Channel state information feedback method and device based on adaptive feedback rate

Abstract

The invention provides a channel state information feedback method and device based on an adaptive feedback rate. The method comprises the following steps: acquiring a channel state information matrix; inputting the sequence into a pre-training marking encoder based on an attention mechanism, and converting the sequence into a symbol sequence with a fixed length; quantizing the symbol sequence into a discrete index sequence through a pre-training vector quantizer and a codebook; determining a target length according to a real-time channel condition or a dynamically allocated system feedback resource based on a preset importance sorting rule obtained by nesting and discarding training, truncating an index sequence arranged according to an importance descending order, and retaining a front-end index to generate a variable-length feedback code word; and sending the variable-length code word to a base station for reconstruction. According to the method, the dynamic self-adaptive optimization of the feedback rate is realized, the spectrum efficiency is effectively improved, the progressive reconstruction is supported by utilizing the inherent importance sorting of the code words, and the robustness of the system is remarkably enhanced.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication and deep learning technology, and in particular to a channel state information feedback method and apparatus based on adaptive feedback rate. Background Technology

[0002] In frequency division duplex massive multiple-input multiple-output (MIMO) communication systems, base stations rely on channel state information (CSI) fed back by user equipment for precise beamforming and resource scheduling to improve system capacity and transmission reliability. However, with the rapid increase in antenna size, direct feedback from high-dimensional CSI matrices will consume a large amount of valuable uplink resources, thus requiring efficient compression.

[0003] Existing CSI feedback schemes mainly follow two technical paths. One is compression methods based on traditional signal processing, and the other is methods based on deep learning. The latter has become mainstream due to its powerful feature learning capabilities, with autoencoder-based architectures being particularly typical: an encoder is deployed on the user side for compression, and a decoder is deployed on the base station side for reconstruction. However, these data-driven models generally face severe generalization problems—when the channel characteristics of the actual deployment environment are inconsistent with the distribution of the training data, their feedback performance will significantly decrease.

[0004] To improve generalization capabilities, the industry has proposed a hybrid approach combining rules and learning. For example, one existing solution employs a manually rule-based importance element selection algorithm. On the user side, key elements and their positions in the CSI matrix are selected and fed back. On the base station side, a pre-trained generator model is used to complete the entire CSI matrix based on the received partial elements. While this approach alleviates the over-reliance on training data to some extent, it still has significant limitations: First, the amount of information fed back is a pre-set fixed value, which cannot be dynamically adjusted according to time-varying channel conditions. This results in insufficient bandwidth utilization for higher accuracy when the channel is good, and potential waste of scarce resources when the channel is poor. Second, its assessment of information importance relies entirely on pre-set manual features such as amplitude and gradient, making it difficult to optimally capture the complex global semantic structure in the CSI and limiting compression efficiency. Third, the base station-side generator receives a sparse matrix with only one or the other elements, failing to utilize the continuity differences between different important elements to more finely guide the reconstruction process.

[0005] Therefore, there is an urgent need in the field for a CSI feedback scheme that can adapt to channel conditions, learn semantic importance in a data-driven manner, and support progressive reconstruction, so as to achieve the optimal balance between feedback overhead and reconstruction accuracy in dynamic wireless environments. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a channel state information feedback method and apparatus based on adaptive feedback rate to solve the problems of fixed feedback rate, reliance on manual rules for importance judgment, and inability to achieve progressive and refined reconstruction in the prior art.

[0007] One aspect of the present invention provides a channel state information feedback method based on adaptive feedback rate, the method comprising the following steps: Obtain the channel state information matrix to be fed back; The channel state information matrix is ​​input into a pre-trained tokenized encoder to obtain a symbol sequence; wherein, the tokenized encoder converts the spatial features of the channel state information matrix into symbol representations with contextual semantic information through attention operations, and each symbol representation corresponds to a feature vector; The symbol sequence is input to a pre-trained vector quantizer, and the symbol sequence is quantized into a discrete index sequence in combination with a pre-trained codebook; wherein, the vector quantizer maps each feature vector to the nearest codeword vector in the codebook and outputs the corresponding codeword index, and the codebook contains multiple codeword vectors; Based on a preset importance ranking rule, the discrete index sequence is adaptively truncated to generate variable-length feedback codewords. The importance ranking rule specifies that the codeword indices in the discrete index sequence are arranged from highest to lowest importance for channel state information reconstruction. The adaptive truncation process includes: determining the target feedback length based on real-time channel conditions or dynamically allocated system feedback resources; retaining the first codeword indices in the discrete index sequence whose number equals the target feedback length; and discarding subsequent codeword indices, so that the length of the feedback codewords can dynamically match the current channel state or resource constraints. The variable-length feedback codeword is transmitted to the base station via the uplink wireless channel, so that the base station can use the corresponding decoder to reconstruct the channel state information based on the variable-length feedback codeword.

