Underwater acoustic communication signal modulation identification method based on multi-feature fusion and CBAM-CNN network

Abstract

The invention discloses an underwater acoustic communication signal modulation identification method based on multi-feature fusion and a CBAM-CNN network, and the method comprises the steps: reading an input signal, and generating Gramer angle field data, Markov transfer field data and recurrence plot data; combining the Gramb angle field data, the Markov transfer field data and the recurrence plot data in a channel direction to generate a mixed image and construct a training set and a test set; constructing a modulation recognition model, wherein the modulation recognition model comprises a CBAM module and a CNN module; inputting the mixed image as a fusion feature into a CBAM-CNN network, and obtaining the probability that the input signal belongs to different modulation types; and selecting the type with the highest probability as the real type of the input signal. According to the method, the structure and dynamic characteristics of the signal can be more comprehensively understood, and the classification accuracy and generalization ability of the network are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and in particular relates to an underwater acoustic communication signal modulation recognition method based on multi-feature fusion and CBAM-CNN network. Background Art

[0002] Underwater acoustic communication is currently the only way to achieve long-distance, high-speed data communication underwater, and modulation signal recognition, as a key link, plays an important role. Traditional modulation signal recognition usually directly maps the IQ signal data sequence into a constellation diagram, a cyclic spectrum, or directly inputs the time series signal into the network. However, the information provided by these methods is relatively simple, which may lead to poor signal classification performance. Summary of the invention

[0003] Purpose of the invention: In order to overcome one or more problems caused by the limitations and defects of the relevant technologies to a certain extent, the present invention provides a method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network.

[0004] Technical solution: A method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network. It includes the following steps:

[0005] Step 1: Read the input signal and generate Gram angle field data, Markov transition field data and recursive graph data;

[0006] Step 2: combining the Gram angle field data, the Markov transition field data and the recursive graph data in the channel direction to generate a mixed image and construct a training set and a test set;

[0007] Step 3: Construct a modulation recognition model, including the CBAM module and the CNN module;

[0008] Step 4: Input the mixed image as a fusion feature into the CBAM-CNN network to obtain the probability that the input signal belongs to different modulation types;

[0009] Step 5: Select the type with the highest probability as the true type of the input signal.

[0010] The present invention also discloses a communication signal modulation recognition system based on a multi-feature fusion network, comprising:

[0011] An acquisition module is configured to: read an input signal, generate Gram angle field data, Markov transition field data, and recurrence graph data;

[0012] The fusion module is configured to extract and fuse the Gram angle field data, the Markov transition field data and the recursive graph data to obtain a mixed image representation, namely, a fusion feature;

[0013] The recognition module is configured as follows: the fused features are fed into a CBAM-CNN network, which is capable of extracting spatiotemporal signal correlations, and finally obtaining the probability that the input signal belongs to different types;

[0014] The selection module is configured to select the type with the highest probability as the true type of the input signal.

[0015] Beneficial effects: On the one hand, the technical solution of the present invention integrates Gram angle field data, Markov transition field data and recursive graph data, which helps to more comprehensively understand the structure and dynamic characteristics of the signal. On the other hand, the network of the technical solution of the present invention fully combines the image feature extraction capability of the convolutional neural network and the global modeling capability of the convolutional block attention mechanism module, realizing the extraction of local deep features and global features of the modulated signal, thereby significantly improving the classification accuracy and generalization ability of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The flowchart of the modulation recognition of underwater acoustic communication signals based on multi-feature fusion and CBAM-CNN network provided by the present invention;

[0017] Figure 2 This is the CBAM module structure diagram;

[0018] Figure 3 This is the structure diagram of the channel attention module;

[0019] Figure 4 This is the structure diagram of the spatial attention module;

[0020] Figure 5 This is the structure diagram of the CNN module. DETAILED DESCRIPTION

[0021] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0022] Embodiment 1:

[0023] like Figure 1 As shown, this embodiment discloses a method for underwater acoustic communication signal modulation recognition using multi-feature fusion and CBAM-CNN network, which mainly includes the following steps:

[0024] Step 1: Read the input signal and generate Gram angle field data, Markov transition field data and recursive graph data;

[0025] The specific operations of the Gram angle field data mentioned in this step include:

[0026] According to the received modulated signal, extract its time series:

[0027] X={x1 ,x 2 ,…,x n} (1)

[0028] Normalize the time series data to the range of (-1, 1):

[0029]

[0030] x i Map to the polar coordinate domain:

[0031]

[0032] In the formula, θ i Represents the arccosine angle, i represents the time step, N is the length of the time series, A is the normalized time series set, and r is the normalized radial distance corresponding to the time step i. This conversion provides a new perspective for understanding time series.

[0033] According to the above results, calculate the GASF matrix:

[0034]

[0035] The Markov transition field data mentioned in this step includes the following operations:

[0036] According to the received modulated signal, extract its time series:

[0037] X={x 1 ,x 2 ,…,x n} (1)

[0038] Normalize the time series data to the range of (-1, 1):

[0039]

[0040] According to the normalized time series X, it is divided into Q intervals, each x i Corresponding to an interval q i .

[0041] Calculate the transition matrix W using a first-order Markov transition chain:

[0042]

[0043] In the formula, w ij Indicates that it is in the interval q i A sample point in the interval q is then j The probability of a sample point following. ij Satisfy wij =p{x t ∈q i ∣x t-1 ∈q j}, while satisfying

[0044] Expand the Markov transition matrix according to the time sequence of the time series data and construct the transition probability matrix M:

[0045]

[0046] Where M ij Indicates that from x i Area i to x j Area j The transition probability.

[0047] The specific operations of the recursive graph data mentioned in this step include:

[0048] According to the received modulated signal, extract its time series:

[0049] X={x 1 ,x 2 ,…,x n} (1)

[0050] Normalize the time series data to the range of (-1, 1):

[0051]

[0052] According to Takens embedding theorem, we reconstruct the phase space, and the reconstructed signal is expressed as:

[0053] X i =[u i ,u i+τ ,…,u i+(m-1)τ ] (7)

[0054] Where i = 1, 2, ..., n-(m-1)τ, m is the embedding dimension, and τ is the time delay. For the values ​​of m and τ, a more appropriate value is selected according to the pseudo-neighborhood method and the average mutual information method.

[0055] According to each pair of embedding vectors X i , calculate the Euclidean distance between them:

[0056] S ij =||x i -x j ||(8)

[0057] In the formula, ||·|| represents the norm,

[0058] By setting a distance threshold ε, the distance matrix S ij Convert to binary recursive matrix R(i,j):

[0059] R(i,j)=θ(ε i -S ij )(9)

[0060] Where θ represents the Heaviside step function, which is expressed as:

[0061]

[0062] Step 2: combining the Gram angle field data, the Markov transition field data and the recursive graph data in the channel direction to generate a mixed image and construct a training set and a test set;

[0063] The combination in the channel direction mentioned in this step is specifically performed as follows: performing pseudo-color mapping on the Gram angle field matrix, the Markov transfer field matrix, and the recursive graph matrix in the channel direction to generate an RGB image.

[0064] Step 3: Construct a modulation recognition model, including the CBAM module and the CNN module;

[0065] The CBAM modules mentioned in this step are as follows: Figure 2 As shown, specifically, there are channel attention module and spatial attention module.

[0066] Furthermore, the channel attention module, such as Figure 3 As shown, specifically:

[0067] First, the input feature layer is subjected to global average pooling and global maximum pooling to capture feature information in a global range. Then, the two pooling results are processed through a shared multi-layer perceptron to reduce the number of parameters and enhance feature expression capabilities. Next, the processed results are added together and a channel attention map is generated through a Sigmoid activation function to assign weights to each channel of the input feature layer. Finally, these weights are used to weight the input feature layer channel by channel to increase the network's attention to key features.

