Time correlation based generative csi feedback method and device, system, storage medium

Abstract

The application discloses a time correlation-based generative CSI feedback method and device, system and storage medium, and adopts a two-stage history-aware generative CSI feedback of "predict, encode, and then feedback", wherein "time correlation" is only used to reduce the encoding bit number of compressed CSI, instead of being directly used for reconstructing current CSI. The application can reduce feedback overhead by utilizing time correlation, while maintaining robustness to UE motion mutation.

Description

Technical Field

[0001] This invention belongs to the technical field of large-scale MIMO systems, specifically relating to a time-dependent generative CSI feedback method, device, system, and storage medium. Background Technology

[0002] Massive MIMO technology is one of the core supporting technologies for next-generation wireless communication systems (5G / 6G). By deploying large-scale antenna arrays on the base station (BS) side, it can significantly improve spectral efficiency and link reliability. To fully leverage the performance advantages of massive MIMO systems, the base station side needs to acquire downlink channel state information (CSI) for beamforming and precoding design.

[0003] In frequency division duplex (FDD) systems, due to the lack of reciprocity between uplink and downlink channels, base stations cannot directly estimate downlink CSI using uplink pilots. Therefore, user equipment (UE) must first estimate the downlink CSI, then quantize and compress it, and report it to the base station via the feedback link. With the increase in antenna size and the number of subcarriers, the dimension of the CSI matrix to be fed back increases dramatically, and feedback overhead has become a key bottleneck limiting the deployment of large-scale MIMO systems.

[0004] To alleviate the feedback overhead problem in FDD massive MIMO systems, academia and industry have proposed a large number of CSI feedback methods based on deep learning (DL). These methods can be broadly divided into two categories: (A) frame-wise compression and (B) temporal CSI feedback.

[0005] (A) Frame-by-frame compression method. Its core idea is to treat a single frame of CSI as an "image," drawing inspiration from image compression in computer vision. It uses CNNs, Transformers, or autoencoders to learn the low-dimensional latent representation of the CSI, and then quantizes it into a bitstream for transmission back to the base station. The limitation of the frame-by-frame compression method is that each frame of CSI is encoded independently, failing to utilize the strong temporal correlation between adjacent time slots. In scenarios where the UE moves continuously along a trajectory, the CSI of several consecutive frames are highly similar; independent encoding leads to a large number of redundant bits, resulting in high feedback overhead.

[0006] (B) Timing-based CSI feedback method. Its core idea is to extrapolate or refine the current CSI using historical CSI, thereby reducing feedback overhead. However, the timing-based feedback method heavily relies on the assumption that "adjacent CSI frames have strong temporal consistency." When the UE experiences sudden changes in motion state such as sharp turns, sudden pauses, or random movements, the temporal correlation between adjacent CSIs is significantly disrupted, and historical CSI no longer has reliable prediction / refinement capabilities for the current CSI, thus introducing significant reconstruction errors.

[0007] In summary, existing frame-by-frame compression methods suffer from excessive feedback overhead, while existing time-series feedback methods suffer from poor robustness to sudden motion changes in the user interface. Summary of the Invention

[0008] To address the problems existing in the prior art, the present invention provides a generative CSI feedback method, apparatus, system, and storage medium based on time correlation.

[0009] To achieve the above objectives, the present invention provides the following solution: A time-correlation-based generative CSI feedback method employs a two-stage history-aware generative CSI feedback approach, consisting of prediction, encoding, and refeedback. The time correlation is used only to reduce the number of encoded bits in compressed CSI, rather than being used directly to reconstruct the current CSI. The first stage involves single-frame CSI compression and tokenization based on ECVQ-VAE, while the second stage involves prediction, encoding, and feedback based on causal Transformer. In the first stage, an entropy-constrained vector quantization variational autoencoder (ECVQ-VAE) is used to encode and quantize the high-dimensional CSI matrix of each frame into CSI tokens; each CSI token is a codeword index in the VQ codebook, thus locking the channel reconstruction error only in the first stage. In the second phase, the UE / BS is based on the historical CSI token sequence. The probability distribution of the current CSI token is predicted using the same causal Transformer. ;based on Construct the Huffman codebook for the current CSI token Variable-length encoding is performed and transmitted back to the base station; the base station decodes using the same probability distribution and the same Huffman codebook, and then reconstructs the code using a CNN decoder. .

