  • singular value decomposition singular value decomposition
  • SVD singular value decomposition
  • the dimension of the second channel eigenmatrix can be the receiving antenna of the network device * the receiving antenna of the network device, or also include ⁇ transmitting antenna of the terminal device, frequency domain, At least one dimension in the time domain ⁇ .
  • the dimension of the second channel eigenmatrix may be the receiving antenna*n of the network device, or also include ⁇ the transmitting antenna of the terminal device, frequency domain, At least one dimension in the time domain ⁇ .
  • the network device may extract some elements from the second channel feature matrix according to the second sparse pattern to obtain a second sparse matrix, where the second sparse matrix includes some elements extracted from the second channel feature matrix. That is to say, the second sparse pattern can be used to extract some elements from the second channel characteristic matrix, so as to obtain the second sparse matrix.
  • the second sparse pattern is predetermined by the network device, and the terminal device also uses the second sparse pattern in the channel state matrix feedback stage, which is the same as the second sparse pattern used by the network device.
  • the second sparse pattern when used to extract elements in at least one dimension of the second channel characteristic matrix, may indicate the index of at least one antenna in the antenna dimension of the network device, An index of at least one antenna in the antenna dimension of the terminal device, an index of at least one frequency domain unit in the frequency domain dimension corresponding to the second reference signal, and an index of at least one time domain unit in the time domain dimension corresponding to the second reference signal At least one item in the index.
  • the network device receives the second reference signal
  • the terminal device sends the second reference signal
  • the uplink channel state matrix corresponding to the uplink channel state information is a 64 ⁇ 120 2-dimensional matrix.
  • the second spatial covariance matrix determined according to the uplink channel state matrix is also a 64 ⁇ 120 2-dimensional matrix
  • the second channel characteristic matrix determined according to the second spatial covariance matrix is also a 64 ⁇ 120 2-dimensional matrix.
  • the second sparse pattern indicates the antennas whose indices are 1, 5, 11, ..., 2n-1, ..., 61 in the network device, n is a positive integer, and the second sparse pattern indicates that the indices are 1, 21, ..., 20m+1 ,...,101 frequency domain units, m is an integer greater than or equal to 0.
  • the second sparse pattern can be used to extract the elements corresponding to the antennas with indices 1, 5, 11, ..., 61 in the second channel feature matrix, and extract the elements corresponding to the antennas with indices 1, 21, ..., 101 in the third matrix
  • the elements corresponding to the frequency domain unit of , and the elements extracted from the second channel feature matrix form a second sparse matrix, and at this time, the second sparse matrix is a 13 ⁇ 6 2-dimensional matrix.
  • the network device sends the second channel characteristic matrix and the second sparse matrix to the AI entity, and correspondingly, the AI entity receives the second channel characteristic matrix and the second sparse matrix.
  • the AI entity trains the neural network according to the second channel feature matrix and the second sparse matrix.
  • the AI entity and the network device are independent of each other as an example, and the AI entity may also be a module of the network device. If the AI entity is a module of the network device, that is, the AI entity is a part of the network device, the network device may not send the second channel characteristic matrix and the second sparse matrix. The second channel feature matrix and the second sparse matrix may be internally delivered to the AI entity in the network device.
  • one second channel feature matrix and its corresponding second sparse matrix can be used as one training data, and the AI entity can obtain multiple training data, and the specific number of training data is not limited.
  • the second sparse matrix can be used as a training sample in the training data, that is, the data input to the neural network when training the neural network; the second channel feature matrix can be used as the sample label in the training data, that is, the neural network expects to obtain
  • the output value of can be understood as the real value corresponding to the training sample.
  • the purpose of training the neural network is to input a second sparse matrix into the neural network, and its output is as close as possible to the second channel feature matrix.
  • the network device can obtain multiple training data for training the neural network, and multiple training data can form a data set.
  • different network deployment environments can use different training data sets, for example, a factory environment uses one data set, an office environment uses another data set, or the training data corresponding to each terminal device A data set is formed, or the training data corresponding to each sparse pattern forms a data set, and different AI models are trained according to different data sets.
  • the collected training data is first clustered, similar training data forms a new data set, and the new data set is used for model training.
  • the collected data set may include an uplink channel state matrix corresponding to the second reference signal sent by the terminal device at different geographic locations.
  • the AI entity after the AI entity obtains the training data, it can select an appropriate AI model and use the training data to train the AI model.
  • the specific structure of the AI model is not limited.
  • FIG. 8 it is a schematic diagram of an AI model provided in an embodiment of the present application.
  • the second sparse matrix input to the neural network is The output fourth matrix is
  • the neural network can be seen as the arrive
  • the mapping function of The purpose of training the neural network is to output with the actual second channel eigenmatrix as close as possible.
  • a loss function can be defined for training the neural network, and the loss function can be the mean square error between the output fourth matrix and the actual second channel feature matrix, namely N is a matrix The number of elements to include.
  • the embodiment of the present application does not limit the specific process of training the AI model.
  • the stochastic gradient descent method may be used for training, and the iterative algorithm may also be used for training to obtain optimal neural network parameters.
  • the AI model can also be updated when new training data is obtained.
  • the update of the AI model can be periodic, for example, the AI entity obtains a new data set every certain period of time, and the AI model is updated based on the data set; it can also be event-triggered, when the error output by the neural network exceeds a threshold , the AI model is updated based on the latest dataset.
  • AI model can also consider the trade-off between complexity and performance. Taking the neural network as an example, when there is enough training data, the more layers and neurons in the neural network, the higher the complexity of the AI model and the better the performance. For a certain amount of training data, there may be too many parameters of the AI model, resulting in overfitting, that is, the model performs well on the training set, but poorly on the test set. Therefore, it is necessary to consider the selection of AI models in combination with actual application scenarios. For network devices with strong computing capabilities, such as macro stations, AI models with more parameters can be used; for network devices with weak computing capabilities, such as small cells and micro cells, AI models with fewer parameters can be used.
  • the network device can train multiple neural networks for different sparse patterns. For example, the network device pre-determines 3 sparse patterns, and then trains the neural network for each sparse pattern to obtain 3 neural networks, namely There is a corresponding relationship between the neural network and the sparse pattern, and the network device can choose to use one of the neural networks to restore the downlink CSI according to the actual situation.
  • the neural network after the neural network in the AI model is trained through the second sparse pattern, the neural network can be used in the downlink channel to recover the feature matrix of the first channel.
  • the terminal device may perform channel estimation on the received downlink reference signal to obtain the first matrix.
  • the terminal device determines the first channel feature matrix according to the first matrix, and extracts elements of the first channel feature matrix in at least one dimension through the second sparse pattern to obtain the first sparse matrix.
  • the first sparse matrix can be considered as the first channel feature The matrix after matrix compression.
  • the terminal device sends the first sparse matrix to the network device, and after the network device obtains the first sparse matrix, the trained neural network may be used to restore the first sparse matrix to obtain the second matrix.
  • the mean square error between the second matrix and the first channel characteristic matrix estimated by the terminal device is smaller than a preset value, that is, the second matrix is close to the first channel characteristic matrix, which will be described in detail below.
  • FIG. 9 it is a schematic flowchart of a channel recovery method provided in the embodiment of the present application.
  • the method includes:
  • S901 The network device sends the second information to the terminal device, and correspondingly, the terminal device receives the second information from the network device.
  • the second information is used to indicate the second sparse pattern.
  • the network device may send the second information to the terminal device through an RRC message or MAC CE or DCI, which is not limited in this embodiment of the present application.
  • the network device can configure multiple sparse patterns to the terminal device, and the terminal device can select a sparse pattern from multiple sparse patterns to use, or the network device can activate a sparse pattern from multiple sparse patterns, and the terminal The device uses the active sparse pattern.
  • the second sparse pattern may be used to sample at least one dimension of the first channel characteristic matrix.
  • the second sparse pattern may indicate the network device's At least one item of an index of at least one antenna in the antenna dimension and an index of at least one frequency domain unit in the frequency domain dimension corresponding to the first reference signal.
  • the dimensions of the uplink state matrix and the downlink state matrix may not be exactly the same, the dimensions of the uplink state matrix include ⁇ transmitting antenna of the terminal device, receiving antenna of the network device, frequency domain, time domain ⁇ , downlink
  • the dimensions of the state matrix include ⁇ the receiving antenna of the terminal device, the transmitting antenna of the network device, the frequency domain, and the time domain ⁇ . Therefore, it is necessary to ensure that the receiving configuration of the terminal device matches the sending configuration, and that the receiving configuration and the sending configuration of the network device match.
  • the receiving configuration includes at least one of receiving antenna, receiving right, and receiving bandwidth.
  • the receiving right can also be called receiving beam or receiving precoding or receiving spatial filter.
  • the sending configuration includes at least one of sending antenna, sending right, and sending bandwidth.
  • the transmission right can also be called a transmission beam or a transmission precoding or a transmission spatial filter, wherein the configuration of the receiving antenna and the transmitting antenna can be established by establishing a first association between the transmitting antenna of the terminal device and the receiving antenna of the terminal device, and /or the second association relationship between the transmitting antenna of the network device and the receiving antenna of the network device is realized.
  • the first association relationship and the second association relationship can be predefined by the protocol, configured by the network device, or preset by the factory. Definition implementation, for example, the first association relationship is determined by the terminal device according to its own internal configuration.
  • the terminal's transmitting antenna 1 corresponds to receiving antenna 1
  • transmitting antenna 3 corresponds to receiving antenna 2.
  • the terminal device can report the first association relationship to the network device.
  • the second association relationship is notified by the network device to the terminal device.
  • the matching of the receiving right and the sending right means that the terminal device and the network device use the same weight when receiving and sending the reference signal.
  • Receiving bandwidth and sending bandwidth refer to the fact that terminal devices and network devices use the same bandwidth when receiving and sending reference signals, or make the bandwidth of the uplink channel greater than or equal to the bandwidth of the downlink channel.
  • the network device sends the first reference signal to the terminal device, and correspondingly, the terminal device receives the first reference signal from the network device.
  • the first reference signal may be a downlink reference signal, and the network device may send multiple first reference signals to the terminal device, which is not limited in this embodiment of the present application.
  • the receiving beam used by the terminal device when receiving the first reference signal is the same as the transmitting beam used when sending the second reference signal, that is, the spatial domain filter used by the terminal device when receiving the first reference signal
  • the filter is the same as the spatial domain filter used for sending the second reference signal.
  • the second precoding used by the terminal device to send the second reference signal is the same as the first precoding used by the terminal device to receive the first reference signal.
  • the receiving beam used by the network device to receive the second reference signal is the same as the transmit beam used when sending the first reference signal, that is, the third precoding used by the network device to send the first reference signal is the same as the third precoding used by the network device to receive the second reference signal.
  • the fourth precoding used by the signals is the same.
  • the method for the terminal device to perform channel estimation is not limited in the present application, and may be a traditional channel estimation algorithm, for example, a channel estimation algorithm, or a channel estimation algorithm based on a neural network.
  • the terminal device determines a first matrix according to the first reference signal, determines a first channel characteristic matrix according to the first matrix, and determines a first sparse matrix according to the first channel characteristic matrix and the second sparse pattern.
  • the terminal device may perform channel estimation according to the first reference signal to obtain the first matrix.
  • the terminal device may determine the first spatial covariance matrix according to the first matrix, and perform singular value decomposition on the first spatial covariance matrix to obtain the first channel feature matrix.
  • reference may be made to the description of the second spatial covariance matrix, which will not be repeated here.
  • the terminal device may extract some elements from the first channel feature matrix according to the second sparse pattern to obtain a first sparse matrix, where the first sparse matrix includes some elements extracted from the first channel feature matrix. That is to say, the second sparse pattern can be used to extract some elements from the first channel characteristic matrix, so as to obtain the first sparse matrix.
  • the first sparse matrix is equivalent to a compressed matrix of the first channel characteristic matrix.
  • the terminal device sends first channel information to the network device, and correspondingly, the network device receives the first channel information from the terminal device.
  • the first channel information is used to indicate a sparse channel state matrix.
  • the first channel information may be a set of values of elements in the sparse channel state matrix, or a set of quantized values of elements in the sparse channel state matrix, or in other forms, which is not limited in this embodiment of the present application. .
  • the network device processes the first sparse matrix through a neural network to obtain a second matrix.
  • the second matrix is the restored value of the first channel characteristic matrix
  • the neural network is trained by the data sampled according to the second sparse pattern, specifically refer to the description in the process shown in FIG. 4 above.
  • the network device may determine parameters such as PMI and RI used when sending data to the terminal device according to the restored second matrix.
  • the network device needs to apply the neural network to restore the first channel feature matrix, so if the training of the neural network is completed on the network device or the AI module located in the network device, it is not necessary to configure the trained neural network in the network device; If the training of the neural network is completed on an independent AI network element, it is necessary to configure the trained neural network into the network device.
  • the specific configuration process is not limited in this embodiment of the present application, and will not be repeated here.
  • the network device can only train the neural network through the second sparse matrix and the second channel feature matrix without the need for the terminal device to participate in the training of the neural network, which can reduce the requirements for the capability of the terminal device and avoid feedback from the terminal device
  • the feedback overhead caused by the complete first channel eigenmatrix caused by the complete first channel eigenmatrix.
  • the bandwidth of the first reference signal may be smaller than or equal to the bandwidth of the second reference signal.
  • the subcarrier spacing corresponding to the first reference signal is greater than or equal to the subcarrier spacing corresponding to the second reference signal.
  • the network device can configure a set of special first reference signals for the terminal device, the bandwidth of the first reference signal is less than or equal to the bandwidth of the second reference signal, and the subcarrier spacing of the first reference signal is greater than or equal to the subcarrier spacing of the second reference signal.
  • the bandwidth and subcarrier spacing of the first reference signal can be independently configured by the network device, that is, the bandwidth of the first reference signal can be greater than or equal to the bandwidth of the uplink active bandwidth part (bandwidth part, BWP) of the terminal device , and/or the subcarrier spacing of the first reference signal has nothing to do with the subcarrier spacing of the uplink active BWP of the terminal device, that is, the subcarrier spacing of the first reference signal and the subcarrier spacing of the uplink active BWP of the terminal device may not be equal, and the uplink The subcarrier spacing of the area including the first reference signal in the active BWP may be different from the subcarrier spacing of the area not including the first reference signal in the uplink active BWP, where the first reference signal may be carried in the uplink active BWP.
  • the first reference signal can be divided into multiple segments in the frequency domain, as long as the total bandwidth of the multiple segments meets the requirements of the network device for recovering downlink channel information.
  • the solution of the embodiment of the present application can also be used when the bandwidth and subcarrier spacing of the first reference signal are inconsistent with the bandwidth and subcarrier spacing of the second reference signal. It can be used, so that the network equipment can use the neural network to restore the complete downlink channel information.
  • the above-mentioned embodiments can be implemented independently, or can also be implemented in combination with each other.
  • the differences between the various embodiments are emphasized. Except for the differences, other contents between different embodiments may refer to each other.
  • the step numbers of the various flow charts described in the above embodiments are only an example of the execution flow, and do not constitute a restriction on the sequence of execution of the steps. In the embodiments of the present application, there is no time sequence dependency between the steps that are not related to each other. Strict order of execution.
  • not all the steps illustrated in each flow chart are steps that must be executed, and some steps may be added or deleted on the basis of each flow chart according to actual needs.
  • the network device, terminal device or the above-mentioned communication device may include a hardware structure and/or a software module in the form of a hardware structure, a software module, or a hardware structure plus a software module. Realize the above functions. Whether one of the above-mentioned functions is executed in the form of a hardware structure, a software module, or a hardware structure plus a software module depends on the specific application and design constraints of the technical solution.
  • each functional module in each embodiment of the present application may be integrated into one processor, or physically exist separately, or two or more modules may be integrated into one module.
  • the above-mentioned integrated modules can be implemented in the form of hardware or in the form of software function modules.
  • the embodiment of the present application further provides a communication device 1000 .
  • the communication apparatus 1000 may be the network device in FIG. 1 , and is configured to implement the method for the network device in the foregoing method embodiments.
  • the communication device may also be the core network device in FIG. 1 , and is used to implement the method corresponding to the core network device in the foregoing method embodiments.
  • the communication device 1000 may include: a processing unit 1001 and a communication unit 1002 .
  • the communication unit may also be referred to as a transceiver unit, and may include a sending unit and/or a receiving unit, respectively configured to perform the sending and receiving steps of the network device or the terminal device in the method embodiments above.
  • the communication device provided by the embodiment of the present application will be described in detail with reference to FIG. 10 to FIG. 11 .
  • the behaviors and functions of the network device in the foregoing method embodiments may be implemented by the communication apparatus 1000, for example, implementing the methods performed by the network device in the embodiments shown in FIG. 4 to FIG. 6 .
  • the communication apparatus 1000 may be a network device, or a component (such as a chip or a circuit) applied in the network device, or a chip or a chipset in the network device, or a part of the chip for performing related method functions.
  • the communication unit 1002 may be used to perform the receiving or sending operation performed by the network device in the embodiments shown in FIG. 4 to FIG. 6
  • the processing unit 1001 may be used to perform An operation performed by a device other than sending and receiving operations.
  • a communication unit configured to receive a first reference signal from a network device
  • a processing unit configured to obtain a first channel state matrix according to a first reference signal; sample the first channel state matrix according to a first sparse pattern to obtain a sparse channel state matrix; the first sparse pattern is configured by a network device, and the first sparse pattern For sampling at least one dimension of the first channel state matrix;
  • a communication unit configured to send first channel information to the network device, where the first channel information is used to indicate a sparse channel state matrix.
  • the communication unit is also used to: send a second reference signal to the network device, the second reference signal is used to train the neural network corresponding to the first sparse pattern, and the neural network is used to restore the first sparse pattern according to the sparse channel state matrix.
  • a channel state matrix A channel state matrix.
  • the second precoding used for sending the second reference signal is the same as the first precoding used for receiving the first reference signal.
  • the bandwidth of the first reference signal is smaller than or equal to the bandwidth of the second reference signal.
  • the subcarrier spacing corresponding to the first reference signal is the same as the subcarrier spacing corresponding to the second reference signal.
  • the dimension of the first channel state matrix includes at least one dimension of the antenna dimension of the terminal device, the antenna dimension of the network device, and the frequency domain dimension and the time domain dimension corresponding to the first reference signal;
  • the first sparse pattern indicates at least one of the following: an index of at least one antenna in the antenna dimension of the terminal device, an index of at least one antenna in the antenna dimension of the network device, an index of at least one frequency domain unit in the frequency domain dimension, and a time Index of at least one temporal domain cell in the domain dimension.
  • a processing unit configured to send the first reference signal to the terminal device through the communication unit
  • the processing unit is configured to receive the first channel information from the terminal device through the communication unit, the first channel information is used to indicate the sparse channel state matrix; the sparse channel state matrix is at least one dimension of the first channel state matrix through the first sparse pattern Obtained by sampling, the first channel state matrix is determined according to the first reference signal; the sparse channel state matrix is processed through the neural network to obtain the second channel state matrix, and the second channel state matrix is the restored value of the first channel state matrix ; the neural network is trained on data sampled from the first sparse pattern.
  • the communication unit is further configured to: receive a second reference signal from the terminal device;
  • the processing unit is also used to: obtain a third channel state matrix according to the second reference signal; obtain a fourth channel state matrix according to the third channel state matrix, the length of the fourth channel state matrix in the subcarrier dimension is equal to the length of the first channel state matrix in the subcarrier
  • the length of the carrier dimension, the length of the fourth channel state matrix in the receiving antenna dimension is equal to the length of the first channel state matrix in the transmitting antenna dimension, the length of the fourth channel state matrix in the transmitting antenna dimension is equal to the length of the first channel state matrix in the receiving antenna dimension length
  • the network device samples at least one dimension of the fourth channel state matrix according to the first sparse pattern to obtain the fifth channel state matrix
  • the neural network is obtained by training multiple fourth channel state matrices and corresponding fifth channel state matrices of.
  • the communication unit is also used to: receive a second reference signal from the terminal device, the second reference signal is used to train the neural network corresponding to the first sparse pattern, and the neural network is used to restore The first channel state matrix.
  • the fourth precoding used for receiving the second reference signal is the same as the third precoding used for sending the first reference signal.
  • the bandwidth of the first reference signal is smaller than or equal to the bandwidth of the second reference signal.
  • the subcarrier spacing corresponding to the first reference signal is the same as the subcarrier spacing corresponding to the second reference signal.
  • the dimension of the first channel state matrix includes at least one dimension of the antenna dimension of the terminal device, the antenna dimension of the network device, and the frequency domain dimension and the time domain dimension corresponding to the first reference signal;
  • the first sparse pattern indicates at least one of the following: an index of at least one antenna in the antenna dimension of the terminal device, an index of at least one antenna in the antenna dimension of the network device, an index of at least one frequency domain unit in the frequency domain dimension, and a time Index of at least one temporal domain cell in the domain dimension.
  • a communication unit configured to receive a first reference signal from a network device
  • the processing unit is configured to perform channel estimation according to the first reference signal to obtain the first channel state matrix; determine the first channel characteristic matrix according to the first channel state matrix; sample the first channel characteristic matrix according to the second sparse pattern to obtain the second channel characteristic matrix A sparse matrix; the second sparse pattern is configured by the network device, and the second sparse pattern is used to sample at least one dimension of the first channel feature matrix;
  • a communication unit configured to send first channel information to the network device, where the first channel information is used to indicate the first sparse matrix.
  • the communication unit is also used to: send a second reference signal to the network device, the second reference signal is used to train the neural network corresponding to the first sparse pattern, and the neural network is used to restore the first sparse pattern according to the first sparse matrix. Second matrix.
  • the second precoding used for sending the second reference signal is the same as the first precoding used for receiving the first reference signal.
  • the bandwidth of the first reference signal is smaller than or equal to the bandwidth of the second reference signal.
  • the subcarrier spacing corresponding to the first reference signal is the same as the subcarrier spacing corresponding to the second reference signal.
  • the dimension of the first channel state matrix includes at least one dimension of the antenna dimension of the terminal device, the antenna dimension of the network device, and the frequency domain dimension and the time domain dimension corresponding to the first reference signal;
  • the second sparse pattern indicates at least one of the following: an index of at least one antenna in the antenna dimension of the terminal device, an index of at least one antenna in the antenna dimension of the network device, an index of at least one frequency domain unit in the frequency domain dimension, and a time Index of at least one temporal domain cell in the domain dimension.
  • a processing unit configured to send the first reference signal to the terminal device through the communication unit
  • the processing unit is configured to receive the first channel information from the terminal device through the communication unit, and the first channel information is used to indicate the first sparse matrix; the first sparse matrix is at least one of the first channel eigenvector matrix through the second sparse pattern
  • the dimension is obtained by sampling, the first channel feature matrix is determined by the first channel state matrix, and the first channel state matrix is determined according to the first reference signal; the first sparse matrix is processed by the neural network to obtain the second matrix, the second The second matrix is the recovery value of the first channel characteristic matrix; the neural network is trained by the data sampled according to the second sparse pattern.
  • the communication unit is also used to: receive a second reference signal from the terminal device, the second reference signal is used to train the neural network corresponding to the second sparse pattern, and the neural network is used to restore second matrix.
  • the fourth precoding used for receiving the second reference signal is the same as the third precoding used for sending the first reference signal.
  • the bandwidth of the first reference signal is smaller than or equal to the bandwidth of the second reference signal.
  • the subcarrier spacing corresponding to the first reference signal is the same as the subcarrier spacing corresponding to the second reference signal.
  • the dimension of the first channel state matrix includes at least one dimension of the antenna dimension of the terminal device, the antenna dimension of the network device, and the frequency domain dimension and the time domain dimension corresponding to the first reference signal;
  • the second sparse pattern indicates at least one of the following: an index of at least one antenna in the antenna dimension of the terminal device, an index of at least one antenna in the antenna dimension of the network device, an index of at least one frequency domain unit in the frequency domain dimension, and a time Index of at least one temporal domain cell in the domain dimension.
  • a communication unit may also be referred to as a transceiver, transceiver, transceiving device, or the like.
  • a processing unit may also be called a processor, a processing board, a processing module, a processing device, and the like.
  • the device in the communication unit 1002 for realizing the receiving function can be regarded as a receiving unit
  • the device in the communication unit 1002 for realizing the sending function can be regarded as a sending unit, that is, the communication unit 1002 includes a receiving unit and a sending unit.
  • the communication unit may sometimes be called a transceiver, a transceiver, or a transceiver circuit and the like.
  • the receiving unit may sometimes be called a receiver, a receiver, or a receiving circuit, etc.
  • the sending unit may sometimes be called a transmitter, a transmitter, or a transmitting circuit, etc.
  • processing unit 1001 and the communication unit 1002 may also perform other functions.
  • processing unit 1001 and the communication unit 1002 may also perform other functions.
  • FIG. 11 a communication device 1100 provided in the embodiment of the present application is shown.
  • the communication device shown in FIG. 11 may be an implementation manner of a hardware circuit of the communication device shown in FIG. 10 .
  • the communication device may be applicable to the flow chart shown above, and execute the functions of the terminal device or the network device in the above method embodiments.
  • FIG. 11 only shows the main components of the communication device.
  • a communication device 1100 includes a processor 1110 and an interface circuit 1120 .
  • the processor 1110 and the interface circuit 1120 are coupled to each other.
  • the interface circuit 1120 may be a transceiver or an input-output interface.
  • the communication device 1100 may further include a memory 1130 for storing instructions executed by the processor 1110 or storing input data required by the processor 1110 to execute the instructions or storing data generated by the processor 1110 after executing the instructions.
  • the processor 1110 is used to implement the functions of the above processing unit 1001
  • the interface circuit 1120 is used to implement the functions of the above communication unit 1002 .
  • the terminal device chip implements the functions of the terminal device in the above method embodiment.
  • the terminal device chip receives information from other modules in the terminal device (such as radio frequency modules or antennas), and the information is sent to the terminal device by the network device; or, the terminal device chip sends information to other modules in the terminal device (such as radio frequency modules or antenna) to send information, which is sent by the terminal device to the network device.
  • the network equipment chip implements the functions of the network equipment in the above method embodiments.
  • the network device chip receives information from other modules in the network device (such as radio frequency modules or antennas), and the information is sent to the network device by the terminal device; or, the network device chip sends information to other modules in the network device (such as radio frequency modules or antenna) to send information, which is sent by the network device to the terminal device.
  • the processor in the embodiments of the present application can be a central processing unit (Central Processing Unit, CPU), and can also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application-specific integrated circuits (Application Specific Integrated Circuit, ASIC), Field Programmable Gate Array (Field Programmable Gate Array, FPGA) or other programmable logic devices, transistor logic devices.
  • a general-purpose processor can be a microprocessor, or any conventional processor.
  • memory can be random access memory (Random Access Memory, RAM), flash memory, read-only memory (Read-Only Memory, ROM), programmable read-only memory (Programmable ROM, PROM), erasable Programmable read-only memory (Erasable PROM, EPROM), electrically erasable programmable read-only memory (Electrically EPROM, EEPROM), registers, hard disk, mobile hard disk or any other form of storage medium known in the art.
  • An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium.
  • the storage medium may also be a component of the processor.
  • the processor and storage medium can be located in the ASIC.
  • the ASIC can be located in a network device or a terminal device. Processors and storage media may also exist in network devices or terminal devices as discrete components.
  • the embodiments of the present application may be provided as methods, systems, or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) having computer-usable program code embodied therein.
  • These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions
  • the device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Mathematical Physics (AREA)
  • Physics & Mathematics (AREA)
  • Power Engineering (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Medical Informatics (AREA)
  • Data Mining & Analysis (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请提供一种信道信息反馈、恢复方法及装置，其中方法包括：终端设备接收来自网络设备的第一参考信号；所述终端设备根据所述第一参考信号进行信道估计，获得第一信道状态矩阵；所述终端设备根据第一稀疏图样对所述第一信道状态矩阵进行抽样，获得稀疏信道状态矩阵；所述第一稀疏图样来自所述由网络设备配置，所述第一稀疏图样用于对所述第一信道状态矩阵的至少一个维度进行抽样；所述终端设备向所述网络设备发送第一信道信息，所述第一信道信息用于指示所述稀疏信道状态矩阵。

