Channel state information coding and decoding method based on double-channel heterogeneity

Abstract

The invention discloses a channel state information coding and decoding method based on double-channel heterogeneity. The method comprises the following steps: firstly, extracting CSI (Channel State Information) features by using a hybrid Transform-CNN (Convolutional Neural Network) compression network by a sending end, and performing unified quantization; according to complexity parameters, quantized symbols are divided into a strong context dependency class and a weak context dependency class, the strong context dependency class is subjected to ASCII discretization and then modeled by a language model and subjected to arithmetic coding, and the weak context dependency class is modeled by a factorization model and subjected to parallel coding. And a receiving end respectively decodes and inversely quantizes the two types of bit streams, and reconstructs the original CSI through a decoder. The method has the advantages of being high in compression rate, low in delay, controllable in structural complexity and the like, is suitable for low-overhead CSI feedback tasks in a large-scale MIMO system, and has wide application prospects.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically, to a dual-path heterogeneous channel state information (CSI) encoding and decoding method for massive MIMO systems, particularly a CSI feedback technique that can achieve both high compression rate and low decoding latency. Background Technology

[0002] Massive MIMO technology is one of the core technologies of fifth-generation (5G) and future wireless communication systems. By deploying a large number of antennas at the base station, it can significantly improve system capacity and spectral efficiency. However, its performance is highly dependent on the base station's accurate acquisition of downlink channel state information (CSI). In frequency division duplex (FDD) systems, since uplink and downlink channels are not reciprocal, the user equipment (UE) must estimate the downlink CSI and send it to the base station through a feedback link. With the rapid increase in the number of antennas, the data dimension of CSI expands dramatically, causing feedback overhead to become a bottleneck restricting system performance. Therefore, researching efficient CSI compression feedback techniques is crucial.

[0003] In recent years, deep learning-based CSI compression schemes have demonstrated tremendous potential. Through end-to-end neural network training, the intrinsic structure of CSI data can be learned, achieving compression performance far exceeding traditional methods. Regarding network architecture design, Convolutional Neural Networks (CNNs) are widely used due to their powerful local feature extraction capabilities and low computational complexity; however, their limited receptive field makes it difficult to capture long-range spatial correlations in the CSI matrix. The Transformer architecture, with its self-attention mechanism, can effectively model global dependencies, but its high computational complexity and memory requirements make it unsuitable for direct deployment on resource-constrained user devices.

[0004] To combine the advantages of both, the hybrid Transformer-CNN architecture emerged, aiming to balance performance and efficiency. However, most existing technologies focus on the design of the lossy compression network itself, while in the subsequent lossless entropy encoding stage, a single, complex autoregressive model (such as a language model) is usually used to pursue the ultimate compression ratio. Although this approach achieves good compression results, its decoding process must be serial, that is, each symbol is recovered one by one, resulting in a decoding latency that increases linearly with the sequence length, making it difficult to meet the requirements of low-latency communication scenarios. Summary of the Invention

[0005] This invention primarily addresses the technical problems of low encoding / decoding efficiency and high communication latency in existing technologies by providing a dual-path heterogeneous channel state information (CSI) encoding / decoding method. The core objective of this method is to significantly reduce the computational complexity and time latency of the decoding process while maintaining high compression ratios through an innovative symbol splitting and heterogeneous processing architecture, achieving a flexible and efficient trade-off between compression performance and decoding efficiency in CSI feedback.