[0008] In some embodiments of the present invention, the tokenization encoder is a neural network encoder based on an attention mechanism, used to convert the channel state information matrix into a symbol sequence with a fixed length, wherein the fixed length is a preset positive integer.

[0009] In some embodiments of the present invention, the vector quantizer determines the nearest codeword vector by calculating the Euclidean distance between the feature vector and the codeword vector in the codebook, and outputs its index; each codeword index in the discrete index sequence is an integer between 1 and M, where M is the total number of codeword vectors in the pre-trained codebook.

[0010] In some embodiments of the present invention, the pre-trained tokenized encoder, the pre-trained vector quantizer, the pre-trained codebook, and the preset importance ranking rule are obtained through the following joint training process, which includes the following steps: Obtain a training dataset, which contains multiple real channel state information matrix samples; Construct an initial model to be trained, the initial model comprising a tokenized encoder, a vector quantizer and a decoder connected in sequence, the vector quantizer being associated with the codebook to be trained; Iterative training is performed on the samples in the training dataset, and each iteration includes the following steps: The sample is input into the initial model and processed by the tokenization encoder and vector quantizer to obtain the corresponding discrete index sequence; Perform nested discard operations, sample cutoff points from a predefined distribution, and truncate the discrete index sequence according to the cutoff points; input the truncated sequence into the decoder, and output a reconstructed channel state information matrix; A reconstruction loss function is constructed based on the deviation between the reconstructed channel state information matrix and the actual channel state information matrix. The reconstruction loss function is minimized, and the error is backpropagated using the gradient descent algorithm to update the trainable parameters of the tokenized encoder, the vector quantizer, the codebook, and the decoder. Repeat the iterative training steps until the model converges or reaches the preset stopping condition to obtain the trained tokenized encoder, vector quantizer, codebook and decoder; wherein, after training, the discrete index sequence output by the vector quantizer is automatically arranged in descending order of information importance, and this arrangement rule is used as the preset importance ranking rule.

[0011] In some embodiments of the present invention, the predefined distribution is a geometric distribution, and the probability of sampling to the cutoff point is calculated as follows: ; in, This represents the probability of sampling a specific truncated index b in a single training iteration; This represents the attenuation rate parameter, and .

[0012] In some embodiments of the present invention, the reconstruction loss function uses mean squared error, and the calculation formula is: ; in, Indicates the reconstruction loss value; For training batch size; It is the first A sample of a real channel state information matrix; This represents the channel state information matrix reconstructed based on the first b indices corresponding to the nth sample; This represents the square of the L2 norm.

[0013] In some embodiments of the present invention, determining the target feedback length based on real-time channel conditions specifically includes: Measure downlink channel quality indication; Based on the preset mapping relationship between channel quality and feedback length, the target feedback length corresponding to the channel quality indication is determined.

[0014] On the other hand, the present invention also provides a channel state information feedback device based on adaptive feedback rate, including a processor, a memory, and a computer program or instructions stored in the memory, wherein the processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the method described in any of the above.

[0015] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the method as described in any of the preceding claims.

