A channel estimation method and system based on orthogonal time-frequency space modulation technique

Abstract

The present disclosure relates to a channel estimation method based on orthogonal time-frequency-space modulation technology. The method comprises establishing a model of orthogonal time-frequency-space channel estimation, which comprises a time-delay Doppler domain received signal, a sparsity adaptive matching pursuit algorithm and a denoising network; processing the time-delay Doppler domain received signal by using the sparsity adaptive matching pursuit algorithm to reconstruct a delay Doppler domain channel; and inputting the delay Doppler domain channel into the denoising network for denoising processing to obtain a denoised channel response. The method can reconstruct the delay Doppler domain channel without relying on prior information of the number of channel paths, and can eliminate the corresponding noise of the delay Doppler domain channel.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of information communication technology, and in particular, to a channel estimation method and system based on an orthogonal time-frequency-space modulation technique. BACKGROUND

[0002] Space-Air-Ground Integrated Networks (SAGINs) can overcome the limitations of ground mobile communication networks, such as limited coverage, insufficient security and reliability, poor survivability, and other shortcomings, and meet the demand for regional coverage, which is a necessary approach to achieve global coverage and high-speed information transmission. The implementation of SAGINs involves many emerging applications such as unmanned aerial vehicles, vehicle-to-everything (V2X), low earth orbit (LEO) satellites, etc.

[0003] However, the high mobility of terminals in V2X and LEO satellite scenarios can cause different degrees of signal attenuation. When the mobile terminal is in high-speed motion, the wireless channel has the characteristics of rapid change due to the existence of Doppler shift. In this fast time-varying channel, the orthogonality of subcarriers in the existing Orthogonal Frequency Division Multiplexing (OFDM) system is destroyed, and the performance is significantly reduced.

[0004] The existing Orthogonal Time Frequency Space (OTFS) modulation technique can reconstruct the time-delay-Doppler domain channel, but it is difficult to eliminate noise on non-zero elements in the presence of time-delay-Doppler domain channel response, thereby causing the problem of insufficient channel estimation performance. In addition, the existing OTFS modulation technique often relies on prior information about the number of channel paths, but in actual communication systems, channel parameters change with factors such as operating frequency and propagation environment, and the number of channel paths is often unknown. Therefore, it is necessary to propose a solution to improve one or more problems in the above related technical solutions.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] A first aspect of the embodiments of the present disclosure provides a channel estimation method based on an orthogonal time-frequency-space modulation technique, comprising the following steps:

[0007] A model of orthogonal time-frequency space channel estimation is established, the model comprising a time-delay Doppler domain received signal, a sparsity adaptive matching pursuit algorithm and a denoising network;

[0008] The time-delay Doppler domain received signal is processed by using the sparsity adaptive matching pursuit algorithm, and a delay Doppler domain channel is reconstructed.

[0009] The delay Doppler domain channel is input into the denoising network for denoising processing, and a denoised channel response is obtained.

[0010] In an example embodiment of the present disclosure, in the step of processing the time-delay Doppler domain received signal by using the sparsity adaptive matching pursuit algorithm and reconstructing a delay Doppler domain channel,

[0011] The initial input information of the sparsity adaptive matching pursuit algorithm comprises a sensing matrix α, an observation vector y, an initial index set C n-1 and an initial support set , wherein n = 1, that is, the initial index set C n-1 is C 0 , and the initial support set is The initial index set C 0 and the initial support set are set as empty sets, and a support set is defined, and the size of the support set is p, the initial size of the initial support set is P = 1, the threshold value T = 0.001, and the observation vector y is an initial residual r0.

[0012] In an example embodiment of the present disclosure, the step of processing the time-delay Doppler domain received signal by using the sparsity adaptive matching pursuit algorithm and reconstructing a delay Doppler domain channel comprises:

[0013] The correlation of all elements in the sensing matrix α and a current residual r n-1 is calculated, and a correlation set μ is obtained, and the expression of the correlation set μ comprises: μ = <r n-1 , α >.

