An OTFS channel estimation method based on a deep residual attention network

Abstract

The application discloses an OTFS channel estimation method based on a deep residual attention network, and comprises the following steps: acquiring a training data set, wherein the training data set comprises multiple training data, each training data comprises an OTFS signal and CSI information corresponding to the OTFS signal; constructing a deep residual attention network, and training the deep residual attention network according to the training data set to obtain a trained deep residual attention network; collecting an OTFS signal to be estimated, and identifying the OTFS signal to be estimated by using the trained deep residual attention network to obtain an OTFS channel estimation result. Through the construction and training of the deep residual attention network, the OTFS signal to be estimated is finally identified by the deep residual attention network, and high-precision OTFS channel estimation is realized.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and specifically to an OTFS channel estimation method based on a deep residual attention network. Background Technology

[0002] Orthogonal Time-Frequency Space (OTFS) has shown promise in handling fast time-varying channels in high-mobility scenarios. Embedding pilot frames can save pilot overhead and achieve higher spectral efficiency, but it introduces interference between data and pilots. Next-generation wireless communication systems require stable performance under high-mobility conditions, such as mobile communications on high-speed trains and drones. Current fifth-generation Orthogonal Frequency Division Multiplexing (OFDM) offers advantages such as multipath resistance and low-complexity equalization. However, its performance degrades under high-Doppler channel conditions in high-mobility environments.

[0003] Orthogonal Time-Frequency Space (OTFS) modulation has attracted much attention due to its robustness to time delay and Doppler shift in high-speed mobile environments. A key difference between OTFS and OFDM lies in the acquisition method of Channel State Information (CSI). Currently, the widely used approach is the embedded pilot channel estimation method. Pilots are directly embedded into the data frame, with guard space inserted to avoid interference between pilot and data symbols. A threshold-based channel estimation method is then used to extract CSI. However, interference between pilots and data is difficult to eliminate effectively. To achieve better channel estimation performance, deep learning-based neural networks have recently shown their potential in communication tasks. Among them, the Residual Shrinking Network (DRSN) employs cascaded denoising blocks with an adaptive thresholding method and performs deep denoising to improve estimation accuracy.

[0004] The DRSN network front-end employs a residual structure for coarse denoising of OTFS symbols, while the back-end uses a learned adaptive threshold to output the final CSI result. However, the network does not fully exploit the correlations between specific features within the received signal. The actual signal consists of a superposition of three signals: pilot signal, data signal, and noise signal. Indiscriminately processing these signals degrades the estimator's performance. Effectively extracting the required OTFS pilot signal from the mixed data and reducing interference from data and noise on the pilot signal becomes a significant challenge. Summary of the Invention

[0005] The purpose of this invention is to provide an OTFS channel estimation method based on deep residual attention networks, which solves the problems existing in the prior art.

[0006] This invention is achieved through the following technical solution:

[0007] An OTFS channel estimation method based on a deep residual attention network includes:

[0008] Obtain a training dataset, which includes multiple training data sets, each of which includes an OTFS signal and the corresponding CSI information of the OTFS signal;

[0009] Construct a deep residual attention network and train it using the training dataset to obtain the trained deep residual attention network.

[0010] The OTFS signal to be estimated is acquired and processed by a trained deep residual attention network to obtain a feature map that separates pilot information from data information. Pilot features are then extracted from the feature map through an attention mechanism, and the OTFS channel estimation result is finally obtained based on the pilot features.

[0011] The OTFS channel estimation results are used to characterize the CSI information corresponding to the OTFS signal to be estimated.

[0012] In one possible implementation, the deep residual attention network includes a residual module, a first attention module, a second attention module, and an attention feature fusion module;

[0013] The residual module receives the OTFS signal and separates the pilot information and data information in the OTFS signal to obtain a feature map of the separation of pilot information and data information.

[0014] After the feature map, which separates the pilot information and the data information, is input into the first attention module and the second attention module, the pilot features are extracted through the first attention module and the second attention module, respectively.

[0015] After fusing and identifying the pilot features output by the first attention module and the pilot features of the second attention module through the attention feature fusion module, the OTFS channel estimation result is obtained.

[0016] In one possible implementation, the first attention module is configured as a positional attention module or a channel attention module, and the second attention module is configured as a channel attention module.

[0017] In one possible implementation, the residual module includes at least one residual unit. When the number of residual units is greater than 1, multiple residual units are connected in series. Each residual unit has the same structure and includes at least one convolutional structure, a third convolutional layer, and a fourth convolutional layer. The input of the at least one convolutional structure serves as the input of the residual unit. The at least one convolutional structure includes a first convolutional layer, a first batch of normalization layers, and a first activation function layer connected in sequence. The output of the first activation function layer serves as the output of the residual unit.

[0018] The convolutional structure and the third convolutional layer share the input of the residual module. That is, after the first data feature composed of the OTFS signal enters the residual module, it is processed by the convolutional structure and the third convolutional layer respectively. The output of the convolutional structure is connected to the input of the fourth convolutional layer. After the output of the third convolutional layer and the output of the fourth convolutional layer are subtracted element by element, the second data feature C×H×W is obtained, and the second data feature is used as the output of the residual module.

[0019] Where C represents the number of channels in the tensor, and each channel contains an H×W matrix, where H represents the length of the matrix and W represents the width of the matrix.

[0020] In one possible implementation, the position attention module includes at least one fifth convolutional layer, at least one sixth convolutional layer, at least one seventh convolutional layer, a first transpose layer, a first reshape layer, a second reshape layer, a third reshape layer, a second activation function layer, and a fourth reshape layer.

[0021] The fifth, sixth, and seventh convolutional layers share the input of the position attention module, that is, the second data feature is used as the input of the position attention module. After the second data feature enters the position attention module, it is processed by the fifth, sixth, and seventh convolutional layers respectively.

[0022] The second data feature of C×H×W is transformed into the third data feature of HW×C after being processed by at least one fifth convolutional layer, a first transpose layer and a first reshape layer connected in sequence.

[0023] The second data feature of C×H×W is transformed into the fourth data feature of C×HW after being processed by at least one sixth convolutional layer and a second reshape layer.

