WO2023116618A1 - 通信方法和通信装置 - Google Patents

通信方法和通信装置 Download PDF

Info

Publication number
WO2023116618A1
WO2023116618A1 PCT/CN2022/139975 CN2022139975W WO2023116618A1 WO 2023116618 A1 WO2023116618 A1 WO 2023116618A1 CN 2022139975 W CN2022139975 W CN 2022139975W WO 2023116618 A1 WO2023116618 A1 WO 2023116618A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
sending end
receiving
sending
matching layer
Prior art date
Application number
PCT/CN2022/139975
Other languages
English (en)
French (fr)
Inventor
胡斌
王坚
李榕
葛屹群
童文
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
Publication of WO2023116618A1 publication Critical patent/WO2023116618A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W16/00—Network planning, e.g. coverage or traffic planning tools; Network deployment, e.g. resource partitioning or cells structures
    • H04W16/02—Resource partitioning among network components, e.g. reuse partitioning
    • H04W16/10—Dynamic resource partitioning
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W28/00—Network traffic management; Network resource management
    • H04W28/02—Traffic management, e.g. flow control or congestion control
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W28/00—Network traffic management; Network resource management
    • H04W28/02—Traffic management, e.g. flow control or congestion control
    • H04W28/0231—Traffic management, e.g. flow control or congestion control based on communication conditions
    • H04W28/0236—Traffic management, e.g. flow control or congestion control based on communication conditions radio quality, e.g. interference, losses or delay
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W72/00—Local resource management
    • H04W72/12—Wireless traffic scheduling
    • H04W72/1263—Mapping of traffic onto schedule, e.g. scheduled allocation or multiplexing of flows

