A multi-slot channel estimation method and device based on neural network
Through the multi-time slot channel estimation method based on neural network, the conditional generation adversarial network and data enhancement technology are used to solve the problem of channel estimation in the single-time slot channel estimation scheme, and achieve higher channel estimation accuracy and robustness.
Patent Information
- Application Number
- CN202211401004.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-09
- Publication Date
- 2025-08-29
- Estimated Expiration
- 2042-11-09
AI Technical Summary
The existing single-slot channel estimation scheme cannot effectively utilize the relevant features of multi-slot channels, resulting in inaccurate channel estimation and it is difficult for traditional methods to extract relevant information between continuous channels.
The multi-time slot channel estimation method based on neural network is adopted to use condition generation and adversarial network to perform multi-time slot channel estimation. Through multi-time slot channel feature extraction, deep channel feature extraction and channel reconstruction, combined with data enhancement technology, the time domain-related features of the channel are fully utilized to improve the accuracy and robustness of channel estimation.
Improves the accuracy and robustness of channel estimation, can adapt to variable channel conditions, reduces computational complexity, and avoids performance degradation caused by mismatch between deployment bandwidth and training bandwidth.
Smart Images
Figure CN115801502B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communication transmission technology, and in particular to a multi-slot channel estimation method and device based on a neural network. Background Art
[0002] The rapid development of the mobile internet has posed new challenges to the design of modern wireless networks. Signal transmission is subject to shadow fading, multipath effects, and Doppler shift, resulting in distortion of the signal's amplitude, phase, and frequency. Accurately estimating channel state information is necessary to restore the original signal as closely as possible and improve data throughput.
[0003] Current channel estimation schemes are generally single-time-slot channel estimation schemes, which do not consider the correlation characteristics of multi-time-slot channels.
[0004] For example, in paper 1 (Deep learning-based channel estimation [Soltani M, Pourahmadi V, Mirzaei A, et al. Deep learning-based channel estimation [J]. IEEE Communications Letters, 2019, 23(4): 652-655.]), the time-frequency response of the channel is regarded as a two-dimensional image, and a network structure that cascades a super-resolution network and a denoising image recovery network is proposed to estimate the channel. In order to obtain a complete channel image, a two-stage training method is proposed: in the first stage, the super-resolution network takes the channel estimation value of the pilot position as a vectorized low-resolution input and estimates the unknown value of the channel response. In the second stage, the image denoising network is cascaded with the super-resolution network to remove the influence of noise. However, the channel model used in this paper is simple and cannot accurately simulate the real wireless communication environment; secondly, its network structure is not further optimized for the channel estimation task.
[0005] Another example is Paper 2 (Deep residual learning meets OFDM channel estimation [Li L, Chen H, Chang HH, et al. Deep residual learning meets OFDM channel estimation [J]. IEEE Wireless Communications Letters, 2019, 9 (5): 615-618.]), which introduces a deep neural network ReEsNet based on residual learning designed specifically for channel estimation. The network input is the estimated channel response at the pilot position, and the output is the total estimated channel response. Due to its compact network structure, the computational cost is greatly reduced. Although the algorithm proposed in this paper greatly reduces the complexity while ensuring the estimation accuracy, it uses a deconvolution layer, which means that the model can only work under a specific bandwidth.
[0006] Continuously received multi-slot pilot signals contain the time-domain correlation characteristics of the channel, which can be used to improve the accuracy of channel estimation and reduce pilot overhead. However, due to the variability of the channel, the correlation information between consecutive channels is difficult to extract using traditional methods.
[0007] Currently, there is no solution that uses neural networks to evaluate the impact of multi-slot pilot information on improving the accuracy of channel estimation in the current time slot. Neural networks have powerful learning and nonlinear fitting capabilities, and can extract multi-slot channel features using learned weights. Furthermore, due to their excellent feature extraction and data generation capabilities, conditional generative adversarial networks are more suitable for deep learning-based channel estimation tasks that treat the channel response at the pilot locations as a low-resolution image.
[0008] Based on the above considerations, the present invention proposes for the first time a multi-slot channel estimation architecture based on a neural network, and uses a conditional generative adversarial network to perform multi-slot channel estimation to obtain more accurate channel state information. Summary of the Invention
[0009] The purpose of the present invention is to provide a multi-slot channel estimation method and device based on neural network to obtain more accurate channel state information.
