Two-stage channel state estimation coding and decoding method and system

Abstract

The invention discloses a two-stage channel state estimation coding and decoding method and system. Comprising the following steps: embedding a real part and an imaginary part of CSI through a complex number input module, and keeping amplitude and phase information of a signal; and the CSI information is adapted to the VQGAN for processing through the data dimension raising module, so that the physical layer structure information of the CSI is ensured to be reserved. A coding potential space is optimized through a quantization process, and the reconstruction precision and generalization ability of the model are improved in generative adversarial training. The method has the advantages that the MIMO channel feedback information can be effectively compressed and transmitted by designing a two-stage encoding and decoding process and combining the VQGAN and the CNN, the data transmission rate is remarkably improved, the transmission loss is reduced, and the method has good adaptability in a complex wireless environment. And support is provided for efficient and reliable channel feedback of a next-generation communication system (such as 6G).

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a MIMO channel feedback information compression and transmission method and system based on a variational autoencoder generative adversarial network (VQGAN). Background Art

[0002] In the field of wireless communications, especially with the approaching sixth-generation mobile communication technology (6G), accurate acquisition and effective feedback of channel state information (CSI) has become one of the core technologies to improve the performance of wireless communication systems. CSI can provide base stations with detailed characteristics of wireless channels, thereby providing key support for communication processes such as beamforming, interference management, and resource allocation, ultimately improving signal quality and system capacity. However, with the continuous expansion of the scale of communication networks, the sharp increase in the number of users, and the explosive growth of data traffic, traditional CSI feedback methods have gradually shown their limitations and huge challenges in meeting the needs of the new generation of communication networks.

[0003] Traditional CSI feedback methods mainly rely on compressed sensing (CS) technology, which uses the sparse characteristics of the channel to reconstruct channel information through iterative algorithms to reduce feedback overhead. The advantage of compressed sensing methods is that they can effectively handle the sparsity of the channel and transmit CSI at a lower bit rate. However, these methods have limited efficiency when processing high-resolution CSI data, and have a strong dependence on the sparsity of the wireless channel, making it difficult to cope with channel feedback requirements in complex environments such as multipath effects and channel changes. In addition, when the sparsity assumption of the channel no longer holds, traditional compressed sensing technology will show lower performance.

[0004] In recent years, deep learning technology, especially the autoencoder framework, has become a research hotspot for solving the CSI feedback problem. Autoencoders can learn low-dimensional representations of CSI, breaking through the reliance of traditional compressed sensing technology on channel sparsity. Through deep learning models, the compression and reconstruction of CSI can not only achieve higher feedback efficiency, but also improve the prediction accuracy of channel state information to a certain extent.

[0005] Some deep learning-based CSI feedback methods have shown their advantages in the compression and reconstruction process. For example, a study proposed a multi-resolution CSI feedback method based on deep neural network (Source: Multi-resolution CSI Feedback with Deep Learning in Massive MIMO System). This method uses deep neural network to achieve efficient compression and feedback of CSI, reducing the feedback overhead in traditional methods. However, although this method is superior in theory, there is still a problem of insufficient prediction accuracy, especially in complex channel environments, the reconstruction accuracy and system performance still cannot meet the needs of actual applications.

[0006] In addition, the CSI compression feedback and recovery method based on the attention mechanism (Source: Attention Model for Massive MIMO CSI Compression Feedback and Recovery IEEE Conference Publication) captures the spatiotemporal correlation of the channel through the self-attention network, improving the accuracy of channel state information prediction. However, these methods are often limited by the training of the model and the amount of data, and cannot effectively adapt to the changing communication environment, and their practicality in different application scenarios is limited.

[0007] Although deep learning methods have made significant progress in the field of CSI feedback in recent years, the existing technology still faces some important problems, mainly including the following aspects:

[0008] 1. Insufficient prediction accuracy: Existing methods fail to fully utilize the complex characteristics of CSI, resulting in low channel prediction accuracy. In complex channel environments, the dynamic changes of the channel are difficult to be accurately captured by traditional methods, which in turn affects the performance and reliability of the communication system.

[0009] 2. Insufficient preservation of CSI physical layer structure information: Existing methods fail to effectively preserve the physical layer structure information of CSI, especially the real and imaginary parts, during the compression and reconstruction process. This results in incomplete channel data, affects the quality and efficiency of data transmission, and may cause signal distortion or reduce signal strength.