Definitions

  • the present application relates to the computer field, and in particular to a communication method and a communication device.
  • Modern communication system design is modular, and the process of signal processing is divided into a series of sub-modules, such as source coding, channel coding, modulation, channel estimation, etc.
  • Each sub-module is modeled based on a specific signal processing algorithm, usually approximated by some simplified linear model.
  • this method of optimizing each sub-module alone cannot guarantee the optimal end-to-end communication of the entire physical layer.
  • traditional end-to-end communication systems introduce more interference effects, such as amplifier distortion and channel impairments, and because of the increased number of factors and parameters to be controlled. Therefore, the complexity of end-to-end optimization using traditional methods is very high.
  • the channels in real communication scenarios are not static, especially in time-vary rayleigh fading channels, the autoencoder (AE) network trained only in specific If there will be a matching error under the predicted channel response (equivalent to the case of outliers in the training data set), retraining and adjustment must be performed, which will bring a large communication overhead at the sending and receiving ends. Therefore, a communication scheme with less communication overhead is studied.
  • AE autoencoder
  • the embodiment of the present application discloses a communication method and a communication device, which can communicate with less overhead.
  • the embodiment of the present application provides a communication method, the method includes: the sending end performs the first encoding process on the first data through the encoding network to obtain the first sending characteristic; the first sending characteristic and the sending end The channel distribution dimension of the environment is related; the sending end performs second encoding processing on the first sending feature through the matching layer to obtain the first feature; the encoding network and the matching layer are independently trained; the The dimension of the first feature is smaller than the dimension of the first sending feature; the sending end sends the first feature to the receiving end; the first feature is used by the receiving end to obtain the first data.
  • the sending end performs first encoding processing on the first data through the encoding network to obtain the first sending feature; and performs second encoding processing on the first sending feature through the matching layer to obtain the first feature. Since the encoding network and the matching layer are trained independently, when the channel changes, only updating the matching layer at the sending end can adapt to the new channel, and the adjustment to the current channel can be realized in a shorter time, and the receiving end is reduced. The overhead required for network training. In addition, since the receiving end does not need to participate in training, it can reduce the demand on the processing capability of the receiving end and prolong the use time of the receiving end.
  • the method further includes: the sending end updates the parameters of the matching layer according to its current channel; the sending end performs first encoding processing on the second data through the encoding network , to obtain the second sending feature; the sending end performs the second encoding process on the second sending feature through the updated matching layer to obtain the second feature; the sending end sends the first sending end to the receiving end Two features; the second feature is used by the receiving end to obtain the second data.
  • the sending end updates the parameters of the matching layer according to its current channel; by updating the matching layer at the sending end, it can adapt to the new channel without updating the encoding network and the decoding network at the receiving end, which can avoid The time overhead and signaling overhead caused by the encoding network and the decoding network at the receiving end.
  • the method before the sending end updates the parameters of the matching layer, the method further includes: the sending end receiving first indication information from the receiving end, the first indication information indicating The sending end updates the parameters of the matching layer.
  • the sending end receives the first indication information from the receiving end, so as to update the parameters of the matching layer in time according to the first indication information.
  • the parameters of the encoding network remain unchanged during the process of updating the parameters of the matching layer at the sending end.
  • the parameters of the encoding network remain unchanged during the process of updating the parameters of the matching layer at the sending end; the current channel can be used only by updating the parameters of the matching layer, which can reduce the amount of computation and improve the efficiency of updating the matching layer .
  • the encoding network is obtained through training under multiple different channels.
  • the encoding network can be regarded as a stack of multiple independent sub-encoding networks.
  • Each sub-encoding network is jointly trained by the sub-encoding network and the decoding network with fixed parameters under a specific channel. Any two sub-encoding networks are in the obtained by training under different specific channels.
  • the encoding network can handle multiple channel situations, that is, the encoding network can be applied to a variety of different channels.
  • the encoding network is applicable to many different channels, if the channel of the sending end changes, the sending end does not need to update the parameters of the encoding network, but only needs to update the parameters of the matching layer.
  • the multiple different channels may be obtained by clustering and dividing channels in the environment where the sending end is currently located.
  • updating the parameters of the matching layer at the sending end according to its current channel includes: updating the matching layer parameters at the sending end according to its current channel, the third sending feature, and the third receiving feature.
  • the parameters of the layer; the third receiving feature includes the feature obtained by the receiving end receiving the third feature sent by the sending end under the first channel, and the third feature includes the sending end using the matching layer pair
  • the third sending feature is a feature obtained by performing the second encoding process, and the current channel of the sending end is different from the first channel.
  • the sending end updates the parameters of the matching layer according to its current channel, the third sending feature, and the third receiving feature; it does not need to interact with the receiving end to obtain information other than the current channel, which can reduce communication overhead.
  • the method before the sending end updates the parameters of the matching layer according to its current channel, the method further includes: the sending end acquiring first information; the sending end according to the The first information is to determine the current channel of the sending end.
  • the sending end determines the current channel of the sending end according to the first information, so as to update the parameters of the matching layer.
  • the first information includes channel information or receiving characteristic offset information from the receiving end; the channel information represents information related to the current channel of the sending end, and the receiving characteristic offset
  • the information represents the difference between the third receiving feature and the fourth receiving feature
  • the third receiving feature includes the feature obtained by the receiving end receiving the third feature sent by the sending end under the first channel
  • the fourth receiving feature It includes the feature obtained by the receiving end receiving the third feature sent by the sending end in the current channel.
  • the first information includes channel information or reception characteristic offset information from the receiving end, so that the sending end uses the first information to obtain its current channel.
  • the first transmission feature includes an L-dimensional vector related to a channel distribution dimension, where L is a product of V and T, and the first data is at least represented by a V-dimensional vector, the T is a channel type obtained by clustering channels in the current environment, the T is an integer greater than or equal to 2, and the V is an integer greater than 0.
  • the first transmission feature includes an L-dimensional vector related to channel distribution dimensions, and may be applicable to different channels.
  • the method further includes: training the encoding network at the sending end when parameters of the matching layer remain unchanged.
  • the sending end trains the coding network; it can not only ensure the independence between the coding network and the matching layer, but also improve the speed of training the coding network.
  • the method further includes: the sending end receiving second instruction information from the receiving end, where the second instruction information instructs the sending end to retrain the encoding network.
  • the sending end may retrain the encoding network according to the second indication information.
  • the sending end receives the second indication information from the receiving end, so as to retrain the encoding network according to the second indication information.
  • the method further includes: when the matching layer is not trained and converged, the sending end sends third indication information to the receiving end, the third indication information indicating that the The receiving end retrains the encoding network.
  • the sending end sends the third indication information to the receiving end when the matching layer is not trained to converge; the training of the matching layer can be stopped in time, so that the retrained encoding network can be used to train a converged matching layer.
  • the matching layer is differentiable.
  • the matching layer is differentiable and thus does not affect gradient backpropagation.
  • the embodiment of the present application provides another communication method, which includes: the receiving end receives the first receiving feature from the sending end; the first receiving feature includes the first receiving feature sent by the sending end and is transmitted through a channel The feature received by the receiving end, the first feature is obtained by the sending end by encoding the first sending feature through the matching layer, and the first sending feature is the encoding network of the sending end encoding the first data Obtained through encoding processing; the encoding network and the matching layer are independently trained; the receiving end performs decoding processing on the first receiving feature through a decoding network to obtain the first data; the decoding The network and the matching layer are trained independently.
  • the receiving end receives the first receiving feature from the sending end. Since the encoding network and the matching layer are trained independently, when the channel changes, only updating the matching layer at the sending end can adapt to the new channel, and the adjustment to the current channel can be realized in a shorter time, and the receiving end is reduced. The overhead required for network training. In addition, since the receiving end does not need to participate in training, it can reduce the demand on the processing capability of the receiving end and prolong the use time of the receiving end.
  • the method further includes: the receiving end sending first indication information to the sending end, where the first indication information instructs the sending end to update parameters of the matching layer.
  • the receiving end sends the first indication information to the sending end, so as to instruct the sending end to update the parameters of the matching layer in time, so as to ensure accurate data transmission under the new channel.
  • the receiving end sending the first indication information to the sending end includes: the receiving end sending The terminal sends the first indication information.
  • the receiving end sends the first indication information to the sending end when the parameter characterizing the degree of channel change is less than or equal to the first threshold, so that the sending end has a lower degree of channel change with the receiving end.
  • the parameters of the matching layer are updated to adapt to the new channel.
  • the method further includes: the receiving end sends second indication information to the sending end when the parameter characterizing the degree of channel change is greater than a first threshold; the second indication The information instructs the sending end to retrain the encoding network.
  • the receiving end sends the second indication information to the sending end when the parameter representing the degree of channel change is greater than the first threshold; it can solve the problem that simply updating the parameters of the matching layer cannot successfully complete data transmission under the new channel question.
  • the method further includes: the receiving end receives third indication information from the sending end, and the third The indication information instructs the receiving end to retrain the encoding network.
  • the receiving end receives the third indication information from the sending end in order to retrain the encoding network, which can solve the problem that simply updating the parameters of the matching layer cannot successfully complete data transmission under the new channel.
  • the method before the receiving end sends the first indication information to the sending end, the method further includes: the receiving end receives fourth indication information from the sending end, and the fourth The indication information instructs the encoding network to complete training.
  • the receiving end receives the fourth indication information from the sending end, and can know in time that the encoding network has completed training.
  • the method further includes: the receiving end sends first information to the sending end, and the first information It is used for the sending end to update the parameters of the matching layer.
  • the receiving end sends the first information to the sending end, so that the sending end uses the first information to update the parameters of the matching layer.
  • the embodiments of the present application provide a communication device, where the communication device has a function of implementing the behaviors in the method embodiments of the first aspect above.
  • the functions described above may be implemented by hardware, or may be implemented by executing corresponding software on the hardware.
  • the hardware or software includes one or more modules or units corresponding to the above functions.
  • the processing module includes a processing module and a transceiver module, wherein: the processing module is configured to perform a first encoding process on the first data through an encoding network to obtain a first sending feature; the first sending feature It is related to the channel distribution dimension of the environment where the sending end is located; the processing module is also used to perform a second encoding process on the first sending feature through the matching layer to obtain the first feature; the encoding network and the matching layer It is obtained through independent training; the dimension of the first feature is smaller than the dimension of the first sending feature; the transceiver module is used to send the first feature to the receiving end; the first feature is used for the receiving The terminal obtains the first data.
  • the processing module is specifically configured to update the parameters of the matching layer according to its current channel; the processing module is also configured to perform first Encoding processing to obtain the second sending feature; performing second encoding processing on the second sending feature through the updated matching layer to obtain the second feature; the transceiver module is also used to send the received feature to the receiving end The second feature; the second feature is used for the receiving end to obtain the second data.
  • the transceiving module is further configured to receive first indication information from the receiving end, where the first indication information instructs the sending end to update parameters of the matching layer.
  • the parameters of the encoding network remain unchanged during the process of updating the parameters of the matching layer at the sending end.
  • the encoding network is obtained through training under multiple different channels.
  • the processing module is specifically configured to update the parameters of the matching layer according to its current channel, a third sending feature, and a third receiving feature;
  • the third receiving feature includes the receiving The end receives the feature obtained by the third feature sent by the sending end under the first channel, and the third feature includes the feature obtained by the sending end using the matching layer to perform a second encoding process on the third sending feature , the current channel of the sending end is different from the first channel.
  • the processing module is further configured to acquire first information; and determine a current channel of the sending end according to the first information.
  • the first information includes channel information or receiving characteristic offset information from the receiving end; the channel information represents information related to the current channel of the sending end, and the receiving characteristic offset
  • the information represents the difference between the third receiving feature and the fourth receiving feature
  • the third receiving feature includes the feature obtained by the receiving end receiving the third feature sent by the sending end under the first channel
  • the fourth receiving feature It includes the feature obtained by the receiving end receiving the third feature sent by the sending end in the current channel.
  • the first transmission feature includes an L-dimensional vector related to a channel distribution dimension, where L is a product of V and T, and the first data is at least represented by a V-dimensional vector, the T is a channel type obtained by clustering channels in the current environment, the T is an integer greater than or equal to 2, and the V is an integer greater than 0.
  • the processing module is further configured to train the encoding network when parameters of the matching layer remain unchanged.
  • the transceiving module is further configured to receive second instruction information from the receiving end, where the second instruction information instructs the sending end to retrain the encoding network.
  • the transceiver module is further configured to send third indication information to the receiving end when the matching layer has not been trained and converged, the third indication information indicating that the receiving end end retrains the encoding network.
  • the matching layer is differentiable.
  • an embodiment of the present application provides a communication device, which has a function of implementing the behavior in the method embodiment of the second aspect above.
  • the functions described above may be implemented by hardware, or may be implemented by executing corresponding software on the hardware.
  • the hardware or software includes one or more modules or units corresponding to the above functions.
  • the transceiver module includes a transceiver module and a processing module, wherein: the transceiver module is configured to receive a first receiving characteristic from a sending end; the first receiving characteristic includes a first characteristic sent by the sending end The feature received by the receiving end (that is, the communication device in the fourth aspect) through channel transmission, the first feature is obtained by the sending end by encoding the first sending feature through the matching layer, and the first sending feature It is obtained by performing encoding processing on the first data for the encoding network at the sending end; the encoding network and the matching layer are independently trained; the processing module is used to perform encoding on the first receiving feature through the decoding network Decoding processing to obtain the first data; the decoding network and the matching layer are obtained through independent training.
  • the transceiving module is further configured to send first indication information to the sending end, where the first indication information instructs the sending end to update parameters of the matching layer.
  • the transceiver module is specifically configured to send the first indication information to the sending end when the parameter characterizing the degree of channel change is less than or equal to a first threshold.
  • the transceiver module is further configured to send second indication information to the sending end when the parameter characterizing the degree of channel change is greater than the first threshold; the second indication information indicates The sending end retrains the encoding network.
  • the transceiving module is further configured to receive third instruction information from the sending end, where the third instruction information instructs the receiving end to retrain the encoding network.
  • the transceiving module is further configured to receive fourth indication information from the sending end, where the fourth indication information indicates that the encoding network has completed training.
  • the transceiving module is further configured to send first information to the sending end, where the first information is used by the sending end to update parameters of the matching layer.
  • the present application provides a communication device, which includes a processor, and the processor can be used to execute computer-executed instructions stored in the memory, so that the above-mentioned first aspect or any possible implementation of the first aspect The shown method is executed, or the method shown in the above second aspect or any possible implementation of the second aspect is executed.
  • the process of sending information in the above method can be understood as the process of outputting information based on the instructions of the processor.
  • the processor In outputting information, the processor outputs the information to the transceiver for transmission by the transceiver. After the information is output by the processor, it may also need to undergo other processing before reaching the transceiver.
  • the processor receives incoming information
  • the transceiver receives that information and inputs it to the processor. Furthermore, after the transceiver receives the information, the information may require other processing before being input to the processor.
  • the above-mentioned processor may be a processor dedicated to performing these methods, or may be a processor that executes computer instructions in a memory to perform these methods, such as a general-purpose processor.
  • the processor may also be used to execute a program stored in the memory, and when the program is executed, the communication device executes the method as shown in the first aspect or any possible implementation manner of the first aspect.
  • the memory is located outside the communication device. In a possible implementation manner, the memory is located in the above communication device.
  • the processor and the memory may also be integrated into one device, that is, the processor and the memory may also be integrated together.
  • the communication device further includes a transceiver, where the transceiver is configured to receive a message or send a message, and the like.
  • the present application provides another communication device, which includes a processing circuit and an interface circuit, where the interface circuit is used to acquire data or output data; the processing circuit is used to perform any of the above-mentioned first aspect or the first aspect.
  • the corresponding method shown in the possible implementation manner, or the processing circuit is configured to execute the corresponding method shown in the second aspect or any possible implementation manner of the second aspect.
  • the present application provides a computer-readable storage medium, which is used to store a computer program, and when it is run on a computer, the above-mentioned first aspect or any possible implementation of the first aspect The shown method is executed, or the method shown in the second aspect or any possible implementation manner of the second aspect is executed.
  • the present application provides a computer program product, the computer program product includes a computer program or computer code, and when it is run on a computer, the above first aspect or any possible implementation of the first aspect shows The method is executed, or the method shown in the second aspect or any possible implementation manner of the second aspect is executed.
  • the present application provides a communication system, including the communication device described in the third aspect or any possible implementation of the third aspect, and the communication device described in the fourth aspect or any possible implementation of the fourth aspect .
  • FIG. 1 is an example of a satellite communication system provided by an embodiment of the present application
  • FIG. 2 is an example of an inter-satellite communication system provided by an embodiment of the present application
  • FIG. 3 is an example of a wireless communication system provided by an embodiment of the present application.
  • FIG. 4 is an example of a relationship diagram representing the relationship between the sending feature, matching layer, transmission feature, and receiving feature provided by the embodiment of the present application;
  • FIG. 5 is a schematic diagram of a sending end independently updating parameters of a matching layer provided by an embodiment of the present application
  • FIG. 6A is a schematic framework diagram of an encoding network and a matching layer at a sending end provided by an embodiment of the present application;
  • FIG. 6B is a schematic diagram of the framework of the coding network and the matching layer at the sending end of the embodiment of the present application;
  • FIG. 7A is a schematic diagram of a process of training an autoencoder provided in an embodiment of the present application.
  • FIG. 7B is a schematic diagram of a process of training a matching layer provided by an embodiment of the present application.
  • FIG. 8 is a flowchart of a communication method provided by an embodiment of the present application.
  • FIG. 9 is a flowchart of another communication method provided by the embodiment of the present application.
  • FIG. 10 is a flowchart of another communication method provided by the embodiment of the present application.
  • FIG. 11 is a flowchart of another communication method provided by the embodiment of the present application.
  • FIG. 12 is a flow chart of another communication method provided by the embodiment of the present application.
  • FIG. 13 is a flow chart of another communication method provided by the embodiment of the present application.
  • FIG. 14 is a flowchart of another communication method provided by the embodiment of the present application.
  • FIG. 15 is a flowchart of another communication method provided by the embodiment of the present application.
  • FIG. 16 is a flow chart of another communication method provided by the embodiment of the present application.
  • FIG. 17 is a flow chart of another communication method provided by the embodiment of the present application.
  • FIG. 18 is a schematic structural diagram of a communication device provided by an embodiment of the present application.
  • FIG. 19 is a schematic structural diagram of another communication device provided by an embodiment of the present application.
  • FIG. 20 is a schematic structural diagram of another communication device 200 provided in an embodiment of the present application.
  • FIG. 21 is a schematic structural diagram of another communication device 210 provided by an embodiment of the present application.
  • an embodiment means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the present application.
  • the occurrences of this phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is understood explicitly and implicitly by those skilled in the art that the embodiments described herein can be combined with other embodiments.
  • the present application provides a communication scheme with less communication overhead.
  • the main principle of the communication scheme provided by this application is that the sending end adjusts the channel changes by updating the parameters of its matching layer through self-training, and the parameters of the decoding network at the receiving end are fixed. That is to say, when the channel between the sending end and the receiving end changes, the sending end adjusts the channel change by updating the parameters of its matching layer through self-training. There is no need to retrain the network at the sending end and the network at the receiving end. Reduce the communication overhead caused by the joint training of the sending end and the receiving end, and improve the training efficiency.
  • the parameters of the decoding network at the receiving end are fixed when the channel between the sending end and the receiving end changes, the communication scheme provided by this application can reduce the demand for the processing capability of the receiving end and prolong the use of the receiving end. duration.
  • Fig. 1 is an example of a satellite communication system provided by an embodiment of the present application.
  • the satellite communication system includes a satellite base station and a terminal type network element, such as the terminal equipment in FIG. 1 .
  • Satellite base stations provide communication services for terminal devices, which may include smart phones, smart watches, tablet computers and other devices.
  • the satellite base station transmits downlink data to the terminal equipment, in which the data is encoded by channel coding, and the channel-coded data is transmitted to the terminal equipment after constellation modulation; the terminal equipment transmits uplink data to the satellite base station, and the uplink data can also be encoded by channel coding.
  • the coded data is transmitted to the satellite base station after constellation modulation.
  • the communication scheme provided by this application can be applied to an inter-satellite communication system.
  • the inter-satellite communication system can be divided into two parts: the space beam acquisition tracking and alignment (acquisition pointing and tracking, APT) subsystem and the communication subsystem.
  • the communication subsystem is responsible for the transmission of inter-satellite information and is the main body of the inter-satellite communication system;
  • the APT system is responsible for the capture, alignment and tracking between satellites, determining the incoming wave direction of the incident signal as capture, and adjusting the direction of the transmitted wave to aim at the receiving direction as the target. Alignment, alignment and capture are continuously adjusted for tracking throughout the entire communication process.
  • FIG. 2 is an example of an inter-satellite communication system provided by an embodiment of the present application.
  • the inter-satellite communication system shown in Figure 2 includes satellite 1 and satellite 2, and both satellite 1 and satellite 2 include: a communication module, a transceiver antenna, an APT module and an APT transmit/receive antenna; wherein, the communication module and the transceiver antenna belong to the communication subsystem , the APT module and the APT transmit/receive belong to the APT subsystem.
  • the APT In order to minimize the attenuation and interference effects in the channel, and at the same time require high confidentiality and transmission rate, the APT must be adjusted in real time to continuously adapt to changes.
  • the APT system may be an optical system.
  • the APT system and the communication subsystem may be separate systems.
  • Fig. 3 is an example of a wireless communication system provided by an embodiment of the present application.
  • the communication system includes: one or more user equipments, and only two user equipments are taken as an example in FIG. 1, and one or more access network devices ( For example, a base station), only one access network device is taken as an example in FIG. 1 .
  • access network devices For example, a base station
  • a user equipment is a device with a wireless transceiver function.
  • the user equipment can communicate with one or more core network (core network, CN) equipment (or called core equipment) via an access network device (or called an access device) in a radio access network (radioaccess network, RAN). communication.
  • Core network CN
  • core equipment or called core equipment
  • RAN radio access network
  • User equipment can be deployed on land, including indoor or outdoor, handheld or vehicle-mounted; it can also be deployed on water (such as ships, etc.); it can also be deployed in the air (such as on aircraft, balloons, and satellites, etc.).
  • the UE may also be referred to as a terminal device, which may be a mobile phone (mobile phone), a mobile station (mobile station, MS), a tablet computer (pad), a computer with a wireless transceiver function, or a virtual reality (virtual reality, VR) terminal equipment, augmented reality (augmented reality, AR) terminal equipment, wireless terminal equipment in industrial control (industrial control), wireless terminal equipment in self driving (self driving), wireless in remote medical (remote medical) Terminal equipment, wireless terminal equipment in smart grid, wireless terminal equipment in transportation safety, wireless terminal equipment in smart city, wireless terminal equipment in smart home, etc.
  • a terminal device may be a mobile phone (mobile phone), a mobile station (mobile station, MS), a tablet computer (pad), a computer with a wireless transceiver function, or a virtual reality (virtual reality, VR) terminal equipment, augmented reality (augmented reality, AR) terminal equipment, wireless terminal equipment in industrial control (industrial control), wireless terminal equipment in self driving (self driving), wireless in remote medical (remote medical) Terminal equipment,
  • the user equipment may be a handheld device with a wireless communication function, a vehicle-mounted device, a wearable device, or a terminal in the Internet of Things, the Internet of Vehicles, a 5G network, or any form of terminal in the future network. Not limited.
  • the access network device may be any device that has a wireless transceiver function and can communicate with the user equipment, for example, a radio access network (radio access network, RAN) node that connects the user equipment to a wireless network.
  • radio access network radio access network
  • RAN nodes include: gNB, transmission reception point (TRP), evolved Node B (evolved Node B, eNB), radio network controller (radio network controller, RNC), Node B (Node B, NB), base station controller (base station controller, BSC), base transceiver station (base transceiver station, BTS), home base station (for example, home evolved NodeB, or home Node B, HNB), base band unit (base band unit , BBU), wireless fidelity (Wifi) access point (access point, AP), integrated access and backhaul (IAB), etc.
  • TRP transmission reception point
  • eNB evolved Node B
  • RNC radio network controller
  • Node B Node B
  • base station controller base station controller
  • a base station is used as an example of an access network device for description.
  • the base station may include a baseband unit (baseband unit, BBU) and a remote radio unit (remote radio unit, RRU).
  • BBU baseband unit
  • RRU remote radio unit
  • the BBU and RRU can be placed in different places, for example: the RRU is remote and placed in areas with high traffic volume, and the BBU is placed in the central equipment room.
  • the BBU and RRU can also be placed in the same equipment room.
  • the BBU and the RRU may also be different components under one rack.
  • the wireless communication systems mentioned in the embodiments of the present application include but are not limited to: narrow band-internet of things (NB-IoT), global system for mobile communications (GSM) , enhanced data rate for GSM evolution (EDGE), wideband code division multiple access (WCDMA), code division multiple access 2000 (code division multiple access, CDMA2000), Time division-synchronization code division multiple access (TD-SCDMA), long term evolution (LTE), universal mobile telecommunications system (UMTS), global interconnected microwave access Access (worldwide interoperability for microwave access, WiMAX) communication system, fourth generation (4th generation, 4G) communication system, non-terrestrial network (non-terrestrial network, NTN) system, fifth generation (5th generation, 5G) communication system Or new radio (new radio, NR) and other communication systems in the future, such as 6G, etc.
  • NB-IoT narrow band-internet of things
  • GSM global system for mobile communications
  • EDGE enhanced data rate for GSM evolution
  • WCDMA wideband code division multiple
  • the communication solution provided in this application involves the network structure of the sending end and the training interaction process of the sending and receiving ends (the sending end and the receiving end).
  • the sending end includes a coding network and a matching layer (matched layer), and the receiving end includes a decoding network.
  • the matching layer may be referred to as a channel matching layer or an adaptive channel coding module.
  • the encoding network and matching layer at the sending end are independently trained. By utilizing the independent characteristics of the encoding network and matching layer, only the matching layer can be updated (the parameters of the encoding network remain unchanged) to solve the problem of time-varying weak channel. The effect of encoder training.
  • the network structure features involved in this application include:
  • the transmission feature design can be understood as the traditional autoencoder design is only optimized for a specific channel, and in order to adapt to channel changes, the transmitter needs to have pre-training results for various channel distribution scenarios in the environment, so the corresponding
  • the features of the traditional autoencoder can also be regarded as adding a dimension related to the channel distribution on the basis of it, and becoming a higher-order tensor.