[0016] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0017] This invention provides a channel state information feedback method and apparatus based on adaptive feedback rate. The method is executed by the user equipment and includes the following steps: acquiring the channel state information matrix to be fed back; using a pre-trained tokenized encoder based on an attention mechanism to convert the spatial features of the high-dimensional matrix into a symbol sequence of fixed length, where each symbol corresponds to a feature vector that integrates local and global contextual semantics; using a pre-trained vector quantizer and its associated pre-trained codebook, matching the nearest codeword vector for each symbol feature vector through a distance metric and outputting its index, thereby mapping the continuous symbol sequence to a discrete index sequence; based on a preset importance ranking rule obtained through a nested dropout training strategy, which ensures that the codeword indices in the discrete index sequence are strictly arranged in descending order of their importance to reconstruction, adaptively determining the target feedback length according to the channel quality measured in real time or the system feedback resources dynamically allocated by the base station, and truncating the ordered index sequence, retaining the corresponding number of indices at the beginning to generate variable-length feedback codewords; finally, transmitting the variable-length feedback codewords to the base station through the uplink wireless channel, whereby the base station uses the decoder corresponding to the user equipment side to reconstruct the signal, thereby completing the complete link from compressed feedback to high-precision channel recovery.

[0018] Compared with existing technologies, this invention enables dynamic adaptive adjustment of the feedback rate based on channel conditions, overcoming the inefficiency of fixed feedback schemes in time-varying channels. Through end-to-end nested dropout training, this invention allows the network to automatically learn the optimal, importance-ordered semantic compression representation for the reconstruction task, replacing the evaluation method that relies on manually designed features, thereby achieving compression results with higher information density and greater generalization ability. Thanks to the inherent structure of the feedback codewords being strictly arranged in descending order of importance, this invention endows the base station with the ability for progressive reconstruction, enabling effective reconstruction even when only partial codewords are received, significantly improving the system's robustness to channel errors and resource fluctuations. Furthermore, this invention employs a lightweight model design on the user equipment side, achieving better compression efficiency while balancing deployment complexity and generalization.

[0019] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0020] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0021] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0022] Figure 1 This is a flowchart illustrating the channel state information feedback method based on adaptive feedback rate according to an embodiment of the present invention.

[0023] Figure 2 This is a flowchart of the channel state information feedback method based on adaptive feedback rate according to another embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the overall architecture and training process of the channel state information feedback method based on adaptive feedback rate according to another embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0026] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0027] In the fields of wireless communication and deep learning technology, existing channel state information (CSI) feedback techniques mainly rely on fixed compression rate coding or hybrid schemes based on manual rules. Facing the core challenges of time-varying channel conditions, high-dimensional CSI matrices, and limited uplink resources in large-scale MIMO systems, existing solutions have three main limitations in practical deployment: First, the amount of feedback information is fixed and cannot be dynamically adjusted according to real-time channel quality, resulting in low resource utilization efficiency; second, importance assessment relies on preset manual features such as amplitude and local variance, making it difficult to effectively capture the complex global semantic structure in CSI and limiting compression performance; third, the feedback data lacks an inherent importance gradient, preventing progressive reconstruction at the base station and limiting the system's robustness and flexibility.

[0028] In view of this, one aspect of the present invention provides a channel state information feedback method based on adaptive feedback rate, such as... Figure 1 As shown, the method includes the following steps S101~S105: S101: Obtain the channel state information matrix to be fed back.

[0029] S102: Input the channel state information matrix into the pre-trained tokenized encoder to obtain a symbol sequence; wherein, the tokenized encoder converts the spatial features of the channel state information matrix into symbol representations with contextual semantic information through attention operations, and each symbol representation corresponds to a feature vector.

[0030] S103: Input the symbol sequence into a pre-trained vector quantizer and combine it with the pre-trained codebook to quantize the symbol sequence into a discrete index sequence; wherein, the vector quantizer maps each feature vector to the nearest codeword vector in the codebook and outputs the corresponding codeword index, and the codebook contains multiple codeword vectors.

[0031] S104: Based on a preset importance ranking rule, the discrete index sequence is adaptively truncated to generate variable-length feedback codewords; wherein, the importance ranking rule stipulates that the codeword indices in the discrete index sequence are arranged from high to low according to their importance to the reconstruction of channel state information; the adaptive truncation process includes: determining the target feedback length according to the real-time channel conditions or dynamically allocated system feedback resources, retaining the codeword indices that are ranked first in the discrete index sequence and whose number is equal to the target feedback length, and discarding the subsequent codeword indices, so that the length of the feedback codeword can dynamically match the current channel state or resource constraints.

