A neural network underwater acoustic channel estimation method based on cluster self-attention mechanism

Abstract

The present invention discloses a neural network underwater acoustic channel estimation method based on cluster self-attention mechanism, the method comprising: receiving a time domain OFDM signal of an orthogonal frequency division multiplexing underwater acoustic communication system and converting it into a frequency domain OFDM signal; performing a channel LS rough estimation on the frequency domain OFDM signal based on pilot information; performing cluster position identification through a reference threshold based on the LS rough estimation result; inputting the LS rough estimation result and the cluster position identification result into a pre-established and trained underwater acoustic channel estimation model to obtain an underwater acoustic channel estimation result; the underwater acoustic channel estimation model adopts a complex neural network based on a cluster self-attention mechanism. The present invention enables the neural network to focus on the features within the cluster; divides the underwater acoustic channel into two components, trend and random, for feature extraction, and obtains more comprehensive characteristic information through feature fusion. The invention realizes a good neural network architecture design that combines the characteristics of the underwater acoustic channel.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater acoustic signal processing, and in particular relates to a neural network underwater acoustic channel estimation method based on a cluster self-attention mechanism. Background Art

[0002] Underwater acoustic channel estimation is an important part of underwater acoustic communication, which directly affects the performance and reliability of underwater acoustic communication. Underwater channels are very different from electromagnetic wave channels, such as severe multipath expansion, strong frequency selectivity, and significant Doppler effect. Due to the influence of reflection and refraction on the sea surface and seabed, as well as the fluctuation of inhomogeneous media in seawater, the underwater acoustic channel formed by the seawater medium has complex, time-varying multipath propagation characteristics. Especially in complex waveguide environments, sound lines propagate in the form of clusters and reach the receiving point. Therefore, the underwater acoustic channel exhibits an obvious cluster sparse structure, that is, the channel impulse response does not appear in a single form on the time delay coordinate, but in the form of clusters.

[0003] For long delay spread channels, channel estimation methods based on traditional algorithms, such as the least squares (LS) and minimum mean square error (MMSE) algorithms, will introduce noise estimation errors at the zero tap. At the same time, the high-order estimator required for channel estimation also requires a large computational complexity.

[0004] In recent years, researchers have introduced machine learning methods into underwater acoustic channel estimation tasks. With its excellent learning ability and strong adaptive ability, it has shown great advantages in complex environmental conditions and problems that are difficult to linearly model. Compared with traditional algorithms, it has made breakthroughs in estimation accuracy and tracking robustness. The underwater acoustic channel estimation model based on machine learning takes the sent and received training sequences (or pilot symbols) as input information and outputs the channel estimation results. Researchers have successively applied machine learning models such as Multilayer Perceptron (MLP), Fully Connected Deep Neural Network (FC-DNN), and Convolutional Neural Network (CNN) to underwater acoustic channel estimation, and have achieved significant improvements in estimation accuracy and bit error rate performance compared to traditional algorithms.

[0005] However, the existing underwater acoustic channel estimation technology based on machine learning does not fully consider and utilize the clustered sparse physical structure of the underwater acoustic channel, so there are certain limitations in performance. It is urgent to develop a machine learning method that cooperates with the inherent physical characteristics of the underwater acoustic channel to give play to the prior advantages. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the prior art and propose a neural network underwater acoustic channel estimation method based on cluster self-attention mechanism.

[0007] The present invention proposes a neural network underwater acoustic channel estimation method based on a cluster self-attention mechanism, the method comprising:

[0008] Step 1) receiving a time domain OFDM signal of an orthogonal frequency division multiplexing water acoustic communication system and converting it into a frequency domain OFDM signal;

[0009] Step 2) Based on the pilot information, a rough LS estimation of the channel is performed on the frequency domain OFDM signal;

[0010] Step 3) Based on the LS rough estimation result, cluster position identification is performed by using a reference threshold;

[0011] Step 4) inputting the LS rough estimation result and the cluster position recognition result into a pre-established and trained underwater acoustic channel estimation model to obtain an underwater acoustic channel estimation result;

[0012] The underwater acoustic channel estimation model adopts a complex neural network based on a cluster self-attention mechanism.

