An OTFS channel estimation method based on SwinTrans-CNN model

Abstract

The present invention discloses an OTFS channel estimation method based on a SwinTrans‑CNN model, which belongs to the technical field of channel estimation; including: determining basic channel parameters according to specific channel scenarios, designing a TDL channel model; designing a delay-Doppler domain signal at a transmitting end using an embedded pilot method; performing two-dimensional cyclic convolution and Gaussian white noise processing on the transmitted delay-Doppler domain signal; constructing a SwinTrans‑CNN denoising model; and using a trained SwinTrans‑CNN denoising model to perform channel state estimation. The present invention adopts a method combining an embedded pilot scheme and a SwinTrans‑CNN model, whereby the transmitting end presets pilots and guard symbols, and the receiving end converts the channel estimation problem into a denoising problem of a two-dimensional image in the delay-Doppler domain, thereby solving the problem of excessively high bit error rate in low signal-to-noise ratio scenarios and improving the accuracy of channel estimation.

Description

Technical Field

[0001] The present invention belongs to the technical field of channel estimation, and in particular relates to an OTFS channel estimation method based on a SwinTrans-CNN model. Background Art

[0002] In future wireless communication scenarios, high-speed mobile scenarios will become an important component, such as high-speed rail communication, drone communication, car autonomous driving and low-orbit satellite communication. These high-speed mobile scenarios will produce the Doppler effect, which will affect the orthogonality between subcarriers in the currently widely used orthogonal frequency division multiplexing (OFDM) modulation system, leading to inter-carrier interference (ICI), which greatly affects the stability and reliability of communication.

[0003] In 2017, Hadani proposed a new modulation scheme and named it OTFS. Different from the traditional channel modeling in the time-frequency domain, the OTFS modulation scheme models the channel in the delay-Doppler domain. The delay-Doppler domain channel changes slowly and can be regarded as a time-invariant two-dimensional convolution channel, which can better handle high-speed mobile scenarios. In addition, the OTFS system has good compatibility with the OFDM system. Adding pre-processing and post-processing modules to the OFDM system can form an OTFS system, saving resources, so the current research on OTFS has become a hot topic.

[0004] At the end of 2022, ChatGPT was born, attracting the attention of the world and attracting more and more people to learn about artificial intelligence. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Deep learning is a new direction in the field of machine learning, which can be regarded as an upgrade of traditional neural networks. After generating and training the model, deep learning can complete feature extraction by itself, and then process the data. It has demonstrated excellent capabilities in processing complex data and tasks. At present, deep learning has achieved great results in different fields such as computer vision, speech recognition, natural language processing, and autonomous driving.

[0005] In the field of communications, deep learning is widely used in technologies such as channel estimation, signal detection, and quantum communication. A large number of experimental results have proved that the use of deep learning algorithms can effectively improve the performance of wireless communication systems.

[0006] In the system model of wireless communication, the signal transmitted by the transmitter will be affected by various obstacles in the wireless channel, causing the amplitude, phase, frequency and other parameters of the signal to change. In order to obtain accurate signals at the receiving end, it is necessary to estimate the channel and obtain the channel state information (CSI) under the current situation. At present, the OTFS channel estimation scheme is basically divided into the traditional scheme based on pilot symbols, the compressed sensing scheme and the deep learning scheme. The traditional pilot scheme is to analyze the channel scene, divide the grid resources at the transmitter, and set the pilot symbol and the protection symbol in the transmitted signal according to the parameters such as the maximum delay and the maximum Doppler shift in the channel, and make a threshold judgment at the receiving end according to the position of the corresponding area, so as to obtain the parameters of different paths; the biggest problem of this scheme is that it is easily interfered by noise. In the case of low signal-to-noise ratio, it is difficult to make a threshold judgment. A high threshold will filter the impulse response of the path, and a low threshold will mistakenly analyze the noise as the impulse response of the path. In real communication scenarios, the noise is often constantly changing, and it is difficult to dynamically set the threshold to remove noise using the threshold judgment method.

[0007] In summary, how to solve the problem of too high bit error rate in low signal-to-noise ratio scenarios, and thus improve the accuracy of OTFS channel estimation, is a technical problem that the present invention aims to solve. Summary of the invention

[0008] The object of the present invention is to provide an OTFS channel estimation method based on the SwinTrans-CNN model to solve the problems raised in the above background technology.