[0032] S105: The variable-length feedback codeword is sent to the base station via the uplink wireless channel so that the base station can use the corresponding decoder to reconstruct the channel state information based on the variable-length feedback codeword.

[0033] In step S101, the user equipment (UE) receives and measures the downlink reference signal transmitted by the base station, and calculates and obtains the current downlink channel state information matrix. The downlink reference signal is a signal with a known structure transmitted periodically or on demand by the base station, used by the UE for channel estimation. The UE calculates the channel frequency response by comparing the received signal with the known reference signal. This matrix is ​​a high-dimensional complex matrix, its dimensions determined by the number of base station transmit antennas and the number of UE receive antennas, fully characterizing the amplitude and phase response of a multi-antenna channel.

[0034] In step S102, the user equipment inputs the channel state information matrix into a pre-trained tokenized encoder. This encoder is a neural network encoder based on an attention mechanism. Its core technology utilizes an internal self-attention mechanism to calculate the correlation weights between any pairs of elements in the input matrix, thereby dynamically aggregating global contextual information. This process transforms the channel features in high-dimensional space into a one-dimensional symbol sequence containing rich semantic relationships.

[0035] The symbol sequence output by the aforementioned tokenization encoder has a fixed length K, where K is a preset positive integer. Each symbol in the sequence corresponds to a multi-dimensional feature vector, which is a compact representation of the local features of the original matrix after global context-weighted fusion.

[0036] In some embodiments, the tokenization encoder is a Transformer-based encoder. This architecture, through stacked multi-head self-attention layers and position-feedforward network layers, is able to more efficiently capture complex dependencies in the channel matrix and generate high-quality feature representations.

[0037] In step S103, the user equipment inputs the symbol sequence into a pre-trained vector quantizer and combines it with a pre-trained codebook to quantize the symbol sequence into a discrete index sequence.

[0038] The core technique in this step lies in discretization compression. The pre-trained codebook stores M codeword vectors, each a vector with the same dimension as the symbol feature vector. For each symbol feature vector in the sequence, the vector quantizer calculates its distance to every codeword vector in the codebook and selects the closest one. The quantizer then outputs the index number of the selected codeword vector in the codebook. After processing the entire symbol sequence, the original continuous symbol sequence is transformed into an index sequence composed of these discrete indices.

[0039] In some embodiments, the vector quantizer determines the nearest codeword vector by calculating the Euclidean distance between the feature vector and the codeword vector. Each codeword index in the discrete index sequence is an integer between 1 and M, where M is the total number of codeword vectors in the pre-trained codebook.

[0040] It is important to emphasize that the pre-trained tokenized encoder and vector quantizer are obtained through a joint training process that includes a nested dropout strategy (see the subsequent training method description for details). The core objective of this training process is to drive the model to learn to encode the most critical information needed to reconstruct the CSI at the beginning of the sequence. Therefore, after the model training is complete, when performing the inference processes S102 and S103 described above, its output discrete index sequence They have been automatically sorted strictly from highest to lowest importance according to the information represented by the index, among which, Most importantly, The least important. This inherent ordering property forms the basis for subsequent adaptive truncation.

[0041] In step S104, the user equipment adaptively truncates the discrete index sequence Z, which is sorted in descending order of importance, to generate variable-length feedback codewords. The importance sorting rule is reflected in the inherent order of sequence Z; that is, the codeword indices are strictly arranged in descending order of their importance for reconstructing channel state information at the base station.

[0042] In some embodiments, the pre-trained tokenized encoder, vector quantizer, codebook, and importance ranking rule are obtained through a joint training process. This process includes: Obtain a training dataset containing multiple real channel state information matrix samples; Construct an initial model consisting of a tokenized encoder, a vector quantizer, and a decoder connected in sequence, wherein the vector quantizer is associated with the codebook to be trained; Iterative training is performed using the training data. In each iteration, after inputting a sample into the model to obtain a discrete index sequence, a nested discard operation is performed: a cutoff point is sampled from a predefined probability distribution, the sequence is truncated based on this point, and only the truncated sequence is input into the decoder to reconstruct the channel matrix; subsequently, a reconstruction loss function is constructed based on the deviation between the reconstructed matrix and the true sample matrix, and this loss is minimized using the gradient descent algorithm to update the parameters of all components in the model. This process is repeated until the model converges. After training, the model naturally learns to encode key information at the beginning of the sequence, thereby forming an importance ranking rule.