[0013] Preferably, the time domain OFDM signal of step 1) is:

[0014]

[0015] in, represents circular convolution, x(n), y(n) and w(n) represent the transmitted signal, received signal and additive white Gaussian noise respectively, h(n) represents the impulse response of the underwater acoustic channel; n represents the time domain;

[0016] The frequency domain OFDM signal is:

[0017] Y(k)=X(k)H(k)+W(k)

[0018] Among them, X(k), Y(k), H(k) and W(k) respectively represent the frequency domain representation of the transmitted signal, the received signal, the underwater acoustic channel impulse response and the additive white Gaussian noise, and k represents the frequency domain.

[0019] Preferably, the channel LS rough estimation result h in step 2) is LS Satisfy the following formula:

[0020]

[0021] Where IDFT stands for Inverse Discrete Fourier Transform, X P (k) represents the pilot, Y P (k) represents the frequency domain received signal corresponding to the pilot.

[0022] Preferably, the step 3) specifically includes:

[0023] According to the following formula, the location and length of the cluster are extracted by thresholding:

[0024]

[0025] Where Ω represents the path delay set of the cluster, N is the channel length, i represents the i-th path delay, || ||2 represents the second norm, ω th is the reference threshold, satisfying the following formula:

[0026]

[0027] Among them, K e is a parameter used to adjust the threshold;

[0028] The cluster position information p of the channel is obtained, where the value 1 represents the cluster distribution area, thereby obtaining the cluster position and range of the underwater acoustic channel.

[0029] Preferably, the input of the underwater acoustic channel estimation model is the channel LS rough estimation result h LS The cluster position information p of the channel is output as the estimation result h of the underwater acoustic channel. CSA , satisfying the following formula:

[0030] h CSA =F({h LS ,p},θ)

[0031] Among them, F() represents the trained underwater acoustic channel estimation model, and θ represents the model parameters.

[0032] Preferably, the underwater acoustic channel estimation model includes a one-dimensional complex convolutional neural network, a cluster self-attention module and a convolution module, wherein the convolution module includes a rectified linear unit and a one-dimensional complex convolutional neural network; the underwater acoustic channel estimation model satisfies the following formula:

[0033] h CSA =

[0034] g L-1 [g L-2 ,r(…concat(g 1,r (g0(h LS ;θ0)-attention(g0(h LS ;θ0));θ 1,r ),res2);θ L-2,r )+g L-2,t (…g 1,t (attention(g0(h LS;θ0));θ 1,t );θ L-2,t );θ L-1 ]

[0035] in

[0036] res l =g l-1,t (…attention(g0(h LS ;θ0));θ l-1,t )-attention(g l-1,t (…attention(g0(h LS ;θ0));θ l-1,t ))

[0037] Represents the training parameters in the neural network, L represents the total number of layers of the complex convolutional neural network, ω l and b l Represent the weight vector and bias vector of the lth complex convolutional neural network layer, θ l represents the parameters of the lth complex convolutional neural network layer, g l represents the output of the lth complex convolutional neural network layer, l∈[0,L-1], the subscripts r and t represent the layers corresponding to the random component and the trend component respectively, attention represents the cluster self-attention module, and res l It represents the residual of the trend component of the l-1th layer before and after the cluster self-attention module, and concat represents the feature map stacking operation, that is, stacking the feature maps of the random component and the residual component.

[0038] Preferably, the method further comprises a training phase of the underwater acoustic channel estimation model, using the Adam algorithm as an optimizer for model training, and the optimization goal of the training phase is:

[0039]

[0040] Among them, lossfunc represents the loss function, N is the channel length, h i , represent the channel impulse response and channel impulse response estimation corresponding to the i-th path delay respectively.

[0041] On the other hand, the present invention proposes a neural network underwater acoustic channel estimation system based on a cluster self-attention mechanism, the system comprising:

[0042] A receiving and converting module, used for receiving a time-domain OFDM signal of an orthogonal frequency division multiplexing water acoustic communication system and converting it into a frequency-domain OFDM signal;

[0043] A channel LS rough estimation module is used to perform a channel LS rough estimation on a frequency domain OFDM signal based on pilot information;

[0044] A channel cluster identification module, used for identifying cluster positions by reference thresholds based on LS rough estimation results; and

[0045] An estimation output module is used to input the LS rough estimation result and the cluster position recognition result into a pre-established and trained underwater acoustic channel estimation model to obtain an underwater acoustic channel estimation result;

[0046] The underwater acoustic channel estimation model adopts a complex neural network based on a cluster self-attention mechanism.