[0024] The second data feature of C×H×W is transformed into the fifth data feature of C×HW after being processed by at least one seventh convolutional layer and a third reshape layer.

[0025] After the third data feature of HW×C is processed by the second activation function layer, it is multiplied by the fourth data feature of C×HW through matrix multiplication to transform it into the sixth data feature of HW×HW.

[0026] After multiplying the fifth data feature of C×HW with the sixth data feature of HW×HW by matrix multiplication and then processing it through the fourth reshape layer, it is transformed into the seventh data feature of C×H×W.

[0027] After summing the second data feature of C×H×W element by element with the seventh data feature of C×H×W, the eighth data feature of C×H×W is obtained, and this eighth data feature is used as the output of the position attention module.

[0028] In one possible implementation, the channel attention module includes a fifth reshape layer, a sixth reshape layer, a second transpose layer, a seventh reshape layer, a third activation function layer, and an eighth reshape layer;

[0029] The fifth reshape layer, the sixth reshape layer, and the second transpose layer share the same input, that is, the second data feature is used as the input of the channel attention module. After the second data feature enters the channel attention module, it is processed by the fifth reshape layer, the sixth reshape layer, and the second transpose layer respectively.

[0030] The second data feature of C×H×W is transformed into the ninth data feature of HW×C after being processed by the second transpose layer and the seventh reshape layer.

[0031] The second data feature of C×H×W is transformed into the tenth data feature of C×HW after being processed by the sixth reshape layer.

[0032] After being processed by the fifth reshape layer, the second data feature of C×H×W is transformed into the eleventh data feature of C×HW.

[0033] After multiplying the ninth data feature of HW×C with the tenth data feature of C×HW by matrix multiplication, and then processing it through the third activation function layer, the twelfth data feature of C×C is obtained.

[0034] After multiplying the eleventh data feature of C×HW with the twelfth data feature of C×C by matrix multiplication, and then processing it through the eighth reshape layer, the thirteenth data feature of C×H×W is obtained.

[0035] After summing the second data feature of C×H×W and the thirteenth data feature of C×H×W element by element, we obtain the fourteenth data feature of C×H×W.

[0036] In one possible implementation, the attention feature fusion module includes a multi-scale channel attention unit and at least one eighth convolutional layer;

[0037] After the eighth data feature of C×H×W output by the position attention module and the fourteenth data feature of C×H×W output by the channel attention module enter the attention feature fusion module, the eighth data feature and the fourteenth data feature are first summed element by element to obtain the initial fusion feature of C×H×W; and the initial fusion feature is processed by the multi-scale channel attention unit to obtain the high-precision fusion feature of C×H×W.

[0038] After subtracting the high-precision fusion feature element by element from the identity matrix with the same dimension as the high-precision fusion feature, the fifteenth data feature of C×H×W is obtained. The Hadamard product between the eighth data feature output by the position attention module and the fifteenth data feature is obtained to obtain the sixteenth data feature of C×H×W.

[0039] The Hadamard product between the fourteenth data feature output by the channel attention module and the high-precision fused feature is obtained to get the seventeenth data feature of C×H×W.

[0040] The sixteenth and seventeenth data features are summed element by element, and then processed through at least one eighth convolutional layer to obtain the 2×H×W OTFS channel estimation result, which is the CSI information corresponding to the OTFS signal.

[0041] In one possible implementation, the multi-scale channel attention unit includes a global average pooling layer, a first point convolutional layer, a third batch normalization layer, a third ReLU activation function layer, a second point convolutional layer, a third point convolutional layer, a fourth batch normalization layer, a fourth ReLU activation function layer, a fourth point convolutional layer, a fifth batch normalization layer, a sixth batch normalization layer, a broadcast layer, and an activation function layer.

[0042] The global average pooling layer and the first point convolutional layer share the same input. That is, after the initial fused features are input into the multi-scale channel attention unit, they are respectively input into the global average pooling layer and the first point convolutional layer for processing.

[0043] The initial fusion features are processed by a global average pooling layer, a third point convolutional layer, a third batch normalization layer, a third ReLU activation function layer, a fourth point convolutional layer, and a fifth batch normalization layer to obtain the eighteenth data feature of C×1×1.

[0044] The initial fusion features are processed by the first point convolutional layer, the fourth batch normalization layer, the fourth ReLU activation function layer, the second point convolutional layer, and the sixth batch normalization layer to obtain the nineteenth data feature of C×H×W.

[0045] The C×H×W feature obtained after processing the eighteenth data feature through the broadcast layer is added element-wise to the nineteenth data feature, and then processed through the activation function layer to obtain the twentieth data feature of C×H×W.

[0046] Obtain the Hadamard product between the initial fusion feature and the twentieth data feature to obtain the high-precision fusion feature C×H×W.

[0047] In one possible implementation, the deep residual attention network is trained based on a training dataset to obtain the trained deep residual attention network, including:

[0048] The OTFS signal of the training data in the training dataset is used as the input of the deep residual attention network, and the CSI information of the training data in the training dataset is used as the expected output of the deep residual attention network. The deep residual attention network is trained by backpropagation algorithm and Adam optimizer until the network converges, and the trained deep residual attention network is obtained.

[0049] In one possible implementation, the criteria for determining network convergence include:

[0050] Obtain the error function value of the deep residual attention network;

[0051] If the error function value of the deep residual attention network decreases from high to low to below the error threshold, and as training progresses, the error function value remains stable below the error threshold and the fluctuation range is less than the set range threshold, then the network is considered to have converged.

[0052] This invention provides an OTFS channel estimation method based on a deep residual attention network. Based on the attention mechanism in deep learning, it achieves the task of accurately extracting pilot features and performing channel estimation from OTFS symbols that mix pilots, data, and noise through a position attention module, a channel attention module, and a multi-scale attention fusion module in deep learning. This deep learning-based residual attention network can adaptively explore the features of different data in OTFS symbols, thereby achieving accurate feature extraction and feature fusion, effectively solving problems such as mutual interference between pilots and data and interference of noise on data and pilots in mixed OTFS symbols. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0054] Figure 1 The flowchart shows an OTFS channel estimation method based on a deep residual attention network provided by the present invention.