[0043] In some embodiments, the nested discard operation uses a predefined distribution, namely a geometric distribution. The probability of sampling the truncated index b is calculated as follows: ; in, This represents the probability of sampling a specific truncated index b in a single training iteration; This represents the attenuation rate parameter, and .

[0044] This distribution ensures sufficient learning of different cutoff depths during training.

[0045] In some embodiments, the reconstruction loss function uses the mean squared error, calculated as follows: ; in, Indicates the reconstruction loss value; For training batch size; It is the first A sample of a real channel state information matrix; This represents the channel state information matrix reconstructed based on the first b indices corresponding to the nth sample; This represents the square of the L2 norm.

[0046] Based on this sequence with inherent importance, the user equipment performs an adaptive truncation operation. First, the user equipment determines a target feedback length L based on real-time channel conditions or dynamically allocated system feedback resources.

[0047] In some embodiments, the user equipment determines L based on real-time channel conditions. Specifically, it measures the downlink channel quality indicator; then, it looks up a pre-configured channel quality-to-feedback length mapping table to determine the target feedback length L corresponding to the current measurement value.

[0048] In other embodiments, the user equipment determines L based on dynamically allocated system feedback resources. Specifically, it receives the uplink feedback resource block size configured by the base station through downlink control signaling, and calculates L based on the resource size and the number of encoded bits for each codeword index.

[0049] After determining L, the user equipment retains the first L codeword indices in the discrete index sequence to form a variable-length feedback codeword, and discards the subsequent indices. This operation allows the feedback overhead to dynamically match the current communication environment and resource constraints.

[0050] In step S105, the user equipment sends the variable-length feedback codeword to the base station through the uplink wireless channel, thereby completing the adaptive feedback process of channel state information.

[0051] The core technical operation in this step is physical layer data transmission. Before actual transmission, the variable-length feedback codeword undergoes channel coding, such as forward error correction coding and modulation, to enhance its reliability in unreliable wireless channels. Subsequently, the user equipment transmits the processed signal to the base station through the physical uplink shared channel or a specific uplink control channel. This process strictly follows the uplink transmission protocol of the wireless communication system.

[0052] After receiving the signal, the base station demodulates and decodes the channel to recover the variable-length feedback codeword. Then, the base station uses a decoder that perfectly corresponds to the user equipment's pre-trained model to process the received codeword. The decoder, based on the received codeword index, queries the same pre-trained codebook as the user equipment side to recover the corresponding symbol sequence and ultimately reconstruct the channel state information matrix.

[0053] On the other hand, the present invention also provides a channel state information feedback device based on adaptive feedback rate, including a processor, a memory, and a computer program or instructions stored in the memory. The processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of any of the methods described above.

[0054] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of any of the methods described above.

[0055] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of any of the above methods.

[0056] The present invention will now be described with reference to a specific embodiment: This embodiment elaborates on the complete technical solution of a channel state information feedback method based on adaptive feedback rate, specifically from the aspects of system principle, core processing flow, model training method, and online adaptive feedback mechanism. The specific technical solution followed in this embodiment is as follows: 1. Basic Principles The technical solution of this invention relates to a method and apparatus based on supporting adaptive channel state information (CSI) feedback rate. This solution processes CSI information through a nested dropout layer, and utilizes information importance ranking and rate control to achieve variable-length CSI feedback, thereby effectively improving the feedback efficiency and transmission reliability of the communication system.

[0057] like Figure 2As shown, the overall process is as follows: the original CSI matrix is ​​converted into a token sequence through tokenization, then mapped to the codebook space through vector quantization, and then the semantic importance is sorted through a nested Dropout layer. Finally, it is adaptively truncated according to the rate requirement to generate variable-length feedback codewords.

[0058] 2. Detailed technical solution To achieve the above objectives, the present invention adopts the following technical solution: Step 1: CSI Matrix Symbolization: At the user equipment (UE) end, the original Channel State Information (CSI) matrix is ​​symbolized. The input is fed into an attention-based tokenizer. This tokenizer transforms the high-dimensional CSI matrix into a token sequence with contextual semantic information. ,in This is a fixed-length sequence. This process transforms the data from matrix form to sequence form, facilitating subsequent processing.

