Underwater acoustic communication signal end-to-end modulation type identification method and system

Abstract

The invention relates to the field of underwater acoustic signal processing, in particular to an underwater acoustic communication signal end-to-end modulation type identification method and system, and the method comprises the steps: detecting a communication signal in an underwater acoustic signal, and obtaining an underwater acoustic communication signal; performing modulation type identification on the obtained underwater acoustic communication signal by using a pre-established and trained underwater acoustic communication signal end-to-end modulation type identification model based on residual attention convolution; wherein the trained underwater acoustic communication signal end-to-end modulation type identification model based on residual attention convolution comprises a feature extraction module and a classification module; wherein the feature extraction module is used for carrying out feature extraction on an underwater acoustic communication signal based on residual attention; and the classification module is used for classifying the extracted features to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type. And a residual structure is combined with a convolution module, so that weak changes of signals can be adapted, the sampling capability of the signals is enhanced, and the recognition reliability is improved.

Description

Technical Field

[0001] The present invention relates to the field of underwater acoustic signal processing, and in particular to a method and system for identifying end-to-end modulation types of underwater acoustic communication signals. Background Art

[0002] At present, underwater acoustic communication technology has achieved unprecedented development. The modulation types of underwater acoustic communication signals are mainly frequency modulation signals, phase modulation signals and spread spectrum signals, which are characterized by low operating frequency and short communication distance. Under the existing communication system, the study of modulation type identification of underwater acoustic communication signals in non-cooperative environments has a decisive impact on the smooth implementation of various underwater tasks. However, the rapid attenuation of sound waves underwater, the multipath effect of seawater, and the Doppler deviation in the ocean will all affect underwater communication signals. The underwater noise is complex, and fish schools, ships, ocean currents, sea surface waves, seismic activities, and tidal noise all bring great challenges to communication signal processing and identification.

[0003] At present, researchers have also conducted extensive research on the modulation recognition of underwater acoustic communication signals. For example, the energy detection method is used to intercept underwater pulse signals, and on this basis, different features such as time domain features, spectrum features, and parameter features of the signal are extracted. In recent years, convolutional neural networks have shown excellent performance in target classification tasks, and have been used in underwater acoustic target recognition such as ship radiation noise classification and sonar image target recognition. They are also widely used in the recognition of modulation methods of underwater acoustic communication signals. Signal recognition can be effectively achieved through expert knowledge bases or classifiers based on machine learning and deep learning. However, for most research methods, the feature extraction process usually uses a single feature, which has great limitations and hinders the high-reliability recognition of communication signals.

[0004] It can be seen that the modulation type identification of non-cooperative underwater acoustic communication signals still faces many difficulties, which also puts forward new requirements for subsequent research. How to perform robust feature extraction and then realize fast, accurate and reliable identification of signal attributes is an important research direction. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and propose a method and system for identifying the end-to-end modulation type of underwater acoustic communication signals.

[0006] To achieve the above object, the present invention is implemented by the following technical solutions:

[0007] On the one hand, a method for identifying an end-to-end modulation type of an underwater acoustic communication signal is provided, comprising:

[0008] Detecting the communication signal in the underwater acoustic signal to obtain the underwater acoustic communication signal;

[0009] The modulation type of the acquired underwater acoustic communication signal is identified by using the pre-established and trained end-to-end modulation type identification model of underwater acoustic communication signal based on residual attention convolution;

[0010] Among them, the trained end-to-end modulation type recognition model of underwater acoustic communication signals based on residual attention convolution includes: a feature extraction module and a classification module; among them,

[0011] The feature extraction module is used to extract features of the underwater acoustic communication signal based on residual attention;

[0012] The classification module is used to classify the extracted features to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type.

[0013] In an improved end-to-end modulation type recognition method for underwater acoustic communication signals, the feature extraction module includes: M pairs of RAC and DC modules consisting of M+1 residual attention convolution blocks RAC and M depth convolution blocks DC and 1 RAC module, which are connected in sequence; wherein the first pair of RAC and DC modules respectively extract features of the communication signal, and connect the output features through a residual structure, and the remaining pairs of RAC and DC modules are used to extract features from the output features; the 1 RAC module is used to continue to extract features from the features output by the last pair of RAC and DC modules.