[0010] In order to achieve the above object, the technical solution of the present invention is as follows:
[0011] In one aspect, the present invention provides a multi-slot channel estimation method based on a neural network, comprising the following steps:
[0012] S1. Obtain the estimated channel response at the pilot position of the i-th time slot by the least square method (LS) Right now: Where i=[t,t-1,…t-T+1], Y p,t is the received signal at the pilot position of the tth time slot, X p,t is the transmitted signal at the pilot position of the t-th time slot, the current time slot is the t-th time slot, and T is the number of consecutive time slots; the subscript p in the formula is the pilot position;
[0013] S2. Store the estimated channel responses at T consecutive pilot positions in a storage queue, expressed as
[0014] S3, the estimated channel response at T consecutive pilot positions Input generator, output actual channel response H for time slot t t Estimated value of
[0015] S4. Using estimated values and the received signal Y at the data position d,t Perform signal detection. The subscript d in the formula is the data position.
[0016] Furthermore, in step S1, the received signal is obtained after the transmitted signal undergoes resource allocation, OFDM modulation, channel and OFDM demodulation.
[0017] Furthermore, the generator in step S3 includes the following processing steps:
[0018] S31. The multi-slot channel feature extraction network extracts the features of the estimated channel response at the pilot positions of the multi-slots and maps them to a high-dimensional feature space, which is expressed as follows:
[0019]
[0020] The output is the multi-slot channel feature F m , f m (·) represents the multi-slot channel feature extraction network;
[0021] S32, deep channel feature extraction network from multi-slot channel features F m Extract deep channel features F d , which is expressed as follows:
[0022] F d =f d (F m )
[0023] where f d (·) represents a deep channel feature extraction network;
[0024] S33, generating an estimated channel response of the pilot position of the current time slot through the channel reconstruction network and the estimated channel response of the current time slot The difference between .
[0025] Furthermore, the generator in step S3 is trained by introducing a discriminator to construct a conditional generative adversarial network, and the discriminator is used to distinguish true (H t , ) and false ( );
[0026] Furthermore, the discriminator in step S3 includes the following processing steps:
[0027] The first step is to fix the generator and find the optimal discriminator;
[0028] The second step is to fix the discriminator and find the optimal generator;
[0029] The above two steps are performed alternately until the Nash equilibrium point is reached.
[0030] Furthermore, the training goal of the conditional generative adversarial network training described in step S3 is to train the generator to generate estimated channel responses to deceive the discriminator, and at the same time train the discriminator to distinguish between the estimated channel responses and the actual channel responses.
[0031] Furthermore, in step S3, the optimal generator G* obtained after training is obtained by the following formula:
[0032]
[0033] Where λ is a constant parameter, which is adjusted by the designer; For traditional optimization goals, The optimization objective of the conditional generative adversarial network is expressed as follows:
[0034]
[0035] Where G and D represent the generator and discriminator respectively. is the mathematical expectation function.
[0036] Furthermore, the generator described in step S3 uses a data augmentation method during training.
[0037] Furthermore, the data enhancement method is: dividing the full-bandwidth channel response into multiple narrowband sub-channel responses.
[0038] On the other hand, the present invention also provides a multi-slot channel estimation device based on a neural network, comprising an LS estimator, a storage queue and a generator, to implement any of the above methods, wherein:
[0039] The LS estimator is used to obtain the estimated channel response at the pilot position of the i-th time slot by the LS algorithm Right now: Where i=[t,t-1,…t-T+1], Y p,t is the received signal at the pilot position of the tth time slot, X p,t is the transmitted signal at the pilot position of the t-th time slot, the current time slot is the t-th time slot, and T is the number of consecutive time slots;
[0040] The storage queue is used to store the estimated channel responses at T consecutive pilot positions, which is expressed as
[0041] The generator, used for channel estimation inference, takes as input the estimated channel responses at T consecutive pilot positions The output is the actual channel response H of the tth time slot t Estimated value of
[0042] Compared with the prior art, the present invention has the following beneficial effects:
[0043] The present invention uses the time domain correlation characteristics of the channel in the continuously received pilot signal to further improve the accuracy of channel estimation. It proposes a multi-slot channel estimation method based on a neural network. Through the steps of multi-slot channel feature extraction, deep channel feature extraction and channel reconstruction, it fully utilizes the time domain correlation characteristics of the channel to improve the accuracy of channel estimation. And by learning a series of parameters through the neural network, it adaptively assigns weights to different channel features, so that it can cope with variable channel conditions. In addition, the present invention is based on the conditional generative adversarial network and can learn the probability distribution characteristics of the channel, rather than just the numerical correspondence, which is more suitable for the reality of variable channels and can improve the robustness of the algorithm to different channels. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention, and those skilled in the art can also obtain other drawings based on these drawings.