[0006] The present invention addresses the aforementioned technical problems primarily through the following technical solution: a channel state information encoding and decoding method based on dual-path heterogeneity, comprising an encoding stage and a decoding stage; the encoding stage includes: S1: Lossy Compression and Quantization: The Channel State Information (CSI) matrix to be fed back is input into a lossy compressor based on a hybrid Transformer-CNN architecture to extract a compact continuous-valued feature representation y; then, the feature representation y is uniformly quantized with n bits to obtain a quantized symbol sequence y', where the quantization step size Δ is based on the maximum value y of the batch of feature representations. max With minimum value y min And the quantization bit number n is calculated to obtain: Δ=(y max -y min ) / (2 n -1); S2: Stream splitting and parallel coding: S21: Probability Distribution Acquisition: The quantized symbol sequence y' is input into a preset language model (LM, autoregressive) and a factorization model (FM), respectively. The language model outputs each symbol y'. j Given the preorder symbol y' <j The conditional probability distribution P under L (y' j |y' <j The factorization model outputs each symbol y' j Discrete probability distribution P based on the statistical independence assumption F (y' j ); S22: Calculate the contextual conditional entropy and discrete entropy: based on the conditional probability distribution P using the Shannon entropy formula. L (y' j |y' <j Calculate the context conditional entropy H L (y' j |y' <j Based on the Shannon entropy formula and the discrete probability distribution P, F (y' j Calculate the discrete entropy H F (y' j ), that is, HF (y' j )=-log2P F (y' j ), H L (y' j |y' <j )=-log2P L (y' j |y' <j ); S23: Stream splitting: Calculate each symbol y' j Entropy difference ΔH j =|H L (y' j |y' <j )-H F (y' j |, that is, the absolute value of the difference between the contextual entropy and the discrete entropy; Choose the entropy difference ΔH j The first C×2 from largest to smallest n The first symbol subsequence is formed by 1 symbol, and the remaining symbols form the second symbol subsequence. C is the complexity control factor, thus constructing a dual-path heterogeneous encoding path. S24: Heterogeneous Coding: For the first symbol subsequence representing strong context dependence, the first bitstream b is generated by arithmetic coding using a conditional probability distribution generated by a language model. l For the second symbol subsequence representing weak context dependence, a discrete probability distribution generated using a factorization model is used to generate the second bitstream b through arithmetic coding. f The first bitstream and the second bitstream together constitute the encoded channel state information. The decoding stage includes: S3: Receive the first bitstream and the second bitstream, and perform dual-channel heterogeneous decoding: S31: Using a parallel decoding method corresponding to the factorization model, the second bitstream b is processed. f Perform parallel arithmetic decoding to quickly recover the second symbol subsequence; S32: Using an autoregressive decoding method corresponding to the language model, process the first bitstream b l Perform serial arithmetic decoding to recover the first symbol subsequence one by one; S4: Data Reconstruction: Based on the correspondence during splitting, the decoded first and second symbol subsequences are concatenated into a complete quantized symbol sequence; the sequence is dequantized to recover the continuous value feature representation; finally, the feature representation is input into a lossy reconstructor based on a hybrid Transformer-CNN architecture to reconstruct the CSI data.

[0007] The core idea of ​​this solution is to decompose the traditional single-path encoding and decoding process into a dual-path heterogeneous architecture. The aim is to "divide and conquer": identify the simple parts (weak context dependencies) and complex parts (strong context dependencies) in the data, and match them with optimal, different (i.e., "heterogeneous") processing models. The simple parts employ a simple model that can be processed in parallel, trading speed for a small loss in compression performance; the complex parts employ a powerful serial model to ensure a high overall compression ratio. Through this design, the parallel decoding path and the serial high-precision decoding path can work together, fundamentally breaking the constraint that high compression ratios inevitably come with high decoding latency. This provides the system with a flexible, adjustable capability between performance and latency, defined by a complexity control factor C.

[0008] In the encoding phase, the lossy compressor in step S1 (such as the hybrid Transformer-CNN architecture detailed later) is responsible for extracting compact feature representations from the high-dimensional CSI data, which is the first step in achieving a high compression ratio. Quantization discretizes the continuous feature values, preparing for subsequent lossless entropy encoding. Step S2 evaluates the compressibility potential of the quantized symbols using two different models (LM and FM), then performs intelligent splitting based on the evaluation results, and finally encodes the two split paths separately.

[0009] During the decoding phase, step S3 fully utilizes the heterogeneous design of the encoding end to perform large-scale parallel decoding of the bitstream from the FM path, greatly shortening the recovery time of this part of the data; at the same time, it performs fine autoregressive decoding on the bitstream from the LM path. Step S4 is responsible for merging the decoding results of the two paths and finally recovering high-quality CSI data.

[0010] Preferably, the encoding stage further includes: generating and transmitting a Boolean indicator matrix S with the same dimension as the quantized symbol sequence y', wherein the elements S of the indicator matrix S... j Used to indicate during the decoding stage whether the corresponding symbol comes from the first part or the second part, so as to facilitate the correct splicing of the decoding results.