[0014] In an improved end-to-end modulation type recognition method for underwater acoustic communication signals, each RAC specifically includes: a first one-dimensional convolutional layer Conv1, a first ReLU6 activation layer, a second one-dimensional convolutional layer Conv2, a one-dimensional convolutional layer Attention with a channel attention mechanism, a second ReLU6 activation layer and a maximum pooling layer MaxPooling connected in sequence;

[0015] The Conv1 is used to use y i =w i ·x+b i The formula extracts features from underwater acoustic communication signals, where x is the input data of the convolution layer, and w i and b i The output features y are i The corresponding convolution kernel and bias, i is the number of convolution layers;

[0016] The first ReLU6 activation layer is used to activate the extracted features to obtain activated features;

[0017] The Conv2 is used to continue to extract features from the activated features using the formula;

[0018] The Attention is used to reduce the dimension of the features extracted by Conv2 using global pooling, output a set of vectors with the same dimension as the number of channels, and then add two dense layers using RELU activation and Sigmoid activation respectively, using the formula To learn the weight of each channel, we get the attention weighted feature map, where p i is the feature output by Conv2, σ i is the channel weight, is the attention weighted feature map;

[0019] The second ReLU6 activation layer is used to activate the attention weighted features to obtain activated features;

[0020] The MaxPooling is used to downsample the output features to 1 / 1 of the original length using a pooling kernel of size l×1, thereby removing redundant information and reducing model parameters.

[0021] In an improved end-to-end modulation type identification method for underwater acoustic communication signals, each DC specifically includes: four ordinary one-dimensional convolutional layers and one ReLU6 activation layer connected in sequence;

[0022] The four common one-dimensional convolutional layers connected in sequence are used to continue extracting features from the underwater acoustic communication signal;

[0023] The ReLU6 activation layer is used to activate the input features using the formula ReLU6(x)=min(6,max(x,0)) to obtain the activated features, where x represents the input value of the activation function, min(·) represents the minimum value function, and max(·) represents the maximum value function.

[0024] In an improved method for end-to-end modulation type recognition of underwater acoustic communication signals, the classification module includes: a global average pooling layer, a fully connected layer and a Softmax activation layer connected in sequence;

[0025] The global average pooling layer is used to perform channel average downsampling on the extracted features;

[0026] The fully connected layer is used to generate a modulation type label of the underwater acoustic communication signal based on the downsampled features;

[0027] The Softmax activation layer is used to calculate the modulation type label of the underwater acoustic communication signal using an activation function to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type.

[0028] In an improved method for identifying end-to-end modulation types of underwater acoustic communication signals, the activation function is:

[0029]

[0030] Among them, Softmax(z f ) is the probability that the underwater acoustic communication signal belongs to each modulation type, z f is the modulation type label of the output of the fth node of the fully connected layer, k is the modulation type index of the underwater acoustic communication signal, and K is the total number of modulation types of the underwater acoustic communication signal.

[0031] In an improved method for identifying end-to-end modulation types of underwater acoustic communication signals, the end-to-end modulation type identification model of underwater acoustic communication signals is trained in the following manner:

[0032] Construct training dataset D train And the test dataset D test : Detect the communication signal in the acquired underwater acoustic signal to obtain N one-dimensional signals X1, X2, ..., X of the same length of K-type underwater acoustic communication signals N Samples and corresponding true label vectors Y1,Y2,...,Y N ; Randomly select N1 samples and corresponding label vectors to form the training data set D train , the remaining N-N1 samples and their corresponding label vectors constitute the test data set D test ;

[0033] Initialize the modulation type recognition model: The parameters of each network layer are Gaussian distributed with a mean of 0 and a variance of 1 (μ,σ 2 ) Random initialization;

[0034] Training the modulation type recognition model: Divide the training data set into batches according to the set batch size; Perform one round of training on all batches of data in the model to obtain the optimized model, and then use all batches of data to perform the next round of training on the optimized model until the number of training rounds reaches the set number of epochs;

[0035] The first round of training for the model includes: selecting a batch from the batch data, inputting the initialized end-to-end modulation type recognition model of underwater acoustic communication signals, combining the true label vector, using the cross entropy loss function to calculate the function loss value Loss of the model, using the error back propagation BP algorithm to update the modulation type recognition model parameters, and using the Adam optimizer with an adaptive learning rate to optimize the model parameters, repeatedly selecting a batch of data from the training set, and inputting the optimized modulation type recognition model until all batches of data are selected;

[0036] Other rounds of training for the model include: selecting a batch from the batch data, inputting the end-to-end modulation type recognition model of underwater acoustic communication signals obtained from the previous round of training, combining the true label vector, using the cross entropy loss function to calculate the function loss value Loss of the model, using the error back propagation BP algorithm to update the modulation type recognition model parameters, and using the Adam optimizer with an adaptive learning rate to optimize the model parameters, repeatedly selecting a batch of data from the training set, and inputting the optimized modulation type recognition model until all batches of data are selected.