[0009] The object of the present invention is achieved by: an OTFS channel estimation method based on the SwinTrans-CNN model, characterized in that the method comprises the following steps:

[0010] Step S1: Determine basic channel parameters according to specific channel scenarios and design a TDL channel model;

[0011] Step S2: Designing the delay-Doppler domain signal of the transmitting end by using the embedded pilot method;

[0012] Step S3: performing two-dimensional circular convolution and Gaussian white noise processing on the delay-Doppler domain signal obtained by transmission to obtain received information;

[0013] Step S4: construct a SwinTrans-CNN denoising model and perform training;

[0014] Step S5: Use the trained SwinTrans-CNN denoising model to perform channel state estimation.

[0015] Preferably, in step S1, basic channel parameters are determined according to a specific channel scenario, and a TDL channel model is designed. The specific operations are:

[0016] Step S1-1: Obtain the maximum moving speed of the mobile carrier in the scene , according to the 5G standard in the area, obtain the carrier frequency of the communication system in the scenario and subcarrier spacing ;

[0017] The subcarrier spacing ∆f is obtained by the symbol period, which is ;

[0018] Step S1-2: Obtain the maximum Doppler shift in the scene and maximum delay spread;

[0019] Maximum Doppler shift ;

[0020] in, is the speed of light, is the maximum moving speed, is the carrier frequency;

[0021] Maximum delay spread ;

[0022] in, is the length of the communication link;

[0023] Step S1-3: According to the size of the time-frequency domain grid resources of the 5G application OFDM technology in the actual scenario, the resource size of the corresponding delay-Doppler domain grid is obtained;

[0024] The delay resolution is obtained by the correspondence between OFDM time-frequency domain grid resources and delay-Doppler domain grid resources. and Doppler resolution ;

[0025] in, is the number of grids in the delay domain, is the number of Doppler domain grids; , ;

[0026] The correspondence between OFDM time-frequency domain grid resources and delay-Doppler domain grid resources is: the number of time domain grids N corresponds to the number of Doppler domain grids , the number of frequency domain grids Corresponding delay domain grid number , the Doppler domain grid length is the Doppler resolution , the delay domain grid length is the delay resolution .

[0027] Preferably, in step S2, the embedded pilot method is used to design the delay-Doppler domain signal of the transmitting end, and the specific operations are:

[0028] Step S2-1: Place the transmission signal in the delay-Doppler domain grid resource allocated at the transmitting end.

[0029] Step S2-2: performing embedded pilot symbol setting according to the determined parameters;

[0030] The setup scheme is as follows:

[0031] The delay value is a non-negative number, and the Doppler frequency shift value is a negative or non-negative number. To avoid mutual influence between the pilot and the transmitted data at the receiving end, the delay domain guard interval is a single maximum delay value, and the Doppler domain guard interval is a double maximum Doppler frequency shift value;

[0032] The frame structure of the delay-Doppler domain signal is:

[0033] ;

[0034] in, is the pilot signal, To send data, is the index of the Doppler frequency shift value on the grid, is the index of the pilot delay value coordinate, is the index of the maximum delay; is the index of the delay value on the grid, is the index of the pilot Doppler value coordinate; is the index of the maximum Doppler shift on the grid.

[0035] Preferably, in step S3, two-dimensional circular convolution and Gaussian white noise processing are performed on the delay-Doppler domain signal obtained by sending to obtain reception information, and the specific operation is as follows:

[0036] Step S3-1: Sending a signal and the channel impulse response matrix Perform a two-dimensional circular convolution to obtain ;

[0037] Step S3-2: Add additive white Gaussian noise to get the received signal ;

[0038] Step S3-3: Send the signal in two dimensions in the delay-Doppler domain and receive signals The relationship is expressed as:

[0039] ;

[0040] in, is the delay-Doppler domain impulse response matrix, is a two-dimensional circular convolution, is the additive white Gaussian noise matrix in the delay-Doppler domain.