[0055] Figure 2 This is a schematic diagram of the structure of the deep residual attention network provided by the present invention.

[0056] Figure 3 This is a schematic diagram of the residual module provided by the present invention.

[0057] Figure 4 A schematic diagram of the position attention module provided by the present invention.

[0058] Figure 5 This is a schematic diagram of the channel attention module provided by the present invention.

[0059] Figure 6 A schematic diagram of the attention feature fusion module provided by the present invention.

[0060] Figure 7 This is a schematic diagram of the structure of the multi-scale channel attention unit provided by the present invention.

[0061] Figure 8 This is a schematic diagram of the first experimental comparison results provided by the present invention.

[0062] Figure 9 This is a schematic diagram of the second experimental comparison results provided by the present invention.

[0063] Figure 10 This is a schematic diagram of the time-delay-Doppler domain transmission signal provided by the present invention.

[0064] Figure 11 The residual module provided by this invention outputs a heat map.

[0065] Figure 12 The single-channel result diagram of the channel attention output provided by this invention.

[0066] Figure 13 The heatmap of the total channel attention output provided by this invention.

[0067] Figure 14 The positional attention output result diagram provided by this invention.

[0068] Figure 15 The heatmap provided by this invention is the result of adding the channel attention output and the position attention output element by element. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0070] Example

[0071] like Figure 1 As shown, an OTFS channel estimation method based on a deep residual attention network includes:

[0072] S101. Obtain the training dataset, which includes multiple training data sets, each of which includes an OTFS signal and the corresponding CSI information of the OTFS signal.

[0073] Optionally, the generation process of the OTFS signal in the training data includes:

[0074] The required channel data was generated using MATLAB, and the proposed network model was trained using offline data to obtain a trained network. The training data was generated under the assumption of biselectivity of the channel, with a carrier frequency of 3 GHz and a subcarrier spacing of 7.5 kHz. OTFS frames with N=32 and M=32 were used, with a single pilot embedded in the center of the OTFS frame. The channel gain was generated according to the distribution CN(0,1P), producing a normalized maximum delay of lmax=5 and a maximum Doppler shift of kmax=4, respectively. The number of paths was assumed to be 6. It should be noted that the actual signal-to-noise ratio (SNR) during the estimation phase is unknown. Therefore, the SNR of the data generated during the training phase is random, ranging from 0 to 20 dB. The generated OTFS symbol data was separated into real and imaginary parts and stacked to construct a dual-channel real data set for training and testing.

[0075] The pilot signals and data are modulated to obtain the corresponding data symbols x[k,l]. k represents the index in the Doppler domain, and l represents the index in the time delay domain. The modulated data symbols x[k,l] are then subjected to OTFS modulation. In the time delay Doppler plane, the data symbols x[k,l] are spaced at intervals of 1 / (MΔf) in the time delay direction and 1 / (NT) in the Doppler frequency shift direction. M represents the number of subcarriers in the frequency domain, N represents the number of OTFS symbols in the time domain, 1 / Δf represents the period of the communication data stream in the time delay Doppler domain in the time delay direction, and 1 / T represents the period of the communication data stream in the Doppler direction in the time delay Doppler domain. The data is passed through a simulated dual-selection channel to obtain the extracted OTFS symbols. The obtained complex OTFS symbols y[k,l] are separated into real and imaginary parts. Dimensions are added to the separated real and imaginary parts, and they are stacked to obtain real data with dimensions (B,C,H,W). Where B is the data volume, C is the data channel dimension, and H and W are the data matrix sizes.

[0076] More specifically, the OTFS signal corresponding to the training data is

[0077] in, X represents a two-dimensional circular convolution. p Indicates a single pilot signal, X d H represents the transmitted data, and N represents noise. dd This represents the channel matrix. x[k,l] is actually equal to X. d +X p , which represents the transmitted signal in the time-delay Doppler domain. Its specific form is as follows: We use an embedded pilot method to embed a single pilot and the transmitted data together into an N×M matrix to obtain the transmitted signal in the time-delay-Doppler domain, as shown in the following formula:

[0078]

[0079] In the above formula, x[k,l] represents the OTFS signal transmitted in the time-delayed Doppler domain, i.e., X d +X p It is an N×M matrix, k∈{0,1,2,......,N-1}, l∈{0,1,2,......,M-1}, k p and l p Both represent position indices. In k p , l p A single pilot signal is placed at position k, which is a complex signal of position 1+1j. p -2k max to k p +2k max Place 0,k within the range maxThis is the value of the normalized maximum Doppler frequency shift. Similarly, in l p -l max to l p +l max Position 0 is set to form a protective interval, l max This is the value of the normalized maximum delay. d [k,l] represents the positions on the N×M matrix excluding pilot signals and guard intervals, where the transmitted data is placed. The remaining positions are used to store the complex data to be transmitted, specifically as follows: Figure 10 As shown.

[0080] H dd This represents the channel matrix generated during simulation. The resulting training signal is the channel matrix H. dd With the transmitted signal (X) d +X p After two-dimensional circular convolution Then add the noise N to get the result.

[0081] The CSI information corresponding to the training data is as follows: The tag data is obtained by performing a two-dimensional circular convolution between the generated channel matrix and a single pilot, without any data manipulation or added noise.

[0082] S102. Construct a deep residual attention network and train it using the training dataset to obtain the trained deep residual attention network.

[0083] S103. Acquire the OTFS signal to be estimated, and process it using a trained deep residual attention network to obtain a feature map that separates pilot information from data information. Extract pilot features from the feature map using an attention mechanism, and finally obtain the OTFS channel estimation result based on the pilot features. The OTFS channel estimation result is used to characterize the CSI information corresponding to the OTFS signal to be estimated.

[0084] By feeding the OTFS signal to be estimated into the trained network in a stacked manner of real and imaginary parts, an accurate estimated CSI channel matrix can be obtained.

[0085] To reduce interference between data and pilot signals, this embodiment proposes a deep learning-based residual attention network (DRAN). DRAN consists of three modules: a channel attention module (CAM), a position attention module (PAM), and an attention feature fusion (AFF) module. These three modules extract pilot features from the received signal to achieve accurate channel estimation.