Description

一种信道信息反馈、恢复方法及装置
相关申请的交叉引用
本申请要求在2021年08月11日提交中国专利局、申请号为202110918725.9、申请名称为“一种信道信息反馈、恢复方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及通信技术领域，尤其涉及一种信道信息反馈、恢复方法及装置。
背景技术
在长期演进(long term evolution，LTE)以及新空口(new radio，NR)等通信系统中，基站需要获取下行信道状态信息(channel state information，CSI)，用于确定调度终端设备的下行数据信道的资源、调制编码方案(modulation and coding scheme，MCS)、预编码矩阵等配置。在时分双工(time division duplex，TDD)模式中，由于上下行信道存在互易性，基站可以通过测量上行参考信号获取上行CSI，进而推测出较为精确的下行CSI，例如将上行CSI用作下行CSI。在频分双工(frequency division duplex，FDD)模式中，上下行信道不一定存在互易性，因此需要从终端设备获取下行CSI。例如，终端设备通过测量参考信号获得下行CSI，并将CSI反馈给基站。
在现有的宽带多天线系统中，CSI反馈需要很大的资源开销。为了降低CSI反馈的开销，终端设备会对CSI进行压缩以及量化等操作，从而反馈经过压缩以及量化的CSI。对CSI进行压缩以及量化等操作，会对CSI的精度造成损失。而基站获取到的下行CSI越不精确，所调度的下行传输就越不能匹配当前的下行信道，下行传输的性能就越差。
综上所述，如何在降低CSI反馈开销的同时，提高CSI反馈的精度，是一个亟待解决的问题。
发明内容
本申请提供一种信道信息反馈、恢复方法及装置，用以在降低信道信息反馈开销的同时，提高信道信息反馈的精度。
第一方面，本申请提供一种信道信息反馈方法，该方法适用于基于AI的信道信息压缩反馈的场景。该方法的执行主体为终端设备或终端设备中的一个模块，这里以终端设备为执行主体为例进行描述。该方法包括：终端设备接收来自网络设备的第一参考信号；根据第一参考信号获得第一信道状态矩阵；根据第一稀疏图样对第一信道状态矩阵进行抽样，获得稀疏信道状态矩阵；第一稀疏图样由网络设备配置，第一稀疏图样用于对第一信道状态矩阵的至少一个维度进行抽样；向网络设备发送第一信道信息，第一信道信息用于指示稀疏信道状态矩阵。
通过实施上面的方法，在不需要终端设备参与训练神经网络，且不需要终端设备反馈大量完整的下行信道信息的情况下，仅由网络设备独立训练神经网络，网络设备在获取到来自终端设备的第一信道信息时，便可以通过神经网络恢复下行信道信息。通过该方法， 可以降低对终端设备能力的要求，且避免终端设备反馈完整下行信道信息导致的反馈开销。
在一种可能的设计中，还包括：向网络设备发送第二参考信号，第二参考信号用于训练第一稀疏图样对应的神经网络，神经网络用于根据稀疏信道状态矩阵恢复第一信道状态矩阵。
在一种可能的设计中，发送第二参考信号所使用的第二预编码与接收第一参考信号所使用的第一预编码相同。
通过上述方法，可以保证发送第二参考信号时使用的发送波束与接收第一参考信号时使用的接收波束相同。
在一种可能的设计中，第一参考信号的带宽小于或者等于第二参考信号的带宽。
在一种可能的设计中，第一参考信号对应的子载波间隔大于或等于第二参考信号对应的子载波间隔。
通过上述方法，本申请实施例方案在第一参考信号带宽、子载波间隔和需要获取的第二信道信息的带宽、子载波间隔不一致的情况也能使用。
在一种可能的设计中，第一信道状态矩阵的维度包含终端设备的天线维度、网络设备的天线维度和第一参考信号对应的频域维度和时域维度中的至少一个维度；
第一稀疏图样指示以下至少一项：终端设备的天线维度中的至少一个天线的索引、网络设备的天线维度中的至少一个天线的索引、频域维度中的至少一个频域单元的索引和时域维度中的至少一个时域单元的索引。
第二方面，本申请还提供一种通信装置，该通信装置具有实现上述第一方面提供的任一方法。该通信装置可以通过硬件实现，也可以通过硬件执行相应的软件实现。该硬件或软件包括一个或多个与上述功能相对应的单元或模块。
在一种可能的实现方式中，该通信装置包括：处理器，该处理器被配置为支持该通信装置执行以上所示方法中相应功能。该通信装置还可以包括存储器，该存储可以与处理器耦合，其保存该通信装置必要的程序指令和数据。可选地，该通信装置还包括接口电路，该接口电路用于支持该通信装置与其它设备之间的通信。
在一种可能的实现方式中，该通信装置可以为终端设备，或者终端设备中的芯片或一个模块。
在一种可能的实现方式中，该通信装置包括相应的功能模块，分别用于实现以上方法中的步骤。功能可以通过硬件实现，也可以通过硬件执行相应的软件实现。硬件或软件包括一个或多个与上述功能相对应的模块。
在一种可能的实施方式中，通信装置的结构中包括处理单元和通信单元，这些单元可以执行上述方法示例中相应功能，具体参见第一方面提供的方法中的描述，此处不做赘述。
第三方面，本申请提供一种信道信息恢复方法，该方法适用于基于AI的信道信息压缩反馈的场景。该方法的执行主体为网络设备或网络设备中的一个模块，这里以网络设备为执行主体为例进行描述。该方法包括：网络设备向终端设备发送第一参考信号；接收来自终端设备的第一信道信息，第一信道信息用于指示稀疏信道状态矩阵；稀疏信道状态矩阵为通过第一稀疏图样对第一信道状态矩阵的至少一个维度进行抽样获得的，第一信道状态矩阵根据第一参考信号确定；通过神经网络对稀疏信道状态矩阵进行处理，获得第二信道状态矩阵，第二信道状态矩阵为对第一信道状态矩阵的恢复值；神经网络是通过根据第一稀疏图样抽样的数据训练的。
在一种可能的设计中，还包括：接收来自终端设备的第二参考信号；根据第二参考信号获得第三信道状态矩阵；根据第三信道状态矩阵获取第四信道状态矩阵，第四信道状态矩阵在子载波维度的长度等于第一信道状态矩阵在子载波维度的长度，第四信道状态矩阵在接收天线维度的长度等于第一信道状态矩阵在发送天线维度的长度，第四信道状态矩阵在发送天线维度的长度等于第一信道状态矩阵在接收天线维度的长度；网络设备根据第一稀疏图样对第四信道状态矩阵的至少一个维度进行抽样获得第五信道状态矩阵；神经网络是通过多个第四信道状态矩阵以及对应的第五信道状态矩阵训练得到的。
在一种可能的设计中，还包括：接收来自终端设备的第二参考信号，第二参考信号用于训练第一稀疏图样对应的神经网络，神经网络用于根据稀疏信道状态矩阵恢复第一信道状态矩阵。
在一种可能的设计中，接收第二参考信号所使用的第四预编码与发送第一参考信号所使用的第三预编码相同。
在一种可能的设计中，第一参考信号的带宽小于或者等于第二参考信号的带宽。
在一种可能的设计中，第一参考信号对应的子载波间隔与第二参考信号对应的子载波间隔相同。
在一种可能的设计中，第一信道状态矩阵的维度包含终端设备的天线维度、网络设备的天线维度和第一参考信号对应的频域维度和时域维度中的至少一个维度；
第一稀疏图样指示以下至少一项：终端设备的天线维度中的至少一个天线的索引、网络设备的天线维度中的至少一个天线的索引、频域维度中的至少一个频域单元的索引和时域维度中的至少一个时域单元的索引。
第四方面，本申请还提供一种通信装置，该通信装置具有实现上述第三方面提供的任一方法。该通信装置可以通过硬件实现，也可以通过硬件执行相应的软件实现。该硬件或软件包括一个或多个与上述功能相对应的单元或模块。
在一种可能的实现方式中，该通信装置包括：处理器，该处理器被配置为支持该通信装置执行以上所示方法中相应功能。该通信装置还可以包括存储器，该存储可以与处理器耦合，其保存该通信装置必要的程序指令和数据。可选地，该通信装置还包括接口电路，该接口电路用于支持该通信装置与其它设备之间的通信。
在一种可能的实现方式中，该通信装置可以为网络设备，或者网络设备中的芯片或一个模块。
在一种可能的实施方式中，通信装置的结构中包括处理单元和通信单元，这些单元可以执行上述方法示例中相应功能，具体参见第三方面提供的方法中的描述，此处不做赘述。
第五方面，本申请提供一种信道信息反馈方法，该方法适用于基于AI的信道信息压缩反馈的场景。该方法的执行主体为终端设备或终端设备中的一个模块，这里以终端设备为执行主体为例进行描述。该方法包括：终端设备接收来自网络设备的第一参考信号；根据第一参考信号进行信道估计，获得第一信道状态矩阵；根据第一信道状态矩阵确定第一信道特征矩阵；根据第二稀疏图样对第一信道特征矩阵进行抽样，获得第一稀疏矩阵；第二稀疏图样由网络设备配置，第二稀疏图样用于对第一信道特征矩阵的至少一个维度进行抽样；向网络设备发送第一信道信息，第一信道信息用于指示第一稀疏矩阵。
通过实施上面的方法，可以在不需要终端设备参与训练神经网络的情况下，利用上下行信道之间相似的相关性，网络设备通过训练神经网络，并通过该神经网络和终端设备反 馈的第一稀疏矩阵恢复完整的第一信道特征矩阵，可以降低对终端设备能力的要求，且避免终端设备反馈完整的第一信道特征矩阵导致的反馈开销。
在一种可能的设计中，还包括：向网络设备发送第二参考信号，第二参考信号用于训练第一稀疏图样对应的神经网络，神经网络用于根据第一稀疏矩阵恢复第二矩阵。
在一种可能的设计中，发送第二参考信号所使用的第二预编码与接收第一参考信号所使用的第一预编码相同。
在一种可能的设计中，第一参考信号的带宽小于或者等于第二参考信号的带宽。
在一种可能的设计中，第一参考信号对应的子载波间隔与第二参考信号对应的子载波间隔相同。
在一种可能的设计中，第一信道状态矩阵的维度包含终端设备的天线维度、网络设备的天线维度和第一参考信号对应的频域维度和时域维度中的至少一个维度；
第二稀疏图样指示以下至少一项：终端设备的天线维度中的至少一个天线的索引、网络设备的天线维度中的至少一个天线的索引、频域维度中的至少一个频域单元的索引和时域维度中的至少一个时域单元的索引。
第六方面，本申请还提供一种通信装置，该通信装置具有实现上述第五方面提供的任一方法。该通信装置可以通过硬件实现，也可以通过硬件执行相应的软件实现。该硬件或软件包括一个或多个与上述功能相对应的单元或模块。
在一种可能的实现方式中，该通信装置包括：处理器，该处理器被配置为支持该通信装置执行以上所示方法中相应功能。该通信装置还可以包括存储器，该存储可以与处理器耦合，其保存该通信装置必要的程序指令和数据。可选地，该通信装置还包括接口电路，该接口电路用于支持该通信装置与其它设备之间的通信。
在一种可能的实现方式中，该通信装置可以为终端设备，或者终端设备中的芯片或一个模块。
在一种可能的实施方式中，通信装置的结构中包括处理单元和通信单元，这些单元可以执行上述方法示例中相应功能，具体参见第五方面提供的方法中的描述，此处不做赘述。
第七方面，本申请提供一种信道信息恢复方法，该方法适用于基于AI的信道信息压缩反馈的场景。该方法的执行主体为网络设备或网络设备中的一个模块，这里以网络设备为执行主体为例进行描述。该方法包括：网络设备向终端设备发送第一参考信号；接收来自终端设备的第一信道信息，第一信道信息用于指示第一稀疏矩阵；第一稀疏矩阵为通过第二稀疏图样对第一信道特征向量矩阵的至少一个维度进行抽样获得的，第一信道特征矩阵为通过第一信道状态矩阵确定的，第一信道状态矩阵根据第一参考信号确定；通过神经网络对第一稀疏矩阵进行处理，获得第二矩阵，第二矩阵为对第一信道特征矩阵的恢复值；神经网络是通过根据第二稀疏图样抽样的数据训练的。
在一种可能的设计中，还包括：接收来自终端设备的第二参考信号，第二参考信号用于训练第二稀疏图样对应的神经网络，神经网络用于根据第一稀疏矩阵恢复第二矩阵。
在一种可能的设计中，接收第二参考信号所使用的第四预编码与发送第一参考信号所使用的第三预编码相同。
在一种可能的设计中，第一参考信号的带宽小于或者等于第二参考信号的带宽。
在一种可能的设计中，第一参考信号对应的子载波间隔与第二参考信号对应的子载波间隔相同。
在一种可能的设计中，第一信道状态矩阵的维度包含终端设备的天线维度、网络设备的天线维度和第一参考信号对应的频域维度和时域维度中的至少一个维度；
第二稀疏图样指示以下至少一项：终端设备的天线维度中的至少一个天线的索引、网络设备的天线维度中的至少一个天线的索引、频域维度中的至少一个频域单元的索引和时域维度中的至少一个时域单元的索引。
第八方面，本申请还提供一种通信装置，该通信装置具有实现上述第七方面提供的任一方法。该通信装置可以通过硬件实现，也可以通过硬件执行相应的软件实现。该硬件或软件包括一个或多个与上述功能相对应的单元或模块。
在一种可能的实现方式中，该通信装置包括：处理器，该处理器被配置为支持该通信装置执行以上所示方法中相应功能。该通信装置还可以包括存储器，该存储可以与处理器耦合，其保存该通信装置必要的程序指令和数据。可选地，该通信装置还包括接口电路，该接口电路用于支持该通信装置与其它设备之间的通信。
在一种可能的实现方式中，该通信装置可以为网络设备，或者网络设备中的芯片或一个模块。
在一种可能的实施方式中，通信装置的结构中包括处理单元和通信单元，这些单元可以执行上述方法示例中相应功能，具体参见第七方面提供的方法中的描述，此处不做赘述。
第九方面，提供了一种通信装置，包括处理器和存储器，存储器中存储计算机程序或指令；该处理器用于执行所述存储器中存储的计算机程序或指令，实现前述第一方面中任意可能的实现方式中的方法。
第十方面，提供了一种通信装置，包括处理器存储器，存储器中存储计算机程序或指令；该处理器用于执行所述存储器中存储的计算机程序或指令，实现前述第三方面中任意可能的实现方式中的方法。
第十一方面，提供了一种通信装置，包括处理器和存储器，存储器中存储计算机程序或指令；该处理器用于执行所述存储器中存储的计算机程序或指令，实现前述第五方面中任意可能的实现方式中的方法。
第十二方面，提供了一种通信装置，包括处理器存储器，存储器中存储计算机程序或指令；该处理器用于执行所述存储器中存储的计算机程序或指令，实现前述第七方面中任意可能的实现方式中的方法。
第十三方面，提供了一种通信装置，包括处理器和接口电路，可选地，还包括存储器，存储器中存储计算机程序或指令；接口电路用于接收来自该通信装置之外的其它通信装置的信号并传输至该处理器或将来自该处理器的信号发送给该通信装置之外的其它通信装置，该处理器用于执行所述存储器中存储的计算机程序或指令，实现前述第一方面或第五方面中任意可能的实现方式中的方法。
第十四方面，提供了一种通信装置，包括处理器和接口电路，可选地，还包括存储器，存储器中存储计算机程序或指令；接口电路用于接收来自该通信装置之外的其它通信装置的信号并传输至该处理器或将来自该处理器的信号发送给该通信装置之外的其它通信装置，该处理器用于执行所述存储器中存储的计算机程序或指令，实现前述第三方面或第七方面中的任意可能的实现方式中的方法。
第十五方面，提供了一种计算机可读存储介质，该计算机可读存储介质中存储有计算机程序或指令，当所述计算机程序或指令在计算机上运行时，使得所述计算机实现前述第 一方面或第三方面或第五方面或第七方面中任意可能的实现方式中的方法。
第十六方面，提供了一种存储有计算机可读指令的计算机程序产品，当所述计算机可读指令在计算机上运行时，使得所述计算机实现前述第一方面或第三方面或第五方面或第七方面中任意可能的实现方式中的方法。
第十七方面，提供一种芯片，该芯片包括处理器，还可以包括存储器，用于执行所述存储器中存储的计算机程序或指令，使得芯片系统实现前述第一方面或第三方面或第五方面或第七方面中任意可能的实现方式中的方法。
第十八方面，提供了一种通信装置，包括处理器和接口电路，接口电路用于接收来自该通信装置之外的其它通信装置的信号并传输至该处理器或将来自该处理器的信号发送给该通信装置之外的其它通信装置，该处理器用于执行计算机程序或指令，实现前述第一方面或第三方面或第五方面或第七方面中任意可能的实现方式中的方法。
第十九方面，提供了一种通信装置，包括用于实现前述第一方面中任意可能的实现方式中方法的模块。
第二十方面，提供了一种通信装置，包括用于实现前述第三方面中任意可能的实现方式中方法的模块。
第二十一方面，提供了一种通信装置，包括用于实现前述第五方面中任意可能的实现方式中方法的模块。
第二十二方面，提供了一种通信装置，包括用于实现前述第七方面中任意可能的实现方式中方法的模块。
第二十三方面，提供一种通信系统，所述系统包括第二方面所述的装置(如终端设备)以及第四方面所述的装置(如网络设备)。
第二十四方面，提供一种通信系统，所述系统包括第六方面所述的装置(如终端设备)以及第八方面所述的装置(如网络设备)。
附图说明
图1为本申请实施例提供的一种网络架构示意图；
图2为本申请实施例提供的一种神经网络的层关系示意图；
图3为适用于本申请实施例的一种AI网络结构示意图；
图4为本申请实施例提供的一种神经网络训练方法流程示意图；
图5为本申请实施例提供的一种AI模型结构示意图；
