A W-Net-based FBMC underwater acoustic communication receiver anti-interference method

Abstract

This invention discloses a W-Net-based anti-interference method for an underwater acoustic communication receiver. The receiving network of the FBMC underwater acoustic communication model includes a W-Net network, which is composed of a multipath cancellation model (UNet) and a serially connected interference suppression model (MS-UNet). A multipath cancellation dataset and an interference suppression dataset are constructed. The UNet network is trained using the multipath cancellation dataset, and the MS-UNet network is trained using the interference suppression dataset. The MS-UNet network includes a UNet portion, three SE modules, and a multi-head self-attention module. This invention improves the anti-interference capability of the underwater acoustic communication system, reduces the bit error rate, and enhances the transmission quality of underwater acoustic signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater acoustic communication, and specifically provides an anti-interference method for an FBMC underwater acoustic communication receiving end based on a W-Net network. Background Art

[0002] Underwater acoustic communication is currently the only means of achieving long-distance, high-speed underwater data transmission. However, if the underwater acoustic channel cannot accurately estimate and balance multipath effects, communication reliability will be affected. Underwater acoustic communication is also susceptible to interference from the non-Gaussian ocean environment, so timely interference suppression is also crucial for the reliability of underwater acoustic systems. Traditional interference suppression algorithms have drawbacks such as requiring prior information and relying on fixed simulated parameters. Summary of the Invention

[0003] Purpose of the invention: In order to reduce the high system complexity problem of the traditional underwater acoustic channel receiver requiring interference suppression, channel estimation and equalization to obtain the transmitted signal, the present invention proposes an anti-interference method for the FBMC underwater acoustic communication receiver based on the attention mechanism combined with the WNet network, which improves the anti-interference ability, reduces the bit error rate, and improves the transmission quality of the underwater acoustic signal.

[0004] Technical solution: A W-Net network-based FBMC underwater acoustic communication receiving end anti-interference method includes the following steps:

[0005] Step 1: Establish an FBMC underwater acoustic communication model. The FBMC underwater acoustic communication model uses the FBMC / OQAM modulation method, inserts a pilot signal, modulates the symbol onto the subcarrier, and sends it into the underwater acoustic channel.

[0006] In the FBMC underwater acoustic communication model, the interference signal is superimposed to form the received signal r(n) expressed as:

[0007] r(n)=s(n)*h(n)+w(n)+j(n)

[0008] Where s(n) represents the transmitted signal, the transmission mode is FBMC, * represents the convolution operator, h(n) is the impulse response of the multipath channel, w(n) is Gaussian white noise, and j(n) is the interference;

[0009] Step 2: Construct a multipath cancellation dataset and an interference suppression dataset. Down-convert the received signal r(n) to a baseband signal, then filter it through a low-pass filter. This signal serves as the input of the multipath cancellation dataset, while the received signal s(n)+j(n), which is not affected by the multipath channel but only by the interference, serves as the output of the multipath cancellation dataset. The received signal s(n)+j(n), which is not affected by the multipath channel but only by the interference, serves as the input of the interference suppression dataset, while the transmitted signal s(n) serves as the output of the interference suppression dataset.

[0010] Step 3: Use the multipath cancellation dataset to train the multipath cancellation network UNet, save the trained network parameters, and obtain the trained underwater acoustic communication multipath cancellation network model. The multipath cancellation model UNet includes a ConvBlock encoder, an UpBlock decoder, a jump connection connecting the encoder and decoder, and a bottleneck; the ConvBlock encoder is composed of convolution, DropBlock, batch normalization, and an activation function; after batch normalization, downsampling is achieved through average pooling, and the UpBlock decoder increases the data size through upsampling to match the dimension of the output data;

[0011] Step 4: Connect the interference suppression model MS-UNet in series after the multipath elimination model UNet, use the interference suppression data set to train the interference suppression network MS-UNet, save the trained network parameters, and obtain the trained underwater acoustic communication interference suppression network model. The interference suppression network MS-UNet includes a UNet part, three SE (Squeeze-and-Excitation module, SE) modules and a multi-head self-attention module (Multi-Head Self-Attention module, MHSA); with the assistance of the attention mechanism, the interference is extracted from the received signal, and then the estimated interference is eliminated from the input signal to achieve the purpose of interference suppression. The processed data is then received by the subsequent module at the receiving end. The W-Net network consists of a multipath elimination model UNet and a serial interference suppression model MS-UNet;