[0037] In an improved method for end-to-end modulation type identification of underwater acoustic communication signals, the Loss calculation formula is:

[0038]

[0039] Among them, N1 represents the number of training set samples, K is the total number of underwater acoustic communication signal modulation types, k is the modulation type index, j is the sample index, q(x j,k ) represents the sample x j,k The predicted modulation type, p(x j,k ) represents the sample x j,k is the probability of the current predicted modulation type.

[0040] On the other hand, a modulation type identification system based on the above-mentioned end-to-end modulation type identification method for underwater acoustic communication signals is provided, comprising:

[0041] An acquisition module, used for detecting the communication signal in the underwater acoustic signal to obtain the underwater acoustic communication signal;

[0042] The recognition module uses a pre-established and trained end-to-end modulation type recognition model for underwater acoustic communication signals based on residual attention convolution to recognize the modulation type of the acquired underwater acoustic communication signal; wherein the trained end-to-end modulation type recognition model for underwater acoustic communication signals based on residual attention convolution specifically includes: a feature extraction module and a classification module; the feature extraction module is used to extract features of the underwater acoustic communication signal based on residual attention; the classification module is used to classify the extracted features to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type.

[0043] In an improved end-to-end modulation type recognition system for underwater acoustic communication signals, the feature extraction module includes: M pairs of RAC and DC modules consisting of M+1 residual attention convolution blocks RAC and M depth convolution blocks DC and 1 RAC module, which are connected in sequence; wherein the first pair of RAC and DC modules respectively extract features from the communication signal, and connect the output features through a residual structure, and the remaining pairs of RAC and DC modules are used to extract features from the output features; the 1 RAC module is used to continue to extract features from the features output by the last pair of RAC and DC modules.

[0044] Compared with the prior art, the present invention has the following advantages:

[0045] 1. Automatically identify multiple types of underwater acoustic communication signals based on intelligent methods. As an important tool for signal processing, convolutional neural networks can automatically learn effective features from training data for different tasks. The use of residual structure combined with convolution modules can adapt to slight changes in signals, enhance the sampling ability of signals, and improve the reliability of modulation type recognition of underwater acoustic communication signals.

[0046] 2. The channel attention mechanism is used in the model. The channel attention mechanism can automatically obtain the importance of each feature channel, assign different weight coefficients to each channel, strengthen the important features obtained in the feature extraction process, and improve the model's ability to represent features.

[0047] 3. The model has fewer parameters and is relatively lightweight. The model takes the number of parameters and computational cost into consideration during the design process, uses fewer network layers for model design, and uses the residual structure to achieve connections between convolutional layers within the module. The model has less complexity and better adaptability in underwater scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of the overall process of the end-to-end modulation type identification method of underwater acoustic communication signals of the present invention;

[0049] Figure 2 It is a schematic diagram of the structure of the deep residual attention module RAC of the present invention;

[0050] Figure 3 This is the end-to-end modulation recognition model diagram of underwater acoustic communication signals based on residual attention convolution of the present invention. DETAILED DESCRIPTION

[0051] The present invention provides a method for identifying an end-to-end modulation type of an underwater acoustic communication signal, comprising:

[0052] Detecting the communication signal in the underwater acoustic signal to obtain the underwater acoustic communication signal;

[0053] The modulation type of the acquired underwater acoustic communication signal is identified by using the pre-established and trained end-to-end modulation type identification model of underwater acoustic communication signal based on residual attention convolution;

[0054] Among them, the trained end-to-end modulation type recognition model of underwater acoustic communication signals based on residual attention convolution includes: a feature extraction module and a classification module; among them,

[0055] The feature extraction module is used to extract features of the underwater acoustic communication signal based on residual attention;

[0056] The classification module is used to classify the extracted features to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type.