[0010] 3. Limited model generalization ability: The existing technology lacks an effective adversarial training mechanism when generating channel data, resulting in insufficient generalization ability of the generated model under changing channel conditions, making it difficult to adapt to application requirements in different environments. Summary of the invention

[0011] In view of the problems existing in the prior art such as insufficient channel estimation accuracy, high computational complexity, redundant feedback data, and low compression efficiency, the present invention proposes a two-stage channel state estimation coding and decoding method and system. Through deep learning technology, especially the VQGAN model, it is possible to maintain high accuracy of channel estimation while efficiently compressing channel state information. Through effective encoding, quantization, and generation of channel data, the present invention can achieve efficient channel feedback and low-loss transmission in MIMO systems, significantly improving the overall performance of the communication system.

[0012] A two-stage channel state estimation coding and decoding method comprises the following steps:

[0013] Complex input: real and imaginary parts of received MIMO channel state information (CSI);

[0014] Data dimensionality increase: The input CSI data is dimensionalized so that it can be read by the variational autoencoder generative adversarial network (VQGAN) model;

[0015] Encoding latent space: Using the convolutional neural network (CNN) in the VQGAN model as an encoder, the input image is encoded into a continuous latent space representation and mapped to a discrete latent space through a quantization process to form a codebook.

[0016] Generate adversarial data: Pass the quantized latent representation through the autoregressive Transformer model to learn the sequence probability distribution of the latent representation, and convert it back to the channel data space through the CNN decoder to generate reconstructed data similar to the input channel data;

[0017] Adversarial training: A discriminator is added to VQGAN. Through adversarial training, the generator learns to generate more diverse and realistic channel data, thereby improving the reconstruction accuracy of channel state information and the generalization ability of the model.

[0018] Furthermore, the complex input module processes the real and imaginary parts of the CSI using 1x1 point convolution to ensure that the amplitude and phase information of the signal are preserved.

[0019] Furthermore, the formula used in the encoding potential space is as follows:

[0020]

[0021] In the formula, z q is the quantized potential representation, which represents the continuous potential space The discrete representation selected from ; is the encoding result of the input signal, that is, the continuous latent space representation generated by the CNN encoder; kis the element in the codebook Z, i.e., the candidate vector in the discrete latent space; Representation calculation and z k The distance between them; h, w represent the height and width of the potential space; n z Represents the number of channels in the latent space.

[0022] Furthermore, the sequence probability distribution formula for learning potential representation is as follows:

[0023]

[0024] In the formula, p(s) represents the probability distribution of the sequence; p(s i |S <i ) means that given all the previous potential representations (i.e., s <i ), the i-th potential representation s i The conditional probability of .

[0025] Furthermore, it is converted back to the channel data space through the CNN decoder to generate the reconstructed data formula similar to the input channel data as follows:

[0026]

[0027] In the formula, is the reconstructed channel data, which represents the channel estimation result output by the generator; G(z q ) is the decoder function, which converts the quantized potential representation z q Convert back to the channel data space; q(E(x)) is to encode the input signal x into E(x), and then quantize it to obtain the quantized potential representation.

[0028] Furthermore, the adversarial loss function in the adversarial training is:

[0029]

[0030] Where D(x) is the discriminator's prediction of the real data x; Generate data for the discriminator pair prediction; E is the encoder; G is the decoder; Z is the codebook; D is the discriminator.

[0031] Furthermore, in the adversarial training, by minimizing the joint optimization goal of the reconstruction error and the adversarial loss, the final optimization goal is:

[0032]

[0033] In the formula, Q * is the optimal model parameter, indicating the optimal codec and generator model obtained through optimization; LVQ (E, G, Z) is the variational quantization loss; λL GAN ({E,G,Z},D) is the adversarial loss; λ is the weight coefficient.

[0034] The present invention also discloses a two-stage channel state estimation coding and decoding system, which can be used to implement the above two-stage channel state estimation coding and decoding method, specifically including:

[0035] A complex input module, used for receiving and processing MIMO channel state information;

[0036] Data dimension upgrading module, used for dimension upgrading of input CSI data;

[0037] Encoding latent space module, used to encode the processed CSI data into latent space;

[0038] Generate adversarial data module, which is used to generate channel state information through the autoregressive Transformer model;

[0039] The discriminator module is used to generate adversarial training and optimize the reconstruction accuracy of channel state information through adversarial learning between the generator and the discriminator.