[0012] Furthermore, the multipath cancellation data set and the interference suppression data set are arranged in such a way that the real and imaginary parts are divided into two layers, and the real and imaginary parts are organized in a serpentine arrangement. In the serpentine arrangement, the odd rows are arranged sequentially, while the even rows are arranged in reverse order. The data is rearranged into a two-layer two-dimensional matrix.

[0013] Furthermore, the multipath cancellation dataset needs to be normalized before entering the multipath cancellation network UNet and the interference suppression dataset needs to be normalized before entering the interference suppression network MS-UNet:

[0014]

[0015] x' is the received data after normalization, x is the received data, x max is the maximum value of the received data, x min It is the minimum value of received data.

[0016] Furthermore, the UNet part of the interference suppression network MS-UNet includes an encoder, a decoder, a jump connection connecting the encoder and the decoder, and a bottleneck.

[0017] Furthermore, the encoder of the UNet part includes an encoder block, and the decoder includes a decoder block. An encoder block consists of 2 convolutional layers, 1 DropBlock layer, 1 batch normalization layer, 1 RELU layer and 1 downsampling layer; a decoder block consists of 1 connection layer, 2 transposed convolutional layers, and 1 upsampling layer.

[0018] Furthermore, adding an SE module to the jump connection connecting the encoder and decoder of the interference suppression network MS-UNet can increase network parameters, prevent falling into local optimality, and accelerate network convergence. The SE module analyzes the input features of the jump connection from the encoder, compresses the features along the spatial dimension through the Squeeze operation, obtains the global receptive field, and then generates weights related to the feature channel correlation through the Excitation operation. The weight of the output of Excitation is regarded as the importance of each channel pair to the interference estimation after feature selection. The weight is normalized channel by channel by multiplication and weighted to the input feature to obtain the output of the SE module, which is connected to the receiving end decoder of the jump connection. The SE attention mechanism can enhance the acquisition of interference features and filter out features that are not related to interference, thereby more accurately estimating interference and enhancing the network's anti-interference performance.

[0019] Furthermore, a multi-head attention mechanism MHSA module is applied at the bottleneck of the interference suppression network MS-UNet. The input features passed to the bottleneck at the end of the encoder path are first processed by the convolution layer. Before entering the multi-head attention mechanism MHSA module, the data is adjusted to a one-dimensional vector sequence. Then, the attention map obtained by the multi-head attention mechanism MHSA module and the input features at the bottleneck are adjusted to the corresponding one-dimensional vector sequence for residual connection and addition. After the connection is obtained, the result is layer normalized and then converted into a two-dimensional format. After upsampling, it is connected to the decoder module.

[0020] Furthermore, the loss function LOSS is constructed to train the interference suppression network MS-UNet as follows:

[0021]

[0022] The loss function LOSS uses the mean square error as the main loss function and the mean absolute error as the secondary loss function, where n represents the length of the signal sequence, α is a hyperparameter that adjusts the weights of the mean square error and the mean absolute error, and y i is the true value, is the predicted value.

[0023] Beneficial Effects: On the one hand, the technical solution of the present invention changes the order of the real and imaginary data to a snake-like arrangement, which is more conducive to capturing the structural characteristics of interference. On the other hand, the network of the technical solution of the present invention fully combines the multi-scale feature extraction capability of the UNET network, the advantage of multi-head self-attention in capturing global dependencies, and the channel attention mechanism's focus on feature correlations between different channels, to better understand the pattern and structural characteristics of interference, thereby improving the effect of interference suppression. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A flow chart for the generation of a multipath elimination and interference suppression module at the FBMC underwater acoustic communication receiving end provided by the invention;

[0025] Figure 2 Schematic diagram of the snake-like arrangement of the data set;

[0026] Figure 3 This is the structure diagram of the SE (Squeeze-and-Excitation) module;