[0057] In an improved end-to-end modulation type recognition method for underwater acoustic communication signals, the feature extraction module includes: M pairs of RAC and DC modules consisting of M+1 residual attention convolution blocks RAC and M depth convolution blocks DC and 1 RAC module, which are connected in sequence; wherein the first pair of RAC and DC modules respectively extract features of the communication signal, and connect the output features through a residual structure, and the remaining pairs of RAC and DC modules are used to extract features from the output features; the 1 RAC module is used to continue to extract features from the features output by the last pair of RAC and DC modules.

[0058] In an improved end-to-end modulation type recognition method for underwater acoustic communication signals, each RAC specifically includes: a first one-dimensional convolutional layer Conv1, a first ReLU6 activation layer, a second one-dimensional convolutional layer Conv2, a one-dimensional convolutional layer Attention with a channel attention mechanism, a second ReLU6 activation layer and a maximum pooling layer MaxPooling connected in sequence;

[0059] The Conv1 is used to use y i =w i ·x+b i The formula extracts features from underwater acoustic communication signals, where x is the input data of the convolution layer, and w i and b i The output features y are i The corresponding convolution kernel and bias, i is the number of convolution layers;

[0060] The first ReLU6 activation layer is used to activate the extracted features to obtain activated features;

[0061] The Conv2 is used to continue to extract features from the activated features using the formula;

[0062] The Attention is used to reduce the dimension of the features extracted by Conv2 using global pooling, output a set of vectors with the same dimension as the number of channels, and then add two dense layers using RELU activation and Sigmoid activation respectively, using the formula To learn the weight of each channel, we get the attention weighted feature map, where p i is the feature output by Conv2, σ i is the channel weight, is the attention weighted feature map;

[0063] The second ReLU6 activation layer is used to activate the attention weighted features to obtain activated features;

[0064] The MaxPooling is used to downsample the output features to 1 / 1 of the original length using a pooling kernel of size l×1, thereby removing redundant information and reducing model parameters.

[0065] In an improved end-to-end modulation type identification method for underwater acoustic communication signals, each DC specifically includes: four ordinary one-dimensional convolutional layers and one ReLU6 activation layer connected in sequence;

[0066] The four common one-dimensional convolutional layers connected in sequence are used to continue extracting features from the underwater acoustic communication signal;

[0067] The ReLU6 activation layer is used to activate the input features using the formula ReLU6(x)=min(6,max(x,0)) to obtain the activated features, where x represents the input value of the activation function, min(·) represents the minimum value function, and max(·) represents the maximum value function.

[0068] In an improved method for end-to-end modulation type recognition of underwater acoustic communication signals, the classification module includes: a global average pooling layer, a fully connected layer and a Softmax activation layer connected in sequence;

[0069] The global average pooling layer is used to perform channel average downsampling on the extracted features;

[0070] The fully connected layer is used to generate a modulation type label of the underwater acoustic communication signal based on the downsampled features;

[0071] The Softmax activation layer is used to calculate the modulation type label of the underwater acoustic communication signal using an activation function to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type.

[0072] In an improved method for identifying end-to-end modulation types of underwater acoustic communication signals, the activation function is:

[0073]

[0074] Among them, Softmax(z f ) is the probability that the underwater acoustic communication signal belongs to each modulation type, z f is the modulation type label of the output of the fth node of the fully connected layer, k is the modulation type index of the underwater acoustic communication signal, and K is the total number of modulation types of the underwater acoustic communication signal.

[0075] In an improved method for identifying end-to-end modulation types of underwater acoustic communication signals, the end-to-end modulation type identification model of underwater acoustic communication signals is trained in the following manner:

[0076] Construct training dataset D train And the test dataset D test : Detect the communication signal in the acquired underwater acoustic signal to obtain N one-dimensional signals X1, X2, ..., X of the same length of K-type underwater acoustic communication signals N Samples and corresponding true label vectors Y1,Y2,...,Y N ; Randomly select N1 samples and corresponding label vectors to form the training data set D train , the remaining N-N1 samples and their corresponding label vectors constitute the test data set D test ;

[0077] Initialize the modulation type recognition model: The parameters of each network layer are Gaussian distributed with a mean of 0 and a variance of 1 (μ,σ 2 ) Random initialization;