[0010] This invention also provides a time-correlation-based generative CSI feedback device, which employs a two-stage history-aware generative CSI feedback process of prediction, encoding, and refeedback. The time correlation is used only to reduce the number of encoded bits in compressed CSI, rather than being directly used to reconstruct the current CSI; including: The first processing module is used for single-frame CSI compression and tokenization based on ECVQ-VAE in the first stage. The second processing module is used for prediction, encoding, and feedback based on causal Transformer in the second stage; The first processing module is used in the first stage to encode and quantize the high-dimensional CSI matrix of each frame into CSI tokens using an entropy-constrained vector quantization variational autoencoder ECVQ-VAE; each CSI token is a codeword index in the VQ codebook, which locks the channel reconstruction error only in the first stage. The second processing module is used in the second phase for the UE / BS to process historical CSI token sequences. The probability distribution of the current CSI token is predicted using the same causal Transformer. ;based on Construct the Huffman codebook for the current CSI token Variable-length encoding is performed and transmitted back to the base station; the base station decodes using the same probability distribution and the same Huffman codebook, and then reconstructs the code using a CNN decoder. .

[0011] The present invention also provides a time-dependent generative CSI feedback system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a time-dependent generative CSI feedback method when executed by the processor.

[0012] The present invention also provides a storage medium storing a computer program that executes a time-dependent generative CSI feedback method during runtime.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a two-stage history-aware generative CSI feedback mechanism of "predict, encode, and then feedback," using "temporal correlation" only to reduce the number of encoded bits in compressed CSI, rather than directly reconstructing the current CSI. Using this invention, the technical solution can simultaneously achieve "low feedback overhead" and "high reconstruction robustness" when the UE moves along any trajectory (including straight lines, random steps, etc.). Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of the generative CSI feedback method based on time correlation according to an embodiment of the present invention; Figure 2 This paper compares the generative CSI feedback framework proposed in this invention with existing time-series feedback methods, where (a) is an existing time-series feedback method and (b) is the generative CSI feedback framework proposed in this invention. Figure 3 For feedback overhead - NMSE performance comparison; where the left is the outdoor scene; the right is the indoor scene. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 like Figure 1 As shown, this invention provides a time-correlation-based generative CSI feedback method, employing a two-stage history-aware generative CSI feedback approach of "predict, encode, and then feedback," specifically as follows: ① In the first stage (single-frame CSI compression and tokenization), the entropy-constrained vector quantization variational autoencoder (ECVQ-VAE) is used to encode and quantize the high-dimensional CSI matrix of each frame into a discrete integer sequence (CSI token), thus locking the "channel reconstruction" error only in this stage; ② In the second stage (Transformer-based prediction, encoding, and feedback), based on historical CSI tokens, a customized causal Transformer is used to predict the probability distribution of the current CSI token. This probability distribution is then used to construct a Huffman codebook to perform variable-length encoding on the current CSI token before transmitting it back to the base station. When the prediction is accurate, only one bit is needed to encode a token; when the prediction is inaccurate, it only increases the number of feedback bits without introducing any additional reconstruction error.

[0019] ③ Further, for the periodic structure of CSI tokens, a cyclic positional embedding is designed to improve the Transformer's ability to model the intra-frame token structure regularity.

