Separate training method for channel compression feedback based on deep learning and related equipment

Abstract

The application provides a deep learning-based channel compression feedback separate training method and related equipment; the method comprises: jointly training an initial encoder and an initial decoder through multiple iterations to obtain a first encoder and a first decoder; the initial decoder is trained through multiple iterations for first separate training, and in each iteration, channel state information is input into the first encoder, and a hidden space vector is output, the hidden space vector is used as the input of the initial decoder in the iteration, and the channel state information is used as the target output by the initial decoder in the iteration; the initial encoder is trained through multiple iterations for second separate training, and in each iteration, the channel state information is used as the input of the initial encoder, the first decoder is used to feed back the gradient in the back propagation to the initial encoder, and the weight of the initial encoder in the iteration is adjusted by taking the minimization of the gradient as the target.

Description

Technical Field

[0001] The embodiments of this application relate to the technical field of communication, and in particular to a deep learning-based channel compression feedback separation training method and related equipment. Background Technology

[0002] In the training of related channel compression feedback, the encoder and decoder are often trained simultaneously and share information with each other, which can lead to potential model and algorithm leakage problems.

[0003] On the other hand, updating the encoder and decoder simultaneously not only leads to low system efficiency, high power consumption, and long operating latency, but also presents complex hardware and software compatibility issues when deploying jointly trained encoders and decoders on different terminals and base stations due to differences in the physical equipment attributes of terminals and base stations.

[0004] Furthermore, the encoder and decoder trained simultaneously need to be retrained when a new terminal or a new base station is added to a one-to-many or many-to-one system, meaning that their generalization performance is not good enough.

[0005] Therefore, a scheme is needed that allows the encoder and decoder to be trained independently. Summary of the Invention

[0006] In view of this, the purpose of this application is to propose a deep learning-based channel compression feedback separation training method and related equipment.

[0007] To achieve the above objectives, this application provides a deep learning-based channel compression feedback separation training method, applied to a communication device, wherein the communication device includes a terminal and a base station, the terminal includes an initial encoder, and the base station includes an initial decoder.

[0008] The method includes:

[0009] By utilizing the channel state information of the communication channel between the initial encoder and the initial decoder, the initial encoder and the initial decoder are jointly trained through multiple iterations to obtain the first encoder and the first decoder;

[0010] The initial decoder is subjected to first separation training through multiple iterations. In each iteration, the channel state information is input to the first encoder and the latent space vector is output. The latent space vector is used as the input of the initial decoder in that iteration, and the channel state information is used as the target output of the initial decoder in that iteration. After the first separation training is completed, the second decoder is obtained.

[0011] The initial encoder is subjected to a second separation training through multiple iterations. In each iteration, the channel state information is used as the input of the initial encoder, and the first decoder feeds back the gradient in the backpropagation to the initial encoder. The weights of the initial encoder in this iteration are adjusted by minimizing the gradient. After the second separation training is completed, the second encoder is obtained.

[0012] Furthermore, the communication device also includes multiple terminals, each terminal including an initial encoder;

[0013] The joint training of the initial encoder and the initial decoder through multiple iterations includes:

[0014] The initial decoder is then jointly trained with each initial encoder to obtain the first encoder for each initial encoder.

[0015] Further, the channel state information is input to the first encoder, and a latent space vector is output, including:

[0016] In each iteration of the first separation training, the channel state information is divided into multiple first channel sub-vectors of the same number according to the number of initial encoders, and each initial encoder is associated with a first channel sub-vector.

[0017] Each first channel subvector is input to the corresponding initial encoder, and the initial encoder is made to output the latent space vector corresponding to the first channel subvector in this iteration.

[0018] Further, using the latent space vector as the input to the initial decoder in this iteration, and using the channel state information as the target output of the initial decoder in this iteration, includes:

[0019] In each iteration of the first separation training, the latent space vector corresponding to each first channel sub-vector is input to the initial decoder;

[0020] Let the initial decoder output the first recovery vector corresponding to the first channel subvector in this iteration;

[0021] Determine the first reconstruction error between each first recovery vector and its corresponding first channel subvector in this iteration round;

[0022] The weights of the initial decoder are adjusted by minimizing each first reconstruction error in the current iteration, and the weighted initial decoder is used for the next iteration.

[0023] When the number of iterations reaches the preset first threshold, the first separation training is considered complete.

[0024] Furthermore, the communication device also includes multiple base stations;

[0025] The joint training of the encoder and the decoder through multiple iterations includes:

[0026] The initial encoder is then jointly trained with each initial decoder to obtain a first decoder for each initial decoder.

[0027] Furthermore, the initial encoder is subjected to a second separation training through multiple iterations. In each iteration, the channel state information is used as the input to the initial encoder, and the first decoder feeds back the gradients from backpropagation to the initial encoder. The weights of the initial encoder in that iteration are adjusted by minimizing these gradients, including:

[0028] In each iteration of the second separation training, the channel state information is divided into multiple second channel sub-vectors of the same number according to the number of initial decoders, and each initial decoder is associated with a second channel sub-vector.

[0029] Each second channel subvector is input into the initial encoder, and the encoding result for each second channel subvector in this iteration is output.

[0030] The initial encoder inputs each encoding result to the corresponding first decoder and decodes the encoding result through backpropagation;

[0031] Determine the gradient of each first decoder during backpropagation in this round, and feed each gradient back to the initial encoder;

[0032] The initial encoder is adjusted based on all gradients in this round, and the adjusted initial encoder is used for the next round of iteration.

[0033] When the number of iterations reaches the preset second threshold, the second separation training is considered complete.