[0041] Preferably, the SwinTrans-CNN denoising model includes an FCN module and a U-Net module, and the SwinTrans-CNN denoising model has dual-channel input and dual-channel output;

[0042] The FCN module includes four convolutional layers, each of which has 32 convolution kernels, the convolution kernel size is 3×3, the padding is 1, there is no bias, the ReLU activation function is used, and the forward propagation process is defined;

[0043] The U-Net module includes a downsampling module, a Swin Transformer Block module and an upsampling module. The Swin Transformer Block module is composed of two continuous blocks, including a first connection block and a second connection block. Both connection blocks contain Layer Normalization and Multi-Layer Perceptron. Layer Normalization is used for normalization operation, which is used for nonlinear mapping and feature extraction.

[0044] The first connection block includes a window self-attention calculation W-MSA, which performs self-attention calculation in each window so that the model pays attention to the features in the region;

[0045] The second connection block contains the translation window self-attention calculation SW-MSA. The moving window enables the features between different windows to interact and enhance the capture of global information. The connection of two consecutive blocks extracts complex global and local information.

[0046] Preferably, the downsampling part uses average pooling and convolution operations to reduce the size of the feature map, and the upsampling module uses a transposed convolution operation to restore the image to its original size, so as to restore image detail information;

[0047] The up-sampling module and the down-sampling module are connected by skip connection to prevent the problems of gradient disappearance and gradient explosion during training.

[0048] Preferably, in step S4, a SwinTrans-CNN denoising model is constructed and trained, and the specific operations are as follows:

[0049] Step S4-1: dividing the two-dimensional complex matrix of the intercepted pilot reception range into two channels of real part and imaginary part;

[0050] Step S4-2: The input data first passes through the FCN module. After multiple convolution operations, the noise level of the input image is extracted and predicted, and a noise feature map is output;

[0051] Step S4-3: Add the noise feature map to the original image, and the added multi-channel image becomes the input of the U-Net module;

[0052] Step S4-4: the original image with the noise feature map is first passed through two consecutive downsampling modules to reduce the resolution of the feature map, facilitate multi-scale fusion and increase the receptive field;

[0053] Step S4-5: The low-resolution image output by the downsampling module is input to the Swin Transformer Block module. By combining local attention and global context, the input features are denoised while retaining important local and global feature information;

[0054] Step S4-6: After the image is output through the Swin Transformer Block module, the denoised low-resolution image is restored to the original image size through two consecutive upsampling modules, and the original image detail information is retained.

[0055] Preferably, when the SwinTrans-CNN denoising model is trained, the delay and Doppler shift range of the pilot signal are intercepted in the area of ​​the receiving end;

[0056] The range of the pilot reception , , the pilot reception range is expressed as follows: ; is the additive white Gaussian noise matrix in the delay-Doppler domain;

[0057] The range of each independent pilot reception at the receiving end is:

[0058] ;

[0059] in, is the path identifier, when the Doppler tap Delay tap When there is a path at ,otherwise ; is the Doppler tap Delay tap The impulse response at For the coordinate grid Noise at is an imaginary number; is the number of grids in the delay domain, is the number of Doppler domain grids;

[0060] The two-dimensional complex matrix of the range area of ​​pilot reception is intercepted as the input of the SwinTrans-CNN model.

[0061] Compared with the prior art, the present invention has the following improvements and advantages:

[0062] 1. By combining the embedded pilot scheme with the SwinTrans-CNN model, the pilot and guard symbols are preset at the transmitter, and the receiver converts the OTFS channel estimation problem into the denoising problem of the two-dimensional image in the delay-Doppler domain, thereby solving the problem of high bit error rate in low signal-to-noise ratio scenarios and improving the accuracy of channel estimation.

[0063] 2. The SwinTrans-CNN model combines the convolutional neural network (CNN) and Transformer, retaining the local perception characteristics of CNN and the advantage of Transformer in capturing the connection between data features, making the OTFS channel estimation method based on this model have higher estimation accuracy and stronger robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Flow chart of the method of the present invention.

[0065] Figure 2 Schematic diagram of the transformation relationship between the time-frequency domain and the corresponding delay-Doppler domain grid resources.

[0066] Figure 3 Schematic diagram of the embedded pilot scheme of the OTFS system transmitter in the present invention.

[0067] Figure 4 The figure is a schematic diagram of the pilot distribution range at the receiving end after the OTFS system transmitting end adopts the embedded pilot solution in the present invention.

