Sparse estimation method of underwater acoustic channel using convolutional neural network channel cluster detection

Abstract

The present invention belongs to the technical field of underwater acoustic signal processing, and specifically relates to a sparse estimation method for underwater acoustic channels using convolutional neural network channel cluster detection, which is used in OFDM underwater acoustic communication systems. The method comprises: obtaining frequency domain signals based on pilot information corresponding to different data blocks; performing coarse channel estimation on the frequency domain signals; inputting the coarse channel estimation results into a pre-established and trained cluster detection model to perform cluster detection and obtain channel cluster position information; and implementing channel estimation based on the channel cluster position information in combination with a sparse estimation algorithm. The present invention utilizes a convolutional neural network to perform cluster detection on underwater acoustic channels, and then combines the cluster detection results with sparse estimation methods to limit the channel search space, reduce the impact of noise on channel estimation, and improve estimation accuracy. Compared with traditional methods, this method has higher channel estimation accuracy, stability, and robustness.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater acoustic signal processing, and in particular relates to an underwater acoustic channel sparse estimation method using convolutional neural network channel cluster detection. Background Art

[0002] Acoustic waves are a crucial carrier for underwater wireless communications. However, the characteristics of underwater acoustic channels are complex and varied, with strong multipath, high noise, large Doppler shifts, and significant spatial and temporal fluctuations. These characteristics pose significant difficulties and challenges for underwater acoustic communication. Underwater acoustic channel estimation is a crucial step in underwater information transmission, and accurate estimation of channel parameters is crucial for improving communication performance.

[0003] Underwater acoustic channels typically exhibit clustered sparsity, meaning that most channel impulse responses are zero or near zero, and channel energy is concentrated in sparse, non-uniformly distributed clusters. Traditional channel estimation methods, such as least squares (LS) and minimum mean square error (MMSE), introduce noise estimation errors at zero taps. Furthermore, the high-order estimators required for channel estimation also present significant computational complexity.

[0004] Given the sparse nature of underwater acoustic channels, sparse estimation methods have been used to improve channel estimation performance. In recent years, compressed sensing-based channel estimation methods have received extensive attention and research due to their superior estimation performance. Among them, the orthogonal matching pursuit (OMP) algorithm has been shown to have significant performance advantages over traditional algorithms. Based on this, the synchronous orthogonal matching pursuit (SOMP) algorithm, leveraging the stable sparse nature of slowly varying underwater acoustic channels, has been applied to solve the joint sparse reconstruction problem of the channel over a period of time. This method further improves algorithm performance by utilizing time gain, but it does not fully exploit the sparse cluster structure of the underwater acoustic channel. Noise estimation errors outside the cluster region still exist, affecting channel estimation accuracy.

[0005] How to detect and utilize the cluster structure of the underwater acoustic channel based on the sparse characteristics of the underwater acoustic channel is the key to improving the performance of sparse estimation methods. Some research methods have been proposed to detect channel cluster structure, such as using rough channel estimation results to determine whether each tap coefficient is greater than the channel mean for cluster grouping. However, when there is large noise interference, this method is prone to estimation errors; the traditional Page detection method can also be applied to underwater acoustic channel cluster detection, but the algorithm needs to adjust parameters according to the environment, and its stability and robustness are poor. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the prior art and propose a sparse estimation method for underwater acoustic channels using convolutional neural network channel cluster detection.

[0007] To achieve the above objectives, the present invention proposes a sparse estimation method for underwater acoustic channels using convolutional neural network channel cluster detection for an orthogonal frequency division multiplexing (OFDM) underwater acoustic communication system. The method comprises:

[0008] Step 1) obtaining frequency domain signals corresponding to different data blocks based on pilot information;

[0009] Step 2) performing a rough channel estimation on the above frequency domain signal;

[0010] Step 3) Inputting the channel coarse estimation result into a pre-established and trained cluster detection model to perform cluster detection and obtain channel cluster location information;

[0011] Step 4) Channel estimation is achieved based on the channel cluster location information combined with a sparse estimation algorithm.