  • the output feature of a traditional autoencoder is a vector with a dimension of V, and the channels in the current environment of the sender are clustered and divided to obtain 10 types of channels.
  • the encoding network output dimension of the sender provided by this application has a dimension of V ⁇ 10 Vector, each channel corresponds to a vector of dimension V.
  • the output characteristics of the traditional autoencoder are the characteristics of the transmitted data under one channel, and the characteristics of the output of the encoding network at the transmitting end provided by the present application include the characteristics of the transmitted data under multiple different channels.
  • the modulation order is determined, the total number of codewords can also be determined. For different codewords, it can be considered that there are n such different transmission characteristics f (n) in total.
  • the receiving characteristic in the embodiment of the present application is described as a fixed receiving characteristic r s which is independent of the channel distribution.
  • the receiving feature can be obtained through the joint training of the encoding network and the decoding network under some specific channel h s .
  • R s can be described as all possible data input to the decoding network at the receiving end under the channel h s from a formal understanding.
  • the sending end can set it as the training target of the matching layer, that is, it is hoped that by adjusting the parameters of the matching layer, the same sending data will have the same reception characteristics obtained at the receiving end after passing through other different channels.
  • the parameters of the matching layer can be expressed as a matrix or tensor description Likewise, it can be determined as n by the total number of codewords.
  • the transmission feature can be understood as the output value of the sending end after passing through all networks (including the encoding network and matching layer), that is, the output of the entire network of the sending end after forward deduction.
  • the memory buffer (memory buff) in the sending end is used to record the values of sending feature f (n) and receiving feature r (n) in the current period of time, as a data set for independent training of the sending end.
  • the memory buffer can be designed as a first-in-first-out queue, which can keep track of the data sent to update the training data at any time, and can obtain coding gain from the distribution of the sent data source.
  • the training set in the memory buffer can well express the high-dimensional continuous space where the sent data is located, and the sending end can use the training set to determine the early stop judgment of training convergence.
  • the input of the encoding network is the transmitted data, and the output is the channel distribution-related dimension vector; the input of the matching layer is the channel distribution-related dimension vector, and the output is the constellation point coordinates corresponding to the modulation order.
  • the output of the encoding network in the original AE is the constellation point coordinates.
  • the encoding network here is equivalent to obtaining the constellation point coordinates under different channels, and corresponding to the coordinates suitable for successful transmission under the current channel through the matching layer.
  • the encoding network is a feature extraction of input data, and can store feature extraction strategies for different channels.
  • the matching layer can be regarded as the attention between the sending feature and the transmission feature (which can be obtained from the receiving feature), and pays attention to some features related to the channel from the input high-dimensional features.
  • h s represents the static channel part, which is determined by the relative distance, location, and surrounding static environment of the two ends of the transceiver. It can be considered that this value is fixed within a certain period of time;
  • ⁇ h d represents the dynamic channel part, which is related to the movement, Blocking and other random events are related, and this part can be estimated by some prior information.
  • the correlation of channel data is used to predict the channel of the autoregressive model, and the transmitting end monitors the environmental change estimation through the sensing device, or obtains the existing channel state information (CSI) feedback estimation, etc.
  • CSI channel state information
  • the dynamic channel information can also be implicitly fed back in the receiving feature offset, which is expressed as the receiving feature r' under the new channel and the fixed receiving feature r s obtained from the self-encoder training
  • ⁇ h m represents the distorted channel part, which is caused by the inevitable measurement uncertainty, but it can be considered that Relative to h s and ⁇ h d are small, it has little impact on the performance after adjustment of the channel matching layer.
  • the problem of autoencoder for time-varying fading channel distortion can be transformed into the matching problem of features extracted by the transceiver network under the influence of ⁇ h d through the matching layer.
  • the optimization of the encoding network (Encoder NN) at the sending end and the decoding network (Decoder NN) at the receiving end under the h s channel is similar to the traditional autoencoder training process, and the specific steps are as follows:
  • the transmission feature is the result of pre-training for scenarios under various channel distributions in the environment, and the channel information of the relevant distribution is stored through the encoding network.
  • the transmission feature can be regarded as adding a dimension T related to the channel distribution on the basis of the output features of the traditional autoencoder, and turning it into a higher-order tensor.
  • the transmission characteristics can be determined by the total number of codeword types as n, and the initial value of the matching layer Can be set to random initialization.
  • the subsequent symbol superscript (n) is omitted to indicate the situation under a specific codeword.
  • the receiving end calculates the mean square error or binary cross entropy
  • N is the batch size of the training
  • p ic is the predicted probability that the decoded data i belongs to category c
  • u ic is the category sign function (0 or 1).
  • the parameters of the transceiver network are trained through gradient feedback. Since the matching layer is a differentiable function, the gradient feedback is not affected.
  • the parameters of the decoding network at the receiving end can be fixed to ensure that the receiving feature rs does not change with the channel, and the pre-training effect for different channels can be achieved by training the encoding network at the sending end.
  • matching layer The function of the matching layer will be described below in combination with the drawings representing the relationship among the sending feature, matching layer, transmission feature, and receiving feature.
  • FIG. 4 is an example of a relational diagram representing a relationship among a sending feature, a matching layer, a transmission feature, and a receiving feature provided in an embodiment of the present application.
  • the mapping relationship between the sending features of the encoding network at the sending end and the receiving features at the receiving end can be obtained as follows:
  • the transmission feature f is mapped to the new transmission feature t' by adjusting the parameters of the matching layer, so that there is In this way, for the new channel environment, the receiving end can complete the correct decoding operation for the fixed receiving characteristics.
  • A' represents the adjusted matching layer parameters. Indicates a compound operation.
  • the operation through the channel or network can be understood as a function operation, so A' is applied to the sending feature f first, and then h' is applied. Can be expressed as h'(A'(f)).
  • the function of the matching layer is to map the transmission feature f to the new transmission feature t' by adjusting the parameters of the matching layer under the new channel environment h', so that there is In other words, by adjusting the parameters of the matching layer, so that when the sending end sends the same data under different channels, the receiving end can receive the same receiving characteristics.
  • the channel information is not limited to using commonly used channel parameters, such as signal strength and CSI, and the offset of the receiving feature can be used as the feedback value.
  • the above-mentioned objective function can also be defined as the receiving feature offset, see above for the specific expression. It can be seen that the channel information at this time is also implicit in the receiving characteristic offset. But relatively speaking, the training of the parameters of the matching layer becomes a way of reinforcement learning, which will receive feature offsets as rewards. Compared with feeding back complete channel information or eigenvectors, the receiving feature offset in this scheme is a scalar and can have fewer values.
  • FIG. 5 is a schematic diagram of a sending end independently updating parameters of a matching layer provided by an embodiment of the present application.
  • a training set D is recorded in the memory buffer, which includes sending features (such as f) and receiving features (such as r); the sending end obtains the target of the receiving end by sampling the receiving features in the training set Receiving feature r; the sending end uses the encoding network to perform the first encoding process on the sent data corresponding to the target receiving feature r to obtain the sending feature f (corresponding to the target receiving feature r), and uses the matching layer to perform the second encoding process on the sending feature f Get the transmission feature t', calculate the receiving feature r' after the transmission feature t' is transmitted through the target channel (current channel); calculate the error between the target receiving feature r and the receiving feature r', and use the calculated error (such as gradient information) to Update the parameters of the matching layer.
  • sending features such as f
  • receiving features such as r
  • the sending end obtains the target of the receiving end by sampling the receiving features in the training set Receiving feature r
  • the sending end uses the encoding network to perform the first encoding
  • the arrow pointing to the matching layer in Fig. 5 indicates the process of updating the matching layer.
  • the dotted box in FIG. 5 shows an example of obtaining the memory buffer training set by recording the sending feature f and the target receiving feature r.
  • the sending end encodes the sent data through the encoding network to obtain the sending feature f and the transmission feature t
  • the target receiving feature r is the reception feature that the transmission feature t is received by the receiving end through the original channel transmission. That is to say, the sending end collects the sending feature f under the original channel and the receiving end receives the target receiving feature r and records them in the memory buffer; the sending end sends the transmission feature t′ under the target channel, and the receiving end receives the receiving feature r '.
  • the sender adjusts the output transmission characteristics by updating the parameters of the matching layer, so that the receiver can receive the target reception characteristic r.
  • the embodiments of the present application provide some implementations in which the sender independently updates the parameters of the matching layer.
  • Method 1 The sender updates its matching layer based on the optimization problem of feature matching.
  • One possible method is as follows:
  • the sending end can record the sending feature f and receiving feature rs obtained under the channel h s at the sending end through the memory buffer.
  • the optimized offset target for D is the distance metric, for example, the square error Feature covariance KL divergence Stiffness and distance mixed error wait.
  • the sender can use limited-memory broyden-fletcher-goldfarb-shan-no (L-BFGS) based on limited memory or stochastic gradient descent to iteratively solve the parameters of the matching layer A' with A 0 as the initial value.
  • L-BFGS limited-memory broyden-fletcher-goldfarb-shan-no
  • FIG. 6A is a schematic diagram of a framework of an encoding network and a matching layer at a sending end provided by an embodiment of the present application.
  • u (1) represents the transmission data
  • f (1) represents the transmission characteristics
  • the rectangular boxes f 1 , f 2 , ..., f L represent the L-dimensional transmission characteristics of the encoding network under different channels
  • A represents the matching layer parameter
  • a 11 represents a parameter in the matching layer
  • t represents the transmission feature.
  • the matching layer at the sending end can be regarded as a readout operation.
  • the sending end can read some features that match the current channel information according to the weight A out Get the transfer characteristic t.
  • the transmission feature t * of the target can be obtained by the inverse operation of the receiving feature r s at the sending end, that is, the transmission feature that the matching layer at the sending end needs to output. So the problem can be changed to solve
  • the sender selects the parameters of the matching layer based on reinforcement learning of stochastic bandits.
  • One possible method is as follows:
  • the sending end can regard the problem of updating the parameters of the matching layer as a random bandits problem, that is, by using the offset objective function value fed back by the receiving end as a reward, to select the optimal matching layer parameters faster (that is, the parameters of the matching layer ).
  • the parameters of the matching layer at the sending end can be pre-set as a linear combination under a limited number of different channel distributions as an action.
  • the A 1 parameter indicates the selection of the transmission feature f extraction method under the h 1 channel
  • the A 2 parameter indicates the selection of the h 2 channel
  • the sending feature f extraction method Therefore, there are finitely many A-parameter designs in theory.
  • FIG. 6B is a schematic diagram of a framework of an encoding network and a matching layer at a sending end according to an embodiment of the present application.
  • f (1) represents the transmission feature
  • the rectangular boxes f 1 , f 2 , ..., f L represent the L-dimensional transmission characteristics of the encoding network under different channels
  • A represents the parameters of the matching layer
  • a 1 is one of A
  • Optional action t represents the transmission characteristic
  • rs represents the reception characteristic in the memory buffer
  • r' represents the reception characteristic of the receiving end calculated by the sending end when the parameter of the matching layer under the current channel is A 1 .
  • the sender can set the reward value to the degree of similarity between the receiving characteristics obtained by using the selected matching layer parameters and the target receiving characteristics, such as the above-mentioned optimized offset target, the higher the similarity, the smaller the offset error, the greater the reward value.
  • the sending end can guide the matching layer to choose the optimal A parameter action through feedback similar to reward information.
  • the sending end can use each matching layer parameter recorded (that is, a limited number of A parameters, such as A1 parameters) to perform a second encoding process on the sending feature to obtain each matching layer parameter Corresponding receiving characteristics; according to the similarity between the receiving characteristics corresponding to each matching layer parameter and the target receiving characteristics, determine the reward obtained by selecting each matching layer parameter; select the matching layer parameter with the largest reward as the parameter of the matching layer .
  • each matching layer parameter recorded that is, a limited number of A parameters, such as A1 parameters
  • the matching layer parameters of the sending end can choose any of A 1 , A 2 , ..., AT ; when the sending end and the receiving end When the channel between changes to h′, the sending end calculates the second encoding process of the sending feature by using each of the matching layer parameters A 1 , A 2 ,..., AT as the matching layer parameters under the channel h′ Receiving feature; if the matching layer parameter A 2 is used to perform the second encoding process on the sending feature under the channel h′, the similarity between the receiving feature and the target receiving feature is higher than that of using any other matching layer parameters to perform the second encoding process on the sending feature According to the similarity between the obtained receiving features and the target receiving features, the matching layer parameter A2 is selected under the channel h' to perform the second encoding process on the sending features.
  • the sender’s matching layer parameters can choose any of A 1 , A 2 , ..., AT , and the A 1 parameter represents the choice For the extraction method of the transmission feature f under the h1 channel, the A2 parameter represents the selection of the transmission characteristic f extraction method under the h2 channel, and the A T parameter represents the selection of the transmission characteristic f extraction method under the h T channel; when the sending end and When the channel between the receiving ends changes from h 1 to h 2 , the sending end selects the A 2 parameter as a parameter of the matching layer to perform the second encoding process on the sending feature.
  • the training of the sending end and the receiving end can be divided into two parts, one part is training the autoencoder, that is, training the encoding network of the sending end and the decoding network of the receiving end, and the other is training the matching layer (or updating the matching layer. parameter).
  • FIG. 7A is a schematic diagram of a process of training an autoencoder provided by an embodiment of the present application.
  • the feedforward process is as follows: the decoding network at the sending end performs the first encoding process on the sending data u, and outputs the sending feature f; the matching layer at the sending end (A represents the parameters of the matching layer in FIG.
  • the transmission feature t is transmitted through the channel after batch regularization processing (optional);
  • the receiving end receives the receiving feature r (that is, the transmission feature t (or the transmission feature t after batch regularization processing) is received by the receiving end through channel transmission);
  • the receiving end uses the decoding network to decode the receiving feature r to obtain the sent data u * ;
  • the backpropagation process includes: the receiving end calculates the objective function according to the sending data u and the sending data u * to obtain gradient information;
  • the receiving end updates the parameters of its decoding network according to the gradient information, and feeds back to the receiving end Gradient information;
  • the sending end updates the parameters of its decoding network according to the gradient information fed back by the receiving end.
  • a possible implementation of training the matching layer is as follows: fix the encoding network at the sending end and the decoding network at the receiving end (that is, the parameters of the encoding network at the sending end and the parameters of the decoding network at the receiving end are fixed), and in the forward iteration process , the receiving feature r i ′ can be calculated by using the current channel combined with the sending feature f obtained from autoencoder training.
  • the sending end obtains the sending feature f under the original channel and the receiving feature corresponding to the sending feature f from the memory buffer
  • the sending end iteratively updates the parameters of the matching layer of the sending end under the new channel, so that the updated matching layer parameters can adapt to the current channel, and the receiving end does not need to adjust the decoding network.
  • Adapting the updated matching layer parameters to the current channel refers to using the updated matching layer parameters to perform the second encoding process on the sending feature f to obtain the receiving feature r s through the new channel transmission.
  • the sending end may use any one of the first, second, and third methods provided in the embodiments of the present application to train the matching layer.
  • Ways for the sender to obtain channel information may include: the sender obtains channel information from prior information estimation, such as autoregressive model channel prediction based on channel data correlation; the sender monitors environmental change estimation through sensing devices; estimates through CSI feedback, etc.
  • FIG. 7B is a schematic diagram of a process of training a matching layer according to an embodiment of the present application.
  • the encoding network at the sending end and the decoding network at the receiving end are fixed, and the sending end trains its matching layer independently.
  • the sending feature f output by the encoding network and the receiving feature r input by the decoding network can be recorded in the memory buffer, and the sending end trains its matching layer according to the sending feature and receiving feature in the memory buffer.
  • FIG. 8 is a flowchart of a communication method provided by an embodiment of the present application. As shown in Figure 8, the method includes:
  • the sending end performs first encoding processing on the first data through an encoding network to obtain a first sending feature.
  • the first sending characteristic is related to the channel distribution dimension of the environment where the sending end is located.
  • the encoding network at the sending end and the decoding network at the receiving end can form an autoencoder.
  • the above-mentioned first transmission feature includes an L-dimensional vector related to the channel distribution dimension, the above-mentioned L is the product of V and T, the above-mentioned first data is represented by at least a V-dimensional vector, and the above-mentioned T is the The channel type obtained by clustering the channels of the environment, the above T is an integer greater than or equal to 2, and the above V is an integer greater than 0.
  • the sending end performs the first encoding process on any sending data (for example, the first data) through the encoding network, and the sending features obtained by the sending end have pre-training results for scenarios under various channel distributions in the environment.
  • the traditional autoencoder design is only optimized for a specific channel.
  • the transmission features output by the encoding network at the sending end have pre-training results for the scenarios under various channel distributions in the environment, so the transmission features correspond to the traditional autoencoder.
  • the encoding network at the sending end is trained under a variety of different channels.
  • the encoding network can be regarded as a stack of multiple independent sub-encoding networks.
  • Each sub-encoding network is jointly trained by the sub-encoding network and the decoding network with fixed parameters under a specific channel. Any two sub-encoding networks are in the obtained by training under different specific channels.
  • Each sub-encoding network can be viewed as an encoding network in a traditional autoencoder.
  • the encoding network since the encoding network is trained under a variety of different channels, the encoding network can handle a variety of channel conditions, that is, the encoding network can be applied to a variety of different channels.
  • the sending end does not need to update the parameters of the encoding network, but only needs to update the parameters of the matching layer.
  • the above-mentioned multiple different channels may be obtained by clustering and dividing the above-mentioned channels in the current environment where the sending end is located.
  • the sending end may be an access network device or a user equipment.
  • the access network device may be any device that has a wireless transceiver function and can communicate with the user equipment, for example, a radio access network (radio access network, RAN) node that connects the user equipment to a wireless network.
  • RAN radio access network
  • examples of some RAN nodes include: gNB, transmission reception point (TRP), evolved Node B (evolved Node B, eNB), radio network controller (radio network controller, RNC), Node B (Node B, NB), base station controller (base station controller, BSC), base transceiver station (base transceiver station, BTS), home base station (for example, home evolved NodeB, or home Node B, HNB), base band unit (base band unit , BBU), wireless fidelity (Wifi) access point (access point, AP), integrated access and backhaul (IAB), etc.
  • a base station is used as an example of an access network device for description.
  • a user equipment is a device with a wireless transceiver function.
  • the user equipment can communicate with one or more core network (core network, CN) equipment (or called core equipment) via an access network device (or called an access device) in a radio access network (radioaccess network, RAN). communication.
  • Core network CN
  • core equipment or called core equipment
  • RAN radio access network
  • User equipment can be deployed on land, including indoor or outdoor, handheld or vehicle-mounted; it can also be deployed on water (such as ships, etc.); it can also be deployed in the air (such as on aircraft, balloons, and satellites, etc.).
  • the UE can also be called a terminal device, which can be a mobile phone, a tablet computer (pad), a computer with a wireless transceiver function, a virtual reality (virtual reality, VR) terminal device, an augmented reality (augmented reality, AR) terminal equipment, wireless terminal equipment in industrial control, wireless terminal equipment in self driving, wireless terminal equipment in remote medical, smart grid Wireless terminal equipment in transportation safety, wireless terminal equipment in smart city, wireless terminal equipment in smart home, etc.
  • the user equipment may be a handheld device with a wireless communication function, a vehicle-mounted device, a wearable device, or a terminal in the Internet of Things, the Internet of Vehicles, a 5G network, or any form of terminal in the future network. Not limited.
  • the sending end performs second encoding processing on the first sending feature through the matching layer to obtain the first feature.
  • the above encoding network and the above matching layer are obtained through independent training.
  • the dimension of the above-mentioned first feature is smaller than the dimension of the above-mentioned first transmission feature.
  • the sending end performs the second encoding process on the first sending feature through the matching layer to satisfy the following formula:
  • the sending end sends the first feature to the receiving end.
  • the above-mentioned first feature is used for the above-mentioned receiving end to obtain the above-mentioned first data.
  • the sending end and the receiving end have completed the training of the encoding network and the decoding network, for example, in the case of fixing the parameters of the matching layer of the sending end, the training of the encoding network and the decoding network train.
  • the sending end performs first encoding processing on the first data through the encoding network to obtain the first sending feature; and performs second encoding processing on the first sending feature through the matching layer to obtain the first feature. Since the encoding network and matching layer are independently trained, when the channel changes, only updating the matching layer at the sending end can adapt to the new channel, which can reduce the overhead required for network training at the receiving end. In addition, since the receiving end does not need to participate in training, it can reduce the demand on the processing capability of the receiving end and prolong the use time of the receiving end.
  • FIG. 9 is a flow chart of another communication method provided by the embodiment of the present application.
  • the method flow in FIG. 9 is a possible implementation of the method described in FIG. 8 .
  • the method flow includes:
  • the sending end performs first encoding processing on the first data through an encoding network to obtain a first sending feature.
  • step 901 refer to step 801.
  • the sending end performs second encoding processing on the first sending feature through the matching layer to obtain the first feature.
  • step 902 refer to step 802.
  • the sending end sends the first feature to the receiving end.
  • step 903 refer to step 803.
  • the sending end and the receiving end have completed the training of the encoding network and the decoding network.
  • Network training The method flow in FIG. 9 can be regarded as a method flow in which the sending end and the receiving end use the trained autoencoder to realize data transmission.
  • the sending end receives the first indication information from the receiving end.
  • the first indication information instructs the sending end to update the parameters of the matching layer.
  • the sending end updates the parameters of the matching layer according to its current channel.
  • step 905 updates the parameters of the matching layer according to its current channel, third sending feature, and third receiving feature.
  • the above-mentioned third receiving feature includes the feature obtained by the above-mentioned receiving end receiving the third feature sent by the above-mentioned sending end under the first channel, and the above-mentioned third feature includes the above-mentioned sending end using the above-mentioned matching layer to perform a second encoding process on the above-mentioned third sending feature obtained features.
  • the current channel of the sending end is different from the first channel.
  • the above-mentioned third sending feature and the above-mentioned third receiving feature are the sending feature and receiving feature recorded in the memory buffer of the sending end.
  • the third sending feature is the above-mentioned first sending feature
  • the third receiving feature is a receiving feature obtained by the receiving end receiving the first feature sent by the sending end through the first channel.
  • the first channel may be a channel between the sending end and the receiving end when the sending end sends the first characteristic. That is to say, the channel between the sending end and the receiving end is originally the first channel (when the first data is sent), and the channel changes from the first channel to the current channel. It can be understood that, before the channel changes, the sending end sends data (for example, the first characteristic) to the receiving end through the first channel.
  • the sending end may record its sending characteristics and receiving characteristics within a recent period of time through the memory buffer, for example, the first sending characteristic and the first receiving characteristic.
  • the first receiving feature is a receiving feature obtained by the receiving end receiving the first feature sent by the sending end through the first channel.
  • the sender can use any one of the methods 1, 2, and 3 described above to update the parameters of the matching layer according to its current channel, which will not be described in detail here.
  • the parameters of the encoding network described above remain unchanged during the process of updating the parameters of the matching layer at the sending end.
  • only the parameters of the matching layer are updated, and the efficiency of updating the matching layer can be improved due to the reduced calculation amount.
  • the sending end may perform the following operations: the sending end acquires first information; the sending end determines the current channel of the sending end according to the first information.
  • the first information includes channel information or receiving characteristic offset information from the receiving end; the channel information represents information related to the current channel of the transmitting end, and the receiving characteristic offset information represents the difference between the third receiving characteristic and the fourth receiving characteristic,
  • the third receiving feature includes the feature obtained by the receiving end receiving the third feature sent by the sending end in the first channel, and the fourth receiving feature includes the receiving end receiving the third feature sent by the sending end in the current channel. obtained features.
  • the sending end before performing step 905, performs the following operation: the sending end estimates the current channel according to prior information. Alternatively, the sending end estimates the current channel by monitoring the environmental changes through the sensing device.
  • the sending end performs first encoding processing on the second data through an encoding network to obtain a second sending feature.
  • step 906 refer to step 901.
  • the sending end performs second encoding processing on the second sending feature through the updated matching layer to obtain the second feature.
  • the updated matching layer is the matching layer where the training converges.
  • the sending end can use any one of the above-mentioned method 1, method 2, and method 3 to update the parameters of the matching layer. Updating the parameters of the matching layer can be regarded as training the matching layer.
  • the condition of matching layer training convergence can be that when the sending end iteratively updates the parameters of the matching layer for less than or equal to the time threshold (for example, 5s), the loss value calculated by the sending end using the parameters of the matching layer and the current channel is less than the loss threshold.
  • the loss value calculated by the sender using the parameters of the matching layer and the current channel can represent the difference between the reception characteristics calculated by the sender using the parameters of the matching layer and the current channel and the target reception characteristics.
  • the target reception characteristic may be the reception characteristic that the sender expects to calculate using the parameters of the matching layer and the current channel, that is, the ideal reception characteristic.
  • the memory buffer of the sending end records the first sending feature and the first receiving feature (corresponding to the first channel); the sending end uses the matching layer to perform the second encoding process on the first sending feature to obtain the feature t; the sending end Calculate the feature t received by the receiving end through the transmission of the current channel to obtain the receiving feature r'; the sending end uses, for example, the square error Feature covariance KL divergence Stiffness and distance mixed error Calculate the loss value of the receiving feature r' and the first receiving feature (as the target receiving feature); if the loss value is less than the loss threshold and the time length for the sending end to iteratively update the parameters of the matching layer is less than or equal to the time threshold, the sending end can determine the matching Layer training converges.
  • Updating the parameters of the matching layer can be regarded as training the matching layer.
  • the condition of matching layer training convergence may be that when the sending end iteratively updates the parameters of the matching layer for less than a preset number of times (for example, 10,000 times), the loss value calculated by the sending end using the parameters of the matching layer and the current channel is less than the loss threshold.
  • a preset number of times for example, 10,000 times
  • the training-signal-to-noise ratio (TSNR) defined below can be used as the evaluation criterion:
  • the sending end sends the second feature to the receiving end.
  • the above-mentioned second feature is used for the above-mentioned receiving end to obtain the above-mentioned second data.
  • the sending end updates the parameters of the matching layer according to its current channel; by updating the matching layer at the sending end, the adaptation to the new channel can be realized without updating the encoding network and the decoding network of the receiving end, which can avoid The time overhead and signaling overhead caused by the encoding network and the decoding network at the receiving end.
  • FIG. 10 is a flow chart of another communication method provided by the embodiment of the present application.
  • the method flow in FIG. 10 is a possible implementation of the method described in FIG. 8 .
  • the method flow includes:
  • the sending end performs first encoding processing on the first data through an encoding network to obtain a first sending characteristic.
  • step 100 refer to step 801.
  • the sending end performs a second encoding process on the first sending feature through the matching layer to obtain the first feature.
  • step 1002 refer to step 802.
  • the sending end sends the first characteristic to the receiving end.
  • step 1003 refer to step 803. Before the receiving end executes step 1001, step 1002, and step 1003, the sending end and the receiving end have completed the training of the encoding network and the decoding network. Network training.
  • the sending end receives first indication information from the receiving end.
  • the first indication information instructs the sending end to update the parameters of the matching layer.
  • Step 1004 may be replaced by: the sending end determines to update the parameters of the matching layer when the channel changes and the degree of channel change is less than a change threshold.
  • the change degree of the channel may be the covariance of the channel after the change and the channel before the change, and the change threshold is set according to actual requirements.
  • the sending end may determine whether to update the parameters of the matching layer according to the covariance between the channel after the change and the channel before the change and the change threshold.
  • the sending end updates the parameters of the matching layer according to its current channel.
  • step 1005 refer to step 905.
  • the sending end sends third indication information to the receiving end when the matching layer has not been trained and converged.