[0068] Figure 5 SwinTrans-CNN model structure diagram designed for the present invention.

[0069] Figure 6 This is the module structure diagram of Swin Transformer Block.

[0070] Figure 7 The module flow chart of the OTFS channel estimation system based on the SwinTrans-CNN model is invented.

[0071] Figure 8 It is a comparison chart of normalized mean square error results between the method of the present invention and other methods. DETAILED DESCRIPTION

[0072] The present invention is further summarized below with reference to the accompanying drawings.

[0073] like Figure 1 , 7 As shown, an OTFS channel estimation method based on the SwinTrans-CNN model includes the following steps:

[0074] Step S1: Determine basic channel parameters according to specific channel scenarios and design a TDL channel model;

[0075] In the process of verifying the feasibility of the method of the present invention, the channel scenario is set to set appropriate simulation parameters under the background of high-speed rail, simulating the signal loss during the transmission process;

[0076] Step S1-1: Obtain the maximum moving speed of the mobile carrier in the scene , according to the 5G standard in the area, obtain the carrier frequency of the communication system in the scenario and subcarrier spacing ;

[0077] The subcarrier spacing ∆f is obtained by the symbol period, which is ;

[0078] The simulation parameters are set as follows: The maximum operating speed of the high-speed rail is set to =350km / h, carrier frequency 4×109Hz, subcarrier spacing 15×103Hz, number of subcarriers M=64, number of OTFS symbols N=64, maximum Doppler shift The Doppler resolution is 234 Hz, and the delay spread is 1296 Hz. 12.5us, delay resolution is 1.04 s, transmission bandwidth B = M × ∆f = 960kHz, symbol period T = 1 / ∆f = 66.7 s.

[0079] The present invention only considers the integer Doppler case, assuming that the delay is the delay domain resolution The Doppler shift is an integer multiple of is the Doppler domain resolution An integer multiple of , ; and are integers, corresponding to the index values ​​of the delay axis and Doppler axis respectively.

[0080] Step S1-2: Obtain the maximum Doppler shift in the scene and maximum delay spread;

[0081] Maximum Doppler shift = ;

[0082] in, is the speed of light, is the maximum moving speed, is the carrier frequency;

[0083] Maximum delay spread ;

[0084] in, is the length of the communication link;

[0085] like Figure 2 As shown, step S1-3: according to the size of the time-frequency domain grid resource of the 5G application OFDM technology in the actual scenario, the resource size of the corresponding delay-Doppler domain grid is obtained;

[0086] The delay resolution is obtained by the correspondence between OFDM time-frequency domain grid resources and delay-Doppler domain grid resources. and Doppler resolution ;

[0087] in, is the number of grids in the delay domain, is the number of Doppler domain grids; , ;

[0088] In the Doppler domain grid resource, The number of grids in the delay domain is equal to is the number of subcarriers, The number of Doppler domain grids is equal to N, the number of OTFS symbols;

[0089] The correspondence between OFDM time-frequency domain grid resources and delay-Doppler domain grid resources is: the number of time domain grids N corresponds to the number of Doppler domain grids , the number of frequency domain grids Corresponding delay domain grid number , the Doppler domain grid length is the Doppler resolution , the delay domain grid length is the delay resolution .

[0090] Step S2: Design the delay-Doppler domain signal of the transmitter using the embedded pilot method. The specific operations are:

[0091] Step S2-1: Place the transmission signal in the delay-Doppler domain grid resource allocated at the transmitting end.

[0092] Step S2-2: performing embedded pilot symbol setting according to the determined parameters;

[0093] The setup scheme is as follows:

[0094] like Figure 3 As shown, the delay value is a non-negative number, and the Doppler frequency shift value is a negative number or a non-negative number. To avoid mutual influence between the pilot and the transmitted data at the receiving end, the delay domain guard interval is a single maximum delay value, and the Doppler domain guard interval is a double maximum Doppler frequency shift value;

[0095] The frame structure of the delay-Doppler domain signal is:

[0096] ;

[0097] in, is the pilot signal, To send data, is the index of the Doppler frequency shift value on the grid, is the index of the pilot delay value coordinate, is the index of the maximum delay; is the index of the delay value on the grid, is the index of the pilot Doppler value coordinate; is the index of the maximum Doppler frequency shift on the grid; the red grid points are pilot , usually set to 1, blue grid points are protection symbols, usually set to 0, and other grid points are sent data Since the delay value is non-negative but the Doppler frequency shift value can be negative, in order to avoid the mutual influence of the pilot and the transmitted data at the receiving end, the delay domain guard interval is single maximum delay value, and the Doppler domain guard interval is double maximum Doppler frequency shift value.