图6为本申请实施例提供的一种信道信息反馈、恢复方法流程示意图；
图7为本申请实施例提供的一种神经网络训练方法流程示意图；
图8为本申请实施例提供的一种AI模型结构示意图；
图9为本申请实施例提供的一种信道信息反馈、恢复方法流程示意图；
图10为本申请实施例提供的一种通信装置结构示意图；
图11为本申请实施例提供的一种通信装置结构示意图。
具体实施方式
下面结合说明书附图对本申请实施例做详细描述。
本申请实施例的技术方案可以应用于各种通信系统，例如：长期演进(long term evolution，LTE)系统、NR系统以及下一代通信系统等，在此不做限制。
本申请实施例中，终端设备，可以为具有无线收发功能的设备或可设置于任一设备中的芯片，也可以称为用户设备(user equipment，UE)、接入终端、用户单元、用户站、移动站、移动台、远方站、远程终端、移动设备、用户终端、无线通信设备、用户代理或用户装置。本申请实施例中的终端设备可以是手机(mobile phone)、平板电脑(Pad)、带无线收发功能的电脑、虚拟现实(virtual reality，VR)终端、增强现实(augmented reality，AR)终端、工业控制(industrial control)中的无线终端、无人驾驶(self driving)中的无线终端等。
在本申请实施例中，网络设备可以为各种制式下无线接入设备，例如可以是NR系统中的下一代基站(next Generation node B，gNB)，可以是演进型节点B(evolved Node B，eNB)、无线网络控制器(radio network controller，RNC)或节点B(Node B，NB)、基站控制器(base station controller，BSC)、基站收发台(base transceiver station，BTS)、家庭基站(例如，home evolved NodeB，或home Node B，HNB)、基带单元(baseband unit，BBU)，无线保真(wireless fidelity，WIFI)系统中的接入点(access point，AP)、无线中继节点、无线回传节点、传输点(transmission and reception point，TRP或者transmission point，TP)等，还可以为5G(NR)系统中的gNB或传输点，5G系统中的基站的一个或一组(包括多个天线面板)天线面板。其中，5G系统中的基站还可以称为发送接收点(transmission reception point，TRP)或下一代节点B(generation Node B，gNB或gNodeB)。本申请实施例中的基站可以是一体化基站，或者可以是包括集中式单元(centralized unit,CU)和分布式单元(distributed unit,DU)的基站。包括CU和DU的基站还可以称为CU和DU分离的基站，如该基站包括gNB-CU和gNB-DU。其中，CU还可以分离为CU控制面(CU control plane,CU-CP)和CU用户面(CU user plane，CU-CP)，如该基站包括gNB-CU-CP、gNB-CU-UP和gNB-DU。
本申请实施例中，用于实现网络设备的功能的装置可以是网络设备；也可以是能够支持网络设备实现该功能的装置，例如芯片系统。该装置可以被安装在网络设备中或者和网络设备匹配使用。下述实施例中，以用于实现网络设备的功能的装置是网络设备，以网络设备是基站为例，描述本申请实施例提供的技术方案。
为了在无线网络中支持机器学习功能，网络中还可能引入专门的人工智能(artificial intelligence，AI)AI网元或模块。如果引入AI网元，则对应一个独立的网元；如果引入AI模块，则可以位于某个网元内部，对应的网元可以是gNB、UE等。
为便于理解本申请实施例，首先以图1中示出的通信系统为例详细说明适用于本申请实施例的通信系统。图1是本申请实施例可以应用提供的通信系统的架构示意图，该通信系统中包括网络设备和终端设备。终端设备可以接入网络设备，并和网络设备进行通信。图1只是示意图，本申请实施例对该通信系统中包括的网络设备和终端设备的数量不做限定。可选地，该通信系统中还可以包括用于实现AI功能的节点，该节点可以和网络设备进行通信，该节点也可以位于网络设备中，是网络设备中的一个模块。
目前，终端设备反馈下行CSI的方法为：网络设备向终端设备发送下行参考信号，终端设备根据下行参考信号进行信道估计，根据信道估计结果从预定义的码本中选择一个与信道最匹配的预编码，并将选择的预编码的信息反馈给网络设备，该预编码的信息为预编 码矩阵索引(precoding matrix indicator，PMI)。此外，终端设备还可通过反馈信道质量指示(channel quality indicator，CQI)指示终端设备判断的当前信道在所能支持的调制编码方式，通过反馈秩指示(rank indicator，RI)指示终端设备建议的下行传输的层数等，PMI、CQI、RI等反馈信息都可用于表征下行CSI。
网络设备获取到的下行CSI越精确，所调度的下行传输就越能匹配当前的下行信道，下行传输的性能就越好。而在现有的宽带多天线系统中，完整的CSI通常很大，为了降低CSI反馈的开销，通常会对CSI进行压缩、量化等操作，这就会对CSI的精度造成损失。CSI反馈需要在反馈开销和反馈精度上进行折中。
由于神经网络(neural network，NN)技术具有自我学习的能力，本申请中将基于神经网络技术来实现恢复压缩的CSI。神经网络是机器学习的一种具体实现形式。根据通用近似定理，神经网络理论上可以逼近任意连续函数，从而使得神经网络具备学习任意映射的能力。因此神经网络可以对复杂的高维度问题进行准确地抽象建模。神经网络的思想来源于大脑组织的神经元结构。每个神经元都对其输入值做加权求和运算，将加权求和结果通过一个激活函数产生输出。
神经网络一般包括多层结构，每层可包括一个或多个神经元。增加神经网络的深度和/或宽度可以提高该神经网络的表达能力，为复杂系统提供更强大的信息提取和抽象建模能力。其中，神经网络的深度可以指神经网络包括的层数，每层包括的神经元个数可以称为该层的宽度。如图2所示，为神经网络的层关系示意图。一种实现中，神经网络包括输入层和输出层。神经网络的输入层将接收到的输入经过神经元处理后，将结果传递给输出层，由输出层得到神经网络的输出结果。另一种实现中，神经网络包括输入层、隐藏层和输出层。神经网络的输入层将接收到的输入经过神经元处理后，将结果传递给中间的隐藏层，隐藏层再将计算结果传递给输出层或者相邻的隐藏层，最后由输出层得到神经网络的输出结果。一个神经网络可以包括一层或多层依次连接的隐藏层，不予限制。神经网络的训练过程中，可以定义损失函数。损失函数描述了神经网络的输出值和理想目标值之间的差距或差异，本申请不限制损失函数的具体形式。神经网络的训练过程就是通过调整神经网络参数，如神经网络的层数、宽度、神经元的权值、和/或神经元的激活函数中的参数等，使得损失函数的值小于阈值门限值或者满足目标需求的过程。
如图3所示，为适用于本申请实施例的一种AI网络结构示意图。图3所示的AI网络结构是基于自编码器(autoencoders，AE)的神经网络。其中编码器(encoder)和解码器(decoder)分别是一个神经网络。本申请实施例中，可以将自编码器的编码器部分部署在终端设备侧，解码器部署在网络设备侧，解码器也可以独立于网络设备。在对编码器和解码器进行训练之后，网络设备可以向终端设备发送下行参考信号。终端设备对下行参考信号进行信道估计，获得信道状态信息。终端设备中的编码器对信道状态信息进行压缩，终端设备将压缩后的信道状态信息发送至解码器。解码器用于对压缩的信道信息进行恢复，通过足够的信道样本训练自编码器，使得解码器输出的信道状态信息与编码器输入的信道信息的差异足够小。
本申请实施例中，神经网络是一种模仿动物神经网络行为特征，进行分布式并行信息处理的数学模型，是AI模型的一种特殊形式。
在本申请的各个实施例中，如果没有特殊说明以及逻辑冲突，不同的实施例之间的术语和/或描述具有一致性、且可以相互引用，不同的实施例中的技术特征根据其内在的逻辑 关系可以组合形成新的实施例。
可以理解的是，在本申请的实施例中涉及的各种数字编号仅为描述方便进行的区分，并不用来限制本申请的实施例的范围。上述各过程的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定。
本申请实施例描述的网络架构以及业务场景是为了更加清楚的说明本申请实施例的技术方案，并不构成对于本申请实施例提供的技术方案的限定，本领域普通技术人员可知，随着网络架构的演变和新业务场景的出现，本申请实施例提供的技术方案对于类似的技术问题，同样适用。
本申请实施例涉及到神经网络的模型训练以及模型部署和应用。神经网络的模型训练需要进行数据收集，收集到的样本数据用于模型训练，数据收集的主体可以是AI网元或AI模块或者专门的数据收集网元或模块。在本申请实施例中，网络设备利用稀疏上行信道状态矩阵和完整的上行信道状态矩阵训练神经网络模型，收集的数据为{稀疏上行信道状态矩阵，上行信道状态矩阵}，其中稀疏上行信道状态矩阵可以作为训练样本，上行信道状态矩阵可以作为样本标签。为了描述方便，本申请中将完整的上行信道状态矩阵简称为上行信道状态矩阵。
本申请实施例中，训练样本指的是在训练神经网络时输入神经网络的数据，样本标签指的是神经网络输入样本时期望得到的输出值，可以理解为训练样本对应的真实值，训练神经网络的目的就是希望输入神经网络一个训练样本，其输出与样本标签尽可能相近。
本申请实施例中，数据收集的具体方法为，终端设备发送上行参考信号，例如探测参考信号(sounding reference signal，SRS)，网络设备根据接收到的上行参考信号进行信道估计，获取上行信道状态矩阵，再根据预先确定好的稀疏图样，从上行信道状态矩阵抽取部分元素，得到稀疏上行信道状态矩阵。稀疏上行信道状态矩阵和其对应的上行信道状态矩阵可以作为训练数据，数据收集阶段需要收集多个训练数据，收集的样本数量可以根据实际情况确定，本申请并不限定。
本申请实施例中，为了描述方便，将网络设备向终端设备发送的下行参考信号称为第一参考信号，将终端设备向网络设备发送的上行参考信号称为第二参考信号，将根据第一参考信号确定的下行信道状态矩阵称为第一信道状态矩阵，将下行信道状态矩阵对应的稀疏下行信道矩阵称为稀疏信道状态矩阵，将通过神经网络对稀疏信道状态矩阵进行处理，获得的矩阵称为第二信道状态矩阵，将网络设备根据第二参考信号确定的上行信道状态矩阵称为第三信道状态矩阵。
本申请实施例中，上行道状态矩阵的维度包括{终端设备的发送天线，网络设备的接收天线，频域，时域}中的至少一个维度，如果某个维度的长度为1，则也可认为这个维度不存在，其中在频域维度的粒度为频域单元，例如可以是子载波或者资源块(resource block，RB)，在时域维度的粒度为时域单元，例如可以是正交频分复用(orthogonal frequency division multiplexing，OFDM)符号或者时隙等。下行道状态矩阵的维度包括{终端设备的接收天线，网络设备的发送天线，频域，时域}中的至少一个维度。
其中，上行信道状态矩阵的维度与上行参考信号对应的资源的维度相同，但上行道状态矩阵是通过信道估计确定出的矩阵，在各个维度上和实际发送上线参考信号的资源并不完全相同。也就是说，上行信道状态矩阵与上行参考信号对应的资源，在各个维度上的具体取值范围可以相同，也可以不同，例如，上行参考信号对应的资源在频域维度的范围是 0～6RB，但上行信道状态矩阵在频域维度的范围可以是0～10RB；再例如，上行参考信号对应的资源在终端设备的发送天线维度包括2根天线，但是上行道状态矩阵在终端设备的发送天线维度具体包括几根天线并不限定。其中，上行参考信号对应的资源的维度，包括发送上行参考信号的天线，接收上行参考信号的天线，时域以及频域等维度。
类似的，下行信道状态矩阵的维度与下行参考信号对应的资源的维度相同，但在每个维度上的具体取值范围可以相同，也可以不同。其中，下行参考信号对应的资源的维度，包括发送下行参考信号的天线，接收下行参考信号的天线，时域以及频域等维度。
为了描述方便，以下的描述中，将发送天线以及接收天线均统称为天线。
如图4所示，为本申请实施例提供的一种神经网络的模型训练流程示意图。该流程涉及终端设备、网络设备以及AI实体之间的交互。其中，AI实体可以为独立的网元，也可以为其他设备中的一个模块，例如可以为网络设备中的一个模块。
S401：终端设备向网络设备发送第二参考信号。
其中，终端设备可以向网络设备发送多个第二参考信号，具体数量并不限定。第二参考信号是上行参考信号，例如可以为SRS，具体类型并不限定。
本申请实施例中，第二参考信号可以用于训练第一稀疏图样对应的神经网络。
S402：网络设备接收来自终端设备的第二参考信号，并根据第二参考信号进行信道估计，获得第三信道状态矩阵。
如前所述，第三信道状态矩阵也就是上行信道状态矩阵。
本申请实施例中，网络设备对第二参考信号进行信道估计可以获得第二信道状态信息，第二信道状态信息可以指示出网络设备和终端设备之间上行信道的信道响应，第二信道状态信息可以为一个多维矩阵，即上行信道状态矩阵。
在本申请实施例中，上行信道的信道响应对应的上行信道状态矩阵，即第三信道状态矩阵的维度可以包括{终端设备的发送天线，网络设备的接收天线，频域，时域}中的至少一个维度，或者包括{终端设备的天线端口，网络设备的天线端口，频域，时域}中的至少一个维度，或者是其他维度。其中，终端设备的天线端口为发送第二参考信号的天线端口，网络设备的天线端口为接收第二参考信号的天线端口。因此，本申请实施例中，上行信道状态矩阵，即第三信道状态矩阵可以是一个多维矩阵。
另外，本申请实施例中，网络设备根据第二参考信号进行信道估计获取第三信道状态矩阵的方法，不做限定，可以是传统信道估计算法，例如，最小均方误差估计算法(minimum mean squareerror estimation，MMSE)，也可以是基于神经网络的信道估计算法。
S403：网络设备根据第三信道状态矩阵确定第五信道状态矩阵。
本申请实施例中，为了获取准确的训练数据，可以将第三信道状态矩阵进行维度转换，获得第四信道状态矩阵。其中，维度转换指的是对第三信道状态矩阵进行操作，使得第三信道状态矩阵与下行信道状态矩阵的维度以及每个维度的长度匹配。
第四信道状态矩阵和下行道状态矩阵的维度对应，即第四信道状态矩阵中终端设备的天线维度与下行道状态矩阵中终端设备的天线维度对应，网络设备的天线维度与下行道状态矩阵中网络设备的天线维度需要对应，频域维度与频域维度对应，时域维度与时域维度对应。例如，第四信道状态矩阵的维度顺序依次为终端设备的天线，网络设备的天线，频域，时域；如果下行道状态矩阵的维度顺序依次为终端设备的天线，网络设备的天线，频域，时域；那么第四信道状态矩阵和下行道状态矩阵的维度是对应的；如果下行道状态矩 阵的维度顺序为其他情况，那么第四信道状态矩阵和下行道状态矩阵的维度不是对应的。
第四信道状态矩阵和下行道状态矩阵的维度对应时，第四信道状态矩阵和下行道状态矩阵在各个维度的长度相同，即第四信道状态矩阵在终端设备的天线维度的长度等于所述下行信道状态矩阵在终端设备的天线维度的长度，第四信道状态矩阵在网络设备的天线维度的长度等于所述下行信道状态矩阵在网络设备的天线维度的长度，所述第四信道状态矩阵在频域维度的长度等于所述下行信道状态矩阵在频域维度的长度，所述第四信道状态矩阵在时域维度的长度等于所述下行信道状态矩阵在时域维度的长度。
网络设备可以根据第一稀疏图样从第四信道状态矩阵中抽取部分元素，获得稀疏上行信道状态矩阵(也可以称为第五信道状态矩阵)，稀疏上行信道状态矩阵包括从第四信道状态矩阵中抽取的部分元素。也就是说第一稀疏图样可以用于从第四信道状态矩阵中抽取部分元素，从而获得第五信道状态矩阵。
该第一稀疏图样为网络设备预先确定的，由于本申请实施例使用第三信道状态矩阵训练神经网络，并将该神经网络应用于下行信道状态矩阵恢复，因此终端设备进行信道状态矩阵反馈阶段也使用第一稀疏图样，与网络设备使用的第一稀疏图样相同。
一种可能的实现方式中，第一稀疏图样可以指示在上行信道状态矩阵(第三信道状态矩阵)或下行信道状态矩阵(终端设备根据下行参考信号进行信道估计获得的矩阵)的至少一个维度上抽取的元素的索引。例如，当第一稀疏图样用于对第四信道状态矩阵的至少一个维度上的元素抽取时，第一稀疏图样可以指示网络设备的天线维度中的至少一个天线的索引、终端设备的天线维度中的至少一个天线的索引、第二参考信号对应的频域维度中的至少一个频域单元的索引，以及第二参考信号对应的时域维度中的至少一个时域单元的索引中的至少一项。