[0020] Consider a single-cell downlink massive MIMO system, where the base station (BS) is equipped with The system uses one antenna, with multiple user equipment (UEs) operating as single-antenna users. The system employs Orthogonal Frequency Division Multiplexing (OFDM) modulation, with the number of subcarriers being [number missing]. Consider a UE moving continuously along a certain trajectory, with the system operating in time slots. In each time slot... The base station needs to obtain the downlink CSI from the UE to design precoding and beamforming. Since uplink and downlink reciprocity is not available in FDD systems, the UE must first estimate the downlink CSI, then compress it and feed it back to the base station.

[0021] Signal model: in the first Time slot, first The signal received by the UE on each subcarrier is (1) in For downlink channel vectors, For precoded vectors, To send symbols, It is additive white Gaussian noise. All The channel vectors on the nth subcarriers are stacked to obtain the nth... Space-frequency channel matrix of time slot .

[0022] Angle-delay domain channel matrix: Note the high correlation between subcarrier and antenna channels, for By performing a two-dimensional DFT transformation, we can obtain (2) in , These are the DFT matrices in the time delay domain and the angle domain, respectively. Since the channel energy in the angle-time delay domain is mainly concentrated in the first few time delay taps, it can be truncated. The former OK( ), to obtain the equivalent channel matrix That is, the feedback object of the present invention.

[0023] Problem description: Even after truncation, each frame It still contains a large number of complex-valued parameters, and complete feedback per slot and per parameter is unacceptable. The goal of this invention is to design a CSI feedback scheme that "utilizes time correlation and is robust to UE motion," so as to minimize the number of feedback bits per frame (bpf) while enabling base station reconstruction. With the original The normalized mean square error (NMSE) should be minimized, i.e., in In a sense, it has near-perfect CSI performance.

[0024] The two-stage history-aware generative CSI feedback framework proposed in this invention consists of two sequential stages, as follows: Phase 1 (Single-Frame CSI Compression and Tokenization): The UE uses a CNN encoder with a vector quantization (VQ) module to compress and tokenize the continuous space-frequency CSI matrix. Compress and quantize into a discrete integer sequence (That is, the so-called "CSI token"). Each CSI token is a codeword index in the VQ codebook.

[0025] Phase 2 (Prediction, Encoding, and Feedback): UE / BS based on historical CSI token sequences The probability distribution of the current CSI token is predicted using the same causal Transformer. ;based on Construct the Huffman codebook for the current CSI token Variable-length encoding is performed and transmitted back to the base station; the base station decodes using the same probability distribution and the same Huffman codebook, and then reconstructs the code using a CNN decoder. .

[0026] Phase 2 only introduces an impact on "coding efficiency" (more accurate prediction → shorter Huffman code), without introducing any additional reconstruction error. When the prediction is inaccurate, only the number of feedback bits increases, while... The reconstruction quality is still determined by stage 1. This "error decoupling" design is the essential reason why this invention remains robust to UE motion mutations.

[0027] Existing methods (LSTM / Markov) directly use historical CSI for current CSI reconstruction, which cannot guarantee reconstruction quality when UE motion changes abruptly. In contrast, this invention first tokenizes the current CSI using the ECVQ-VAE module, and then performs probability prediction and Huffman variable-length coding on the current token based on the historical CSI token. Even if the UE motion changes abruptly, it will only cause an increase in the number of feedback bits, but the reconstruction quality can still be guaranteed.

[0028] Phase 1: Single-frame CSI compression and tokenization based on ECVQ-VAE The goal of Phase 1 is to generate a high-dimensional continuous CSI matrix for each frame. Convert to a fixed-length discrete integer sequence This facilitates sequence modeling and encoding based on Transformer in stage 2. The present invention adopts an entropy-constrained vector quantization-variable autoencoder (ECVQ-VAE) architecture [6] to achieve this purpose, which includes three sub-modules: CNN encoder, vector quantization (VQ) layer, and CNN decoder. The specific network structure is shown in Table 1.

[0029] Table 1 The workflow for Phase 1 is as follows: Step 1: The CNN encoder will (Real part, imaginary part → 2 channels, dimensions) Layer-by-layer extraction of continuous latent characterization ,in Number of tokens per frame (in the experiment) ), Feature dimensions for each token (in experiments) ).