[0034] Further, the initial encoder is instructed to input each encoding result into the corresponding first decoder, including:

[0035] After the initial encoder, a corresponding adaptation layer is set for each first decoder, wherein each adaptation layer includes multiple fully connected layers, and each fully connected layer is set with its own activation function;

[0036] In each iteration, the initial encoder inputs the encoding result corresponding to each second channel sub-vector into the corresponding adaptation layer;

[0037] Each adaptation layer is instructed to fit the nonlinear transformation of the corresponding first decoder to its received encoding result, thereby obtaining the fitting result of each adaptation layer in this round, and inputting it into the corresponding first decoder.

[0038] Based on the same inventive concept, this application also provides a separation training device for channel compression feedback based on deep learning, including: a joint training module, a first separation training module, and a second separation training module;

[0039] The joint training module is configured to use the channel state information of the communication channel between the initial encoder and the initial decoder to jointly train the initial encoder and the initial decoder through multiple iterations to obtain the first encoder and the first decoder.

[0040] The first separation training module is configured to perform first separation training on the initial decoder through multiple iterations, and in each iteration, input the channel state information to the first encoder and output the latent space vector, use the latent space vector as the input of the initial decoder in that iteration, use the channel state information as the target output of the initial decoder in that iteration, and obtain the second decoder after completing the first separation training.

[0041] The second separation training module is configured to perform second separation training on the initial encoder through multiple iterations. In each iteration, the channel state information is used as the input of the initial encoder, and the first decoder feeds back the gradient in the backpropagation to the initial encoder. The weights of the initial encoder in this iteration are adjusted by minimizing the gradient as the objective. After the second separation training is completed, the second encoder is obtained.

[0042] Based on the same inventive concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the deep learning-based channel compression feedback separation training method as described above.

[0043] Based on the same inventive concept, this application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions for causing the computer to execute the separation training method for channel compression feedback based on deep learning as described above.

[0044] As can be seen from the above, this application provides a separation training method and related device based on deep learning channel compression feedback. Based on the latent space vector of the encoder and the gradient of the backpropagation of the decoder, it avoids the leakage of data information between them, and utilizes the structure of the adaptation layer so that the second encoder trained separately can adapt to the mapping space of each first encoder. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a deep learning-based channel compression feedback separation training method according to an embodiment of this application;

[0047] Figure 2 This is a one-to-one application scenario diagram of an embodiment of this application;

[0048] Figure 3 This is a logic diagram of a one-to-one separation training method according to an embodiment of this application;

[0049] Figure 4 This is a diagram illustrating a many-to-one application scenario in the embodiments of this application;

[0050] Figure 5 This is a logic diagram of the many-to-one separation training method according to an embodiment of this application;

[0051] Figure 6 This is a diagram illustrating a one-to-many application scenario according to an embodiment of this application.

[0052] Figure 7 This is a logic diagram of a one-to-many separation training method according to an embodiment of this application;

[0053] Figure 8 This is a schematic diagram of the adaptation layer in an embodiment of this application;

[0054] Figure 9 This is a schematic diagram of the separation training device based on deep learning channel compression feedback according to an embodiment of this application.

[0055] Figure 10 This is a schematic diagram of the electronic device structure according to an embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0057] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0058] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0059] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0060] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0061] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0062] As described in the background section, the existing deep learning-based channel compression feedback training methods are still insufficient to meet the needs of compressing channel state information in the spatial frequency domain of communication channels.

[0063] In the process of implementing this application, the applicant discovered that the main problem with the relevant deep learning-based channel compression feedback training methods is that the encoder and decoder are often trained simultaneously and share information with each other, which can lead to potential model and algorithm leakage problems.

[0064] On the other hand, updating the encoder and decoder simultaneously not only leads to low system efficiency, high power consumption, and long operating latency, but also presents complex hardware and software compatibility issues when deploying jointly trained encoders and decoders on different terminals and base stations due to differences in the physical equipment attributes of terminals and base stations.

[0065] Furthermore, the encoder and decoder trained simultaneously need to be retrained when a new terminal or a new base station is added to a one-to-many or many-to-one system, meaning that their generalization performance is not good enough.

[0066] Based on this, one or more embodiments of this application provide a separation training method for channel compression feedback based on deep learning.

[0067] The embodiments of this application are described in detail below with reference to the accompanying drawings.

[0068] In embodiments of this application, the communication device may include a terminal and a base station; in some other scenarios, the communication device may also include, for example, a terminal and multiple base stations; in other scenarios, the communication device may also include, for example, multiple terminals and a base station.

[0069] The terminal can be a user terminal such as a mobile phone.

[0070] Furthermore, in the embodiments of this application, each terminal is provided with an encoder and a quantization layer, and each base station is provided with a decoder and a dequantization layer.

[0071] Furthermore, each terminal can obtain channel state information of the communication channel with each base station. In this application, the channel state information can also be referred to as the channel vector.

[0072] Furthermore, each terminal can use an encoder to encode the acquired channel vector, that is, compress the channel vector to reduce its dimension, and use a quantization layer to quantize the compressed channel vector.

[0073] Furthermore, after quantization, a bit sequence represented in binary can be obtained.

[0074] Furthermore, the terminal can send the bit sequence to the base station, and after the base station obtains the bit sequence, it can dequantize it using the dequantization layer.

[0075] Furthermore, the base station can use a decoder to decode the dequantized bit sequence, that is, attempt to recover the dequantized bit sequence and obtain the recovery vector.

[0076] refer to Figure 1 This application discloses a deep learning-based channel compression feedback separation training method, applicable to a communication device having a terminal and a base station, and specifically includes the following steps:

[0077] Step S101: Using the channel state information of the communication channel between the initial encoder and the initial decoder, the initial encoder and the initial decoder are jointly trained through multiple iterations to obtain the first encoder and the first decoder.