[0098] In step S3, the delay-Doppler domain signal obtained by transmission is subjected to two-dimensional circular convolution and Gaussian white noise processing to obtain received information. The specific operation is as follows:

[0099] Step S3-1: Sending a signal and the channel impulse response matrix Perform a two-dimensional circular convolution to obtain ;

[0100] Step S3-2: Add additive white Gaussian noise to get the received signal ;

[0101] Step S3-3: Send the signal in two dimensions in the delay-Doppler domain and receive signals The relationship is expressed as:

[0102] ;

[0103] in, is the delay-Doppler domain impulse response matrix, is a two-dimensional circular convolution, is the additive white Gaussian noise matrix in the delay-Doppler domain.

[0104] Step S4: Build the SwinTrans-CNN denoising model and train it. The specific operations are as follows:

[0105] The SwinTrans-CNN denoising model includes the FCN module and the U-Net module. The SwinTrans-CNN denoising model has dual-channel input and dual-channel output.

[0106] like Figure 5 As shown in the figure, the SwinTrans-CNN denoising model has dual-channel input and dual-channel output. The two-dimensional complex matrix of the yellow grid area intercepted by the input is divided into two channels of real and imaginary parts. The input data first passes through the FCN module. The FCN module consists of four convolutional layers. Each convolutional layer has 32 convolutional kernels with a convolutional kernel size of 3×3, padded with 1, without bias, using the ReLU activation function, and defining the forward propagation process. The number of output channels after the input passes through each convolutional layer is 32, 16, 8, and 4 respectively.

[0107] The U-Net module is a classic "U"-shaped symmetrical structure network, including a downsampling module, a Swin Transformer Block module, and an upsampling module. The Swin Transformer Block module is composed of two consecutive blocks, including a first connection block and a second connection block. Both connection blocks contain Layer Normalization and Multi-Layer Perceptron. Layer Normalization is used for normalization operations, which are used for nonlinear mapping and feature extraction.

[0108] like Figure 6 As shown, the first connection block contains the window self-attention calculation W-MSA, which performs self-attention calculation in each window so that the model pays attention to the features in the region;

[0109] The second connection block contains the translation window self-attention calculation SW-MSA. The moving window enables the features between different windows to interact and enhance the capture of global information. The connection of two consecutive blocks extracts complex global and local information.

[0110] The sampling part uses average pooling and convolution operations to reduce the size of the feature map. The upsampling module uses a transposed convolution operation to restore the image to its original size to restore image detail information. The upsampling module and the downsampling module use a skip connection for residual connection to prevent the problems of gradient disappearance and gradient explosion during training.

[0111] The specific operation of the SwinTrans-CNN denoising model for denoising tasks is:

[0112] Step S4-1: dividing the two-dimensional complex matrix of the intercepted pilot reception range into two channels of real part and imaginary part;

[0113] Step S4-2: The input data first passes through the FCN module. After multiple convolution operations, the noise level of the input image is extracted and predicted, and a noise feature map is output;

[0114] Step S4-3: Add the noise feature map to the original image, and the added multi-channel image becomes the input of the U-Net module;

[0115] Step S4-4: the original image with the noise feature map is first passed through two consecutive downsampling modules to reduce the resolution of the feature map, facilitate multi-scale fusion and increase the receptive field;

[0116] Step S4-5: The low-resolution image output by the downsampling module is input to the Swin Transformer Block module. By combining local attention and global context, the input features are denoised while retaining important local and global feature information;

[0117] Step S4-6: After the image is output through the Swin Transformer Block module, the denoised low-resolution image is restored to the original image size through two consecutive upsampling modules, and the original image detail information is retained.

[0118] Step S5: Use the trained SwinTrans-CNN denoising model to perform channel state estimation;

[0119] At the receiving end, the test set data is used as the input of the trained model, and the output result is the denoised delay-Doppler domain impulse response matrix, from which the current channel state information can be obtained.