[0014] A plurality of maximum value elements are extracted from the correlation set μ, wherein the index values corresponding to the plurality of maximum value elements constitute an index value set A n , A n = {A n1 , A n2 ,... A nP}, the index value set A n is merged with the initial index set C n -1 to obtain an index set Cn , C n =C n-1 ∪A n ; wherein n is an integer greater than or equal to 1, when n = 1, C 0 is the initial index set; the number of extracted maximum elements corresponds to the size p of the support set ;

[0015] adding the vector in the corresponding perception matrix a of the index set C n to the support set , obtaining

[0016] reconstructing the delay-Doppler domain channel from the support set and the observation vector y The expression of the delay-Doppler domain channel includes:

[0017]

[0018] wherein T represents a threshold value; -1 represents an inverse matrix;

[0019] The item with the largest absolute value in the delay-Doppler domain channel is taken as the estimation result of the current iteration and the estimation residual r n of the delay-Doppler domain channel of the current iteration is calculated n The expression of the estimation residual r includes:

[0020]

[0021] If the estimation residual r n satisfies: ||r n ||≤T, the iteration is stopped;

[0022] If the estimation residual r n satisfies: ||r n ||≥||r n-1 ||, the iteration is continued after the size p of the support set is increased by 1;

[0023] In other cases, the next iteration is directly performed.

[0024] In an example embodiment of the present disclosure, the step of inputting the delay-Doppler domain channel into the denoising network for denoising processing to obtain a denoised channel response includes:

[0025] The denoising network is established, and the denoising network comprises an Inception-ResNet denoising network model, the Inception-ResNet denoising network model comprises an initial convolutional layer, a plurality of Inception-ResNet modules with the same structure and a full connection layer;

[0026] The plurality of Inception-ResNet modules with the same structure comprise a first Inception-ResNet module, a second Inception-ResNet module and a third Inception-ResNet module connected in sequence, each of the Inception-ResNet modules comprises a splicing layer, the initial convolutional layer is connected with the first Inception-ResNet module, and the full connection layer is connected with the third Inception-ResNet module.

[0027] The estimation result is input into the Inception-ResNet denoising network model, and the denoised channel response is obtained.

[0028] In an example embodiment of the present disclosure, the step of inputting the estimation result into the Inception-ResNet denoising network model to obtain the denoised channel response comprises:

[0029] The estimation result is input into the initial convolutional layer to obtain the extraction feature F0.

[0030] The extraction feature F0 is input into the plurality of Inception-ResNet modules, and the splicing and compression feature F is obtained after splicing and compression processing of the plurality of Inception-ResNet modules.

[0031] The splicing and compression feature F is input into the full connection layer, the full connection layer performs integration processing on the splicing and compression feature F, and the denoised channel response is obtained.

[0032] In an example embodiment of the present disclosure, the step of inputting the estimation result into the initial convolutional layer to obtain the extraction feature F0 comprises:

[0033] The expression of the extraction feature F0 comprises:

[0034] ​​​​​​

[0035] wherein F0 represents the extracted feature of the initial convolutional layer; f conv represents the role of the convolutional layer; represents the estimation result of iteration; the extracted feature F0 is taken as the input of the Inception-ResNet module.

[0036] In an example embodiment of the present disclosure, in the step of inputting the extracted feature F0 into a plurality of Inception-ResNet modules, after the splicing and compression processing of the plurality of Inception-ResNet modules, a splicing and compression feature is obtained.

[0037] Each of the Inception-ResNet modules includes two feature extraction paths respectively.

[0038] The first feature extraction path includes a first convolutional layer, a second convolutional layer and a third convolutional layer; wherein the first convolutional layer is a 3x3 convolutional layer including 32 convolutional kernels; the second convolutional layer is a 1x7 convolutional layer including 48 convolutional kernels; and the third convolutional layer is a 7x1 convolutional layer including 64 convolutional kernels.

[0039] The second feature extraction path includes a fourth convolutional layer, which is a 3x3 convolutional layer including 64 convolutional kernels.