[0012] As an improvement to the above method, the frequency domain signals corresponding to different data blocks based on pilot information in step 1) satisfy the following formula:

[0013] Y p,l =diag(X p,l )F p h l +W p,l

[0014] Among them, Y p,l is the frequency domain received signal of the lth data block based on the pilot information, l=1,2,…,L, L is the total number of data blocks, diag(X p,l ) is the corresponding pilot symbol X p,l The diagonal matrix composed of p is the corresponding Fourier transform matrix, W p,l is the corresponding frequency domain additive noise, h l ∈N×1 is the channel corresponding to the lth data block, and N is the channel length.

[0015] As an improvement to the above method, step 2) is specifically as follows:

[0016] The frequency domain signal Y based on pilot information of different data blocks is calculated using the following formula: p,l Perform rough channel estimation to obtain the rough channel estimation result of the lth data block

[0017]

[0018] in, (·) H represents the conjugate transpose, K p Indicates the number of pilot frequencies.

[0019] As an improvement to the above method, the input of the cluster detection model is the channel rough estimation result, and the output is the channel cluster location information; a convolutional neural network with a "U-net" architecture is adopted, including a contraction path and an expansion path, and the two paths form a symmetrical structure, wherein,

[0020] The contraction path is used to extract required features through feature dimensionality reduction;

[0021] The extended path is used to decode the extracted features, compare each value in the decoded vector with the set threshold, and obtain the corresponding cluster position information Ψ n , n∈[1,N], and then obtain the channel cluster position information [Ψ1,…Ψ n ,…Ψ N ].

[0022] As an improvement to the above method, the corresponding cluster position information is n Satisfy the following formula:

[0023]

[0024] in, It represents the probability that the corresponding position has an intra-cluster channel impulse response. If the probability is greater than the set threshold of 0.5, it is considered that the position has an intra-cluster channel impulse response. Otherwise, the position does not have an intra-cluster channel impulse response.

[0025] As an improvement to the above method, step 4) specifically includes:

[0026] Construct the cluster region constraint matrix Ψ based on the obtained channel cluster position information:

[0027] Ψ=diag([Ψ1,…Ψ n ,…Ψ N ]),Ψ n ∈{0,1}

[0028] Among them, diag() represents a diagonal matrix;

[0029] The compressed sensing sparse reconstruction algorithm is used to solve the received signal Y in the frequency domain of the lth data block based on the pilot information under the cluster constraint condition. p,l , simplified into the following form:

[0030] Y p,l =Φ l Ψh l +W p,l

[0031] Among them, Φ l is the perception matrix, Φ l =diag(X p,l )Fp , l=1,2,…,L;

[0032] Based on the joint sparse model, the joint dictionary matrix Λ=diag(Φ l The dictionary atoms whose positions in the corresponding cluster region constraint matrix are 1 remain unchanged, and the remaining dictionary atoms are 0 vectors. The corresponding optimization problem is expressed as:

[0033]

[0034] in, Represents the estimation result matrix of L channels, is the receiving matrix composed of pilot information of L data blocks, (·) T represents transpose; δ is the minimum residual allowed;

[0035] A sparse channel estimation algorithm is used to solve the channel joint sparse reconstruction problem.

[0036] As an improvement to the above method, the method further includes a cluster detection model training step; specifically, the steps include:

[0037] Select historical data of surrounding sea areas to perform channel measurement and extract cluster structure to generate a training set;

[0038] The training set is input into the convolutional neural network, and the output vector is N is the channel length. Each value in the vector represents the probability that a channel impulse response exists within the cluster at the corresponding position. Binary cross entropy is used as the loss function. Training is performed through supervised learning until the training requirements are met, resulting in a trained cluster detection model.

[0039] As an improvement to the above method, the method selects historical data of the surrounding sea area to perform channel measurement and extract cluster structure to generate a training set; specifically includes:

[0040] The channel coarse estimation result and the known cluster position information are used as the input data and training labels of the training set respectively. The training labels are represented by vectors, where the positions with impulse responses are 1 and the rest are 0.

[0041] On the other hand, the present invention also proposes a sparse estimation system for underwater acoustic channels using convolutional neural network channel cluster detection, the system comprising:

[0042] A receiving module, configured to obtain frequency domain signals corresponding to different data blocks based on pilot information;

[0043] A channel coarse estimation module, used to perform coarse channel estimation on the above frequency domain signal;

[0044] A cluster detection module is used to input the channel coarse estimation result into a pre-established and trained cluster detection model to perform cluster detection and obtain channel cluster location information; and

[0045] The channel estimation module is used to realize channel estimation based on the channel cluster location information combined with the sparse estimation algorithm.