[0120] like Figure 4 As shown, the red grid points are the locations of the pilots at the transmitter, and the yellow grid points are , is the range of pilot reception, in the delay-Doppler domain, the two-dimensional transmission signal and receive signals The relationship can be expressed as: .in, is the delay-Doppler domain impulse response matrix, is a two-dimensional circular convolution, is the additive white Gaussian noise matrix in the delay-Doppler domain;

[0121] The yellow grid area can be represented as a whole as follows: , if the noise can be eliminated, the delay-Doppler domain channel impulse response can be obtained , which is the channel state information required for channel estimation. Therefore, by setting pilot symbols and protection symbols in advance at the transmitting end, the channel estimation problem of OTFS can be converted into a two-dimensional image denoising problem at the receiving end for processing. The channel estimation task of OTFS can be completed by subsequent design and training of the denoising neural network model.

[0122] In the process of verifying the feasibility of the method of the present invention, the simulation experiment of the present invention is implemented using the pytorch framework:

[0123] Design TDL channel model:

[0124] The channel estimation scheme designed by the present invention is modeled on the delay-Doppler domain channel to simulate a high-speed motion scene with 6 multipaths. It is hoped that the channel parameters can be flexibly designed according to the simulation parameters and the actual channel scene. Therefore, the tapped delay line TDL channel model is adopted to randomly generate the pulse response of the multipath, and the channel gain of each multipath is a time-varying function, and the speed of change is related to the relative speed of the scatterer represented by the multipath. The greater the speed, the faster the change, and the smaller the speed, the slower the change. At the same time, the impulse response matrix of the random channel is obtained. , the impulse response matrix The size is 13*13. In order to facilitate subsequent calculations, the channel gain is normalized for Hw, and the impulse response sample data corresponding to the 6 multipaths of Hw is generated as follows:

[0125] -0.13562687-0.54733448j;

[0126] 0.33070939+0.06927313j;

[0127] 0.44903339+0.01067447j;

[0128] 0.0242942 +0.05105517j;

[0129] -0.33244615-0.41147718j;

[0130] -0.28249367-0.05731045j.

[0131] Designed to send signal:

[0132] Delay-Doppler domain transmitted signal For a 16QAM symbol of size 64*64, the grid resources in the delay-Doppler domain are set according to the set parameters, and Figure 1 The pilot symbols and guard symbols are set in a manner.

[0133] Get the received signal:

[0134] The discrete two-dimensional convolution channel impulse response matrix Hw, plus additive Gaussian white noise under a 20dB signal-to-noise ratio environment, obtains the received signal ,according to Figure 2 The channel estimation area is intercepted in a manner as the channel estimation area.

[0135] Generate training dataset:

[0136] Generate 1000 channel estimation regions in batches and divide them into real and imaginary parts as training input data sets. Generate 1000 corresponding channel estimation regions as training labels. The training indicators are output and The mean square error between the two, the ratio of the number of training data and test data is 8:2.

[0137] Offline training of SwinTrans-CNN model:

[0138] The input is two 2D grayscale images of the real and imaginary parts of the channel estimation area. They first go through four convolutional layers in sequence. The number of output channels after each convolutional layer is 32, 16, 8 and 4 respectively. Each convolutional layer has 32 convolution kernels with a convolution kernel size of 3×3 and a padding of 1. There is no bias, and the ReLU activation function is used. The forward propagation process is defined. The output 4-channel feature image plus the original two-channel input form a six-channel input to the subsequent U-Net module. The downsampling part of the U-Net module uses average pooling and convolution operations to reduce the feature map size, and the upsampling part uses transposed convolution operations to restore the image to its original size to restore image detail information. The skip connection is used for residual connection between upsampling and downsampling. In the Swin Transformer Block, the window size is 7*7, the number of input channels is 256, the input feature map size is 3×3, the number of multi-heads in the multi-head attention mechanism is 8, and the LN layer is used to standardize the input features. The model loss function is the mean square error function, the learning rate is set to 0.001, and the number of training epochs is 100.