  • the third indication information instructs the receiving end to retrain the encoding network.
  • the situation where the matching layer training does not converge can be that when the sending end iteratively updates the parameters of the matching layer for a period greater than or equal to the time threshold (for example, 5s), the loss value calculated by the sending end using the parameters of the matching layer and the current channel is greater than or equal to the loss threshold .
  • the case where the matching layer training does not converge can be that when the number of times the sending end iteratively updates the parameters of the matching layer is equal to or greater than the preset number of times (for example, 10,000 times), the loss value calculated by the sending end using the parameters of the matching layer and the current channel is equal to or greater than the loss threshold.
  • the sending end and the receiving end train the encoding network of the sending end.
  • the sending end and the receiving end may train the encoding network of the sending end by: training the encoding network of the sending end under the condition that the parameters of the decoding network of the receiving end are fixed. When the parameters of the matching layer remain unchanged, the sending end trains the encoding network.
  • the sending end updates the parameters of the matching layer according to its current channel.
  • the situation where the matching layer training does not converge can be that when the sending end iteratively updates the parameters of the matching layer for a period greater than or equal to the time threshold (for example, 5s), the loss value calculated by the sending end using the parameters of the matching layer and the current channel is greater than or equal to the loss threshold .
  • the case where the matching layer training does not converge can be that when the number of times the sending end iteratively updates the parameters of the matching layer is equal to or greater than the preset number of times (for example, 10,000 times), the loss value calculated by the sending end using the parameters of the matching layer and the current channel is equal to or greater than the loss threshold.
  • the sending end may repeatedly execute steps 1006 to 1008 until the matching layer training converges. After the sending end executes step 1008, if the matching layer is not trained to converge, then the sending end executes step 1006; if the matching layer training converges, then the sending end executes step 1009.
  • the sending end performs first encoding processing on the second data through an encoding network to obtain a second sending feature.
  • step 1009 refer to step 801.
  • the sending end uses the updated matching layer to perform a second encoding process on the second sending feature to obtain the second feature.
  • step 1010 refer to step 907.
  • the sending end sends the second characteristic to the receiving end.
  • step 1011 refer to step 908.
  • the sending end when the matching layer has not been trained and converged, sends third indication information to the receiving end, so as to instruct the receiving end to retrain the encoding network through the third indication information. After retraining the encoding network, the sending end updates the parameters of the matching layer according to its current channel so that the training of the matching layer can converge.
  • FIG. 11 is a flow chart of another communication method provided by the embodiment of the present application.
  • the method flow in FIG. 11 is a possible implementation of the method described in FIG. 8 .
  • the method flow includes:
  • the sending end performs first encoding processing on the first data through an encoding network to obtain a first sending characteristic.
  • step 110 refer to step 801.
  • the sending end performs second encoding processing on the first sending feature through the matching layer to obtain the first feature.
  • step 1102 refer to step 802.
  • the sending end sends the first characteristic to the receiving end.
  • step 1103 refer to step 803.
  • the sending end and the receiving end have completed the training of the encoding network and the decoding network, for example, in the case of fixing the parameters of the matching layer of the sending end, the encoding network and decoding network are completed. Network training.
  • the sending end receives second indication information from the receiving end.
  • Step 1104 may be replaced by: the sending end determines to update the parameters of the encoding network when the channel changes and the degree of channel change is greater than or equal to the first threshold.
  • the change degree of the channel may be the covariance of the channel after the change and the channel before the change, and the first threshold is set according to actual requirements.
  • the sending end may determine whether to update the parameters of the coding network according to the covariance between the channel after the change and the channel before the change and the first threshold.
  • the sending end and the receiving end train the encoding network of the sending end.
  • the sending end and the receiving end may train the encoding network of the sending end by: training the encoding network of the sending end under the condition that the parameters of the decoding network of the receiving end are fixed.
  • the sending end updates the parameters of the matching layer according to its current channel.
  • step 1106 refer to step 1008.
  • the sending end executes step 1106, if the matching layer is not trained to converge, then the sending end can execute steps 1105 and 1106 multiple times until the matching layer training converges; if the matching layer training converges, then the sending end executes step 1107.
  • the sending end performs first encoding processing on the second data through an encoding network to obtain a second sending feature.
  • step 1107 refer to step 801.
  • the sending end uses the updated matching layer to perform a second encoding process on the second sending feature to obtain the second feature.
  • step 1108 refer to step 907.
  • the sending end sends the second characteristic to the receiving end.
  • step 1109 refer to step 908.
  • the sending end trains the coding network of the sending end with the receiving end.
  • the sending end updates the parameters of the matching layer according to its current channel; the encoding network and matching layer suitable for the current channel can be trained faster.
  • FIG. 8 to FIG. 11 describe the flow of the method executed by the sending end in the communication solution provided by the present application.
  • the flow of the method executed by the receiving end in the communication solution provided by the present application will be described below with reference to the accompanying drawings.
  • FIG. 12 is a flowchart of a communication method provided by an embodiment of the present application. As shown in Figure 12, the method includes:
  • the receiving end receives a first reception characteristic from the sending end.
  • the first receiving feature includes a feature that the first feature sent by the sending end is received by the receiving end through channel transmission.
  • the first feature is obtained by the sending end encoding the first sending feature through the matching layer, and the first sending feature is obtained by encoding the first data through the encoding network of the sending end.
  • the above encoding network and the above matching layer are obtained through independent training.
  • the receiving end may be an access network device or a user equipment.
  • the receiving end performs decoding processing on the first receiving feature through a decoding network to obtain first data.
  • the above-mentioned decoding network and the above-mentioned matching layer are obtained through independent training.
  • the sending end and the receiving end have completed the training of the encoding network and the decoding network, for example, in the case of fixing the parameters of the matching layer of the sending end, the training of the encoding network and the decoding network train.
  • the decoding network and the matching layer are independently trained, and the encoding network and the matching layer are independently trained, when the channel between the sending end and the receiving end changes, only the The matching layer can realize the adaptation to the new channel, which can reduce the overhead required for network training at the receiving end.
  • the receiving end since the receiving end does not need to participate in training, it can reduce the demand on the processing capability of the receiving end and prolong the use time of the receiving end.
  • FIG. 13 is a flow chart of another communication method provided by the embodiment of the present application.
  • the method flow in FIG. 13 is a possible implementation of the method described in FIG. 12 .
  • the method flow includes:
  • the receiving end receives a first reception characteristic from the sending end.
  • step 1301 refer to step 1201.
  • the receiving end performs decoding processing on the first receiving feature through the decoding network to obtain the first data.
  • step 1302 refer to step 1202. Before the receiving end performs step 1301 and step 1302, the sending end and the receiving end have completed the training of the encoding network and the decoding network.
  • the receiving end sends first indication information to the sending end when the parameter characterizing the degree of channel change is less than or equal to the first threshold.
  • the first indication information instructs the sending end to update the parameters of the matching layer.
  • the receiving end periodically detects channel changes.
  • the parameter characterizing the variation degree of the channel may be the covariance of the channel.
  • the receiving end detects the covariance between the current channel and the last detected channel every 10ms; if it detects that the covariance between the current channel and the last detected channel is greater than the second threshold and less than the first threshold, then send The terminal sends the first indication information; wherein, the second threshold is smaller than the first threshold, and both the first threshold and the second threshold are real numbers greater than 0 set according to actual needs.
  • the receiving end may further perform the following operation: send first information to the sending end, where the first information is used by the sending end to update the parameters of the matching layer.
  • the first information includes channel information or receiving characteristic offset information from the receiving end; the channel information represents information related to the current channel of the transmitting end, and the receiving characteristic offset information represents the difference between the third receiving characteristic and the fourth receiving characteristic,
  • the third receiving feature includes the feature obtained by the receiving end receiving the third feature sent by the sending end in the first channel
  • the fourth receiving feature includes the receiving end receiving the third feature sent by the sending end in the current channel. obtained features.
  • the receiving end before the receiving end sends the first indication information to the sending end, the following operations may be performed: the receiving end receives the fourth indication information from the sending end, and the fourth indication information instructs the encoding network to complete the training .
  • the receiving end when the parameter representing the degree of channel change is less than or equal to the first threshold, the receiving end sends the first indication information to the sending end; the sending end can be instructed to update the parameters of the matching layer in time, so that after the channel changes , the data transfer can still complete successfully.
  • FIG. 14 is a flow chart of another communication method provided by the embodiment of the present application.
  • the method flow in FIG. 14 is a possible implementation of the method described in FIG. 12 .
  • the method flow includes:
  • the receiving end receives a first receiving characteristic from the sending end.
  • step 1401 refer to step 1201.
  • the receiving end performs decoding processing on the first receiving feature through the decoding network to obtain the first data.
  • step 1402 refer to step 1202.
  • the receiving end sends second indication information to the sending end when the parameter characterizing the degree of channel change is greater than the first threshold.
  • the second instruction information instructs the sending end to retrain the encoding network.
  • the sending end and the receiving end have completed the training of the encoding network and the decoding network, for example, when the parameters of the matching layer of the sending end are fixed, the encoding network and decoding Network training.
  • the receiving end periodically detects channel changes.
  • the parameter characterizing the variation degree of the channel may be the covariance of the channel. For example, the receiving end detects the covariance between the current channel and the last detected channel every 10 ms; if it detects that the covariance between the current channel and the last detected channel is greater than the first threshold, a second indication is sent to the sending end Information; wherein, the first threshold is a real number greater than 0 set according to actual needs.
  • the receiving end and the sending end train the encoding network of the sending end.
  • the receiving end and the sending end may train the encoding network of the sending end by fixing the parameters of the decoding network unchanged, and training the encoding network of the sending end under the current channel.
  • the receiving end when the parameter representing the degree of channel change is greater than the first threshold, the receiving end sends the second indication information to the sending end, so as to instruct the sending end to retrain the encoding network; it can solve the problem that simply updating the matching layer parameters cannot Issues with successfully completing data transfers under a new channel.
  • FIG. 15 is a flow chart of another communication method provided by the embodiment of the present application.
  • the method flow in FIG. 15 is a possible implementation of the method described in FIG. 12 .
  • the method flow includes:
  • the receiving end receives a first receiving characteristic from the sending end.
  • step 1501 refer to step 1201.
  • step 1502 refer to step 1202. Before the receiving end performs step 1501 and step 1502, the sending end and the receiving end have completed the training of the encoding network and the decoding network.
  • the receiving end sends first indication information to the sending end when the parameter characterizing the degree of channel change is less than or equal to the first threshold.
  • step 1503 refer to step 1302.
  • the receiving end sends the first information to the sending end.
  • the first information is used by the sending end to update the parameters of the matching layer.
  • the first information includes channel information or receiving characteristic offset information from the receiving end; the channel information represents information related to the current channel of the transmitting end, and the receiving characteristic offset information represents the difference between the third receiving characteristic and the fourth receiving characteristic,
  • the third receiving feature includes the feature obtained by the receiving end receiving the third feature sent by the sending end in the first channel, and the fourth receiving feature includes the receiving end receiving the third feature sent by the sending end in the current channel. obtained features.
  • Step 1501 is optional.
  • the sending end may update the parameters of the matching layer according to the first information from the receiving end, or may update the parameters of the matching layer in other ways (without using the first information), for example, the sending end estimates the current channel according to prior information.
  • the receiving end receives third indication information from the sending end.
  • the third indication information instructs the receiving end to retrain the encoding network.
  • the receiving end and the sending end train the encoding network of the sending end.
  • step 1506 refer to step 1404.
  • the receiving end when the parameter representing the degree of channel change is less than or equal to the first threshold, the receiving end sends the first indication information to the sending end; it can make the sending end update the parameters of the matching layer in time, so as to ensure that the data is successfully transmission.
  • the receiving end After the receiving end receives the third indication information from the sending end, it trains the encoding network of the sending end with the sending end; it can solve the problem that simply updating the parameters of the matching layer cannot successfully complete the data transmission under the new channel.
  • FIG. 8 to FIG. 11 describe the flow of the method executed by the sending end after the autoencoder completes the training.
  • FIG. 12 to FIG. 15 describe the flow of the method executed by the receiving end after the self-encoder completes the training. The following describes the process of first training the autoencoder at the sending end and the receiving end and then training the matching layer with reference to the accompanying drawings.
  • FIG. 16 is a flowchart of another communication method provided by the embodiment of the present application.
  • the method in FIG. 16 describes the training process and signaling interaction of a single transceiver (ie, a transmitter and a receiver). As shown in Figure 16, the method includes:
  • the sending end initializes the encoding network and the matching layer.
  • the sender can use any method to initialize the parameters of its encoding network, which is not limited in this application.
  • the receiver initializes the decoding network.
  • the receiving end can use any method to initialize the parameters of its decoding network, which is not limited in this application.
  • the sequence of step 1601 and step 1602 is not limited.
  • the sending end sends the first training feature to the receiving end, and saves the transmission task in its memory buffer.
  • step 1603 is as follows: the sending end performs first encoding processing on the first sending data through the encoding network to obtain the first training sending feature; the sending end performs second encoding processing on the first training sending feature through the matching layer, The first training feature is obtained; the sending end sends the first training feature to the receiving end; the sending end saves the transmission task of sending the first sending data in its memory buffer. For example, the sending end stores the first training feature through a memory buffer.
  • the receiving end decodes the received first training reception features through the decoding network to obtain first reception data.
  • the first training reception feature is a feature that the first training feature is received by the receiving end through channel transmission.
  • the receiving end calculates the loss value and gradient information of the objective function according to the first received data and the first sent data.
  • the receiving end may pre-store the first sending data (ie training data).
  • the objective function can be the mean squared error or binary cross entropy wait.
  • N is the number of training batches, are the first sent data at the sending end and the first received data decoded at the receiving end, p ic is the predicted probability that the decoded data i belongs to category c, and u ic is the category sign function (0 or 1).
  • the receiving end trains the decoding network according to the gradient information.
  • the receiving end feeds back the gradient information to the sending end.
  • the receiving end may also feed back channel information ⁇ h d to the sending end. If the sender has the ability to obtain ⁇ h d through sensing or channel prediction, the receiver does not need feedback.
  • the purpose of receiving channel information feedback at the receiving end is to enable the sending end to perform training on the channel distribution dimension T, and to obtain the approximate distribution of the channel by performing statistical clustering analysis on the current channel, determine the dimension T, and perform training on each dimension T. channel for training.
  • the sending end trains the encoding network according to the gradient information.
  • the sending end and the receiving end may repeatedly execute step 1603 to step 1608 until the calculated loss value of the objective function is less than or equal to the loss threshold. That is to say, the sending end and the receiving end can use different training data to conduct training under the same channel until the loss value of the objective function calculated by the receiving end is less than or equal to the loss threshold.
  • the sending end saves the sending characteristic and the receiving characteristic through a memory buffer.
  • the memory buffer can use the first-in-first-out queue to store data, and the sending end can track the sending characteristics and receiving characteristics at any time through the memory buffer to update the training data.
  • the receiving end fixes the decoding network, and continues to jointly train the encoding network under different channel distributions with the sending end until the pre-training is completed.
  • Steps 1601 to 1610 are the steps of training the autoencoder at the sending end and the receiving end. That is to say, the sending end and the receiving end have completed the training of the autoencoder under different channel distributions by performing steps 1601 to 1610.
  • the receiving end determines a training method according to the channel change degree.
  • step 1611 is as follows: if the parameter representing the degree of channel change is less than or equal to the first threshold, the receiver instructs the sender to retrain the encoding network under the current channel; otherwise, instructs the sender to update the parameters of the matching layer.
  • the receiving end sends the first indication information to the sending end.
  • the first indication information instructs the sending end to update the parameters of the matching layer.
  • step 1612 refer to step 1503.
  • the receiving end sends the first information to the sending end.
  • step 1613 refer to step 1504.
  • the sending end updates the parameters of the matching layer according to the sending feature and receiving feature recorded in the memory buffer and the current channel.
  • the sending end can adopt any one of the above-mentioned method 1, method 2, and method 3 to update the parameters of the matching layer according to the sending characteristics and receiving characteristics recorded in the memory buffer and the current channel.
  • the sending end can obtain the optimal matching layer parameters through an iterative method.
  • the sending end performs data transmission when the matching layer training converges.
  • Step 1615 may be replaced by: the sending end sends third indication information to the receiving end when the matching layer training has not converged.
  • the third indication information instructs the receiving end to retrain the encoding network.
  • the autoencoder is trained first, and then the matching layer is trained.
  • the matching layer is trained.
  • the channel changes only the parameters of the matching layer need to be updated, which can save communication overhead.
  • the following simulates the transceiver training in the time-varying channel scenario, that is, the training of the sending end and the receiving end.
  • the performance under the time-varying fading channel can reach the same performance as that under the specific channel h s .
  • the iterative update of the matching layer can realize the adjustment of the current channel in a shorter time.
  • the communication scheme provided by this application can be implemented as the selection and switching of different receivers by the sender, and also make each receiver The network does not require frequent training.
  • the ability to extract channel features in multiple scenarios is adaptively adjusted by the matching layer according to the receiving feature offset or channel information such as CSI fed back by the sending end, so as to select the receiving end that is currently communicating with it Corresponding characteristics of the sender, for communication transmission.
  • FIG. 17 is a flowchart of another communication method provided by the embodiment of the present application.
  • the method in FIG. 17 describes the training process and signaling interaction of multiple transceivers (ie, one transmitter and multiple receivers). As shown in Figure 17, the method includes:
  • the sending end initializes the coding network and the matching layer.
  • step 1701 refer to step 1601.
  • the receiver initializes the decoding network.
  • step 1702 refer to step 1602.
  • the sending end trains the autoencoder under the current channel with different receiving ends respectively.
  • the sending end and receiving end 1 train the encoding network of the sending end and the decoding network of the receiving end 1
  • the sending end and the receiving end 2 train the encoding network of the sending end and the decoding network of the receiving end 2.
  • the sending end respectively records and updates the transmission task data with different receiving ends through the memory buffer.
  • the transmission task data may include sending characteristics of the sending end and receiving characteristics of the receiving end.
  • the receiving end instructs the sending end to complete the training of the autoencoder.
  • the receiving end 1 sends the first indication information to the sending end.
  • step 1707 refer to step 1503.
  • the receiving end 1 sends the first information to the sending end.
  • step 1708 refer to step 1504.
  • the sending end updates the parameters of the matching layer according to the sending feature and receiving feature recorded in the memory buffer and the current channel.
  • the sending end performs data transmission when the matching layer training converges.
  • the sending end sends communication data through the encoding network and the updated matching layer, and the receiving end decodes the data through the original decoding network.
  • the autoencoder is trained first, and then the matching layer is trained.
  • the matching layer is trained.
  • the channel changes only the parameters of the matching layer need to be updated, which can save communication overhead.
  • FIG. 18 is a schematic structural diagram of a communication device provided by an embodiment of the present application.
  • the communication device in FIG. 18 may be the sending end in the foregoing embodiments.
  • a communication device 1800 includes: a processing module 1801 and a transceiver module 1802 .
  • the processing module 1801 is configured to perform a first encoding process on the first data through an encoding network to obtain a first transmission feature; the above-mentioned first transmission feature is related to the channel distribution dimension of the environment where the sending end is located;
  • the processing module 1801 is further configured to perform a second encoding process on the above-mentioned first transmission feature through the matching layer to obtain the first feature; the above-mentioned coding network and the above-mentioned matching layer are obtained through independent training; the dimension of the above-mentioned first feature is smaller than the above-mentioned first feature Dimensions of sending features;
  • the transceiver module 1802 is configured to send the above-mentioned first feature to the receiving end; the above-mentioned first feature is used for the above-mentioned receiving end to obtain the above-mentioned first data.
  • the processing module 1801 is specifically configured to update the parameters of the above-mentioned matching layer according to its current channel; the processing module 1801 is also used to perform the first coding process on the second data through the above-mentioned coding network to obtain The second transmission feature; the second encoding process is performed on the second transmission feature through the updated matching layer to obtain the second feature; the above-mentioned transceiver module is also used to send the above-mentioned second feature to the above-mentioned receiving end; the above-mentioned second feature It is used for the receiving end to obtain the second data.
  • the transceiver module 1802 is further configured to receive first indication information from the receiving end, where the first indication information instructs the sending end to update the parameters of the matching layer.
  • the processing module 1801 is specifically configured to update the parameters of the matching layer according to the current channel, the third sending feature, and the third receiving feature;
  • the third receiving feature includes The feature obtained by receiving the third feature sent by the sending end under the channel, the third feature includes the feature obtained by the sending end using the matching layer to perform the second encoding process on the third sending feature, the current channel of the sending end is the same as the first One channel is different.
  • the processing module 1801 is further configured to acquire first information; and determine the current channel of the sending end according to the first information.
  • the processing module 1801 is further configured to train the encoding network when the parameters of the matching layer remain unchanged.
  • the transceiver module 1802 is further configured to receive second instruction information from the receiving end, where the second instruction information instructs the sending end to retrain the encoding network.
  • the transceiver module 1802 is further configured to send third indication information to the receiving end when the matching layer has not been trained and converged, and the third indication information instructs the receiving end to retrain the encoding network.
  • FIG. 19 is a schematic structural diagram of another communication device provided by an embodiment of the present application.
  • the communication device in FIG. 19 may be the receiving end in the foregoing embodiments.
  • a communication device 1900 includes: a transceiver module 1901 and a processing module 1902 .
  • the transceiver module 1901 is configured to receive the first receiving feature from the sending end; the first receiving feature includes the feature that the first feature sent by the sending end is received by the receiving end through channel transmission, and the first feature is that the sending end passes matching
  • the layer performs encoding processing on the first sending feature, and the first sending feature is obtained by encoding the first data through the encoding network at the sending end; the encoding network and the matching layer are obtained through independent training;
  • the processing module 1902 is configured to perform decoding processing on the above-mentioned first received feature through a decoding network to obtain the above-mentioned first data; the above-mentioned decoding network and the above-mentioned matching layer are obtained through independent training.
  • the transceiver module 1901 is further configured to send first indication information to the sending end, where the first indication information instructs the sending end to update the parameters of the matching layer.
  • the transceiver module 1901 is specifically configured to send the above-mentioned first indication information to the above-mentioned sending end when the parameter characterizing the degree of channel change is less than or equal to a first threshold.
  • the transceiver module 1901 is further configured to send second indication information to the above-mentioned sending end when the parameter characterizing the degree of channel change is greater than the first threshold; the above-mentioned second indication information indicates that the above-mentioned sending end Retrain the above encoding network.
  • the transceiver module 1901 is further configured to receive third instruction information from the sending end, where the third instruction information instructs the receiving end to retrain the encoding network.
  • the transceiver module 1901 is further configured to receive fourth indication information from the sending end, where the fourth indication information indicates that the encoding network has completed training.
  • the transceiver module 1901 is further configured to send first information to the sending end, where the first information is used by the sending end to update the parameters of the matching layer.
  • FIG. 20 is a schematic structural diagram of another communication device 200 provided in an embodiment of the present application.
  • the communication device in Fig. 20 may be the above-mentioned sending end.
  • the communication device in FIG. 20 may be the above-mentioned receiving end.
  • the communication device 200 includes at least one processor 2020 and a transceiver 2010 .
  • the processor 2020 and the transceiver 2010 may be configured to perform the above-mentioned functions or operations performed by the sending end.
  • the processor 2020 may perform one or more of the following operations: Step 801, Step 802 in FIG. 8, Step 901, Step 902, Step 905, Step 906, Step 907 in FIG. Step 1002, step 1005, step 1007, step 1008, step 1009, step 1010, step 1101, step 1102, step 1105, step 1106, step 1107, step 1108 in FIG.
  • the transceiver 2010 may perform one or more of the following operations: step 803 in FIG. 8, step 903, step 904, and step 908 in FIG. 9, step 1003, step 1004, step 1006, and step 1011 in FIG. Step 1103, step 1104, and step 1109 in 11.
  • the processor 2020 and the transceiver 2010 may be configured to perform the above functions or operations performed by the receiving end.
  • the processor 2020 may perform one of the following multiple operations: step 1202 in FIG. 12 , step 1302 in FIG. 13 , step 1402 and step 1404 in FIG. 14 , and step 1502 and step 1506 in FIG. 15 .
  • the transceiver 2010 may perform one or more of the following operations: step 1201 in FIG. 12 , step 1301 and step 1303 in FIG. 13 , step 1401 and step 1403 in FIG. 14 , step 1501 and step 1503 in FIG. 15 , Step 1504, step 1505.
  • the transceiver 2010 is used to communicate with other devices/apparatus through transmission media.
  • the processor 2020 uses the transceiver 2010 to send and receive data and/or signaling, and is used to implement the methods in the foregoing method embodiments.
  • the processor 2020 can realize the function of the processing module 1801 , and the transceiver 2010 can realize the function of the transceiver module 1802 .
  • the communication device 200 may further include at least one memory 2030 for storing program instructions and/or data.
  • the memory 2030 is coupled to the processor 2020 .
  • the coupling in the embodiments of the present application is an indirect coupling or a communication connection between devices, units or modules, which may be in electrical, mechanical or other forms, and is used for information exchange between devices, units or modules.
  • Processor 2020 may cooperate with memory 2030 .
  • Processor 2020 may execute program instructions stored in memory 2030 . At least one of the at least one memory may be included in the processor.
  • the specific connection medium among the transceiver 2010, the processor 2020, and the memory 2030 is not limited.
  • the memory 2030, the processor 2020, and the transceiver 2010 are connected through the bus 2040.
  • the bus is represented by a thick line in FIG. 20, and the connection between other components is only for schematic illustration. , is not limited.
  • the bus can be divided into address bus, data bus, control bus and so on. For ease of representation, only one thick line is used in FIG. 20 , but it does not mean that there is only one bus or one type of bus.
  • the processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, and may implement or Execute the methods, steps and logic block diagrams disclosed in the embodiments of the present application.
  • a general purpose processor may be a microprocessor or any conventional processor or the like. The steps of the methods disclosed in connection with the embodiments of the present application may be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.
  • FIG. 21 is a schematic structural diagram of another communication device 210 provided by an embodiment of the present application.
  • the communication device shown in FIG. 21 includes a logic circuit 2101 and an interface 2102 .
  • the processing module 1801 in FIG. 18 can be realized by a logic circuit 2101
  • the transceiver module 1802 in FIG. 18 can be realized by an interface 2102 .
  • the processing module 1902 in FIG. 19 can be realized by a logic circuit 2101
  • the transceiver module 1901 in FIG. 19 can be realized by an interface 2102 .
  • the logic circuit 2101 may be a chip, a processing circuit, an integrated circuit or a system on chip (SoC) chip, etc.
  • the interface 2102 may be a communication interface, an input-output interface, or the like.
  • the logic circuit and the interface may also be coupled to each other. The embodiment of the present application does not limit the specific connection manner of the logic circuit and the interface.
  • the logic circuit and the interface may be used to perform the above-mentioned functions or operations performed by the sending end.
  • the logic circuit and the interface may be used to perform the functions or operations performed by the receiving end described above.
  • the present application also provides a computer-readable storage medium, where computer codes are stored in the computer-readable storage medium, and when the computer codes are run on the computer, the computer is made to execute the methods of the above-mentioned embodiments.
  • the present application also provides a computer program product.
  • the computer program product includes computer code or computer program.
  • the communication method in the above-mentioned embodiments is executed.
  • the present application also provides a communication system, including the above-mentioned receiving end and the above-mentioned sending end.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请实施例公开了一种通信方法和通信装置，该方法包括：发送端通过编码网络对第一数据做第一编码处理，得到第一发送特征；第一发送特征与发送端所处环境的信道分布维度相关；发送端通过匹配层对第一发送特征做第二编码处理，得到第一特征；编码网络和匹配层是独立训练得到的；发送端向接收端发送第一特征；第一特征用于接收端获得第一数据。由于编码网络和匹配层是独立训练得到的，因此当信道发生变化时，仅更新发送端的匹配层就能实现对新信道的适应，可用更短时间实现对当前信道的调整，并减少因接收端的网络训练所需要的开销。另外，由于接收端不需要参与训练，可以降低对接收端处理能力的需求，延长接收端的使用时长。