[0040] Wherein, after each convolutional layer in each of the feature extraction paths, a Gaussian error linear unit is connected respectively.

[0041] The first feature extraction path and the second feature extraction path are connected with the splicing layer respectively, and the splicing layer is used for splicing processing of the first feature extraction path and the second feature extraction path to obtain a splicing feature; the splicing layer is connected with a fifth convolutional layer, which is a 1x1 convolutional layer including 64 convolutional kernels, and the fifth convolutional layer is used for compression processing of the splicing feature.

[0042] The plurality of Inception-ResNet modules adopt a connection mode of skipping each convolutional layer in the feature extraction path to take the splicing and compression feature obtained after the splicing and compression processing as the output of the plurality of Inception-ResNet modules.

[0043] In an example embodiment of the present disclosure, in the step of inputting the extracted feature F0 into a plurality of Inception-ResNet modules, after the splicing and compression processing of the plurality of Inception-ResNet modules, a splicing and compression feature is obtained.

[0044] inputting the first extracted feature F into the second Inception-ResNet module to obtain an output second extracted feature F

[0045]

[0046] wherein F c1 represents a compressed feature of the first Inception-ResNet module;

[0047] inputting the second extracted feature F into the second Inception-ResNet module to obtain an output second extracted feature F

[0048]

[0049] wherein F c2 represents a compressed feature of the second Inception-ResNet module;

[0050] inputting the second feature F into the third Inception-ResNet module to obtain an output spliced compressed feature F

[0051]

[0052] wherein F c3 represents a compressed feature of the third Inception-ResNet module.

[0053] In an example embodiment of the present disclosure, inputting the spliced compressed feature F into the full connection layer, the full connection layer integrates the spliced compressed feature F to obtain a denoised channel response H

[0054]

[0055]

[0056] wherein w FC represents a weight of the full connection layer; and b FC ​​​​​​​​bias of the fully connected layer.

[0057] A second aspect of the embodiments of the present disclosure provides a channel estimation system based on an orthogonal time frequency space modulation technology. The channel estimation system comprises a signal subsystem, a sparsity adaptive matching pursuit subsystem and a denoising network subsystem; the sparsity adaptive matching pursuit subsystem is connected with the signal subsystem and the denoising network subsystem respectively; wherein,

[0058] The signal subsystem is configured to receive a time delay Doppler domain received signal; and input the time delay Doppler domain received signal to the sparsity adaptive matching pursuit subsystem;

[0059] The sparsity adaptive matching pursuit subsystem is configured to provide a sparsity adaptive matching pursuit algorithm, and process the time delay Doppler domain received signal by using the sparsity adaptive matching pursuit algorithm to reconstruct a delay Doppler domain channel;

[0060] The denoising network subsystem is configured to perform denoising processing on the input delay Doppler domain channel to obtain a denoised channel response.

[0061] The technical solution provided by the present disclosure can include the following beneficial effects: the channel estimation method based on the orthogonal time frequency space modulation technology proposed by the present disclosure reconstructs the delay Doppler domain channel by establishing an OTFS channel estimation model and using the sparsity adaptive matching pursuit algorithm, and puts the constructed delay Doppler domain channel into the denoising network for denoising processing to obtain the denoised channel response, thereby realizing the reconstruction of the delay Doppler domain channel without relying on the prior information of the number of channel paths, and eliminating the noise of the initial OTFS channel response. BRIEF DESCRIPTION OF DRAWINGS

[0062] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure and serve to explain the principles of the present disclosure. It is apparent that the accompanying drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0063] Figure 1 A step schematic diagram of the channel estimation method based on the orthogonal time frequency space modulation technology in the exemplary embodiments of the present disclosure is shown;

[0064] Figure 2 A structure schematic diagram of the Inception-ResNet denoising network model in the exemplary embodiments of the present disclosure is shown;

[0065] Figure 3A structural schematic diagram of each Inception-ResNet module in the Inception-ResNet denoising network model of the exemplary embodiment of the present disclosure is shown.