[0078] Training the modulation type recognition model: Divide the training data set into batches according to the set batch size; Perform one round of training on all batches of data in the model to obtain the optimized model, and then use all batches of data to perform the next round of training on the optimized model until the number of training rounds reaches the set number of epochs;

[0079] The first round of training for the model includes: selecting a batch from the batch data, inputting the initialized end-to-end modulation type recognition model of underwater acoustic communication signals, combining the true label vector, using the cross entropy loss function to calculate the function loss value Loss of the model, using the error back propagation BP algorithm to update the modulation type recognition model parameters, and using the Adam optimizer with an adaptive learning rate to optimize the model parameters, repeatedly selecting a batch of data from the training set, and inputting the optimized modulation type recognition model until all batches of data are selected;

[0080] Other rounds of training for the model include: selecting a batch from the batch data, inputting the end-to-end modulation type recognition model of underwater acoustic communication signals obtained from the previous round of training, combining the true label vector, using the cross entropy loss function to calculate the function loss value Loss of the model, using the error back propagation BP algorithm to update the modulation type recognition model parameters, and using the Adam optimizer with an adaptive learning rate to optimize the model parameters, repeatedly selecting a batch of data from the training set, and inputting the optimized modulation type recognition model until all batches of data are selected.

[0081] In an improved method for end-to-end modulation type identification of underwater acoustic communication signals, the Loss calculation formula is:

[0082]

[0083] Among them, N1 represents the number of training set samples, K is the total number of underwater acoustic communication signal modulation types, k is the modulation type index, j is the sample index, q(x j,k ) represents the sample x j,k The predicted modulation type, p(x j,k ) represents the sample x j,k is the probability of the current predicted modulation type.

[0084] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0085] Example

[0086] like Figure 1 As shown, an embodiment of the present invention proposes an end-to-end modulation recognition method for underwater acoustic communication signals, and its implementation method includes the following steps:

[0087] Step 1: Process the acquired underwater acoustic signal, detect the communication signal in the underwater acoustic signal, and obtain different types of underwater acoustic communication signals. The specific implementation is as follows:

[0088] Step 1-1) performing communication signal detection on the acquired underwater acoustic signal, and extracting and saving the detected communication pulse signal;

[0089] Step 1-2) The extracted underwater acoustic communication signal is framed, and data is intercepted according to 4096 sampling points per frame, and two adjacent frames overlap 2048 sampling points, and the duration of each frame signal is about 0.2 seconds;

[0090] Step 1-3) Signal normalization: perform zero mean and normalization processing on each frame of data. Assuming the signal s(n), n=1,2,…,4096, the processed sample s'(n) is

[0091]

[0092] After processing in steps 1-4), a total of N samples of K-type underwater acoustic communication signals are obtained as the data set for subsequent experiments.

[0093] Step 2: Construct an end-to-end modulation recognition model for underwater acoustic communication signals based on residual attention convolution. The specific implementation is as follows:

[0094] like Figure 3 As shown, it includes modules such as input layer, feature extraction module, classification module and output layer. The input layer InputLayer preprocesses the underwater acoustic communication signal, inputs the preprocessed underwater acoustic communication signal into the feature extraction module (Feature Extraction) for feature extraction, obtains the deep abstract features of the signal, and then generates the signal modulation type label through the classification module (Classification), and finally outputs the signal modulation type label through the output layer Output Layer; the feature extraction module includes multiple residual attention convolutions (RAC) and multiple depth convolutions (DC), among which the first RAC and the first DC extract features of the communication signal respectively, and connect the output features through the residual structure, and extract the output features through the remaining RAC and DC in turn;

[0095] The processing of the remaining RAC and the remaining DC pair features is similar to the processing of the first RAC and the first DC pair features, which will not be repeated here.

[0096] In the feature extraction module, there are 4 RAC units, all convolution kernels are 5×1 in size, all convolution steps are 1, and all pooling kernels are 4×1 in size; there are 3 DC units in the depth convolution block, which consists of 4 standard convolution layers and 1 ReLU activation function layer, and is connected by multi-convolution layer jump connection. Feature downsampling is achieved through 4 convolution kernels with a step size of 2; the convolution layers are all activated using the function ReLU6(x), and the expression is:

[0097] ReLU6(x)=min(6,max(x,0))

[0098] Where x represents the input value of the activation function, min(·) represents the minimum value function, and max(·) represents the maximum value function;

[0099] The classification module includes a global average pooling layer GAP, a fully connected layer Dense and a Softmax activation layer. The GAP layer performs channel average downsampling on the features generated by the feature extraction module. The Dense layer generates a signal modulation type label, that is, the probability that the signal belongs to a certain category, and uses the Softmax activation function to activate the output value.