Description

通信方法和通信装置
本申请要求于2021年12月22日提交中国国家知识产权局、申请号为202111583487.7、申请名称为“通信方法和通信装置”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
本申请涉及计算机领域，尤其涉及一种通信方法和通信装置。
背景技术
现代的通信系统设计是模块化的，信号处理的过程被分为一系列子模块，例如信源编码、信道编码、调制、信道估计等。每个子模块都是基于特定的信号处理算法建模，通常是近似为一些简化的线性模型。然而，用这种单独优化每个子模块的方式并不能保证整个物理层端到端的通信最优。相反，传统的端到端通信系统引入了更多的干扰效应，如放大器失真和信道损伤，并且因为要控制的因素和参数数量的增加。因此，使用传统方法来进行端到端优化的复杂性非常高。
随着深度学习技术的发展，有研究人员提出通过基于自编码器来取代传统通信收发机设计，将发送端和接收端用神经网络的方式建模，并通过大量训练样本学习数据的分布，然后用来预测结果。这样的端到端学习方式能够做到联合优化，相比现有方法可以做到更优的效果。然而，端到端通信系统中还有信道这个外在环境因素的影响。真实通信场景中的信道并非一成不变的，特别是在时变瑞利衰弱信道(time-vary rayleigh fading channel)下，仅在特定信道下训练好的自编码器(autoencoder，AE)网络在遇到不可预测信道响应下将出现匹配错误的情况(等同于训练数据集出现异常值的情况)，则必须要进行重新训练调整，这将带来较大的收发端通信开销。因此研究通信开销较少的通信方案。
发明内容
本申请实施例公开了一种通信方法和通信装置，可通信开销较少。
第一方面，本申请实施例提供一种通信方法，该方法包括：发送端通过编码网络对第一数据做第一编码处理，得到第一发送特征；所述第一发送特征与所述发送端所处环境的信道分布维度相关；所述发送端通过匹配层对所述第一发送特征做第二编码处理，得到第一特征；所述编码网络和所述匹配层是独立训练得到的；所述第一特征的维度小于所述第一发送特征的维度；所述发送端向接收端发送所述第一特征；所述第一特征用于所述接收端获得所述第一数据。
本申请实施例中，发送端通过编码网络对第一数据做第一编码处理，得到第一发送特征；通过匹配层对该第一发送特征做第二编码处理，得到第一特征。由于编码网络和匹配层是独立训练得到的，因此当信道发生变化时，仅更新发送端的匹配层就能实现对新信道的适应，可用更短时间实现对当前信道的调整，并减少因接收端的网络训练所需要的开销。另外，由于接收端不需要参与训练，可以降低对接收端处理能力的需求，延长接收端的使用时长。
在一种可能的实现方式中，所述方法还包括：所述发送端根据其当前信道，更新所述匹配层的参数；所述发送端通过所述编码网络对第二数据做第一编码处理，得到第二发送特征；所述发送端通过更新后的所述匹配层对所述第二发送特征做第二编码处理，得到第二特征；所述发送端向所述接收端发送所述第二特征；所述第二特征用于所述接收端获得所述第二数据。
在该实现方式中，发送端根据其当前信道，更新匹配层的参数；通过更新发送端的匹配层就能实现对新信道的适应，不需要更新编码网络和接收端的译码网络，可以避免因更新编码网络和接收端的译码网络造成的时间开销和信令开销。
在一种可能的实现方式中，所述发送端更新所述匹配层的参数之前，所述方法还包括：所述发送端接收来自所述接收端的第一指示信息，所述第一指示信息指示所述发送端更新所述匹配层的参数。
在该实现方式中，发送端接收来自接收端的第一指示信息，以便根据该第一指示信息及时更新匹配层的参数。
在一种可能的实现方式中，所述编码网络的参数在所述发送端更新所述匹配层的参数的过程中，保持不变。
在该实现方式中，编码网络的参数在发送端更新匹配层的参数的过程中，保持不变；仅更新匹配层的参数就可使用当前信道，可以减少运算量，并提高更新匹配层的效率。
在一种可能的实现方式中，所述编码网络是在多种不同信道下训练得到的。
编码网络可视为多个独立的子编码网络的堆叠，每个子编码网络为该子编码网络在一种特定信道下与参数固定不变的译码网络联合训练得到，任意两个子编码网络是在不同特定信道下训练得到的。
在该实现方式中，由于编码网络是在多种不同信道下训练得到的，因此该编码网络能够处理多个信道情况，即该编码网络可适用于多种不同的信道。当编码网络可适用于多种不同的信道时，若发送端的信道发生变化，则该发送端不必更新编码网络的参数，仅需更新匹配层的参数。所述多种不同信道可以是对所述发送端当前所处环境的信道做聚类划分得到。
在一种可能的实现方式中，所述发送端根据其当前信道，更新所述匹配层的参数包括：所述发送端根据其当前信道、第三发送特征、第三接收特征，更新所述匹配层的参数；所述第三接收特征包括所述接收端在第一信道下接收所述发送端发送的第三特征得到的特征，所述第三特征包括所述发送端利用所述匹配层对所述第三发送特征做第二编码处理得到的特征，所述发送端的当前信道与所述第一信道不同。
在该实现方式中，发送端根据其当前信道、第三发送特征、第三接收特征，更新匹配层的参数；不需要与接收端交互来获取除用于获得当前信道之外的信息，可以减少通信开销。
在一种可能的实现方式中，在所述发送端根据其当前信道，更新所述匹配层的参数之前，所述方法还包括：所述发送端获取第一信息；所述发送端根据所述第一信息，确定所述发送端的当前信道。
在该实现方式中，发送端根据第一信息，确定发送端的当前信道，以便更新匹配层的参数。
在一种可能的实现方式中，所述第一信息包括来自所述接收端的信道信息或者接收特征偏移信息；所述信道信息表征所述发送端的当前信道的相关信息，所述接收特征偏移信息表征第三接收特征和第四接收特征的差异，所述第三接收特征包括所述接收端在第一信道下接收所述发送端发送的第三特征得到的特征，所述第四接收特征包括所述接收端在当前信道下 接收所述发送端发送的所述第三特征得到的特征。
在该实现方式中，第一信息包括来自接收端的信道信息或者接收特征偏移信息，以便发送端利用该第一信息获得其当前信道。
在一种可能的实现方式中，所述第一发送特征包括与信道分布维度相关的L维向量，所述L为V和T的乘积，所述第一数据至少用V维向量表征，所述T由对当前环境的信道做聚类得到的信道种类，所述T为大于或等于2的整数，所述V为大于0的整数。
在该实现方式中，第一发送特征包括与信道分布维度相关的L维向量，可以适用于不同的信道。
在一种可能的实现方式中，所述方法还包括：在所述匹配层的参数不变的情况下，所述发送端训练所述编码网络。
在该实现方式中，在匹配层的参数不变的情况下，发送端训练编码网络；既能保证编码网络和匹配层之间的独立，又能提高训练编码网络的速度。
在一种可能的实现方式中，所述方法还包括：所述发送端接收来自所述接收端的第二指示信息，所述第二指示信息指示所述发送端重新训练所述编码网络。所述发送端可根据所述第二指示信息重新训练所述编码网络。
在该实现方式中，发送端接收来自接收端的第二指示信息，以便根据该第二指示信息重新训练编码网络。
在一种可能的实现方式中，所述方法还包括：所述发送端在所述匹配层未训练收敛的情况下，向所述接收端发送第三指示信息，所述第三指示信息指示所述接收端重新训练所述编码网络。
在该实现方式中，发送端在匹配层未训练收敛的情况下，向接收端发送第三指示信息；可及时停止训练匹配层，以便通过重新训练的编码网络来训练得到收敛的匹配层。
在一种可能的实现方式中，所述匹配层是可微的。
在该实现方式中，匹配层是可微的，因此不影响梯度回传。
第二方面，本申请实施例提供另一种通信方法，该方法包括：接收端接收来自发送端的第一接收特征；所述第一接收特征包括所述发送端发送的第一特征经过信道传输被所述接收端接收到的特征，所述第一特征为所述发送端通过匹配层对第一发送特征做编码处理得到，所述第一发送特征为所述发送端编码网络对第一数据做编码处理得到；所述编码网络和所述匹配层是独立训练得到的；所述接收端通过译码网络对所述第一接收特征做译码处理，得到所述第一数据；所述译码网络和所述匹配层是独立训练得到的。
本申请实施例中，接收端接收来自发送端的第一接收特征。由于编码网络和匹配层是独立训练得到的，因此当信道发生变化时，仅更新发送端的匹配层就能实现对新信道的适应，可用更短时间实现对当前信道的调整，并减少因接收端的网络训练所需要的开销。另外，由于接收端不需要参与训练，可以降低对接收端处理能力的需求，延长接收端的使用时长。
在一种可能的实现方式中，所述方法还包括：所述接收端向所述发送端发送第一指示信息，所述第一指示信息指示所述发送端更新所述匹配层的参数。
在该实现方式中，接收端向发送端发送第一指示信息，以便及时指示该发送端更新匹配层的参数，从而保证在新信道下可准确地传输数据。
在一种可能的实现方式中，所述接收端向所述发送端发送第一指示信息包括：所述接收端在表征信道变化程度的参数小于或等于第一阈值的情况下，向所述发送端发送所述第一指示信息。
在该实现方式中，接收端在表征信道变化程度的参数小于或等于第一阈值的情况下，向发送端发送第一指示信息，以便该发送端在与该接收端之间的信道变化程度较小时，更新匹配层的参数，从而适应新信道。
在一种可能的实现方式中，所述方法还包括：所述接收端在表征信道变化程度的参数大于第一阈值的情况下，向所述发送端发送第二指示信息；所述第二指示信息指示所述发送端重新训练所述编码网络。
在该实现方式中，接收端在表征信道变化程度的参数大于第一阈值的情况下，向发送端发送第二指示信息；可以解决单纯更新匹配层的参数无法在新信道下成功完成数据传输的问题。
在一种可能的实现方式中，所述接收端向所述发送端发送第一指示信息之后，所述方法还包括：所述接收端接收来自所述发送端的第三指示信息，所述第三指示信息指示所述接收端重新训练所述编码网络。
在该实现方式中，接收端接收来自发送端的第三指示信息，以便重新训练编码网络，可以解决单纯更新匹配层的参数无法在新信道下成功完成数据传输的问题。
在一种可能的实现方式中，所述接收端向所述发送端发送第一指示信息之前，所述方法还包括：所述接收端接收来自所述发送端的第四指示信息，所述第四指示信息指示所述编码网络完成训练。
在该实现方式中，接收端接收来自发送端的第四指示信息，可以及时获知编码网络完成训练。
在一种可能的实现方式中，所述接收端向所述发送端发送第一指示信息之后，所述方法还包括：所述接收端向所述发送端发送第一信息，所述第一信息用于所述发送端更新所述匹配层的参数。
在该实现方式中，接收端向发送端发送第一信息，以便该发送端利用该第一信息更新匹配层的参数。
第三方面，本申请实施例提供一种通信装置，该通信装置具有实现上述第一方面方法实施例中的行为的功能。所述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。所述硬件或软件包括一个或多个与上述功能相对应的模块或单元。在一种可能的实现方式中，包括处理模块和收发模块，其中：所述处理模块，用于通过编码网络对第一数据做第一编码处理，得到第一发送特征；所述第一发送特征与发送端所处环境的信道分布维度相关；所述处理模块，还用于通过匹配层对所述第一发送特征做第二编码处理，得到第一特征；所述编码网络和所述匹配层是独立训练得到的；所述第一特征的维度小于所述第一发送特征的维度；所述收发模块，用于向接收端发送所述第一特征；所述第一特征用于所述接收端获得所述第一数据。
在一种可能的实现方式中，所述处理模块，具体用于根据其当前信道，更新所述匹配层的参数；所述处理模块，还用于通过所述编码网络对第二数据做第一编码处理，得到第二发送特征；通过更新后的所述匹配层对所述第二发送特征做第二编码处理，得到第二特征；所述收发模块，还用于向所述接收端发送所述第二特征；所述第二特征用于所述接收端获得所述第二数据。
在一种可能的实现方式中，所述收发模块，还用于接收来自所述接收端的第一指示信息，所述第一指示信息指示所述发送端更新所述匹配层的参数。
在一种可能的实现方式中，所述编码网络的参数在所述发送端更新所述匹配层的参数的 过程中，保持不变。
在一种可能的实现方式中，所述编码网络是在多种不同信道下训练得到的。
在一种可能的实现方式中，所述处理模块，具体用于根据其当前信道、第三发送特征、第三接收特征，更新所述匹配层的参数；所述第三接收特征包括所述接收端在第一信道下接收所述发送端发送的第三特征得到的特征，所述第三特征包括所述发送端利用所述匹配层对所述第三发送特征做第二编码处理得到的特征，所述发送端的当前信道与所述第一信道不同。
在一种可能的实现方式中，所述处理模块，还用于获取第一信息；根据所述第一信息，确定所述发送端的当前信道。
在一种可能的实现方式中，所述第一信息包括来自所述接收端的信道信息或者接收特征偏移信息；所述信道信息表征所述发送端的当前信道的相关信息，所述接收特征偏移信息表征第三接收特征和第四接收特征的差异，所述第三接收特征包括所述接收端在第一信道下接收所述发送端发送的第三特征得到的特征，所述第四接收特征包括所述接收端在当前信道下接收所述发送端发送的所述第三特征得到的特征。
在一种可能的实现方式中，所述第一发送特征包括与信道分布维度相关的L维向量，所述L为V和T的乘积，所述第一数据至少用V维向量表征，所述T由对当前环境的信道做聚类得到的信道种类，所述T为大于或等于2的整数，所述V为大于0的整数。
在一种可能的实现方式中，所述处理模块，还用于在所述匹配层的参数不变的情况下，训练所述编码网络。
在一种可能的实现方式中，所述收发模块，还用于接收来自所述接收端的第二指示信息，所述第二指示信息指示所述发送端重新训练所述编码网络。
在一种可能的实现方式中，所述收发模块，还用于在所述匹配层未训练收敛的情况下，向所述接收端发送第三指示信息，所述第三指示信息指示所述接收端重新训练所述编码网络。
在一种可能的实现方式中，所述匹配层是可微的。
关于第三方面的各种可能的实施方式所带来的技术效果，可参考对于第一方面或第一方面的各种可能的实施方式的技术效果的介绍。
第四方面，本申请实施例提供一种通信装置，该通信装置具有实现上述第二方面方法实施例中的行为的功能。所述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。所述硬件或软件包括一个或多个与上述功能相对应的模块或单元。在一种可能的实现方式中，包括收发模块和处理模块，其中：所述收发模块，用于接收来自发送端的第一接收特征；所述第一接收特征包括所述发送端发送的第一特征经过信道传输被接收端(即第四方面中的通信装置)接收到的特征，所述第一特征为所述发送端通过匹配层对第一发送特征做编码处理得到，所述第一发送特征为所述发送端编码网络对第一数据做编码处理得到；所述编码网络和所述匹配层是独立训练得到的；所述处理模块，用于通过译码网络对所述第一接收特征做译码处理，得到所述第一数据；所述译码网络和所述匹配层是独立训练得到的。
在一种可能的实现方式中，所述收发模块，还用于向所述发送端发送第一指示信息，所述第一指示信息指示所述发送端更新所述匹配层的参数。
在一种可能的实现方式中，所述收发模块，具体用于在表征信道变化程度的参数小于或等于第一阈值的情况下，向所述发送端发送所述第一指示信息。
在一种可能的实现方式中，所述收发模块，还用于在表征信道变化程度的参数大于第一阈值的情况下，向所述发送端发送第二指示信息；所述第二指示信息指示所述发送端重新训练所述编码网络。
在一种可能的实现方式中，所述收发模块，还用于接收来自所述发送端的第三指示信息，所述第三指示信息指示所述接收端重新训练所述编码网络。
在一种可能的实现方式中，所述收发模块，还用于接收来自所述发送端的第四指示信息，所述第四指示信息指示所述编码网络完成训练。
在一种可能的实现方式中，所述收发模块，还用于向所述发送端发送第一信息，所述第一信息用于所述发送端更新所述匹配层的参数。
关于第四方面的各种可能的实施方式所带来的技术效果，可参考对于第二方面或第二方面的各种可能的实施方式的技术效果的介绍。
第五方面，本申请提供一种通信装置，该通信装置包括处理器，该处理器可以用于执行存储器所存储的计算机执行指令，以使上述第一方面或第一方面的任意可能的实现方式所示的方法被执行，或者以使上述第二方面或第二方面的任意可能的实现方式所示的方法被执行。
本申请实施例中，在执行上述方法的过程中，上述方法中有关发送信息的过程，可以理解为基于处理器的指令进行输出信息的过程。在输出信息时，处理器将信息输出给收发器，以便由收发器进行发射。该信息在由处理器输出之后，还可能需要进行其他的处理，然后到达收发器。类似的，处理器接收输入的信息时，收发器接收该信息，并将其输入处理器。更进一步的，在收发器收到该信息之后，该信息可能需要进行其他的处理，然后才输入处理器。
对于处理器所涉及的发送和/或接收等操作，如果没有特殊说明，或者，如果未与其在相关描述中的实际作用或者内在逻辑相抵触，则可以一般性的理解为基于处理器的指令输出。
在实现过程中，上述处理器可以是专门用于执行这些方法的处理器，也可以是执行存储器中的计算机指令来执行这些方法的处理器，例如通用处理器等。例如，处理器还可以用于执行存储器中存储的程序，当该程序被执行时，使得该通信装置执行如上述第一方面或第一方面的任意可能的实现方式所示的方法。在一种可能的实现方式中，存储器位于上述通信装置之外。在一种可能的实现方式中，存储器位于上述通信装置之内。
本申请实施例中，处理器和存储器还可能集成于一个器件中，即处理器和存储器还可能被集成于一起。
在一种可能的实现方式中，通信装置还包括收发器，该收发器，用于接收报文或发送报文等。
第六方面，本申请提供另一种通信装置，该通信装置包括处理电路和接口电路，该接口电路用于获取数据或输出数据；处理电路用于执行如上述第一方面或第一方面的任意可能的实现方式所示的相应的方法，或者处理电路用于执行如上述第二方面或第二方面的任意可能的实现方式所示的相应的方法。
第七方面，本申请提供一种计算机可读存储介质，该计算机可读存储介质用于存储计算机程序，当其在计算机上运行时，使得上述第一方面或第一方面的任意可能的实现方式所示的方法被执行，或者使得上述第二方面或第二方面的任意可能的实现方式所示的方法被执行。
第八方面，本申请提供一种计算机程序产品，该计算机程序产品包括计算机程序或计算机代码，当其在计算机上运行时，使得上述第一方面或第一方面的任意可能的实现方式所示的方法被执行，或者使得上述第二方面或第二方面的任意可能的实现方式所示的方法被执行。第九方面，本申请提供一种通信系统，包括上述第三方面或第三方面的任意可能的实现方式所述通信装置、上述第四方面或第四方面的任意可能的实现方式所述通信装置。
附图说明
为了更清楚地说明本申请实施例或背景技术中的技术方案，下面将对本申请实施例或背景技术中所需要使用的附图进行说明。
图1为本申请实施例提供的一种卫星通信系统的示例；
图2为本申请实施例提供的一种星间通信系统的示例；
图3为本申请实施例提供的一种无线通信系统的示例；
图4为本申请实施例提供的一种表征发送特征、匹配层、传输特征、接收特征之间的关系的关系图的示例；
图5为本申请实施例提供的一种发送端独自更新匹配层的参数的示意图；
图6A为本申请实施例提供的一种发送端的编码网络和匹配层的框架示意图；
图6B为本申请实施例发送端的编码网络和匹配层的框架示意图；
图7A为本申请实施例提供的一种训练自编码器的过程示意图；
图7B为本申请实施例提供的一种训练匹配层的过程示意图；
图8为本申请实施例提供的一种通信方法流程图；
图9为本申请实施例提供的另一种通信方法流程图；
图10为本申请实施例提供的另一种通信方法流程图；
图11为本申请实施例提供的另一种通信方法流程图；
图12为本申请实施例提供的另一种通信方法流程图；
图13为本申请实施例提供的另一种通信方法流程图；
图14为本申请实施例提供的另一种通信方法流程图；
图15为本申请实施例提供的另一种通信方法流程图；
图16为本申请实施例提供的另一种通信方法流程图；
图17为本申请实施例提供的另一种通信方法流程图；
图18为本申请实施例提供的一种通信装置的结构示意图；
图19为本申请实施例提供的另一种通信装置的结构示意图；
图20为本申请实施例提供的另一种通信装置200的结构示意图；
图21为本申请实施例提供的另一种通信装置210的结构示意图。
具体实施方式
本申请的说明书、权利要求书及附图中的术语“第一”和“第二”等仅用于区别不同对象，而不是用于描述特定顺序。此外，术语“包括”和“具有”以及它们的任何变形，意图在于覆盖不排他的包含。例如包含了一系列步骤或单元的过程、方法、系统、产品或设备等，没有限定于已列出的步骤或单元，而是可选地还包括没有列出的步骤或单元等，或可选地还包括对于这些过程、方法、产品或设备等固有的其它步骤或单元。
在本文中提及的“实施例”意味着，结合实施例描述的特定特征、结构或特性可以包含在本申请的至少一个实施例中。在说明书中的各个位置出现该短语并不一定均是指相同的实施例，也不是与其它实施例互斥的独立的或备选的实施例。本领域技术人员可以显式地和隐式地理解的是，本文所描述的实施例可以与其它实施例相结合。
本申请以下实施例中所使用的术语只是为了描述特定实施例的目的，而并非旨在作为对本申请的限制。如在本申请的说明书和所附权利要求书中所使用的那样，单数表达形式“一个”、 “一种”、“所述”、“上述”、“该”和“这一”旨在也包括复数表达形式，除非其上下文中明确地有相反指示。还应当理解，本申请中使用的术语“和/或”是指并包含一个或多个所列出项目的任何或所有可能组合。例如，“A和/或B”可以表示：只存在A，只存在B以及同时存在A和B三种情况，其中A，B可以是单数或者复数。本申请中使用的术语“多个”是指两个或两个以上。
如背景技术部分所述，目前需要研究通信开销较少的通信方案。本申请提供了通信开销较少的通信方案。本申请提供的通信方案的主要原理是发送端通过自训练更新其匹配层的参数来对信道变化做出调整，接收端的译码网络的参数固定不变。也就是说，在发送端和接收端之间的信道发生变化时，发送端通过自训练更新其匹配层的参数来对信道变化做出调整，不需要重新训练发送端的网络和接收端的网络，可减少联合训练发送端和接收端带来的通信开销，并提高训练效率。另外，由于在发送端和接收端之间的信道发生变化时，接收端的译码网络的参数固定不变，因此采用本申请提供的通信方案能够降低对接收端的处理能力的需求，延长接收端的使用时长。
以下将详细介绍本申请涉及的网络架构。
本申请提供的通信方案可应用于卫星通信等通信系统。图1为本申请实施例提供的一种卫星通信系统的示例。如图1所示，卫星通信系统包括卫星基站以及终端类型网元，例如图1中的终端设备。卫星基站为终端设备提供通信服务，终端设备可包括智能手机、智能手表、平板电脑等设备。卫星基站向终端设备传输下行数据，其中数据采用信道编码进行编码，信道编码后的数据经过星座调制后传输给终端设备；终端设备向卫星基站传输上行数据，上行数据也可以采用信道编码进行编码，编码后的数据经过星座调制后传输给卫星基站。
本申请提供的通信方案可应用于星间通信系统。星间通信系统可以分为：空间光束捕获跟踪对准(acquisition pointing and tracking，APT)子系统和通信子系统两大部分。通信子系统负责星间信息的传输，是星间通信系统的主体；APT系统负责卫星之间的捕获、对准和跟踪，确定入射信号的来波方向为捕获，调整发射波瞄准接收方向为对准，在整个通信过程中，不断调整对准和捕获为跟踪。图2为本申请实施例提供的一种星间通信系统的示例。图2所示的星间通信系统包括卫星1和卫星2，卫星1和卫星2均包括：通信模块、收发天线、APT模块和APT发射/接收天线；其中，通信模块和收发天线属于通信子系统，APT模块和APT发射/接收属于APT子系统。为了尽量减少信道中的衰减和干扰影响，同时要求具有较高的保密性和传输率，必须实时的调整APT来不断适应变化。APT系统可为光学系统。APT系统和通信子系统可为独立的系统。
本申请提供的通信方案可应用于5G、卫星通信等无线通信系统中。图3为本申请实施例提供的一种无线通信系统的示例。如图3所示，该通信系统包括：一个或多个用户设备，图1中仅以2个用户设备为例，以及可为该用户设备提供的通信服务的一个或多个接入网设备(例如基站)，图1中仅以一个接入网设备为例。
用户设备(user equipment，UE)是一种具有无线收发功能的设备。用户设备可经无线接入网(radioaccess network，RAN)中的接入网设备(或者称为接入设备)与一个或多个核心网(core network，CN)设备(或者称为核心设备)进行通信。用户设备可以部署在陆地上，包括室内或室外、手持或车载；也可以部署在水面上(如轮船等)；还可以部署在空中(例如飞机、气球和卫星上等)。本申请实施例中，UE也可称为终端设备，可以是手机(mobile phone)、移动台(mobile station，MS)、平板电脑(pad)、带无线收发功能的电脑、虚拟现实(virtual reality，VR)终端设备、增强现实(augmented reality，AR)终端设备、工业控制(industrial  control)中的无线终端设备、无人驾驶(self driving)中的无线终端设备、远程医疗(remote medical)中的无线终端设备、智能电网(smart grid)中的无线终端设备、运输安全(transportation safety)中的无线终端、智慧城市(smart city)中的无线终端设备、智慧家庭(smart home)中的无线终端等。可选的，用户设备可以是具有无线通信功能的手持设备、车载设备、可穿戴设备或物联网、车联网中的终端、5G网络以及未来网络中的任意形态的终端等，本申请对此并不限定。
接入网设备可以是任意一种具有无线收发功能且能和用户设备通信的设备，例如将用户设备接入到无线网络的无线接入网(radio access network，RAN)节点。目前，一些RAN节点的举例包括：gNB、传输接收点(transmission reception point，TRP)、演进型节点B(evolved Node B，eNB)、无线网络控制器(radio network controller，RNC)、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(base band unit，BBU)、无线保真(wireless fidelity，Wifi)接入点(access point，AP)、接入回传一体化(integrated access and backhaul，IAB)等。本申请中以基站作为接入网设备的示例进行描述。基站可包含基带单元(baseband unit，BBU)和远端射频单元(remote radio unit，RRU)。BBU和RRU可以放置在不同的地方，例如：RRU拉远，放置于高话务量的区域，BBU放置于中心机房。BBU和RRU也可以放置在同一机房。BBU和RRU也可以为一个机架下的不同部件。
需要说明的是，本申请实施例提及的无线通信系统包括但不限于：窄带物联网系统(narrow band-internet of things，NB-IoT)、全球移动通信系统(global system for mobile communications，GSM)、增强型数据速率GSM演进系统(enhanced data rate for GSM evolution，EDGE)、宽带码分多址系统(wideband code division multiple access，WCDMA)、码分多址2000系统(code division multiple access，CDMA2000)、时分同步码分多址系统(time division-synchronization code division multiple access，TD-SCDMA)、长期演进系统(long term evolution，LTE)、通用移动通信系统(universal mobile telecommunication system，UMTS)、全球互联微波接入(worldwide interoperability for microwave access，WiMAX)通信系统、第四代(4th generation，4G)通信系统、非陆地通信网络(non-terrestrial network，NTN)系统、第五代(5th generation，5G)通信系统或新无线(new radio，NR)以及未来的其他通信系统，如6G等。
本申请提供的通信方案涉及发送端的网络结构和收发端(发送端和接收端)训练交互流程。本申请实施例提供的通信方案中，发送端包括编码网络与匹配层(matched layer)，接收端包括译码网络。匹配层可称为信道匹配层或自适应信道编码模块。本申请中，发送端的编码网络和匹配层是独立训练的，通过利用编码网络和匹配层相互独立的特性，可仅更新匹配层(编码网络的参数保持不变)以解决时变衰弱信道对自编码器训练的影响。本申请涉及的网络结构特征包括：
1)、发送特征：发送特征是发送数据通过位于发送端训练完成的编码网络得到的输出，其设置为跟信道分布维度相关L维特征向量f＝[f 1,f 2,…,f L]。更具体的，可以视为将待发送数据u经过编码网络f(u；θ enc)做升维处理，得到的T维特征空间
Figure PCTCN2022139975-appb-000001
每个特征向量的维度为V，即
Figure PCTCN2022139975-appb-000002