[0066] Figure 4 A schematic diagram of a channel estimation system based on an orthogonal time-frequency space modulation technique in the exemplary embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0067] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these examples are provided so that this disclosure will be thorough and complete, and will fully convey the inventive concept to those skilled in the art. The described features, structures, or characteristics can be combined in one or more implementations.

[0068] In addition, the accompanying drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification, illustrate embodiments described herein, and together with the description serve to explain principles of the present disclosure. Features, elements, and / or properties that are the same in the various embodiments are labeled with the same reference numbers throughout the drawings, and descriptions of these parts will not be repeated. Some of the block diagrams in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. These functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0069] A first aspect of the present example implementation provides a channel estimation method based on an orthogonal time-frequency space (OTFS) modulation technique, as shown in Figure 1 The method can include the following steps:

[0070] Step S101: A model of orthogonal time-frequency space channel estimation is established, which includes a time-delay Doppler domain received signal, a sparsity adaptive matching pursuit algorithm, and a denoising network.

[0071] Step S102: The time-delay Doppler domain received signal is processed using a sparsity adaptive matching pursuit (SAMP) algorithm to reconstruct a delay-Doppler (DD) domain channel.

[0072] Step S103: The delay-Doppler domain channel is input into the denoising network for denoising processing to obtain a denoised channel response.

[0073] The first aspect of the embodiments of the present disclosure provides a channel estimation method based on an orthogonal time frequency space modulation technology. The method reconstructs a delay Doppler domain channel by establishing an OTFS channel estimation model and using a sparsity adaptive matching pursuit algorithm, and puts the constructed delay Doppler domain channel into a denoising network for denoising processing, so as to obtain a denoised channel response, thereby realizing reconstruction of the delay Doppler domain channel without relying on prior information of a channel path number, and being capable of eliminating corresponding noise of the delay Doppler domain channel.

[0074] In the following, each step of the above method in the present example embodiment will be described in more detail.

[0075] In step S101, in order to be able to solve the above technical problem, the channel estimation method based on the OTFS modulation technology proposed by the embodiments of the present disclosure first establishes an OTFS channel estimation model, which includes a time delay Doppler domain received signal, a sparsity adaptive matching pursuit algorithm and a denoising network. Then, the channel estimation method is divided into the following two stages: the first stage is step S102, and the second stage is step S103.

[0076] In step S102, in the first stage, the SAMP algorithm is used to process the time delay Doppler domain received signal. The algorithm compares the residual after each iteration with the residual of the previous iteration, and uses a threshold to estimate the sparsity of the OTFS channel. Therefore, reconstruction of the DD channel can be realized without prior information about the number of channel paths.

[0077] Specifically, the initial input information of the SAMP algorithm includes: a perception matrix α, an observation vector y, an initial index set C n-1 and an initial support set Wherein, n = 1, that is, the initial index set C n-1 is C 0 , the initial support set is The initial index set C 0 and the initial support set are set to empty sets, and the support set is set to have a size of p. Therefore, the initial size of the initial support set is P = 1, the threshold T = 0.001, and the observation vector y is the initial residual r0.

[0078] The step S102 includes the following sub-steps:

[0079] Sub-step S1021: calculate the correlation of all elements in the perception matrix α and the current residual r n-1 , to obtain a correlation set μ. The expression of the correlation set μ includes: μ = <r n-1 , α >.

[0080] Sub-step S1022: extracting several maximum value elements from the correlation set μ, wherein the index values corresponding to the several maximum value elements constitute an index value set A n , A n ={A n1 ,A n2 ,...A nP}; merging the index value set A n with the initial index set C n-1 to obtain an index set C n , C n =C n-1 ∪A n ; wherein n is an integer greater than or equal to 1, and when n=1, C 0 is the initial index set; the number of the maximum value elements corresponds to the size p of the support set ;

[0081] Sub-step S1023: adding the vector in the perception matrix α corresponding to the index set C n to the support set to obtain

[0082] Sub-step S1024: reconstructing the delay-Doppler domain channel from the support set and the observation vector y The expression of the delay-Doppler domain channel