[0045] Figure 1 A flowchart of a multi-slot channel estimation method based on a neural network provided in an embodiment of the present invention.
[0046] Figure 2 A schematic diagram of a generator flow provided in an embodiment of the present invention.
[0047] Figure 3A schematic diagram of the discriminator flow provided by an embodiment of the present invention.
[0048] Figure 4 An overall view of the generator provided by an embodiment of the present invention.
[0049] Figure 5 This is a structural diagram of the Swin Transformer residual block (RSTB) provided in an embodiment of the present invention.
[0050] Figure 6 This is a structural diagram of the Swin Transformer Layer (STL) provided in an embodiment of the present invention.
[0051] Figure 7 A schematic diagram of the discriminator structure provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0052] We propose a neural network-based multi-slot channel estimation architecture and design a multi-slot channel estimation scheme based on a generative adversarial network. Specifically, the proposed neural network-based multi-slot channel estimation method fully utilizes the time-domain correlation characteristics of the channel through steps such as multi-slot channel feature extraction, deep channel feature extraction, and channel reconstruction, thereby improving the accuracy of channel estimation. Furthermore, to reduce the computational complexity of the training process, improve the robustness of the model, and avoid performance degradation caused by a mismatch between the deployment bandwidth and the training bandwidth, a data augmentation method is proposed that divides the full-bandwidth channel response into multiple narrowband sub-channel responses.
[0053] In order to better understand the present technical solution, the method of the present invention is described in detail below with reference to the accompanying drawings.
[0054] In the embodiment of the present invention, the system model adopts an Orthogonal Frequency Division Multiplexing (OFDM) system. OFDM is a multi-carrier transmission scheme that divides a transmission channel into multiple orthogonal sub-channels. Consider a size N f ×N s OFDM system, where N f is the number of subcarriers, N s is the number of OFDM symbols in each time slot. In the time-frequency domain, the relationship between the received signal and the transmitted signal can be expressed as:
[0055]
[0056] in Represent the time-frequency response of the received signal, transmitted signal and channel respectively; represents a mean of zero and a variance of σ 2 Additive Gaussian white noise; stands for Hadamard product.
[0057] The present invention provides a multi-slot channel estimation method based on neural network, the overall architecture of which is referenced Figure 1 As shown, X d,t , X p,t , Y d,t and Y p,t The transmitted signal at the data position, the transmitted signal at the pilot position, the received signal at the data position, and the received signal at the pilot position in time slot t, respectively. The received signal is obtained after the transmitted signal undergoes resource allocation, OFDM modulation, channel and OFDM demodulation. The multi-slot channel estimation device consists of a generator, an LS estimator, and a storage queue.
[0058] The method specifically comprises the following steps:
[0059] S1, the estimated channel response at the pilot position of the i-th time slot is obtained by the LS algorithm Right now: Where i=[t,t-1,…t-T+1], Y p,t is the received signal at the pilot position of the tth time slot, X p,t is the transmitted signal at the pilot position of the t-th time slot, the current time slot is the t-th time slot, and T is the number of consecutive time slots;
[0060] S2. Store the estimated channel responses at T consecutive pilot positions in a storage queue, expressed as The size of T is determined by the user based on the actual application environment, mainly considering performance and computational complexity.
[0061] S3, the estimated channel response at T consecutive pilot positions Input generator, output actual channel response H for time slot t t Estimated value of The generator is trained by introducing a discriminator to construct a conditional generative adversarial network. The discriminator is used to distinguish true (H t , ) and false After channel estimation, we use and Y d,t Perform signal detection.
[0062] To train the generator network, a discriminator is introduced to construct a conditional generative adversarial network. The specific network structure and training objectives are described below.
[0063] 1. Generator
[0064] like Figure 2 As shown, the generator includes the following processing steps:
[0065] In the first step, the multi-slot channel feature extraction neural network extracts the features of the estimated channel response at the pilot position of the multi-slot and maps it to a high-dimensional feature space. This process is expressed as follows:
[0066]
[0067] The output is the multi-slot channel feature F m , f m (·) represents the multi-slot channel feature extraction network;
[0068] The second step is to obtain the multi-slot channel characteristics F m Extract deep channel features F d , which is expressed as follows:
[0069] F d =f d (F m )
[0070] where f d (·) represents a deep channel feature extraction network;
[0071] The third step is to generate the estimated channel response of the pilot position of the current time slot through the channel reconstruction network. and the estimated channel response of the current time slot The difference between .