[0086] DRAN is primarily used for channel estimation in Orthogonal Time-Frequency Space (OTFS) communication systems. In all communication systems, signals pass through a medium (channel). As signals pass through the channel, distortion or various noises are added to them. Correctly decoding the received signal without introducing too many errors involves removing the distortion and noise imposed by the channel. To do this, the first step is to determine the characteristics of the channel the signal passes through. This process is called "channel estimation." DRAN is suitable for channel estimation in OTFS systems. DRAN can effectively estimate channel parameters such as channel gain, delay, and Doppler in OTFS systems, allowing the results of channel estimation to be used to compensate and correct the received OTFS symbol data.

[0087] The deep residual attention network includes a residual module, a first attention module, a second attention module, and an attention feature fusion module. The residual module receives the OTFS signal and separates the pilot information from the data information in the OTFS signal to obtain a feature map showing the separation of pilot and data information. This feature map is then input into the first and second attention modules, where pilot features are extracted. Finally, the attention feature fusion module fuses and identifies the pilot features output from the first and second attention modules to obtain the OTFS channel estimation result.

[0088] like Figure 2 As shown, the deep residual attention network in this embodiment preferably includes a residual module, a location attention module, a channel attention module, and an attention feature fusion module. The input of the residual module receives the OTFS signal, and its output is connected to the inputs of both the location attention module and the channel attention module. The outputs of the location attention module and the channel attention module are connected to the first and second inputs of the attention feature fusion module, respectively. The output of the attention feature fusion module outputs the OTFS channel estimation result.

[0089] The residual module includes at least one residual unit. When the number of residual units is greater than one, multiple residual units are connected in series. Each residual unit has the same structure and includes at least one convolutional structure, a third convolutional layer, and a fourth convolutional layer. The input of the at least one convolutional structure serves as the input of the residual unit. The at least one convolutional structure includes a first convolutional layer, a first batch of normalization layers, and a first activation function layer connected in sequence. The output of the first activation function layer serves as the output of the residual unit. The convolutional structure and the third convolutional layer share the input of the residual module. That is, after the first data feature composed of the OTFS signal enters the residual module, it is processed by the convolutional structure and the third convolutional layer respectively. The output of the convolutional structure is connected to the input of the fourth convolutional layer. The output of the third convolutional layer and the output of the fourth convolutional layer are subtracted element-wise to obtain a second data feature of C×H×W, and this second data feature is used as the output of the residual module. Here, C represents the number of channels of the tensor, each channel contains an H×W matrix, H represents the length of the matrix, and W represents the width of the matrix. The first activation function layer is preferably a ReLU activation function layer.

[0090] like Figure 3 As shown, in this embodiment, the residual module preferably comprises three residual units, and each residual unit preferably includes two convolutional structures. One convolutional structure includes a first convolutional layer, a first batch normalization layer, and a first ReLU activation function layer, and the other convolutional structure includes a second convolutional layer, a second batch normalization layer, and a second ReLU activation function layer. That is, the residual module includes a first convolutional layer, a first batch normalization layer, a first ReLU activation function layer, a second convolutional layer, a second batch normalization layer, a second ReLU activation function layer, a third convolutional layer, and a fourth convolutional layer.

[0091] The first and third convolutional layers share the same input to the residual module. The first data feature, composed of the OTFS signal, enters the residual module and is then processed by the first and third convolutional layers, respectively. The first convolutional layer, the first batch normalization layer, the first ReLU activation function layer, the second convolutional layer, the second batch normalization layer, the second ReLU activation function layer, and the fourth convolutional layer are connected sequentially. After element-wise subtraction between the third and fourth convolutional layers, a second data feature of C×H×W is obtained, and this second data feature is used as the output of the residual module.

[0092] Where C represents the number of channels in the tensor, and each channel contains an H×W matrix, where H represents the length of the matrix and W represents the width of the matrix.

[0093] like Figure 4As shown, the positional attention module includes at least one fifth convolutional layer, at least one sixth convolutional layer, at least one seventh convolutional layer, a first transpose layer, a first reshape layer, a second reshape layer, a third reshape layer, a second activation function layer, and a fourth reshape layer. In this embodiment, the fifth, sixth, and seventh convolutional layers are preferably all present in one layer. The second activation function layer is preferably a softmax activation function or a sigmoid activation function layer.

[0094] The fifth, sixth, and seventh convolutional layers share the input of the position attention module, that is, the second data feature is used as the input of the position attention module. After the second data feature enters the position attention module, it is processed by the fifth, sixth, and seventh convolutional layers respectively.

[0095] The second data feature of C×H×W is transformed into the third data feature of HW×C after being processed by the fifth convolutional layer, the first transpose layer, and the first reshape layer.

[0096] The second data feature of C×H×W is transformed into the fourth data feature of C×HW after being processed by the sixth convolutional layer and the second reshape layer.

[0097] The second data feature of C×H×W is transformed into the fifth data feature of C×HW after being processed by the seventh convolutional layer and the third reshape layer.

[0098] After the third data feature of HW×C is processed by the second activation function layer, it is multiplied by the fourth data feature of C×HW through matrix multiplication to transform it into the sixth data feature of HW×HW.

[0099] After multiplying the fifth data feature of C×HW with the sixth data feature of HW×HW by matrix multiplication and then processing it through the fourth reshape layer, it is transformed into the seventh data feature of C×H×W.

[0100] After summing the second data feature of C×H×W element by element with the seventh data feature of C×H×W, the eighth data feature of C×H×W is obtained, and this eighth data feature is used as the output of the position attention module.

[0101] like Figure 5As shown, the channel attention module includes a fifth reshape layer, a sixth reshape layer, a second transpose layer, a seventh reshape layer, a third activation function layer, and an eighth reshape layer. The third activation function layer is preferably a softmax activation function or a sigmoid activation function layer.

[0102] The fifth reshape layer, the sixth reshape layer, and the second transpose layer share a common input, that is, the second data feature is used as the input of the channel attention module. After the second data feature enters the channel attention module, it is processed by the fifth reshape layer, the sixth reshape layer, and the second transpose layer, respectively.

