Lightweight OFDM channel estimation method and system based on attention mechanism

Abstract

The invention discloses a lightweight OFDM (Orthogonal Frequency Division Multiplexing) channel estimation method and system based on an attention mechanism. The method comprises the steps that an improved LMMSE algorithm is adopted to carry out preliminary estimation on a pilot frequency matrix, and adversarial noise disturbance is introduced to enhance estimation robustness; lMMSE estimation is used as input, and global and local features of a frequency domain and a time domain are captured through processing of a Transform encoder and a lightweight residual network in sequence; and super-resolution-level channel recovery is realized based on up-sampling. Compared with a traditional MMSE and an existing attention model, the method has the advantages that more than 81.14% of performance improvement is kept under the condition of a high signal-to-noise ratio, and better robustness and generalization ability are shown under different Doppler frequency shift scenes; the method has the advantages of being simple in structure, low in calculation overhead and suitable for the mobile terminal with limited resources, and the requirements of 5G and subsequent mobile communication systems for real-time high-precision channel estimation are met.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication channel estimation, and in particular relates to a lightweight OFDM channel estimation method and system based on an attention mechanism. Background Art

[0002] Orthogonal frequency division multiplexing (OFDM) technology has been widely used in wireless communication fields such as the fourth and fifth generation mobile communication systems (4G / 5G) due to its advantages in spectrum utilization and resistance to multipath fading. The performance of OFDM systems depends largely on the accurate estimation of wireless channel state information (CSI). In order to obtain CSI, traditional methods usually use pilot-assisted, least squares (LS), minimum mean square error (MMSE) or linear minimum mean square error (LMMSE) algorithms for channel estimation. However, existing channel estimation algorithms generally have the following shortcomings:

[0003] Weak noise resistance: The estimation accuracy of traditional LS, MMSE, and LMMSE methods decreases significantly in low signal-to-noise ratio environments, making it difficult to meet the requirements for high accuracy and high robustness in actual mobile communication scenarios.

[0004] High model complexity: Some high-performance channel estimation algorithms rely on complex prior channel statistics and matrix operations. As the system scales up, the computational complexity and storage overhead increase rapidly, making it difficult to implement efficiently in resource-constrained mobile terminals or embedded devices.

[0005] New challenges of deep learning methods: In recent years, deep learning methods have been introduced into the field of channel estimation. Although they have improved the ability to extract nonlinear features, existing deep learning models generally have large parameters, strong dependence on training data, insufficient adaptability to low SNR and dynamic channel environments, and are prone to information loss when capturing long-range correlations and global features.

[0006] Limitations of attention mechanism models: Attention-based models such as Transformer demonstrate good global modeling capabilities in channel estimation. However, existing solutions have a large number of parameters, are computationally complex, and offer limited accuracy improvements at low signal-to-noise ratios, making them difficult to meet the real-time and lightweight requirements of mobile devices.

[0007] Therefore, there is an urgent need for a new OFDM channel estimation method that combines high accuracy, low latency, low complexity, and good generalization capabilities, especially in complex wireless environments such as low signal-to-noise ratio and high-speed mobility, and that can overcome the performance bottlenecks of traditional methods and existing deep learning methods. Therefore, studying a lightweight channel estimation model based on the attention mechanism has significant practical value. Summary of the Invention

[0008] In order to solve the problems of the above-mentioned prior art, the purpose of the present invention is to propose a lightweight OFDM channel estimation method and system based on the attention mechanism. The method comprehensively utilizes the pilot distribution structure and super-resolution reconstruction idea under the 5G New Radio (5G NR) standard, combined with the improved LMMSE initial estimation and Transformer encoder, to achieve high-precision and high-efficiency estimation of the OFDM channel. By introducing a lightweight residual network with depthwise separable convolution and H-Swish activation function, the number of model parameters and computational complexity are greatly reduced, making the present invention suitable for resource-constrained terminal devices. The present invention has the advantages of simple structure, low computational overhead, and suitability for resource-constrained mobile terminals, meeting the needs of 5G and subsequent mobile communication systems for real-time high-precision channel estimation, and solving the problems of low estimation accuracy, high model complexity, and large computational resource consumption of existing OFDM channel estimation methods in low signal-to-noise ratio environments.