举例来说，终端设备在发送第二参考信号时，终端设备采用的天线数为N T UE＝4，网络设备在接收第二参考信号时，网络设备采用的天线数为N R BS＝64，第二参考信号对应的频域单元数为Nsc＝120，第二参考信号对应的时域单元数为1，则第四信道状态矩阵为4×64×120的3维矩阵。例如，第一稀疏图样可以为{终端设备的天线的索引:1,3；网络设备的天线的索引:1,5,11,…,61；频域单元的索引:1,21,…,101}，其中，终端设备的天线的索引，表示索引为1和3的天线；网络设备的天线的索引，表示索引为1,5,11,…,2n-1,…,61的天线，n为正整数；频域单元的索引，表示索引为1,21,…,20m+1,…,101的频域单元，m为大于或等于0的整数。该第一稀疏图样可以用于抽取第四信道状态矩阵中与索引为1和3的天线对应的元素，抽取第四信道状态矩阵中与索引为1,5,11,…,61的天线对应的元素，以及抽取第四信道状态矩阵中与索引为1,21,…,101的子载波对应的元素，从第四信道状态矩阵中抽取的元素构成第五信道状态矩阵，即稀疏上行信道状态矩阵，此时稀疏上行信道状态矩阵为2×13×6的3维矩阵。
S404：网络设备向AI实体发送第四信道状态矩阵和第五信道状态矩阵。
另一种可能的实现方式中，网络设备可以发送第三信道状态矩阵和第五信道状态矩阵。
S405：AI实体接收第四信道状态矩阵和第五信道状态矩阵，并根据第四信道状态矩阵和第五信道状态矩阵对神经网络进行训练。
其中，第五信道状态矩阵可以作为训练数据中的训练样本，即是在训练神经网络时输入神经网络的数据；第四信道状态矩阵可以作为训练数据中的样本标签，即神经网络输入样本时期望得到的输出值，可以理解为训练样本对应的真实值。训练神经网络的目的就是 希望输入神经网络一个第四信道状态矩阵，其输出与第五信道状态矩阵尽可能相近。
前面的流程中，以AI实体与网络设备相互独立为例描述，AI实体还可以为网络设备的一个模块。如果AI实体为网络设备的一个模块，即AI实体为网络设备的一部分，那么网络设备可以不发送第四信道状态矩阵和第五信道状态矩阵。第四信道状态矩阵和第五信道状态矩阵可以通过内部传递至网络设备中的AI实体。
其中，一种可能的实现方式中，一个第四信道状态矩阵和其对应的第五信道状态矩阵可以作为一个训练数据。AI实体可以获得多个训练数据，训练数据的具体数量并不限定。
另一种可能的实现方式中，可以将一个第三信道状态矩阵和其对应的第五信道状态矩阵可以作为一个训练数据，由AI实体将第三信道状态矩阵转换为第四信道状态矩阵，此时AI实体获取到的训练数据为第三信道状态矩阵和第五信道状态矩阵。
通过上面的过程，网络设备可以获得用于训练神经网络的多个训练数据，多个训练数据可以构成一个数据集。本申请实施例中，一种实现方式中，不同的网络部署环境可以采用不同的训练数据集，例如工厂环境采用一个数据集，办公室环境采用另一个数据集，或者每个终端设备对应的训练数据构成一个数据集，或者每个稀疏图样对应的训练数据构成一个数据集，根据不同数据集训练得到不同的AI模型。另一种实现方式中，将收集得到的训练数据先进行聚类，相类似的训练数据形成一个新的数据集，并利用新的数据集进行模型训练。收集得到的数据集可能包括终端设备在不同地理位置上发送的第二参考信号对应的上行信道状态矩阵。
本申请实施例中，AI实体获得训练数据后，可以选择合适的AI模型，并利用训练数据对AI模型进行训练。AI模型的具体结构并不限定，例如，如图5所示，为本申请实施例提供的一种AI模型示意图。
该AI模型中的神经网络的输入为稀疏上行信道状态矩阵
Figure PCTCN2022111628-appb-000001
输出为恢复的上行信道状态矩阵
Figure PCTCN2022111628-appb-000002
该神经网络可以看做是从
Figure PCTCN2022111628-appb-000003
到
Figure PCTCN2022111628-appb-000004
的映射函数，即
Figure PCTCN2022111628-appb-000005
训练该神经网络的目的是希望输出的
Figure PCTCN2022111628-appb-000006
与输入的稀疏上行信道状态矩阵对应的完整的上行信道状态矩阵
Figure PCTCN2022111628-appb-000007
尽可能相近。训练神经网络需要定义损失函数，损失函数的选取与任务的目标有关，常用的损失函数有均方误差、交叉熵(cross entropy)等，例如，本申请实施例的损失函数可以用恢复的上行信道状态矩阵与完整的上行信道状态矩阵之间的均方误差，即
Figure PCTCN2022111628-appb-000008
N为矩阵
Figure PCTCN2022111628-appb-000009
包括的元素的数量。
对于AI模型的训练的具体过程，本申请实施例对此并不限定，例如可以采用的随机梯度下降法进行训练，还可以通过迭代算法进行训练，以得到最优的神经网络参数。当获得新的训练数据时，还可以对AI模型进行更新。AI模型的更新可以是周期性的，例如AI实体每隔一定时长获得一个新的数据集，基于该数据集对AI模型进行更新；也可以是事件触发的，当输出的误差超过一个阈值，则基于最新的数据集对AI模型进行更新。
AI模型的选择还可以考虑复杂度和性能的折中。以神经网络为例，在训练数据足够多的情况下，神经网络的层数和神经元越多，AI模型的复杂度越高，性能通常也更好。对于训练数据一定，则可能出现AI模型的参数过多，出现过拟合的情况，即模型在训练集上性能很好，但在测试集上性能不好。因此，需要结合实际应用场景考虑AI模型的选择。对于计算能力强的网络设备，例如宏站，可以采用参数更多的AI模型；对于计算能力弱的网络设备，例如小站、微站等，可以采用参数更少的AI模型。
本申请实施例中，网络设备可以针对不同的稀疏图样训练多个神经网络，例如，网络设备预先确定了3个稀疏图样，则针对每个稀疏图样分别训练神经网络，得到3个神经网络，即神经网络与稀疏图样存在对应关系，网络设备可以根据实际情况，选择使用其中一个神经网络进行下行CSI恢复。
上面的过程中，网络设备通过对接收到的第二参考信号进行信道估计，获得第三信道状态矩阵，在将第三信道状态矩阵转换为第四信道状态矩阵。再通过第一稀疏图样对第四信道状态矩阵在至少一个维度进行元素抽取，获得第五信道状态矩阵，第五信道状态矩阵是由从第四信道状态矩阵中抽取的元素构成的，相当于通过第一稀疏图样对第四信道状态矩阵进行了压缩，第五信道状态矩阵可以认为是第四信道状态矩阵压缩后的矩阵。进一步的，通过第四信道状态矩阵和第五信道状态矩阵对神经网络进行训练，使得神经网络对第五信道状态矩阵进行恢复处理后获得的信道状态矩阵，与第四信道状态矩阵之间的差异足够小，例如他们之间的均方误差小于预设值，即恢复的信道状态矩阵接近于第四信道状态矩阵。
本申请实施例中，通过第一稀疏图样对AI模型中的神经网络训练完成之后，在下行信道中，可以采用该神经网络，恢复下行信道状态矩阵。具体的，终端设备可以对接收到的下行参考信号进行信道估计，获得下行信道状态矩阵。终端设备再通过第一稀疏图样对下行信道状态矩阵在至少一个维度进行元素抽取，获得稀疏下行信道状态矩阵，稀疏下行信道状态矩阵可以认为是下行信道状态矩阵压缩后的矩阵。进一步的，终端设备向网络设备发送稀疏下行信道状态矩阵，网络设备获得稀疏下行信道状态矩阵之后，可以采用训练好的神经网络对稀疏下行信道状态矩阵进行恢复处理，获得恢复的下行信道状态矩阵。恢复的下行信道状态矩阵与终端设备估计得到的下行信道状态矩阵之间的差异足够小，例如它们之间的均方误差小于预设值，即恢复的上行信道状态矩阵接近于终端设备估计得到的信道状态矩阵，下面详细描述。
如图6所示，为本申请实施例提供的一种信道恢复方法流程示意图。该方法包括：
可选地，S601：网络设备向终端设备发送第一信息，第一信息用于指示第一稀疏图样。
可选地，S602：终端设备接收来自网络设备的第一信息。
网络设备可以通过无线资源控制(radio resource control，RRC)消息或者媒体接入控制(medium access control，MAC)控制元素(control element，CE)或者下行控制信息(downlink control information，DCI)向终端设备发送第一信息，本申请实施例并不限定。
本申请实施例中，网络设备可以向终端设备配置一个稀疏图样，也可以配置多个稀疏图样，终端设备可以从多个稀疏图样中选择一个稀疏图样使用，也可以由网络设备从多个稀疏图样中激活一个稀疏图样，终端设备使用激活的稀疏图样。
本申请实施例中，第一稀疏图样可以指示在下行信道状态矩阵的至少一个维度上的抽取的元素的索引，下行信道状态矩阵为终端设备根据下行参考信号进行信道估计获得的矩阵。也就是说，第一稀疏图样可以用于对下行信道状态矩阵的至少一个维度进行抽样。
举例来说，以下行参考信号为第一参考信号进行描述，下行信道状态矩阵的维度包含终端设备的天线(对应接收第一参考信号的天线)维度、网络设备的天线(对应发送第一参考信号的天线)维度和第一参考信号对应的频域维度时，第一稀疏图样可以指示终端设备的天线维度中的至少一个天线的索引、网络设备的天线维度中的至少一个天线的索引以及第一参考信号对应的频域维度中的至少一个频域单元的索引中的至少一项。其中，网络 设备的天线用于发送第一参考信号，终端设备的天线用于接收第一参考信号，第一参考信号对应的频域单元为承载第一参考信号的频域单元。
例如，网络设备在发送第一参考信号时，发送的天线数为N T BS＝64，终端设备在接收第一参考信号时，接收的天线数为N R US＝4，第一参考信号在频域对应的频域单元数为Nsc＝120，第一参考信号在时域对应的时域单元数为1，则终端设备根据第一参考信号估计得到的下行信道状态矩阵可以为4×64×120的3维矩阵。假如第一稀疏图样指示终端设备的接收天线中索引为1和3的天线，以及指示网络设备发送第一参考信号的天线中索引为1,5,11,…,2n-1,…,61的天线，以及指示索引为1,21,…,20m+1,…,101的频域单元，其中n为正整数，m为大于或等于0的整数。那么根据第一稀疏图样，需要抽取下行信道状态矩阵中与索引为1,5,11,…,61的天线对应的元素，抽取下行信道状态矩阵中与索引为1和3的天线对应的元素，以及抽取下行信道状态矩阵中与索引为1,21,…,101的频域单元对应的元素，从下行信道状态矩阵中抽取的元素构成稀疏下行道状态矩阵，即稀疏下行信道状态矩阵为2×13×6的3维矩阵。
另外，由于上行道状态矩阵和下行道状态矩阵的维度可能不完全相同，其中，上行道状态矩阵的维度包括{终端设备的发送天线，网络设备的接收天线，频域，时域}，下行道状态矩阵的维度包括{终端设备的接收天线，网络设备的发送天线，频域，时域}，因此，需要保证终端设备的接收配置和发送配置匹配，以及网络设备的接收配置和发送配置匹配，接收配置包括接收天线、接收权、接收带宽中的至少一项，接收权也可以叫做接收波束或接收预编码或接收的空域滤波器，发送配置包括发送天线、发送权、发送带宽中的至少一项，发送权也可以叫做发送波束或发送预编码或发送的空域滤波器，其中接收天线和发送天线配置可通过建立终端设备的发送天线和终端设备的接收天线之间的第一关联关系，和/或网络设备的发送天线和网络设备的接收天线之间的第二关联关系实现，第一关联关系和第二关联关系可以通过协议预定义，也可以通过网络设备配置，也可以通过设备出厂预定义实现，例如第一关联关系为终端设备根据自己的内部配置确定的，终端的发送天线1对应接收天线1，发送天线3对应接收天线2，终端设备可将第一关联关系上报给网络设备，又例如第二关联关系为网络设备通知给终端设备的。接收权和发送权的匹配指的是终端设备和网络设备在接收和发送参考信号时使用相同的权值。接收带宽和发送带宽指的是终端设备和网络设备在接收和发送参考信号时使用相同的带宽，或者使得上行信道的带宽大于或等于下行信道的带宽，则在使用上行信道训练神经网络时，可通过在上行信道的带宽中抽取部分带宽，使得抽取后上行信道的带宽与下行信道的带宽相等。
本申请实施例中，为了描述方便，将下行信道状态矩阵称为第一信道状态矩阵，将下行信道状态矩阵对应的稀疏下行信道矩阵称为稀疏信道状态矩阵。
S603：网络设备向终端设备发送第一参考信号，相应的，终端设备接收来自网络设备的第一参考信号。
第一参考信号可以为下行参考信号，网络设备可以向终端设备发送多个第一参考信号，本申请实施例对此并不限定。
一种可能的实现方式中，终端设备接收第一参考信号时使用的接收波束与发送第二参考信号时使用的发送波束相同，也就是说，终端设备在接收第一参考信号时使用的空域滤波器，与发送第二参考信号使用的空域滤波器相同。为了实现上述目的，终端设备发送第二参考信号所使用的第二预编码与终端设备接收第一参考信号所使用的第一预编码相同。
同样的，网络设备接收第二参考信号时使用的接收波束与发送第一参考信号时使用的发送波束相同，即网络设备发送第一参考信号所使用的第三预编码与网络设备接收第二参考信号所使用的第四预编码相同。
举例来说，终端设备发送第二参考信号时，网络设备接收到的第二参考信号可以表示为
Figure PCTCN2022111628-appb-000010
其中
Figure PCTCN2022111628-appb-000011
为网络设备接收第二参考信号时使用的第四预编码，
Figure PCTCN2022111628-appb-000012
为终端设备发送第二参考信号时使用的第二预编码，如果中
Figure PCTCN2022111628-appb-000013
和
Figure PCTCN2022111628-appb-000014
是单位矩阵，则表示未采用预编码，H ul为上行信道矩阵，s ul为第二参考信号，n为噪声，网络设备根据第二参考信号估计的上行信道状态矩阵为
Figure PCTCN2022111628-appb-000015
即训练神经网络时使用的上行信道状态矩阵为包含预编码的等效信道矩阵。
为了实现通过稀疏信道状态矩阵恢复第一信道状态矩阵，本申请实施例中，输入神经网络的稀疏信道状态矩阵与
Figure PCTCN2022111628-appb-000016
可以存在相关性。具体的，网络设备在发送第一参考信号时，终端设备接收到的第一参考信号可以表示为
Figure PCTCN2022111628-appb-000017
其中
Figure PCTCN2022111628-appb-000018
为终端设备接收第一参考信号时使用的第一预编码，
Figure PCTCN2022111628-appb-000019
分别为网络设备发送第一参考信号时使用的第三预编码，H dl为下行信道矩阵，s dl为第一参考信号，终端设备根据第一参考信号估计的第一信道状态矩阵(即上行信道状态矩阵)为
Figure PCTCN2022111628-appb-000020
上行信道矩阵H ul和下行信道矩阵H dl之间存在相似的相关性，当
Figure PCTCN2022111628-appb-000021
和
Figure PCTCN2022111628-appb-000022
时，
Figure PCTCN2022111628-appb-000023
和
Figure PCTCN2022111628-appb-000024
之间也存在相关性，也就是第二预编码与第一预编码相同，第三预编码与第四预编码相同，即终端设备发送第二参考信号的发送波束和接收第二参考信号的接收波束相同，网络设备发送第一参考信号的发送波束和接收第一参考信号的接收波束相同。
本申请实施例中，终端设备进行信道估计的方法，本申请并不限定，可以是传统信道估计算法，例如，MMSE算法，也可以是基于神经网络的信道估计算法。
S604：终端设备根据第一参考信号和第一稀疏图样确定稀疏信道状态矩阵。
第一种可能的实现方式中，终端设备可以根据第一参考信号获得第一信道状态矩阵，具体的，终端设备可以根据第一参考信号进行信道估计，获得第一信道状态矩阵。第一信道状态信息可以指示出网络设备和终端设备之间下行信道的信道响应，第一信道状态信息可以为一个多维矩阵，例如，第一信道状态矩阵的维度包含终端设备的天线维度、网络设备的天线维度和第一参考信号对应的子载波维度中的至少一个维度。终端设备可以根据第一稀疏图样和第一信道状态矩阵，获得稀疏信道状态矩阵，例如根据第一稀疏图样对第一信道状态矩阵进行抽取，获得稀疏信道状态矩阵，或者根据第一稀疏图样对第一信道状态矩阵进行抽取，并对未包括在第一信道状态矩阵中的元素进行估计，获得稀疏信道状态矩阵。稀疏信道状态矩阵相当于对第一信道状态矩阵进行压缩后的矩阵。
第二种可能的实现方式中，网络设备无需向终端设备发送第一信息，即网络设备无需向终端设备发送第一稀疏图样，则终端设备可认为根据第一参考信号获得第一信道状态矩阵也是稀疏信道状态矩阵。例如，网络设备在发送第一参考信号时，使用的发送天线和频域资源已经与预先确定的第一稀疏图样一致，则终端设备根据第一参考信号估计第一参考信号所对应的资源上的信道，即为稀疏信道状态矩阵。
第三种可能的实现方式中，终端设备可以使用第一稀疏图样所指示的终端设备天线索引对应的天线接收第一参考信号，和/或，对第一稀疏图样所指示的网络设备天线索引对应的天线以及频域索引对应的频域资源上的信道进行估计，直接得到稀疏信道状态矩阵。