[0139] Online application model:

[0140] The model inputs a specific area of ​​the receiving end signal, and the denoised delay-Doppler domain impulse response matrix can be obtained. If the matrix If the value is not zero, it means that the Doppler tap and delay tap There is a path where the delay is , the Doppler value is , the complex channel gain is , at this time the acquisition of channel state information is completed.

[0141] like Figure 8 As shown in FIG. 1 , the present invention is compared with the least square method, the threshold-based embedded pilot method and the OTFS channel estimation scheme based on the DnCNN model, and the normalized mean square error between the channel estimation results and the true value of these four methods in the case of 5dB to 30dB signal-to-noise ratio is compared. The simulation results are shown in FIG. Figure 8 As shown, by comparing the results, it can be concluded that the OTFS channel estimation accuracy based on the deep learning scheme represented by the SwinTrans-CNN model and the DnCNN model of the present invention is better than the traditional channel estimation scheme, and the performance of the SwinTrans-CNN model of the present invention is better than the DnCNN model.

[0142] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. An OTFS channel estimation method based on the SwinTrans-CNN model, characterized by: The method comprises the following steps: Step S1: Determine basic channel parameters according to specific channel scenarios and design a TDL channel model; Step S2: Designing the delay-Doppler domain signal of the transmitting end by using the embedded pilot method; Step S3: performing two-dimensional circular convolution and Gaussian white noise processing on the delay-Doppler domain signal obtained by transmission to obtain received information; Step S4: construct a SwinTrans-CNN denoising model and perform training; The SwinTrans-CNN denoising model includes an FCN module and a U-Net module, and the SwinTrans-CNN denoising model has dual-channel input and dual-channel output; The FCN module includes four convolutional layers, each of which has 32 convolution kernels, a convolution kernel size of 3×3, a padding of 1, no bias, a ReLU activation function, and defines a forward propagation process; The U-Net module includes a downsampling module, a Swin Transformer Block module and an upsampling module. The Swin Transformer Block module is composed of two continuous blocks, including a first connection block and a second connection block. Both connection blocks contain Layer Normalization and Multi-Layer Perceptron. Layer Normalization is used for normalization operation, which is used for nonlinear mapping and feature extraction. The first connection block includes a window self-attention calculation W-MSA, which performs self-attention calculation in each window so that the model pays attention to the features in the region; The second connection block contains the translation window self-attention calculation SW-MSA. The moving window enables the features between different windows to interact and enhance the capture of global information. The connection of two consecutive blocks extracts complex global and local information. Step S5: Use the trained SwinTrans-CNN denoising model to perform channel state estimation.

2. The OTFS channel estimation method based on the SwinTrans-CNN model according to claim 1, characterized in that: In step S1, basic channel parameters are determined according to a specific channel scenario, and a TDL channel model is designed. The specific operations are: Step S1-1: Obtain the maximum moving speed V of the mobile carrier in the scene max , according to the 5G standard in the area, the carrier frequency f of the communication system in the scenario is obtained c and subcarrier spacing Δf; The subcarrier spacing Δf is obtained by the symbol period, and the symbol period T = 1 / Δf; Step S1-2: Obtain the maximum Doppler frequency shift f in the scenario dmax and maximum delay spread; Maximum Doppler shift Where c is the speed of light, V max is the maximum moving speed, fc is the carrier frequency; Maximum delay spread Where L is the length of the communication link; Step S1-3: According to the size of the time-frequency domain grid resources of the 5G application OFDM technology in the actual scenario, the resource size of the corresponding delay-Doppler domain grid is obtained; The delay resolution is obtained by the correspondence between OFDM time-frequency domain grid resources and delay-Doppler domain grid resources. and Doppler resolution Where M is the number of grids in the delay domain, and N is the number of grids in the Doppler domain; The correspondence between OFDM time-frequency domain grid resources and delay-Doppler domain grid resources is: the number of time domain grids N corresponds to the number of Doppler domain grids N, the number of frequency domain grids M corresponds to the number of delay domain grids M, the Doppler domain grid length is the Doppler resolution Δv, and the delay domain grid length is the delay resolution Δτ.