[0027] Figure 4 This is the structural diagram of the multipath elimination network UNet;

[0028] Figure 5 This is the structural diagram of the interference suppression network MS-UNet;

[0029] Figure 6 This is the module structure diagram of the multi-head attention mechanism MHSA. DETAILED DESCRIPTION

[0030] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, the present invention provides a technical solution:

[0032] A W-Net network-based FBMC underwater acoustic communication receiving end anti-interference method includes the following steps:

[0033] Step 1: Establish an FBMC underwater acoustic communication model. This FBMC underwater acoustic communication model uses the FBMC / OQAM modulation scheme. After inserting a pilot signal, the symbol is modulated onto a subcarrier and sent into the underwater acoustic channel. Additive interference signals are added through the FBMC underwater acoustic channel model to form the target received signal. In communication systems, interference signals are divided into artificial interference and non-artificial interference. Artificial interference primarily comes from ship signals in the ocean and is simulated using SαS pulses. Non-artificial interference primarily comes from natural biological activities such as killer whale echolocation clicks and is simulated using actual killer whale echolocation clicks. Both are sources of pulse interference in the ocean environment.

[0034] In the FBMC underwater acoustic communication model, the interference signal is superimposed to form the received signal r(n) expressed as:

[0035] r(n)=s(n)*h(n)+w(n)+j(n)

[0036] Where s(n) represents the transmitted signal, the transmission mode is FBMC, * represents the convolution operator, h(n) is the impulse response of the multipath channel, w(n) is Gaussian white noise, and j(n) is the interference;

[0037] Step 2: Construct a multipath cancellation dataset and an interference suppression dataset. Down-convert the received signal r(n) to a baseband signal, then filter it through a low-pass filter. This signal serves as the input of the multipath cancellation dataset, while the received signal s(n)+j(n), which is not affected by the multipath channel but only by the interference, serves as the output of the multipath cancellation dataset. The received signal s(n)+j(n), which is not affected by the multipath channel but only by the interference, serves as the input of the interference suppression dataset, while the transmitted signal s(n) serves as the output of the interference suppression dataset.

[0038] Step 3: Use the multipath elimination dataset to train the multipath elimination network UNet, save the trained network parameters, and obtain the trained underwater acoustic communication multipath elimination network model. The multipath elimination model UNet includes a ConvBlock encoder, an UpBlock decoder, a jump connection connecting the encoder and decoder, and a bottleneck; a lightweight network is used as the basic architecture based on the amount of data. The ConvBlock encoder consists of convolution, DropBlock, batch normalization, and activation functions; after batch normalization, downsampling is achieved through average pooling, and the UpBlock decoder increases the data size through upsampling to finally match the dimension of the output data;

[0039] In order to reduce overfitting, the DropBlock layer is added after the conv block of the ConvBlock encoder part. Batch normalization of the data can speed up the network convergence, and then nonlinearity is introduced through the activation function. Among them, DropBlock has a different shielding strategy compared to the traditional Dropout technology. Dropout shields by randomly selecting a single feature, while DropBlock shields by randomly selecting multiple continuous areas. This method effectively reduces the redundant connections in the neural network and helps to improve the generalization ability of the model. Multipath elimination network UNet is visible Figure 4 .

[0040] Step 4: Connect the interference suppression model MS-UNet in series after the multipath elimination model UNet, use the interference suppression data set to train the interference suppression network MS-UNet, save the trained network parameters, and obtain the trained underwater acoustic communication interference suppression network model. The interference suppression network MS-UNet includes a UNet part, three SE modules, and a multi-head self-attention module; with the assistance of the attention mechanism, the interference is extracted from the received signal, and then the estimated interference is eliminated from the input signal to achieve the purpose of interference suppression. The processed data is then received by the subsequent modules at the receiving end. The W-Net network consists of the multipath elimination model UNet and the serial interference suppression model MS-UNet; the overall interference suppression network MS-UNet can be seen Figure 5 .