[0059] Step 2: Vector quantization of the token sequence: Quantize the symbol sequence obtained in Step 1... Input vector quantizer. This codebook. Stored Each codeword vector Each codeword vector is a symbol feature vector. Real vectors of the same dimension. For each symbol vector in the sequence. The vector quantizer calculates its value relative to the codebook. All codeword vectors in The distance between them is used to find the codeword vector that is closest to them. Then, the quantizer outputs the index number of this codeword vector in the codebook, denoted as . After traversing and processing the entire sequence, the original continuous sequence of symbols is... It is then converted into a discrete index sequence , where each index Both are in a middle arrive Integers between [a certain value]. The core function of this step is discretization compression: by mapping high-dimensional continuous features to a finite-size [equation]. The discrete codebook space greatly eliminates information redundancy and significantly reduces the amount of data to be fed back. At the same time, the discrete index representation also enhances the error resistance of the feedback information during transmission.

[0060] Step 3: Semantic importance ranking based on nested Dropout: This step is the core of the training process and aims to force the model to learn an ordered representation, that is, the tokens at the beginning of the sequence contain more important information.

[0061] (1) Random truncation: In each training iteration, from a predefined distribution Sampling a truncated index This distribution applies to all indices. All have support (i.e.) This invention employs a geometric distribution, and the calculation formula is: ; in, This represents the probability of sampling a specific truncated index b in a single training iteration; This represents the attenuation rate parameter, and .

[0062] (2) Forward computation: based on the sampled Discrete sequence Cut off, keeping only the first part Each token forms a truncated representation. .

[0063] (3) Reconstruction and loss calculation: The truncated representation Input decoder to reconstruct CSI matrix Calculate the reconstruction loss for this iteration. The loss function uses the mean squared error, and the formula is: ; in, Indicates the reconstruction loss value; For training batch size; It is the first A sample of a real channel state information matrix; This represents the channel state information matrix reconstructed based on the first b indices corresponding to the nth sample; This represents the square of the L2 norm.

[0064] (4) Parameter update: Update the model parameters by minimizing the expected reconstruction loss through the optimizer. The formula for calculating the expected reconstruction loss is: ; in, Indicates the expected reconstruction loss; Indicates a truncated index In its probability distribution The expected value of the following; This represents the truncation index of the random sample taken in each training iteration; This indicates the predefined probability distribution followed by the truncated index; This indicates that the index is truncated. The reconstruction losses at that time; This represents the set of model parameters to be optimized, including trainable parameters in the tokenized encoder, vector quantizer, codebook, and decoder.

[0065] Minimize the above-mentioned expected reconstruction loss using an optimization algorithm. For model parameters Iterative updates drive the model to prioritize encoding the most critical information needed to reconstruct channel state information at the beginning of the sequence, thereby ensuring that the output discrete index sequence forms a strictly descending order of semantic importance, i.e., the index... The most important information is the corresponding index. The corresponding information is the least important.

[0066] The above nested Dropout training process is as follows: Figure 3 As shown in the "Training Phase" section: After the training data is labeled and quantized, it is processed according to the distribution in each iteration. Random truncation, using only the truncated sequence to calculate the reconstruction loss. And update the model parameters accordingly. This process is repeated until a fully trained model is obtained.

[0067] Step 4: CSI Encoding and Sequence Sorting (Inference Phase): After model training is complete, the inference phase begins. The user device acquires the new CSI matrix and sequentially performs CSI matrix tokenization, token sequence vector quantization, and semantic importance sorting to obtain the discrete token sequence. At this point, the sequence has been ordered from highest to lowest importance. arrive arrangement.

[0068] Step 5: Adaptive Rate Control and Feedback: The user equipment or base station determines a target feedback length based on the current channel conditions or the feedback rate required by the system. ( ). For the sorted sequence Truncate the section, keeping only the first part. The most important token, discarded after Each token generates a variable-length feedback code. The user equipment then sends this variable-length codeword back to the base station.

[0069] Step 6: CSI Reconstruction: The base station receives variable-length codewords Then, the CSI matrix is ​​reconstructed using a decoder corresponding to the user equipment. .