[0040] Furthermore, the generative adversarial data module includes an autoregressive Transformer model and a decoder, which is used to convert the quantized latent representation back to the channel data space and generate reconstructed data similar to the input channel data.

[0041] The present invention also discloses a computer device, including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above two-stage channel state estimation coding and decoding method when executing the program.

[0042] The present invention also discloses a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the two-stage channel state estimation coding and decoding method is implemented.

[0043] Compared with the prior art, the advantages of the present invention are:

[0044] 1. Improve prediction accuracy: Through the design of the complex input module, the present invention can process the real and imaginary parts of CSI at the same time, ensuring that the amplitude and phase information of the signal are completely preserved. Compared with the traditional method, the 1x1 point convolution processing method can more accurately capture and retain the physical characteristics of CSI, thereby significantly improving the prediction accuracy. Especially in complex channel environments, it can better adapt to the dynamic changes of the channel and improve the prediction accuracy of the system.

[0045] 2. Effectively retain CSI physical layer structure information: The present invention uses a data dimensioning module to enable the original two-dimensional CSI channel information to be effectively processed by the VQGAN model. Data dimensioning optimizes the structure of CSI channel information, allowing VQGAN to better understand and retain the physical layer structure information of CSI, especially when processing the complex characteristics of the channel, it can better maintain the integrity of amplitude and phase. This not only enhances the expression ability of channel information, but also helps to reduce information loss in signal transmission and improve the accuracy of channel reconstruction.

[0046] 3. Optimize the coding latent space: In the present invention, the CNN encoder of VQGAN maps the input CSI data to a continuous latent space representation and converts it into a discrete latent space, i.e., a codebook, through a quantization process. This process optimizes the retention of CSI information while compressing the data, reduces the transmission requirements of redundant data, and improves the efficiency and effect of coding. This not only reduces the burden on the channel, but also speeds up the data transmission process.

[0047] 4. Enhance the generalization ability of the model: The present invention combines the VQGAN model with the adversarial training mechanism. When the generator generates channel data, it can generate more realistic channel data through adversarial learning between the discriminator and the generated data. This adversarial training mechanism not only improves the reconstruction accuracy of the model, but also significantly improves the generalization ability of the model, enabling it to maintain high performance under diverse channel conditions and adapt to different application scenarios and environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a framework diagram of a two-stage channel state estimation encoding and decoding method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0050] The present invention proposes a MIMO channel feedback information compression and transmission method based on variational autoencoder generative adversarial network (VQGAN), which mainly realizes efficient compression and transmission of MIMO channel state information through the following steps and modules.

[0051] 1. System background and signal reception

[0052] In this embodiment, a single-cell system is considered, in which a base station (BS) is equipped with N t The user equipment (UE) is equipped with N antennas. r antennas. The received signal y can be described by the following formula:

[0053] y=Ax+z

[0054] Among them, A is the channel matrix, which represents the channel transmission characteristics from the base station to the user; x is the transmission symbol, which represents the signal sent by the base station; z is the complex Gaussian noise, which represents the noise part of the received signal. In order to further process the feedback information of the MIMO channel, we use the two-dimensional discrete Fourier transform (DFT) to convert the channel state information (CSI) in the space-frequency domain into the angle-delay domain for subsequent channel estimation and compression.

[0055] 2. Complex input

[0056] The present invention processes the real and imaginary parts of CSI data by designing a complex input module. Since CSI is a complex-valued channel coefficient, it contains the amplitude and phase information of different signal paths. In order to effectively process this information, this embodiment first uses a fake complex-valued input layer to embed the real and imaginary parts into the network.

[0057] Specifically, 1x1 point convolution is used to process the real and imaginary data, which ensures that the amplitude and phase information of the signal are preserved throughout the network. In this way, the complex input module can accurately extract the physical characteristics of CSI and provide high-quality input data for subsequent channel estimation and compression.

[0058] 3. Data Dimensionality

[0059] The two-dimensional CSI channel information data extracted from the complex input module is upgraded to fit the large image model VQGAN. By upgrading the dimension, the CSI channel data is mapped from the two-dimensional space to a higher-dimensional space, which can fully utilize the image processing capabilities of VQGAN and improve the model's ability to represent channel data.