[0030] Step 2: The VQ layer is based on a learnable codebook (In the experiment) For each subvector Select codewords and output the corresponding codeword index. : (3) in It's typing The probability of its use It is a Lagrange multiplier that balances bit rate and distortion. The two terms in Equation (3) correspond to "the more frequently the codeword is used, the fewer the number of encoded bits (entropy constraint term)" and "the smaller the quantization error, the better". Taking both into account, the whole system is optimal in the sense of bit rate and distortion.

[0031] Step 3: The UE will quantize the resulting discrete integer sequence. The (i.e., CSI token) is sent to Phase 2 for variable-length encoding and feedback. After receiving the feedback, the base station first decodes it. Then look up the corresponding codeword in the table. Finally, the CSI matrix is ​​obtained by feeding it into the CNN decoder. .

[0032] Phase 1 end-to-end training loss: (4) in, For stopping gradient operators; To rebuild MSE; It is an entropy constraint term; This includes commitment loss for the VQ layer and codebook update loss.

[0033] Phase 2: Prediction, Encoding, and Feedback Based on Causal Transformers The goal of Phase 2 is to leverage the temporal correlation of historical CSI tokens to further reduce the impact of current CSI tokens. The number of encoded bits. Its structure consists of three parts: a CSI token embedding module, Layered causal Transformer stacking, linear projection + softmax probability prediction layer.

[0034] (1) CSI token embedding. All historical CSI tokens within the context window are embedded. Mapped to continuous embeddings via a learnable token embedding table. ,in Historical CSI frame count (in the experiment) ), For model dimensions (in experiments) ).

[0035] Observation 1 (Cyclic positional embedding): Unlike natural language tokens, CSI tokens inherently possess "... The "periodic" structure—each frame is always divided into... The token, the first Locations ( The semantic roles are the same in each frame. Based on this, the present invention designs a cyclic positional encoding: (5) in ( The cyclic positional encoding (CSI) table is a learnable embedding table shared by all frames. This allows the Transformer to be explicitly aware of the repeating structure of sub-positions within a frame, thus enabling more efficient modeling of CSI token sequences. The final embedding input to the Transformer is composed of a combination of token embeddings, cyclic positional embeddings, and standard absolute positional embeddings.

[0036] (2) Causal Transformer stacking. go through After layer "masked multi-head self-attention + feedforward network", the hidden state is output at the end of the sequence. (In the experiment) (Number of heads to focus on = 8)

[0037] (3) Probability Prediction. Linear projection + softmax gives the probability distribution of the current CSI token: (6) (4) Huffman variable-length coding. Based on Constructing the Huffman codebook The UE uses this codebook to index the current token. The encoded data is transmitted back to the base station as a variable-length bit string; simultaneously, since the UE and BS deploy identical Transformers and historical tokens, they can independently generate identical data. and Based on this, the base station can correctly decode the feedback bit stream and recover the data. Then, it is restored by the CNN decoder in stage 1. When the prediction is very accurate (the probability of a true token is close to 1), Huffman coding only requires 1 bit to encode the token; when the prediction is inaccurate, the token only needs to be encoded with a longer Huffman codeword, but no additional reconstruction error is introduced.

[0038] Figure 1 In the first stage, the 32×32×2 real-valued CSI matrix is ​​projected through a CNN encoder and quantized into a 16-dimensional latent representation through a VQ layer, resulting in 16 integer CSI token indices. The UE / BS uses a Transformer to predict the probability distribution of the current token, synchronously constructs a Huffman codebook, and performs variable-length encoding and decoding of the token (the figure demonstrates the case where the current token is predicted to be hit and only 1 bit is needed for encoding). The base station recovers the CSI matrix based on the decoded token index using a VQ reverse lookup table and a CNN decoder.