[0070] The apparatus for implementing the above method can be deployed in user equipment or base stations for uplink or downlink channel status information feedback. The apparatus includes a processor, memory, and a communication interface.

[0071] (1) The processor is configured to call and execute a computer program stored in memory to control the various modules of the device to work together to implement the method described in steps 1 to 6 above.

[0072] (2) The communication interface is used to transmit data with the peer communication equipment (such as between user equipment and base station), especially to send variable length CSI feedback codewords and receive related control signaling.

[0073] (3) The memory stores computer programs and pre-trained codebooks. and model parameters .

[0074] In terms of functional modules, the device specifically includes: (1) Tokenization module: Coupled with the processor, it performs step 1. It receives the original CSI matrix. Furthermore, by employing an attention-based sequence modeling structure, the CSI matrix is ​​transformed into a token sequence with semantic information. .

[0075] (2) Vector quantization module: Coupled with the tokenization module and the processor, it is used to execute step 2. It receives the token sequence. The sequence is discretized into an index sequence by querying a codebook pre-stored in memory. .

[0076] (3) Nested Dropout sorting module: Coupled with the vector quantization module and the processor, this module performs step 3 during the training phase. During training, this module employs a nested dropout strategy, by sorting the data from a geometric distribution... Mid-sample truncated index and only use the previous Each token is reconstructed and mean squared error loss is calculated, forcing the representation layer units of the network to be arranged in order of information importance.

[0077] (4) Rate Control Module: Coupled with the nested Dropout sorting module (during the inference phase) or the vector quantization module (during post-training inference) and the processor, it performs step 5. It dynamically determines the truncation length based on the channel conditions or target feedback rate obtained through the communication interface. The sorted discrete token sequence is truncated, low-importance tokens are discarded, and variable-length CSI feedback codewords are generated. And send it through the communication interface.

[0078] The deployment and collaboration data flow of the above functional modules are as follows: Figure 3 As shown in the "User Equipment Side" and "Base Station Side" sections: The original CSI matrix is ​​processed sequentially by the tokenization module, vector quantization module, nested Dropout sorting module, and rate control module, outputting variable-length feedback codewords. The CSI decoder on the base station side receives the codeword and reconstructs the CSI matrix. .

[0079] The aforementioned functional modules are interconnected through the processor's control and system bus, working together to complete the entire process from CSI compression and sorting to adaptive feedback.

[0080] In summary, the present invention provides a channel state information feedback method and apparatus based on adaptive feedback rate. This method, executed by the user equipment, includes the following steps: acquiring the channel state information matrix to be fed back; using a pre-trained tokenized encoder based on an attention mechanism to convert the spatial features of the high-dimensional matrix into a fixed-length symbol sequence, where each symbol corresponds to a feature vector that integrates local and global contextual semantics; using a pre-trained vector quantizer and its associated pre-trained codebook, matching the nearest codeword vector for each symbol feature vector through a distance metric and outputting its index, thereby mapping the continuous symbol sequence to a discrete index sequence; based on a preset importance ranking rule obtained through a nested dropout training strategy, which ensures that the codeword indices in the discrete index sequence are strictly arranged in descending order of their importance to reconstruction, adaptively determining the target feedback length according to the channel quality measured in real time or the system feedback resources dynamically allocated by the base station, and truncating the ordered index sequence, retaining the corresponding number of indices at the beginning to generate variable-length feedback codewords; finally, transmitting the variable-length feedback codewords to the base station through the uplink wireless channel, where the base station uses the decoder corresponding to the user equipment side for reconstruction, thus completing the complete link from compressed feedback to high-precision channel recovery.