[0011] This scheme provides the decoding end with the necessary information to accurately assemble the decoded symbol sequence with minimal signaling overhead. The Boolean indicator matrix S is equivalent to a map, recording which path each symbol was assigned to during splitting. After receiving this map, the decoding end can accurately place the parallel-decoded symbols and serially decoded symbols back into their correct positions in the original sequence in steps S3 and S4, thus ensuring error-free reassembly of the dual-path decoding results. This is the key link that enables the closed-loop implementation of the entire scheme, and the additional overhead (1 bit per symbol) is far less than the improvement in decoding efficiency it brings.

[0012] Preferably, the lossy compressor comprises two parallel convolutional branches: The first branch includes, in sequence, a convolution-batch normalization-ReLU module with a 9×9 convolution kernel, a convolution-batch normalization-ReLU module with a 1×15 convolution kernel, and a convolution-batch normalization module with a 15×1 convolution kernel. The second branch includes a convolution-batch normalization-ReLU module with a 9×9 convolution kernel and a TransConv module; The output of the first branch is fused with the output of the second branch, and then the feature is deepened by a subsequent 1×1 convolutional module and another TransConv module. Finally, the compact continuous value feature representation is output through a fully connected layer.

[0013] This approach employs two parallel, structurally distinct convolutional branches to collaboratively extract multi-dimensional features from the CSI matrix. The first branch, utilizing a large receptive field convolution (9×9) and two directional asymmetric convolutions (1×15 and 15×1), excels at capturing spatially and frequency-related local texture features within the CSI matrix. The second branch directly combines the large receptive field convolution with the TransConv module (described in detail later), focusing on introducing global dependency modeling early on. The fusion of the outputs from both branches results in a richer feature representation that deeply understands both the local details and global structure of the CSI matrix, providing high-quality input for subsequent compression and reconstruction.

[0014] Preferably, the internal processing of the TransConv module includes: First, the input features are expanded using 1×1 convolutions. Then, the expanded channels are divided into local and global branches. The local branches are processed by at least one convolutional layer based on a 3×3 kernel and residually connected to the original input to enhance local features. The global branches are processed by a Swin Transformer module to extract long-range dependencies. Finally, the processing results of the local and global branches are concatenated, and after feature fusion using 1×1 convolutions, residually connected to the original input.

[0015] The TransConv module implements a deep fusion of CNN and Transformer within a unified module. By segmenting feature channels into local and global branches and processing them separately using convolution and Swing Transformer, it achieves simultaneous capture and efficient fusion of local texture details and global dependencies. The residual connection design ensures lossless information flow and avoids the gradient vanishing problem in deep networks. Compared to simply concatenating CNN and Transformer, this parallel, internally fused design can more efficiently balance computational complexity and feature representation capability.

[0016] Preferably, the lossy reconstructor comprises, in sequence: A fully connected layer is used to map the feature representation of the input back to the dimensions of the original CSI data; A 9×9 convolutional module and a TransConv module are used to initially fuse global and local information; At least three TCBlock modules, designed based on the ResNet residual structure and connected in series, are used to refine and reconstruct the features; Finally, a sigmoid activation function is used to ensure that the numerical range of the reconstructed CSI matrix is ​​consistent with that of the original CSI.

[0017] The goal of the lossy reconstructor is to recover the original CSI matrix from the compact feature representation with high quality. Fully connected layers are responsible for restoring the size; 9×9 convolutional layers and TransConv modules handle the initial reconstruction; multiple cascaded TCBlock modules (based on ResNet and also containing TransConv) progressively refine the reconstructed CSI and correct residuals, effectively recovering high-frequency details; the final sigmoid activation function is a crucial post-processing step, ensuring that the output CSI values ​​are constrained within a reasonable physical range, thus improving the fidelity of the reconstruction.

[0018] Preferably, the language model (LM) is a multi-layer multi-head self-attention Transformer model containing 4 stacked layers, each containing 8 attention heads, with a vocabulary size of 256 and an embedding layer dimension of 256.

[0019] The language model accurately captures long-range temporal dependencies between quantized symbols. These hyperparameters (4 layers, 8 heads, etc.) are optimally configured and validated in experiments, achieving a good performance-efficiency balance in CSI compression tasks. The model exhibits autoregressive properties, meaning that predicting the current symbol depends only on previously generated historical symbols (P(y')). j |y' <j This is the fundamental reason why it can be used for sequence-generated decoding tasks.

[0020] Preferably, the discrete probability distribution P of the factorization model (FM) is... F (y' j The global marginal probability of each quantized symbol is obtained by compressing and quantizing a large-scale CSI training dataset.