[0083]

[0084] wherein T represents a threshold value; -1 represents an inverse matrix;

[0085] Sub-step S1025: taking the item with the largest absolute value in the DD domain channel as the estimation result of the current iteration and calculating the estimation residual r n of the DD domain channel of the current iteration n The expression of the estimation residual r

[0086]

[0087] According to the calculation result of the estimation residual r n , there can be three cases:

[0088] If the estimation residual r n satisfies: ||r n ||≤T, then the iteration is stopped;

[0089] If the estimation residual rn satisfies: ||r n ||≥||r n-1 ||, then the support set After the size P of the scale is increased by 1, the iteration is continued;

[0090] Otherwise, the next iteration is directly performed.

[0091] Step S102 realizes that the DD domain channel can be reconstructed without prior information about the number of channels.

[0092] In step S103, an Inception-ResNet denoising network model is designed here, as shown in Figure 2 The model includes an initial convolutional layer, a plurality of Inception-ResNet modules with the same structure, and a fully connected layer.

[0093] The plurality of Inception-ResNet modules with the same structure include a first Inception-ResNet module, a second Inception-ResNet module, and a third Inception-ResNet module connected in sequence; each Inception-ResNet module includes a splicing layer; the initial convolutional layer is connected to the first Inception-ResNet module; and the fully connected layer is connected to the third Inception-ResNet module.

[0094] The estimation result is input into the Inception-ResNet denoising network model to obtain a denoised channel response

[0095] Further, step S103 includes the following sub-steps:

[0096] Sub-step S1031: input the estimation result into the initial convolutional layer to obtain extracted features F0;

[0097] Sub-step S1032: input the extracted features F0 into the plurality of Inception-ResNet modules, and obtain spliced and compressed features

[0098] Sub-step S1033: input the spliced and compressed features into the fully connected layer, and the fully connected layer integrates the spliced and compressed features to obtain the denoised channel response

[0099] Further, in sub-step S1031, the estimation result is input into the initial convolutional layer to obtain the extraction feature F0 in the step of obtaining the expression of the extraction feature F0, which includes:

[0100]

[0101] wherein F0 represents the extraction feature of the convolutional layer; f conv represents the effect of the convolutional layer; represents the iterative estimation result; the extraction feature F0 is taken as the input of the plurality of Inception-ResNet modules.

[0102] In sub-step S1032, as shown in Figure 3 , each Inception-ResNet module includes two feature extraction paths respectively.

[0103] wherein Figure 3 It can be seen that each Inception-ResNet module designed in this way connects the convolutional layers with different sizes of convolutional kernels in parallel, which can fully mine the detailed features of the input data. Specifically, the first feature extraction path includes a first convolutional layer, a second convolutional layer and a third convolutional layer, wherein the first convolutional layer is a 3x3 convolutional layer including 32 convolutional kernels; the second convolutional layer is a 1x7 convolutional layer including 48 convolutional kernels; and the third convolutional layer is a 7x1 convolutional layer including 64 convolutional kernels. The second feature extraction path includes a fourth convolutional layer, which is a 3x3 convolutional layer including 64 convolutional kernels. It can be seen that the Inception-ResNet module designed in this way combines the extraction features of different convolutional layers in the deep dimension, which can improve the accuracy of the deep learning network and prevent overfitting. In addition, by using asymmetric kernels, the amount of calculation can be reduced without changing the size of the receptive field.

[0104] At the same time, each convolutional layer in each feature extraction path is followed by a Gaussian Error Linear Unit (GELU) as an activation function. GELU means that the idea of introducing random rules is introduced as a probability description of the neuron input, which makes it have better generalization ability than the Rectified Linear Unit (RELU).

[0105] The first feature extraction path and the second feature extraction path are connected with a splicing layer, and the splicing layer is configured to perform splicing processing on the first feature extraction path and the second feature extraction path to obtain spliced features; the splicing layer is connected with a fifth convolutional layer, the fifth convolutional layer is a 1*1 convolutional layer including 64 convolutional kernels, and the fifth convolutional layer is configured to perform compression processing on the spliced features;

[0106] The plurality of Inception-ResNet modules adopt a jump connection mode to obtain spliced and compressed features from the spliced and compressed features as the output of the plurality of Inception-ResNet modules.

