Codebook construction method and device based on pseudo image expression, equipment and storage medium

Abstract

The invention belongs to the technical field of wireless communication, and discloses a codebook construction method and device based on pseudo image expression, equipment and a storage medium. The method comprises the following steps: respectively mapping channel state information to different color channels of a blank image according to real and imaginary parts to reconstruct image data, then extracting features of the image by using an image feature extraction model and carrying out feature clustering to generate an initial codebook, and then embedding the codebook into a generative adversarial network. And calculating and minimizing cosine similarity loss between codebook vectors, and finally outputting an optimized codebook. The spatial structure characteristics of channel information are fully mined through pseudo-image processing, and the direction of the codebook vector is constrained by using the cosine similarity loss, so that the generated codebook can be more accurately matched with the real channel direction when being used for CSI feedback, the quantization error is remarkably reduced, and the quantization efficiency is improved. And finally, the system spectrum efficiency and communication reliability are improved.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a codebook construction method, apparatus, device, and storage medium based on pseudo-image representation. Background Technology

[0002] With the rapid development of wireless communication technology, modern communication systems widely adopt massively multi-input multiple-output (MIMO) technology to meet the ever-increasing demands for data transmission rates and system capacity. By configuring a large number of antennas on the base station side, the system can significantly improve spectrum utilization and link reliability by utilizing spatial degrees of freedom, becoming one of the key implementation methods for fifth-generation and future mobile communication systems.

[0003] In the aforementioned massive MIMO systems, base stations rely on user equipment to accurately feed back downlink channel state information for optimization operations such as beamforming and precoding, thereby fully utilizing the multi-antenna gain. However, with the significant increase in the number of antennas, the dimensionality of the CSI data to be fed back increases dramatically, posing a severe challenge to traditional codebook-based feedback mechanisms. Specifically, to ensure channel reconstruction accuracy, the required codebook size continuously expands, while a larger codebook introduces considerable feedback index overhead, creating a prominent contradiction between accuracy and overhead.

[0004] Therefore, how to construct an efficient codebook that is suitable for large-scale antenna systems, so as to significantly reduce feedback overhead while maintaining high-precision channel reconstruction capability, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The main objective of this application is to provide a codebook construction method, apparatus, device, and storage medium based on pseudo-image representation, aiming to solve the technical problem in the prior art of how to construct an efficient codebook that is suitable for large-scale antenna systems, so as to significantly reduce feedback overhead while maintaining high-precision channel reconstruction capability.

[0006] To achieve the above objectives, this application proposes a codebook construction method based on pseudo-image representation, the method comprising: Based on the original channel state information, a channel state mapping image is obtained; Image feature extraction is performed on the channel state mapping image to obtain a set of feature vectors; Clustering is performed based on the set of feature vectors and the channel state information, and the clustering results are used as the initial codebook. The initial codebook is input into the codebook optimization generative adversarial network model, and the cosine similarity between the codebook vectors of the initial codebook is calculated to optimize directional consistency, thereby obtaining the optimized codebook.

[0007] Optionally, obtaining the channel state mapping image based on the original channel state information includes: Based on the original channel state information, the spatial frequency domain channel state matrix is ​​obtained; The spatial frequency domain channel state matrix is ​​transformed to obtain a sparse angle-delay domain channel matrix. The sparse angle-delay domain channel matrix is ​​separated into a real part matrix and an imaginary part matrix; The real and imaginary matrices are mapped to their respective color channels in the initial image to obtain the channel state mapping image.

[0008] Optionally, the step of performing a domain transformation on the space-frequency domain channel state matrix to obtain a sparse angle-delay domain channel matrix includes: Perform a first-dimensional inverse Fourier transform on the column vectors of the spatial frequency domain channel state matrix to convert the frequency domain information to the time delay domain, thereby obtaining the time delay domain channel matrix; Perform a second-dimensional inverse Fourier transform on the row vectors of the time-delay domain channel matrix after the first-dimensional transformation to convert the spatial domain information to the angular domain, thus obtaining the angular-time-delay domain channel matrix. The sparse angle-delay domain matrix is ​​obtained by sparsifying and pruning the angle-delay domain matrix.

[0009] Optionally, the step of extracting image features from the channel state mapping image to obtain a feature vector set includes: Semantic features are extracted from the channel state mapping image to obtain a first feature vector; Local structural features are extracted from the channel state mapping image to obtain a second feature vector; The feature vector set is obtained based on the first feature vector and the second feature vector.

[0010] Optionally, the step of clustering based on the feature vector set and the channel state information, and using the clustering result as the initial codebook, includes: The feature vector set is initialized with parameters to obtain multiple initial cluster centers; Based on the initial cluster centers, the feature vector set is iteratively optimized to obtain an optimized cluster center set. The optimized cluster center set is subjected to convergence judgment processing until the change in the cluster centers is less than a preset threshold or the number of iterations reaches a preset upper limit, and the iterative optimization result of the cluster centers is obtained. The iterative optimization results of the cluster centers are structured to obtain the initial codebook.

[0011] Optionally, the step of inputting the initial codebook into a codebook optimization generative adversarial network model, calculating the cosine similarity between the codebook vectors of the initial codebook to perform directional consistency optimization, and obtaining an optimized codebook includes: Configure the initial codebook as the training parameters for the vector quantization layer in the codebook-optimized generative adversarial network model; The original channel state information is input into the encoder of the generative adversarial network to obtain the latent feature vector; Based on the vector quantization layer, the potential feature vector is quantized and mapped to obtain the quantized feature vector; The quantized feature vector is reduced in dimension by a projector and then input into a decoder for reconstruction to obtain the reconstructed channel state information. Through end-to-end joint training, the parameters of the encoder, vector quantization layer, projector and decoder are updated, and the codebook vector in the vector quantization layer is optimized synchronously during training. The optimized codebook is obtained from the vector quantization layer in the converged model.

[0012] Optionally, the goal of the end-to-end joint training is to minimize a preset composite loss function; The composite loss function includes at least a reconstruction loss term and a feature fidelity loss term based on cosine similarity, wherein the feature fidelity loss term based on cosine similarity is used to constrain the directional consistency between the reconstructed channel state information and the original channel state information in the feature space.

[0013] Furthermore, to achieve the above objectives, this application also proposes a codebook construction apparatus based on pseudo-image representation, the pseudo-image representation-based codebook construction apparatus comprising: The image conversion module is used to obtain a channel state mapping image based on the original channel state information; The feature extraction module is used to extract image features from the channel state mapping image to obtain a set of feature vectors; The codebook construction module is used to perform clustering based on the feature vector set and the channel state information, and use the clustering result as the initial codebook. The codebook optimization and evaluation module is used to input the initial codebook into the codebook optimization generative adversarial network model, calculate the cosine similarity between the codebook vectors of the initial codebook to perform directional consistency optimization, and obtain the optimized codebook.

[0014] Furthermore, to achieve the above objectives, this application also proposes a codebook construction device based on pseudo-image representation, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the codebook construction method based on pseudo-image representation as described above.

[0015] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the codebook construction method based on pseudo-image representation as described above.

[0016] The technical solution of this application includes: obtaining a channel state mapping image based on the original channel state information; extracting image features from the channel state mapping image to obtain a set of feature vectors; clustering the feature vector set with the channel state information and using the clustering result as the initial codebook; inputting the initial codebook into a codebook optimization generative adversarial network model, calculating the cosine similarity between the codebook vectors of the initial codebook to perform directional consistency optimization, and obtaining the optimized codebook.