[0021] The factorization model is based on the assumption of statistical independence, assuming that the probability of each symbol's occurrence is independent of other symbols and is determined solely by its frequency of occurrence in the entire dataset. This model requires no complex context modeling, is extremely fast to compute, and because its probability distribution is static and context-independent, the corresponding bitstream can be decoded in complete parallel, which is one of the core aspects of this invention's decoding acceleration.

[0022] Preferably, the autoregressive decoding method for the first bitstream in step S32 specifically includes: Using a predefined start symbol as initial context input to the language model, predict the probability distribution of the first symbol; By combining this probability distribution with the first bit stream, the first symbol is recovered using an arithmetic decoder; The recovered first symbol is appended to the current context to form a new context, which is then used to predict the probability distribution of the second symbol. The above process is repeated iteratively until the entire first symbol subsequence is recovered.

[0023] The decoding process of the language model path is a progressive generation process: starting with a known start symbol, the model predicts the probability of the first symbol, and the decoder combines the bitstream to recover the symbol; then, this newly recovered symbol is fed back into the model, becoming part of the context for predicting the second symbol. This prediction-decoding-updating context loop ensures that each decoding step utilizes all known historical information, thus enabling the accurate recovery of symbol sequences with strong temporal dependencies.

[0024] Preferably, in step S24, before encoding using a language model (LM), the symbols in the first symbol subsequence are mapped to unique integer indices.

[0025] The purpose of this step is to convert the quantized symbols (e.g., numerical values ​​from 0 to 255) into an input format that the language model can process, namely, a token ID.

[0026] The substantial effects of this invention are: 1. Breaking through the inherent bottlenecks of performance and latency, achieving controllable complexity encoding and decoding: The core contribution of this invention lies in its innovative symbol splitting and dual-path heterogeneous processing architecture based on entropy difference evaluation. For the first time in the CSI compression field, it intelligently decomposes the decoding task into a simple, parallelizable part and a complex part that must be executed serially. This fundamentally breaks the inherent constraint of traditional solutions that pursuing high compression rates inevitably leads to high decoding latency. By adjusting the complexity control factor C, this invention allows the system to flexibly and smoothly balance compression performance and decoding efficiency according to actual business needs (such as low latency and high reliability), a significant improvement not found in existing single-architecture solutions.

[0027] 2. Improved entropy coding efficiency and further optimized rate-distortion performance: This invention is the first to apply a powerful Language Model (LM) to entropy coding of CSI quantized symbols. Compared to traditional methods that rely on simple independence assumptions for entropy coding, LM can more accurately capture long-distance dependencies between symbols, thus providing more accurate probability predictions. This allows arithmetic coding to represent information with fewer bits. Meanwhile, symbols with very weak contextual dependencies are handled by an efficient Factorization Model (FM), avoiding unnecessary model overfitting and computational overhead. This tailored coding strategy enables overall lossless compression efficiency to surpass traditional "one-size-fits-all" methods, achieving a lower final feedback bit rate while maintaining the same lossy compression quality, thus optimizing the overall rate-distortion performance of the system.

[0028] 3. Balancing advanced network architecture with practical deployment requirements, the solution possesses high practicality: In the lossy compression section, this invention employs an advanced hybrid Transformer-CNN architecture, fully leveraging the local feature extraction efficiency of CNNs and the global dependency modeling capabilities of Transformers to ensure high-precision feature extraction. Simultaneously, by introducing modules such as the lightweight Swing Transformer, computational complexity is effectively controlled. Combined with a subsequent complexity-controlled decoding mechanism, the entire solution proposed in this invention not only demonstrates excellent simulation performance but also adapts to asymmetric hardware environments ranging from resource-constrained user terminals to powerful base stations during practical deployment, showcasing high engineering practical value. Attached Figure Description

[0029] Figure 1 This is a flowchart of a channel state information encoding and decoding method based on dual heterogeneous channels according to the present invention; Figure 2 This is a schematic diagram of an encoding / decoding system structure according to the present invention; Figure 3 This is a schematic diagram of a lossy compressor and reconstructor structure based on a hybrid Transformer-CNN architecture according to the present invention; Figure 4 This is a comparison chart of the compression performance of embodiments of the present invention and CSI compression methods in recent years in indoor scenes of the COST2100 dataset; Figure 5 This is a comparison chart of the compression performance of embodiments of the present invention and CSI compression methods in recent years in outdoor scenarios on the COST2100 dataset. Detailed Implementation