[0103] The second data feature of C×H×W is transformed into the ninth data feature of HW×C after being processed by the second transpose layer and the seventh reshape layer.

[0104] The second data feature of C×H×W is transformed into the tenth data feature of C×HW after being processed by the sixth reshape layer.

[0105] The second data feature of C×H×W is transformed into the eleventh data feature of C×HW after being processed by the fifth reshape layer.

[0106] After multiplying the ninth data feature of HW×C with the tenth data feature of C×HW by matrix multiplication, and then processing it through the third activation function layer, the twelfth data feature of C×C is obtained.

[0107] After multiplying the eleventh data feature of C×HW with the twelfth data feature of C×C by matrix multiplication, and then processing it through the eighth reshape layer, the thirteenth data feature of C×H×W is obtained.

[0108] After summing the second data feature of C×H×W and the thirteenth data feature of C×H×W element by element, we obtain the fourteenth data feature of C×H×W.

[0109] The actual signal is composed of a superposition of three signals: pilot signal, data signal, and noise signal. Before the residual module, these three types of signals are mixed together, making it difficult to distinguish and extract the required pilot signal.

[0110] After the OTFS symbol is input into the residual module, the output of the residual module is as follows: Figure 11 As shown. After passing through the residual module, the difference between the pilot feature region and the data and noise regions in the OTFS signal becomes obvious. The output data of the residual module is filled with three types of data and their superposition, such as... Figure 11 (a) shows the data-enhanced heatmap, as follows: Figure 11 (b) shows a heatmap of enhanced pilot characteristics. These three types of signals account for different proportions in the output. Figure 11 The feature maps of the H×W matrix in different channels C are shown. It can be observed that the distinction between data and pilot signals becomes more pronounced in different channels. However, the feature maps reveal blurred edges between the data and pilot signals, and the pilot enhancement signal is blurred.

[0111] At this point, the signal dimension becomes C×H×W, where H and W are the length and width of the corresponding matrix, and C represents the C channels, or C such matrices. Within these C channels, the three types of signals can be distinguished; each channel retains only one of the three types of signals mentioned earlier: either a pilot signal, a data signal, or a noise signal. However, the proportions of these three types of signals in the C channels differ. Subsequently, through a position attention module and a channel attention module, the required pilot information can be extracted from the C channels.

[0112] By exploring the interdependencies of feature maps in these channels of the residual module's output data, the channel attention module can extract pilot features. The channel attention module essentially eliminates irrelevant features and noise by weighting the channels from the residual module's output data, resulting in sharper edges. However, it can be observed that the enhancement of the pilot feature regions is not flat enough, leading to a decrease in the amplitude estimation performance of the channel coefficients. Therefore, a positional attention module is introduced to supplement this. The output results of a single channel in the channel attention module are shown below. Figure 12 As shown, the heatmap of the total output of the channel attention module is as follows. Figure 13 As shown.

[0113] After the output of the residual module is fed into the channel attention module, the output of the channel attention module has dimensions of C×H×W. The C channels are filled with selected pilot features, extracting a large number of clear pilot features. By exploring the interdependencies of these channel graphs in the output data of the residual module, the channel attention module can extract pilot features. Figure 12 This displays the output heatmap of the H×W matrix in one channel. Figure 13 The sum of all C channels is displayed. It can be observed that the channel attention module effectively eliminates irrelevant features and noise by weighting the channels from the residual module's output data. Figure 11 In comparison, sharper edges were obtained. However, it can be observed that the enhancement of the pilot feature region is not flat enough, leading to a decrease in the amplitude estimation performance of the channel coefficients. Therefore, a position attention module is introduced here to supplement the results.

[0114] The positional attention module (PI) essentially provides enhancement across different channels C by weighting all grids in the H×W dimension of the residual module's output data. Each channel is essentially a type of feature selected by the PI, and the PI exhibits varying degrees of channel importance, thus generating weights across the entire spatial dimension. Therefore, the PI helps the channel attention module to smoothly enhance pilot feature regions. The output of the PI is as follows: Figure 14 As shown, the position attention module actually provides enhancements in different channels C by weighting all grids in the H×W dimension of the residual module's output data. Figure 14 The output of the position attention module is displayed. Figure 14 The y-axis represents the index of each channel C, and the x-axis represents the elements of the corresponding channel's H×W matrix transformed into a single row. It can be observed that each channel is actually a type of feature selected by the positional attention module. Simultaneously, the positional attention module exhibits varying degrees of channel importance, thus generating weights across the entire spatial dimension. Therefore, the positional attention module helps the channel attention module to smoothly enhance the pilot feature region. Figure 15 The heatmap shown, after element-wise addition of the channel attention output and the position attention output, exhibits a smoother enhancement in the pilot feature region, which means that a filter with better in-band flatness is obtained for pilot feature extraction.

[0115] In this embodiment, feature A∈R C×H×W They are fed into a convolutional layer (1×1) to generate two new feature maps {B,C}∈R. C×H×W Then reshape it into R. C×HW , where HW is the number of pixels.

[0116] Where A∈R C×H×W Let A represent a tensor of dimension (b, C, H, W), where b is the number of tensors, C is the number of channels, and each channel contains a matrix of size H × W. H and W are the length and width of the matrix, respectively. Generally, the dimension b is omitted by default. Therefore, A represents a tensor of size H × W with C channels. Similarly, {B, C} ∈ R C×H×W This represents two tensors of size H×W with dimension C. Then, B and C are transformed into R. C×(HW) This represents a C×(HW) matrix obtained by compressing an H×W matrix into a single column. The Softmax layer is also a commonly used normalization layer in neural networks.

[0117] The correlation between C and B is represented by matrix multiplication of their transposes, and the spatial attention map S∈R is computed using a softmax layer. HW×HW :

[0118]

[0119] Where exp() represents the natural exponential function, B i Represents the i-th column of the previous tensor B, C j Similarly, let S represent the j-th column in the previous tensor C. By transposing B and performing matrix multiplication with C, we obtain S, S... ji Let represent an element in the j-th row and i-th column of the obtained S matrix, and the size of S is (HW)×(HW).