[0039] Phase 2 end-to-end training loss (training the Transformer based on the pre-trained ECVQ-VAE): (7) That is, the true current token index for each sampling window. Maximize its position in the predicted distribution The log-likelihood loss is used. Since each window has only one true token index available for supervision during training, negative log-likelihood loss is used instead of the commonly used multi-class cross-entropy loss.

[0040] This invention differs fundamentally in structure from existing temporal CSI feedback methods (such as CsiNet-LSTM and MarkovNet), for example... Figure 2 As shown: Existing methods use "temporal correlation" for "channel reconstruction"—that is, directly using historical CSI to predict / refine the current CSI. In this case, if the prediction / refinement error is large, it will inevitably directly manifest as a deterioration in the reconstruction error.

[0041] This invention applies "temporal correlation" to "Huffman coding of tokens"—that is, using only the statistical regularity of historical CSI tokens to help compress the bit representation of the current token. The "channel reconstruction error" of this invention is determined only by Phase 1 (ECVQ-VAE). When there are sudden changes in UE motion or inaccurate probability predictions in Phase 2, the Huffman codeword will become longer (i.e., the number of feedback bits increases), but... The decoding is still completely error-free, therefore The reconstruction quality is exactly the same as that of "frame-by-frame compression without temporal information". This is "error decoupling".

[0042] Furthermore, since the UE and BS deploy the same Transformer and the same historical token, they can independently predict exactly the same results. This generates an identical Huffman codebook, avoiding the extra overhead of "codebook synchronization" transmission.

[0043] To verify the effectiveness of this invention, it was evaluated on indoor (I1_2p5) and outdoor (O1_3p5) scenes of the DeepMIMOv3 dataset. The main parameters are shown in Table 2. Wherein: BS antenna / number of delay taps. Potential dimensions codebook size Transformer layer count: 2, width: 256, attention head count: 8; historical frame count: Dropout=0.1; Optimizer AdamW, Scheduler OneCycle; Learning rate Weight decay The batch size was 160, and the training lasted for 100 epochs. The indoor and outdoor speeds were 0.1 m / s and 2 m / s, respectively. To evaluate robustness, in addition to the straight-line trajectory, this invention also sets a random walking trajectory: each step has a probability of 0.5 for going straight, 0.25 for turning left forward, and 0.25 for turning right forward.

[0044] Table 2 The comparison baselines include: baseline 1 (ECVQ-VAE), baseline 2 (VQ-VAE), baseline 3 (CsiNet-LSTM), and baseline 4 (TransNet). Among them, the original implementations of baselines 3 and 4 do not contain quantization layers. For fair comparison, this invention adds a uniform quantizer and a pass-through estimator (STE) to their encoder output to enable end-to-end training and measure the feedback overhead using bitwise FPS.

[0045] The horizontal axis represents the number of feedback bits per frame (bpf), and the vertical axis represents the NMSE (dB). Across all feedback overhead values, this invention (solid red line) consistently achieves the lowest NMSE, significantly outperforming baselines 1 through 4. Key observations are as follows: First, from... Figure 3 Looking at the bitrate-distortion curves, this invention consistently achieves the lowest NMSE across all feedback overhead ranges, significantly outperforming the four baseline methods. Secondly, in the low overhead range, baseline 3 (CsiNet-LSTM) achieves an NMSE comparable to this invention, but its performance quickly saturates as feedback overhead increases, eventually even being surpassed by frame-by-frame VQ-VAE / ECVQ-VAE that doesn't utilize temporal correlations. There are two reasons for this: (i) LSTM has limited ability to model complex temporal correlations; (ii) CsiNet-LSTM and TransNet rely on scalar quantization rather than vector quantization, resulting in lower efficiency.