[0078] In the embodiments of this application, such as Figure 2 As shown, in a communication scenario involving a single terminal and a single base station, the terminal includes an encoder and a quantization layer, while the base station includes a decoder and a dequantization layer. Here, V represents the channel state information acquired by the terminal, i.e., the channel vector. This represents the recovery vector output by the decoder.

[0079] In this embodiment, the encoder and decoder that have not yet been trained are used as the initial encoder and initial decoder. The initial encoder and initial decoder are jointly trained first, and after joint training, the first encoder of the corresponding initial encoder and the first decoder of the corresponding initial decoder are obtained.

[0080] In specific examples, such as Figure 3 As shown, step S301, initialization, can be executed first.

[0081] Specifically, during the initialization process, the structure of the initial encoder and the structure of the initial decoder can be determined, and the quantization method of the quantization layer in the terminal and the dequantization method of the dequantization layer in the base station can be further determined.

[0082] Furthermore, the number of feedback bits when the terminal feeds back the bit sequence to the base station can be determined, and the number of quantization bits can be determined accordingly.

[0083] Furthermore, the terminal can be instructed to collect the channel vector V of the downlink channel transmitted to it by the base station.

[0084] Further, step S302 can be performed to determine the structure of the joint training.

[0085] Specifically, the input of the initial encoder, the initial encoder, the quantization layer, the dequantization layer, the initial decoder, and the output of the initial decoder can be used as the overall structure for joint training, and the above order can be used as the order of the data flow during joint training.

[0086] Furthermore, joint training can be performed through multiple iterations. Based on this, a limit can be set for the number of iterations of joint training, and the joint training is considered complete when the number of iterations of joint training reaches the limit.

[0087] Based on this, the joint training iteration can be initiated, and step S303 can be further executed to determine whether the limit for the number of iterations has been reached.

[0088] Based on the judgment of step S303, if the judgment result is negative, then step S304 can be further executed to perform joint training.

[0089] In this embodiment, in each iteration of the joint training, the channel vector V can be randomly divided into multiple different data v. b And input it into the structure of the above joint training.

[0090] Specifically, each data v can be... b Input to the initial encoder, causing the initial encoder to process each data v b After compression, the signal passes through the quantization layer and is then sent to the base station.

[0091] Furthermore, after dequantizing the received data, the dequantization layer in the base station uses the initial decoder to recover the dequantized data, obtaining the recovered vector data for joint training of this iteration.

[0092] Based on this, each data v can be determined. b With the corresponding recovery vector The error between them is called the reconstruction error.

[0093] Furthermore, in this round, the weights of the initial encoder and the initial decoder are adjusted using each reconstruction error. In this embodiment, the weights of each encoder can be regarded as parameters of that encoder, and the weights of each decoder can be regarded as parameters of that decoder.

[0094] It can be seen that in each iteration, when the channel vector V is divided into n data v b When this happens, n corresponding recovery vectors will be obtained. Then, n reconstruction errors are determined, and based on these, the weights of the initial encoder and the initial decoder can be adjusted n times.

[0095] Specifically, the weights of the initial encoder and initial decoder can be adjusted by minimizing each reconstruction error, so that the initial encoder can better address each data v. b The compression, combined with the initial decoder to recover the recovery vector The two processes are inverses of each other.

[0096] Furthermore, after completing the joint training for this iteration, the process returns to step S303 to determine whether the iteration limit has been reached. If not, step S304 is executed again, that is, the joint training is performed again in the next iteration.

[0097] In this embodiment, during each joint training, the gradients of the initial encoder and initial decoder during backpropagation can be set to 1.

[0098] Furthermore, if the judgment result of step S303 is yes, it can be considered that the joint training of the initial encoder and the initial decoder has been completed, and step S305 is further executed to determine the first encoder and the first decoder.

[0099] Specifically, the initial encoder that has completed joint training can be designated as the first encoder E, and the initial decoder that has completed joint training can be designated as the first decoder D.

[0100] Step S102: Perform first separation training on the initial decoder through multiple iterations. In each iteration, input the channel state information to the first encoder and output the latent space vector. Use the latent space vector as the input of the initial decoder in that iteration and use the channel state information as the target output of the initial decoder in that iteration. After completing the first separation training, obtain the second decoder.

[0101] In this embodiment, based on the determined first encoder, the initial decoder can be separated and further trained separately, that is, the initial decoder is subjected to first separation training to obtain the second decoder.

[0102] Specifically, such as Figure 3 As shown, based on the above step S305, step S306 can be further executed to determine the latent space vector.

[0103] The channel vector V can be input into the first encoder E after joint training. Through the inference of the first encoder E, the latent space vector L corresponding to the channel vector V can be obtained. The latent space vector L represents the spatial transformation method of the first encoder E for the channel vector V.

[0104] Based on this, the initial decoder can be trained in a supervised manner with L as input and V as output, that is, the first separation training can be performed.

[0105] Specifically, step S307 can be further performed to determine the structure of the first separation training.

[0106] In this step, the input of the initial decoder, the dequantization layer, the initial decoder, and the output of the initial decoder can be used as the overall structure of the first separation training, and the above order can be used as the order of the data stream in the first separation training.

[0107] Furthermore, the first separation training can be performed through multiple iterations. Based on this, a first threshold number of iterations for the first separation training can be set, and the first separation training is considered complete when the iterations of the first separation training reach this first threshold number.

[0108] Based on this, the first separation training iteration can be started, and step S308 can be further executed to determine whether the first number threshold has been reached.

[0109] Based on the judgment in step S308, if the judgment result is negative, then step S309 is further executed to perform the first separation training.