[0120] The output of PAM can be written as:

[0121]

[0122] Where α represents a variable, S ji As shown above, the elements of matrix S, D i Let A represent the i-th column of signal D in the DRAN structure diagram, where the size of D is C×(HW). j Let E represent the j-th column in A. j Let E represent the element in the j-th column of E. The size of E is C×(HW), which is then expanded to C×H×W.

[0123] By leveraging the interdependencies between channel graphs, interdependent feature maps can be emphasized, improving the feature representation of specific semantics. Therefore, CAM is used to select and enhance information from specific channels. Input A is reshaped into R. C×(HW) Then, matrix multiplication is performed between A and its transpose, and finally, a softmax layer is applied to obtain channel attention. Figure X ∈R C×(HW) :

[0124]

[0125] Here, the dimension of A is transformed into C×(HW), A i Let A represent the i-th row in A. j Let X represent the j-th row in matrix A. By multiplying A by its transpose, we obtain X, which has a size of (C × C). ji This represents the element in the j-th row and i-th column of X.

[0126] The output of CAM can be written as:

[0127]

[0128] Among them, X ji Let β be an element, and let A be a variable. i Let A represent the i-th row in A. j This represents the j-th row in A. F jLet F represent the element in the j-th row of F, where F is of size C×(HW), and then expanded to C×H×W.

[0129] like Figure 6 As shown, the attention feature fusion module includes a multi-scale channel attention unit and at least one eighth convolutional layer; in this embodiment, the number of eighth convolutional layers is preferably one.

[0130] The eighth data feature of C×H×W output by the position attention module and the fourteenth data feature of C×H×W output by the channel attention module are fed into the attention feature fusion module. First, the eighth data feature and the fourteenth data feature are summed element by element to obtain the initial fused feature of C×H×W. Then, the initial fused feature is processed by the multi-scale channel attention unit to obtain the high-precision fused feature of C×H×W.

[0131] After subtracting the high-precision fusion feature element by element from the identity matrix with the same dimension as the high-precision fusion feature, the fifteenth data feature of C×H×W is obtained. The Hadamard product between the eighth data feature output by the position attention module and the fifteenth data feature is obtained to obtain the sixteenth data feature of C×H×W.

[0132] The Hadamard product between the fourteenth data feature output by the channel attention module and the high-precision fusion feature is obtained to get the seventeenth data feature of C×H×W.

[0133] The sixteenth and seventeenth data features are summed element by element, and then processed through the eighth convolutional layer to obtain the 2×H×W OTFS channel estimation result, which is the CSI information corresponding to the OTFS signal.

[0134] Optionally, the attention feature fusion module may consist of only the eighth convolutional layer, which sums the eighth data feature and the fourteenth data feature element by element, and then processes them through the eighth convolutional layer to obtain the 2×H×W OTFS channel estimation result.

[0135] like Figure 7 As shown, the multi-scale channel attention unit includes a global average pooling layer, a first point-wise convolutional layer, a third batch normalization layer, a third ReLU activation function layer, a second point-wise convolutional layer, a third point-wise convolutional layer, a fourth batch normalization layer, a fourth ReLU activation function layer, a fourth point-wise convolutional layer, a fifth batch normalization layer, a sixth batch normalization layer, a broadcast layer, and an activation function layer.

[0136] The global average pooling layer and the first point convolutional layer share the same input. That is, after the initial fused features are input into the multi-scale channel attention unit, they are respectively input into the global average pooling layer and the first point convolutional layer for processing.

[0137] The initial fusion features are processed by a global average pooling layer, a third point convolutional layer, a third batch normalization layer, a third ReLU activation function layer, a fourth point convolutional layer, and a fifth batch normalization layer to obtain the eighteenth data feature of C×1×1.

[0138] The initial fusion features are processed by the first point convolutional layer, the fourth batch normalization layer, the fourth ReLU activation function layer, the second point convolutional layer, and the sixth batch normalization layer to obtain the nineteenth data feature of C×H×W.

[0139] The C×H×W feature obtained after processing the eighteenth data feature through the broadcast layer is added element-wise to the nineteenth data feature, and then processed through the activation function layer to obtain the twentieth data feature of C×H×W.

[0140] Obtain the Hadamard product between the initial fusion feature and the twentieth data feature to obtain the high-precision fusion feature C×H×W.

[0141] Optionally, the activation function layer can be a Sigmoid activation function, a ReLU activation function, or a LeakyReLU activation function. Figure 7 In this context, 'r' represents the value of the reduced channel.

[0142] To fully utilize the characteristics of both paths, the following approach is adopted: Figure 7 The AFF module based on Multi-Scale Channel Attention Unit (MS-CAM) shown selectively fuses features from two paths. Compared to a fixed linear aggregation that only provides feature maps by directly adding features from PAM and CAM, MS-CAM extracts attention information and aggregates multi-scale contextual information through two branches at different scales. This achieves a non-linear fusion, improving fusion accuracy. Figure 5 As shown, the input X will pass through two distinct branches. The upper branch preserves the shape of X and extracts features from the spatial dimension using pointwise convolution. This branch can be viewed as a feature selection filter in the H×W dimension. The lower branch will compress the input X to obtain a global view in C×1 columns. This can be seen as a filter that selects different feature channels. The results of the upper and lower branches will be fused through a direct element-wise summation operation. Finally, a softmax layer is used to obtain the attention matrix, and the output is an element-wise weighted sum of the attention matrix over the input X. The output can be written as:

[0143]

[0144] in, The symbol represents element-wise summation, ⊙ represents the Hadamard product, M represents the MS-CAM operation, and Z represents the fused features. The output of the AFF is actually a weighted average between E and F. Finally, the output of the AFF is compressed to the original dimension of 2×H×W, with two channels representing the real and imaginary parts of the channel matrix. The output is the required channel matrix CSI information.

[0145] In one possible implementation, the deep residual attention network is trained based on a training dataset to obtain the trained deep residual attention network, including:

[0146] The iteration count was set to 1500, the training learning rate to 0.0003, the batch size to 128, and the random seed to 42. The OTFS signal of the training data in the training dataset was used as the input of the deep residual attention network, and the CSI information of the training data in the training dataset was used as the expected output of the deep residual attention network. The deep residual attention network was trained using the backpropagation algorithm and the Adam optimizer until the network converged, resulting in the trained deep residual attention network.