[0068] Furthermore, the spatial attention module, such as Figure 4 As shown, specifically:

[0069] First, the input feature map is subjected to global average pooling and global maximum pooling along the channel direction to extract the global information of each feature point in the channel dimension. Then, the two pooling results are stacked along the channel dimension to generate an effective feature descriptor. It is then processed through a standard convolutional layer to capture relevant features in the spatial dimension. Next, a 2D spatial attention map is generated using the Sigmoid activation function to assign weights to each feature point in the input feature map. Finally, these weights are weighted channel by channel to the input feature map to enhance important features in the spatial dimension.

[0070] The CNN modules mentioned in this step are as follows: Figure 5 As shown, it includes 5 convolutional layers, 5 batch normalization layers, 5 ReLU layers, 5 maximum pooling layers, 1 random inactivation layer, 1 flattening layer, and 1 Softmax layer.

[0071] Furthermore, the sizes of the convolution kernels of the five convolution layers are 7×7, 5×5, 3×3, 3×3, and 3×3, respectively. The number of convolution kernels is set to 64, 128, 256, 256, and 256, respectively.

[0072] Step 4: Input the mixed image as a fusion feature into the CBAM-CNN network to obtain the probability that the input signal belongs to different modulation types;

[0073] The CBAM-CNN network mentioned in this step specifically adds a CBAM module after each batch normalization layer of CNN, for a total of 5 modules.

[0074] Step 5: Select the type with the highest probability as the true type of the input signal.

[0075] The present invention extracts the features of time domain signals through Gram angle field, Markov transfer field and recursive graph algorithm, and combines these three features into an image as the input of the network model. Furthermore, by combining the convolutional block attention module and the convolutional neural network module, the image information is captured more deeply and the representation ability of the network is improved. Compared with direct input of time domain signals, the signal recognition accuracy is significantly improved, with stronger anti-noise ability and less susceptible to noise.

[0076] Embodiment 2:

[0077] This embodiment discloses a system for underwater acoustic communication signal modulation recognition using multi-feature fusion and CBAM-CNN network, including:

[0078] An acquisition module is configured to: read an input signal, generate Gram angle field data, Markov transition field data, and recurrence graph data;

[0079] The fusion module is configured to extract and fuse the Gram angle field data, the Markov transition field data and the recursive graph data to obtain a mixed image representation, namely, a fusion feature;

[0080] The recognition module is configured as follows: the fused features are fed into a CBAM-CNN network, which is capable of extracting spatiotemporal signal correlations, and finally obtaining the probability that the input signal belongs to different types;

[0081] The selection module is configured to select the type with the highest probability as the true type of the input signal.

[0082] The sub-technical solution of this embodiment captures the transient characteristics of the modulated signal through the Gram angle field, Markov transition field and recursive graph algorithm. At the same time, normalization preprocessing is performed during the processing to improve the stability and convergence speed of the model. Through the feature fusion module, the three types of data are extracted and fused, and the obtained mixed image representation is the fusion feature, which captures the relationship between the time series data points and effectively mines the time series structure. The fused features are sent to the recognition module, and the convolution block attention module can extract the spatiotemporal signal correlation, further enhancing the feature extraction capability of the convolutional neural network. The final selection module outputs the result to obtain the predicted modulation signal category.

[0083] The steps involved in the above embodiment 2 correspond to those in embodiment 1. For the specific implementation method, please refer to the relevant description part of embodiment 1.

Claims

1. A method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network, characterized in that: The following steps are involved: Step 1: Read the input signal and generate Gram angle field data, Markov transition field data and recursive graph data; Step 2: combining the Gram angle field data, the Markov transition field data and the recursive graph data in the channel direction to generate a mixed image and construct a training set and a test set; Step 3: Construct a modulation recognition model, including the CBAM module and the CNN module; Step 4: Input the mixed image as a fusion feature into the CBAM-CNN network to obtain the probability that the input signal belongs to different modulation types; Step 5: Select the type with the highest probability as the true type of the input signal.