[0009] The specific technical solutions of the present invention are as follows.

[0010] The present invention provides a lightweight OFDM channel estimation method based on an attention mechanism, comprising the following steps:

[0011] Step 1: Embed the combined comb and block pilots into the channel matrix in the OFDM symbol slot according to the 5G NR resource block allocation scheme.

[0012] Step 2: Based on the improved LMMSE algorithm, perform initial channel estimation on the sparse pilot signal and enhance the estimation robustness by counteracting noise perturbations;

[0013] Step 3: Use the multi-head self-attention mechanism to model the global correlation of the LMMSE estimation matrix and extract global features in the time and frequency domains;

[0014] Step 4: Use a lightweight residual network based on depthwise separable convolution and H-Swish activation function to refine local features of global features and alleviate gradient disappearance;

[0015] Step 5: Use the bilinear interpolation method and convolution layer to restore the low-resolution feature map output by the lightweight residual network to a high-resolution channel estimation result of the same size as the original OFDM channel matrix.

[0016] In the present invention, in step 1, each time slot contains 14 OFDM symbols, each OFDM symbol contains 72 subcarriers, and pilots are inserted at the odd subcarrier positions and even subcarrier positions of the 1st and 13th OFDM symbols, respectively. The remaining subcarriers corresponding to these two pilot symbols are filled with 0, and the remaining 12 OFDM symbols are filled with modulation symbols, and their subcarriers serve as data subcarriers.

[0017] In the present invention, in step 1, the cyclic prefix (CP) length of each OFDM symbol is 16 OFDM samples, and Quadrature Phase Shift Keying (QPSK) modulation is used.

[0018] In the present invention, in step 1, the channel model can be a generalized multipath fading channel model, covering various typical channel models such as Rayleigh fading, Rician fading, Nakagami fading, Weibull fading, etc.; in a specific embodiment, the channel model is a Rayleigh fading channel model.

[0019] In the present invention, in step 2, the actual pilot matrix estimated based on the LS method includes the pilot matrix H p And the channel noise component N:

[0020]

[0021] Therefore, the cross-correlation between the actual pilot matrix and the actual channel matrix estimated by LS is:

[0022]

[0023] Compared with the traditional LMMSE algorithm based on ideal and noiseless pilot different, By introducing noise perturbations to disturb the LS estimation, adaptive adjustments can be made to different SNRs and dynamic channel conditions, thereby improving the accuracy and robustness of the actual channel estimation. Therefore, the improved LMMSE algorithm is:

[0024]

[0025] in is the autocorrelation of the ideal pilot matrix, SNR is the actual signal-to-noise ratio, and β = 1 is the modulation constant.

[0026] In the present invention, step 3 specifically includes:

[0027] The input LMMSE estimation matrix is ​​linearly projected through the Transformer encoder, and the attention head is set to 6 to obtain the query, key, and value vectors respectively. The attention weights are calculated through the scaled dot product attention mechanism, and the global features are weightedly fused to capture the long-range correlation between OFDM symbols and subcarriers. The final output dimension is the same as the feature representation of the original input.

[0028] In the present invention, the output of the Transformer encoder is connected to a lightweight residual network through a convolutional layer containing 12 convolution kernels. After processing this layer,

[0029] In the present invention, in step 4, the lightweight residual network adopts a depth-wise separable convolution structure, and realizes sparse extraction of feature channels through depth-wise convolution layer, H-Swish activation function and point-by-point convolution layer; the first depth-wise convolution layer in the lightweight residual convolution network uses 12 depth-wise convolutions with 5×5 convolution kernels, and the second point-by-point convolution layer uses 12 point-by-point convolutions with 1×1×12 convolution kernels.