例如，频域上的频域单元为子载波，假如第一稀疏图样指示终端设备的天线中索引为 1和3的天线，以及指示网络设备的天线中索引为1,5,11,…,2n-1,…,61的天线，以及指示索引为1,21,…,20m+1,…,101的子载波，其中n为正整数，m为大于或等于0的整数。那么终端设备在接收第一参考信号时，使用索引为1和3的天线，且估计索引为1,21,…,101的子载波和网络设备中索引为1,5,11,…,2n-1,…,61的天线对应的信道，终端设备可以得到维度为2×13×6的稀疏信道状态矩阵。
本申请实施例中，第一信道状态矩阵在子载波维度的长度等于第四信道状态矩阵在子载波维度的长度，第一信道状态矩阵在接收天线维度的长度等于第四信道状态矩阵在发送天线维度的长度，第一信道状态矩阵在发送天线维度的长度等于第四信道状态矩阵在接收天线维度的长度。
S605：终端设备向网络设备发送第一信道信息，相应的，网络设备接收来自终端设备的第一信道信息。
其中，第一信道信息用于指示稀疏信道状态矩阵。
其中，第一信道信息可以是稀疏信道状态矩阵中元素的取值的集合，也可以是其它形式，本申请实施例对此并不限定。
S606：网络设备通过神经网络对第一信道信息指示的稀疏信道状态矩阵进行处理，获得第二信道状态矩阵。
其中，第二信道状态矩阵为对第一信道状态矩阵的恢复值；神经网络是通过根据第一稀疏图样抽样的数据训练的，具体参考前面图4所示的流程中的描述。
网络设备可以根据恢复的第二信道状态矩阵确定向终端设备发送数据时所使用的预编码、资源等参数。
另外，网络设备需要应用神经网络恢复第一信道状态矩阵，因此如果神经网络的训练是在网络设备或者位于网络设备的AI模块上完成的，则不需要在网络设备中配置训练好的神经网络；如果神经网络的训练是在独立的AI网元上完成的，则需要将训练好的神经网络配置到网络设备中，具体配置过程，本申请实施例对此并不限定，在此不再赘述。
通过上述方法，可以在不需要终端设备参与神经网络训练的情况下，网络设备仅通过上行信道状态矩阵训练神经网络，并通过该神经网络和终端设备反馈的稀疏下行信道状态矩阵恢复完整的下行信道状态矩阵，可以降低对终端设备能力的要求，且避免终端设备反馈完整的下行信道状态矩阵导致的反馈开销。
本申请实施例还提供一种方法，可以用于终端设备反馈信道特征矩阵，下面将详细描述。本申请实施例中，为了描述方便，以下描述中，将网络设备向终端设备发送的下行参考信号称为第一参考信号，将终端设备向网络设备发送的上行参考信号称为第二参考信号，将根据第一参考信号确定的下行信道状态矩阵称为第一矩阵，将根据下行信道状态矩阵确定的信道特征矩阵称为第一信道特征矩阵，将第一信道特征矩阵对应的稀疏下行信道特征矩阵称为第一稀疏矩阵；将通过神经网络对第一稀疏矩阵进行处理，获得的矩阵称为第二矩阵，将网络设备根据第二参考信号确定的上行信道状态矩阵称为第三矩阵，将根据上行信道状态矩阵确定的信道特征矩阵称为第二信道特征矩阵，将对第二信道特征矩阵进行元素抽取获得的矩阵称为第二稀疏矩阵，将通过神经网络对第二稀疏矩阵进行处理，获得的矩阵称为第四矩阵。
其中，上行道状态矩阵的维度包括{终端设备的发送天线，网络设备的接收天线，频域，时域}中的至少一个维度，如果某个维度的长度为1，则也可认为这个维度不存在，其 中频域的粒度可以是子载波或RB，时域的粒度可以是OFDM符号或者时隙等。下行道状态矩阵的维度包括{终端设备的接收天线，网络设备的发送天线，频域，时域}中的至少一个维度。
如图7所示，为本申请实施例提供的一种神经网络的模型训练流程示意图。该流程涉及终端设备、网络设备以及AI实体之间的交互。其中，AI实体可以为独立的网元，也可以位于网络设备中。
图7所示的流程中，数据收集的具体方法为，终端设备发送第二参考信号，例如SRS，网络设备根据接收到的第二参考信号进行信道估计，获取第三矩阵。根据第三矩阵获得第二信道特征矩阵，再根据预先确定好的第二稀疏图样，从第二信道特征矩阵抽取部分元素，得到第二稀疏矩阵。第二稀疏矩阵和第二信道特征矩阵可以作为一个训练数据，数据收集阶段需要收集多个训练数据，收集的样本数量可以根据实际情况确定，本申请并不限定。
S701：终端设备向网络设备发送第二参考信号，相应的，网络设备接收来自终端设备的第二参考信号。
其中，终端设备可以向网络设备发送多个第二参考信号，具体数量并不限定。第二参考信号是上行参考信号，例如可以为SRS，具体类型并不限定。
S702：网络设备根据第二参考信号进行信道估计，获得上行信道状态信息。
网络设备对每个第二参考信号进行信道估计可以获得上行信道状态信息，上行信道状态信息可以指示出网络设备和终端设备之间上行信道的信道响应，上行信道状态信息的维度可以包括{终端设备的发送天线，网络设备的接收天线，频域，时域}中的至少一个维度。其中，对于上行信道，网络设备的天线，是指接收第二参考信号的接收天线；子载波，是指第二参考信号对应的子载波。
举例来说，网络设备在接收第二参考信号时，网络设备采用的接收天线数为N R BS＝64，第二参考信号对应的子载波数为Nsc＝120，则第三矩阵为64×120的2维矩阵。
S703：网络设备根据上行信道状态信息确定第二稀疏矩阵。
具体的，以上行信道状态信息为上行信道状态矩阵为例，网络设备可以根据上行信道状态信息确定第二空域协方差矩阵。
第二空域协方差矩阵可以由一个上行信道计算得到，例如第二空域协方差矩阵可以满足以下形式：
Figure PCTCN2022111628-appb-000025
其中，R i表示每个终端天线、频域单元、时域单元对应的空域协方差矩阵，H i表示每个终端天线、频域单元、时域单元对应的信道向量。当上行信道状态信息的维度为a*b*c*网络设备天线时，得到的第二空域协方差矩阵的维度为a*b*c*网络设备天线*网络设备天线。第二空域协方差矩阵也可对除了网络设备天线以外的其他维度取平均，例如对b维度取平均，则得到的第二空域协方差矩阵的维度为a*c*网络设备天线*网络设备天线，又如对a、b、c维度都取平均，则得到的第二空域协方差矩阵的维度为网络设备天线*网络设备天线。
第二空域协方差矩阵也可以由多个上行信道状态矩阵计算得到，例如对一段时间内的多个上行信道状态矩阵的空域协方差矩阵求平均得到，本申请实施例不做限制。
网络设备可以对第二空域协方差矩阵进行特征值分解(eigen value decomposition，EVD)，获得第二信道特征矩阵，或者，也可以直接对上行信道状态矩阵进行奇异值分解(singular  value decomposition，SVD)，获得第二信道特征矩阵。当保留所有特征值(或奇异值)对应的特征向量时，第二信道特征矩阵的维度可以是网络设备的接收天线*网络设备的接收天线，或者还包括{终端设备的发送天线，频域，时域}中的至少一个维度。当仅保留最大的n个特征值(或奇异值)对应的特征向量时，第二信道特征矩阵的维度可以是网络设备的接收天线*n，或者还包括{终端设备的发送天线，频域，时域}中的至少一个维度。
进一步的，网络设备可以根据第二稀疏图样从第二信道特征矩阵中抽取部分元素，获得第二稀疏矩阵，第二稀疏矩阵包括从第二信道特征矩阵中抽取的部分元素。也就是说第二稀疏图样可以用于从第二信道特征矩阵中抽取部分元素，从而获得第二稀疏矩阵。
该第二稀疏图样为网络设备预先确定的，终端设备进行信道状态矩阵反馈阶段也使用第二稀疏图样，与网络设备使用的第二稀疏图样相同。
一种可能的实现方式中，当第二稀疏图样用于对第二信道特征矩阵的至少一个维度上的元素抽取时，第二稀疏图样可以指示网络设备的天线维度中的至少一个天线的索引、终端设备的天线维度中的至少一个天线的索引、第二参考信号对应的频域维度中的至少一个频域单元的索引，以及第二参考信号对应的时域维度中的至少一个时域单元的索引中的至少一项。
举例来说，网络设备在接收第二参考信号时，网络设备采用的天线数为N R BS＝64，第二参考信号在频域对应的频域单元数为Nsc＝120，终端设备在发送第二参考信号时，终端设备采用的天线数为1，第二参考信号在时域对应的时域单元数为1，则上行信道状态信息对应的上行信道状态矩阵为64×120的2维矩阵。根据上行信道状态矩阵确定的第二空域协方差矩阵也为64×120的2维矩阵，同样的，根据第二空域协方差矩阵确定的第二信道特征矩阵也为64×120的2维矩阵。第二稀疏图样指示网络设备中索引为1,5,11,…,2n-1,…,61的天线，n为正整数，以及第二稀疏图样指示索引为1,21,…,20m+1,…,101的频域单元，m为大于或等于0的整数。那么该第二稀疏图样可以用于抽取第二信道特征矩阵中与索引为1,5,11,…,61的天线对应的元素，以及抽取第三矩阵中与索引为1,21,…,101的频域单元对应的元素，从第二信道特征矩阵中抽取的元素构成第二稀疏矩阵，此时第二稀疏矩阵为13×6的2维矩阵。
S704：网络设备向AI实体发送第二信道特征矩阵和第二稀疏矩阵，相应的，AI实体接收第二信道特征矩阵和第二稀疏矩阵。
S705：AI实体根据第二信道特征矩阵和第二稀疏矩阵对神经网络进行训练。
前面的流程中，以AI实体与网络设备相互独立为例描述，AI实体还可以为网络设备的一个模块。如果AI实体为网络设备的一个模块，即AI实体为网络设备的一部分，那么网络设备可以不发送第二信道特征矩阵和第二稀疏矩阵。第二信道特征矩阵和第二稀疏矩阵可以通过内部传递至网络设备中的AI实体。
其中，一个第二信道特征矩阵和其对应的第二稀疏矩阵可以作为一个训练数据，AI实体可以获得多个训练数据，训练数据的具体数量并不限定。其中，第二稀疏矩阵可以作为训练数据中的训练样本，即是在训练神经网络时输入神经网络的数据；第二信道特征矩阵可以作为训练数据中的样本标签，即神经网络输入样本时期望得到的输出值，可以理解为训练样本对应的真实值。训练神经网络的目的就是希望输入神经网络一个第二稀疏矩阵，其输出与第二信道特征矩阵尽可能相近。
通过上面的过程，网络设备可以获得用于训练神经网络的多个训练数据，多个训练数 据可以构成一个数据集。本申请实施例中，一种实现方式中，不同的网络部署环境可以采用不同的训练数据集，例如工厂环境采用一个数据集，办公室环境采用另一个数据集，或者每个终端设备对应的训练数据构成一个数据集，或者每个稀疏图样对应的训练数据构成一个数据集，根据不同数据集训练得到不同的AI模型。另一种实现方式中，将收集得到的训练数据先进行聚类，相类似的训练数据形成一个新的数据集，并利用新的数据集进行模型训练。收集得到的数据集可能包括终端设备在不同地理位置上发送的第二参考信号对应的上行信道状态矩阵。
本申请实施例中，AI实体获得训练数据后，可以选择合适的AI模型，并利用训练数据对AI模型进行训练。AI模型的具体结构并不限定，例如，如图8所示，为本申请实施例提供的一种AI模型示意图。
该神经网络输入的第二稀疏矩阵为
Figure PCTCN2022111628-appb-000026
输出的第四矩阵为
Figure PCTCN2022111628-appb-000027
该神经网络可以看做是从
Figure PCTCN2022111628-appb-000028
到
Figure PCTCN2022111628-appb-000029
的映射函数，即
Figure PCTCN2022111628-appb-000030
训练该神经网络的目的是希望输出的
Figure PCTCN2022111628-appb-000031
与实际的第二信道特征矩阵
Figure PCTCN2022111628-appb-000032
尽可能相近。本申请实施例中，可以为训练神经网络定义损失函数，损失函数可以为输出的第四矩阵与实际的第二信道特征矩阵之间的均方误差，即
Figure PCTCN2022111628-appb-000033
N为矩阵
Figure PCTCN2022111628-appb-000034
包括的元素的数量。
对于AI模型的训练的具体过程，本申请实施例对此并不限定，例如可以采用的随机梯度下降法进行训练，还可以通过迭代算法进行训练，以得到最优的神经网络参数。当获得新的训练数据时，还可以对AI模型进行更新。AI模型的更新可以是周期性的，例如AI实体每隔一定时长获得一个新的数据集，基于该数据集对AI模型进行更新；也可以是事件触发的，当神经网络输出的误差超过一个阈值，则基于最新的数据集对AI模型进行更新。
AI模型的选择还可以考虑复杂度和性能的折中。以神经网络为例，在训练数据足够多的情况下，神经网络的层数和神经元越多，AI模型的复杂度越高，性能通常也更好。对于训练数据一定，则可能出现AI模型的参数过多，出现过拟合的情况，即模型在训练集上性能很好，但在测试集上性能不好。因此，需要结合实际应用场景考虑AI模型的选择。对于计算能力强的网络设备，例如宏站，可以采用参数更多的AI模型；对于计算能力弱的网络设备，例如小站、微站等，可以采用参数更少的AI模型。
本申请实施例中，网络设备可以针对不同的稀疏图样训练多个神经网络，例如，网络设备预先确定了3个稀疏图样，则针对每个稀疏图样分别训练神经网络，得到3个神经网络，即神经网络与稀疏图样存在对应关系，网络设备可以根据实际情况，选择使用其中一个神经网络进行下行CSI恢复。
本申请实施例中，通过第二稀疏图样对AI模型中的神经网络训练完成之后，在下行信道中，可以采用该神经网络，恢复第一信道特征矩阵。具体的，终端设备可以对接收到的下行参考信号进行信道估计，获得第一矩阵。终端设备再根据第一矩阵确定第一信道特征矩阵，通过第二稀疏图样对第一信道特征矩阵在至少一个维度进行元素抽取，获得第一稀疏矩阵，第一稀疏矩阵可以认为是第一信道特征矩阵压缩后的矩阵。进一步的，终端设备向网络设备发送第一稀疏矩阵，网络设备获得第一稀疏矩阵之后，可以采用训练好的神经网络对第一稀疏矩阵进行恢复处理，获得第二矩阵。第二矩阵与终端设备估计得到的第一信道特征矩阵之间的均方误差，小于预设值，即第二矩阵接近于第一信道特征矩阵，下 面详细描述。
如图9所示，为本申请实施例提供的一种信道恢复方法流程示意图。该方法包括：
可选地，S901：网络设备向终端设备发送第二信息，相应的，终端设备接收来自网络设备的第二信息。
其中，第二信息用于指示第二稀疏图样。
网络设备可以通过RRC消息或者MAC CE或者DCI向终端设备发送第二信息，本申请实施例并不限定。
本申请实施例中，网络设备可以向终端设备配置多个稀疏图样，终端设备可以从多个稀疏图样中选择一个稀疏图样使用，也可以由网络设备从多个稀疏图样中激活一个稀疏图样，终端设备使用激活的稀疏图样。
本申请实施例中，第二稀疏图样可以用于对第一信道特征矩阵的至少一个维度进行抽样。
举例来说，以下行参考信号为第一参考信号进行描述，第一信道特征矩阵的维度包含网络设备的天线维度和第一参考信号对应的频域维度时，第二稀疏图样可以指示网络设备的天线维度中的至少一个天线的索引以及第一参考信号对应的频域维度中的至少一个频域单元的索引中的至少一项。
另外，由于上行道状态矩阵和下行道状态矩阵的维度可能不完全相同，其中，上行道状态矩阵的维度包括{终端设备的发送天线，网络设备的接收天线，频域，时域}，下行道状态矩阵的维度包括{终端设备的接收天线，网络设备的发送天线，频域，时域}，因此，需要保证终端设备的接收配置和发送配置匹配，以及网络设备的接收配置和发送配置匹配，接收配置包括接收天线、接收权、接收带宽中的至少一项，接收权也可以叫做接收波束或接收预编码或接收的空域滤波器，发送配置包括发送天线、发送权、发送带宽中的至少一项，发送权也可以叫做发送波束或发送预编码或发送的空域滤波器，其中接收天线和发送天线配置可通过建立终端设备的发送天线和终端设备的接收天线之间的第一关联关系，和/或网络设备的发送天线和网络设备的接收天线之间的第二关联关系实现，第一关联关系和第二关联关系可以通过协议预定义，也可以通过网络设备配置，也可以通过设备出厂预定义实现，例如第一关联关系为终端设备根据自己的内部配置确定的，终端的发送天线1对应接收天线1，发送天线3对应接收天线2，终端设备可将第一关联关系上报给网络设备，又例如第二关联关系为网络设备通知给终端设备的。接收权和发送权的匹配指的是终端设备和网络设备在接收和发送参考信号时使用相同的权值。接收带宽和发送带宽指的是终端设备和网络设备在接收和发送参考信号时使用相同的带宽，或者使得上行信道的带宽大于或等于下行信道的带宽，则在使用上行信道训练神经网络时，可通过在上行信道的带宽中抽取部分带宽，使得抽取后上行信道的带宽与下行信道的带宽相等。
S902：网络设备向终端设备发送第一参考信号，相应的，终端设备接收来自网络设备的第一参考信号。
第一参考信号可以为下行参考信号，网络设备可以向终端设备发送多个第一参考信号，本申请实施例对此并不限定。
一种可能的实现方式中，终端设备接收第一参考信号时使用的接收波束与发送第二参考信号时使用的发送波束相同，也就是说，终端设备在接收第一参考信号时使用的空域滤波器，与发送第二参考信号使用的空域滤波器相同。为了实现上述目的，终端设备发送第 二参考信号所使用的第二预编码与终端设备接收第一参考信号所使用的第一预编码相同。