[0046] Compared with the prior art, the advantages of the present invention are:

[0047] The present invention utilizes the characteristic that the slowly varying underwater acoustic channel presents a relatively stable sparse cluster structure within a certain period of time, and uses a convolutional neural network to detect the underwater acoustic channel cluster structure, thereby achieving accurate positioning of the channel cluster position. Furthermore, the cluster position information and the channel sparse estimation method are combined to construct a channel sparse estimation algorithm based on cluster constraints. By limiting the channel search space, the impact of noise on channel estimation is reduced. At the same time, the joint estimation is used to obtain time gain, further improving the channel estimation accuracy and having good stability and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 2. It is a flow chart of the underwater acoustic channel sparse estimation method using convolutional neural network channel cluster detection in the present invention;

[0049] Figure 2 Schematic diagram of the convolutional neural network cluster detection model structure of Example 1;

[0050] Figure 3 1 is a diagram of the channel estimation result of Example 1;

[0051] Figure 4 This is a comparison chart of the bit error rate-signal-to-noise ratio simulation performance of Example 1. DETAILED DESCRIPTION

[0052] This paper proposes a sparse estimation method for underwater acoustic channels using convolutional neural network channel cluster detection. By taking advantage of the sparse cluster structure of underwater acoustic channels and the stable sparse characteristics of slowly varying channels, a coarse estimation of the time-varying channel is first performed. A convolutional neural network is then used to perform cluster detection on the rough channel estimation results. The cluster location information is then combined with the channel sparse estimation method to limit the search space for channel estimation and improve estimation accuracy. This method is used in OFDM underwater acoustic communication systems and includes:

[0053] Step 1) obtaining frequency domain signals corresponding to different data blocks based on pilot information;

[0054] Step 2) performing a rough channel estimation on the above frequency domain signal;

[0055] Step 3) Inputting the channel coarse estimation result into a pre-established and trained cluster detection model to perform cluster detection and obtain channel cluster location information;

[0056] Step 4) Channel estimation is achieved based on the channel cluster location information combined with a sparse estimation algorithm.

[0057] Specifically include:

[0058] (1) Select historical data of the surrounding sea area for channel measurement and extract the cluster structure to generate a training dataset.

[0059] More specifically, the channel coarse estimation result and the known cluster location information are used as the input data and training labels of the training dataset, respectively. The training labels are specifically represented as vectors, where the positions with impulse responses are 1 and the rest are 0.

[0060] (2) Use the training dataset obtained in step 1 to train the underwater acoustic channel cluster detection model. The cluster detection model uses a convolutional neural network with a "U-net" architecture. The network can be viewed as two parts: a contraction path and an expansion path. The contraction path extracts the required features through feature dimensionality reduction, while the symmetrical expansion path decodes the extracted features to obtain channel cluster location information.

[0061] More specifically, the convolutional neural network has 14 layers, 12 of which are hidden layers. In the expansion path, the convolution kernel size is 3×1. Each convolution layer is preceded by 1×1 border padding, followed by ReLU activation and 2×1 maximum pooling. Except for the first layer, the number of channels is doubled after each convolution. The expansion path and contraction path are symmetrical. In the expansion path, the feature map is doubled upsampled before each convolution layer, followed by 1×1 border padding, followed by convolution that halves the number of feature channels, and then concatenated with the corresponding feature map in the contraction path. Then, 1×1 border padding and convolution with a convolution kernel size of 3×1 that halves the number of channels are performed, followed by ReLU activation. The final layer is a convolution that maps the number of channels to 1, with a 1×1 convolution kernel size and a Sigmoid activation function.

[0062] The output vector of the network is N is the channel length. Each value in this vector represents the probability that an intra-cluster channel impulse response exists at the corresponding location. Binary cross entropy is used as the network loss function, and training is performed through supervised learning. If the probability is greater than a threshold (threshold = 0.5), it is considered that an intra-cluster channel impulse response exists at this location. The corresponding cluster location information can be obtained as follows:

[0063]

[0064] (3) Use the trained model to perform cluster detection on the channel, and then use the sparse estimation algorithm to estimate the channel. Specifically:

[0065] 3.1 More specifically, in an orthogonal frequency division multiplexing (OFDM) underwater acoustic communication system, for an OFDM signal with L data blocks, the frequency domain received signal based on the pilot information can be obtained at the receiving end.