上述信道分布维度的一种具体表现是对当前环境的信道做聚类划分所确定的最大数目T＝A×φ，其中，A表示振幅划分数目，φ表示相位划分数目，A和φ均为大于0的整数。编码网络可视为多个独立的子编码网络的堆叠，每个子编码网络对应于一种信道。若传统自编码器(对应一个子编码网络)的发送端输出维度为V的向量，则本 申请提供的发送端的编码网络输出总维度为L＝V×T的向量(发送特征)。该发送特征设计可以理解为传统的自编码器设计只针对于特定信道进行优化，而为了适应信道变化，发送端需要对所处环境内各种信道分布下的场景均有预训练结果，故对应于传统自编码器的特征也可视为在其基础上增加对信道分布相关的维度，变成高一阶的张量。举例来说，传统的自编码器输出的特征是维度为V的向量，对发送端当前环境的信道做聚类划分得到10种信道，本申请提供的发送端的编码网络输出维度为V×10的向量，每种信道对应维度为V的向量。可理解，传统的自编码器输出的特征是发送数据在一种信道下的特征，本申请提供的发送端的编码网络输出的特征包括发送数据在多种不同信道下的特征。另外由于调制阶数确定，码字的总数量也可确定，针对不同的码字可以认为总共存在n个这样的不同的发送特征f (n)。
2)、接收特征：接收特征为发送数据在位于接收端训练完成的译码网络的输入，其设置为K维接收向量r＝[r 1,r 2,…,r k]。本申请实施例中的接收特征描述为与信道分布无关的固定接收特征r s。该接收特征可通过某些特定信道h s下的编码网络和译码网络联合训练得到。r s从形式上理解可以描述为在信道h s下输入接收端的译码网络的所有可能数据。发送端可将其设定为匹配层的训练目标，即希望通过调整匹配层的参数，令同一发送数据经过其他不同信道后在接收端得到的接收特征相同。
3)、匹配层的参数：匹配层的参数可表示为矩阵
Figure PCTCN2022139975-appb-000003
或者张量描述
Figure PCTCN2022139975-appb-000004
同样可由码字总数确定为n个。匹配层的输入为发送特征f，输出为K维传输特征t＝[t 1,t 2,…,t K],有
Figure PCTCN2022139975-appb-000005
σ为非线性函数，例如批量正则化操作。传输特征可理解为发送端通过所有网络(包括编码网络和匹配层)后的输出值，即发送端的整个网络前向推演后的输出。由对信道线性传输刻画可知，r＝ht+n，故新的信道影响下的接收特征可确定为r′＝h′σ(Af)+n，对匹配层的参数的调整可视为对原接收特征r 0与新信道影响下的接收特征r′之间的相关性优化，例如令两个特征之间的内积最大。
4)、发送端中的内存缓冲器(memory buff)用来记录当前一段时间内的发送特征f (n)与接收特征r (n)的值，来作为发送端独立训练的数据集。内存缓冲器可设计为先入先出队列，可以随时跟踪发送数据更新训练数据，并可从发送数据源的分布中获得的编码增益。内存缓冲器中的训练集能够很好的表达发送数据所在的高维连续空间，并且发送端可通过该训练集来确定训练收敛的早停判断。
编码网络与匹配层的特征区别如下：
输入输出维度：
编码网络的输入是发送数据，输出是信道分布相关维度向量；匹配层的输入是信道分布相关维度向量，输出是根据调制阶数对应的星座点坐标。理解上，原AE中的编码网络的输出是星座点坐标，这里的编码网络相当于是得到不同信道下的星座点坐标，通过匹配层来对应到适合当前信道下成功传输的那个坐标。
功能上：
编码网络是对输入数据的特征提取，并且能够存储针对不同信道下的特征提取策略。
匹配层是可看做是发送特征与传输特征(可由接收特征得到)之间的注意力(attention)，从输入的高维特征中关注到与信道相关的部分特征。
下面结合信道模型来描述通过更新匹配层的参数可使用信道变化的原理。
信道可近似使用公式表示：h＝h s+Δh d+δh m。其中，h s表示静态信道部分，是由收发两端的相对距离、位置、周围静态环境决定的，可认为在一定时间内该值是固定的；Δh d表 示动态信道部分，是与设备的移动、阻挡等随机事件相关，该部分可以通过一些先验信息估计得到。例如信道数据相关性做自回归模型信道预测，发送端通过感知设备来监视环境变化估计，或现有的信道状态信息(channel state information，CSI)反馈估计等方式得到。除此之外，动态信道信息也可隐含在接收特征偏移量中反馈，上述接收特征偏移量表示为新信道下的接收特征r′与自编码器训练得到的固定接收特征r s的差值Δr＝r′-r s。该部分主要通过匹配层来克服该部分的影响，从而保证系统对时变衰弱信道环境下的有效性；δh m表示失真信道部分，是不可避免的测量不确定性导致的，但可认为该部分相对于h s和Δh d较小，对信道匹配层调整处理后的性能影响不大。
在上述信道模型假设下，可将自编码器对时变衰弱信道失真的问题转为通过匹配层对Δh d影响下的收发端网络提取的特征之间匹配问题。
对于自编码器部分，即发送端的编码网络(Encoder NN)和接收端的译码网络(Decoder NN)在h s信道下的优化，和传统自编码器训练过程类似，具体步骤如下：
1)、发送端通过编码网络对发送数据做第一编码处理，输出跟信道分布维度相关的L维发送特征f＝[f 1,f 2,…,f L]。该发送特征是对环境内各种信道分布下的场景均做预训练得到的结果，并通过编码网络存储相关分布的信道信息。该发送特征在表现上相对于传统自编码器的输出特征可视为在传统自编码器的输出特征的基础上增加对信道分布相关的维度T，变成更高一阶的张量。由于通信的训练集不同于图像数据集，在调制阶数m确定下，可认为发送数据的码字类型总数量固定为n＝2 m，故可以针对每种码字设计，总共存在n个这样不同的发送特征f (n)。
2)、发送端将发送特征f输入至匹配层
Figure PCTCN2022139975-appb-000006
运算(第二编码处理)得到K维传输特征t＝[t 1,t 2,…,t K]，
Figure PCTCN2022139975-appb-000007
传输特征可由码字类型总数确定为n个，匹配层的初始值
Figure PCTCN2022139975-appb-000008
可设为随机初始化。
3)、接收端接收发送端通过信道h发送的传输特征t，得到K维接收特征r＝[r 1,r 2,…,r k],
Figure PCTCN2022139975-appb-000009
接收端将该K维接收特征r输入自编码器在接收端的译码网络得到接收端的码字信息，为书写方便，后续符号上标(n)省略，表示在特定码字下的情况。
4)、接收端通过计算均方误差
Figure PCTCN2022139975-appb-000010
或者二进制交叉熵
Figure PCTCN2022139975-appb-000011
等任意目标函数以得到损失值。其中N为训练的批量(batch)大小，
Figure PCTCN2022139975-appb-000012
分别为发送端的发送数据与接收端译码后的数据，p ic为译码数据i属于类别c的预测概率，u ic为类别符号函数(0或1)。最后再通过梯度回传来训练收发端网络的参数，由于匹配层都是可微函数，所以不影响梯度回传。在完成特定信道h s的收发端网络训练后，接收端的译码网络参数可以固定，从而确保接收特征r s不随信道变化，而通过训练发送端的编码网络来实现对不同信道的预训练效果。
下面结合表征发送特征、匹配层、传输特征、接收特征之间的关系的附图来介绍匹配层的作用。
图4为本申请实施例提供的一种表征发送特征、匹配层、传输特征、接收特征之间的关系的关系图的示例。如图4所示，通过训练发送端的编码网络和接收端的译码网络，可以得到发送端的编码网络的发送特征与接收端的接收特征之间的映射关系有
Figure PCTCN2022139975-appb-000013
而在新的信道环境h′下，通过调整匹配层的参数将发送特征f映射到新的传输特征t′上，从而使得有
Figure PCTCN2022139975-appb-000014
这样对于新信道环境下，接收端可以对固定的接收特征完成正确的译码 操作。A′表示调整后的匹配层参数。
Figure PCTCN2022139975-appb-000015
表示复合运算，此处可将经过信道或者网络的操作理解成函数运算，故对发送特征f先作用A’再作用h’，
Figure PCTCN2022139975-appb-000016
可表示为h′(A′(f))。可理解，匹配层的作用就是在新的信道环境h′下，通过调整匹配层的参数将发送特征f映射到新的传输特征t′上，从而使得有
Figure PCTCN2022139975-appb-000017
或者说，通过调整匹配层的参数，以便发送端在不同信道下发送相同数据时，接收端可接收到相同的接收特征。
在一些实施例中，若发送端需要接收端反馈信道信息，该信道信息不局限与采用常用的信道参数，例如信号强度，CSI，可将接收特征的偏移量作为反馈值。
由于在AE训练过程中，接收端训练得到固定的接收特征r s,所以不同信道信息可以隐含在接收特征偏移量中。一种反映接收特征偏移量和信道信道之间的关系的公式如下：
Figure PCTCN2022139975-appb-000018
同样，也可将上述的目标函数定义为接收特征偏移量，具体表达式见上。可见此时的信道信息同样隐含在该接收特征偏移量中。但相对而言，对匹配层的参数的训练变为强化学习的方式，将接收特征偏移量作为奖励。相对于反馈完整的信道信息或者特征向量，该方案下接收特征偏移量为标量，可以有更少的值。
下面结合附图介绍本申请实施例提供的一种发送端独自更新匹配层的参数的流程。图5为本申请实施例提供的一种发送端独自更新匹配层的参数的示意图。如图5所示，内存缓冲器中记录有训练集D，该训练集D包含发送特征(例如f)和接收特征(例如r)；发送端通过在该训练集中采样接收特征来获得接收端的目标接收特征r；发送端利用编码网络对与该目标接收特征r对应的发送数据做第一编码处理得到发送特征f(对应于目标接收特征r)，利用匹配层对发送特征f做第二编码处理得到传输特征t′，计算传输特征t′经目标信道(当前信道)传输后的接收特征r′；计算目标接收特征r和接收特征r′的误差，并利用计算的误差(例如梯度信息)来更新匹配层的参数。图5中指向匹配层的箭头表示更新匹配层的过程。图5的虚线框中示出了通过记录发送特征f与目标接收特征r得到内存缓存器训练集的示例。在该示例中，发送端通过编码网络对发送数据做编码处理得到发送特征f与传输特征t，目标接收特征r为该传输特征t经过原始信道传输被接收端接收的接收特征。也就是说，发送端收集原始信道下的发送特征f与接收端接收到目标接收特征r并记录到内存缓冲器中；发送端在目标信道下发送传输特征t′，接收端接收到接收特征r′。发送端通过更新匹配层的参数来调整输出的传输特征，以便使接收端接收到目标接收特征r。
本申请实施例提供了一些发送端独自更新匹配层的参数的实现方式。
方式一、发送端基于特征匹配的优化问题对其匹配层进行更新，一种可能的方式如下：
发送端通过内存缓冲器，可以在发送端记录在信道h s下得到的发送特征f和接收特征r s。在新的Δh d下，可得h′＝h s+Δh d，故新时刻t下的接收特征为r′＝h′σ(A 0f)+n，则特征匹配等价于求解优化问题的优化偏移目标
Figure PCTCN2022139975-appb-000019
D为距离度量，具体可例如平方误差
Figure PCTCN2022139975-appb-000020
特征协方差KL散度
Figure PCTCN2022139975-appb-000021
刚度与距离混合误差
Figure PCTCN2022139975-appb-000022
等。最后发送端可通过基于有限内存的BFGS(limited-memory broyden-fletcher-goldfarb-shan-no，L-BFGS)或者随机梯度下降法等以A 0为初值迭代求解匹配层A′的参数。
方式二、发送端基于储蓄池计算(reservoir computing)对其匹配层进行更新，一种可能的方式如下：
由于发送端在更新匹配层的参数时，发送端的编码网络的参数是固定不动的，故可将发 送端的编码网络(encoder NN)看做储蓄池结构(reservoir)，其通常满足非线性，并且能够通过循环连接单元来存储信息，例如回声状态网络(echo state network,ESN)。因此可以理解储蓄池结构下的编码网络能够存储不同信道的信息，输入不同信道下的L维发送特征。图6A为本申请实施例提供的一种发送端的编码网络和匹配层的框架示意图。图6A中，u (1)表示发送数据，f (1)表示发送特征，矩形框f 1、f 2、…、f L表示编码网络在不同信道下的L维发送特征，A表示匹配层的参数，a 11表示匹配层中的一个参数，t表示传输特征。
发送端的匹配层可看做是读出(readout)操作，发送端通过对储蓄池的输出(即发送特征f)进行线性变换操作，可将与当前信道信息相匹配的部分特征按权重A out读取得到传输特征t。
发送端通过对接收特征r s的逆运算，可以得到目标的传输特征t *，即发送端的匹配层需要输出的传输特征。故问题可变为求解
Figure PCTCN2022139975-appb-000023
对接收特征r s的逆运算可以是利用r s＝h′t+n，逆运算得到传输特征t *，h′表示新信道。
发送端利用线性回归A out＝tf T(ff T) +或者岭回归A out＝tf T(ff T+λI) -1计算，可以直接求得所需的匹配层的权重A out。‘+’表示Moore-Penrose伪逆。
方式三、发送端基于随机老虎机问题(stochastic bandits)的强化学习(reinforcement learning)对其匹配层参数进行选择，一种可能的方式如下：
发送端可将更新匹配层的参数的问题视为一个随机bandits问题，即通过将接收端反馈的偏移目标函数值作为奖励，来更快的选择最优的匹配层参数(即匹配层的参数)。
发送端的匹配层参数可预先设置为有限种不同信道分布下的线性组合作为动作，例如A 1参数表示选择对h 1信道下的发送特征f提取方式，例如A 2参数表示选择对h 2信道下的发送特征f提取方式。故理论上存在有限多个A参数设计。图6B为本申请实施例发送端的编码网络和匹配层的框架示意图。图6B中，f (1)表示发送特征，矩形框f 1、f 2、…、f L表示编码网络在不同信道下的L维发送特征，A表示匹配层的参数，A 1为A的一个可选动作，t表示传输特征，r s表示内存缓冲器中的接收特征，r’表示发送端计算得到的在当前信道下匹配层的参数为A 1时，接收端的接收特征。发送端可将奖励值设置为利用选择的匹配层参数得到的接收特征与目标接收特征的相似程度，例如上述优化偏移目标，相似程度越高，偏移误差越小则奖励值越大。发送端可通过类似对奖励信息的反馈，指导匹配层选择最优的A参数动作。
实现上可选择解决bandits问题常用的指数权重的探索-利用方法exploration-exploitation with exponential weights(EXP3)或者例如分级乐观优化方法(hierarchical optimistic optimization)的方式在
Figure PCTCN2022139975-appb-000024
(N为线性组合总个数)实现最优匹配层参数选择。在一种可能的实现方式中，发送端可分别使用其记录的每种匹配层参数(即有限多个A参数，例如A 1参数)对发送特征做第二编码处理以得到每种匹配层参数对应的接收特征；根据每种匹配层参数对应的接收特征和目标接收特征的相似程度，确定选择每种匹配层参数获得的奖励；选择获得的奖励最大的一种匹配层参数作为匹配层的参数。发送端基于随机bandits问题的强化学习对其匹配层参数进行选择的一个举例如下：发送端的匹配层参数可选择A 1、A 2、…、A T中的任一种；当发送端与接收端之间的信道变为h′时，发送端计算在信道h′下利用匹配层参数A 1、A 2、…、A T中的每种作为匹配层参数对发送特征做第二编码处理得到的接收特征；若在信道h′下利用匹配层参数A 2对发送特征做第二编码处理得到的接收特征与目标接收特征的相似度高于利用任意其他匹配层参数对发送特征做第二编码处理得到的接收特征与目标接收特征的相似度，则在信道h′下选择匹配层参数A 2对发送特征做第二编码处理。发送端基于随机bandits问题的强化学习对其匹配层参数进行选择的另一个举例如下：发送端的匹配层参 数可选择A 1、A 2、…、A T中的任一种，A 1参数表示选择对h 1信道下的发送特征f提取方式，A 2参数表示选择对h 2信道下的发送特征f提取方式，A T参数表示选择对h T信道下的发送特征f提取方式；当发送端与接收端之间的信道从h 1变为h 2时，发送端选择A 2参数作为匹配层的参数对发送特征做第二编码处理。
本申请实施例中，发送端和接收端的训练可分为两部分，一部分是训练自编码器，即训练发送端的编码网络和接收端的译码网络，一部分是训练匹配层(或者说更新匹配层的参数)。
训练自编码器一种可能的实现方式如下：固定发送端的匹配层参数A＝A 0，A 0是匹配层的随机初始值。由于匹配层中的运算可导，故可以采用任意训练自编码器的方式训练当前信道分布
Figure PCTCN2022139975-appb-000025
下的收发端的编码网络和接收端的译码网络。图7A为本申请实施例提供的一种训练自编码器的过程示意图。如图7A所示，在训练自编码器时，发送端的匹配层的参数固定不变；前馈(feedforward)过程依次为：发送端的译码网络对发送数据u做第一编码处理，输出发送特征f；发送端的匹配层(图7A中A表示匹配层的参数)对发送特征f做第二编码处理，输出传输特征t；传输特征t经批正则化处理(可选的)之后通过信道传输；接收端接收到接收特征r(即传输特征t(或批正则化处理后的传输特征t)经过信道传输被接收端接收到的特征)；接收端利用译码网络对接收特征r做译码处理以得到发送数据u *；反向传播过程包括：接收端根据发送数据u和发送数据u *，计算目标函数以得到梯度信息；接收端根据梯度信息更新其译码网络的参数，向接收端反馈梯度信息；发送端根据接收端反馈的梯度信息，更新其译码网络的参数。当发送端的编码网络和接收端的译码网络训练收敛，说明自编码器能在给定信道下达到最优性能，由此得到最优的接收特征r s。
训练匹配层一种可能的实现方式如下：固定发送端的编码网络和接收端的译码网络(即发送端的编码网络的参数和接收端的译码网络的参数均固定不变)，在前向迭代过程中,可利用当前信道结合自编码器训练得到的发送特征f，计算得到接收特征r i′。发送端从内存缓冲器获取原信道下发送特征f和该发送特征f对应的接收特征
Figure PCTCN2022139975-appb-000026
发送端迭代更新新信道下发送端的匹配层的参数，以便更新后的匹配层参数可以适应当前信道，并且接收端无需调整译码网络。更新后的匹配层参数适应当前信道是指利用更新后的匹配层参数对发送特征f做第二编码处理得到的传输特征经新信道传输得到接收特征r s。
发送端可采用本申请实施例提供的方式一、方式二、方式三种的任一种来训练匹配层。发送端获取信道信息的方式可包括：发送端由先验信息估计得到信道信息，例如根据信道数据相关性做自回归模型信道预测；发送端通过感知设备来监视环境变化估计；通过CSI反馈估计等方式得到；根据接收端反馈的接收特征偏移量，确定信道信息，上述接收特征偏移量表示为新信道下的接收特征r'与自编码器训练得到的固定接收特征r s的差值Δr＝r′-r s。图7B为本申请实施例提供的一种训练匹配层的过程示意图。如图7B所示，在训练匹配层时，固定发送端的编码网络和接收端的译码网络，发送端独自训练其匹配层。如图7B所述，编码网络输出的发送特征f和译码网络输入的接收特征r可记录于内存缓冲器，发送端根据内存缓冲器中的发送特征和接收特征来训练其匹配层。
下面结合附图介绍本申请实施例提供的通信方案。
图8为本申请实施例提供的一种通信方法流程图。如图8所示，该方法包括：
801、发送端通过编码网络对第一数据做第一编码处理，得到第一发送特征。
上述第一发送特征与上述发送端所处环境的信道分布维度相关。发送端的编码网络与接收端的译码网络可组成一个自编码器。
在一种可能的实现方式中，上述第一发送特征包括与信道分布维度相关的L维向量，上 述L为V和T的乘积，上述第一数据至少用V维向量表征，上述T为对当前环境的信道做聚类得到的信道种类，上述T为大于或等于2的整数，上述V为大于0的整数。发送端通过编码网络对任意发送数据(例如第一数据)做第一编码处理得到的发送特征对其所处环境内各种信道分布下的场景均有预训练结果。传统的自编码器设计只针对于特定信道进行优化，发送端的编码网络输出的发送特征对其所处环境内各种信道分布下的场景均有预训练结果，故发送特征对应于传统自编码器的发送特征也可视为在原基础上增加对信道分布相关的维度，变成高一阶的张量。举例来说，若传统自编码器的发送端输出的发送特征的维度为V(大于1的整数)，则本申请实施例中的发送端的编码网络输出的发送特征的维度为L＝V×T，T为对当前环境的信道做聚类得到的信道种类。
在一种可能的实现方式中，发送端的编码网络是在多种不同信道下训练得到的。编码网络可视为多个独立的子编码网络的堆叠，每个子编码网络为该子编码网络在一种特定信道下与参数固定不变的译码网络联合训练得到，任意两个子编码网络是在不同特定信道下训练得到的。每个子编码网络可视为传统自编码器中的编码网络。在该实现方式中，由于编码网络是在多种不同信道下训练得到的，因此该编码网络能够处理多种信道情况，即该编码网络可适用于多种不同的信道。当编码网络可适用于多种不同的信道时，若发送端的信道发生变化，则该发送端不必更新编码网络的参数，仅需更新匹配层的参数。上述多种不同信道可以是对上述发送端当前所处环境的信道做聚类划分得到。
发送端可以是接入网设备，也可以是用户设备。接入网设备可以是任意一种具有无线收发功能且能和用户设备通信的设备，例如将用户设备接入到无线网络的无线接入网(radio access network，RAN)节点。目前，一些RAN节点的举例包括：gNB、传输接收点(transmission reception point，TRP)、演进型节点B(evolved Node B，eNB)、无线网络控制器(radio network controller，RNC)、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(base band unit，BBU)、无线保真(wireless fidelity，Wifi)接入点(access point，AP)、接入回传一体化(integrated access and backhaul，IAB)等。本申请中以基站作为接入网设备的示例进行描述。用户设备(user equipment，UE)是一种具有无线收发功能的设备。用户设备可经无线接入网(radioaccess network，RAN)中的接入网设备(或者称为接入设备)与一个或多个核心网(core network，CN)设备(或者称为核心设备)进行通信。用户设备可以部署在陆地上，包括室内或室外、手持或车载；也可以部署在水面上(如轮船等)；还可以部署在空中(例如飞机、气球和卫星上等)。本申请实施例中，UE也可称为终端设备，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)终端设备、增强现实(augmented reality，AR)终端设备、工业控制(industrial control)中的无线终端设备、无人驾驶(self driving)中的无线终端设备、远程医疗(remote medical)中的无线终端设备、智能电网(smart grid)中的无线终端设备、运输安全(transportation safety)中的无线终端、智慧城市(smart city)中的无线终端设备、智慧家庭(smart home)中的无线终端等。可选的，用户设备可以是具有无线通信功能的手持设备、车载设备、可穿戴设备或物联网、车联网中的终端、5G网络以及未来网络中的任意形态的终端等，本申请对此并不限定。
802、发送端通过匹配层对第一发送特征做第二编码处理，得到第一特征。
上述编码网络和上述匹配层是独立训练得到的。上述第一特征的维度小于上述第一发送特征的维度。
在一种可能的实现方式中，发送端通过匹配层对第一发送特征做第二编码处理满足如下公式：
t＝A·f；(2)
其中，t表示第一特征，t＝[t 1,t 2,…,t K]，t i＝σ(∑ la l,if l)；A表示匹配层的参数，A＝[a l,i] L×K；f表示第一发送特征，f＝[f 1,f 2,…,f L]。
803、发送端向接收端发送第一特征。
上述第一特征用于上述接收端获得上述第一数据。
发送端在执行图8中的方法流程之前，发送端和接收端已完成编码网络和译码网络的训练，例如在固定发送端的匹配层的参数的情况下，完成对编码网络和译码网络的训练。
本申请实施例中，发送端通过编码网络对第一数据做第一编码处理，得到第一发送特征；通过匹配层对该第一发送特征做第二编码处理，得到第一特征。由于编码网络和匹配层是独立训练得到的，因此当信道发生变化时，仅更新发送端的匹配层就能实现对新信道的适应，可以减少因接收端的网络训练所需要的开销。另外，由于接收端不需要参与训练，可以降低对接收端处理能力的需求，延长接收端的使用时长。
图9为本申请实施例提供的另一种通信方法流程图。图9中的方法流程为图8中描述的方法的一种可能的实现方式。如图9所示，该方法流程包括：
901、发送端通过编码网络对第一数据做第一编码处理，得到第一发送特征。
步骤901可参阅步骤801。
902、发送端通过匹配层对第一发送特征做第二编码处理，得到第一特征。
步骤902可参阅步骤802。
903、发送端向接收端发送第一特征。
步骤903可参阅步骤803。接收端在执行步骤901、步骤902、步骤903之前，发送端和接收端已完成编码网络和译码网络的训练，例如在固定发送端的匹配层的参数的情况下，完成对编码网络和译码网络的训练。图9的方法流程可视为发送端和接收端利用训练好的自编码器实现数据传输的方法流程。
904、发送端接收来自接收端的第一指示信息。
上述第一指示信息指示上述发送端更新上述匹配层的参数。
905、发送端根据其当前信道，更新匹配层的参数。
步骤905一种可能的实现方式如下：发送端根据其当前信道、第三发送特征、第三接收特征，更新上述匹配层的参数。上述第三接收特征包括上述接收端在第一信道下接收上述发送端发送的第三特征得到的特征，上述第三特征包括上述发送端利用上述匹配层对上述第三发送特征做第二编码处理得到的特征。上述发送端的当前信道与上述第一信道不同。上述第三发送特征和上述第三接收特征为发送端的内存缓冲器中记录的发送特征和接收特征。例如，第三发送特征为上述第一发送特征，第三接收特征为接收端接收发送端通过第一信道发送的第一特征得到的接收特征。第一信道可为发送端发送第一特征时，发送端和接收端之间的信道。也就是说，发送端和接收端之间的信道原来为第一信道(发送第一数据时)，信道发生变化后从第一信道变为当前信道。可理解，发送端在信道发生变化之前，通过第一信道向接收端发送数据(例如第一特征)。发送端可通过内存缓冲器记录其最近一段时间内的发送特征和接收特征，例如第一发送特征和第一接收特征。第一接收特征为接收端接收发送端通过第一信道发送的第一特征得到的接收特征。
发送端可采用前面介绍的方式一、方式二、方式三中的任一种来根据其当前信道更新匹 配层的参数，这里不再详述。
在一种可能的实现方式中，上述编码网络的参数在发送端更新匹配层的参数的过程中，保持不变。在该实现方式中，仅更新匹配层的参数，由于减少了计算量，因此可提高更新匹配层的效率。
在一种可能的实现方式中，发送端在执行步骤905之前，可执行如下操作：发送端获取第一信息；上述发送端根据上述第一信息，确定上述发送端的当前信道。上述第一信息包括来自上述接收端的信道信息或者接收特征偏移信息；上述信道信息表征上述发送端的当前信道的相关信息，上述接收特征偏移信息表征第三接收特征和第四接收特征的差异，上述第三接收特征包括上述接收端在第一信道下接收上述发送端发送的第三特征得到的特征，上述第四接收特征包括上述接收端在当前信道下接收上述发送端发送的上述第三特征得到的特征。
在一种可能的实现方式中，发送端在执行步骤905之前，执行如下操作：发送端根据先验信息估计当前信道。或者，发送端通过感知设备来监视环境变化来估计当前信道。
906、发送端通过编码网络对第二数据做第一编码处理，得到第二发送特征。
步骤906可参阅步骤901。
907、发送端在匹配层训练收敛的情况下，通过更新后的匹配层对第二发送特征做第二编码处理，得到第二特征。
更新后的匹配层为训练收敛的匹配层。发送端可采用前述介绍的方式一、方式二、方式三中的任意一种来更新匹配层的参数。更新匹配层的参数可视为训练匹配层。匹配层训练收敛的情况可以是发送端在迭代更新匹配层的参数的时长小于或等于时间阈值(例如5s)时，发送端利用匹配层的参数与当前信道计算的损失值小于损失阈值。发送端利用匹配层的参数与当前信道计算的损失值可表征发送端利用匹配层的参数与当前信道计算得到的接收特征与目标接收特征之间的差异程序。目标接收特征可为发送端期望利用匹配层的参数与当前信道计算得到的接收特征，即理想接收特征。举例来说，发送端的内存缓冲器记录有第一发送特征和第一接收特征(对应于第一信道)；发送端利用匹配层对该第一发送特征做第二编码处理得到特征t；发送端计算该特征t经过当前信道的传输被接收端接收的特征，得到接收特征r′；发送端采用例如平方误差
Figure PCTCN2022139975-appb-000027
特征协方差KL散度
Figure PCTCN2022139975-appb-000028
刚度与距离混合误差
Figure PCTCN2022139975-appb-000029
计算接收特征r′和第一接收特征(作为目标接收特征)的损失值；若该损失值小于损失阈值且发送端迭代更新匹配层的参数的时长小于或等于时间阈值，则发送端可确定匹配层训练收敛。更新匹配层的参数可视为训练匹配层。匹配层训练收敛的情况可以是发送端在迭代更新匹配层的参数的次数小于预设次数(例如一万次)时，发送端利用匹配层的参数与当前信道计算的损失值小于损失阈值。或者可以将下面定义的训练信噪比(training-signalto-noise ratio,TSNR)作为评估标准：
Figure PCTCN2022139975-appb-000030