[0046] Table 3 Third, as shown in Table 3, the present invention achieves an NMSE of -13.34 dB with a resolution of 18.06 bpf on a straight trajectory, reducing feedback overhead by approximately 68.7% compared to baseline 1 (57.68 bpf, -13.34 dB). On a random walking trajectory, the present invention still achieves an NMSE of -13.29 dB with a resolution of 26.15 bpf, maintaining the same reconstruction accuracy as baseline 1 (57.74 bpf, -13.29 dB), while still reducing feedback overhead by approximately 54.7%. Conversely, baseline 3 (CsiNet-LSTM) performs reasonably well on a straight trajectory (-12.43 dB), but catastrophically deteriorates to -1.79 dB (a degradation of approximately 10 dB) on a random walking trajectory, highlighting the vulnerability of time-series methods to sudden changes in UE motion. These two sets of data collectively validate that the present invention significantly reduces feedback overhead while maintaining reconstruction accuracy and is highly robust to sudden changes in UE motion.

[0047] This invention has the following innovative features: (1) A two-stage historical perception generative CSI feedback framework of "prediction, encoding, and feedback". In stage 1, each frame of high-dimensional continuous CSI is compressed and quantized into a discrete integer sequence (CSI token); in stage 2, based on the historical CSI token, the probability distribution of the current CSI token is predicted by Transformer, and Huffman codebook is constructed to encode the current CSI token with variable length before being transmitted back to the base station. The core design concept of this two-stage serial structure of "prediction-encoding-feedback" and "using temporal correlation to reduce the number of encoded bits rather than directly reconstructing CSI" is the key.

[0048] (2) Feedback mechanism for "error decoupling". The "channel reconstruction error" of this invention is completely locked in stage 1 and decoupled from the prediction accuracy of stage 2: when the prediction of stage 2 is inaccurate, only the number of feedback bits is increased, without introducing any additional reconstruction error. This error decoupling mechanism is the fundamental reason for achieving "robustness to UE motion mutation".

[0049] (3) Stage 1: CSI tokenization based on ECVQ-VAE. A CNN encoder + vector quantization (VQ) layer + CNN decoder structure is used to compress and quantize the continuous space-frequency CSI matrix into a discrete integer sequence. The VQ layer considers both quantization error and codeword usage entropy constraint in codeword selection (Equation 3); the end-to-end training loss (Equation 4) combines MSE, entropy constraint and VQ commitment loss to make the discrete tokens output in Stage 1 have both low reconstruction distortion and high entropy coding efficiency.

[0050] (4) Stage 2: Causal Transformer + Huffman Variable Length Encoding. The historical CSI tokens are used as context input to the causal Transformer. The probability distribution of the current CSI token is predicted by softmax projection (Equation 6). Based on this, the Huffman codebook is constructed to perform variable length encoding and feedback on the current token.

[0051] (5) Huffman codebook construction method for UE / BS synchronization. Since the UE and BS share the same Transformer model and the same historical CSI token, they can independently infer the exact same probability distribution and thus construct the exact same Huffman codebook, thereby achieving codebook synchronization with "zero additional overhead".

[0052] (6) Cyclic positional embedding (Equation 5) designed for the periodic structure of CSI tokens. By reusing the sub-position embeddings modulo "with a period of N", the periodic structure of "each frame consists of N tokens and the semantics of the same sub-positions are consistent" is explicitly modeled. This positional encoding significantly improves the Transformer's ability to probabilistically model CSI token sequences, which is an important innovation of this invention that distinguishes it from directly applying NLPTransformer.

[0053] (7) The training objective of stage 2 is to use the negative log-likelihood (Equation 7) as the training loss, which is the "lower bound of the Huffman coding length". This loss is directly aligned with the objective of stage 2, "minimizing the number of feedback bits", so that the probability distribution learned by the Transformer is best suited for Huffman coding, rather than simple multi-class classification.

[0054] (8) End-to-end robustness. Under different motion trajectories such as straight line and random step, the present invention can maintain the same reconstructed NMSE as the frame-by-frame compression method ECVQ-VAE, while significantly reducing feedback overhead (approximately 68.7% reduction under straight line and approximately 54.7% reduction under random step).