[0041] In the FBMC system, the transmitter first performs symbol mapping on the input bit stream to generate a complex symbol sequence. Subsequently, the complex symbols are grouped according to the real and imaginary parts and modulated using OQAM (Offset Quadrature Amplitude Modulation), with the real and imaginary parts interleaved for half a symbol period. After inserting pilot symbols at the specified positions of the modulated signal, the signal undergoes IFFT transformation and is then processed by a polyphase filter bank composed of prototype filters to achieve signal filtering and shaping. Finally, each symbol is modulated onto the corresponding subcarrier and transmitted to the underwater acoustic channel. Interference is added to the underwater acoustic channel to form a multipath cancellation data set and an interference suppression data set.

[0042] Preferably, the multipath cancellation data set and the interference suppression data set are arranged in such a way that the real and imaginary parts are divided into two layers, and the real and imaginary parts are organized in a serpentine arrangement, wherein the odd rows of the serpentine arrangement are arranged sequentially and the even rows are arranged in reverse order, and the data is rearranged into a 2-layer 2D matrix.

[0043] If the conventional reshape method is used, the direct distance d between sequences may fluctuate greatly as the signal scale increases, and the structural signal characteristics may not be well captured and maintained. The data is organized in a serpentine arrangement, where odd rows are arranged in order and even rows are arranged in reverse order. The specific arrangement can be seen in Figure 2 The signal points are rearranged into a 2-layer 2D matrix to facilitate the attention mechanism to capture the correlation between the previous and next data sequences. The performance is improved by converting the sequence signal anti-interference task into an image anti-interference task to meet the needs of local sampling and perception range of the convolution kernel. Figure 2The permutation method is intuitively understood: each channel contains all the real and imaginary parts of the received data. Rei(n) represents the real part of the i-th received data in the n-th FBMC block. Imi(n) represents the imaginary part of the i-th received data in the n-th FBMC block. Compared to the simple reshape permutation method, the snake permutation method is more suitable for training neural network models for multipath resistance and interference suppression.

[0044] Preferably, the multipath cancellation data set needs to be normalized before entering the multipath cancellation network UNet and the interference suppression data set needs to be normalized before entering the interference suppression network MS-UNet:

[0045]

[0046] x' is the received data after normalization, x is the received data, x max is the maximum value of the received data, x min It is the minimum value of received data.

[0047] Preferably, the UNet part of the interference suppression network MS-UNet includes an encoder, a decoder, a jump connection connecting the encoder and the decoder, and a bottleneck.

[0048] Preferably, the encoder of the UNet part includes an encoder block, and the decoder includes a decoder block. An encoder block consists of 2 convolutional layers, 1 DropBlock layer, 1 batch normalization layer, 1 RELU layer and 1 downsampling layer; a decoder block consists of 1 connection layer, 2 transposed convolutional layers and 1 upsampling layer.

[0049] Preferably, adding a SE module at the jump connection connecting the encoder and decoder of the interference suppression network MS-UNet can increase network parameters, prevent falling into local optimality, and accelerate network convergence. The SE module analyzes the input features of the jump connection from the encoder, compresses the features along the spatial dimension through the Squeeze operation, obtains the global receptive field, and then generates weights related to the feature channel correlation through the Excitation operation. The weight of the output of Excitation is regarded as the importance of each channel pair to the interference estimation after feature selection, and the weight is normalized channel by channel by multiplication and weighted to the input feature to obtain the output of the SE module, which is connected to the receiving end decoder of the jump connection. The SE attention mechanism can enhance the acquisition of interference features, filter out features that are not related to interference, thereby more accurately estimating interference and enhancing the network's anti-interference performance. The structure of the SE attention mentioned in this step is as follows: Figure 3 shown.

[0050] Preferably, a multi-head attention mechanism MHSA module is applied at the bottleneck of the interference suppression network MS-UNet, and the input features passed to the bottleneck at the end of the encoder path are first processed by the convolution layer. Before entering the multi-head attention mechanism MHSA module, the data is first adjusted to a one-dimensional vector sequence, and then the attention map obtained by the multi-head attention mechanism MHSA module and the input features at the bottleneck are adjusted to the corresponding one-dimensional vector sequence for residual connection and addition. After the connection is obtained, the result is layer normalized and then converted into a two-dimensional format. After upsampling, it is connected to the decoder module.