[0147] In one possible implementation, the criteria for determining network convergence include:

[0148] Obtain the error function value of the deep residual attention network:

[0149]

[0150] Where NMSE represents the error function value, H~ represents the estimated channel matrix, which is the actual output of the deep residual attention network obtained after the TFS signal in the training data is input into the deep residual attention network. H represents the true channel matrix, which is the CSI information corresponding to the OTFS signal in the training data.

[0151] If the error function value of the deep residual attention network decreases from high to low to below the error threshold, and as training progresses, the error function value remains stable below the error threshold and the fluctuation range is less than the set range threshold, then the network is considered to have converged.

[0152] like Figure 8As shown, comparing the network structure proposed in this embodiment with existing technologies reveals that DRSN exhibits superior channel estimation performance. In terms of utilizing CAM, PAM, and AFF, the proposed DRAN outperforms DRSN. It can be observed that the proposed DRAN, i.e., PAM+CAM+AFF, achieves the highest accuracy, proving the effectiveness of PAM and CAM. This embodiment also replaces AFF with a direct summation operation, denoted as CAM+PAM+ADD, and it can be seen that AFF demonstrates better fusion capabilities. An interesting observation arises from comparing CAM+AFF and PAM+AFF: CAM+AFF provides quite good performance, validating the effectiveness of CAM, but the performance of direct PAM is far inferior to CAM+AFF. This is because PAM+AFF alone destroys the spatial structure information between data.

[0153] The signal A output by the residual module consists of three types of signals: data enhancement signal, pilot enhancement signal, and their superposition. Furthermore, the proportions of these three types of signals differ in different channels of A. The distinction between data and pilot signals becomes more pronounced in different channels. However, the feature map reveals blurred edges between the data and pilot signals, and the pilot enhancement signal is also blurred.

[0154] By exploring the interdependencies of these channel maps of A, CAM can extract pilot features. CAM essentially eliminates irrelevant features and noise by weighting the channels from the coarsely denoised data, resulting in sharper edges. However, the enhancement of the pilot feature region is not flat enough, leading to a decrease in the amplitude estimation performance of the channel coefficients.

[0155] PAM essentially provides enhancements across different channels by weighting all H×W grids of A. Each channel exhibits the same features selected by PAM, but loses the original spatial structure. As mentioned earlier in this embodiment, directly using PAM incurs a significant performance loss. This is because spatial structure information (H×W) is discarded, which is crucial for channel estimation. Nevertheless, PAM exhibits varying degrees of channel importance, thus generating weights across the entire spatial dimension. Therefore, PAM helps CAM to smoothly enhance the pilot feature region, resulting in a smoother enhancement in the pilot feature region, which means a filter with better in-band flatness is obtained for pilot feature extraction.

[0156] like Figure 9 As shown, the impact of different channel path numbers on the model was examined. Although the DRAN proposed in this embodiment was only trained on P=6 channel data, it showed strong adaptability to various channel path conditions. It can be found that the performance is mainly related to the number of paths. From Figure 8 and Figure 9It can be observed that DRAN can process both mixed SNR and mixed channel path data simultaneously.

[0157] In one possible implementation, the OTFS signal to be estimated is acquired, and a trained deep residual attention network is used to identify the OTFS signal to be estimated, thereby obtaining the OTFS channel estimation result, including:

[0158] The OTFS signal to be estimated is acquired and used as the input to the residual module in the deep residual attention network.

[0159] The OTFS signal is initially processed by the residual module. Then, the features of the initially processed signal are extracted by the position attention module and the channel attention module respectively. Finally, the features extracted by the position attention module and the channel attention module are fused by the attention feature fusion module to obtain the CSI information, that is, the OTFS channel estimation result.