[0017] The proposed one or more technical solutions in this application have at least the following technical effects: By converting the original channel state information into a channel state mapping image and extracting high-dimensional features using image processing methods, the spatiotemporal structure and modal characteristics in the channel can be captured more comprehensively and precisely. Combined with feature clustering and generative adversarial networks for joint optimization, redundancy and conflicts between codebook vectors are effectively reduced, enhancing the codebook's ability to represent complex channel environments, thereby improving the accuracy of CSI reconstruction. By optimizing the initial codebook with cosine similarity, a reasonable directional distribution of codebook vectors is achieved, significantly compressing the codebook size while maintaining high reconstruction quality. This method enables the final codebook to express richer channel states with the same number of feedback bits, thereby reducing uplink channel resource consumption and improving the spectral efficiency of large-scale MIMO systems. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an embodiment of the codebook construction method based on pseudo-image representation in this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the codebook construction method based on pseudo-image representation in this application; Figure 3This is a flowchart illustrating Embodiment 3 of the codebook construction method based on pseudo-image representation in this application. Figure 4 This is a comparison chart of the channel state information compression performance of different compression feedback methods provided in Embodiment 3 of the codebook construction method based on pseudo-image representation in this application; Figure 5 This is a schematic diagram of the module structure of the codebook construction device based on pseudo-image representation in an embodiment of this application; Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the codebook construction method based on pseudo-image representation in the embodiments of this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] The main solution of this application embodiment is: to obtain the training tasks in the training task queue; to extract training features from the training tasks to obtain the training features of the training tasks; to obtain the communication operation workflow of the training tasks and the priority of each communication operation in the communication operation workflow according to the training features; to obtain the communication scheduling scheme of the training tasks according to the priority of each communication operation; and to execute the communication scheduling of the training tasks according to the communication scheduling scheme.

[0025] In this embodiment, for ease of description, the codebook construction device based on pseudo-image expression will be used as the execution subject for the following description.

[0026] The main solution of this application embodiment is as follows: obtain a channel state mapping image based on the original channel state information; extract image features from the channel state mapping image to obtain a set of feature vectors; cluster the feature vector set with the channel state information and use the clustering result as the initial codebook; input the initial codebook into the codebook optimization generative adversarial network model, calculate the cosine similarity between the codebook vectors of the initial codebook to perform directional consistency optimization, and obtain the optimized codebook.

[0027] Currently, while massive MIMO base stations can improve system capacity and spectral efficiency, they require user equipment to feed back downlink channel state information (CSI) to the base station for precoding and other optimizations. With the surge in the number of antennas, traditional codebook feedback schemes face severe challenges: the amount of CSI data to be fed back increases dramatically, and to ensure reconstruction accuracy, the codebook size must expand exponentially, leading to a significant increase in feedback index overhead and resulting in huge uplink resource consumption, which has become a bottleneck restricting system performance.

[0028] Traditional codebook schemes struggle to fundamentally overcome the aforementioned problems due to inherent limitations: First, based on predefined static codebook structures, traditional codebook schemes have a limited number of codewords. In extremely high-dimensional channels, maintaining accuracy necessitates a dramatic increase in codebook size and index bits, resulting in enormous storage and feedback overhead. Second, codebooks are largely based on ideal channel model assumptions, making them ill-suited to real, complex environments and variable user distributions. They lack adaptability to channel structure characteristics, leading to significant quantization errors. Simply increasing codebook size further exacerbates overhead. Third, they lack efficient feature extraction and semantic compression mechanisms, failing to fully utilize structural correlations in the channel's space-time-frequency domain, making it difficult to approach the information theory lower bound. Therefore, in existing codebook technologies, incremental improvements such as increasing codebook size, optimizing codeword design, or adjusting quantization strategies are insufficient to resolve the fundamental contradiction between accuracy and overhead. Constructing efficient codebooks adaptable to large-scale antenna systems, significantly reducing feedback overhead while maintaining high-precision channel reconstruction capabilities, is a pressing technical challenge in this field.

[0029] This application transforms raw channel state information into a channel state mapping image and extracts high-dimensional features using image processing methods. This enables a more comprehensive and refined capture of the spatiotemporal structure and modal characteristics of the channel. Combined with feature clustering and generative adversarial networks for joint optimization, it effectively reduces redundancy and conflicts among codebook vectors, enhancing the codebook's ability to represent complex channel environments and thus improving the accuracy of CSI reconstruction. By optimizing the initial codebook with cosine similarity, a reasonable directional distribution of codebook vectors is achieved, significantly compressing the codebook size while maintaining high reconstruction quality. This method allows the final codebook to express richer channel states with the same number of feedback bits, thereby reducing uplink channel resource consumption and improving the spectral efficiency of large-scale MIMO systems.

[0030] This application provides a solution aimed at addressing the technical problem of the lack of a global resource scheduling mechanism in the prior art, which can dynamically perceive task characteristics, coordinate computing and communication resources, and thus achieve overall performance optimization.

[0031] It should be noted that the execution subject in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a codebook construction device based on pseudo-image representation capable of implementing the above functions. The following description uses a codebook construction device based on pseudo-image representation as the execution subject to illustrate this embodiment and the subsequent embodiments.

[0032] Based on this, embodiments of this application provide a codebook construction method based on pseudo-image representation, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the codebook construction method based on pseudo-image representation in this application.

[0033] In this embodiment, the codebook construction method based on pseudo-image representation includes steps S10 to S40: Step S10: Obtain the channel state mapping image based on the original channel state information.

[0034] It should be noted that the raw channel state information refers to the channel response data measured at the receiver, which can be represented as a complex matrix containing amplitude and phase information. This information reflects the characteristics of the wireless channel in multiple dimensions, such as frequency domain subcarrier response, spatial domain antenna element correlation, or time domain variations. The channel state mapping image refers to converting the aforementioned raw channel data into a two-dimensional image representation through specific preprocessing and transformation, enabling it to be effectively identified and its features extracted by subsequent image processing algorithms (such as convolutional neural networks).

[0035] Understandably, this mapping process aims to transform abstract and complex channel parameters into an image structure with spatial local correlations. For example, each element of the channel response matrix is ​​typically a complex number containing real and imaginary parts, representing the amplitude variations and phase shifts experienced by the signal on a specific subcarrier and antenna pair. The real and imaginary parts of the channel matrix can be decomposed into two channels, mapping each element in the matrix to a corresponding pixel in the image. For instance, the real data can be mapped to the R channel of an RGB image, and the imaginary data to the B channel, thus achieving the transformation from the original channel state information to the mapped image while preserving complete complex domain information. Through such transformations, hidden structural features in the original signal (such as spatial beamform or frequency smoothness) are explicitly expressed, making them suitable for feature learning and dimensionality reduction. Similarly, the channel gains of different antenna pairs in a large-scale MIMO system can be arranged into a grayscale image matrix, where the grayscale value of each pixel represents the channel amplitude or phase information of the corresponding antenna pair on a specific subcarrier.