[0030] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0031] Example: A channel state information encoding and decoding method based on dual heterogeneous paths, the overall process of which is as follows: Figure 1 As shown, this includes an encoding phase performed at the transmitting end (e.g., user equipment) and a decoding phase performed at the receiving end (e.g., base station). This method operates as follows: Figure 2 The encoding and decoding system shown consists of a compression module, a unified quantization module, a complexity control module, a language model (LM), a factorization probability model (FM), and an arithmetic encoder based on a hybrid Transformer-CNN architecture.

[0032] The coding phase specifically includes the following steps: S1: Lossy Compression and Quantization First, the CSI matrix H to be fed back is input into a lossy compressor, which outputs a compact, significantly reduced-dimensional continuous-valued feature representation y. This process can be represented as: y=TCC(H); Where TCC(·) represents the lossy compression operator.

[0033] Then, the feature representation y is uniformly quantized using n bits (n can be set as needed) to obtain the quantized symbol sequence y'. The quantization process is as follows: First, calculate the maximum value y of the feature of this batch of samples. max and minimum value y min And calculate the quantization step size Δ according to the formula: Δ=(y max -y min ) / (2 n -1); Then, each element in y is quantized to obtain the corresponding symbol in y': y' j =round((y j -y min ) / Δ); The round(·) function represents the round-down operation.

[0034] S2: Shunting and Parallel Encoding S21: Probability Distribution Obtaining The complete quantized symbol sequence y' is sent into two preset models simultaneously: a language model (LM) and a factorization model (FM).

[0035] LM is a multi-layer Transformer model. For each symbol y' in y' j , LM outputs a conditional probability distribution P L (y' j |y' <j ) based on all its previous symbols y'<j.

[0036] FM, whose probability distribution P F (y' j ) is the global marginal probability obtained by pre-statistics on a large amount of training data, and it is fixed for all positions j.

[0037] S22 & S23: Entropy Calculation and Shunting First, use the Shannon entropy formula to calculate the two entropies of each symbol:[[]] H F (y' j ) = -log2P F (y' j ); H L (y' j |y' <j ) = -log2P L (y' j |y' <j ); Then calculate the entropy difference ΔH j = |H L (y' j |y' <j ) - H F (y' j )|. The larger the entropy difference, the more obvious the role of context information in reducing the uncertainty of the symbol, that is, the stronger the context dependence of the symbol.

[0038] According to the complexity control factor C ∈ [0, 1], select the first C × 2 n symbols with the largest entropy difference to form the first symbol subsequence (strong dependence), and the remaining symbols form the second symbol subsequence (weak dependence). At the same time, generate a Boolean indicator matrix S. The elements of S are only 0 and 1, and the dimension is the same as the number of CSI symbols (i.e., 2 n ), so it will not bring a large overhead.

[0039] The complexity control factor C is expressed as: ; Where H and W are the width and height of y', respectively, and S j The indicator matrix is ​​for either independent tokens or context tokens, based on the conditional entropy matrix pre-estimated by LM and the discrete entropy H pre-estimated by FM. f (y j The difference and the predefined C determine this.

[0040] S24: Heterogeneous coding For the first symbol subsequence, first map its symbols to integer indices, then use the conditional probability distribution P generated by LM. L The first bitstream b is generated by an arithmetic encoder. l .

[0041] For the second symbol subsequence, the discrete probability distribution P generated directly using FM is used. F The second bitstream b is generated by an arithmetic encoder. f .

[0042] Ultimately, b l ,b f Together with the indicator matrix S, they are transmitted as encoded data.

[0043] The decoding stage specifically includes the following steps: S3: Dual-channel heterogeneous decoding The receiving end first parses the bit stream b l ,b f And the indicator matrix S.

[0044] S31: Parallel Decoding Since each symbol in the FM path is statistically independent, the decoder can initiate multiple processing units for the second bitstream b. f Parallel arithmetic decoding is performed on different segments simultaneously to quickly recover the entire second symbol subsequence q. f .

[0045] S32: Autoregressive Decoding For the first bit stream b l The decoding process must be sequential. The decoder takes a start symbol as the initial context input LM, obtains the probability distribution of the first symbol, and combines it with b. l The first symbol is recovered; then the recovered symbol is added to the context, and the second symbol is predicted. This process is repeated until the first symbol subsequence q is obtained. l It has been fully restored.