[0055] This invention employs a two-stage generative feedback framework that reduces coded bits based on temporal correlation rather than directly reconstructing the CSI, overcoming the shortcomings of traditional temporal feedback methods that couple prediction errors to reconstruction errors and have poor robustness to sudden changes in UE motion. Specifically, in the first stage, ECVQ-VAE is used to compress and quantize the CSI of each frame into discrete tokens; in the second stage, a causal Transformer is used to predict the probability distribution of the current token based on historical tokens, and a Huffman codebook is constructed accordingly to perform variable-length encoding on the current token before feeding it back to the base station. Since the second stage does not introduce any reconstruction error, even if the prediction is inaccurate, only the number of feedback bits increases. Thus, this invention significantly reduces the feedback overhead (from 57.68 bpf to 18.06 bpf for straight-line trajectories, a reduction of approximately 68.7%; from 57.74 bpf to 26.15 bpf for random-walking trajectories) while maintaining the same reconstruction accuracy as frame-by-frame ECVQ-VAE. (Reduces bpf by approximately 54.7%) and is highly robust to UE motion abrupt changes, enabling FDD massive MIMO systems to achieve near-lossless CSI reconstruction performance with lower feedback overhead.

[0056] Example 2 This invention also provides a time-correlation-based generative CSI feedback device, which employs a two-stage history-aware generative CSI feedback process of prediction, encoding, and refeedback. The time correlation is used only to reduce the number of encoded bits in compressed CSI, rather than being directly used to reconstruct the current CSI; including: The first processing module is used for single-frame CSI compression and tokenization based on ECVQ-VAE in the first stage. The second processing module is used for prediction, encoding, and feedback based on causal Transformer in the second stage; The first processing module is used in the first stage to encode and quantize the high-dimensional CSI matrix of each frame into CSI tokens using an entropy-constrained vector quantization variational autoencoder ECVQ-VAE; each CSI token is a codeword index in the VQ codebook, which locks the channel reconstruction error only in the first stage. The second processing module is used in the second phase for the UE / BS to process historical CSI token sequences. The probability distribution of the current CSI token is predicted using the same causal Transformer. ;based on Construct the Huffman codebook for the current CSI token Variable-length encoding is performed and transmitted back to the base station; the base station decodes using the same probability distribution and the same Huffman codebook, and then reconstructs the code using a CNN decoder. .

[0057] Example 3 The present invention also provides a time-dependent generative CSI feedback system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a time-dependent generative CSI feedback method when executed by the processor.

[0058] Example 4 The present invention also provides a storage medium storing a computer program that executes a time-dependent generative CSI feedback method during runtime.

[0059] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A generative CSI feedback method based on time correlation, characterized in that, A two-stage history-aware generative CSI feedback approach is adopted, which involves prediction, encoding, and feedback. Temporal correlation is used only to reduce the number of encoded bits for compressed CSI. The first stage is single-frame CSI compression and tokenization based on ECVQ-VAE. The second stage is prediction, encoding, and feedback based on causal Transformer. In the first stage, an entropy-constrained vector quantization variational autoencoder (ECVQ-VAE) is used to encode and quantize the high-dimensional CSI matrix of each frame into CSI tokens; each CSI token is a codeword index in the VQ codebook. In the second phase, the UE / BS is based on the historical CSI token sequence. The probability distribution of the current CSI token is predicted using the same causal Transformer. ;based on Construct the Huffman codebook for the current CSI token Variable-length encoding is performed and transmitted back to the base station; the base station decodes using the same probability distribution and the same Huffman codebook, and then reconstructs the code using a CNN decoder. .