[0110] In this embodiment, in each iteration of the first separation training, the channel vector V can be randomly divided into multiple different first channel sub-vectors v′. b Based on this, the latent space vector L is divided into multiple different latent space data l, that is, each first channel subvector v′ b Each corresponds to a latent space data l.

[0111] Furthermore, each latent space data l can be input to the dequantization layer of the initial decoder, and after dequantization processing in the dequantization layer, each dequantized latent space data l can be input to the initial decoder for recovery, thus obtaining the first recovered data corresponding to each latent space data l.

[0112] Based on this, the relationship between each first channel subvector v′ and the given information can be determined. b The corresponding latent space data l can then be used to determine each first channel subvector v′. b The corresponding first recovery vector This allows us to determine each first channel subvector v′. b With the corresponding first recovery vector The error between them is called the first reconstruction error.

[0113] Furthermore, the weights of the initial decoder are adjusted using each first reconstruction error in this round.

[0114] It can be seen that in each iteration, when the latent space vector L is divided into n latent space data l, n corresponding first recovery vectors are obtained. Then, n first reconstruction errors are determined, and based on these, the weights of the initial decoder can be adjusted n times.

[0115] Specifically, the weights of the initial decoder can be adjusted by minimizing each of the first reconstruction errors, so that the first encoder can optimize each of the first channel sub-vectors v′. b The compression, combined with the initial decoder, recovers the first recovery vector. The two processes are inverses of each other.

[0116] Furthermore, after completing the first separation training of this iteration, the process returns to step S308 to determine whether the first iteration threshold has been reached. If not, step S309 is executed, that is, the next iteration of the first separation training is performed.

[0117] Furthermore, if the judgment result of step S308 is yes, it can be considered that the first separation training of the initial decoder has been completed, and step S310 is further executed to obtain the second decoder and apply it.

[0118] Specifically, the initial decoder that has completed the first separation training can be used as the second decoder D. * And will use the second decoder D when applying it. * Deployed at the base station, the first encoder E is deployed at the terminal, utilizing the first encoder E and the second decoder D. * This enables communication between the terminal and the base station.

[0119] Step S103: Perform second separation training on the initial encoder through multiple iterations. In each iteration, use the channel state information as the input of the initial encoder, and let the first decoder feed back the gradient in the backpropagation to the initial encoder. Adjust the weight of the initial encoder in this iteration by minimizing the gradient.

[0120] In another embodiment of this application, based on the determined first decoder, the initial encoder can be separated and further trained separately, that is, the initial encoder is subjected to second separation training to obtain the second encoder.

[0121] Specifically, such as Figure 4 As shown, based on the above step S305, step S311 can be further executed to fix the weight of the first decoder.

[0122] Specifically, the weights of the first decoder D after the joint training can be kept constant, and based on this, supervised learning can be performed on the initial encoder, that is, the second separation training can be performed.

[0123] Specifically, step S312 can be further performed to determine the structure of the second separation training.

[0124] In this step, the input of the initial encoder, the initial encoder, the quantization layer, the dequantization layer, the first decoder, and the output of the first decoder can be used as the overall structure of the second separation training, and the above order can be used as the order of the data stream in the second separation training.

[0125] Furthermore, the second separation training can be performed through multiple iterations. Based on this, a second threshold number of iterations for the second separation training can be set, and the second separation training is considered to be completed when the iterations of the second separation training reach this second threshold number.

[0126] Based on this, the second separation training iteration can be started, and step S313 can be further executed to determine whether the second threshold has been reached.

[0127] Based on the judgment in step S313, if the judgment result is negative, then step S314 is further executed to perform the second separation training.

[0128] In this embodiment, in each iteration of the second separation training, the channel vector V can be randomly divided into multiple different second channel sub-vectors v″. b .

[0129] Furthermore, each second channel sub-sub can be directed to v″ b The input is fed into the initial encoder, which compresses it and then outputs the compressed second channel sub-direction v″. b The input is fed to the quantization layer for quantization, and after quantization is complete, it is sent to the dequantization layer.

[0130] Furthermore, after dequantization, the dequantization layer can instruct the first decoder to recover each dequantized data, thereby obtaining the corresponding second channel subdirectory v″. b Their respective second recovery vectors

[0131] In this embodiment, the second recovery vector output by the first decoder can be... With the corresponding second channel subdirection v″ b Minimizing the reconstruction error between the two is the training objective for the initial encoder; however, to avoid misreconstructing the second recovery vector of the first decoder... Sending the second recovery vector directly to the initial encoder carries the risk of leaking information related to the first encoder. Alternative methods can be used to replace transmitting the second recovery vector.

[0132] Specifically, during the recovery process of the first decoder, a backpropagation operation is required. Based on this, the first decoder can determine the gradient of the backpropagation process in this iteration and feed the gradient back to the initial encoder.

[0133] In this embodiment, the gradient minimization can be used as the objective to adjust the weights of the initial encoding.

[0134] Specifically, minimizing the gradient means that when adjusting the weights of the initial encoder, the standard for adjustment is that the output of the initial encoder after adjusting the weights should minimize the gradient fed back by the first decoder.

[0135] It can be seen that in each iteration, when the channel vector V is divided into n second channel sub-vectors v″ b When this happens, n corresponding second recovery vectors will be obtained. Among them, all second recovery vectors All of these are obtained based on the recovery operation under a common gradient in this iteration.

[0136] Based on this, the weights of the initial encoder are adjusted using gradients, so that the initial decoder adjusts the weights of each second channel sub-direction v″. b The compression process, and the first decoder recovers the second recovery vector. The two processes are inverses of each other.

[0137] Furthermore, after completing the second separation training of this iteration, the process returns to step S313 to determine whether the second iteration threshold has been reached. If not, step S314 is executed, that is, the next iteration of the second separation training is performed.