[0066] Y p,l =diag(X p,l )F p h l +W p,l (2)

[0067] Where l = 1, 2, ..., L, Y p,l is the received signal in the frequency domain of the first data block based on the pilot information, diag(X p,l ) is the corresponding diagonal matrix composed of pilot symbols, F p is the corresponding Fourier transform matrix, W p,l is the corresponding frequency domain additive noise, h l ∈N×1 is the channel corresponding to the lth data block, and N is the channel length.

[0068] The following formula is used to perform rough channel estimation for different data blocks:

[0069]

[0070] in, (·) H represents the conjugate transpose, K p Indicates the number of pilot frequencies.

[0071] 3.2 Each column of the channel estimation result obtained in step 3.1 is used as a different channel input of the cluster detection model. The output of the model is then used to obtain the channel cluster location information through formula (1). The obtained cluster location information is used to construct the cluster area constraint matrix. Specifically, define Ψ as the cluster area constraint matrix

[0072] Ψ=diag([Ψ1,...,Ψ n ,...,Ψ N ]),Ψ n ∈{0,1} (4)

[0073] 3.3 Use the cluster constraint matrix obtained in step 3.2 and the sparse estimation method to estimate the underwater acoustic channel.

[0074] Specifically, by utilizing the common sparse characteristics of different communication data blocks under slowly varying channels, a cluster-constrained joint channel sparse model is constructed based on the cluster constraint matrix, and the joint dictionary matrix is ​​derived. More specifically, due to the sparse characteristics of the channel, most values ​​in h can be regarded as 0, so the problem in equation (3) can be solved using the compressed sensing sparse reconstruction algorithm. Under the cluster constraint condition, the problem is further simplified to the following form

[0075] Y p,l =Φ l Ψh l +W p,l (5)

[0076] Among them, Φ l =diag(X p,l )F p is the perception matrix. Under the joint sparse model (JSM2), the joint sparse reconstruction problem of the OFDM channel of L data blocks can be expressed as:

[0077]

[0078] in, is the receiving matrix composed of pilot information of L data blocks; Λ is the cluster-constrained joint dictionary matrix; Φ l ,l∈[1,L] is the perception matrix of a single data block, and the pilots on different data blocks can be different. Therefore, under the action of the cluster constraint matrix, the dictionary atoms with the corresponding cluster region constraint matrix position of 1 in the joint dictionary matrix Λ remain unchanged, and the remaining dictionary atoms are 0 vectors. The corresponding optimization problem is expressed as:

[0079]

[0080] Among them, δ is the minimum residual allowed.

[0081] Finally, a sparse channel estimation algorithm is used to solve the channel joint sparse reconstruction problem.

[0082] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0083] Example 1

[0084] like Figure 1 As shown, embodiment 1 of the present invention proposes a sparse estimation method for underwater acoustic channels using convolutional neural network channel cluster detection. This embodiment uses the OFDM underwater acoustic communication system as the application background, and adopts the SOMP algorithm as a sparse channel estimation method combined with cluster information. The effectiveness of the present invention is verified through simulation.

[0085] Specifically include:

[0086] Step 1: Select historical data of the surrounding sea area to perform channel measurement and extract cluster structure to generate a training dataset.

[0087] 1.1 Preprocess the real sea trial data and extract the cluster structure of the channel estimation results as the channel used in the simulation experiment.

[0088] 1.2 Using the channel obtained in 1.1, simulate the underwater acoustic communication process and perform a rough estimation of the time-varying channel based on the received signal and the prior sequence.

[0089] Specifically, the simulation parameters are as follows: the transmitted signal is an OFDM signal, the bandwidth is 100Hz, the total number of subcarriers is 256, the subcarrier spacing is 0.39Hz, 64 pilots are inserted at equal intervals, the guard interval length is 0.44s, the modulation method is 4-order QAM, and a complete OFDM signal contains 10 data blocks. The training set simulation signal-to-noise ratio is [-5,15]dB, and the test set simulation signal-to-noise ratio is [-515]dB. The signal-to-noise ratio here is defined as

[0090]

[0091] At the receiving end of the OFDM underwater acoustic communication system, a rough estimation of the channel is obtained using the received signal and pilot information.