同样的，网络设备接收第二参考信号时使用的接收波束与发送第一参考信号时使用的发送波束相同，即网络设备发送第一参考信号所使用的第三预编码与网络设备接收第二参考信号所使用的第四预编码相同。
本申请实施例中，终端设备进行信道估计的方法，本申请并不限定，可以是传统信道估计算法，例如，算法，也可以是基于神经网络的信道估计算法。
S903：终端设备根据第一参考信号确定第一矩阵，根据第一矩阵确定第一信道特征矩阵，根据第一信道特征矩阵和第二稀疏图样确定第一稀疏矩阵。
一种可能的实现方式中，终端设备可以根据第一参考信号进行信道估计，获得第一矩阵。终端设备可以根据第一矩阵确定第一空域协方差矩阵，并对第一空域协方差矩阵进行奇异值分解，获得第一信道特征矩阵。根据第一矩阵确定第一空域协方差矩阵的具体过程，可以参考第二空域协方差矩阵的描述，在此不再赘述。
进一步的，终端设备可以根据第二稀疏图样从第一信道特征矩阵中抽取部分元素，获得第一稀疏矩阵，第一稀疏矩阵包括从第一信道特征矩阵中抽取的部分元素。也就是说第二稀疏图样可以用于从第一信道特征矩阵中抽取部分元素，从而获得第一稀疏矩阵。第一稀疏矩阵相当于对第一信道特征矩阵进行压缩后的矩阵。
S904：终端设备向网络设备发送第一信道信息，相应的，网络设备接收来自所述终端设备的第一信道信息。
其中，第一信道信息用于指示稀疏信道状态矩阵。
其中，第一信道信息可以是稀疏信道状态矩阵中元素的取值的集合，也可以是稀疏信道状态矩阵中元素的量化值的集合，也可以是其它形式，本申请实施例对此并不限定。
S905：网络设备通过神经网络对第一稀疏矩阵进行处理，获得第二矩阵。
其中，第二矩阵为对第一信道特征矩阵的恢复值；神经网络是通过根据第二稀疏图样抽样的数据训练的，具体参考前面图4所示的流程中的描述。
网络设备可以根据恢复的第二矩阵确定向终端设备发送数据时所使用的PMI以及RI等参数。
另外，网络设备需要应用神经网络恢复第一信道特征矩阵，因此如果神经网络的训练是在网络设备或者位于网络设备的AI模块上完成的，则不需要在网络设备中配置训练好的神经网络；如果神经网络的训练是在独立的AI网元上完成的，则需要将训练好的神经网络配置到网络设备中，具体配置过程，本申请实施例对此并不限定，在此不再赘述。
通过上述方法，可以在不需要终端设备参与神经网络训练的情况下，网络设备仅通过第二稀疏矩阵和第二信道特征矩阵训练神经网络，可以降低对终端设备能力的要求，且避免终端设备反馈完整的第一信道特征矩阵导致的反馈开销。
一种可能的实现方式中，在图3至图8的流程中，第一参考信号的带宽可以小于或者等于第二参考信号的带宽。第一参考信号对应的子载波间隔大于或者等于第二参考信号对应的子载波间隔。
在该实现方式中，网络设备可以给终端设备配置一组特殊的第一参考信号，该第一参考信号的带宽小于或者等于第二参考信号的带宽，该第一参考信号的子载波间隔大于或者等于第二参考信号的子载波间隔。
另一种实现方式中，第一参考信号的带宽和子载波间隔可以为网络设备独立配置的， 即第一参考信号的带宽可以大于或者等于终端设备的上行激活带宽部分(bandwidth part，BWP)的带宽，和/或第一参考信号的子载波间隔与终端设备的上行激活BWP的子载波间隔无关，即第一参考信号的子载波间隔与终端设备的上行激活BWP的子载波间隔可以不相等，上行激活BWP中包括第一参考信号的区域的子载波间隔，与上行激活BWP中不包括第一参考信号的区域的子载波间隔可以不同，其中，第一参考信号可以位于上行激活BWP中承载。第一参考信号在频域可以分为多段，只要多段的总带宽满足网络设备恢复下行信道信息的要求即可。
通过配置第一参考信号的带宽和子载波间隔与第二参考信号相同，使得本申请实施例方案在第一参考信号的带宽、子载波间隔和第二参考信号的带宽、子载波间隔不一致的情况也能使用，使得网络设备能够使用神经网络恢复出完整的下行信道信息。
上述各个实施例可以分别单独实施，或者也可以相互结合实施。上文中，在不同实施例中，侧重描述了各个实施例的区别之处，除区别之处的其它内容，不同实施例之间的其它内容可以相互参照。上述实施例所描述的各个流程图的步骤编号仅为执行流程的一种示例，并不构成对步骤执行的先后顺序的限制，本申请实施例中相互之间没有时序依赖关系的步骤之间没有严格的执行顺序。此外，各个流程图中所示意的步骤并非全部是必须执行的步骤，可以根据实际需要在各个流程图的基础上增添或者删除部分步骤。
为了实现上述本申请实施例提供的方法中的各功能，网络设备、终端设备或上述通信装置可以包括硬件结构和/或软件模块，以硬件结构、软件模块、或硬件结构加软件模块的形式来实现上述各功能。上述各功能中的某个功能以硬件结构、软件模块、还是硬件结构加软件模块的方式来执行，取决于技术方案的特定应用和设计约束条件。
本申请实施例中对模块的划分是示意性的，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式。另外，在本申请各个实施例中的各功能模块可以集成在一个处理器中，也可以是单独物理存在，也可以两个或两个以上模块集成在一个模块中。上述集成的模块既可以采用硬件的形式实现，也可以采用软件功能模块的形式实现。
与上述构思相同，如图10所示，本申请实施例还提供一种通信装置1000。所述通信装置1000可以是图1中的网络设备，用于实现上述方法实施例中对于网络设备的方法。所述通信装置也可以是图1中的核心网设备，用于实现上述方法实施例中对应于核心网设备的方法。具体的功能可以参见上述方法实施例中的说明。
具体的，通信装置1000可以包括：处理单元1001和通信单元1002。本申请实施例中，通信单元也可以称为收发单元，可以包括发送单元和/或接收单元，分别用于执行上文方法实施例中网络设备或终端设备发送和接收的步骤。以下，结合图10至图11详细说明本申请实施例提供的通信装置。
一些可能的实施方式中，上述方法实施例中网络设备的行为和功能可以通过通信装置1000来实现，例如实现图4至图6的实施例中网络设备执行的方法。例如通信装置1000可以为网络设备，也可以为应用于网络设备中的部件(例如芯片或者电路)，也可以是网络设备中的芯片或芯片组或芯片中用于执行相关方法功能的一部分。通信单元1002可以用于执行图4至图6所示的实施例中由网络设备所执行的接收或发送操作，处理单元1001可以用于执行如图4至图6所示的实施例中由网络设备所执行的除了收发操作之外的操作。
该通信装置实现图4至图6所示的流程中的终端设备的功能时：
通信单元，用于接收来自网络设备的第一参考信号；
处理单元，用于根据第一参考信号获得第一信道状态矩阵；根据第一稀疏图样对第一信道状态矩阵进行抽样，获得稀疏信道状态矩阵；第一稀疏图样由网络设备配置，第一稀疏图样用于对第一信道状态矩阵的至少一个维度进行抽样；
通信单元，用于向网络设备发送第一信道信息，第一信道信息用于指示稀疏信道状态矩阵。
在一种可能的设计中，通信单元还用于：向网络设备发送第二参考信号，第二参考信号用于训练第一稀疏图样对应的神经网络，神经网络用于根据稀疏信道状态矩阵恢复第一信道状态矩阵。
在一种可能的设计中，发送第二参考信号所使用的第二预编码与接收第一参考信号所使用的第一预编码相同。
在一种可能的设计中，第一参考信号的带宽小于或者等于第二参考信号的带宽。
在一种可能的设计中，第一参考信号对应的子载波间隔与第二参考信号对应的子载波间隔相同。
在一种可能的设计中，第一信道状态矩阵的维度包含终端设备的天线维度、网络设备的天线维度和第一参考信号对应的频域维度和时域维度中的至少一个维度；
第一稀疏图样指示以下至少一项：终端设备的天线维度中的至少一个天线的索引、网络设备的天线维度中的至少一个天线的索引、频域维度中的至少一个频域单元的索引和时域维度中的至少一个时域单元的索引。
该通信装置实现图4至图6所示的流程中的网络设备的功能时：
处理单元，用于通过通信单元向终端设备发送第一参考信号；
处理单元，用于通过通信单元接收来自终端设备的第一信道信息，第一信道信息用于指示稀疏信道状态矩阵；稀疏信道状态矩阵为通过第一稀疏图样对第一信道状态矩阵的至少一个维度进行抽样获得的，第一信道状态矩阵根据第一参考信号确定；通过神经网络对稀疏信道状态矩阵进行处理，获得第二信道状态矩阵，第二信道状态矩阵为对第一信道状态矩阵的恢复值；神经网络是通过根据第一稀疏图样抽样的数据训练的。
在一种可能的设计中，通信单元还用于：接收来自终端设备的第二参考信号；
处理单元还用于：根据第二参考信号获得第三信道状态矩阵；根据第三信道状态矩阵获取第四信道状态矩阵，第四信道状态矩阵在子载波维度的长度等于第一信道状态矩阵在子载波维度的长度，第四信道状态矩阵在接收天线维度的长度等于第一信道状态矩阵在发送天线维度的长度，第四信道状态矩阵在发送天线维度的长度等于第一信道状态矩阵在接收天线维度的长度；网络设备根据第一稀疏图样对第四信道状态矩阵的至少一个维度进行抽样获得第五信道状态矩阵；神经网络是通过多个第四信道状态矩阵以及对应的第五信道状态矩阵训练得到的。
在一种可能的设计中，通信单元还用于：接收来自终端设备的第二参考信号，第二参考信号用于训练第一稀疏图样对应的神经网络，神经网络用于根据稀疏信道状态矩阵恢复第一信道状态矩阵。
在一种可能的设计中，接收第二参考信号所使用的第四预编码与发送第一参考信号所使用的第三预编码相同。
在一种可能的设计中，第一参考信号的带宽小于或者等于第二参考信号的带宽。
在一种可能的设计中，第一参考信号对应的子载波间隔与第二参考信号对应的子载波 间隔相同。
在一种可能的设计中，第一信道状态矩阵的维度包含终端设备的天线维度、网络设备的天线维度和第一参考信号对应的频域维度和时域维度中的至少一个维度；
第一稀疏图样指示以下至少一项：终端设备的天线维度中的至少一个天线的索引、网络设备的天线维度中的至少一个天线的索引、频域维度中的至少一个频域单元的索引和时域维度中的至少一个时域单元的索引。
该通信装置实现图7至图9所示的流程中的终端设备的功能时：
通信单元，用于接收来自网络设备的第一参考信号；
处理单元，用于根据第一参考信号进行信道估计，获得第一信道状态矩阵；根据第一信道状态矩阵确定第一信道特征矩阵；根据第二稀疏图样对第一信道特征矩阵进行抽样，获得第一稀疏矩阵；第二稀疏图样由网络设备配置，第二稀疏图样用于对第一信道特征矩阵的至少一个维度进行抽样；
通信单元，用于向网络设备发送第一信道信息，第一信道信息用于指示第一稀疏矩阵。
在一种可能的设计中，通信单元还用于：向网络设备发送第二参考信号，第二参考信号用于训练第一稀疏图样对应的神经网络，神经网络用于根据第一稀疏矩阵恢复第二矩阵。
在一种可能的设计中，发送第二参考信号所使用的第二预编码与接收第一参考信号所使用的第一预编码相同。
在一种可能的设计中，第一参考信号的带宽小于或者等于第二参考信号的带宽。
在一种可能的设计中，第一参考信号对应的子载波间隔与第二参考信号对应的子载波间隔相同。
在一种可能的设计中，第一信道状态矩阵的维度包含终端设备的天线维度、网络设备的天线维度和第一参考信号对应的频域维度和时域维度中的至少一个维度；
第二稀疏图样指示以下至少一项：终端设备的天线维度中的至少一个天线的索引、网络设备的天线维度中的至少一个天线的索引、频域维度中的至少一个频域单元的索引和时域维度中的至少一个时域单元的索引。
该通信装置实现图7至图9所示的流程中的网络设备的功能时：
处理单元，用于通过通信单元向终端设备发送第一参考信号；
处理单元，用于通过通信单元接收来自终端设备的第一信道信息，第一信道信息用于指示第一稀疏矩阵；第一稀疏矩阵为通过第二稀疏图样对第一信道特征向量矩阵的至少一个维度进行抽样获得的，第一信道特征矩阵为通过第一信道状态矩阵确定的，第一信道状态矩阵根据第一参考信号确定；通过神经网络对第一稀疏矩阵进行处理，获得第二矩阵，第二矩阵为对第一信道特征矩阵的恢复值；神经网络是通过根据第二稀疏图样抽样的数据训练的。
在一种可能的设计中，通信单元还用于：接收来自终端设备的第二参考信号，第二参考信号用于训练第二稀疏图样对应的神经网络，神经网络用于根据第一稀疏矩阵恢复第二矩阵。
在一种可能的设计中，接收第二参考信号所使用的第四预编码与发送第一参考信号所使用的第三预编码相同。
在一种可能的设计中，第一参考信号的带宽小于或者等于第二参考信号的带宽。
在一种可能的设计中，第一参考信号对应的子载波间隔与第二参考信号对应的子载波 间隔相同。
在一种可能的设计中，第一信道状态矩阵的维度包含终端设备的天线维度、网络设备的天线维度和第一参考信号对应的频域维度和时域维度中的至少一个维度；
第二稀疏图样指示以下至少一项：终端设备的天线维度中的至少一个天线的索引、网络设备的天线维度中的至少一个天线的索引、频域维度中的至少一个频域单元的索引和时域维度中的至少一个时域单元的索引。
应理解，装置实施例的描述与方法实施例的描述相互对应，如图4至图9中的用于实现终端设备和网络设备的装置结构也可以参照通信装置1000，因此，未详细描述的内容可以参见上文方法实施例，为了简洁，这里不再赘述。
通信单元也可以称为收发器、收发机、收发装置等。处理单元也可以称为处理器，处理单板，处理模块、处理装置等。可选的，可以将通信单元1002中用于实现接收功能的器件视为接收单元，将通信单元1002中用于实现发送功能的器件视为发送单元，即通信单元1002包括接收单元和发送单元。通信单元有时也可以称为收发机、收发器、或收发电路等。接收单元有时也可以称为接收机、接收器、或接收电路等。发送单元有时也可以称为发射机、发射器或者发射电路等。
以上只是示例，处理单元1001和通信单元1002还可以执行其他功能，更详细的描述可以参考图4至图9所示的方法实施例中相关描述，这里不加赘述。
如图11所示为本申请实施例提供的通信装置1100，图11所示的通信装置可以为图10所示的通信装置的一种硬件电路的实现方式。该通信装置可适用于前面所示出的流程图中，执行上述方法实施例中终端设备或者网络设备的功能。为了便于说明，图11仅示出了该通信装置的主要部件。
如图11所示，通信装置1100包括处理器1110和接口电路1120。处理器1110和接口电路1120之间相互耦合。可以理解的是，接口电路1120可以为收发器或输入输出接口。可选的，通信装置1100还可以包括存储器1130，用于存储处理器1110执行的指令或存储处理器1110运行指令所需要的输入数据或存储处理器1110运行指令后产生的数据。
当通信装置1100用于实现图4至图9所示的方法时，处理器1110用于实现上述处理单元1001的功能，接口电路1120用于实现上述通信单元1002的功能。
当上述通信装置为应用于终端设备的芯片时，该终端设备芯片实现上述方法实施例中终端设备的功能。该终端设备芯片从终端设备中的其它模块(如射频模块或天线)接收信息，该信息是网络设备发送给终端设备的；或者，该终端设备芯片向终端设备中的其它模块(如射频模块或天线)发送信息，该信息是终端设备发送给网络设备的。
当上述通信装置为应用于网络设备的芯片时，该网络设备芯片实现上述方法实施例中网络设备的功能。该网络设备芯片从网络设备中的其它模块(如射频模块或天线)接收信息，该信息是终端设备发送给网络设备的；或者，该网络设备芯片向网络设备中的其它模块(如射频模块或天线)发送信息，该信息是网络设备发送给终端设备的。
可以理解的是，本申请的实施例中的处理器可以是中央处理单元(Central Processing Unit，CPU)，还可以是其它通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field Programmable Gate Array，FPGA)或者其它可编程逻辑器件、晶体管逻辑器件。通用处理器可以是微处理器，也可以是任何常规的处理器。
本申请的实施例中存储器可以是随机存取存储器(Random Access Memory，RAM)、闪存、只读存储器(Read-Only Memory，ROM)、可编程只读存储器(Programmable ROM，PROM)、可擦除可编程只读存储器(Erasable PROM，EPROM)、电可擦除可编程只读存储器(Electrically EPROM，EEPROM)、寄存器、硬盘、移动硬盘或者本领域熟知的任何其它形式的存储介质中。一种示例性的存储介质耦合至处理器，从而使处理器能够从该存储介质读取信息，且可向该存储介质写入信息。当然，存储介质也可以是处理器的组成部分。处理器和存储介质可以位于ASIC中。另外，该ASIC可以位于网络设备或终端设备中。处理器和存储介质也可以作为分立组件存在于网络设备或终端设备中。
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
显然，本领域的技术人员可以对本申请进行各种改动和变型而不脱离本申请的范围。这样，倘若本申请的这些修改和变型属于本申请权利要求及其等同技术的范围之内，则本申请也意图包含这些改动和变型在内。