[0047] Compared with the prior art, the advantages of the present invention are:

[0048] The present invention uses the prior information of the underwater acoustic channel and proposes a cluster self-attention module, which enables the neural network to focus on the features within the cluster. Subsequently, combined with the multi-time scale variation characteristics of the underwater acoustic channel, the underwater acoustic channel is divided into two components, trend and random, for feature extraction, and more comprehensive characteristic information is obtained through feature fusion. This invention realizes a good neural network architecture design that combines the characteristics of the underwater acoustic channel. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] FIG1( a ) is a schematic diagram of a neural network underwater acoustic channel estimation method based on a cluster self-attention mechanism of the present invention;

[0050] FIG1( b ) is a schematic diagram of input and output signals of the method of the present invention;

[0051] Figure 2 It is a neural network structure diagram based on cluster self-attention mechanism of the present invention;

[0052] Figure 3 is a comparison diagram of the bit error rate-signal-to-noise ratio simulation performance of Example 1;

[0053] Figure 4 This is an example diagram of the channel estimation result of Example 1. DETAILED DESCRIPTION

[0054] The present invention proposes a neural network underwater acoustic channel estimation method based on cluster self-attention mechanism (CSA), which is used in an orthogonal frequency division multiplexing (OFDM) underwater acoustic communication system. The method comprises:

[0055] Step 1) In the training phase, a frequency domain OFDM signal based on pilot information is received;

[0056] Step 2) Based on the pilot information, a channel LS rough estimation is performed on the above frequency domain signal;

[0057] Step 3) Based on the channel LS rough estimation result, cluster position identification is performed by thresholding;

[0058] Step 4) inputting the channel LS rough estimate and the cluster position recognition result into the cluster self-attention neural network for training until the model parameters are fixed after converging to the error threshold;

[0059] Step 5) The rough channel LS estimation result and cluster position identification result of the deployment phase are input into the underwater acoustic channel estimation model trained in step (4), so as to achieve accurate estimation of the underwater acoustic channel, as shown in Figure 1(a) and Figure 1(b).

[0060] (1) Combining the transmitting data of the underwater acoustic OFDM communication system with the corresponding receiving data, a machine learning model data set is constructed.

[0061] More specifically, at the transmitting end, the frequency domain OFDM transmission signal X(k) is composed of pilot X P (k) and data bit X d (k). The parallel data strings are then transformed by inverse discrete Fourier transform (IDFT) to obtain the time domain OFDM signal x(n). After passing through the channel, the underwater acoustic OFDM signal at the receiving end can be expressed as:

[0062]

[0063] in represents circular convolution, x(n), y(n) and w(n) represent the transmitted signal, received signal and additive white Gaussian noise respectively, h(n) represents the impulse response of the underwater acoustic channel, and Y(k), X(k), H(k) and W(k) correspond to the frequency domain representation respectively.

[0064] (ii) Based on the pilot information, LS channel estimation is performed on the frequency domain underwater acoustic communication signal in step (i). The method is implemented as follows:

[0065]

[0066] Where IDFT stands for Inverse Discrete Fourier Transform, X P (k) represents the pilot, Y P (k) represents the frequency domain received signal corresponding to the pilot.

[0067] (iii) Based on the LS rough estimation result of step (ii), cluster location information is identified.

[0068] Specifically, as shown in Figure 1(b), although the rough estimation result of the underwater acoustic channel is not accurate enough, the location and length of the cluster can be extracted by the threshold demarcation method, that is:

[0069]

[0070] Where Ω represents the path delay set of the cluster, N is the channel length, i represents the i-th path delay of the channel, || ||2 represents the second norm, ω th is the reference threshold, and its value can be determined by the following formula:

[0071]

[0072] The parameter K e is a parameter used to adjust the threshold. Through the above method, we can obtain the channel prior information vector p, which is a vector composed of 1 and 0, where 1 represents the impulse response value area, that is, the cluster distribution area:

[0073]

[0074] From this we extract the cluster position and range of the underwater acoustic channel.