[0036] It should be understood that this step is not limited to a fixed mapping rule, and its specific implementation can be flexibly adjusted according to system architecture, antenna configuration, or spectrum parameters. During execution, by separating the real / imaginary parts or amplitude / phase, the originally abstract complex matrix is ​​converted into pixels with clearly defined numerical values. During the conversion process, the dimensional structure of the original matrix is ​​strictly maintained, directly mapping the matrix's rows (e.g., subcarrier indices) and columns (e.g., transmit-receive antenna pair indices) to the "height" and "width" of the image, thereby converting the spatial correlation and frequency correlation between different antennas and subcarriers into spatial proximity relationships between pixels in the image. After the above pseudo-image processing, the resulting two-dimensional image data format is fully compatible with the input format required by computer vision tools such as Convolutional Neural Networks (CNNs), allowing channel information to be treated as a special image with "texture" and "pattern" for processing. This conversion of channel state information into a pseudo-image successfully transforms the communication problem into a visual problem, enabling traditional communication problems to overcome the bottleneck of traditional channel feedback using the power of AI, providing key technical support for the practical application of advanced communication technologies such as large-scale MIMO.

[0037] Step S20: Extract image features from the channel state mapping image to obtain a set of feature vectors.

[0038] It should be noted that the image feature extraction method used in this step is not limited to a specific type of model or algorithm. Its core objective is to effectively capture and compress key information about the channel state from the image structure.

[0039] Understandably, conventional image feature extraction methods, through multi-layer convolution and pooling operations, can abstract increasingly sophisticated features layer by layer. Shallow networks might identify edges and textures (corresponding to spatial-frequency variations in the channel), while deep networks can combine these basic features to form more complex structural patterns (such as specific beamforms and multipath clusters). Feature extraction networks can extract the most essential information representing the core features of the original channel (such as main beam direction, delay spread, spatial-frequency correlation, etc.) and compress it into a very small feature vector. This vector can be seen as a digital "fingerprint" or "summary" of the channel information, with its data volume reduced by several orders of magnitude compared to the original image, thus greatly reducing feedback overhead.

[0040] It should be understood that, regardless of the model used, the final set of feature vectors must be robust, highly representative, and low in redundancy to support subsequent downstream tasks such as channel reconstruction, user identification, or beam management.

[0041] Step S30: Cluster the feature vector set and the channel state information, and use the clustering result as the initial codebook.

[0042] It should be noted that this step uses a clustering algorithm to perform unsupervised grouping of the feature vector set, aiming to uncover inherent patterns and typical features of the channel state to form representative codebook vectors. Specifically, the clustering process divides the sample points in the high-dimensional feature space into several categories, and the center point of each category corresponds to a codebook entry. The initial codebook is composed of the set of these center points.

[0043] Understandably, the clustering process utilizes the channel structure semantics inherent in the feature vectors, enabling the generated initial codebook to not only retain the key statistical characteristics of the original channel state but also possess good generalization and robustness. Since the feature vectors have already denoised and abstracted the original CSI data, the clustering operation can more clearly separate different channel modes (such as LOS, NLOS scenarios, or user location differences), thereby improving the codebook's expressive efficiency. Furthermore, the codebook size (i.e., the number of clusters) can be flexibly adjusted according to system capacity and accuracy requirements to balance feedback overhead and reconstruction quality.

[0044] In one feasible implementation, step S30 may include steps A10 to A40: Step A10: Perform parameter initialization processing on the feature vector set to obtain multiple initial cluster centers.

[0045] It should be noted that the specific initialization process can employ the probabilistic selection mechanism of the K-means++ algorithm, prioritizing the selection of sample points that are relatively far apart as initial centers; or, based on the density peak assumption, selecting representative sample points from high-density regions of the feature space. These methods ensure that the initial centers better cover the distribution range of the feature vectors.

[0046] Understandably, since feature vectors are abstract representations extracted from channel state information, they already possess good discriminative power. Initialization methods based on data distribution characteristics can more accurately capture the multimodal features of channel states. Especially for channel clustering tasks in large-scale MIMO systems, different channel states often exhibit significant clustering in the feature space. A reasonable selection of initial centers can significantly improve the convergence speed and result stability of the subsequent clustering process.

[0047] It should be understood that the quality of the initial cluster centers directly affects the convergence direction of the entire clustering process and the performance of the final codebook. Appropriate initialization not only enables the clustering algorithm to find the globally optimal or near-optimal partition more quickly, but also generates more representative codebook vectors. These initial centers will serve as the starting point for subsequent iterative optimizations, continuously adjusted by the clustering algorithm until a stable cluster partition is formed.

[0048] Step A20: Based on the initial cluster centers, perform iterative optimization on the feature vector set to obtain the optimized cluster center set.

[0049] It should be noted that this step iteratively adjusts the positions of cluster centers and the categories to which feature vectors belong, aiming to minimize intra-cluster distances and maximize inter-cluster discriminability. Specifically, each iteration includes two key operations: first, each feature vector is assigned to the cluster center with the nearest Euclidean distance to form a temporary category; then, the mean of all feature vectors in each category is recalculated as the new cluster center. By alternately executing the assignment and update steps, the cluster centers gradually migrate to denser areas of the data distribution.

[0050] Understandably, since the initial cluster centers are only preliminary estimates of the data distribution, the positions of the centers can be continuously corrected through multiple iterations, eventually converging to the vicinity of the true centers of each category. This optimization process is particularly suitable for clustering tasks of channel state feature vectors, because the time-varying characteristics of wireless channels mean that the feature vector distribution may have complex patterns, requiring iterative methods to gradually capture its inherent clustering structure.

[0051] Optionally, the iterative optimization process includes: performing nearest neighbor assignment on the feature vector set based on the current cluster centers to obtain cluster assignment results; and performing cluster center update on the feature vector set based on the cluster assignment results to obtain a new generation of cluster centers.

[0052] It should be understood that the optimized set of cluster centers obtained after sufficient iterations can represent the distribution characteristics of the original feature vectors to the greatest extent, and each cluster center corresponds to a typical channel state pattern. These optimized centers will constitute the core elements of the codebook, providing a reference vector for subsequent channel state quantization, and ensuring that the system can compress and reconstruct channel information with minimal distortion.

[0053] Step A30: Perform convergence judgment processing on the optimized cluster center set until the change in the cluster centers is less than a preset threshold or the number of iterations reaches a preset upper limit, and obtain the iterative optimization result of the cluster centers.

[0054] It should be noted that in practical implementation, the difference between the cluster center positions of the current round and the previous round is usually measured using Euclidean distance or squared error. For example, the mean or maximum value of the movement distance of all cluster centers can be calculated. When this value is lower than a preset threshold for several consecutive iterations, the algorithm is considered to have converged. At the same time, a maximum number of iterations is set as a forced termination condition to prevent infinite loops in cases where convergence is not complete or is too slow.

[0055] It is understandable that when the algorithm converges normally, the obtained cluster centers can fully represent the distribution characteristics of the feature vectors, forming a stable and effective codebook structure; even if it terminates due to reaching the upper limit of iteration, the obtained suboptimal solution can usually meet the needs of practical applications.

[0056] Step A40: The iterative optimization results of the cluster centers are structured to obtain the initial codebook.

[0057] It should be noted that the structured processing includes operations such as normalization, sorting, index allocation, and metadata annotation of the cluster center vectors. The purpose of this step is to construct the discrete cluster center set obtained after convergence judgment into a codebook data structure usable by the system according to a specific format and organization method.

[0058] Understandably, all cluster center vectors can be sorted in ascending order according to their Euclidean norm and assigned unique index numbers. Simultaneously, the cluster size or usage frequency statistics corresponding to each codeword can be recorded, forming a codebook matrix with fixed dimensions and format. This process transforms the abstract mathematical clustering results into communication resource units with clear engineering significance.