在已知信道信噪比SNR下，可以认为当
Figure PCTCN2022139975-appb-000031
时训练收敛，其中
Figure PCTCN2022139975-appb-000032
L为上述提到的损失函数，λ为大于0的常数。
908、发送端向接收端发送第二特征。
上述第二特征用于上述接收端获得上述第二数据。
本申请实施例中，发送端根据其当前信道，更新匹配层的参数；通过更新发送端的匹配层就能实现对新信道的适应，不需要更新编码网络和接收端的译码网络，可以避免因更新编码网络和接收端的译码网络造成的时间开销和信令开销。
图10为本申请实施例提供的另一种通信方法流程图。图10中的方法流程为图8中描述的方法的一种可能的实现方式。如图10所示，该方法流程包括：
1001、发送端通过编码网络对第一数据做第一编码处理，得到第一发送特征。
步骤1001可参阅步骤801。
1002、发送端通过匹配层对第一发送特征做第二编码处理，得到第一特征。
步骤1002可参阅步骤802。
1003、发送端向接收端发送第一特征。
步骤1003可参阅步骤803。接收端在执行步骤1001、步骤1002、步骤1003之前，发送端和接收端已完成编码网络和译码网络的训练，例如在固定发送端的匹配层的参数的情况下，完成对编码网络和译码网络的训练。
1004、发送端接收来自接收端的第一指示信息。
上述第一指示信息指示上述发送端更新上述匹配层的参数。步骤1004可替换为：发送端在信道发生变化且信道的变化程度小于变化阈值的情况下，确定更新匹配层的参数。信道的变化程度可以是变化后的信道与变化前的信道的协方差，变化阈值为根据实际需求设置的。发送端可根据变化后的信道与变化前的信道的协方差以及变化阈值，确定是否更新匹配层的参数。
1005、发送端根据其当前信道，更新匹配层的参数。
步骤1005可参阅步骤905。
1006、发送端在匹配层未训练收敛的情况下，向接收端发送第三指示信息。
上述第三指示信息指示上述接收端重新训练上述编码网络。
匹配层训练未收敛的情况可以是发送端在迭代更新匹配层的参数的时长大于或等于时间阈值(例如5s)时，发送端利用匹配层的参数与当前信道计算的损失值大于或等于损失阈值。匹配层训练未收敛的情况可以是发送端在迭代更新匹配层的参数的次数等于或大于预设次数(例如一万次)时，发送端利用匹配层的参数与当前信道计算的损失值等于或大于损失阈值。
1007、发送端与接收端训练其发送端的编码网络。
发送端与接收端训练其发送端的编码网络可以是：在接收端的译码网络的参数固定不变的情况下，训练发送端的编码网络。在匹配层的参数不变的情况下，发送端训练编码网络。
1008、发送端在其编码网络训练收敛的情况下，根据其当前信道，更新匹配层的参数。
匹配层训练未收敛的情况可以是发送端在迭代更新匹配层的参数的时长大于或等于时间阈值(例如5s)时，发送端利用匹配层的参数与当前信道计算的损失值大于或等于损失阈值。匹配层训练未收敛的情况可以是发送端在迭代更新匹配层的参数的次数等于或大于预设次数(例如一万次)时，发送端利用匹配层的参数与当前信道计算的损失值等于或大于损失阈值。
应理解，发送端可重复执行步骤1006至步骤1008，直到匹配层训练收敛。发送端执行步骤1008之后，若匹配层未训练收敛，则发送端执行步骤1006；若匹配层训练收敛，则发送端执行步骤1009。
1009、发送端通过编码网络对第二数据做第一编码处理，得到第二发送特征。
步骤1009可参阅步骤801。
1010、发送端通过更新后的匹配层对第二发送特征做第二编码处理，得到第二特征。
步骤1010可参阅步骤907。
1011、发送端向接收端发送第二特征。
步骤1011可参阅步骤908。
本申请实施例中，发送端在匹配层未训练收敛的情况下，向接收端发送第三指示信息，以便通过该第三指示信息来指示接收端重新训练编码网络。发送端在重新训练编码网络之后，根据其当前信道，更新匹配层的参数，以便匹配层训练收敛。
图11为本申请实施例提供的另一种通信方法流程图。图11中的方法流程为图8中描述的方法的一种可能的实现方式。如图11所示，该方法流程包括：
1101、发送端通过编码网络对第一数据做第一编码处理，得到第一发送特征。
步骤1101可参阅步骤801。
1102、发送端通过匹配层对第一发送特征做第二编码处理，得到第一特征。
步骤1102可参阅步骤802。
1103、发送端向接收端发送第一特征。
步骤1103可参阅步骤803。接收端在执行步骤1101、步骤1102、步骤1103之前，发送端和接收端已完成编码网络和译码网络的训练，例如在固定发送端的匹配层的参数的情况下，完成对编码网络和译码网络的训练。
1104、发送端接收来自接收端的第二指示信息。
第二指示信息指示上述发送端重新训练上述编码网络。步骤1104可替换为：发送端在信道发生变化且信道的变化程度大于或等于第一阈值的情况下，确定更新编码网络的参数。信道的变化程度可以是变化后的信道与变化前的信道的协方差，第一阈值为根据实际需求设置的。发送端可根据变化后的信道与变化前的信道的协方差以及第一阈值，确定是否更新编码网络的参数。
1105、发送端与接收端训练其发送端的编码网络。
发送端与接收端训练其发送端的编码网络可以是：在接收端的译码网络的参数固定不变的情况下，训练发送端的编码网络。
1106、发送端在其编码网络训练收敛的情况下，根据其当前信道，更新匹配层的参数。
步骤1106可参阅步骤1008。发送端执行步骤1106之后，若匹配层未训练收敛，则发送端可多次执行步骤1105和步骤1106直到匹配层训练收敛；若匹配层训练收敛，则发送端执行步骤1107。
1107、发送端通过编码网络对第二数据做第一编码处理，得到第二发送特征。
步骤1107可参阅步骤801。
1108、发送端通过更新后的匹配层对第二发送特征做第二编码处理，得到第二特征。
步骤1108可参阅步骤907。
1109、发送端向接收端发送第二特征。
步骤1109可参阅步骤908。
本申请实施例中，发送端接收来自接收端的第二指示信息之后，与接收端训练其发送端的编码网络。发送端在编码网络训练收敛的情况下，根据其当前信道，更新匹配层的参数；能够更快地训练好适用于当前信道的编码网络和匹配层。
图8至图11描述了在本申请提供的通信方案中，发送端执行的方法流程。下面结合附图描述在本申请提供的通信方案中，接收端执行的方法流程。
图12为本申请实施例提供的一种通信方法流程图。如图12所示，该方法包括：
1201、接收端接收来自发送端的第一接收特征。
上述第一接收特征包括上述发送端发送的第一特征经过信道传输被上述接收端接收到的特征。上述第一特征为上述发送端通过匹配层对第一发送特征做编码处理得到，上述第一发 送特征为上述发送端编码网络对第一数据做编码处理得到。上述编码网络和上述匹配层是独立训练得到的。
接收端可以是接入网设备，也可以是用户设备。
1202、接收端通过译码网络对第一接收特征做译码处理，得到第一数据。
上述译码网络和上述匹配层是独立训练得到的。接收端在执行图12中的方法流程之前，发送端和接收端已完成编码网络和译码网络的训练，例如在固定发送端的匹配层的参数的情况下，完成对编码网络和译码网络的训练。
本申请实施例中，由于译码网络和匹配层是独立训练得到的，并且编码网络和匹配层是独立训练得到的，因此当发送端与接收端之间的信道发生变化时，仅更新发送端的匹配层就能实现对新信道的适应，可以减少因接收端的网络训练所需要的开销。另外，由于接收端不需要参与训练，可以降低对接收端处理能力的需求，延长接收端的使用时长。
图13为本申请实施例提供的另一种通信方法流程图。图13中的方法流程为图12中描述的方法的一种可能的实现方式。如图13所示，该方法流程包括：
1301、接收端接收来自发送端的第一接收特征。
步骤1301可参阅步骤1201。
1302、接收端通过译码网络对第一接收特征做译码处理，得到第一数据。
步骤1302可参阅步骤1202。接收端在执行步骤1301和步骤1302之前，发送端和接收端已完成编码网络和译码网络的训练。
1303、接收端在表征信道变化程度的参数小于或等于第一阈值的情况下，向发送端发送第一指示信息。
上述第一指示信息指示上述发送端更新上述匹配层的参数。
在一种可能的实现方式中，接收端周期性检测信道变化。表征信道变化程度的参数可以是信道的协方差。举例来说，接收端每隔10ms检测一次当前信道与上一次检测的信道的协方差；若检测到当前信道与上一次检测的信道的协方差大于第二阈值且小于第一阈值，则向发送端发送第一指示信息；其中，该第二阈值小于该第一阈值，该第一阈值和该第二阈值均为根据实际需求设置的大于0的实数。
在一种可能的实现方式中，接收端还可以执行如下操作：向上述发送端发送第一信息，上述第一信息用于上述发送端更新上述匹配层的参数。上述第一信息包括来自上述接收端的信道信息或者接收特征偏移信息；上述信道信息表征上述发送端的当前信道的相关信息，上述接收特征偏移信息表征第三接收特征和第四接收特征的差异，上述第三接收特征包括上述接收端在第一信道下接收上述发送端发送的第三特征得到的特征，上述第四接收特征包括上述接收端在当前信道下接收上述发送端发送的上述第三特征得到的特征。
在一种可能的实现方式中，接收端在向发送端发送第一指示信息之前，可执行如下操作：接收端接收来自上述发送端的第四指示信息，上述第四指示信息指示上述编码网络完成训练。
本申请实施例中，接收端在表征信道变化程度的参数小于或等于第一阈值的情况下，向发送端发送第一指示信息；可以及时指示发送端更新匹配层的参数，以便在信道变化后，仍可以成功完成数据传输。
图14为本申请实施例提供的另一种通信方法流程图。图14中的方法流程为图12中描述的方法的一种可能的实现方式。如图14所示，该方法流程包括：
1401、接收端接收来自发送端的第一接收特征。
步骤1401可参阅步骤1201。
1402、接收端通过译码网络对第一接收特征做译码处理，得到第一数据。
步骤1402可参阅步骤1202。
1403、接收端在表征信道变化程度的参数大于第一阈值的情况下，向发送端发送第二指示信息。
上述第二指示信息指示上述发送端重新训练上述编码网络。接收端在执行步骤1401、步骤1402、步骤1403之前，发送端和接收端已完成编码网络和译码网络的训练，例如在固定发送端的匹配层的参数的情况下，完成对编码网络和译码网络的训练。
在一种可能的实现方式中，接收端周期性检测信道变化。表征信道变化程度的参数可以是信道的协方差。举例来说，接收端每隔10ms检测一次当前信道与上一次检测的信道的协方差；若检测到当前信道与上一次检测的信道的协方差大于第一阈值，则向发送端发送第二指示信息；其中，该第一阈值为根据实际需求设置的大于0的实数。
1404、接收端与发送端训练发送端的编码网络。
接收端与发送端训练发送端的编码网络可以是：固定译码网络的参数不变，在当前信道下，训练发送端的编码网络。
本申请实施例中，接收端在表征信道变化程度的参数大于第一阈值的情况下，向发送端发送第二指示信息，以便指示发送端重新训练编码网络；可以解决单纯更新匹配层的参数无法在新信道下成功完成数据传输的问题。
图15为本申请实施例提供的另一种通信方法流程图。图15中的方法流程为图12中描述的方法的一种可能的实现方式。如图15所示，该方法流程包括：
1501、接收端接收来自发送端的第一接收特征。
步骤1501可参阅步骤1201。
1502、接收端通过译码网络对第一接收特征做译码处理，得到第一数据。
步骤1502可参阅步骤1202。接收端在执行步骤1501、步骤1502之前，发送端和接收端已完成编码网络和译码网络的训练。
1503、接收端在表征信道变化程度的参数小于或等于第一阈值的情况下，向发送端发送第一指示信息。
步骤1503可参阅步骤1302。
1504、接收端向发送端发送第一信息。
上述第一信息用于上述发送端更新上述匹配层的参数。上述第一信息包括来自上述接收端的信道信息或者接收特征偏移信息；上述信道信息表征上述发送端的当前信道的相关信息，上述接收特征偏移信息表征第三接收特征和第四接收特征的差异，上述第三接收特征包括上述接收端在第一信道下接收上述发送端发送的第三特征得到的特征，上述第四接收特征包括上述接收端在当前信道下接收上述发送端发送的上述第三特征得到的特征。
步骤1501是可选的。发送端可根据来自接收端的第一信息更新匹配层的参数，也可以通过其他方式(不利用第一信息)更新匹配层的参数，例如发送端根据先验信息估计当前信道。
1505、接收端接收来自发送端的第三指示信息。
上述第三指示信息指示上述接收端重新训练上述编码网络。
1506、接收端与发送端训练发送端的编码网络。
步骤1506可参阅步骤1404。
本申请实施例中，接收端在表征信道变化程度的参数小于或等于第一阈值的情况下，向发送端发送第一指示信息；可以使得发送端及时更新匹配层的参数，以便保证数据被成功传 输。接收端接收来自发送端的第三指示信息之后，与发送端训练发送端的编码网络；可以解决单纯更新匹配层的参数无法在新信道下成功完成数据传输的问题。
图8至图11描述的是发送端在自编码器完成训练之后，执行的方法流程。图12至图15描述的是接收端在自编码器完成训练之后，执行的方法流程。下面结合附图介绍发送端和接收端先训练自编码器，再训练匹配层的方法流程。
图16为本申请实施例提供的另一种通信方法流程。图16中的方法描述了单一收发端(即一个发送端和一个接收端)的训练过程与信令交互。如图16所示，该方法包括：
1601、发送端初始化编码网络和匹配层。
发送端可根据信道分布维度初始化其匹配层的参数，例如对当前环境下信道聚类确定信道分布维度T，则定义匹配层A＝[a l,i] L×K中的参数初值为
Figure PCTCN2022139975-appb-000033
也可以随机初始化其匹配层的参数。发送端可采用任意方式初始化其编码网络的参数，本申请不作限定。
1602、接收端初始化译码网络。
接收端可采用任意方式初始化其译码网络的参数，本申请不作限定。步骤1601和步骤1602的先后顺序不作限定。
1603、发送端向接收端发送第一训练特征，并将传输任务保存在其内存缓冲器中。
步骤1603一种可能的实现方式如下：发送端通过编码网络对第一发送数据做第一编码处理，得到第一训练发送特征；发送端通过匹配层对第一训练发送特征做第二编码处理，得到第一训练特征；发送端向接收端发送第一训练特征；发送端将发送第一发送数据的传输任务保存在其内存缓冲器中。例如，发送端通过内存缓冲器存储第一训练特征。
1604、接收端通过译码网络对接收到的第一训练接收特征做译码处理，得到第一接收数据。
第一训练接收特征是第一训练特征经信道传输被接收端接收到的特征。
1605、接收端根据第一接收数据和第一发送数据，计算目标函数的损失值以及梯度信息。
接收端可预先存储有第一发送数据(即训练数据)。目标函数可以是均方误差
Figure PCTCN2022139975-appb-000034
或者二进制交叉熵
Figure PCTCN2022139975-appb-000035
等。其中N为训练的batch个数，
Figure PCTCN2022139975-appb-000036
分别为发送端的第一发送数据与接收端译码后的第一接收数据，p ic为译码数据i属于类别c的预测概率，u ic为类别符号函数(0或1)。
1606、接收端根据梯度信息训练译码网络。
1607、接收端向发送端反馈梯度信息。
可选的，接收端还可向发送端反馈信道信息Δh d。发送端若有能力通过感知或信道预测方式获得Δh d,则不需要接收端反馈的操作。接收端反馈信道信息的目的是让发送端能做针对信道分布维度T上的训练，从实现通过对当前信道做统计聚类分析，得到信道的大致分布，确定维度T，并对每个维度上的信道进行训练。
1608、发送端根据梯度信息训练编码网络。
发送端和接收端可重复执行步骤1603至步骤1608，直到计算得到的目标函数的损失值小于或等于损失阈值。也就是说，发送端和接收端可使用不同的训练数据在同一信道下做训练，直到接收端计算得到的目标函数的损失值小于或等于损失阈值。
1609、发送端通过内存缓冲器保存发送特征和接收特征。
内存缓冲器可采用先入先出队列来存储数据，发送端通过内存缓冲器可以随时跟踪发送特征和接收特征以更新训练数据。
1610、接收端在译码网络训练完成时，固定译码网络，继续与发送端共同训练不同信道分布下的编码网络，直至预训练完成。
可理解，若接收端在某种信道分布下训练编码网络时，计算得到的目标函数的损失值小于或等于损失阈值，则继续在下一种信道分布下训练编码网络。
步骤1601至步骤1610为发送端和接收端训练自编码器的步骤。也就是说，发送端和接收端通过执行步骤1601至步骤1610已完成自编码器在不同信道分布下的训练。
1611、接收端在信道发生变化时，根据信道变化程度确定训练方法。
步骤1611可能的实现方式如下：若表征信道变化程度的参数小于或等于第一阈值，则接收端指示发送端重新训练当前信道下的编码网络；反之，指示发送端更新匹配层的参数。
1612、接收端向发送端发送第一指示信息。
上述第一指示信息指示上述发送端更新上述匹配层的参数。步骤1612可参阅步骤1503。
1613、接收端向发送端发送第一信息。
步骤1613可参阅步骤1504。
1614、发送端根据内存缓冲器中记录的发送特征和接收特征以及当前信道，更新匹配层的参数。
发送端可采用上述介绍的方式一、方式二、方式三中的任一种，根据内存缓冲器中记录的发送特征和接收特征以及当前信道，更新匹配层的参数。举例来说，发送端可采样训练集D＝{f,r}，将r作为目标接收特征，f通过匹配层和信道得到估计的接收特征，求解偏移目标优化问题
Figure PCTCN2022139975-appb-000037
发送端可通过迭代方法得到最优的匹配层参数。
1615、发送端在匹配层训练收敛的情况下，进行数据传输。
步骤1615可替换为：发送端在匹配层训练未收敛的情况下，向上述接收端发送第三指示信息。上述第三指示信息指示上述接收端重新训练上述编码网络。
本申请实施例中，先训练自编码器，再训练匹配层。当信道发生变化时，仅需更新匹配层的参数，可以节省通信开销。
下面模拟在时变信道场景下的收发机训练，即发送端和接收端的训练。
设定调制阶数为B＝4，发送端发送的训练数据的长度为K＝256，在固定SNR下，先训练特定信道h s下的Autoencoder网络(即编码网络和译码网络)，并将其作为基准(baseline)，然后在测试数据中每隔200次传输改变信道h t＝h 0+Δh d，其中Δh d～CN(0,0.3)。
通过匹配层的迭代更新，可以使得时变fading信道下的性能达到和特定信道h s下一样的性能。
另外对比本申请提供的通信方法跟重新训练收发端的自编码器的耗时，如下表1所示，匹配层的迭代更新可以用更短时间实现对当前信道的调整。
表1
Method AE Ada_Enc
Training time 30.8688 2.8119
SER 0.00351 0.002310
在同样测试场景，时变信道每100步变化下，对比现有的接收端信道均衡操作，本方案能在发送端达到更优效果。
相比于依赖收发两端联合训练的操作，仅需要发送端进行对环境的自适应调整，减少了接收端的训练开销，降低对接收端设备处理能力的需求，延长接收端设备的使用时长。
若在同场景中存在多个接收端，若视不同接收端与发送端之间的信道状态不同，可将本申请提供的通信方案作为发送端对不同接收端的选择切换实施，同样使各接收端的网络无需频繁训练。基于发送端的神经网络的匹配层对多场景下的信道特征提取能力，由匹配层根据发送端反馈的接收特征偏移量或者信道信息例如CSI，来自适应调整，从而选择当前与之通信的接收端所对应的发送端特征，进行通信传输。
图17为本申请实施例提供的另一种通信方法流程。图17中的方法描述了多收发端(即一个发送端和多个接收端)的训练过程与信令交互。如图17所示，该方法包括：
1701、发送端初始化编码网络和匹配层。
步骤1701可参阅步骤1601。
1702、接收端初始化译码网络。
步骤1702可参阅步骤1602。
1703、发送端分别与不同接收端训练当前信道下的自编码器。
例如，发送端与接收端1训练发送端的编码网络和接收端1的译码网络，发送端与接收端2训练发送端的编码网络和接收端2的译码网络。
1704、发送端通过内存缓冲器分别记录和更新与不同接收端之间的传输任务数据。
传输任务数据可包括发送端的发送特征和接收端的接收特征。
1705、在接收端计算得到的目标函数的损失值小于或等于损失阈值时，接收端指示发送端完成自编码器的训练。
1706、发送端与接收端1完成自编码器的训练之后，与接收端1进行数据传输。
可理解，发送端与任意接收端完成自编码器的训练之后，可与该任意接收端进行数据传输。
1707、接收端1向发送端发送第一指示信息。
步骤1707可参阅步骤1503。
1708、接收端1向发送端发送第一信息。
步骤1708可参阅步骤1504。
1709、发送端根据内存缓冲器中记录的发送特征和接收特征以及当前信道，更新匹配层的参数。
1710、发送端在匹配层训练收敛的情况下，进行数据传输。
发送端通过编码网络与更新后的匹配层发送通信数据，接收端通过原先的译码网络解码数据。
本申请实施例中，先训练自编码器，再训练匹配层。当信道发生变化时，仅需更新匹配层的参数，可以节省通信开销。
图18为本申请实施例提供的一种通信装置的结构示意图。图18中的通信装置可以是前述实施例中的发送端。如图18所示，通信装置1800包括：处理模块1801和收发模块1802。
处理模块1801，用于通过编码网络对第一数据做第一编码处理，得到第一发送特征；上述第一发送特征与发送端所处环境的信道分布维度相关；
处理模块1801，还用于通过匹配层对上述第一发送特征做第二编码处理，得到第一特征；上述编码网络和上述匹配层是独立训练得到的；上述第一特征的维度小于上述第一发送特征的维度；
收发模块1802，用于向接收端发送上述第一特征；上述第一特征用于上述接收端获得上述第一数据。
在一种可能的实现方式中，处理模块1801，具体用于根据其当前信道，更新上述匹配层的参数；处理模块1801，还用于通过上述编码网络对第二数据做第一编码处理，得到第二发送特征；通过更新后的上述匹配层对上述第二发送特征做第二编码处理，得到第二特征；上述收发模块，还用于向上述接收端发送上述第二特征；上述第二特征用于上述接收端获得上述第二数据。
在一种可能的实现方式中，收发模块1802，还用于接收来自上述接收端的第一指示信息，上述第一指示信息指示上述发送端更新上述匹配层的参数。
在一种可能的实现方式中，处理模块1801，具体用于根据其当前信道、第三发送特征、第三接收特征，更新上述匹配层的参数；上述第三接收特征包括上述接收端在第一信道下接收上述发送端发送的第三特征得到的特征，上述第三特征包括上述发送端利用上述匹配层对上述第三发送特征做第二编码处理得到的特征，上述发送端的当前信道与上述第一信道不同。
在一种可能的实现方式中，处理模块1801，还用于获取第一信息；根据上述第一信息，确定上述发送端的当前信道。
在一种可能的实现方式中，处理模块1801，还用于在上述匹配层的参数不变的情况下，训练上述编码网络。
在一种可能的实现方式中，收发模块1802，还用于接收来自上述接收端的第二指示信息，上述第二指示信息指示上述发送端重新训练上述编码网络。
在一种可能的实现方式中，收发模块1802，还用于在上述匹配层未训练收敛的情况下，向上述接收端发送第三指示信息，上述第三指示信息指示上述接收端重新训练上述编码网络。
图19为本申请实施例提供的另一种通信装置的结构示意图。图19中的通信装置可以是前述实施例中的接收端。如图19所示，通信装置1900包括：收发模块1901和处理模块1902。
收发模块1901，用于接收来自发送端的第一接收特征；上述第一接收特征包括上述发送端发送的第一特征经过信道传输被接收端接收到的特征，上述第一特征为上述发送端通过匹配层对第一发送特征做编码处理得到，上述第一发送特征为上述发送端编码网络对第一数据做编码处理得到；上述编码网络和上述匹配层是独立训练得到的；
处理模块1902，用于通过译码网络对上述第一接收特征做译码处理，得到上述第一数据；上述译码网络和上述匹配层是独立训练得到的。
在一种可能的实现方式中，收发模块1901，还用于向上述发送端发送第一指示信息，上述第一指示信息指示上述发送端更新上述匹配层的参数。
在一种可能的实现方式中，收发模块1901，具体用于在表征信道变化程度的参数小于或等于第一阈值的情况下，向上述发送端发送上述第一指示信息。
在一种可能的实现方式中，收发模块1901，还用于在表征信道变化程度的参数大于第一阈值的情况下，向上述发送端发送第二指示信息；上述第二指示信息指示上述发送端重新训练上述编码网络。
在一种可能的实现方式中，收发模块1901，还用于接收来自上述发送端的第三指示信息，上述第三指示信息指示上述接收端重新训练上述编码网络。
在一种可能的实现方式中，收发模块1901，还用于接收来自上述发送端的第四指示信息，上述第四指示信息指示上述编码网络完成训练。
在一种可能的实现方式中，收发模块1901，还用于向上述发送端发送第一信息，上述第一信息用于上述发送端更新上述匹配层的参数。
图20为本申请实施例提供的另一种通信装置200的结构示意图。图20中的通信装置可 以是上述发送端。图20中的通信装置可以是上述接收端。
如图20所示。该通信装置200包括至少一个处理器2020和收发器2010。
在本申请的一些实施例中，处理器2020和收发器2010可以用于执行上述发送端执行的功能或操作等。处理器2020例如可执行如下一项或多项操作：图8中的步骤801、步骤802，图9中的步骤901、步骤902、步骤905、步骤906、步骤907,图10中的步骤1001、步骤1002、步骤1005、步骤1007、步骤1008、步骤1009、步骤1010，图11中的步骤1101、步骤1102、步骤1105、步骤1106、步骤1107、步骤1108。收发器2010可执行如下一项或多项操作：图8中的步骤803，图9中的步骤903、步骤904、步骤908，图10中的步骤1003、步骤1004、步骤1006、步骤1011，图11中的步骤1103、步骤1104、步骤1109。
在本申请的另一些实施例中，处理器2020和收发器2010可以用于执行上述接收端执行的功能或操作等。处理器2020可执行如下一项多项操作：图12中的步骤1202，图13中的步骤1302，图14中的步骤1402、步骤1404，图15中的步骤1502、步骤1506。收发器2010可执行如下一项或多项操作：图12中的步骤1201，图13中的步骤1301、步骤1303，图14中的步骤1401、步骤1403，图15中的步骤1501、步骤1503、步骤1504、步骤1505。
收发器2010用于通过传输介质和其他设备/装置进行通信。处理器2020利用收发器2010收发数据和/或信令，并用于实现上述方法实施例中的方法。处理器2020可实现处理模块1801的功能，收发器2010可实现收发模块1802的功能。
可选的，通信装置200还可以包括至少一个存储器2030，用于存储程序指令和/或数据。存储器2030和处理器2020耦合。本申请实施例中的耦合是装置、单元或模块之间的间接耦合或通信连接，可以是电性，机械或其它的形式，用于装置、单元或模块之间的信息交互。处理器2020可能和存储器2030协同操作。处理器2020可能执行存储器2030中存储的程序指令。该至少一个存储器中的至少一个可以包括于处理器中。
本申请实施例中不限定上述收发器2010、处理器2020以及存储器2030之间的具体连接介质。本申请实施例在图20中以存储器2030、处理器2020以及收发器2010之间通过总线2040连接，总线在图20中以粗线表示，其它部件之间的连接方式，仅是进行示意性说明，并不引以为限。该总线可以分为地址总线、数据总线、控制总线等。为便于表示，图20中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
在本申请实施例中，处理器可以是通用处理器、数字信号处理器、专用集成电路、现场可编程门阵列或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件，可以实现或者执行本申请实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者任何常规的处理器等。结合本申请实施例所公开的方法的步骤可以直接体现为硬件处理器执行完成，或者用处理器中的硬件及软件模块组合执行完成。
图21为本申请实施例提供的另一种通信装置210的结构示意图。如图21所示，图21所示的通信装置包括逻辑电路2101和接口2102。图18中的处理模块1801可以用逻辑电路2101实现，图18中的收发模块1802可以用接口2102实现。图19中的处理模块1902可以用逻辑电路2101实现，图19中的收发模块1901可以用接口2102实现。其中，该逻辑电路2101可以为芯片、处理电路、集成电路或片上系统(system on chip，SoC)芯片等，接口2102可以为通信接口、输入输出接口等。本申请实施例中，逻辑电路和接口还可以相互耦合。对于逻辑电路和接口的具体连接方式，本申请实施例不作限定。
在本申请的一些实施例中，该逻辑电路和接口可用于执行上述发送端执行的功能或操作等。
在本申请的另一些实施例中，该逻辑电路和接口可用于执行上述接收端执行的功能或操作等。
本申请还提供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代码，当计算机代码在计算机上运行时，使得计算机执行上述实施例的方法。
本申请还提供一种计算机程序产品，该计算机程序产品包括计算机代码或计算机程序，当该计算机代码或计算机程序在计算机上运行时，使得上述实施例中的通信方法被执行。
本申请还提供一种通信系统，包括上述接收端和上述发送端。
以上上述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以上述权利要求的保护范围为准。