[0046] S4: Data Restructuring Based on the indicator matrix S, the two subsequences obtained from parallel decoding and serial decoding are accurately concatenated back into a complete quantized symbol sequence, i.e., y. r =R(q l ,q f ,S), where R(·) represents the splicing process, that is, when S is 1, q is taken. l When S is 0, take q. f .

[0047] Then, an inverse quantization operation is performed to recover the continuous-value feature representation y^: y^=y r ×Δ+y min +Δ / 2; Finally, the recovered y^ is input into a lossy reconstructor (decoder) based on a hybrid Transformer-CNN architecture. The reconstructor structure is as follows: Figure 3 As shown. After refinement through fully connected layers, a preliminary reconstruction module, and multiple TCBlocks, the reconstructed CSI matrix H is finally output through the Sigmoid activation function. rec , can be represented as: H rec =TCD(y^), where TCD(·) represents the reconstruction operator.

[0048] For a lossy compressor (encoder) based on a hybrid Transformer-CNN architecture, its framework is as follows: Figure 3 As shown, CSI is input into two parallel convolutional branches. One branch is designed as follows: the input passes through three cascaded convolutional modules to progressively extract features—the first module uses a 9×9 CBR unit (i.e., convolution + batch normalization + ReLU activation function), followed by a 1×15 CBR module and a 15×1 CB module (containing only convolution and batch normalization). After processing by the above convolutional layers, the data is further passed to the TransConv module to capture both global and local features simultaneously. Branch 2 consists of a 9×9 CBR unit cascaded with TransConv, and its output is fused with the output of branch 1 via a Concat operation. The fused tensor is input to a 1×1 CBR module to extract element-level features, and then processed by TransConv. Finally, based on the compression ratio, the fully connected layer generates a sequence of dimensionality-reduced feature vectors.

[0049] After the CSI feature vectors enter the decoder, they are first processed by a fully connected layer to recover the original data size. Then, a 9×9 CBR and a TransConv module are used to integrate global and local information. Subsequently, CSI reconstruction is completed through three TCBlock modules. The TCBlock modules are based on the ResNet residual design. The TransConv module has different parameter selections for the convolution and attention dimensions compared to the encoder. This is because the decoder needs to obtain the reconstructed CSI information in a higher dimension, which is beneficial for the global information reconstruction of the Transformer. Then, a 1×1 CBR is used to supplement the detailed information. After three TCBlocks, the reconstructed CSI matrix is ​​finally output through Sigmoid activation.

[0050] Trans&Conv, as an important hybrid Transformer-CNN module, employs a "W" working mode at the encoding end, with the Transformer implementation based on the SwinT Block (W-MSA); and a "SW" working mode at the decoding end, with the Transformer implementation based on the SwinT Block (SW-MSA). SwinT Block uses a windowing method instead of the Transformer's attention mechanism, which significantly reduces computational complexity and is suitable for lightweight network implementations. For the operation of TransConv, let the input be I, and the 1×1 convolution operator be Conv. 1×1 (·), then after convolution, we get I'=Conv 1×1 (I); Subsequently, I' is segmented using a split operation, with the segments used for global and local feature extraction. For TCNet-lossy, there are two segmentation scenarios: when I' has 2 channels, the segmentation method is Channel_conv=1, Channel_trans=1; When I' has 11 channels, the segmentation method is Channel_conv=5, Channel_trans=6; Where Channel_conv and Channel_trans represent the number of dimensions used for feature extraction via convolution and attention mechanisms, respectively. After splitting, the data is divided into: I cnn ,I trans =split(I'); Where I cnn and I trans These represent the dimensions for subsequent local and global feature extraction, respectively. cnn After two local feature extractions based on small convolutional kernels and superimposing the input data, we obtain: I'cnn =Relu(Conv 3×3 (Relu(Conv 3×3 (I cnn ))))+I cnn ; And I trans The SwinT module will then perform global feature extraction, resulting in: I' trans =Swin(I trans ); Where Swin(·) represents the global feature extraction of the SwinT module. The two are then concatenated with a 1×1 convolution and superimposed on the original input to obtain the final output: I out =Conv 1×1 (concat(I' trans ,I' cnn ))+I.