[0059] It should be understood that normalization ensures all codewords meet power constraints or unit modulus requirements; sorting and indexing establish an efficient codeword retrieval mechanism; and adding metadata provides information support for subsequent codebook optimization and dynamic adjustment. This process allows machine learning-based clustering results to be seamlessly integrated into traditional communication frameworks, retaining the advantages of data-driven methods while meeting the standardized requirements of communication systems for resource formats.

[0060] Step S40: Input the initial codebook into the codebook optimization generative adversarial network model, calculate the cosine similarity between the codebook vectors of the initial codebook to perform directional consistency optimization, and obtain the optimized codebook.

[0061] It should be noted that this step introduces the adversarial training mechanism of Generative Adversarial Network (GAN) to fine-tune the initial codebook. By calculating the cosine similarity between codebook vectors, the directional redundancy between codewords can be quantified, and adjustments are made based on this to ensure that the codebook can uniformly cover the entire channel space.

[0062] Understandably, directional consistency optimization aims to address the uneven codeword distribution problem that may arise from traditional clustering methods, where channel states in some directions are overrepresented while others are underrepresented. Cosine similarity effectively captures the spatial directional differences of high-dimensional vectors. By minimizing the maximum similarity pairs through adversarial training, the overall diversity and representational power of the codebook can be systematically improved. This optimization does not simply change the vector magnitude but focuses on adjusting its phase or direction, which is crucial for modern wireless communication systems that rely on beamforming, as it directly affects the gain and interference suppression capabilities of multi-antenna arrays. For example, in practice, by constructing a cosine similarity-based regularization loss term, codebook vectors that are too close in direction are penalized, prompting the generator to push them away from each other, ultimately forming an isotropic and uniformly distributed codebook.

[0063] It should be understood that a codebook optimized for directional consistency can more accurately quantify various possible channel states, thereby enabling the receiver to reconstruct channel information more accurately with limited feedback overhead. This directly translates to more accurate beamforming and higher spectral efficiency in downlink transmission.

[0064] This embodiment decomposes the real and imaginary parts of the channel state information and maps them to image channels to form a two-dimensional representation with spatial correlation. Then, a condensed set of feature vectors is obtained through a feature extraction network. Next, an improved clustering algorithm is used to group the feature vectors and generate a structured initial codebook. Finally, the cosine similarity between codebook vectors is calculated through a generative adversarial network to optimize the directional distribution uniformity of the codebook in an adversarial training manner, and finally a codebook that meets the needs of large-scale communication is obtained.

[0065] In summary, this technical solution, through image processing and convolutional feature extraction, can fully exploit the spatial-frequency structure features in the channel state, significantly improving feature representation capabilities compared to traditional scalar quantization methods. By generating an initial codebook through clustering, the codebook can adaptively learn channel statistical characteristics, avoiding the dependence of traditional codebooks on prior channel knowledge. In particular, by optimizing the directional consistency of the codebook through generative adversarial networks, and utilizing cosine similarity to measure and reduce directional redundancy between codewords, the codebook uniformly covers the entire channel space. This innovative approach effectively solves the problem of uneven codebook distribution in traditional methods. Ultimately, the optimized codebook can provide more accurate channel quantization performance with limited feedback overhead, thereby improving beamforming accuracy and system spectral efficiency, providing an effective solution for the practical application of large-scale MIMO systems.

[0066] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2Step S10 in the codebook construction method based on pseudo-image representation includes steps S101 to S104: Step S101: Obtain the spatial frequency domain channel state matrix based on the original channel state information.

[0067] It should be noted that the raw channel state information is typically a set of channel parameters extracted from the time domain or from the received signal after preliminary processing. Through specific signal processing transformations, or by organizing and arranging different antenna dimensions, a matrix that can simultaneously reflect the channel characteristics in both spatial and frequency dimensions is ultimately constructed. The rows and columns of this matrix usually correspond to different physical dimensions.

[0068] Understandably, projecting the multidimensional characteristics of the channel (such as multipath delay and spatial beam direction) into a structured two-dimensional matrix representation lays the foundation for subsequent image mapping. For example, in a massive MIMO system, the original channel impulse response, after being transformed to the frequency domain by an FFT, yields a three-dimensional tensor with dimensions of "number of receive antennas × number of transmit antennas × number of subcarriers". By arranging the responses of different transmit-receive antenna pairs on the same subcarrier into rows or columns of a matrix, the final spatial-frequency domain channel state matrix can be obtained, thus preserving the joint spatial and frequency variation characteristics of the channel.

[0069] Step S102: Perform a domain transformation on the spatial frequency domain channel state matrix to obtain a sparse angle-delay domain channel matrix.

[0070] In one feasible implementation, step S102 may include steps B10 to B30: Step B10: Perform a first-dimensional inverse Fourier transform on the column vectors of the spatial frequency domain channel state matrix to convert the frequency domain information to the time delay domain, thereby obtaining the time delay domain channel matrix.

[0071] It should be noted that in the specific operation, it is performed along the column vectors of the spatial frequency domain matrix (usually corresponding to the frequency dimension). The aim is to use Fourier transform to convert the subcarrier channel coefficients in the frequency domain into tap gains in the time delay domain, thereby obtaining the energy distribution characteristics of the channel in the multipath time delay dimension.

[0072] Understandably, this transformation effectively compresses the representation of channel information and highlights its sparsity. In wireless communication, multipath channels typically contain significant energy in only a few major taps in the time-delay domain. For example, in an OFDM system, performing an IDFT on the frequency-domain channel vector of each receiving antenna converts it into a time-delay domain representation where non-zero values ​​are concentrated on a few taps. This significantly reduces the data dimensionality for subsequent processing and highlights key features.

[0073] It should be understood that the time-delay domain channel matrix not only preserves the spatial characteristics of the original channel (represented by row vectors), but also obtains structured information in the time-delay dimension through transformation. This sparse representation helps improve channel quantization efficiency and reduce feedback overhead.

[0074] Step B20: Perform a second-dimensional inverse Fourier transform on the row vectors of the time-delay domain channel matrix after the first-dimensional transformation to convert the spatial domain information to the angular domain, thereby obtaining the angular-time-delay domain channel matrix.

[0075] It should be noted that this step, based on the completed time-delay domain transformation, further applies an inverse Fourier transform to the row vectors of the matrix (usually corresponding to spatial dimensions, such as antenna arrays). The purpose of this operation is to transform the spatial channel response to its corresponding angular domain representation, thereby utilizing the sparsity characteristics of the wireless channel in the angular domain.

[0076] Understandably, the spatial-to-angular domain transformation essentially maps the antenna array's response to an angular spectrum along the direction of arrival (DoD) or direction of departure (DoA). For example, in a uniform linear array (ULA) system, performing an IDFT on this dimension is equivalent to performing a virtual angular domain decomposition on the spatial channel, such that each transformed output dimension corresponds to a specific propagation direction. When combined with the previous time-delay domain, the channel energy is concentrated on a few significant coefficients in the angular-time-delay domain, forming a highly structured two-dimensional sparse representation.

[0077] It should be understood that the final angle-delay domain channel matrix reveals the essential characteristics of the channel from the two key physical dimensions of time delay and angle. This dual-domain transformation not only greatly compresses the data, providing a highly condensed and distinctive input for subsequent generation of image representations, but is also a key prerequisite for achieving efficient feature extraction and codebook optimization, significantly improving the accuracy and efficiency of channel state information feedback.