[0107] Here, the jump connection refers to a connection mode of each convolutional layer in the jump feature extraction path of the plurality of Inception-ResNet.

[0108] Further, the extracted features F0 are input into the first Inception-ResNet module to obtain output first extracted features The expression of the first extracted features includes:

[0109]

[0110] wherein F0 represents the extracted features of the convolutional layer, F c1 represents the compressed features of the first Inception-ResNet module;

[0111] The first extracted features are input into the second Inception-ResNet module to obtain output second extracted features The expression of the second extracted features includes:

[0112]

[0113] wherein F c2 represents the compressed features of the second Inception-ResNet module;

[0114] The second features are input into the third Inception-ResNet module to obtain output spliced and compressed features The expression of the spliced and compressed features includes:

[0115]

[0116] wherein F c3compressed features representing the third Inception-ResNet module.

[0117] In step S1033, the spliced compressed features are input to a full connection layer, which integrates the spliced compressed features to obtain a denoised channel response In the step of

[0118] the denoised channel response has an expression including:

[0119]

[0120] where w FC represents a weight of the full connection layer; and b FC represents a bias of the full connection layer.

[0121] The second aspect of the embodiments of the present disclosure proposes a channel estimation system based on an orthogonal time-frequency space modulation technology, as shown in Figure 4

[0122] The channel estimation system includes a signal subsystem, a sparsity adaptive matching pursuit subsystem, and a denoising network subsystem; the sparsity adaptive matching pursuit subsystem is connected with the signal subsystem and the denoising network subsystem; wherein

[0123] The signal subsystem is configured to receive a time-delay Doppler domain received signal; and input the time-delay Doppler domain received signal to the sparsity adaptive matching pursuit subsystem

[0124] The sparsity adaptive matching pursuit subsystem is configured to provide a sparsity adaptive matching pursuit algorithm, and utilize the sparsity adaptive matching pursuit algorithm to process the time-delay Doppler domain received signal, and reconstruct a time-delay Doppler domain channel.

[0125] The denoising network subsystem is configured to perform denoising processing on the input time-delay Doppler domain channel to obtain a denoised channel response.

[0126] It should be noted that, although several units of the system for action execution are mentioned in the foregoing detailed description, such division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided into several units embodied. Part or all of the units can be selected to achieve the purpose of the present disclosure according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0127] ​Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the features disclosed herein. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