[0060] This dimensionality-enhancing module converts CSI data into a format suitable for VQGAN processing, enabling VQGAN to encode and decode channel information more accurately, thereby achieving more efficient channel estimation and feedback compression.

[0061] 4. Encoding Latent Space

[0062] The convolutional neural network (CNN) of VQGAN is used as an encoder. The dimension-upgraded input data is encoded into a continuous latent space representation by CNN. This latent representation is a compact encoding of the channel data and can capture the key characteristics of the channel.

[0063] Next, the latent space representation of the encoder output After quantization, it is mapped to a discrete latent space (i.e., codebook Z). The quantization process is done by finding the closest The vector z i To complete, the specific mathematical expression is:

[0064]

[0065] The quantized latent representation z q As input to the next step of generating modules.

[0066] 5. Generate Adversarial Data

[0067] The quantized latent representation z q is fed into an autoregressive Transformer model. The model is responsible for learning the probability distribution of the latent space representation sequence:

[0068]

[0069] Specifically, the model learns the dependencies between each latent representation and the previous one, thereby generating a sequence of channel state information.

[0070] The generated latent space representation is converted back to the channel data space by the decoder (also a CNN) to obtain the reconstructed channel data. The goal of the generator is to reconstruct a signal as similar to the original input channel data as accurately as possible. The generated channel data It can be expressed as:

[0071]

[0072] Among them, E(x) represents the encoding process of the input signal, and G is the decoder function.

[0073] 6. Adversarial Training

[0074] In order to improve the quality and authenticity of the generated channel data, the present invention introduces an adversarial training module, including a generator (Decoder) and a discriminator (Discriminator). The function of the discriminator is to distinguish the generated channel data from the real channel data and provide feedback to the generator.

[0075] During training, the generator optimizes the generated channel data by minimizing the reconstruction error (e.g., L1 or L2 loss), quantization loss, and KL divergence in the latent space. At the same time, the generator and discriminator are optimized by adversarial loss (e.g., Wasserstein GAN loss). The specific adversarial loss function is:

[0076]

[0077] Among them, D(x) is the discriminator's prediction of the real data, is the discriminator's prediction on the generated data.

[0078] Finally, by minimizing the joint optimization objective of reconstruction error and adversarial loss, the generator can learn to generate more realistic and diverse channel data. The final optimization objective is:

[0079]

[0080] Among them, L VQ is the VQ loss, and λ is the weight controlling the adversarial loss.

[0081] 7. Optimal Compression Model

[0082] After adversarial training and optimization, the present invention can obtain an optimal channel feedback compression model. The model uses VQGAN to efficiently compress and transmit MIMO channel state information, reducing feedback delay and redundant data, while maintaining high-precision channel estimation and significantly improving data transmission rate.

[0083] The present invention combines VQGAN and Generative Adversarial Network (GAN) technology to achieve efficient compression and accurate reconstruction of MIMO channel state information in a two-stage encoding and decoding process. Through this design, we effectively reduce the redundant data and delay of channel feedback, and improve the communication efficiency and channel estimation accuracy of the system. In addition, the present invention can improve the quality of generated data through adversarial training without increasing the computational complexity, and has a high application prospect.

[0084] In yet another embodiment of the present invention, a two-stage channel state estimation coding and decoding system is provided. The system can be used to implement the above two-stage channel state estimation coding and decoding method, specifically including:

[0085] A complex input module, used for receiving and processing MIMO channel state information;

[0086] Data dimension upgrading module, used for dimension upgrading of input CSI data;

[0087] Encoding latent space module, used to encode the processed CSI data into latent space;

[0088] Generate adversarial data module, which is used to generate channel state information through the autoregressive Transformer model;

[0089] The discriminator module is used to generate adversarial training and optimize the reconstruction accuracy of channel state information through adversarial learning between the generator and the discriminator.

[0090] The generative adversarial data module includes an autoregressive Transformer model and a decoder, which is used to convert the quantized latent representation back to the channel data space and generate reconstructed data similar to the input channel data.

[0091] In another embodiment of the present invention, a terminal device is provided, the terminal device includes a processor and a memory, the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, which are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding functions; the processor described in the embodiment of the present invention can be used for the operation of the two-stage channel state estimation coding and decoding method.