Claims (40)

  1. 一种通信方法，其特征在于，包括：
    发送端通过编码网络对第一数据做第一编码处理，得到第一发送特征；所述第一发送特征与所述发送端所处环境的信道分布维度相关；
    所述发送端通过匹配层对所述第一发送特征做第二编码处理，得到第一特征；所述编码网络和所述匹配层是独立训练得到的；所述第一特征的维度小于所述第一发送特征的维度；
    所述发送端向接收端发送所述第一特征；所述第一特征用于所述接收端获得所述第一数据。
  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：
    所述发送端根据其当前信道，更新所述匹配层的参数；
    所述发送端通过所述编码网络对第二数据做第一编码处理，得到第二发送特征；
    所述发送端通过更新后的所述匹配层对所述第二发送特征做第二编码处理，得到第二特征；
    所述发送端向所述接收端发送所述第二特征；所述第二特征用于所述接收端获得所述第二数据。
  3. 根据权利要求2所述的方法，其特征在于，所述发送端更新所述匹配层的参数之前，所述方法还包括：
    所述发送端接收来自所述接收端的第一指示信息，所述第一指示信息指示所述发送端更新所述匹配层的参数。
  4. 根据权利要求2所述的方法，其特征在于，所述编码网络的参数在所述发送端更新所述匹配层的参数的过程中，保持不变。
  5. 根据权利要求2至4任一项所述的方法，其特征在于，所述发送端根据其当前信道，更新所述匹配层的参数包括：
    所述发送端根据其当前信道、第三发送特征、第三接收特征，更新所述匹配层的参数；所述第三接收特征包括所述接收端在第一信道下接收所述发送端发送的第三特征得到的特征，所述第三特征包括所述发送端利用所述匹配层对所述第三发送特征做第二编码处理得到的特征，所述发送端的当前信道与所述第一信道不同。
  6. 根据权利要求5所述的方法，其特征在于，在所述发送端根据其当前信道，更新所述匹配层的参数之前，所述方法还包括：
    所述发送端获取第一信息；
    所述发送端根据所述第一信息，确定所述发送端的当前信道。
  7. 根据权利要求6所述的方法，其特征在于，所述第一信息包括来自所述接收端的信道信息或者接收特征偏移信息；所述信道信息表征所述发送端的当前信道的相关信息，所述接收特征偏移信息表征第三接收特征和第四接收特征的差异，所述第三接收特征包括所述接收端在第一信道下接收所述发送端发送的第三特征得到的特征，所述第四接收特征包括所述接收端在当前信道下接收所述发送端发送的所述第三特征得到的特征。
  8. 根据权利要求1至7任一项所述的方法，其特征在于，所述第一发送特征包括与信道分布维度相关的L维向量，所述L为V和T的乘积，所述第一数据至少用V维向量表征，所述T为对当前环境的信道做聚类得到的信道种类，所述T为大于或等于2的整数，所述V为大于0的整数。
  9. 根据权利要求1至7任一项所述的方法，其特征在于，所述方法还包括：
    在所述匹配层的参数不变的情况下，所述发送端训练所述编码网络。
  10. 根据权利要求1至7任一项所述的方法，其特征在于，所述方法还包括：
    所述发送端接收来自所述接收端的第二指示信息，所述第二指示信息指示所述发送端重新训练所述编码网络。
  11. 根据权利要求1至7任一项所述的方法，其特征在于，所述方法还包括：
    所述发送端在所述匹配层未训练收敛的情况下，向所述接收端发送第三指示信息，所述第三指示信息指示所述接收端重新训练所述编码网络。
  12. 一种通信方法，其特征在于，包括：
    接收端接收来自发送端的第一接收特征；所述第一接收特征包括所述发送端发送的第一特征经过信道传输被所述接收端接收到的特征，所述第一特征为所述发送端通过匹配层对第一发送特征做编码处理得到，所述第一发送特征为所述发送端编码网络对第一数据做编码处理得到；所述编码网络和所述匹配层是独立训练得到的；
    所述接收端通过译码网络对所述第一接收特征做译码处理，得到所述第一数据；所述译码网络和所述匹配层是独立训练得到的。
  13. 根据权利要求12所述的方法，其特征在于，所述方法还包括：
    所述接收端向所述发送端发送第一指示信息，所述第一指示信息指示所述发送端更新所述匹配层的参数。
  14. 根据权利要求13所述的方法，其特征在于，所述接收端向所述发送端发送第一指示信息包括：
    所述接收端在表征信道变化程度的参数小于或等于第一阈值的情况下，向所述发送端发送所述第一指示信息。
  15. 根据权利要求12至14任一项所述的方法，其特征在于，所述方法还包括：
    所述接收端在表征信道变化程度的参数大于第一阈值的情况下，向所述发送端发送第二指示信息；所述第二指示信息指示所述发送端重新训练所述编码网络。
  16. 根据权利要求13或14所述的方法，其特征在于，所述接收端向所述发送端发送第一指示信息之后，所述方法还包括：
    所述接收端接收来自所述发送端的第三指示信息，所述第三指示信息指示所述接收端重新训练所述编码网络。
  17. 根据权利要求13或14所述的方法，其特征在于，所述接收端向所述发送端发送第一指示信息之前，所述方法还包括：
    所述接收端接收来自所述发送端的第四指示信息，所述第四指示信息指示所述编码网络完成训练。
  18. 根据权利要求13或14所述的方法，其特征在于，所述接收端向所述发送端发送第一指示信息之后，所述方法还包括：
    所述接收端向所述发送端发送第一信息，所述第一信息用于所述发送端更新所述匹配层的参数。
  19. 一种通信装置，其特征在于，包括：
    处理模块，用于通过编码网络对第一数据做第一编码处理，得到第一发送特征；所述第一发送特征与发送端所处环境的信道分布维度相关；
    所述处理模块，还用于通过匹配层对所述第一发送特征做第二编码处理，得到第一特征；所述编码网络和所述匹配层是独立训练得到的；所述第一特征的维度小于所述第一发送特征 的维度；
    收发模块，用于向接收端发送所述第一特征；所述第一特征用于所述接收端获得所述第一数据。
  20. 根据权利要求19所述的装置，其特征在于，
    所述处理模块，还用于根据其当前信道，更新所述匹配层的参数；通过所述编码网络对第二数据做第一编码处理，得到第二发送特征；通过更新后的所述匹配层对所述第二发送特征做第二编码处理，得到第二特征；
    所述收发模块，还用于向所述接收端发送所述第二特征；所述第二特征用于所述接收端获得所述第二数据。
  21. 根据权利要求20所述的装置，其特征在于，
    所述收发模块，还用于接收来自所述接收端的第一指示信息，所述第一指示信息指示所述发送端更新所述匹配层的参数。
  22. 根据权利要求20所述的装置，其特征在于，所述编码网络的参数在所述发送端更新所述匹配层的参数的过程中，保持不变。
  23. 根据权利要求20至22任一项所述的装置，其特征在于，
    所述处理模块，具体用于根据其当前信道、第三发送特征、第三接收特征，更新所述匹配层的参数；所述第三接收特征包括所述接收端在第一信道下接收所述发送端发送的第三特征得到的特征，所述第三特征包括所述发送端利用所述匹配层对所述第三发送特征做第二编码处理得到的特征，所述发送端的当前信道与所述第一信道不同。
  24. 根据权利要求23所述的装置，其特征在于，
    所述处理模块，还用于获取第一信息；根据所述第一信息，确定所述发送端的当前信道。
  25. 根据权利要求24所述的装置，其特征在于，所述第一信息包括来自所述接收端的信道信息或者接收特征偏移信息；所述信道信息表征所述发送端的当前信道的相关信息，所述接收特征偏移信息表征第三接收特征和第四接收特征的差异，所述第三接收特征包括所述接收端在第一信道下接收所述发送端发送的第三特征得到的特征，所述第四接收特征包括所述接收端在当前信道下接收所述发送端发送的所述第三特征得到的特征。
  26. 根据权利要求19至25任一项所述的装置，其特征在于，所述第一发送特征包括与信道分布维度相关的L维向量，所述L为V和T的乘积，所述第一数据至少用V维向量表征，所述T为对当前环境的信道做聚类得到的信道种类，所述T为大于或等于2的整数，所述V为大于0的整数。
  27. 根据权利要求19至25任一项所述的装置，其特征在于，
    所述处理模块，还用于在所述匹配层的参数不变的情况下，训练所述编码网络。
  28. 根据权利要求19至25任一项所述的装置，其特征在于，
    所述收发模块，还用于接收来自所述接收端的第二指示信息，所述第二指示信息指示所述发送端重新训练所述编码网络。
  29. 根据权利要求19至25任一项所述的装置，其特征在于，
    所述收发模块，还用于在所述匹配层未训练收敛的情况下，向所述接收端发送第三指示信息，所述第三指示信息指示所述接收端重新训练所述编码网络。
  30. 一种通信装置，其特征在于，包括：
    收发模块，用于接收来自发送端的第一接收特征；所述第一接收特征包括所述发送端发送的第一特征经过信道传输被接收端接收到的特征，所述第一特征为所述发送端通过匹配层 对第一发送特征做编码处理得到，所述第一发送特征为所述发送端编码网络对第一数据做编码处理得到；所述编码网络和所述匹配层是独立训练得到的；
    处理模块，用于通过译码网络对所述第一接收特征做译码处理，得到所述第一数据；所述译码网络和所述匹配层是独立训练得到的。
  31. 根据权利要求30所述的装置，其特征在于，
    所述收发模块，还用于向所述发送端发送第一指示信息，所述第一指示信息指示所述发送端更新所述匹配层的参数。
  32. 根据权利要求31所述的装置，其特征在于，
    所述收发模块，具体用于在表征信道变化程度的参数小于或等于第一阈值的情况下，向所述发送端发送所述第一指示信息。
  33. 根据权利要求30至32任一项所述的装置，其特征在于，
    所述收发模块，还用于在表征信道变化程度的参数大于第一阈值的情况下，向所述发送端发送第二指示信息；所述第二指示信息指示所述发送端重新训练所述编码网络。
  34. 根据权利要求31或32所述的装置，其特征在于，
    所述收发模块，还用于接收来自所述发送端的第三指示信息，所述第三指示信息指示所述接收端重新训练所述编码网络。
  35. 根据权利要求31或32所述的装置，其特征在于，
    所述收发模块，还用于接收来自所述发送端的第四指示信息，所述第四指示信息指示所述编码网络完成训练。
  36. 根据权利要求31或32所述的装置，其特征在于，
    所述收发模块，还用于向所述发送端发送第一信息，所述第一信息用于所述发送端更新所述匹配层的参数。
  37. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质中存储有计算机程序，所述计算机程序包括程序指令，所述程序指令当被处理器执行时，使所述处理器执行权利要求1至18任意一项所述的方法。
  38. 一种通信装置，其特征在于，所述通信装置包括处理电路和接口电路，所述接口电路用于获取数据或输出数据；所述处理电路用于执行如权利要求1至18任意一项所述的方法。
  39. 一种计算机程序产品，其特征在于，所述计算机程序产品包括计算机程序或计算机代码，当其在计算机上运行时，使得如权利要求1至18任意一项所述的方法被执行。
  40. 一种通信系统，其特征在于，包括如权利要求19至29任意一项所述的通信装置，以及如权利要求30至36任意一项所述的通信装置。
PCT/CN2022/139975 2021-12-22 2022-12-19 通信方法和通信装置 WO2023116618A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111583487.7 2021-12-22
CN202111583487.7A CN116367167A (zh) 2021-12-22 2021-12-22 通信方法和通信装置

Publications (1)

Publication Number Publication Date
WO2023116618A1 true WO2023116618A1 (zh) 2023-06-29

Family

ID=86901332

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/139975 WO2023116618A1 (zh) 2021-12-22 2022-12-19 通信方法和通信装置

Country Status (2)

Country Link
CN (1) CN116367167A (zh)
WO (1) WO2023116618A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109039531A (zh) * 2018-04-20 2018-12-18 电子科技大学 一种基于机器学习调整lt码编码长度的方法
CN110581732A (zh) * 2019-09-30 2019-12-17 山东建筑大学 基于神经网络的室内可见光通信多目标优化系统及方法
WO2021041551A2 (en) * 2019-08-26 2021-03-04 Board Of Regents, The University Of Texas System Autoencoder-based error correction coding for low-resolution communication
CN113379040A (zh) * 2021-07-07 2021-09-10 东南大学 基于语义编码的混合重传方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109039531A (zh) * 2018-04-20 2018-12-18 电子科技大学 一种基于机器学习调整lt码编码长度的方法
WO2021041551A2 (en) * 2019-08-26 2021-03-04 Board Of Regents, The University Of Texas System Autoencoder-based error correction coding for low-resolution communication
CN110581732A (zh) * 2019-09-30 2019-12-17 山东建筑大学 基于神经网络的室内可见光通信多目标优化系统及方法
CN113379040A (zh) * 2021-07-07 2021-09-10 东南大学 基于语义编码的混合重传方法

Also Published As

Publication number Publication date
CN116367167A (zh) 2023-06-30

Similar Documents

Publication Publication Date Title
US9258089B2 (en) System and methods for compress and forward relaying schemes in joint signal processing
US8340207B2 (en) Differential feedback scheme for closed-loop MIMO beamforming
US11329700B2 (en) Method and apparatus for downlink transmission in a cloud radio access network
WO2020088489A1 (en) Channel Prediction for Adaptive Channel State Information (CSI) Feedback Overhead Reduction
US20150146565A1 (en) Method and apparatus for downlink transmission in a cloud radio access network
CN104937858B (zh) 用于先验解码的系统和方法
US20240214329A1 (en) Data processing method and apparatus
US20220263546A1 (en) Uplink single user multiple input multiple output (su-mimo) precoding in wireless cellular systems
US11431543B2 (en) Facilitating a two-stage downlink control channel in a wireless communication system
US11411600B2 (en) Processing of uplink data streams
WO2023116618A1 (zh) 通信方法和通信装置
WO2023071683A1 (zh) 用于反馈信道状态的方法和装置
CN110785944B (zh) 下位无线基站、上位无线基站及无线基站系统
KR20190093019A (ko) 클라우드 무선 접속 네트워크의 수신 간섭 상관도에 따른 신호 송수신 방법 및 장치
CN117730330A (zh) 用于边缘学习中的联合计算和通信的在线优化
WO2022236785A1 (zh) 信道信息的反馈方法、收端设备和发端设备
US20230361820A1 (en) Downlink symbol-level precoding as part of multiuser miso communication
WO2023282804A1 (en) Classification of csi compression quality
JP6222761B2 (ja) 無線通信システム
WO2022036682A1 (en) Methods, apparatuses, and computer readable media for controlling transmit power
US20240171429A1 (en) Communication method and apparatus
WO2024103352A1 (zh) 一种通信方法、装置及系统
US20240048207A1 (en) Method and apparatus for transmitting and receiving feedback information based on artificial neural network
WO2023065150A1 (en) Channel state based beamforming enhancement
US20240013031A1 (en) Method and apparatus for artificial neural network based feedback

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: 22909931

Country of ref document: EP

Kind code of ref document: A1