[0138] Furthermore, if the judgment result of step S313 is yes, it can be considered that the second separation training of the initial encoder has been completed, and step S315 is further executed to obtain the second encoder and apply it.

[0139] Specifically, the initial encoder that has completed the second separation training can be used as the second encoder E. * The first decoder D will be deployed at the base station during application, and the second encoder E will be deployed at the base station. * Deployed on the terminal, using a two-encoder E * The communication between the terminal and the base station is achieved using decoder D.

[0140] In another embodiment of this application, the communication device may further include multiple terminals and a base station.

[0141] In this embodiment, as Figure 4 As shown, there are n terminals on one side of the terminal, and they are arranged in the following order: terminal 1... terminal i... terminal n. Each terminal is equipped with an initial encoder and a quantization layer. The base station is equipped with a dequantization layer and an initial decoder. Each terminal can communicate with the base station independently, that is, each terminal can send its own bit sequence to the base station.

[0142] Furthermore, each terminal can obtain its own channel vector when communicating with the base station, and the channel vectors are in the following order: v1…v i …v n Each of these corresponds to a specific terminal.

[0143] Based on this, when training the initial encoder of each terminal and the initial decoder of the base station, the initial encoder and initial decoder can be jointly trained first, and after the joint training is completed, the initial decoder can be separately trained. In this way, the first encoder and second decoder corresponding to each terminal can be obtained and applied to the communication between each terminal and the base station.

[0144] In this embodiment, each initial encoder and the initial decoder can be jointly trained one-to-one according to the above step S101. That is, a total of n joint trainings are performed to obtain the first encoder corresponding to each initial encoder and the first decoder corresponding to each initial decoder.

[0145] like Figure 5 As shown, based on the above one-to-one joint training, step S501 can be performed to determine multiple first encoders and first decoders.

[0146] Specifically, after each initial encoder is jointly trained one-to-one with the initial decoder, the resulting multiple first encoders are encoders with different structures, and can be represented in sequence as: E1…E i …E n .

[0147] Furthermore, the resulting first decoder can be represented as D.

[0148] In this embodiment, during joint training, since training is performed separately for each of the first encoders, the first decoder can have different parameters for each first encoder with a different structure, specifically represented as: D1…D i …D n Among them, D1 to D n They all use the same first decoder, have the same structure, and differ only in their weights.

[0149] Specifically, different weights cause each first encoder to apply different weights to each data v. i b The compression, and the recovery vector recovered by the first decoder under the corresponding weights. The two processes are inverses of each other, where data v i b This indicates that the channel vector v iAny one of the multiple data points that are divided.

[0150] Based on this, the initial decoder can be further subjected to independent first separation training.

[0151] Specifically, such as Figure 5 As shown, based on the above step S501, step S502 can be further executed to determine the latent space vector set.

[0152] Specifically, each channel vector v1 to v n The inputs are respectively fed into the first encoder E1 to E2 after joint training. n In the process, through the inference of each first encoder, the corresponding channel vectors v1 to v2 can be obtained. n The respective latent space vectors are represented as l1…l in order. i …l n , where any latent space vector l i This characterizes the corresponding first encoder E i For the corresponding channel vector v i Spatial transformation methods.

[0153] Furthermore, the various latent space vectors can be merged to obtain the latent space vector set L. s and labeled as L s =(l1…l i …l n ).

[0154] Based on this, we can use the latent space vector set L s The initial decoder is trained in a supervised manner, with the spatial vectors corresponding to each first encoder as the target output. This is also known as the first separation training.

[0155] Specifically, step S503 can be further performed to determine the structure of the first separation training in a many-to-one relationship.

[0156] In this step, the input of the initial decoder, the dequantization layer, the initial decoder, and the output of the initial decoder can be used as the overall structure of the first separation training, and the above order can be used as the order of the data stream in the first separation training.

[0157] Furthermore, the first separation training can be performed through multiple iterations. Based on this, a first threshold number of iterations for the first separation training can be set, and the first separation training is considered complete when the iterations of the first separation training reach this first threshold number.

[0158] Based on this, the first separation training iteration can be started, and step S504 can be further executed to determine whether the first threshold has been reached.

[0159] Based on the judgment in step S504, if the judgment result is negative, then step S505 is further executed to perform the first separation training of many to one.

[0160] In this embodiment, in each iteration of the first separation training, the channel vectors corresponding to each first encoder can be merged to obtain a channel vector set V. s And marked as V s =(v1…v i …v n ); and can further convert the channel vector set V s Randomly divided into multiple different first channel sub-vectors v′ b Based on this, the latent space vector set L s The data is divided into multiple different latent space data l, that is, each first channel subvector v′ b Each corresponds to a latent space data l.

[0161] Furthermore, each latent space data l can be input to the dequantization layer of the initial decoder, and after dequantization processing in the dequantization layer, each dequantized latent space data l can be input to the initial decoder for recovery, thus obtaining the first recovered data corresponding to each latent space data l.

[0162] Based on this, the relationship between each first channel subvector v′ and the given information can be determined. b The corresponding latent space data l can then be used to determine each first channel subvector v′. b The corresponding first recovery vector This allows us to determine each first channel subvector v′. b With the corresponding first recovery vector The error between them is called the first reconstruction error.

[0163] Furthermore, the weights of the initial decoder are adjusted using each first reconstruction error in this round.

[0164] It can be seen that in each iteration, when the latent space vector set L... s When the data is divided into n latent space data l, n corresponding first recovery vectors will be obtained. Then, n first reconstruction errors are determined, and based on these, the weights of the initial decoder can be adjusted n times.