[0092] 1.3 The channel rough estimation results and known cluster location information are used as the input data and training labels of the training data set respectively. There are 400 sets of training data in total.

[0093] Step 2: Use the training data set obtained in step 1 to train the convolutional neural network model. The model parameters and structure are as follows: Figure 2 shown.

[0094] Step 3: Use the trained model to perform cluster detection on the channel, and then use the sparse estimation algorithm to estimate the channel.

[0095] Specifically, if Figure 3 As shown, assume that the channel parameter h is [-0.250+0.153i,0,0,0.715+0.425i,0,-0.265+0.054i,0,0.92-0.389i,0,0,-0.223-0.230i,zeros(1,78),-0.141--0.028i,0,0,0.361+0.41i,0,0,0,0.201-0.169i,0,0,-0.103-0.04i,zeros(1,60)]; the simulation signal-to-noise ratio is

[15] dB. The channel rough estimation result is taken as input, and the network outputs the channel cluster position information vector is [1,1,1,1,1,1,1,1,1,1,1,1,zeros(1,78),1,1,1,1,1,1,1,1,1,1,zeros(1,60)]; the cluster-constrained SOMP algorithm is used for channel estimation and the output is It is [-0.236+0.200i,0.057-0.046i,0,0.740+0.378i,0,-0.253-0.007i,0,0.88-0.429j,0,0,-0.232-0.244i,zeros(1,79),-0.047-0.162i,0,0.272+0.357i,0-0.105+0.018i,0,0.132-0.127i,0,-0.019+0.098i,-0.030-0.015i,zeros(1,60)]. For comparison, under the same conditions, the BER after equalization of the channel estimation results of the LS algorithm, OMP algorithm, SOMP algorithm and the proposed SOMP algorithm using deep learning convolutional neural network cluster detection (DL-SOMP) and the BER when the complete channel state information (CSI) is known are 0.066, 0.043, 0.033, 0.029 and 0.027 respectively. Furthermore, simulations are performed in the signal-to-noise ratio range of [-5, 15] dB, and the results are Figure 4 The bit error rate curve.

[0096] Example 2

[0097] Embodiment 2 of the present invention proposes a system for sparse estimation of underwater acoustic channels using convolutional neural network channel cluster detection, which is implemented based on the method of embodiment 1. The system includes:

[0098] A receiving module, configured to obtain frequency domain signals corresponding to different data blocks based on pilot information;

[0099] A channel coarse estimation module, used to perform coarse channel estimation on the above frequency domain signal;

[0100] The cluster detection module is used to input the channel rough estimation results into a pre-established and trained cluster detection model to perform cluster detection and obtain channel cluster location information;

[0101] The channel estimation module is used to realize channel estimation based on the channel cluster location information combined with the sparse estimation algorithm.

[0102] The present invention proposes a sparse estimation method for underwater acoustic channels using convolutional neural network channel cluster detection. First, a convolutional neural network is used to perform cluster detection on the underwater acoustic channel. Then, the cluster detection results are combined with sparse estimation methods to limit the channel search space, reduce the impact of noise on channel estimation, and improve estimation accuracy. Compared with traditional methods, this method has higher channel estimation accuracy, stability and robustness.