Claims (21)

  1. 一种信道信息反馈方法，其特征在于，应用于终端设备，包括：
    接收来自网络设备的第一参考信号；
    根据所述第一参考信号获得第一信道状态矩阵；
    根据第一稀疏图样对所述第一信道状态矩阵进行抽样，获得稀疏信道状态矩阵；所述第一稀疏图样由网络设备配置，所述第一稀疏图样用于对所述第一信道状态矩阵的至少一个维度进行抽样；
    向所述网络设备发送第一信道信息，所述第一信道信息用于指示所述稀疏信道状态矩阵。
  2. 根据权利要求1所述的方法，其特征在于，所述方法还包括：
    向所述网络设备发送第二参考信号，所述第二参考信号用于训练所述第一稀疏图样对应的神经网络，所述神经网络用于根据所述稀疏信道状态矩阵恢复所述第一信道状态矩阵。
  3. 根据权利要求2所述的方法，其特征在于，所述发送所述第二参考信号所使用的第二预编码与接收所述第一参考信号所使用的第一预编码相同。
  4. 根据权利要求2或3所述的方法，其特征在于，所述第一参考信号的带宽小于或者等于所述第二参考信号的带宽。
  5. 根据权利要求2或3所述的方法，其特征在于，所述第一参考信号对应的子载波间隔与所述第二参考信号对应的子载波间隔相同。
  6. 根据权利要求1至5任一所述的方法，其特征在于，所述第一信道状态矩阵的维度包含所述终端设备的天线维度、所述网络设备的天线维度和所述第一参考信号对应的频域维度和时域维度中的至少一个维度；
    所述第一稀疏图样指示以下至少一项：所述终端设备的天线维度中的至少一个天线的索引、所述网络设备的天线维度中的至少一个天线的索引、所述频域维度中的至少一个频域单元的索引和所述时域维度中的至少一个时域单元的索引。
  7. 一种信道信息恢复方法，其特征在于，应用于网络设备，包括：
    向终端设备发送第一参考信号；
    接收来自所述终端设备的第一信道信息，所述第一信道信息用于指示稀疏信道状态矩阵；所述稀疏信道状态矩阵为通过第一稀疏图样对第一信道状态矩阵的至少一个维度进行抽样获得的，所述第一信道状态矩阵根据所述第一参考信号确定；
    通过神经网络对所述稀疏信道状态矩阵进行处理，获得第二信道状态矩阵，所述第二信道状态矩阵为对所述第一信道状态矩阵的恢复值；所述神经网络是通过根据所述第一稀疏图样抽样的数据训练的。
  8. 根据权利要求7所述的方法，其特征在于，所述方法还包括：
    接收来自终端设备的第二参考信号；
    根据所述第二参考信号获得第三信道状态矩阵；
    根据第三信道状态矩阵获取第四信道状态矩阵，所述第四信道状态矩阵在子载波维度的长度等于所述第一信道状态矩阵在子载波维度的长度，所述第四信道状态矩阵在接收天线维度的长度等于所述第一信道状态矩阵在发送天线维度的长度，所述第四信道状态矩阵在发送天线维度的长度等于所述第一信道状态矩阵在接收天线维度的长度；
    根据第一稀疏图样对第四信道状态矩阵的至少一个维度进行抽样获得第五信道状态矩阵；所述神经网络是通过多个第四信道状态矩阵以及对应的第五信道状态矩阵训练得到的。
  9. 根据权利要求7所述的方法，其特征在于，所述方法还包括：
    接收来自所述终端设备的第二参考信号，所述第二参考信号用于训练所述第一稀疏图样对应的神经网络，所述神经网络用于根据所述稀疏信道状态矩阵恢复所述第一信道状态矩阵。
  10. 根据权利要求9所述的方法，其特征在于，所述接收所述第二参考信号所使用的第四预编码与发送所述第一参考信号所使用的第三预编码相同。
  11. 根据权利要求9或10所述的方法，其特征在于，所述第一参考信号的带宽小于或者等于所述第二参考信号的带宽。
  12. 根据权利要求9或10所述的方法，其特征在于，所述第一参考信号对应的子载波间隔与所述第二参考信号对应的子载波间隔相同。
  13. 根据权利要求7至12任一所述的方法，其特征在于，所述第一信道状态矩阵的维度包含所述终端设备的天线维度、所述网络设备的天线维度和所述第一参考信号对应的频域维度和时域维度中的至少一个维度；
    所述第一稀疏图样指示以下至少一项：所述终端设备的天线维度中的至少一个天线的索引、所述网络设备的天线维度中的至少一个天线的索引、所述频域维度中的至少一个频域单元的索引和所述时域维度中的至少一个时域单元的索引。
  14. 一种信道信息反馈方法，其特征在于，应用于终端设备，包括：
    接收来自网络设备的第一参考信号；
    根据所述第一参考信号进行信道估计，获得第一信道状态矩阵；
    根据所述第一信道状态矩阵确定第一信道特征矩阵；
    根据第二稀疏图样对所述第一信道特征矩阵进行抽样，获得第一稀疏矩阵；所述第二稀疏图样由所述网络设备配置，所述第二稀疏图样用于对所述第一信道特征矩阵的至少一个维度进行抽样；
    向所述网络设备发送第一信道信息，所述第一信道信息用于指示所述第一稀疏矩阵。
  15. 一种信道信息恢复方法，其特征在于，应用于网络设备，包括：
    向终端设备发送第一参考信号；
    接收来自所述终端设备的第一信道信息，所述第一信道信息用于指示所述第一稀疏矩阵；所述第一稀疏矩阵为通过第二稀疏图样对第一信道特征向量矩阵的至少一个维度进行抽样获得的，所述第一信道特征矩阵为通过第一信道状态矩阵确定的，所述第一信道状态矩阵根据所述第一参考信号确定；
    通过神经网络对所述第一稀疏矩阵进行处理，获得第二矩阵，所述第二矩阵为对所述第一信道特征矩阵的恢复值；所述神经网络是通过根据所述第二稀疏图样抽样的数据训练的。
  16. 一种通信装置，其特征在于，用于实现权利要求1至6、14中任一项所述的方法。
  17. 一种通信装置，其特征在于，用于实现权利要求7至13、15中任一项所述的方法。
  18. 一种通信装置，其特征在于，包括处理器，所述处理器和存储器耦合；
    所述处理器，用于执行所述存储器中存储的计算机程序或指令，使得所述通信装置实 现权利要求1至6、14中任意一项所述的方法。
  19. 一种通信装置，其特征在于，包括处理器，所述处理器和存储器耦合；
    所述处理器，用于执行所述存储器中存储的计算机程序或指令，使得所述通信装置实现权利要求7至13、15中任意一项所述的方法。
  20. 一种计算机可读存储介质，其特征在于，存储有计算机程序或指令，当所述计算机程序或指令在计算机上运行时，使得所述计算机实现如权利要求1至6、14中任意一项所述的方法，或者使得所述计算机实现如权利要求7至13、15中任意一项所述的方法。
  21. 一种计算机程序产品，其特征在于，包括指令，当所述指令在计算机上运行时，使得所述计算机实现如权利要求1至6、14中任意一项所述的方法，或者使得所述计算机实现如权利要求7至13、15中任意一项所述的方法。
PCT/CN2022/111628 2021-08-11 2022-08-11 一种信道信息反馈、恢复方法及装置 WO2023016508A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP22855488.7A EP4376479A1 (en) 2021-08-11 2022-08-11 Channel information feedback and recovery method and device
US18/437,789 US20240187283A1 (en) 2021-08-11 2024-02-09 Channel information feedback method, channel information recovery method, and apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110918725.9 2021-08-11
CN202110918725.9A CN115915266A (zh) 2021-08-11 2021-08-11 一种信道信息反馈、恢复方法及装置