[0165] Specifically, the weights of the initial decoder can be adjusted by minimizing each of the first reconstruction errors, so that the first encoder can optimize each of the first channel sub-vectors v′. b The compression, combined with the initial decoder, recovers the first recovery vector. The two processes are inverses of each other.

[0166] Furthermore, after completing the first separation training of this iteration, the process returns to step S504 to determine whether the first iteration threshold has been reached. If not, step S505 is executed, that is, the next iteration of the first separation training is performed.

[0167] Furthermore, if the judgment result of step S504 is yes, it can be considered that the first separation training of the initial decoder has been completed, and step S506 is further executed to obtain the second decoder and apply it.

[0168] Specifically, the initial decoder that has completed the first separation training can be used as the second decoder D. * And will use the second decoder D when applying it. * Deployed at the base station, each of the first encoders E1 to E n Deployed on corresponding terminals 1 to n respectively, and when any terminal i communicates with the base station, the corresponding first encoder E is used. i Second decoder D * This enables communication between the terminal and the base station.

[0169] In another embodiment of this application, the communication device may further include a terminal and multiple base stations.

[0170] In this embodiment, as Figure 6 As shown, there are n base stations on one side of the base station, and they are arranged in the following order: base station 1...base station i...base station n. Each base station is equipped with an initial decoder and a dequantization layer. The terminal is equipped with a quantization layer and an initial encoder. Each base station can communicate independently with the terminal, that is, the terminal sends the corresponding bit sequence to each base station.

[0171] Furthermore, the terminal can obtain its respective channel vector when communicating with its respective base station, and the channel vectors are in the following order: v1…v i …v n Each of these corresponds to a specific base station.

[0172] Based on this, when training the initial encoder of the terminal and the initial decoders of each base station, the initial encoder and each initial decoder can be jointly trained first. After the joint training is completed, the initial encoder can be separately trained first. Based on this, the second encoder and the first decoder corresponding to each base station can be obtained and applied to the communication between the terminal and each base station.

[0173] In this embodiment, each initial encoder can be jointly trained one-to-one with the initial decoder according to the above step S101. That is, a total of n joint trainings are performed to obtain the first encoder corresponding to the initial encoder and the first decoder corresponding to each initial decoder.

[0174] like Figure 7 As shown, based on the above one-to-one joint training, step S701 can be performed to determine the first encoder and the first decoder.

[0175] Specifically, after each initial encoder is jointly trained one-to-one with each initial decoder, the resulting multiple first decoders are structurally different decoders and can be represented in sequence as: D1…D i …D n .

[0176] Furthermore, the resulting first encoder can be represented as E.

[0177] In this embodiment, during joint training, since training is performed separately for each of the first decoders, the first encoder can have different parameters for each first decoder with a different structure, specifically represented as: E1…E i …E n Among them, E1 to E n They all use the same first encoder, have the same structure, and differ only in weights.

[0178] Specifically, the first encoder under different weights processes each data v i b The compression, and the recovery vector recovered by the corresponding first decoder. The two processes are inverses of each other, where data v i b This indicates that the channel vector v i Any one of the multiple data points that are divided.

[0179] Based on this, the initial encoder can be further subjected to independent first separation training.

[0180] Specifically, such as Figure 7 As shown, based on the above step S701, step S702 can be further executed to fix the weights of each first decoder.

[0181] Specifically, the weights of each first decoder after joint training can be kept constant, and based on this, supervised learning can be performed on the initial encoder, that is, second separation training can be performed.

[0182] Specifically, step S703 can be further performed to determine the structure of the second separation training in a one-to-many manner.

[0183] In this step, such as Figure 8 As shown, in the second separation training of the initial encoder, in order to enable the trained second encoder to be applicable to different mapping spaces of decoders with different structures, an adaptation layer can be designed for the initial encoder. The structure of the adaptation layer includes three cascaded fully connected layers, each with a different activation function.

[0184] Based on this, the input of the initial encoder, the initial encoder, the adaptation layer, the quantization layer, the dequantization layer, the first decoder, and the output of the first decoder can be taken as the overall structure of the second separation training, and the above order can be taken as the order of the data stream in the second separation training.

[0185] Furthermore, the second separation training can be performed through multiple iterations. Based on this, a second threshold number of iterations for the second separation training can be set, and the second separation training is considered to be completed when the iterations of the second separation training reach this second threshold number.

[0186] Based on this, the second separation training iteration can be started, and step S704 can be further executed to determine whether the second threshold has been reached.

[0187] Based on the judgment in step S704, if the judgment result is negative, then step S705 is further executed to perform a second separation training of one-to-many.

[0188] In this embodiment, in each iteration of the second separation training, corresponding to any one of the first decoders v i The corresponding channel vector v can be used. i Randomly divided into multiple different second channel sub-vectors v″ b .

[0189] Furthermore, each second channel sub-sub can be directed to v″ b The input is fed into the initial encoder, which compresses it and then outputs the compressed second channel subvector v″. b That is, the encoded result is input into the adaptation layer of each corresponding first decoder.

[0190] Furthermore, each adaptation layer will fit the received encoding results through a nonlinear transformation and obtain the fitting result.

[0191] Furthermore, each fitting result is input into the quantization layer for quantization, and after quantization is completed, it is sent to the dequantization layer.

[0192] Furthermore, after dequantization, the dequantization layer can make the first decoder vi To recover the data after each solution quantization, thereby obtaining the corresponding second channel subdirectory v″. b Their respective second recovery vectors

[0193] In this embodiment, the second recovery vector output by each first decoder can be... With the corresponding second channel subdirection v″ b Minimizing the reconstruction error between the two is the training objective for the initial encoder; however, to avoid misreconstructing the second recovery vector of the first decoder... Sending the second recovery vector directly to the initial encoder carries the risk of leaking information related to the first encoder. Alternative methods can be used to replace transmitting the second recovery vector.