[0030] In the present invention, in step 5, the bilinear interpolation method is used to upsample the feature map output by the lightweight residual network, and the low-resolution feature map is restored to a high-resolution output that is the same as the original OFDM channel matrix; the output features are further fine-tuned through the convolution layer to achieve channel response estimation for all subcarriers in the entire time slot.

[0031] The present invention also provides a lightweight OFDM channel estimation system based on an attention mechanism, which is used to implement the above-mentioned OFDM channel estimation method, and includes:

[0032] Pilot signal embedding: In the OFDM symbol time slot, according to the 5G NR resource block allocation scheme, the pilot structure combining combo pilot and grid pilot is embedded into the channel matrix;

[0033] The initial estimation module performs initial channel estimation on sparse pilot signals based on the improved LMMSE algorithm and enhances the estimation robustness by counteracting noise disturbances.

[0034] The Transformer encoder module uses a multi-head self-attention mechanism to model the global correlation of the LMMSE estimation matrix and extract global features in the time and frequency domains;

[0035] The lightweight residual network module, based on depthwise separable convolution and H-Swish activation function, refines the local features of the Transformer encoder output and alleviates gradient vanishing;

[0036] The upsampling reconstruction module uses bilinear interpolation and convolution layers to restore the low-resolution feature map to a high-resolution channel estimation result of the same size as the original OFDM channel matrix.

[0037] In the present invention, a convolutional layer is set between the Transformer encoder module and the lightweight residual network module, and the upsampling reconstruction module includes a fully connected layer and a convolutional layer.

[0038] In the above, the present invention applies a new LMMSE algorithm to the pilot matrix of the traditional LS estimation to improve the initial estimation accuracy. The Transformer attention mechanism is used to fully capture the global and local features of the low-resolution channel matrix. A lightweight residual network is used to process the encoder output to improve the model feature extraction capability and alleviate the gradient disappearance. Finally, an upsampling method is used to restore the complete channel matrix. Compared with the existing technology, the present invention has the following advantages:

[0039] (1) Significantly improves the channel estimation accuracy in low signal-to-noise ratio and high dynamic environments, and enhances the system's anti-interference capability and robustness;

[0040] (2) The number of model parameters and computational complexity are greatly reduced, meeting the requirements of mobile terminals, IoT devices, etc. for lightweight and low latency;

[0041] (3) Through adaptive noise perturbation and multi-head self-attention mechanism, the model's generalization ability and adaptability to actual complex channels are effectively enhanced;

[0042] In summary, the present invention not only improves the performance and efficiency of OFDM channel estimation, but also provides key basic technical support for the next generation of intelligent wireless communication systems. Experimental results show that compared with existing solutions, the CE-TrEM model proposed in the present invention has stronger generalization ability and robustness under different signal-to-noise ratios and different Doppler frequency shift conditions. Under the extended SNR conditions on the extended typical urban channel model (ETU) channel, the model shows excellent low MSE compared with traditional MMSE and existing deep learning methods; when the SNR is 10dB and the Doppler frequency shift is transferred from 0Hz to 194Hz, the model is always better than other methods, and shows slower degradation under high Doppler frequency shift; in order to evaluate the end-to-end communication performance, the bit error rate (BER) is used as a metric, and the experimental results also show that the model provides relatively more accurate channel estimation, maintains a low bit error rate and better performance, especially at lower signal-to-noise ratio levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only part of the embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 It is a block diagram of the CE-TrEM channel estimation system of the present invention.

[0045] Figure 2 This is a flowchart of the lightweight OFDM channel estimation method based on the attention mechanism of the present invention.

[0046] Figure 3 This is the MSE comparison between the present invention and the existing solution under the extended SNR condition.

[0047] Figure 4 The present invention is compared with the existing solution MSE under extended Doppler frequency shift.