3. The OTFS channel estimation method based on the SwinTrans-CNN model according to claim 1, characterized in that: In step S2, the embedded pilot method is used to design the delay-Doppler domain signal of the transmitting end, and the specific operation is: Step S2-1: Place the transmission signal in the delay-Doppler domain grid resource allocated at the transmitting end. Step S2-2: performing embedded pilot symbol setting according to the determined parameters; The setup scheme is as follows: The delay value is a non-negative number, and the Doppler frequency shift value is a negative or non-negative number. To avoid mutual influence between the pilot and the transmitted data at the receiving end, the delay domain guard interval is a single maximum delay value, and the Doppler domain guard interval is a double maximum Doppler frequency shift value; The frame structure of the delay-Doppler domain signal is: Among them, x p is the pilot frequency, x d is the data to be sent, l is the index of the Doppler frequency shift value on the grid, l p is the index of the pilot delay value coordinate, l max is the index of the maximum delay; k is the index of the delay value on the grid, k p is the index of the pilot Doppler value coordinate; k max is the index of the maximum Doppler shift on the grid.

4. The OTFS channel estimation method based on the SwinTrans-CNN model according to claim 1, characterized in that: In step S3, the delay-Doppler domain signal obtained by transmission is subjected to two-dimensional circular convolution and Gaussian white noise processing to obtain received information. The specific operation is as follows: Step S3-1: Send signal X DD With the channel impulse response matrix H W Perform a two-dimensional circular convolution to get Y DD _no_noise; Step S3-2: For Y DD _no_noise plus additive Gaussian white noise to get the received signal Y DD ; Step S3-3: In the delay-Doppler domain, the two-dimensional transmission signal Y DD and receive signal X DD The relationship is expressed as: Y DD =X DD *H W +N DD ; Among them, H W is the delay-Doppler domain impulse response matrix, * is the two-dimensional circular convolution, N DD is the additive white Gaussian noise matrix in the delay-Doppler domain.

5. The OTFS channel estimation method based on the SwinTrans-CNN model according to claim 1, characterized in that: The downsampling module uses average pooling and convolution operations to reduce the size of the feature map, and the upsampling module uses a transposed convolution operation to restore the image to its original size to restore image detail information; The up-sampling module and the down-sampling module are connected by skip connection to prevent the problems of gradient disappearance and gradient explosion during training.

6. The OTFS channel estimation method based on the SwinTrans-CNN model according to claim 1, characterized in that: In step S4, a SwinTrans-CNN denoising model is constructed and trained. The specific operations are as follows: Step S4-1: dividing the two-dimensional complex matrix of the intercepted pilot reception range into two channels of real part and imaginary part; Step S4-2: The input data first passes through the FCN module. After multiple convolution operations, the noise level of the input image is extracted and predicted, and a noise feature map is output; Step S4-3: Add the noise feature map to the original image, and the added multi-channel image becomes the input of the U-Net module; Step S4-4: the original image with the noise feature map is first passed through two consecutive downsampling modules to reduce the resolution of the feature map, facilitate multi-scale fusion and increase the receptive field; Step S4-5: The low-resolution image output by the downsampling module is input to the Swin Transformer Block module. By combining local attention and global context, the input features are denoised while retaining important local and global feature information; Step S4-6: After the image is output through the Swin Transformer Block module, the denoised low-resolution image is restored to the original image size through two consecutive upsampling modules, and the original image detail information is retained.

7. The OTFS channel estimation method based on the SwinTrans-CNN model according to claim 1, characterized in that: When the SwinTrans-CNN denoising model is trained, the delay and Doppler shift range of the pilot signal are intercepted in the area of ​​the receiving end; The range of pilot reception is l∈(l p , l p +l max ), k∈(k p -k max , k p +k max ), the pilot reception range is represented as δ(l p ,k p )*H W +N DD =H W +N DD ; N DD is the additive white Gaussian noise matrix in the delay-Doppler domain; The range of each independent pilot reception at the receiving end is: Among them, b[kk p ,ll p ] is the path identifier, when the Doppler tap kk p Delay tap ll p When there is a path at p ,ll p ]=1, otherwise b[kk p ,ll p ]=0;h[kk p ,ll p ] is the Doppler tap kk p Delay tap ll p The impulse response at the coordinate grid point (k, l) is n(k, l); j is an imaginary number; M is the number of grids in the delay domain, and N is the number of grids in the Doppler domain; The two-dimensional complex matrix of the range area of ​​pilot reception is intercepted as the input of the SwinTrans-CNN model.

Images