[0051] MHSA Modules:

[0052] MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O

[0053] h is the number of heads, that is, the number of parallel heads in the multi-head attention mechanism. O It is a weight matrix of the output linear transformation, which is used to merge the outputs of different heads together. i The calculation formula is as follows:

[0054] head i =Attention(QW i Q ,KW i K ,VW i V )

[0055] Q represents the query vector, K represents the key vector, and V represents the value vector. i Q ,W i K ,W i V are the projection matrices of query, key, and value, which are trainable parameters used to map the input to the new representation space. Q, K, V are first transformed by different linear transformations W. i Q ,W i K ,W i V Map to the new subspace.

[0056] Attention calculation for each head:

[0057]

[0058] T represents transpose, d kRepresents the dimension of K, which is used to scale the dot product to prevent the gradient vanishing problem. Attention is calculated within the sequence, focusing on the relationship between different parts of the sequence.

[0059] The MHSA module is applied in the bottleneck passed to the decoder to obtain the spatial attention distribution of the input sequence, thereby more comprehensively capturing the potential multiple associations in the sequence. The MHSA attention module mentioned in this step is as follows Figure 6 shown.

[0060] Preferably, the loss function MSE is constructed to train the multipath elimination network.

[0061]

[0062] It is used to input the received signal affected by multipath into the multipath elimination network, train the multipath elimination network based on the constructed loss function MSE, and save the trained network parameters to obtain a network model that eliminates the influence of multipath.

[0063] Preferably, the loss function LOSS is constructed to train the interference suppression network MS-UNet as follows:

[0064]

[0065] The loss function LOSS uses the mean square error as the main loss function and the mean absolute error as the secondary loss function, where n represents the length of the signal sequence, α is a hyperparameter that adjusts the weights of the mean square error and the mean absolute error, and y i is the true value, is the predicted value.

[0066] The W-Net network is an end-to-end processing module that performs multipath distortion correction and interference suppression in FBMC underwater acoustic communication systems without the need for traditional channel estimation and equalization algorithms. Requiring no prior knowledge, the integrated model is designed to process received signals in underwater acoustic communication environments with reduced multipath effects, enhanced anti-interference capabilities, and improved communication performance.

[0067] According to the above description, the present invention provides a method of anti-interference through deep learning. Although the present invention has been described in detail through specific embodiments, it should be understood that the embodiments are only for illustration and not for limitation. Those skilled in the art may modify and change the embodiments without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope defined by the claims.

Claims

1. A W-Net network-based FBMC underwater acoustic communication receiving end anti-interference method, characterized in that: The following steps are involved: Step 1: Establish an FBMC underwater acoustic communication model. The FBMC underwater acoustic communication model uses the FBMC / OQAM modulation method. After inserting the pilot, the symbol is modulated onto the subcarrier and sent into the underwater acoustic channel. The interference signal is superimposed in the FBMC underwater acoustic communication model to form a received signal r(n) expressed as: r(n)=s(n)*h(n)+w(n)+j(n) Where s(n) represents the transmitted signal, the transmission mode is FBMC, * represents the convolution operator, h(n) is the impulse response of the multipath channel, w(n) is Gaussian white noise, and j(n) is the interference; Step 2: Construct a multipath cancellation dataset and an interference suppression dataset. Down-convert the received signal r(n) to a baseband signal, then filter it through a low-pass filter. This signal serves as the input of the multipath cancellation dataset, while the received signal s(n)+j(n), which is not affected by the multipath channel but only by the interference, serves as the output of the multipath cancellation dataset. The received signal s(n)+j(n), which is not affected by the multipath channel but only by the interference, serves as the input of the interference suppression dataset, while the transmitted signal s(n) serves as the output of the interference suppression dataset. Step 3: Use the multipath cancellation dataset to train the multipath cancellation model UNet. The multipath cancellation model UNet includes a ConvBlock encoder, an UpBlock decoder, a skip connection connecting the encoder and decoder, and a bottleneck. The ConvBlock encoder is composed of convolution, DropBlock, batch normalization, and an activation function. After batch normalization, downsampling is achieved through average pooling, and the UpBlock decoder increases the data size by upsampling to match the output data dimension. Step 4: Use the interference suppression dataset to train the interference suppression network MS-UNet; the interference suppression network MS-UNet includes a UNet part, three SE modules and a multi-head self-attention module; With the help of the attention mechanism, interference is extracted from the received signal, and the estimated interference is then eliminated from the input signal. The processed data is then received by the subsequent modules at the receiving end. The interference suppression model MS-UNet is connected in series after the multipath elimination model UNet to form a W-Net network as the underwater acoustic communication receiving end. The multipath cancellation data set and the interference suppression data set are arranged in such a way that the real and imaginary parts are divided into two layers, and the real and imaginary parts are organized in a serpentine arrangement. In the serpentine arrangement, the odd rows are arranged sequentially, while the even rows are arranged in reverse order. The data is rearranged into a two-layer two-dimensional matrix.