[0048] Figure 5 This is a comparison of the BER of the present invention with that of the existing solution under the extended SNR condition. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0050] Example 1

[0051] This embodiment takes into account that OFDM urgently needs a new OFDM channel estimation method with high precision, low latency, low complexity, and good generalization capability, especially in complex wireless environments such as low signal-to-noise ratio and high-speed movement, which can break through the performance bottlenecks of traditional methods and existing deep learning methods.

[0052] Combine Figure 1 The lightweight OFDM channel estimation system based on the attention mechanism (hereinafter referred to as "CE-TrEM") shown in the figure provides a specific implementation of the method described in the present invention. The system mainly includes:

[0053] Pilot signal embedding: In the OFDM symbol time slot, according to the 5G NR resource block allocation scheme, the pilot structure combining combo pilot and grid pilot is embedded into the channel matrix;

[0054] The initial estimation module performs initial channel estimation on sparse pilot signals based on the improved LMMSE algorithm and enhances the estimation robustness by counteracting noise disturbances.

[0055] The Transformer encoder module uses a multi-head self-attention mechanism to model the global correlation of the LMMSE estimation matrix and extract global features in the time and frequency domains;

[0056] A lightweight residual network module, based on depthwise separable convolution and H-Swish activation function, refines local features of the encoder output and alleviates gradient vanishing;

[0057] The upsampling reconstruction module uses bilinear interpolation and convolution layers to restore the low-resolution feature map to a high-resolution channel estimation result of the same size as the original OFDM channel matrix.

[0058] Further, such as Figure 2 As shown in FIG, the lightweight OFDM channel estimation method based on the attention mechanism includes the following steps:

[0059] Step 1: Use the comb pilot and grid pilot structures to embed pilots in the OFDM channel matrix. Reference the resource block allocation method of the 5G NR (New Radio) standard and establish a channel estimation model based on the concept of super-resolution reconstruction.

[0060] Step 2: Introduce a LMMSE algorithm that includes the cross-correlation between the true channel matrix and the actual pilot matrix to make an initial channel estimate. The result is regarded as a low-resolution image, and the complete channel matrix is ​​restored using the image reconstruction method;

[0061] Step 3: Use the Transformer-based encoder to capture the correlation between the carrier vectors in the pilot matrix obtained by LMMSE, extract fine global features, improve computational efficiency, and reduce information loss.

[0062] Step 4: Use a lightweight residual network to process the encoder output to improve feature extraction capabilities and alleviate gradient disappearance;

[0063] Step 5: Use the bilinear interpolation method to construct the output of the residual network into a complete channel matrix.

[0064] The following are specific examples.

[0065] In this embodiment, to minimize model complexity and runtime overhead while ensuring channel estimation accuracy, we also jointly optimize the hyperparameters of each module, including pilot interval, noise perturbation amplitude, number of attention heads, residual network convolution kernel size, and interpolation strategy, to achieve real-time, high-precision channel estimation in low signal-to-noise ratio and high-speed mobility scenarios. This method includes the following steps:

[0066] Step 1:

[0067] Pilot structure design:

[0068] In the OFDM channel matrix, a pilot structure combining comb pilot (comb pilot) and grid pilot (block pilot) is used to improve the accuracy of channel estimation and anti-interference capability.

[0069] Channel model and pilot distribution adaptability:

[0070] This method is applicable to generalized multipath fading channel models, including Rayleigh, Rician, Nakagami, and Weibull fading. It is particularly well-suited for complex multipath environments, such as urban and indoor environments, and non-line-of-sight (NLOS) propagation scenarios. The pilot structure and resource allocation mechanism utilize a dynamic parameter adaptation design. Through a channel feature identification module, the system detects key parameters such as delay spread, Doppler shift, and frequency selectivity in real time, enabling adaptive matching of channel statistics and time-frequency domain characteristics under different propagation environments.