[0194] Specifically, during the recovery process of each first decoder, a backpropagation operation is required. Based on this, the first decoder can determine the gradient of the backpropagation process in this iteration and feed the gradient back to the initial encoder.

[0195] In this embodiment, the gradient minimization can be used as the objective to adjust the weights of the initial encoding.

[0196] Specifically, minimizing the gradient means that when adjusting the weights of the initial encoder, the standard for adjustment is that the output of the initial encoder after adjusting the weights should minimize the gradient fed back by the first decoder.

[0197] It can be seen that in each iteration, when the channel vector V is divided into n second channel sub-vectors v″ b When this happens, n corresponding second recovery vectors will be obtained. Among them, all second recovery vectors All of these are obtained based on the recovery operation under a common gradient in this iteration.

[0198] Based on this, the weights of the initial encoder are adjusted using gradients, so that the initial decoder adjusts the weights of each second channel sub-direction v″. b The compression process, and the first decoder recovers the second recovery vector. The two processes are inverses of each other.

[0199] In this embodiment, the gradients of each first decoder obtained by the initial encoder can be merged to obtain the global gradient for all first decoders.

[0200] Based on this, the initial encoder can use this global gradient to adjust its own weights.

[0201] Furthermore, after completing the second separation training of this iteration, the process returns to step S704 to determine whether the second iteration threshold has been reached. If not, step S705 is executed, that is, the next iteration of the second separation training is performed.

[0202] Furthermore, if the judgment result of step S704 is yes, it can be considered that the second separation training of the initial encoder has been completed, and step S706 is further executed to obtain the second encoder and apply it.

[0203] Specifically, the initial encoder that has completed the second separation training can be used as the second encoder E. * And during application, each first decoder will be deployed at the base station, and the second encoder E will be deployed... * Deployed on the terminal.

[0204] Furthermore, after completing the second separation training for one-to-many relationships, the corresponding adaptation layers for each first decoder can be determined, and each adaptation layer can be deployed on the terminal.

[0205] Based on this, a two-encoder E can be used * The terminal communicates with each base station through each first decoder D.

[0206] As can be seen, the deep learning-based channel compression feedback separation training method of the embodiments of this application avoids the leakage of data information between each other based on the latent space vector of the encoder and the gradient of the backpropagation of the decoder, and utilizes the structure of the adaptation layer to enable the second encoder trained separately to adapt to the mapping space of each first encoder.

[0207] It should be noted that the method of the embodiments of this application can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of the embodiments of this application, and the multiple devices will interact with each other to complete the method described.

[0208] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0209] Based on the same inventive concept, and corresponding to the methods of any of the above embodiments, the embodiments of this application also provide a separation training device for channel compression feedback based on deep learning.

[0210] refer to Figure 9 The deep learning-based channel compression feedback separation training device includes: a joint training module 901, a first separation training module 902, and a second separation training module 903.

[0211] The joint training module 901 is configured to use the channel state information of the communication channel between the initial encoder and the initial decoder to perform joint training on the initial encoder and the initial decoder through multiple iterations to obtain the first encoder and the first decoder.

[0212] The first separation training module 902 is configured to perform first separation training on the initial decoder through multiple iterations, and in each iteration, input the channel state information to the first encoder and output the latent space vector, use the latent space vector as the input of the initial decoder in that iteration, use the channel state information as the target of the output of the initial decoder in that iteration, and obtain the second decoder after completing the first separation training;

[0213] The second separation training module 903 is configured to perform second separation training on the initial encoder through multiple iterations. In each iteration, the channel state information is used as the input of the initial encoder, and the first decoder feeds back the gradient in the backpropagation to the initial encoder. The weights of the initial encoder in this iteration are adjusted by minimizing the gradient as the objective. After the second separation training is completed, the second encoder is obtained.

[0214] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, in implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware.

[0215] The apparatus of the above embodiments is used to implement the corresponding deep learning-based channel compression feedback separation training method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0216] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the deep learning-based channel compression feedback separation training method as described in any of the above embodiments.

[0217] Figure 10 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0218] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0219] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this application are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0220] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0221] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0222] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0223] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this application, and not necessarily all the components shown in the figures.

[0224] The apparatus of the above embodiments is used to implement the corresponding deep learning-based channel compression feedback separation training method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0225] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the deep learning-based channel compression feedback separation training method as described in any of the above embodiments.