[0081] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0082] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0083] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A channel state information feedback method based on adaptive feedback rate, characterized in that, The method includes the following steps: Obtain the channel state information matrix to be fed back; The channel state information matrix is ​​input into a pre-trained tokenized encoder to obtain a symbol sequence; wherein, the tokenized encoder converts the spatial features of the channel state information matrix into symbol representations with contextual semantic information through attention operations, and each symbol representation corresponds to a feature vector; The symbol sequence is input to a pre-trained vector quantizer, and the symbol sequence is quantized into a discrete index sequence in combination with a pre-trained codebook; wherein, the vector quantizer maps each feature vector to the nearest codeword vector in the codebook and outputs the corresponding codeword index, and the codebook contains multiple codeword vectors; Based on a preset importance ranking rule, the discrete index sequence is adaptively truncated to generate variable-length feedback codewords. The importance ranking rule specifies that the codeword indices in the discrete index sequence are arranged from highest to lowest importance for channel state information reconstruction. The adaptive truncation process includes: determining the target feedback length based on real-time channel conditions or dynamically allocated system feedback resources; retaining the first codeword indices in the discrete index sequence whose number equals the target feedback length; and discarding subsequent codeword indices, so that the length of the feedback codewords can dynamically match the current channel state or resource constraints. The variable-length feedback codeword is transmitted to the base station via the uplink wireless channel, so that the base station can use the corresponding decoder to reconstruct the channel state information based on the variable-length feedback codeword.

2. The channel state information feedback method based on adaptive feedback rate according to claim 1, characterized in that, The tokenization encoder is a neural network encoder based on an attention mechanism, used to convert the channel state information matrix into a symbol sequence with a fixed length, wherein the fixed length is a preset positive integer.

3. The channel state information feedback method based on adaptive feedback rate according to claim 1, characterized in that, The vector quantizer determines the nearest codeword vector by calculating the Euclidean distance between the feature vector and the codeword vectors in the codebook, and outputs its index. Each codeword index in the discrete index sequence is an integer between 1 and M, where M is the total number of codeword vectors in the pre-trained codebook.

4. The channel state information feedback method based on adaptive feedback rate according to claim 1, characterized in that, The pre-trained tokenized encoder, the pre-trained vector quantizer, the pre-trained codebook, and the preset importance ranking rule are obtained through the following joint training process, which includes the following steps: Obtain a training dataset, which contains multiple real channel state information matrix samples; Construct an initial model to be trained, the initial model comprising a tokenized encoder, a vector quantizer and a decoder connected in sequence, the vector quantizer being associated with the codebook to be trained; Iterative training is performed on the samples in the training dataset, and each iteration includes the following steps: The sample is input into the initial model and processed by the tokenization encoder and vector quantizer to obtain the corresponding discrete index sequence; Perform nested discard operations, sample cutoff points from a predefined distribution, and truncate the discrete index sequence according to the cutoff points; input the truncated sequence into the decoder, and output a reconstructed channel state information matrix; A reconstruction loss function is constructed based on the deviation between the reconstructed channel state information matrix and the actual channel state information matrix. The reconstruction loss function is minimized, and the error is backpropagated using the gradient descent algorithm to update the trainable parameters of the tokenized encoder, the vector quantizer, the codebook, and the decoder. Repeat the iterative training steps until the model converges or reaches the preset stopping condition to obtain the trained tokenized encoder, vector quantizer, codebook and decoder; wherein, after training, the discrete index sequence output by the vector quantizer is automatically arranged in descending order of information importance, and this arrangement rule is used as the preset importance ranking rule.

5. The channel state information feedback method based on adaptive feedback rate according to claim 4, characterized in that, The predefined distribution is a geometric distribution, and the probability of sampling to the cutoff point is calculated as follows: ； in, This represents the probability of sampling a specific truncated index b in a single training iteration; This represents the attenuation rate parameter, and .

6. The channel state information feedback method based on adaptive feedback rate according to claim 4, characterized in that, The reconstruction loss function uses the mean squared error, and its calculation formula is: ； in, Indicates the reconstruction loss value; For training batch size; It is the first A sample of a real channel state information matrix; This represents the channel state information matrix reconstructed based on the first b indices corresponding to the nth sample; This represents the square of the L2 norm.

7. The channel state information feedback method based on adaptive feedback rate according to claim 1, characterized in that, The determination of the target feedback length based on real-time channel conditions specifically includes: Measure downlink channel quality indicators; Based on the preset mapping relationship between channel quality and feedback length, the target feedback length corresponding to the channel quality indication is determined.

8. A channel state information feedback device based on adaptive feedback rate, comprising a processor, a memory, and a computer program or instructions stored in the memory, characterized in that, The processor is configured to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

Images