[0075] (iv) The LS channel rough estimation result h obtained in step (ii) is LS Together with the cluster position information p obtained in step (iii), it is input into the cluster self-attention neural network for training. The optimization goal of the training phase is:

[0076]

[0077] Among them, lossfunc represents the loss function, h i , represent the channel impulse response and channel impulse response estimation corresponding to the i-th path delay respectively.

[0078] Specifically, the machine learning model in the present invention adopts a complex neural network, which can directly process the complex form of the hydroacoustic signal. The complex neural network is expressed as:

[0079] h CSA =F({h LS ,p},θ)=

[0080]

[0081] in

[0082] Represents the training parameters in the neural network, ω l and b l Represent the weight vector and bias vector of the lth layer respectively, g lrepresents the output of the lth complex convolutional neural network layer. The subscripts r and t represent the layers corresponding to the random and trend components respectively. Attention means the output after the cluster self-attention module. l It represents the residual of the l-1th layer trend before and after the cluster self-attention module, and concat represents the feature map stacking operation, that is, stacking the random component and the residual component feature map.

[0083] More specifically, if Figure 2 As shown in the figure, the neural network input is a two-dimensional LS channel estimation result matrix. After passing through a one-dimensional complex convolutional neural network, the output is passed through a cluster self-attention module to extract the features of the slow-changing component (trend), and then the upper layer output is subtracted from the trend to obtain the fast-changing feature component (random). The trend feature component and the random feature component are then respectively subjected to feature extraction and "noise reduction" through a convolution module. The convolution module is composed of a rectified linear unit (Relu) and a one-dimensional complex convolutional neural network. After that, the trend feature component is passed through the output of the convolution module and then through the cluster self-attention module to continue to calculate the residual to obtain the residual random or noise component. The feature map of this residual and the random feature component is stacked to prevent information loss. The same operation is continued, and then the extracted random and trend feature components are added and passed through the convolution module to complete the scale transformation and feature fusion to obtain the estimated channel.

[0084] The neural network self-attention mechanism is a method of weighted combination of elements in an input sequence by learning to calculate the relative importance of each element in the entire sequence. The key to the self-attention mechanism is to use query (Query, Q), key (Key, K) and value (Value, V) to establish the association between elements. For an input sequence x = [x1, x2, ... x n ], where x i is the i-th element of the input sequence. The self-attention mechanism contains three matrices w q 、w k 、w v , which correspond to the linear transformations of the query Q, key K, and value V, respectively. The element x is calculated by the following calculation i Attention coefficient relative to the entire sequence:

[0085]

[0086] Among them, Q(x i ) and K(x j ) represent the input element x i and x j The query vector and key vector obtained by linear transformation, represents the dimension of the key vector, T represents the transpose operation of the matrix, and the attention coefficient is obtained after the softmax function, which represents the value of each element x in the input sequence. i With other elements x j The relative importance between the two, the sum of the attention coefficients is equal to 1. Finally, the attention coefficient is combined with the value V(x i ) to perform weighted summation and obtain element x i The output represents:

[0087]

[0088] The output represents the sum of each element x in the input sequence. i It is obtained by weighted summing the values ​​of the entire sequence.

[0089] The cluster self-attention mechanism is an improvement for the special properties of underwater acoustic channels. This method uses the channel cluster vector information p to replace the three matrices w that need to be learned. q 、w k 、w v , so as to better utilize the information in the cluster vector, focus on the features within the underwater acoustic channel cluster, and perform self-attention calculation on the elements within the cluster, making the model more efficient and accurate.

[0090] The model training process is to optimize the loss function in equation (6), continuously iterate and update the model parameters, and finally expect to converge to the global optimal solution. This paper uses the Adam algorithm as the optimizer for model training. The algorithm strategy can be expressed as:

[0091]

[0092] Where m t and v t are the first-order momentum term and the second-order momentum term, β1 and β2 are exponential decay rates, with default values ​​of 0.9 and 0.999, respectively. The default value of the regularization term ε is 1×10 -8 , and are their respective correction values.