[0071] Refer to the resource block allocation method of the 5G NR standard:

[0072] The pilot embedding method refers to the resource block allocation scheme in the 5G New Radio (NR) standard. Specifically, each time slot contains 14 OFDM symbols, each of which contains 72 subcarriers. The first and 13th OFDM symbols are reserved for pilot embedding. The first OFDM symbol inserts a pilot on the odd-numbered subcarriers, and the 13th OFDM symbol inserts a pilot on the even-numbered subcarriers. The remaining subcarriers of these two pilot symbols are filled with zeros. The remaining 12 OFDM symbols are all data symbols, and all data subcarriers are filled using QPSK modulation.

[0073] System configuration parameter description:

[0074] The cyclic prefix (CP) length of each OFDM symbol is 16 OFDM sampling points; the modulation method adopts gray coding (Gray Coding) QPSK to support efficient data transmission and channel estimation.

[0075] Step 2:

[0076] The actual pilot matrix estimated by the traditional least squares (LS) method Contains the pilot matrix H p And the channel noise component N:

[0077]

[0078] therefore:

[0079]

[0080] Compared with the traditional LMMSE algorithm based on ideal and noiseless pilot different, By introducing noise perturbations to disturb the LS estimate, adaptive adjustments can be made to different signal-to-noise ratios (SNRs) and dynamic channel conditions, thereby improving the accuracy and robustness of the actual channel estimate. Therefore, the modified LMMSE algorithm is:

[0081]

[0082] Where β=1, which is a constant of QPSK modulation.

[0083] Step 3:

[0084] Using the multi-head self-attention mechanism adopted by the Transformer encoder, the input linear minimum mean square error estimation (LMMSE) estimation matrix is ​​linearly projected, and the attention head is set to 6. The query, key, and value vectors are obtained respectively. The attention weights are calculated through the scaled dot product attention mechanism, and the global features are weightedly fused to capture the long-range correlation between orthogonal frequency division multiplexing (OFDM) symbols and subcarriers. The final output dimension is the same as the feature representation of the original input.

[0085] Step 4:

[0086] A lightweight residual network is connected after the encoder, which is based on a depth-separable convolutional structure. It is composed of a depth-wise convolutional layer, an H-Swish activation function, and a point-by-point convolutional layer to achieve efficient sparse extraction of feature channels. In this invention, the first convolutional layer before the residual network contains 12 convolution kernels, and the output of the encoder is After processing this layer, The residual module consists of 3 residual convolutional network stacks, each stack uses a depthwise convolution kernel of size 5×5. In the standard residual network, each residual block uses two standard convolutional layers, each layer contains 12 convolution kernels of size 5×5×12, and a total of 7200 parameters. The present invention replaces the first convolutional layer with a depthwise convolution using 12 convolution kernels of size 5×5, and replaces the second convolutional layer with a pointwise convolution using 12 convolution kernels of size 1×1×12. This modification reduces the number of parameters to 444, which is only 6.17% of the standard residual design, while retaining the representation capacity, alleviating the gradient vanishing problem through residual connections, and improving network training stability and generalization ability.

[0087] Step 5:

[0088] The low-dimensional pilot matrix output by the lightweight residual network is upsampled using bilinear interpolation, which includes a fully connected layer and a convolutional layer. The low-resolution feature map is restored to a high-resolution output equivalent to the original OFDM channel matrix. The convolutional layer further fine-tunes the output features to estimate the channel response for all subcarriers within the entire time slot.

[0089] To measure the performance of the algorithm, the present invention compares the channel matrix obtained through steps 1 to 5 with the existing research [1-6] Compare the estimated results in Figure 3 、 Figure 4 、 Figure 5 shown.

[0090] This example generates a test set with an SNR range from -10dB to 40dB on the Extended Typical Urban Channel Model (ETU) channel over a wide signal-to-noise ratio range, maintaining the same Doppler shift conditions as the training set. Each SNR is evaluated using 2500 independent channel realizations. The CE-TrEM of the present invention outperforms all other models in terms of mean square error (MSE) performance across the entire SNR range. Even at the extremely low SNR of -10dB, the MSE of CE-TrEM is 0.203 lower than the minimum mean square error (MMSE) estimate and 0.412 lower than the offline Channelformer. At higher SNR levels, the performance of all methods converges to the theoretical limit, with CE-TrEM consistently maintaining a 0.0006 advantage over Channelformer.