2. The anti-interference method for an FBMC underwater acoustic communication receiving end based on a W-Net network according to claim 1 is characterized in that: The multipath elimination dataset needs to be normalized before entering the multipath elimination model UNet and the interference suppression dataset needs to be normalized before entering the interference suppression network MS-UNet: x' is the received data after normalization, x is the received data, x max is the maximum value of the received data, x min It is the minimum value of received data.

3. The anti-interference method for the FBMC underwater acoustic communication receiving end based on the W-Net network according to claim 1 is characterized in that: The UNet part of the interference suppression network MS-UNet includes an encoder, a decoder, a jump connection connecting the encoder and the decoder, and a bottleneck.

4. The anti-interference method for the FBMC underwater acoustic communication receiving end based on the W-Net network according to claim 3 is characterized in that: The encoder of the UNet part includes an encoder block, and the decoder includes a decoder block. An encoder block consists of 2 convolutional layers, 1 DropBlock layer, 1 batch normalization layer, 1 RELU layer and 1 downsampling layer; a decoder block consists of 1 connection layer, 2 transposed convolutional layers, and 1 upsampling layer.

5. The anti-interference method for the FBMC underwater acoustic communication receiving end based on the W-Net network according to claim 3 is characterized in that: An SE module is added to the jump connection connecting the encoder and decoder of the interference suppression network MS-UNet. The SE module analyzes the input features of the jump connection from the encoder, compresses the features along the spatial dimension through the Squeeze operation, obtains the global receptive field, and then generates weights related to the feature channel correlation through the Excitation operation. The weight of the output of Excitation is regarded as the importance of each channel pair to the interference estimation after feature selection. The weight is normalized channel by channel by multiplication and weighted to the input feature to obtain the output of the SE module, which is connected to the receiving decoder of the jump connection.

6. The anti-interference method for the FBMC underwater acoustic communication receiving end based on the W-Net network according to claim 3 is characterized in that: A multi-head attention mechanism MHSA module is applied at the bottleneck of the interference suppression network MS-UNet. The input features passed to the bottleneck at the end of the encoder path are first processed by the convolution layer. Before entering the multi-head attention mechanism MHSA module, the data is adjusted to a one-dimensional vector sequence. Then, the attention map obtained by the multi-head attention mechanism MHSA module and the input features at the bottleneck are adjusted to the corresponding one-dimensional vector sequence for residual connection and addition. After the connection is obtained, the result is layer normalized and then converted into a two-dimensional format. After upsampling, it is connected to the decoder module.

7. The anti-interference method for an FBMC underwater acoustic communication receiving end based on a W-Net network according to claim 1 is characterized in that: Construct the loss function LOSS to train the interference suppression network MS-UNet as follows: The loss function LOSS uses the mean square error as the main loss function and the mean absolute error as the secondary loss function, where n represents the length of the signal sequence, α is a hyperparameter that adjusts the weights of the mean square error and the mean absolute error, and y i is the true value, is the predicted value.

8. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, an anti-interference method for an FBMC underwater acoustic communication receiving end based on a W-Net network is implemented as described in any one of claims 1 to 7.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements an anti-interference method for an FBMC underwater acoustic communication receiving end based on a W-Net network according to any one of claims 1 to 7.

Images