[0078] Step B30: Perform sparsification pruning on the angle-delay domain matrix to obtain the sparse angle-delay domain channel matrix.

[0079] It should be noted that the pruning in this step is based on a preset energy threshold or retention coefficient criterion. Elements in the diagonal-time-delay domain matrix are filtered, retaining only the most significant values ​​with the largest amplitudes, while setting other minor, weaker matrix elements to zero. After pruning, the data can be further compressed.

[0080] Understandably, in real-world large-scale MIMO channels, the signal propagation path is finite, and its energy is mainly distributed in a few key delay taps and the angle of arrival / departure angle. By setting a threshold (such as retaining a coefficient that accounts for 95% of the total energy) or directly specifying the coefficients that retain the top K largest amplitude values, noise and minor components can be effectively removed. The sparse angle-delay domain channel matrix obtained after pruning greatly reduces data dimensionality and storage overhead while preserving the key information that determines channel quality to the maximum extent.

[0081] Step S103: Separate the sparse angle-delay domain channel matrix into a real part matrix and an imaginary part matrix.

[0082] It is understandable that any complex matrix can be uniquely composed of the real and imaginary parts of all its elements. For example, for a sparse complex matrix containing primary path information, the separated real part matrix reflects the in-phase component of the channel response, while the imaginary part matrix reflects the quadrature component. These two parts of information physically characterize the amplitude and phase relationship of the signal. Separation processing allows complex complex data to be transformed into a real number form that is more in line with conventional image processing paradigms (such as RGB dual channels).

[0083] Optionally, a first real matrix and a second real matrix with the same dimension as the sparse angle-delay domain channel matrix are created; the real part value of each element in the sparse angle-delay domain channel matrix is ​​assigned to the corresponding position of the first real matrix to generate the real part matrix; the imaginary part value of each element in the sparse angle-delay domain channel matrix is ​​assigned to the corresponding position of the second real matrix to generate the imaginary part matrix.

[0084] Step S104: Map the real part matrix and the imaginary part matrix to their respective color channels in the initial image to obtain the channel state mapping image.

[0085] It should be noted that the values ​​of the two real matrices (real part matrix and imaginary part matrix) obtained in the aforementioned steps, after normalization and other preprocessing, are used to fill two predefined color channels in a multi-channel image. The initial image here is usually an empty image template in which all pixel initial values ​​can be set to 0, and its spatial dimensions are consistent with the real or imaginary part matrix.

[0086] Understandably, after normalizing all values ​​of the real part matrix to the interval [0, 255], it is filled into the first channel of the image (e.g., the simulated "R" channel), while the imaginary part matrix is ​​similarly processed and filled into the second channel of the image (e.g., the simulated "G" channel), thus generating a two-dimensional image that can characterize the spatial distribution features of the channel amplitude and phase. The resulting channel state mapping image is a highly structured data representation. It transforms the originally complex one-dimensional channel response information into a standard image format that can be directly processed by mature computer vision models. This greatly facilitates the subsequent automatic extraction of deep features using deep learning techniques.

[0087] In this embodiment, a spatial frequency domain channel state matrix is ​​constructed using the original channel state information. Then, it is transformed to the angle-delay domain using a two-dimensional inverse Fourier transform and sparsed to obtain a highly condensed sparse angle-delay domain channel matrix. This complex matrix is ​​then separated into two real matrices, a real part and an imaginary part, and finally mapped to the two color channels of the image to generate a channel state mapping image.

[0088] In summary, this technical solution transforms complex channel state information into a structured image format, converting high-dimensional complex channel data, which is inherently difficult to process directly using deep learning models, into image data that conforms to the input standards of visual models such as convolutional neural networks. This transformation not only fully preserves the dual sparsity characteristics of the channel in both the angular and time-delay domains but also significantly improves the expressiveness and recognizability of features through image representation. This greatly enhances the efficiency and accuracy of channel feature extraction, providing a higher-quality data foundation for subsequent deep learning-based channel quantization, feedback, and reconstruction, thereby improving the overall spectral efficiency and reliability of the communication system.

[0089] Based on the first embodiment of this application, in the third embodiment of this application, the content that is the same as or similar to that in the second embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 Step S40 in the codebook construction method based on pseudo-image representation includes steps C10 to C50: Step C10: Configure the initial codebook as the training parameters for the vector quantization layer in the codebook-optimized generative adversarial network model.

[0090] It should be noted that this step is a crucial operation in the model initialization phase. Its core content is to set the initial codebook generated in the previous steps as the training parameters for the vector quantization layer in the generative adversarial network. The initial codebook is usually generated by random distribution or obtained by clustering historical data (such as the K-means algorithm). In the implementation, it is directly registered as the weight tensor of the VQL layer to ensure that its value can be updated in subsequent backpropagation.

[0091] Understandably, during training, the codebook vector position is continuously adjusted by quantizing the loss gradient. In the 5G MassiveMIMO scenario, the codebook significantly focuses on the core channel feature region after multiple rounds of training. This mechanism is the core guarantee for achieving low-overhead feedback.

[0092] Step C20: Input the original channel state information into the encoder of the generative adversarial network to obtain the latent feature vector.

[0093] It's important to note that this step is the first stage in the model's forward propagation process. Its core is inputting the raw channel state information data into the encoder component of the generative adversarial network (GAN). The raw channel state information typically includes multidimensional channel gain, phase, and delay data, represented as a complex matrix, for example, in a 5G MIMO system. This encoder is usually composed of convolutional neural network layers, designed to compress and transform the high-dimensional input into a low-dimensional space by automatically learning feature maps. Its output latent feature vector is a fixed-dimensional vector that characterizes the essential features of the original data, such as path loss and multipath effects, while preserving crucial spatial information.

[0094] Understandably, this step is not only a crucial step in signal preprocessing but also a fundamental supporting mechanism for the entire generative adversarial network. In the field of electronic communications, the original CSI is typically massive (e.g., on the order of 10^4 elements). By compressing it into a latent vector through an encoder (e.g., reducing it to the order of 10^2), redundant data and transmission overhead are significantly reduced. In an example, consider an LTE system in an urban microcell scenario: the original CSI may contain complex features of inter-cell interference. After being input into the encoder, it becomes a vectorized representation, which facilitates the efficient retrieval of the discrete codebook by the vector quantization layer in subsequent steps.

[0095] Step C30: Quantize and map the latent feature vector based on the vector quantization layer to obtain the quantized feature vector.

[0096] It should be noted that the latent feature vectors are discretized and mapped using a predefined codebook in the Vector Quantization Layer (VQL) to generate quantized feature vectors suitable for transmission or storage. The codebook consists of a set of learnable vectors, such as 256 64-dimensional embedding terms. The quantization process calculates the distance between the input vector and each vector in the codebook, and selects the nearest neighbor codeword as the quantization result, thereby achieving a low-dimensional discrete representation of high-dimensional continuous features.

[0097] Understandably, this method effectively compresses the data size and improves the robustness of features, making it particularly suitable for communication scenarios such as channel state information feedback. For example, in a 5G NR system, the high-dimensional features output by the encoder are mapped to an 8-bit index via VQL. Only this index needs to be transmitted instead of the original data, significantly reducing feedback overhead while maintaining reconstruction accuracy.

[0098] It should be understood that quantization mapping not only reduces resource consumption, but also enhances the model's tolerance to noise and interference through discrete representation, which can combat signal distortion caused by multipath fading and improve the overall performance of the system in complex channel environments.