[0091] At a 10dB signal-to-noise ratio, the Doppler shift ranged from 0Hz to 194Hz. Each test scenario was implemented using 2500 independent channels. Simulation results show that our CE-TrEM consistently outperforms other methods. While MMSE performs well at low Doppler shifts, its performance degrades as the Doppler frequency increases, falling behind the offline Channelformer at 116Hz. In contrast, CE-TrEM maintains excellent performance at high Doppler shifts, with a slower performance degradation.

[0092] This example also demonstrates bit error rate (BER) performance at different signal-to-noise ratios. Simulation results show that at a -10dB signal-to-noise ratio, the CE-TrEM of our invention outperforms the offline Channelformer by 0.018903 and outperforms the MMSE by 0.003. As the signal-to-noise ratio increases, the CE-TrEM maintains its excellent performance. At higher signal-to-noise ratios, ReEsNet surpasses the offline Channelformer, but still lags behind the CE-TrEM.

[0093] References

[0094] [1]M.Soltani,V.Pourahmadi,A.Mirzaei,and H.Sheikhzadeh,“Deeplearning-based channel estimation,”IEEE Commun.Lett.,pp.652–655,Apr.2019.

[0095] [2]L.Li,H.Chen,H.-H.Chang,and L.Liu,“Deep residual learning meetsOFDMchannel estimation,”IEEE Wireless Commun.Lett.,p.615–618,May 2020.

[0096] [3]D.Luan and J.Thompson,“Low complexity channel estimationwithneural network solutions,”in Proc.25th Int.ITG Workshop SmartAntennas,2021,pp.1–6.

[0097] [4]A.Vaswani,N.Shazeer,N.Parmar,J.Uszkoreit,L.Jones,A.Gomez,L.Kaiser,and I.Polosukhin,“Attention is all you need,”in Proc.Adv.NeuralInf.Process.Syst.,pp.5998–6008,Jun.2017.

[0098] [5]D.Luan and J.Thompson,“Attention based neural networks forwirelesschannel estimation,”in in Proc.IEEE 95th Veh.Technol.Conf.,Apr.2022,pp.1–5.

[0099] [6]D.Luan and J.S.Thompson,“Channelformer:Attention basedneuralsolution for wirelesschannel estimation and effective online training,”IEEE Trans.Wireless Commun.,vol.22,no.10,pp.6562–6577,2023。

Claims

1. A lightweight OFDM channel estimation method based on attention mechanism, characterized in that: The following steps are involved: Step 1: Embed the combined comb and block pilots into the channel matrix in the OFDM symbol slot according to the 5G NR resource block allocation scheme. Step 2: Based on the improved LMMSE algorithm, perform initial channel estimation on the sparse pilot signal and enhance the estimation robustness by counteracting noise perturbations; Step 3: Use the multi-head self-attention mechanism to model the global correlation of the LMMSE estimation matrix and extract global features in the time and frequency domains; Step 4: Use a lightweight residual network based on depthwise separable convolution and H-Swish activation function to refine local features of global features and alleviate gradient disappearance; Step 5: Use the bilinear interpolation method and convolution layer to restore the low-resolution feature map output by the lightweight residual network to a high-resolution channel estimation result of the same size as the original OFDM channel matrix.

2. The lightweight OFDM channel estimation method based on the attention mechanism according to claim 1 is characterized in that In step 1, each time slot contains 14 OFDM symbols, and each OFDM symbol contains 72 subcarriers. Pilots are inserted at the odd subcarrier positions and even subcarrier positions of the 1st and 13th OFDM symbols, respectively. The remaining subcarriers corresponding to these two pilot symbols are filled with 0, and the remaining 12 OFDM symbols are filled with modulation symbols, and their subcarriers are used as data subcarriers.