[0092] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understandable that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as at least one disk storage.

[0093] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the two-stage channel state estimation encoding and decoding method in the above embodiment; one or more instructions in the computer-readable storage medium are loaded and executed by the processor.

[0094] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0096] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0097] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0098] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the implementation methods of the present invention, and should be understood that the protection scope of the present invention is not limited to such special statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A two-stage channel state estimation coding and decoding method, characterized in that: The following steps are involved: Complex input: real and imaginary parts of received MIMO channel state information CSI; Data dimensionality increase: The input CSI data is processed to increase its dimension so that it can be read by the variational autoencoder generative adversarial network (VQGAN) model; Encoding latent space: Use the convolutional neural network (CNN) in the VQGAN model as an encoder to encode the input image into a continuous latent space representation and map it to a discrete latent space through a quantization process to form a codebook. Generate adversarial data: Pass the quantized latent representation through the autoregressive Transformer model to learn the sequence probability distribution of the latent representation, and convert it back to the channel data space through the CNN decoder to generate reconstructed data similar to the input channel data; Adversarial training: A discriminator is added to VQGAN. Through adversarial training, the generator learns to generate more diverse and realistic channel data, thereby improving the reconstruction accuracy of channel state information and the generalization ability of the model.

2. The two-stage channel state estimation coding and decoding method according to claim 1, characterized in that: The complex input module processes the real and imaginary parts of the CSI using 1x1 point convolution to ensure that the amplitude and phase information of the signal are preserved.

3. The two-stage channel state estimation coding and decoding method according to claim 1, characterized in that: The formula used in the encoding latent space is as follows: In the formula, z q is the quantized potential representation, which represents the continuous potential space The discrete representation selected from ; is the encoding result of the input signal, that is, the continuous latent space representation generated by the CNN encoder; k is the element in the codebook Z, i.e., the candidate vector in the discrete latent space; Representation calculation and z k The distance between them; h, w represent the height and width of the potential space; n z Represents the number of channels in the latent space.

4. The two-stage channel state estimation coding and decoding method according to claim 1, characterized in that: The sequence probability distribution formula for learning potential representation is as follows: In the formula, p(s) represents the probability distribution of the sequence; p(s i |s <i ) indicates that given all the previous potential representations, the i-th potential representation s i The conditional probability of .

5. The two-stage channel state estimation coding and decoding method according to claim 1, characterized in that: The CNN decoder converts it back to the channel data space and generates the reconstructed data formula similar to the input channel data as follows: In the formula, is the reconstructed channel data, which represents the channel estimation result output by the generator; G(z q ) is the decoder function, which converts the quantized potential representation z q Convert back to the channel data space; q(E(x)) is to encode the input signal x into E(x), and then quantize it to obtain the quantized potential representation.

6. The two-stage channel state estimation coding and decoding method according to claim 1, characterized in that: The adversarial loss function in the adversarial training is: Where D(x) is the discriminator's prediction of the real data x; Generate data for the discriminator pair prediction; E is the encoder; G is the decoder; Z is the codebook; D is the discriminator.

7. The two-stage channel state estimation coding and decoding method according to claim 6, characterized in that: In the adversarial training, the joint optimization goal of minimizing the reconstruction error and the adversarial loss is finally optimized as follows: In the formula, Q * is the optimal model parameter, indicating the optimal codec and generator model obtained through optimization; L VQ (E, G, Z) is the variational quantization loss; λL GAN ({E,G,Z},D) is the adversarial loss; λ is the weight coefficient.

8. A two-stage channel state estimation coding and decoding system, characterized in that: The system can be used to implement the two-stage channel state estimation coding and decoding method according to any one of claims 1 to 7, specifically comprising: A complex input module, used for receiving and processing MIMO channel state information; Data dimension upgrading module, used for dimension upgrading of input CSI data; Encoding latent space module, used to encode the processed CSI data into latent space; Generate adversarial data module, which is used to generate channel state information through the autoregressive Transformer model; The discriminator module is used to generate adversarial training and optimize the reconstruction accuracy of channel state information through adversarial learning between the generator and the discriminator.

9. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the two-stage channel state estimation coding and decoding method as claimed in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the two-stage channel state estimation encoding and decoding method described in one of claims 1 to 7 is implemented.

Images