[0093] (V) The channel LS rough estimation result h obtained in the deployment phase LS And the cluster position information p is input into the underwater acoustic channel estimation model trained in step (iv), and accurate estimation of the underwater acoustic channel can be achieved:

[0094] h CSA =F({h LS ,p},θ) (12)

[0095] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0096] Example 1

[0097] As shown in FIG1 , Embodiment 1 of the present invention proposes a neural network underwater acoustic channel estimation method based on a cluster self-attention mechanism, as shown in FIG1 . This embodiment uses an OFDM underwater acoustic communication system as the application background, and illustrates the effectiveness of the present invention through simulation verification. Specifically, it includes:

[0098] Step 1: Select historical data of the surrounding sea area to measure the channel and generate the channel library required for training and testing. Use the obtained channel to simulate the underwater acoustic OFDM communication process and perform rough estimation of the underwater acoustic channel based on the received signal and pilot sequence.

[0099] Specifically, the simulation parameters are: 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 4th order QAM, and a complete OFDM signal contains 10 data blocks. The training set simulation signal-to-noise ratio is [0,20]dB, and the test set simulation signal-to-noise ratio is [-1520]dB. The signal-to-noise ratio here is defined as

[0100]

[0101] Among them, y signal (n) represents the received signal in the time domain, y noise (n) represents the corresponding noise signal.

[0102] Step 2: At the receiving end of the OFDM underwater acoustic communication system, the LS rough estimation result of the channel is obtained by using the received signal and pilot information.

[0103] Step 3: Based on the channel LS rough estimation result, cluster position identification is performed through threshold delineation.

[0104] Step 4: Input the channel LS rough estimate and the cluster position recognition result into the cluster self-attention neural network for training until the model parameters are fixed after converging to the error threshold. The neural network structure based on cluster self-attention is as follows: Figure 2 shown.

[0105] Step 5, take the channel LS rough estimation result and cluster position identification result in the deployment phase and input them into the underwater acoustic channel estimation model trained in step (4), so as to achieve accurate estimation of the underwater acoustic channel.

[0106] Furthermore, the channel structure is obtained according to the LS algorithm, compressed sensing OMP algorithm, compressed sensing SOMP algorithm, conventional DnCNN and the proposed channel estimation algorithm based on cluster self-attention mechanism under the same conditions, and the bit error rate after equalization based on the zero forcing algorithm (Zero Force, ZF) and the bit error rate when the complete channel state information (CSI) is known are 0.0210, 0.0100, 0.0080, 0.0073, 0.0054, and 0.0032 (the signal-to-noise ratio is set to 20dB), respectively.

[0107] Furthermore, simulation was performed in the common range of underwater acoustic communication with a signal-to-noise ratio of [-15, 20] dB, and the results were Figure 3 The bit error rate curve. Figure 4 The results of channel estimation using the proposed method are shown in the case of a signal-to-noise ratio of 20dB.

[0108] Example 2

[0109] Embodiment 2 of the present invention proposes a neural network underwater acoustic channel estimation system based on a cluster self-attention mechanism. The system is implemented based on the method of embodiment 1 and specifically includes:

[0110] A receiving and converting module, used for receiving a time-domain OFDM signal of an orthogonal frequency division multiplexing water acoustic communication system and converting it into a frequency-domain OFDM signal;

[0111] A channel LS rough estimation module is used to perform a channel LS rough estimation on a frequency domain OFDM signal based on pilot information;

[0112] A channel cluster identification module is used to identify cluster positions based on LS rough estimation results by using a reference threshold;

[0113] An estimation output module is used to input the LS rough estimation result and the cluster position recognition result into a pre-established and trained underwater acoustic channel estimation model to obtain an underwater acoustic channel estimation result;