2. According to the method of multi-feature fusion and CBAM-CNN network for underwater acoustic communication signal modulation recognition in claim 1, it is characterized in that: Generate Gram angle field data. The specific operations include: According to the received modulated signal, extract its time series: X={x1,x2,…,x n } Normalize the time series data to the range of (-1, 1): x i Map to the polar coordinate domain: In the formula, θ i represents the arccosine angle, i represents the time step, N is the length of the time series, A is the normalized time series set, r is the normalized radial distance corresponding to the time step i, According to the above results, calculate the GASF matrix:

3. According to the method of underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network in claim 1, it is characterized in that: Generate Markov transition field data. The specific operations include: According to the received modulated signal, extract its time series: X={x1,x2,…,x n } Normalize the time series data to the range of (-1, 1): According to the normalized time series X, it is divided into Q intervals, each x i Corresponding to an interval q i , Calculate the transition matrix W using a first-order Markov transition chain: In the formula, w ij Indicates that it is in the interval q i A sample point in the interval q is then j The probability of a sample point in following; w ij Satisfy w ij =p{x t ∈q i ∣x t-1 ∈q j }, while satisfying Expand the Markov transition matrix according to the time sequence of the time series data and construct the transition probability matrix M: Where M ij Indicates that from x i Area i to x j Area j The transition probability.

4. According to claim 1, a method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network is characterized in that: Generate recursive graph data. The specific operations include: According to the received modulated signal, extract its time series: X={x1,x2,…,x n } Normalize the time series data to the range of (-1, 1): According to Takens embedding theorem, the phase space is reconstructed and the reconstructed signal is expressed as: X i =[u i ,u i+τ ,…,u i+(m-1)τ ] Where i=1,2,…,n-(m-1)τ, m is the embedding dimension, and τ is the time delay.

5. The method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network according to claim 4 is characterized in that: For the values ​​of m and τ, a more appropriate value is selected according to the pseudo-neighborhood method and the average mutual information method; According to each pair of embedding vectors X i , calculate the Euclidean distance between them: S ij =||x i -x j || In the formula, ||·|| represents the norm, By setting a distance threshold ε, the distance matrix S ij Convert to binary recursive matrix R(i,j): R(i,j)=θ(ε i -S ij ) Where θ represents the Heaviside step function, which is expressed as:

6. The method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network according to claim 1 is characterized in that: The Gram angle field data, the Markov transfer field data and the recursive graph data are combined in the channel direction. The specific operation is: performing pseudo-color mapping on the Gram angle field matrix, the Markov transfer field matrix and the recursive graph matrix in the channel direction to generate an RGB image.

7. The method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network according to claim 1 is characterized in that: The CBAM module includes a channel attention module and a spatial attention module; The channel attention module first performs global average pooling and global maximum pooling on the input feature layer to capture feature information in a global range, then processes the two pooling results through a shared multi-layer perceptron, then adds the processed results and generates a channel attention map through a Sigmoid activation function to assign weights to each channel of the input feature layer, and finally, uses these weights to weight the input feature layer channel by channel to enhance the network's attention to key features.

8. The method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network according to claim 7 is characterized in that: The spatial attention module is specifically: First, global average pooling and global maximum pooling are performed on the input feature map along the channel direction to extract the global information of each feature point in the channel dimension. Then, the two pooling results are stacked along the channel dimension to generate an effective feature descriptor, which is then processed through a standard convolutional layer to capture the relevant features of the spatial dimension. Then, the Sigmoid activation function is used to generate a two-dimensional spatial attention map to assign weights to each feature point in the input feature map. Finally, these weights are weighted to the input feature map channel by channel to enhance the important features in the spatial dimension.

9. The method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network according to claim 1 is characterized in that: The CNN module includes 5 convolutional layers, 5 batch normalization layers, 5 ReLU layers, 5 maximum pooling layers, 1 random inactivation layer, 1 flattening layer, and 1 Softmax layer.

10. The method for underwater acoustic communication signal modulation recognition based on multi-feature fusion and CBAM-CNN network according to claim 1 is characterized in that: The CBAM-CNN network specifically adds a CBAM module after each batch normalization layer of CNN.

Images