[0103] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A method for sparse estimation of underwater acoustic channels using convolutional neural network channel cluster detection, for use in OFDM underwater acoustic communication systems, the method comprising: Step 1) obtaining frequency domain signals corresponding to different data blocks based on pilot information; Step 2) performing a rough channel estimation on the above frequency domain signal; Step 3) Inputting the channel coarse estimation result into a pre-established and trained cluster detection model to perform cluster detection and obtain channel cluster location information; Step 4) realizing channel estimation based on the channel cluster location information combined with a sparse estimation algorithm; The frequency domain signals corresponding to different data blocks in step 1) based on the pilot information satisfy the following formula: Y p,l =diag(X p,l )F p h l +W p,l Among them, Y p,l is the frequency domain received signal of the lth data block based on the pilot information, l=1,2,…,L, L is the total number of data blocks, diag(X p,l ) is the corresponding pilot symbol X p,l The diagonal matrix composed of p is the corresponding Fourier transform matrix, W p,l is the corresponding frequency domain additive noise, h l ∈N×1 is the channel corresponding to the lth data block, and N is the channel length; The step 2) is specifically as follows: The frequency domain signal Y based on pilot information of different data blocks is calculated using the following formula: p,l Perform rough channel estimation to obtain the rough channel estimation result of the lth data block in, (·) H represents the conjugate transpose, K p Indicates the number of pilots; The input of the cluster detection model is the rough channel estimation result, and the output is the channel cluster location information; the convolutional neural network using the "U-net" architecture includes a contraction path and an expansion path, and the two paths form a symmetrical structure, wherein, The contraction path is used to extract required features through feature dimensionality reduction; The extended path is used to decode the extracted features, compare each value in the decoded vector with the set threshold, and obtain the corresponding cluster position information Ψ n , n∈[1,N], and then obtain the channel cluster position information [Ψ1,…Ψ n ,…Ψ N ].

2. The underwater acoustic channel sparse estimation method using convolutional neural network channel cluster detection according to claim 1 is characterized in that: The corresponding cluster location information is Ψ n Satisfy the following formula: in, It represents the probability that the corresponding position has an intra-cluster channel impulse response. If the probability is greater than the set threshold of 0.5, it is considered that the position has an intra-cluster channel impulse response. Otherwise, the position does not have an intra-cluster channel impulse response.

3. The underwater acoustic channel sparse estimation method using convolutional neural network channel cluster detection according to claim 2 is characterized in that: The step 4) specifically includes: Construct the cluster region constraint matrix Ψ based on the obtained channel cluster position information: Ψ=diag([Ψ1,…Ψ n ,…Π N ]),Ψ n ∈{0,1} Among them, diag() represents a diagonal matrix; The compressed sensing sparse reconstruction algorithm is used to solve the received signal Y in the frequency domain of the lth data block based on the pilot information under the cluster constraint condition. p,l , simplified into the following form: Y p,l =Φ l Ψh l +W p,l Among them, Φ l is the perception matrix, Φ l =diag(X p,l )F p , l=1,2,…,L; Based on the joint sparse model, the joint dictionary matrix Λ=diag(Φ l The dictionary atoms whose positions in the corresponding cluster region constraint matrix are 1 remain unchanged, and the remaining dictionary atoms are 0 vectors. The corresponding optimization problem is expressed as: in, Represents the estimation result matrix of L channels, is the receiving matrix composed of pilot information of L data blocks, (·) T represents transpose; δ is the minimum residual allowed; A sparse channel estimation algorithm is used to solve the channel joint sparse reconstruction problem.

4. The underwater acoustic channel sparse estimation method using convolutional neural network channel cluster detection according to claim 2 is characterized in that: The method also includes a cluster detection model training step; specifically, including: Select historical data of surrounding sea areas to perform channel measurement and extract cluster structure to generate a training set; The training set is input into the convolutional neural network, and the output vector is N is the channel length. Each value in the vector represents the probability that a channel impulse response exists within the cluster at the corresponding position. Binary cross entropy is used as the loss function. Training is performed through supervised learning until the training requirements are met, resulting in a trained cluster detection model.

5. The underwater acoustic channel sparse estimation method using convolutional neural network channel cluster detection according to claim 4 is characterized in that: The method of selecting historical data of the surrounding sea area to perform channel measurement and extract cluster structure to generate a training set specifically includes: The channel coarse estimation result and the known cluster position information are used as the input data and training labels of the training set respectively. The training labels are represented by vectors, where the positions with impulse responses are 1 and the rest are 0.

6. A system based on the underwater acoustic channel sparse estimation method using convolutional neural network channel cluster detection according to claim 1, characterized in that: The system comprises: A receiving module, configured to obtain frequency domain signals corresponding to different data blocks based on pilot information; A channel coarse estimation module, used to perform coarse channel estimation on the above frequency domain signal; A cluster detection module is used to input the channel coarse estimation result into a pre-established and trained cluster detection model to perform cluster detection and obtain channel cluster location information; and The channel estimation module is used to realize channel estimation based on the channel cluster location information combined with the sparse estimation algorithm.

Images