[0099] Step C40: The quantized feature vector is reduced in dimension by a projector and then input into the decoder for reconstruction to obtain the reconstructed channel state information.

[0100] It should be noted that the quantized feature vectors undergo dimensionality reduction using a projector to further reduce data redundancy and extract core information. The projector is typically implemented using a linear transformation layer or a lightweight neural network. Its role is to map high-dimensional discrete features to a more representative low-dimensional latent space, thereby providing the decoder with a more compact and information-density-rich input.

[0101] Understandably, the dimensionality-reduced features are progressively reconstructed by the decoder to recover accurate channel state information. The decoder gradually expands the feature dimension and reconstructs channel details through cascaded deconvolution or upsampling operations. For example, it can restore the complete channel matrix of a multi-antenna system from compressed features, ensuring that the reconstructed signal maintains high fidelity in both spatial and frequency dimensions.

[0102] It should be understood that this structure, through coordinated dimensionality reduction and reconstruction operations, significantly reduces communication overhead while maintaining the integrity of channel information. For example, in dense multi-user environments, the system can achieve high-quality channel reconstruction with only minimal feedback data, effectively improving spectrum utilization and enhancing the system's robustness against interference.

[0103] Step C50: Through end-to-end joint training, update the parameters of the encoder, vector quantization layer, projector and decoder, and simultaneously optimize the codebook vector in the vector quantization layer during training, and obtain the optimized codebook from the vector quantization layer in the converged model.

[0104] It should be noted that this step synchronously optimizes the parameters of the encoder, vector quantization layer, projector, and decoder through end-to-end joint training, ensuring that each module works collaboratively. During training, the model calculates the reconstruction error gradient using the backpropagation algorithm and updates the network weights using gradient descent. The codebook vector in the vector quantization layer participates in the optimization as a learnable parameter. The codebook continuously updates its embedding representation using exponential moving average or gradient descent to better adapt it to the distribution of input features, for example, gradually approximating the true cluster centers of channel features, thus improving the accuracy of the quantization representation. After training converges, the optimized codebook can be extracted from the vector quantization layer, which can efficiently represent the core patterns of channel features.

[0105] In a feasible embodiment, the goal of end-to-end joint training is to minimize a preset composite loss function; the composite loss function includes at least a reconstruction loss term and a feature fidelity loss term based on cosine similarity, wherein the feature fidelity loss term based on cosine similarity is used to constrain the directional consistency between the reconstructed channel state information and the original channel state information in the feature space.

[0106] It's important to note that the reconstruction loss term aims to ensure numerical accuracy. It measures the mean squared error of the numerical differences (amplitude and phase) between the reconstructed channel matrix and the original channel matrix at each element point. Minimizing this loss allows the reconstructed result to be as numerically close as possible to the original signal. The feature fidelity loss term based on cosine similarity aims to ensure structural consistency. It measures the alignment of two vectors in direction, i.e., the cosine of the angle between them, and is insensitive to absolute amplitude changes. For CSI, phase information determines the beamforming direction. Maximizing cosine similarity (i.e., minimizing its corresponding loss) aims to ensure that the key structural features of the reconstructed channel (especially the phase relationship) remain consistent with the original channel.

[0107] Understandably, the core advantage of introducing a cosine similarity-based loss term lies in its ability to effectively measure and constrain the structural similarity of high-dimensional complex signals in the feature space. This is crucial for signals with specific spatial structures, such as channel state information, because key channel information is often contained in its relative phase and direction characteristics, rather than absolute amplitude values. This loss term can guide the model to learn and preserve these essential feature relationships, thereby improving reconstruction quality.

[0108] It should be understood that this design significantly enhances the model's robustness and generalization ability under adverse channel conditions. For example, when the channel is subjected to interference such as amplitude scaling, the loss calculated based on amplitude will produce a large error, while cosine similarity, due to its inherent invariance to vector size, can ignore this interference and still accurately assess the structural fidelity of the signal. This ensures that the trained compression and feedback system can operate more stably and reliably, ultimately improving the performance of the entire communication system.

[0109] like Figure 4 As shown, Figure 4 This is a comparison chart of the channel state information compression performance of different compression feedback methods according to embodiments of the present invention.

[0110] It's important to note that LASSO (Least Absolute Shrinkage and Selection Operator) is a regression analysis method applied in statistics and signal processing. In compressed sensing, L1 regularization (i.e., constraining the sum of the absolute values ​​of coefficients) promotes sparsity, thereby reconstructing the original signal from finite measurement data. Essentially, it solves a convex optimization problem, automatically performing feature selection (i.e., making unimportant coefficients zero) while maintaining model prediction accuracy. BM3D-AMP (Block-Matching and 3D Filtering - Approximate Message Passing) is a hybrid algorithm that combines the powerful denoising mechanism (BM3D) from traditional image processing with the modern compressed sensing reconstruction framework (AMP). The AMP algorithm produces a noisy signal estimate in each iteration, while BM3D is used as the denoising mechanism in this iteration, utilizing its image block matching and collaborative filtering capabilities to clean the signal. BM3D-AMP excels at reconstructing image signals with rich texture and detail from highly compressed measurements. TVAL3 (Total Variation Augmented Lagrangian and Alternating Direction Algorithm) is an optimization algorithm for compressed sensing signal reconstruction. Its core is Total Variation (TV) regularization, which assumes the signal is sparse in the gradient domain (i.e., the image or signal is piecewise smooth). It uses the augmented Lagrangian method and the alternating direction multiplier method to solve the optimization problem, making it very suitable for reconstructing images with sharp edges and patchy continuity, effectively suppressing noise and preserving edge structure. CSINE (Compressive Sensing Image Reconstruction Network) is a deep learning model specifically designed for image reconstruction. It utilizes a neural network architecture to learn prior knowledge of the image and directly recover high-quality images from compressed measurements. The core advantage of this type of method lies in its data-driven approach, which is trained on a large amount of data. The network learns how to better handle specific types of images, thus achieving better reconstruction results than general algorithms, especially at low sampling rates. CS-CSNET, short for Compressed Sensing Based Compressed Sampling Network, is a deep learning-based compressed sensing reconstruction framework.It is an end-to-end neural network whose structure is designed to simulate the two processes of compressed sensing: sampling and reconstruction. Unlike traditional iterative optimization algorithms, CS-CSNet learns the optimal mapping relationship from compressed measurements to the original signal through training data. Once trained, reconstruction is just a forward propagation process, thus having an extremely fast reconstruction speed compared to other algorithms.

[0111] Understandably, this performance comparison chart compares the performance of the proposed method with traditional compressed sensing methods (such as LASSO, BM3D-AMP, and TVAL3) and deep learning-based compressed sensing methods (such as CSINE and CS-CSNet) in the channel state information (CSI) compression and reconstruction task from multiple dimensions. Specifically, signal reconstruction tests were conducted using datasets from two physical environments: indoor and outdoor datasets. The normalized mean square error (NMSE) and cosine similarity (ρ) indices were calculated at different compression ratios to quantitatively evaluate the performance of each algorithm.

[0112] It should be understood that NMSE measures numerical / energy accuracy, representing the proportion of energy in the reconstruction error. A larger (negative) value is better, indicating that the reconstructed channel coefficients are accurate in amplitude, which is important for applications such as power control. Cosine similarity measures structural / phase consistency, representing the directional consistency between the reconstructed signal and the original signal. The closer to 1, the better, indicating that the reconstructed channel coefficients are accurate in phase. This is crucial for beamforming because beamforming weights heavily depend on the phase relationship of the signals. A channel estimate with a phase error will cause the beam to be pointed in the wrong direction, severely degrading communication quality.