3. The lightweight OFDM channel estimation method based on the attention mechanism according to claim 2, characterized in that In step 1, the cyclic prefix (CP) length of each OFDM symbol is 16 OFDM samples, and gray-coded QPSK modulation is used.

4. The lightweight OFDM channel estimation method based on the attention mechanism according to claim 2, characterized in that In step 2, the actual pilot matrix estimated based on the LS method contains the pilot matrix H p And the channel noise component N: Therefore, the cross-correlation between the actual pilot matrix and the actual channel matrix estimated by LS is: Compared with the traditional LMMSE algorithm based on ideal and noiseless pilot different, By introducing noise perturbations to disturb the LS estimate, adaptive adjustments can be made to different signal-to-noise ratios and dynamic channel conditions, thereby improving the accuracy and robustness of the actual channel estimation. Therefore, the improved LMMSE algorithm is: in is the autocorrelation of the ideal pilot matrix, SNR is the actual signal-to-noise ratio, and β = 1 is the modulation constant.

5. The lightweight OFDM channel estimation method based on the attention mechanism according to claim 1, characterized in that Step 3 specifically includes: The input LMMSE estimation matrix is ​​linearly projected through the Transformer encoder, and the attention head is set to 6 to obtain the query, key, and value vectors respectively. The attention weights are calculated through the scaled dot product attention mechanism, and the global features are weightedly fused to capture the long-range correlation between OFDM symbols and subcarriers. The final output dimension is the same as the feature representation of the original input.

6. The lightweight OFDM channel estimation method based on the attention mechanism according to claim 5, characterized in that The output of the Transformer encoder is connected to a lightweight residual network through a convolutional layer containing 12 convolution kernels. After processing this layer, 7. The lightweight OFDM channel estimation method based on the attention mechanism according to claim 1, characterized in that In step 4, the lightweight residual network adopts a depth-wise separable convolution structure to achieve sparse extraction of feature channels through depth-wise convolutional layers, H-Swish activation functions, and point-wise convolutional layers. The first depth-wise convolutional layer in the lightweight residual convolutional network uses 12 depth-wise convolutions with 5×5 kernels, and the second point-wise convolutional layer uses 12 point-wise convolutions with 1×1×12 kernels.

8. The lightweight OFDM channel estimation method based on the attention mechanism according to claim 1, characterized in that In step 5, the bilinear interpolation method is used to upsample the feature map output by the lightweight residual network, and the low-resolution feature map is restored to a high-resolution output with the same high-resolution as the original OFDM channel matrix; the output features are further fine-tuned through the convolution layer to achieve channel response estimation for all subcarriers in the entire time slot.

9. A lightweight OFDM channel estimation system based on attention mechanism, characterized in that: It is used to implement the method in claim 1, characterized in that it includes: Pilot signal embedding: In the OFDM symbol time slot, according to the 5G NR resource block allocation scheme, the pilot structure combining combo pilot and grid pilot is embedded into the channel matrix; The initial estimation module performs initial channel estimation on sparse pilot signals based on the improved LMMSE algorithm and enhances the estimation robustness by counteracting noise disturbances. The Transformer encoder module uses a multi-head self-attention mechanism to model the global correlation of the LMMSE estimation matrix and extract global features in the time and frequency domains; The lightweight residual network module, based on depthwise separable convolution and H-Swish activation function, refines the local features of the Transformer encoder output and alleviates gradient vanishing; The upsampling reconstruction module uses bilinear interpolation and convolution layers to restore the low-resolution feature map to a high-resolution channel estimation result of the same size as the original OFDM channel matrix.

10. The lightweight OFDM channel estimation system based on attention mechanism according to claim 9, characterized in that A convolutional layer is set between the Transformer encoder module and the lightweight residual network module, and the upsampling reconstruction module includes a fully connected layer and a convolutional layer.

Images