Claims

1. A method for channel estimation based on orthogonal time and frequency space modulation technique, characterized in that, The method comprises the following steps: A model of orthogonal time-frequency space channel estimation is established, which comprises a time-delay Doppler domain received signal, a sparsity adaptive matching pursuit algorithm and a denoising network; The time-delay Doppler domain received signal is processed by using the sparsity adaptive matching pursuit algorithm to reconstruct a delay Doppler domain channel; The delay Doppler domain channel is input into the denoising network for denoising processing to obtain a denoised channel response, which comprises: The denoising network is established, which comprises an Inception-ResNet denoising network model, the Inception-ResNet denoising network model comprising an initial convolutional layer, a plurality of Inception-ResNet modules with the same structure and a fully connected layer; The plurality of Inception-ResNet modules with the same structure comprise a first Inception-ResNet module, a second Inception-ResNet module and a third Inception-ResNet module connected in sequence; each Inception-ResNet module comprises a splicing layer; the initial convolutional layer is connected with the first Inception-ResNet module; The fully connected layer is connected with the third Inception-ResNet module; The estimation result is input into the Inception-ResNet denoising network model to obtain the denoised channel response , comprising: The estimation result is input into the initial convolutional layer to obtain an extraction feature The extraction feature is input into a plurality of Inception-ResNet modules to obtain a spliced and compressed feature after splicing and compression processing of the plurality of Inception-ResNet modules The spliced and compressed feature is input into a plurality of Inception-ResNet modules to obtain a spliced and compressed feature after splicing and compression processing of the plurality of Inception-ResNet modules Each Inception-ResNet module includes two feature extraction paths​ A first feature extraction path comprises a first convolutional layer, a second convolutional layer and a third convolutional layer; The first convolutional layer is a 3*3 convolutional layer comprising 32 convolutional kernels; the second convolutional layer is a 1*7 convolutional layer comprising 48 convolutional kernels; and the third convolutional layer is a 7*1 convolutional layer comprising 64 convolutional kernels; A second feature extraction path comprises a fourth convolutional layer, which is a 3*3 convolutional layer comprising 64 convolutional kernels; Each convolutional layer in each feature extraction path is connected with a Gaussian error linear unit respectively; the first feature extraction path and the second feature extraction path are connected with the splicing layer, which is used for splicing the first feature extraction path and the second feature extraction path to obtain spliced features; the splicing layer is connected with a fifth convolutional layer, which is a 1*1 convolutional layer comprising 64 convolutional kernels, and the fifth convolutional layer is used for compressing the spliced features; The plurality of Inception-ResNet modules employ a skip connection manner of jumping each convolutional layer in the feature extraction path to obtain the spliced compressed features after the splicing compression processing as the output of the plurality of Inception-ResNet modules The concatenated compressed features to the fully connected layer, which integrates the concatenated compressed features to obtain a de-noised channel response .

2. The channel estimation method of claim 1, wherein The step of processing the time delay Doppler domain received signal by using the sparsity adaptive matching pursuit algorithm to reconstruct the delay Doppler domain channel, the initial input information of the sparsity adaptive matching pursuit algorithm includes: a sensing matrix , an observation vector y , an initial index set and an initial support set , wherein , the initial index set is , the initial support set is ; the initial index set and the initial support set are set as empty sets, the size of the support set is P , the initial size of the initial support set is , a threshold value , and the observation vector y is an initial residual .

3. The channel estimation method of claim 2, wherein The step of processing the delay-Doppler domain received signal by the sparsity adaptive matching pursuit algorithm to reconstruct the delay-Doppler domain channel comprises: calculating the correlation of all elements in the sensing matrix with the current residual , obtaining a correlation set The expression of the correlation set includes: ;​​ extracting a number of maximum value elements from the correlation set , wherein a set of index values corresponding to the number of maximum value elements constitutes an index value set , ; merging the index value set with the initial index set to obtain an index set , ; wherein n is an integer of , when , is the initial index set; the number of extracted maximum value elements corresponds to the size of the support set P . add to the support set the corresponding perception matrix a vector in the support set , resulting in ; According to the support set and the observation vector y reconstructing a delay-Doppler domain channel , the expression of the delay-Doppler domain channel includes: （1） wherein, T denotes a threshold value; -1 denotes an inverse matrix; said delay-Doppler domain channel the term with the largest absolute value as the estimate of this iteration and compute an estimated residual error of said delay-Doppler domain channel of this iteration said estimated residual error whose expression comprises: （2） if the estimated residual error rn satisfies: then the iteration is stopped. if the estimated residual rn satisfies: then the support set is continued with an increase of the size of the scale P by 1. In other cases, the next iteration is directly performed.

4. The channel estimation method of claim 3, wherein The step of obtaining the extracted features Inputting the estimation result into the initial convolutional layer to obtain the extracted features The expression of the extracted features includes: （3） wherein, represents an extracted feature of an initial convolutional layer; represents an effect of a convolutional layer; represents an estimated result of iteration; the extracted feature as an input to the Inception-ResNet module.