[0100] The end-to-end modulation recognition model processing process of underwater acoustic communication signals based on residual attention convolution specifically includes:

[0101] The input layer Input inputs the preprocessed underwater acoustic communication signal into the residual attention convolution block RAC and the depth convolution block DC in the feature extraction module;

[0102] The residual attention convolution block RAC and the depth convolution block DC extract features of the communication signal respectively, and connect the output features through the residual structure, and then pass the output features through the residual attention convolution block RAC and the depth convolution block DC for feature extraction in turn;

[0103] like Figure 2 As shown in the figure, the residual attention convolution block RAC specifically includes: 2 ordinary one-dimensional convolution layers Conv1 and Conv2, 2 ReLU6 activation layers, 1 MaxPooling layer, and 1 one-dimensional convolution layer Attention with channel attention mechanism;

[0104] The depth convolution block DC specifically includes: 4 ordinary one-dimensional convolution layers and 1 ReLU6 activation layer connected in sequence;

[0105] The feature extraction module inputs the extracted features into the global average pooling GAP layer of the classification module for downsampling and inputs them into the Dense layer;

[0106] The Dense layer generates the signal modulation type label, that is, the probability that the signal belongs to a certain category, and uses the Softmax activation function to activate the output value.

[0107] Step 2-3) In the residual attention convolution block RAC, the ordinary one-dimensional convolution layer is used to extract features from the underwater communication signal. The convolution layer operation is expressed as

[0108] y i =w i ·x+b i

[0109] In the above formula, x is the input data of the convolutional layer, w i and b i The output features y are i The corresponding convolution kernel and bias;

[0110] The MaxPooling layer uses a pooling kernel of size l×1 to downsample the features to 1 / l of the original length, removing redundant information and reducing model parameters;

[0111] In the one-dimensional convolutional layer Attention with channel attention mechanism, global pooling is first used to reduce the dimension of the features and output a set of vectors with the same dimension as the number of channels. Then two dense layers are added using ReLU activation and Sigmoid activation respectively to learn the weights of each channel. The process is expressed as

[0112]

[0113] In the formula, p i is the output feature map of the convolutional layer, σ i is the channel weight, is the attention weighted feature map.

[0114] Step 2-4) The feature extraction module inputs the extracted features into the global average pooling GAP layer of the classification module for downsampling, and inputs them into the Dense layer to convert the features into one dimension, and uses the Softmax activation function to classify the signal and output the result, that is, the probability that the signal belongs to a certain modulation type.

[0115] Step 3: Divide the acquired underwater acoustic communication signals into training sets and validation sets. The specific implementation is as follows:

[0116] Step 3-1) After step 1) is completed, a total of N one-dimensional signals X1, X2, ..., X with the same length are obtained. N And the corresponding label vectors Y1,Y2,...,Y N ;

[0117] Step 3-2) Randomly select N1 samples and corresponding label vectors from the data set in step 3-1) to form a training data set D train , the remaining N-N1 samples and their corresponding label vectors constitute the test data set D test .

[0118] Step 4: Use the underwater acoustic communication signal training set to iteratively train the end-to-end recognition model. The specific implementation is as follows:

[0119] Step 4-1) For the constructed end-to-end modulation recognition model of underwater acoustic communication signals based on residual attention convolution, the parameters of each network layer are subjected to a Gaussian distribution G(μ,σ) with a mean of 0 and a variance of 1. 2 ) Random initialization;

[0120] Step 4-2) Use the Adam optimizer or the Optimizers optimizer to update the network parameters so that the error between the output signal and the label converges to the optimal level;

[0121] Step 4-3) Calculate the function loss value Loss of the model through the cross entropy loss function. The calculation formula is:

[0122]

[0123] Among them, N1 represents the number of training set samples, K represents the total number of target modulation types, q(x j,k ) represents the sample x j,k The predicted modulation type, p(x j,k ) represents the sample x j,k is the probability of the current predicted modulation type.