[0072] This network is used for channel estimation inference. Convolutional neural networks, Transformers, various activation functions, Dropout, and other technologies can be used to build this generator network.
[0073] 2. Discriminator
[0074] like Figure 3 As shown, the discriminator includes the following processing steps:
[0075] The first step is to fix the generator and find the optimal discriminator;
[0076] The second step is to fix the discriminator and find the optimal generator;
[0077] The above two steps are performed alternately until the Nash equilibrium point is reached.
[0078] 3. Training objectives
[0079] The training goal of the conditional generative adversarial network training is to train the generator to produce estimated channel responses to deceive the discriminator, and at the same time train the discriminator to distinguish between the estimated channel responses and the actual channel responses.
[0080] Optimization Objectives for Conditional Generative Adversarial Network Training as follows:
[0081]
[0082] Where G and D represent the generator and discriminator respectively. is the optimization objective of the conditional generative adversarial network, is the mathematical expectation function.
[0083] By combining this optimization objective with the traditional optimization objective, the generator is limited to avoid uncertain output. Finally, we get the optimal generator G after training through the following formula: * :
[0084]
[0085] Where λ is a constant parameter, which is adjusted by the designer; It is a traditional optimization objective, such as L1 norm, L2 norm, etc.
[0086] Furthermore, during training, the generator uses data augmentation to partition the full-bandwidth channel response into multiple narrowband sub-channel responses. This allows us to generate more data from limited data, increasing the diversity of training samples. This approach reduces the computational cost of training and improves model robustness, while also avoiding performance degradation caused by a mismatch between deployment and training bandwidths.
[0087] Example 1
[0088] This example consists of a generator and a discriminator. The generator is used for channel estimation inference, and the discriminator is used to train the generator network. The value of T is determined by the user based on the actual application environment, primarily considering performance and computational complexity. After simulation verification, this example sets T to 5.
[0089] 1 Generator
[0090] To avoid the conversion of input and output dimensions, zero is added at the data position. The shape becomes C in ×T×N f ×N s . C in is equal to 2 and represents the real and imaginary parts of the channel response.
[0091] 1.1 Multi-slot channel feature extraction
[0092] like Figure 4 As shown in Figure 1, the first layer is a 3D convolutional layer that uses C filters of size T×3×3. Through parameter adjustment, C is set to 81. It can extract the features of the estimated channel response at the pilot position of multiple time slots and map them to a high-dimensional feature space. This process can be described as:
[0093]
[0094] The output is the multi-slot channel characteristics f c o nv3d (·) represents a 3D convolutional layer.
[0095] 1.2 Deep channel feature extraction
[0096] The deep features are obtained from F m Extracted from: F d =f d (F m ), where f d (·) consists of three Residual Swin Transformer Blocks (RSTB) and a 2D convolutional layer using C filters of size 1×1. Through parameter adjustment, N is set to 1, 1, and 2 respectively. Figure 5 As shown in , RSTB(N×) is a residual block with N Swin Transformer layers (STL) and a 2D convolutional layer using C 1×1 filters. Figure 6 As shown in the figure, each STL consists of two SwIn Transformer encoders: one is a Window Multi-head Self-attention (W-MSA) encoder and the other is a Shift-window Multi-head Self-attention (SW-MSA) encoder. The W-MSA encoder reduces computational complexity but constrains features within the window. The SW-MSA encoder allows features to flow between windows, expanding the receptive field. The window size is set to 4×2.
[0097] 1.3 Channel Reconstruction
[0098] After deep feature extraction, the estimated channel response is reconstructed by fusing multi-slot channel features and deep channel features through 2D convolutional layers and skip connections. For stable training, the estimated channel response of the pilot position of the current time slot is generated using global residual learning. and the estimated channel response of the current time slot The reconstructed estimated channel can be expressed as: where f conv2d (·) represents a 2D convolutional layer.
[0099] 2 Discriminator
[0100] The discriminator is used to distinguish true (H t , ) and false like Figure 7 As shown, first of all, Through a 3D convolution using two filters of size T×3×3. Then, the output is or H t This is followed by a 2D convolutional layer with 64 4×4 filters and a Leaky ReLU activation function. The next two layers are 2D convolutional layers with 128 and 256 4×4 filters, followed by a BatchNorm layer and a Leaky ReLU activation function. The final layer is a 2D convolutional layer with a 4×4 filter.