[0113] It is important to emphasize that, as shown in the data in the figure, regardless of whether it is an indoor or outdoor scenario, or under different compression ratios, the NMSE / dB value of the method of this invention is significantly lower than that of all the comparison schemes. Secondly, as shown in the cosine similarity data listed in the table, the ρ value of the reconstruction result of the method of this invention is closest to 1, far exceeding other comparison algorithms. This indicates that the channel vector reconstructed by this method is almost completely identical to the original channel vector in direction, proving that the method of this invention has excellent phase recovery capability. Finally, under different compression ratios (e.g., 1 / 4, 1 / 8, 1 / 16, 1 / 32), the performance degradation of the method of this invention is much smaller than that of traditional schemes. Even at extremely high compression ratios (e.g., CR=1 / 32), this invention can still maintain acceptable or even superior reconstruction quality. This reflects the extremely high compression efficiency of the method of this invention, which means that user equipment can send back sufficiently accurate CSI information to the base station with less feedback overhead. This greatly improves the uplink bandwidth utilization efficiency, saves valuable spectrum resources, and leaves space for more users to access or transmit service data.

[0114] In this embodiment, a codebook optimization generative adversarial network model is constructed and initialized, and the initial codebook is configured as a trainable parameter of the vector quantization layer in the model. Then, the original channel state information is input into the encoder to obtain the latent feature vector. The vector is then subjected to nearest neighbor quantization mapping using the vector quantization layer. The quantized features are then projected and dimensionality reduced before being input into the decoder to reconstruct the channel information. Finally, the network parameters and codebook vector are synchronously optimized through end-to-end joint training, and the optimized codebook is obtained from the vector quantization layer of the converged model.

[0115] In summary, this technical solution significantly improves quantization accuracy by enabling the codebook to adaptively approximate the true distribution of channel features through a joint training mechanism. Furthermore, the quantization operation transforms high-dimensional continuous features into low-dimensional discrete representations, greatly reducing the amount of feedback data. The introduction of a cosine similarity-based loss function effectively constrains the consistency of the reconstructed signal with the original signal in phase and structure, ensuring the directional accuracy required for beamforming. Therefore, this method ultimately outperforms traditional schemes in both NMSE and cosine similarity metrics, achieving the beneficial effect of maintaining high-precision reconstruction with extremely low feedback overhead and significantly improving the system's spectral efficiency.

[0116] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the codebook construction method based on pseudo-image expression in this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0117] This application also provides a codebook construction device based on pseudo-image representation, please refer to... Figure 5 The codebook construction device based on pseudo-image representation includes: Image conversion module 10 is used to obtain a channel state mapping image based on the original channel state information; Feature extraction module 20 is used to extract image features from the channel state mapping image to obtain a set of feature vectors; The codebook construction module 30 is used to perform clustering based on the feature vector set and the channel state information, and use the clustering result as the initial codebook. The codebook optimization and evaluation module 40 is used to input the initial codebook into the codebook optimization generative adversarial network model, calculate the cosine similarity between the codebook vectors of the initial codebook to perform directional consistency optimization, and obtain the optimized codebook.

[0118] In one embodiment, the image conversion module 10 is further configured to: obtain a spatial-frequency domain channel state matrix based on the original channel state information; perform a domain transformation on the spatial-frequency domain channel state matrix to obtain a sparse angle-delay domain channel matrix; separate the sparse angle-delay domain channel matrix into a real part matrix and an imaginary part matrix; and map the real part matrix and the imaginary part matrix to their respective color channels in the initial image to obtain the channel state mapping image.

[0119] In one embodiment, the image conversion module 10 is further configured to perform a first-dimensional inverse Fourier transform on the column vectors of the spatial-frequency domain channel state matrix to convert the frequency domain information to the time delay domain, thereby obtaining a time delay domain channel matrix; perform a second-dimensional inverse Fourier transform on the row vectors of the time delay domain channel matrix after the first-dimensional transform to convert the spatial domain information to the angular domain, thereby obtaining an angular-time delay domain channel matrix; and perform sparsification pruning on the angular-time delay domain matrix to obtain the sparse angular-time delay domain channel matrix.

[0120] In one embodiment, the feature extraction module 20 is further configured to perform semantic feature extraction on the channel state mapping image to obtain a first feature vector; perform local structural feature extraction on the channel state mapping image to obtain a second feature vector; and obtain the feature vector set based on the first feature vector and the second feature vector.

[0121] In one embodiment, the codebook construction module 30 is further configured to perform parameter initialization processing on the feature vector set to obtain multiple initial cluster centers; based on the initial cluster centers, perform iterative optimization processing on the feature vector set to obtain an optimized cluster center set; perform convergence judgment processing on the optimized cluster center set until the change in the cluster centers is less than a preset threshold or the number of iterations reaches a preset upper limit to obtain the iterative optimization result of the cluster centers; and perform structured processing on the iterative optimization result of the cluster centers to obtain the initial codebook.

[0122] In one embodiment, the codebook optimization and evaluation module 40 is further configured to: configure the initial codebook as the training parameters of the vector quantization layer in the codebook-optimized generative adversarial network model; input the original channel state information into the encoder of the generative adversarial network to obtain a latent feature vector; perform quantization mapping on the latent feature vector based on the vector quantization layer to obtain a quantized feature vector; input the quantized feature vector into the decoder after dimensionality reduction by the projector to reconstruct the reconstructed channel state information; update the parameters of the encoder, vector quantization layer, projector, and decoder through end-to-end joint training, and simultaneously optimize the codebook vector in the vector quantization layer during training, and obtain the optimized codebook from the vector quantization layer in the converged model.

[0123] In one embodiment, the codebook optimization evaluation module 40 is further configured to minimize a preset composite loss function as the objective of the end-to-end joint training; the composite loss function includes at least a reconstruction loss term and a feature fidelity loss term based on cosine similarity, wherein the feature fidelity loss term based on cosine similarity is used to constrain the directional consistency between the reconstructed channel state information and the original channel state information in the feature space.

[0124] The codebook construction apparatus based on pseudo-image representation provided in this application, employing the codebook construction method based on pseudo-image representation in the above embodiments, can solve the technical problem in the prior art of how to construct an efficient codebook adapted to large-scale antenna systems, significantly reducing feedback overhead while maintaining high-precision channel reconstruction capabilities. Compared with the prior art, the beneficial effects of the codebook construction apparatus based on pseudo-image representation provided in this application are the same as those of the codebook construction method based on pseudo-image representation provided in the above embodiments, and other technical features in the codebook construction apparatus based on pseudo-image representation are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0125] This application provides a codebook construction device based on pseudo-image representation. The codebook construction device based on pseudo-image representation includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the codebook construction method based on pseudo-image representation in the above embodiment 1.