[0226] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0227] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the deep learning-based channel compression feedback separation training method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0228] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0229] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0230] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0231] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A separation training method based on deep learning channel compression feedback, characterized in that, Applied to a communication device, the communication device including a terminal and a base station, the terminal including an initial encoder and the base station including an initial decoder; The method includes: By utilizing the channel state information of the communication channel between the initial encoder and the initial decoder, the initial encoder and the initial decoder are jointly trained through multiple iterations to obtain the first encoder and the first decoder; The initial decoder is trained by multiple iterations. In each iteration, the channel state information is input to the first encoder and the latent space vector is output. The latent space vector is used as the input of the initial decoder in that iteration, and the channel state information is used as the target output of the initial decoder in that iteration. The second decoder is obtained after the first separation training is completed; The initial encoder is subjected to a second separation training through multiple iterations. In each iteration, the channel state information is used as the input of the initial encoder, and the first decoder feeds back the gradient in the backpropagation to the initial encoder. The weights of the initial encoder in this iteration are adjusted by minimizing the gradient. The second encoder is obtained after the second separation training is completed; The communication device also includes multiple terminals, each terminal including an initial encoder; The joint training of the initial encoder and the initial decoder through multiple iterations includes: The initial decoder is jointly trained with each initial encoder to obtain the first encoder for each initial encoder. The step of inputting the channel state information to the first encoder and outputting the latent space vector includes: In each iteration of the first separation training, the channel state information is divided into multiple first channel sub-vectors of the same number according to the number of initial encoders, and each initial encoder is associated with a first channel sub-vector. Each first channel subvector is input to the corresponding initial encoder, and the initial encoder is made to output the latent space vector corresponding to the first channel subvector in this iteration; The step of using the latent space vector as input to the initial decoder in this iteration and using the channel state information as the target output of the initial decoder in this iteration includes: In each iteration of the first separation training, the latent space vector corresponding to each first channel sub-vector is input to the initial decoder; Let the initial decoder output the first recovery vector corresponding to the first channel subvector in this iteration; Determine the first reconstruction error between each first recovery vector and its corresponding first channel subvector in this iteration round; The weights of the initial decoder are adjusted by minimizing each first reconstruction error in the current iteration, and the weighted initial decoder is used for the next iteration. When the number of iterations reaches the preset first threshold, the first separation training is considered complete. The communication device also includes multiple base stations; The joint training of the encoder and the decoder through multiple iterations includes: The initial encoder and each initial decoder are jointly trained to obtain a first decoder for each initial decoder. The second separation training of the initial encoder is performed through multiple iterations. In each iteration, the channel state information is used as the input to the initial encoder, and the first decoder feeds back the gradients from backpropagation to the initial encoder. The weights of the initial encoder in that iteration are adjusted by minimizing the gradients. In each iteration of the second separation training, the channel state information is divided into multiple second channel sub-vectors of the same number according to the number of initial decoders, and each initial decoder is associated with a channel sub-vector. Each channel subvector is input into the initial encoder, and the encoding result for each channel subvector in this iteration is output. The initial encoder inputs each encoding result to the corresponding first decoder and decodes the encoding result through backpropagation; Determine the gradient of each first decoder during backpropagation in this round, and feed each gradient back to the initial encoder; The initial encoder is adjusted based on all gradients in this round, and the adjusted initial encoder is used for the next round of iteration. When the number of iterations reaches the preset second threshold, the second separation training is considered complete.

2. The method according to claim 1, characterized in that, The step of having the initial encoder input each encoding result to the corresponding first decoder includes: After the initial encoder, a corresponding adaptation layer is set for each first decoder, wherein each adaptation layer includes multiple fully connected layers, and each fully connected layer is set with its own activation function; In each iteration, the initial encoder inputs the encoding result corresponding to each channel subvector into the corresponding adaptation layer; Each adaptation layer is instructed to fit the nonlinear transformation of the corresponding first decoder to its received encoding result, thereby obtaining the fitting result of each adaptation layer in this round, and inputting it into the corresponding first decoder.

3. A separation training device based on deep learning channel compression feedback, characterized in that, include: Joint training module, first separate training module and second separate training module; The joint training module is configured to use the channel state information of the communication channel between the initial encoder and the initial decoder to jointly train the initial encoder and the initial decoder through multiple iterations to obtain the first encoder and the first decoder. The first separation training module is configured to perform first separation training on the initial decoder through multiple iterations, and in each iteration, input the channel state information to the first encoder and output the latent space vector, use the latent space vector as the input of the initial decoder in that iteration, use the channel state information as the target of the output of the initial decoder in that iteration, and obtain the second decoder after completing the first separation training. The second separation training module is configured to perform second separation training on the initial encoder through multiple iterations. In each iteration, the channel state information is used as the input of the initial encoder, and the first decoder feeds back the gradient in the backpropagation to the initial encoder. The weights of the initial encoder in this iteration are adjusted by minimizing the gradient as the objective. After the second separation training is completed, the second encoder is obtained. The communication device also includes multiple terminals, each terminal including an initial encoder; The joint training module is configured as follows: The initial decoder is jointly trained with each initial encoder to obtain the first encoder for each initial encoder. The first separation training module is configured as follows: In each iteration of the first separation training, the channel state information is divided into multiple first channel sub-vectors of the same number according to the number of initial encoders, and each initial encoder is associated with a first channel sub-vector. Each first channel subvector is input to the corresponding initial encoder, and the initial encoder is made to output the latent space vector corresponding to the first channel subvector in this iteration; In each iteration of the first separation training, the latent space vector corresponding to each first channel sub-vector is input to the initial decoder; Let the initial decoder output the first recovery vector corresponding to the first channel subvector in this iteration; Determine the first reconstruction error between each first recovery vector and its corresponding first channel subvector in this iteration round; The weights of the initial decoder are adjusted by minimizing each first reconstruction error in the current iteration, and the weighted initial decoder is used for the next iteration. When the number of iterations reaches the preset first threshold, the first separation training is considered complete. The communication device also includes multiple base stations; The joint training module is configured as follows: The initial encoder and each initial decoder are jointly trained to obtain a first decoder for each initial decoder. The second separation training module is configured as follows: In each iteration of the second separation training, the channel state information is divided into multiple second channel sub-vectors of the same number according to the number of initial decoders, and each initial decoder is associated with a channel sub-vector. Each channel subvector is input into the initial encoder, and the encoding result for each channel subvector in this iteration is output. The initial encoder inputs each encoding result to the corresponding first decoder and decodes the encoding result through backpropagation; Determine the gradient of each first decoder during backpropagation in this round, and feed each gradient back to the initial encoder; The initial encoder is adjusted based on all gradients in this round, and the adjusted initial encoder is used for the next round of iteration. When the number of iterations reaches the preset second threshold, the second separation training is considered complete.

4. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 2.

5. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method according to any one of claims 1 to 2.

Images