[0160] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An OTFS channel estimation method based on a deep residual attention network, characterized in that, include: Obtain a training dataset, which includes multiple training data sets, each of which includes an OTFS signal and the corresponding CSI information of the OTFS signal; Construct a deep residual attention network and train it using the training dataset to obtain the trained deep residual attention network. The OTFS signal to be estimated is acquired and processed by a trained deep residual attention network to obtain a feature map that separates pilot information from data information. Pilot features are then extracted from the feature map through an attention mechanism, and the OTFS channel estimation result is finally obtained based on the pilot features. The OTFS channel estimation result is used to characterize the CSI information corresponding to the OTFS signal to be estimated. The deep residual attention network includes a residual module, a first attention module, a second attention module, and an attention feature fusion module; The residual module receives the OTFS signal and separates the pilot information and data information in the OTFS signal to obtain a feature map of the separation of pilot information and data information. After the feature map, which separates the pilot information and the data information, is input into the first attention module and the second attention module, the pilot features are extracted through the first attention module and the second attention module, respectively. After fusing and identifying the pilot features output by the first attention module and the pilot features of the second attention module through the attention feature fusion module, the OTFS channel estimation result is obtained. The first attention module is configured as either a positional attention module or a channel attention module, and the second attention module is configured as a channel attention module. The residual module includes at least one residual unit. When the number of residual units is greater than 1, multiple residual units are connected in series. Each residual unit has the same structure and includes at least one convolutional structure, a third convolutional layer, and a fourth convolutional layer. The input of the at least one convolutional structure serves as the input of the residual unit. The at least one convolutional structure includes a first convolutional layer, a first batch of normalization layers, and a first activation function layer connected in sequence. The output of the first activation function layer serves as the output of the residual unit. The convolutional structure and the third convolutional layer share the input of the residual module. After the first data feature composed of the OTFS signal enters the residual module, it is processed by the convolutional structure and the third convolutional layer respectively. The output of the convolutional structure is connected to the input of the fourth convolutional layer. After the output of the third convolutional layer and the output of the fourth convolutional layer are subtracted element by element, the second data feature C×H×W is obtained, and the second data feature is used as the output of the residual module. Where C represents the number of channels of the tensor, and each channel contains an H×W matrix, where H represents the length of the matrix and W represents the width of the matrix; The position attention module includes at least one fifth convolutional layer, at least one sixth convolutional layer, at least one seventh convolutional layer, a first transpose layer, a first reshape layer, a second reshape layer, a third reshape layer, a second activation function layer, and a fourth reshape layer; The fifth, sixth, and seventh convolutional layers share the input of the position attention module, that is, the second data feature is used as the input of the position attention module. After the second data feature enters the position attention module, it is processed by the fifth, sixth, and seventh convolutional layers respectively. The second data feature of C×H×W is transformed into the third data feature of HW×C after being processed by at least one fifth convolutional layer, a first transpose layer and a first reshape layer connected in sequence. The second data feature of C×H×W is transformed into the fourth data feature of C×HW after being processed by at least one sixth convolutional layer and a second reshape layer. The second data feature of C×H×W is transformed into the fifth data feature of C×HW after being processed by at least one seventh convolutional layer and a third reshape layer. After the third data feature of HW×C is processed by the second activation function layer, it is multiplied by the fourth data feature of C×HW through matrix multiplication to transform it into the sixth data feature of HW×HW. After multiplying the fifth data feature of C×HW with the sixth data feature of HW×HW by matrix multiplication and then processing it through the fourth reshape layer, it is transformed into the seventh data feature of C×H×W. After summing the second data feature of C×H×W element by element with the seventh data feature of C×H×W, the eighth data feature of C×H×W is obtained, and this eighth data feature is used as the output of the position attention module. The channel attention module includes a fifth reshape layer, a sixth reshape layer, a second transpose layer, a seventh reshape layer, a third activation function layer, and an eighth reshape layer; The fifth reshape layer, the sixth reshape layer, and the second transpose layer share the same input, that is, the second data feature is used as the input of the channel attention module. After the second data feature enters the channel attention module, it is processed by the fifth reshape layer, the sixth reshape layer, and the second transpose layer respectively. The second data feature of C×H×W is transformed into the ninth data feature of HW×C after being processed by the second transpose layer and the seventh reshape layer. The second data feature of C×H×W is transformed into the tenth data feature of C×HW after being processed by the sixth reshape layer. After being processed by the fifth reshape layer, the second data feature of C×H×W is transformed into the eleventh data feature of C×HW. After multiplying the ninth data feature of HW×C with the tenth data feature of C×HW by matrix multiplication, and then processing it through the third activation function layer, the twelfth data feature of C×C is obtained. After multiplying the eleventh data feature of C×HW with the twelfth data feature of C×C by matrix multiplication, and then processing it through the eighth reshape layer, the thirteenth data feature of C×H×W is obtained. After summing the second data feature of C×H×W and the thirteenth data feature of C×H×W element by element, we obtain the fourteenth data feature of C×H×W.

2. The OTFS channel estimation method based on deep residual attention networks according to claim 1, characterized in that, The attention feature fusion module includes a multi-scale channel attention unit and at least one eighth convolutional layer; After the eighth data feature of C×H×W output by the position attention module and the fourteenth data feature of C×H×W output by the channel attention module enter the attention feature fusion module, the eighth data feature and the fourteenth data feature are summed element by element to obtain the initial fusion feature of C×H×W. The initial fusion features are then processed through a multi-scale channel attention unit to obtain high-precision fusion features of C×H×W. After subtracting the high-precision fusion feature element by element from the identity matrix with the same dimension as the high-precision fusion feature, the fifteenth data feature of C×H×W is obtained. The Hadamard product between the eighth data feature output by the position attention module and the fifteenth data feature is obtained to obtain the sixteenth data feature of C×H×W. The Hadamard product between the fourteenth data feature output by the channel attention module and the high-precision fused feature is obtained to get the seventeenth data feature of C×H×W. The sixteenth and seventeenth data features are summed element by element, and then processed through at least one eighth convolutional layer to obtain the 2×H×W OTFS channel estimation result, which is the CSI information corresponding to the OTFS signal.

3. The OTFS channel estimation method based on deep residual attention networks according to claim 2, characterized in that, The multi-scale channel attention unit includes a global average pooling layer, a first point convolutional layer, a third batch normalization layer, a third ReLU activation function layer, a second point convolutional layer, a third point convolutional layer, a fourth batch normalization layer, a fourth ReLU activation function layer, a fourth point convolutional layer, a fifth batch normalization layer, a sixth batch normalization layer, a broadcast layer, and an activation function layer. The global average pooling layer and the first point convolutional layer share the same input. That is, after the initial fused features are input into the multi-scale channel attention unit, they are respectively input into the global average pooling layer and the first point convolutional layer for processing. The initial fusion features are processed by a global average pooling layer, a third point convolutional layer, a third batch normalization layer, a third ReLU activation function layer, a fourth point convolutional layer, and a fifth batch normalization layer to obtain the eighteenth data feature of C×1×1. The initial fusion features are processed by the first point convolutional layer, the fourth batch normalization layer, the fourth ReLU activation function layer, the second point convolutional layer, and the sixth batch normalization layer to obtain the nineteenth data feature of C×H×W. The C×H×W feature obtained after processing the eighteenth data feature through the broadcast layer is added element-wise to the nineteenth data feature, and then processed through the activation function layer to obtain the twentieth data feature of C×H×W. Obtain the Hadamard product between the initial fusion feature and the twentieth data feature to obtain the high-precision fusion feature C×H×W.

4. The OTFS channel estimation method based on deep residual attention networks according to any one of claims 1-3, characterized in that, The deep residual attention network is trained using the training dataset to obtain the trained deep residual attention network, including: The OTFS signal of the training data in the training dataset is used as the input of the deep residual attention network, and the CSI information of the training data in the training dataset is used as the expected output of the deep residual attention network. The deep residual attention network is trained by backpropagation algorithm and Adam optimizer until the network converges, and the trained deep residual attention network is obtained.

5. The OTFS channel estimation method based on deep residual attention networks according to claim 4, characterized in that, The criteria for determining network convergence include: Obtain the error function value of the deep residual attention network; If the error function value of the deep residual attention network decreases from high to low to below the error threshold, and as training progresses, the error function value remains stable below the error threshold and the fluctuation range is less than the set range threshold, then the network is considered to have converged.

Images