[0126] The following is for reference. Figure 6This document illustrates a schematic diagram of a codebook construction device based on pseudo-image representation suitable for implementing embodiments of this application. The codebook construction device based on pseudo-image representation in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 6 The codebook construction device based on pseudo-image representation shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0127] like Figure 6 As shown, the codebook construction device based on pseudo-image representation may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the codebook construction device based on pseudo-image representation. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the pseudo-image representation-based codebook building device to communicate wirelessly or wiredly with other devices to exchange data. Although pseudo-image representation-based codebook building devices with various systems are shown in the figures, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0128] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0129] The codebook construction device based on pseudo-image representation provided in this application, employing the codebook construction method based on pseudo-image representation in the above embodiments, can solve the technical problem in the prior art of how to construct an efficient codebook adapted to large-scale antenna systems, significantly reducing feedback overhead while maintaining high-precision channel reconstruction capabilities. Compared with the prior art, the beneficial effects of the codebook construction device based on pseudo-image representation provided in this application are the same as those of the codebook construction method based on pseudo-image representation provided in the above embodiments, and other technical features in this codebook construction device based on pseudo-image representation are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0130] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0131] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0132] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the codebook construction method based on pseudo-image representation in the above embodiments.

[0133] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), or any suitable combination thereof.

[0134] The aforementioned computer-readable storage medium may be included in a codebook construction device based on pseudo-image representation; or it may exist independently and not assembled into a codebook construction device based on pseudo-image representation.

[0135] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a codebook construction device based on pseudo-image representation, the codebook construction device based on pseudo-image representation: obtains a channel state mapping image based on the original channel state information; extracts image features from the channel state mapping image to obtain a set of feature vectors; clusters the feature vector set with the channel state information and uses the clustering result as an initial codebook; inputs the initial codebook into a codebook optimization generative adversarial network model, calculates the cosine similarity between the codebook vectors of the initial codebook to perform directional consistency optimization, and obtains an optimized codebook.

[0136] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0137] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0138] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0139] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described codebook construction method based on pseudo-image representation. This solves the technical problem in the prior art of how to construct an efficient codebook suitable for large-scale antenna systems, significantly reducing feedback overhead while maintaining high-precision channel reconstruction capabilities. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the codebook construction method based on pseudo-image representation provided in the above embodiments, and will not be repeated here.

[0140] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the codebook construction method based on pseudo-image representation as described above.

[0141] The computer program product provided in this application solves the technical problem in the prior art of how to construct an efficient codebook adapted to large-scale antenna systems, thereby significantly reducing feedback overhead while maintaining high-precision channel reconstruction capabilities. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the codebook construction method based on pseudo-image representation provided in the above embodiments, and will not be elaborated upon here.

[0142] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A codebook construction method based on pseudo-image representation, characterized in that, The codebook construction method based on pseudo-image representation includes: Based on the original channel state information, a channel state mapping image is obtained; Image feature extraction is performed on the channel state mapping image to obtain a set of feature vectors; Clustering is performed based on the set of feature vectors and the channel state information, and the clustering results are used as the initial codebook. The initial codebook is input into the codebook optimization generative adversarial network model, and the cosine similarity between the codebook vectors of the initial codebook is calculated to optimize directional consistency, thereby obtaining the optimized codebook.

2. The codebook construction method based on pseudo-image representation according to claim 1, characterized in that, The step of obtaining the channel state mapping image based on the original channel state information includes: Based on the original channel state information, the spatial frequency domain channel state matrix is ​​obtained; The spatial frequency domain channel state matrix is ​​transformed to obtain a sparse angle-delay domain channel matrix. The sparse angle-delay domain channel matrix is ​​separated into a real part matrix and an imaginary part matrix; The real and imaginary matrices are mapped to their respective color channels in the initial image to obtain the channel state mapping image.

3. The codebook construction method based on pseudo-image representation according to claim 2, characterized in that, The step of performing a domain transformation on the space-frequency domain channel state matrix to obtain a sparse angle-delay domain channel matrix includes: Perform a first-dimensional inverse Fourier transform on the column vectors of the spatial frequency domain channel state matrix to convert the frequency domain information to the time delay domain, thereby obtaining the time delay domain channel matrix; Perform a second-dimensional inverse Fourier transform on the row vectors of the time-delay domain channel matrix after the first-dimensional transformation to transform the spatial domain information to the angular domain, thus obtaining the angular-time-delay domain channel matrix. The sparse angle-delay domain matrix is ​​obtained by sparsifying and pruning the angle-delay domain matrix.

4. The codebook construction method based on pseudo-image representation according to claim 1, characterized in that, The step of extracting image features from the channel state mapping image to obtain a feature vector set includes: Semantic features are extracted from the channel state mapping image to obtain a first feature vector; Local structural features are extracted from the channel state mapping image to obtain a second feature vector; The feature vector set is obtained based on the first feature vector and the second feature vector.

5. The codebook construction method based on pseudo-image representation according to claim 1, characterized in that, The step of clustering based on the feature vector set and the channel state information, and using the clustering result as the initial codebook, includes: The feature vector set is initialized with parameters to obtain multiple initial cluster centers; Based on the initial cluster centers, the feature vector set is iteratively optimized to obtain an optimized cluster center set. The optimized cluster center set is subjected to convergence judgment processing until the change in the cluster centers is less than a preset threshold or the number of iterations reaches a preset upper limit, and the iterative optimization result of the cluster centers is obtained. The iterative optimization results of the cluster centers are structured to obtain the initial codebook.

6. The codebook construction method based on pseudo-image representation according to claim 1, characterized in that, The process of inputting the initial codebook into a codebook optimization generative adversarial network model, calculating the cosine similarity between the codebook vectors of the initial codebook for directional consistency optimization, and obtaining the optimized codebook includes: Configure the initial codebook as the training parameters for the vector quantization layer in the codebook-optimized generative adversarial network model; The original channel state information is input into the encoder of the generative adversarial network to obtain the latent feature vector; Based on the vector quantization layer, the potential feature vector is quantized and mapped to obtain the quantized feature vector; The quantized feature vector is reduced in dimension by a projector and then input into a decoder for reconstruction to obtain the reconstructed channel state information. Through end-to-end joint training, the parameters of the encoder, vector quantization layer, projector and decoder are updated, and the codebook vector in the vector quantization layer is optimized synchronously during training. The optimized codebook is obtained from the vector quantization layer in the converged model.

7. The codebook construction method based on pseudo-image representation according to claim 6, characterized in that, The goal of the end-to-end joint training is to minimize the preset composite loss function; The composite loss function includes at least a reconstruction loss term and a feature fidelity loss term based on cosine similarity, wherein the feature fidelity loss term based on cosine similarity is used to constrain the directional consistency between the reconstructed channel state information and the original channel state information in the feature space.

8. A codebook construction device based on pseudo-image representation, characterized in that, The codebook construction device based on pseudo-image representation includes: The image conversion module is used to obtain a channel state mapping image based on the original channel state information; The feature extraction module is used to extract image features from the channel state mapping image to obtain a set of feature vectors; The codebook construction module is used to perform clustering based on the feature vector set and the channel state information, and use the clustering result as the initial codebook. The codebook optimization and evaluation module is used to input the initial codebook into the codebook optimization generative adversarial network model, calculate the cosine similarity between the codebook vectors of the initial codebook to perform directional consistency optimization, and obtain the optimized codebook.

9. A codebook construction device based on pseudo-image representation, characterized in that, The pseudo-image representation-based codebook construction device includes: a memory, a processor, and a pseudo-image representation-based codebook construction program stored in the memory and executable on the processor, wherein the pseudo-image representation-based codebook construction program is configured to implement the steps of the pseudo-image representation-based codebook construction method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a codebook construction program based on pseudo-image representation, which, when executed by a processor, implements the steps of the codebook construction method based on pseudo-image representation as described in any one of claims 1 to 7.

Images