2. The generative CSI feedback method based on time correlation as described in claim 1, characterized in that, For the periodic structure of CSI tokens, a cyclic position encoding is designed; specifically: In the Time slot, first The signal received by the UE on each subcarrier is in, For downlink channel vectors, For precoded vectors, To send symbols, Additive white Gaussian noise; all The channel vectors on the nth subcarriers are stacked to obtain the nth... Space-frequency channel matrix of time slot ; right Performing a two-dimensional DFT transformation yields: in, , These are the DFT matrices in the time delay domain and the angle domain, respectively; Cut The former OK( ), to obtain the equivalent channel matrix ; Design a CSI feedback scheme to reduce base station reconstruction costs. With the original The normalized mean square error.

3. The generative CSI feedback method based on time correlation as described in claim 2, characterized in that, The first phase includes: Step (1): The CNN encoder will Layer-by-layer extraction as a continuous latent characterization ,in The number of tokens per frame. Feature dimensions for each token; Step (2): The VQ layer is based on a learnable codebook For each subvector Select codewords and output the corresponding codeword index. : in, It's typing The probability of its use It is a Lagrange multiplier that balances bit rate and distortion; Step (3): The UE will quantize the resulting discrete integer sequence The data is sent to the second stage for variable-length encoding and feedback; after receiving the feedback, the base station first decodes it. Then look up the corresponding codeword in the table. Finally, the CSI matrix is ​​obtained by feeding it into the CNN decoder. .

4. The generative CSI feedback method based on time correlation as described in claim 3, characterized in that, The loss during the first phase of end-to-end training is: in, For stopping gradient operators; To rebuild MSE; It is an entropy constraint term; This includes commitment loss for the VQ layer and codebook update loss.

5. The generative CSI feedback method based on time correlation as described in claim 4, characterized in that, The second phase specifically includes: Step (1) CSI token embedding: embed all historical CSI tokens in the context window. Mapped to continuous embeddings via a learnable token embedding table. ,in, For historical CSI frame counts, For model dimensions; Step (2) Causal Transformer Stacking: go through After layer "masked multi-head self-attention + feedforward network", the hidden state is output at the end of the sequence. ; Step (3) Probability Prediction: Linear projection + softmax gives the probability distribution of the current CSI token: Step (4) Huffman variable-length coding: based on Constructing the Huffman codebook The UE uses this codebook to index the current token. The encoded data is transmitted back to the base station after being converted into a variable-length bit string; simultaneously, based on identical Transformers and historical tokens deployed by the UE and BS, identical tokens are generated. and Based on this, the base station can correctly decode the feedback bit stream and recover the data. Then, it is restored by the first-stage CNN decoder. .

6. The generative CSI feedback method based on time correlation as described in claim 5, characterized in that, The end-to-end training loss in the second phase is: The true current token index for each sampling window Maximize its position in the predicted distribution The log-likelihood in.

7. A generative CSI feedback device based on time correlation, characterized in that, A two-stage history-aware generative CSI feedback mechanism, consisting of prediction, encoding, and refeedback, is employed, using temporal correlation solely to reduce the number of encoded bits in compressed CSI; this includes: The first processing module is used for single-frame CSI compression and tokenization based on ECVQ-VAE in the first stage; it uses an entropy-constrained vector quantization variational autoencoder ECVQ-VAE to encode and quantize the high-dimensional CSI matrix of each frame into CSI tokens; each CSI token is a codeword index in the VQ codebook, which locks the channel reconstruction error only in the first stage. The second processing module is used for prediction, encoding, and feedback based on causal Transformer in the second stage; the UE / BS is based on historical CSI token sequences. The probability distribution of the current CSI token is predicted using the same causal Transformer. ;based on Construct the Huffman codebook for the current CSI token Variable-length encoding is performed and transmitted back to the base station; the base station decodes using the same probability distribution and the same Huffman codebook, and then reconstructs the code using a CNN decoder. .

8. A generative CSI feedback system based on time correlation, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program, when executed by the processor, performs the time-dependent generative CSI feedback method as described in any one of claims 1-6.

9. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, performs the time-correlation-based generative CSI feedback method as described in any one of claims 1-6.

Images