[0051] To demonstrate the performance of this scheme compared to the current CSI compression network, the following experiments were conducted on the COST2100 dataset: The simulation parameters for the communication system are set as follows: the base station is equipped with a uniform linear array consisting of 32 antennas. In FDD mode, the frequency domain is divided into 1024 subcarriers, and the angular resolution is set to 32. The COST2100 dataset includes 150,000 channel samples, of which 100,000 are used for training, 30,000 for validation, and 20,000 for testing.

[0052] The PyTorch network framework parameters are set as follows: Xavier is used for initialization. The Adam optimizer is used, with its hyperparameters kept at default, and the epsilon value set to 1e-8. Mean squared error is chosen as the loss function. Cosine annealing is used for learning rate scheduling, with an initial learning rate of 0.002 and a minimum learning rate of 0.00005. The model is trained for a total of 500 epochs, with the first 20 epochs serving as a warm-up period.

[0053] The Transformer language model training settings are as follows: vocabulary size is set to 256, embedding layer dimension is 256, and parameters are initialized to follow a normal distribution with a standard deviation of 0.02. The model consists of four stacked multi-head self-attention structures, each containing eight attention heads. The hidden layer width of the feedforward neural network is four times the embedding dimension, i.e., 1024 dimensions. The batch size is set to 16, meaning that 16 input sequences are processed in parallel during each training iteration.

[0054] We selected some advanced compression networks from recent years for comparison, namely: Z. Lu, J. Wang, and J. Song, “Multi-resolution csi feedback with deeplearning in massive mimo system,” in ICC 2020-2020 IEEEinternationalconference on communications (ICC). IEEE, 2020, pp. 1–6. S. Ji and M. Li, “Clnet: Complex input lightweight neural network designed for massive mimo csi feedback,” IEEE Wireless Communications Letters, vol.10, no. 10, pp. 2318–2322, 2021. J. Guo, C.-K. Wen, S. Jin, and GY Li, “Convolutional neural network based multiple-rate compressive sensing for massive mimo csi feedback: Design, simulation, and analysis,” IEEE Transactions on WirelessCommunications, vol. 19, no. 4, pp. 2827–2840, 2020. Transnet: Full attention network for csi feedback in fdd massive mimosystem,” IEEE Wireless Communications Letters, vol. 11, no. 5, pp. 903–907, 2022. Table 1 shows a comparison of the compression performance of the present invention and the methods described above in indoor scenes on the COST2100 dataset. The superior reconstruction accuracy of the present invention can be seen under various settings.

[0055] Table 1 Figure 4The diagram illustrates the BR-NMSE of this invention and other methods in a COST2100 indoor scenario. It can be seen that this invention has the best BR-NMSE trade-off and performs best under most complexity settings.

[0056] Figure 5 The diagram illustrates the BR-NMSE of this invention and other methods in an outdoor COST2100 scenario. It can be seen that this invention still has the best BR-NMSE trade-off and performs best under all complexity settings.

[0057] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0058] Although this document uses various terms extensively, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of the invention; interpreting them as any additional limitation would contradict the spirit of the invention.