5. The method for channel estimation according to claim 4, wherein, The step of inputting the extracted feature F to the plurality of Inception-ResNet modules, and obtaining spliced and compressed features after splicing and compression processing of the plurality of Inception-ResNet modules The step of inputting the extracted feature to the first Inception-ResNet module to obtain output first extracted features , and the expression of the first extracted features includes: （4） wherein, represents the extracted features of the convolutional layer, represents the compressed features of the first Inception-ResNet module; The first extracted feature is obtained by inputting the first feature into the first Inception-ResNet module The second extracted feature is obtained by inputting the second feature into the second Inception-ResNet module The expression of the second extracted feature includes （5） wherein, denotes the compressed features of the second Inception-ResNet module; The second feature is input into the third Inception-ResNet module to obtain a spliced compressed feature The expression of the spliced compressed feature includes: （6） wherein, denotes the compressed features of the third Inception-ResNet module.

6. The channel estimation method of claim 5, wherein, The concatenated compressed features are input to the fully connected layer, which integrates the concatenated compressed features to obtain a denoised channel response, wherein an expression of the denoised channel response includes: （7） wherein, denotes the weights of the fully connected layer; denotes the bias of the fully connected layer.

7. A channel estimation system based on orthogonal time and frequency space modulation technique characterized in that, The channel estimation system comprises a signal subsystem, a sparsity adaptive matching pursuit subsystem and a denoising network subsystem; the sparsity adaptive matching pursuit subsystem is connected with the signal subsystem and the denoising network subsystem; wherein The signal subsystem is used for receiving a time-delay Doppler domain received signal and inputting the time-delay Doppler domain received signal into the sparsity adaptive matching pursuit subsystem; The sparsity adaptive matching pursuit subsystem is used for providing a sparsity adaptive matching pursuit algorithm and processing the time-delay Doppler domain received signal by using the sparsity adaptive matching pursuit algorithm to reconstruct a delay Doppler domain channel; The denoising network subsystem is configured to denoise the input delay-Doppler domain channel to obtain a denoised channel response, comprising: The denoising network is established, and the denoising network comprises an Inception-ResNet denoising network model, the Inception-ResNet denoising network model comprising an initial convolutional layer, a plurality of Inception-ResNet modules with the same structure and a fully connected layer; The plurality of Inception-ResNet modules with the same structure comprise a first Inception-ResNet module, a second Inception-ResNet module and a third Inception-ResNet module connected in sequence; each Inception-ResNet module comprises a splicing layer; the initial convolutional layer is connected to the first Inception-ResNet module; The fully connected layer is connected to the third Inception-ResNet module; The estimation result is input into the Inception-ResNet denoising network model to obtain the denoised channel response , comprising: The estimation result is input into the initial convolutional layer to obtain an extraction feature The extraction feature is input into a plurality of Inception-ResNet modules, and a splicing and compression feature is obtained after splicing and compression processing of the plurality of Inception-ResNet modules The splicing and compression feature is input into a plurality of Inception-ResNet modules, and a splicing and compression feature is obtained after splicing and compression processing of the plurality of Inception-ResNet modules Each Inception-ResNet module includes two feature extraction paths​ The first feature extraction path comprises a first convolutional layer, a second convolutional layer and a third convolutional layer; The first convolutional layer is a 3×3 convolutional layer comprising 32 convolutional kernels; the second convolutional layer is a 1×7 convolutional layer comprising 48 convolutional kernels; and the third convolutional layer is a 7×1 convolutional layer comprising 64 convolutional kernels; The second feature extraction path comprises a fourth convolutional layer, which is a 3×3 convolutional layer comprising 64 convolutional kernels; Each convolutional layer in each feature extraction path is connected to a Gaussian error linear unit respectively; the first feature extraction path and the second feature extraction path are connected to the splicing layer, the splicing layer being configured to splice the first feature extraction path and the second feature extraction path to obtain spliced features; the splicing layer is connected to a fifth convolutional layer, which is a 1×1 convolutional layer comprising 64 convolutional kernels, the fifth convolutional layer being configured to compress the spliced features; The plurality of Inception-ResNet modules employ a skip connection manner of jumping each convolutional layer in the feature extraction path to obtain the spliced compressed features after the splicing compression processing as the output of the plurality of Inception-ResNet modules The concatenated compressed features to the fully connected layer, which integrates the concatenated compressed features to obtain a de-noised channel response .

Images