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/437,789 Continuation US20240187283A1 (en) 2021-08-11 2024-02-09 Channel information feedback method, channel information recovery method, and apparatus

Publications (1)

Publication Number Publication Date
WO2023016508A1 true WO2023016508A1 (zh) 2023-02-16

Family

ID=85200558

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/111628 WO2023016508A1 (zh) 2021-08-11 2022-08-11 一种信道信息反馈、恢复方法及装置

Country Status (4)

Country Link
US (1) US20240187283A1 (zh)
EP (1) EP4376479A1 (zh)
CN (1) CN115915266A (zh)
WO (1) WO2023016508A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116055264A (zh) * 2023-04-03 2023-05-02 西南交通大学 稀疏信道的信号估计方法、装置、设备及可读存储介质
CN117097382A (zh) * 2023-08-30 2023-11-21 华南理工大学 一种基于人工智能的csi反馈增强方法、装置及介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109474316A (zh) * 2018-11-22 2019-03-15 东南大学 一种基于深度循环神经网络的信道信息压缩反馈方法
WO2020047774A1 (zh) * 2018-09-05 2020-03-12 华为技术有限公司 信道信息处理方法和装置
CN111464220A (zh) * 2020-03-10 2020-07-28 西安交通大学 一种基于深度学习的信道状态信息重建方法
WO2020180221A1 (en) * 2019-03-06 2020-09-10 Telefonaktiebolaget Lm Ericsson (Publ) Compression and decompression of downlink channel estimates

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020047774A1 (zh) * 2018-09-05 2020-03-12 华为技术有限公司 信道信息处理方法和装置
CN109474316A (zh) * 2018-11-22 2019-03-15 东南大学 一种基于深度循环神经网络的信道信息压缩反馈方法
WO2020180221A1 (en) * 2019-03-06 2020-09-10 Telefonaktiebolaget Lm Ericsson (Publ) Compression and decompression of downlink channel estimates
CN111464220A (zh) * 2020-03-10 2020-07-28 西安交通大学 一种基于深度学习的信道状态信息重建方法

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116055264A (zh) * 2023-04-03 2023-05-02 西南交通大学 稀疏信道的信号估计方法、装置、设备及可读存储介质
CN117097382A (zh) * 2023-08-30 2023-11-21 华南理工大学 一种基于人工智能的csi反馈增强方法、装置及介质
CN117097382B (zh) * 2023-08-30 2024-10-11 华南理工大学 一种基于人工智能的csi反馈增强方法、装置及介质

Also Published As

Publication number Publication date
CN115915266A (zh) 2023-04-04
EP4376479A1 (en) 2024-05-29
US20240187283A1 (en) 2024-06-06

Similar Documents

Publication Publication Date Title
WO2023016508A1 (zh) 一种信道信息反馈、恢复方法及装置
CN114499608B (zh) 用信号发送端口信息
US20240137082A1 (en) Communication method and apparatus
US20240348478A1 (en) Communication method and apparatus
US20240171429A1 (en) Communication method and apparatus
WO2024164641A1 (zh) 一种csi发送和接收方法及装置
WO2021155609A1 (zh) 信号发送方法及装置
WO2023126007A1 (zh) 信道信息传输方法及装置
CN117356050A (zh) 神经网络辅助的通信技术
EP4447524A1 (en) Communication method and apparatus
WO2024094177A1 (zh) 信道状态信息确定方法、装置及可读存储介质
WO2023066346A1 (zh) 一种通信方法及装置
WO2024192970A1 (en) A scalar quantization based channel state information feedback mechanism
EP4447334A1 (en) Uplink precoding method and apparatus
US20240356595A1 (en) Uplink precoding method and apparatus
WO2024153004A1 (zh) 一种通信方法及设备
WO2024027424A1 (zh) 一种数据质量度量方法及装置
US20240357560A1 (en) Data processing method for model, and apparatus
EP4443835A1 (en) Data processing method and apparatus for model
US20240235897A1 (en) Communication method based on quantum channel estimation and base station
EP4422094A1 (en) Calibration method and apparatus
CN118509014A (zh) 一种通信的方法和通信装置
CN118784033A (zh) 一种通信的方法和通信装置
WO2024150208A1 (en) Improving accuracy of artificial intelligence/machine learning (ai/ml) based channel state information (csi) feedback
CN115549739A (zh) 一种获得预编码矩阵的方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22855488

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2022855488

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2022855488

Country of ref document: EP

Effective date: 20240220

NENP Non-entry into the national phase

Ref country code: DE