Claims

1. A dual-path heterogeneous based channel state information encoding and decoding method, characterized in that, Comprising an encoding stage and a decoding stage; the encoding stage comprises: S1: lossy compression and quantization: input the channel state information matrix to be fed back, i.e. the CSI matrix, into a lossy compressor based on a hybrid Transformer-CNN architecture, and extract a compact continuous-valued feature representation y; Then, the feature representation y is uniformly quantized by n bits to obtain a quantized symbol sequence y'; S2: splitting and parallel encoding: S21: probability distribution acquisition: input the quantized symbol sequence y' into a preset language model and a factorization model respectively, the language model outputs the conditional probability distribution P(y' |y') of each symbol y' j under the given preceding symbol y' <j . L (y' j |y' <j ), and the factorization model outputs the discrete probability distribution P(y' ) based on the statistical independence assumption. j F (y' j ) of each symbol y'.​ S22: Calculate the context conditional entropy and the discrete entropy: Calculate the context conditional entropy H L (y' j | y <j ) according to the Shannon entropy formula based on the conditional probability distribution P L (y' j | y <j ) and calculate the discrete entropy H F (y' j ) according to the Shannon entropy formula based on the discrete probability distribution P F (y' j ). S23： Split: Calculate the entropy difference ΔH of each symbol y' j j = |H L (y' j |y' <j )- H F (y' j ) |;​ Choose the entropy difference ΔH j The first C×2 from largest to smallest n The first symbol subsequence is formed by 1 symbol, and the remaining symbols form the second symbol subsequence. C is the complexity control factor. S24: Isomorphic encoding: for the first subsequence of symbols representing strong context dependencies, a first bit stream b is generated by arithmetic encoding using the conditional probability distribution generated by the language model l ; For the second subsequence of symbols representing weak context dependency, a discrete probability distribution generated by a factorization model is employed to generate a second bit stream b f ; the first bit stream and the second bit stream jointly constitute the encoded channel state information; The decoding stage comprises: S3: receiving the first bit stream and the second bit stream, and performing dual-path heterogeneous decoding: S31: using a parallel decoding mode corresponding to the factorization model, performing parallel arithmetic decoding on the second bit stream b f to quickly recover a second symbol subsequence. S32: serially arithmetically decoding the first bit stream b l serially arithmetically decoding the first bit stream b S4: data reconstruction: according to the corresponding relationship during splitting, the decoded first and second symbol subsequences are spliced into a complete quantized symbol sequence; the sequence is dequantized to recover the continuous-valued feature representation; finally, the feature representation is input into a lossy reconstructor based on a hybrid Transformer-CNN architecture to reconstruct the CSI data.

2. The method of claim 1, wherein, The encoding stage also comprises generating and transmitting a Boolean indication matrix S of the same dimension as the sequence of quantized symbols y', the elements S j for indicating at the decoding stage whether the corresponding symbol is from the first part or the second part, in order to facilitate correct stitching of the decoding result.

3. The method of claim 1, wherein, The lossy compressor comprises two parallel convolution branches: The first branch comprises, in sequence, a convolution-batch normalization-ReLU module using a 9x9 convolution kernel, a convolution-batch normalization-ReLU module using a 1x15 convolution kernel, and a convolution-batch normalization module using a 15x1 convolution kernel; The second branch comprises a convolution-batch normalization-ReLU module using a 9x9 convolution kernel and a TransConv module; The output of the first branch is fused with the output of the second branch, and is subjected to subsequent 1x1 convolution modules and another TransConv module for feature deepening, and finally outputs the compact continuous-valued feature representation through a fully connected layer.

4. The method of claim 3, wherein, The lossy reconstructor comprises, in sequence: A fully connected layer for mapping the input feature representation back to the size of the original CSI data; A 9x9 convolution module and a TransConv module for preliminary fusion of global and local information; At least three TCBlock modules based on the ResNet residual structure design, which are connected in sequence, for fine reconstruction of features; Finally, a Sigmoid activation function is used to ensure that the reconstructed CSI matrix has the same numerical range as the original CSI.

5. The method of claim 4, wherein, The internal processing process of the TransConv module comprises: First, the input features are expanded in channels through 1x1 convolution; then the expanded channels are divided into local branches and global branches; the local branches are processed through at least one convolution layer based on a 3x3 convolution kernel, and are connected in residual with the original input; the global branches are processed through a Swin Transformer module; finally, the processing results of the local branches and the global branches are spliced, and are connected in residual with the original input after feature fusion through 1x1 convolution.

6. The dual-path heterogeneous-based channel state information encoding and decoding method of claim 1 or 3, wherein, The language model is a multi-layer multi-head self-attention Transformer model comprising 4 stacked layers, each layer comprising 8 attention heads, with a vocabulary size of 256 and an embedding layer dimension of 256.

7. The method of claim 1, wherein, the discrete probability distribution P of the factorization model F (y' j ) is the global marginal probability of each quantized symbol obtained by compressing and quantizing a large-scale CSI training dataset.

8. The method of Claim 6, wherein, The autoregressive decoding mode of the first bit stream in step S32 comprises: a predefined starting symbol is input to the language model as an initial context to predict a probability distribution of a first symbol; a first symbol is recovered from the first bit stream by an arithmetic decoder in combination with the probability distribution; the recovered first symbol is appended to a current context to form a new context and used to predict a probability distribution of a second symbol; the above process is iteratively performed until all the first symbol subsequence is recovered.

9. The method of claim 1, wherein, The step S24 further comprises mapping the symbols in the first symbol subsequence to unique integer indexes before encoding using the language model.

Images