[0101] 3 Training goals
[0102] The goal of the generative adversarial network is to train the generator to produce estimated channel responses to deceive the discriminator, while training the discriminator to distinguish between the estimated channel responses and the actual channel responses. We add the L1 norm to the optimization objective of the conditional generative adversarial network. The mathematical expression of this process is as follows:
[0103]
[0104] Where G and D represent the generator and discriminator respectively; is the mathematical expectation function; λ is a constant parameter, which is set to 10 through parameter adjustment; represents the L1 norm; G* is the optimal generator obtained after training.
[0105] 4 Data Augmentation
[0106] During model training, the size N f ×N s The full bandwidth channel response is randomly divided into multiple 32×N s This method performs data augmentation based on the narrowband subchannel response. The specific value of the partition size is determined by the user. This method can reduce the computational cost of training and improve the robustness of the model, while also avoiding performance degradation caused by a mismatch between the deployment bandwidth and the training bandwidth.
[0107] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.
Claims
1. A multi-slot channel estimation method based on a neural network, characterized in that: The steps include: S1, the estimated channel response at the pilot position of the i-th time slot is obtained by the least squares method Right now: Where i=[t, t-1,...t-T+1], Y p,i is the received signal at the pilot position of the i-th time slot, X p,i is the transmitted signal at the pilot position of the i-th time slot, the current time slot is the t-th time slot, and T is the number of consecutive time slots; S2. Store the estimated channel responses at T consecutive pilot positions in a storage queue, expressed as S3, the estimated channel response at T consecutive pilot positions Input generator, output actual channel response H for time slot t t Estimated value of The generator includes the following processing steps: S31. The multi-slot channel feature extraction network extracts the features of the estimated channel response at the pilot positions of the multi-slots and maps them to a high-dimensional feature space, which is expressed as follows: The output is the multi-slot channel feature F m , f m (·) represents the multi-slot channel feature extraction network; S32, deep channel feature extraction network from multi-slot channel features F m Extract deep channel features F d , which is expressed as follows: F d =f d (F m ) where f d (·) represents a deep channel feature extraction network; S33, generating an estimated channel response of the pilot position of the tth time slot through the channel reconstruction network The actual channel response H of the tth time slot t Estimated value of The difference between S4. Using estimated values and the received signal Y at the data position d,t Perform signal detection.
2. A multi-slot channel estimation method based on neural network according to claim 1, characterized in that: In step S3, the generator is trained by introducing a discriminator to construct a conditional generative adversarial network. The discriminator is used to distinguish true and false 3. A multi-slot channel estimation method based on neural network according to claim 2, characterized in that: The discriminator described in step S3 includes the following processing steps: The first step is to fix the generator and find the optimal discriminator; The second step is to fix the discriminator and find the optimal generator; The above two steps are performed alternately until the Nash equilibrium point is reached.
4. The multi-slot channel estimation method based on a neural network according to claim 2, characterized in that: The training goal of the conditional generative adversarial network training described in step S3 is to train the generator to generate estimated channel responses to deceive the discriminator, and at the same time train the discriminator to distinguish between the estimated channel responses and the actual channel responses.
5. The multi-slot channel estimation method based on neural network according to claim 1, characterized in that: In step S3, the optimal generator G obtained after training is obtained by the following formula: * : Where λ is a constant parameter, which is adjusted by the designer; For traditional optimization goals, The optimization objective of the conditional generative adversarial network is expressed as follows: Where G and D represent the generator and discriminator respectively. is the mathematical expectation function.
6. The multi-slot channel estimation method based on neural network according to claim 1, characterized in that: The generator described in step S3 uses data augmentation methods during training.
7. The multi-slot channel estimation method based on neural network according to claim 6, characterized in that: The data enhancement method is to divide the full-bandwidth channel response into multiple narrowband sub-channel responses.
8. A multi-slot channel estimation device based on a neural network, characterized in that: The method comprises an LS estimator, a storage queue, and a generator to implement the method according to any one of claims 1 to 7, wherein: The LS estimator is used to obtain the estimated channel response at the pilot position of the i-th time slot by the LS algorithm Right now: Where i=[t, t-1,...t-T+1], Y p,t is the received signal at the pilot position of the tth time slot, X p,t is the transmitted signal at the pilot position of the t-th time slot, the current time slot is the t-th time slot, and T is the number of consecutive time slots; The storage queue is used to store the estimated channel responses at T consecutive pilot positions, which is expressed as The generator, used for channel estimation inference, takes as input the estimated channel responses at T consecutive pilot positions The output is the actual channel response H of the tth time slot t Estimated value of