[0114] The underwater acoustic channel estimation model adopts a complex neural network based on a cluster self-attention mechanism.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A neural network underwater acoustic channel estimation method based on cluster self-attention mechanism, the method comprising: Step 1) receiving a time domain OFDM signal of an orthogonal frequency division multiplexing water acoustic communication system and converting it into a frequency domain OFDM signal; Step 2) Based on the pilot information, a rough LS estimation of the channel is performed on the frequency domain OFDM signal; Step 3) Based on the LS rough estimation result, cluster position identification is performed by using a reference threshold; Step 4) inputting the LS rough estimation result and the cluster position recognition result into a pre-established and trained underwater acoustic channel estimation model to obtain an underwater acoustic channel estimation result; The underwater acoustic channel estimation model adopts a complex neural network based on a cluster self-attention mechanism; The input of the underwater acoustic channel estimation model is the channel LS rough estimation result h LS The cluster position information vector p of the channel is output as the estimation result h of the underwater acoustic channel. CSA , satisfying the following formula: Where F() represents the trained underwater acoustic channel estimation model, and θ represents the model parameters; The underwater acoustic channel estimation model includes a one-dimensional complex convolutional neural network, a cluster self-attention module and a convolution module, wherein the convolution module includes a rectified linear unit and a one-dimensional complex convolutional neural network; the underwater acoustic channel estimation model satisfies the following formula: in res l =g l-1,t (…attention(g0(h LS ;θ0));θ l-1,t )-attention(g l-1,t (…attention(g0(h LS ;θ0));θ l-1,t )) represents the training parameters in the neural network, L represents the total number of layers of the complex convolutional neural network, ω l and b l Represent the weight vector and bias vector of the lth complex convolutional neural network layer, θ l represents the parameters of the lth complex convolutional neural network layer, g l represents the output of the lth complex convolutional neural network layer, l∈[0,L-1], the subscripts r and t represent the layers corresponding to the random component and the trend component respectively, attention represents the cluster self-attention module, and res l It represents the residual of the trend component of the l-1th layer before and after the cluster self-attention module, and concat represents the feature map stacking operation, that is, stacking the feature maps of the random component and the residual component.

2. According to claim 1, the neural network underwater acoustic channel estimation method based on cluster self-attention mechanism is characterized in that: The time domain OFDM signal of step 1) is: in, represents circular convolution, x(n), y(n) and w(n) represent the transmitted signal, received signal and additive white Gaussian noise respectively, h(n) represents the impulse response of the underwater acoustic channel; n represents the time domain; The frequency domain OFDM signal is: Y(k)=X(k)H(k)+W(k) Among them, X(k), Y(k), H(k) and W(k) respectively represent the frequency domain representation of the transmitted signal, the received signal, the underwater acoustic channel impulse response and the additive white Gaussian noise, and k represents the frequency domain.

3. The neural network underwater acoustic channel estimation method based on cluster self-attention mechanism according to claim 2 is characterized in that: The channel LS rough estimation result h of step 2) LS Satisfy the following formula: Where IDFT stands for Inverse Discrete Fourier Transform, X P (k) represents the pilot, Y P (k) represents the frequency domain received signal corresponding to the pilot.

4. The neural network underwater acoustic channel estimation method based on cluster self-attention mechanism according to claim 3 is characterized in that: The step 3) specifically includes: According to the following formula, the location and length of the cluster are extracted by thresholding: Where Ω represents the path delay set of the cluster, N is the channel length, i represents the i-th path delay, || ||2 represents the second norm, ω th is the reference threshold, satisfying the following formula: Among them, K e is a parameter used to adjust the threshold; The cluster position information vector p of the channel is obtained. This vector is composed of 1 and 0, where 1 represents the impulse response value area, that is, the cluster distribution area: The cluster position and range of the underwater acoustic channel are thus extracted.

5. The neural network underwater acoustic channel estimation method based on cluster self-attention mechanism according to claim 4 is characterized in that: The method also includes a training phase of the underwater acoustic channel estimation model, using the Adam algorithm as an optimizer for model training, and the optimization goal of the training phase is: Among them, lossfunc represents the loss function, N is the channel length, h i , represent the channel impulse response and channel impulse response estimation corresponding to the i-th path delay respectively.

6. A system based on the neural network underwater acoustic channel estimation method based on cluster self-attention mechanism of claim 1, characterized in that: The system comprises: A receiving and converting module, used for receiving a time-domain OFDM signal of an orthogonal frequency division multiplexing water acoustic communication system and converting it into a frequency-domain OFDM signal; A channel LS rough estimation module is used to perform a channel LS rough estimation on a frequency domain OFDM signal based on pilot information; A channel cluster identification module, used for identifying cluster positions by reference thresholds based on LS rough estimation results; and An estimation output module is used to input the LS rough estimation result and the cluster position recognition result into a pre-established and trained underwater acoustic channel estimation model to obtain an underwater acoustic channel estimation result; The underwater acoustic channel estimation model adopts a complex neural network based on a cluster self-attention mechanism.

Images