[0124] Step 4-4) Use the Softmax classification function to convert the output of the last layer of the network, that is, the fully connected layer, into the probability Softmax (z i ), and the final recognition result is obtained. The calculation formula is:

[0125]

[0126] Among them, Softmax(z f ) is the probability that the underwater acoustic communication signal belongs to each modulation type, z f is the modulation type label of the output of the fth node of the fully connected layer, k is the modulation type index of the underwater acoustic communication signal, and K is the total number of modulation types of the underwater acoustic communication signal. After training, the trained modulation type recognition model can be compared with other different network models, and different evaluation indicators can be used to measure the final classification effect of each model including the recognition model provided by the embodiment of the present invention.

[0127] Step 5: Use the trained modulation type recognition model to identify the modulation type of the underwater acoustic communication signal.

[0128] Based on any of the above-mentioned methods for identifying the end-to-end modulation type of underwater acoustic communication signals, an embodiment of the present invention further provides a system for identifying the end-to-end modulation type of underwater acoustic communication signals, comprising:

[0129] An acquisition module, used for detecting the communication signal in the underwater acoustic signal to obtain the underwater acoustic communication signal;

[0130] The recognition module uses a pre-established and trained end-to-end modulation type recognition model for underwater acoustic communication signals based on residual attention convolution to recognize the modulation type of the acquired underwater acoustic communication signal; wherein the trained end-to-end modulation type recognition model for underwater acoustic communication signals based on residual attention convolution specifically includes: a feature extraction module and a classification module; the feature extraction module is used to extract features of the underwater acoustic communication signal based on residual attention; the classification module is used to classify the extracted features to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type.

[0131] Preferably, the feature extraction module includes: M pairs of RAC and DC modules consisting of M+1 residual attention convolution blocks RAC and M deep convolution blocks DC, and 1 RAC module, which are connected in sequence; wherein the first pair of RAC and DC modules respectively extract features from the communication signal, and connect the output features through a residual structure, and the remaining pairs of RAC and DC modules are used to extract features from the output features; the 1 RAC module is used to continue to extract features from the features output by the last pair of RAC and DC modules.

[0132] The following is a further description of the technical effects of the present invention in conjunction with simulation experiments:

[0133] Conditions and contents of simulation experiment:

[0134] The simulation experiment platform is Windows system, and the network is built based on Keras environment with TensorFlow background. The CPU of the experimental computer is i7-10750H, and the GPU is RTX2070. The experimental data are LFM, single-carrier chaotic frequency modulation, single-carrier chaotic phase modulation, multi-carrier OFDM, single-carrier PSK5 types of underwater acoustic communication signals, with a total of 10560 samples, which are used to assist in training the model.

[0135] In order to verify the performance of the proposed method, 80% of the samples were randomly selected as the training set, and the remaining 20% ​​of the samples were used as the test set. The network cost function selected the classification cross entropy, the optimizer was set to Adam, the batch size was set to 256, and the epochs was set to 100. In the iteration process, the error back propagation (BP) algorithm was used to update the model parameters, and the Adam optimizer with adaptive learning rate was used to optimize the model parameters. The initial learning rate of the Adam optimizer was 0.001. The exponential decay rate of the first-order moment estimate was set to 0.9, and the exponential decay rate of the second-order moment estimate was set to 0.999. After 100 classification experiments, the final model classification accuracy was 96.4%.

[0136] Table 1 Classification performance of different models / %

[0137]

[0138] 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 method for identifying end-to-end modulation type of underwater acoustic communication signals, comprising: Detecting the communication signal in the underwater acoustic signal to obtain the underwater acoustic communication signal; The modulation type of the acquired underwater acoustic communication signal is identified by using the pre-established and trained end-to-end modulation type identification model of underwater acoustic communication signal based on residual attention convolution; Among them, the trained end-to-end modulation type recognition model of underwater acoustic communication signals based on residual attention convolution includes: a feature extraction module and a classification module; among them, The feature extraction module is used to extract features of the underwater acoustic communication signal based on residual attention; The classification module is used to classify the extracted features to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type.

2. The method for identifying end-to-end modulation type of underwater acoustic communication signals according to claim 1, characterized in that: The feature extraction module includes: M pairs of RAC and DC modules consisting of M+1 residual attention convolution blocks RAC and M depth convolution blocks DC and 1 RAC module, which are connected in sequence; wherein the first pair of RAC and DC modules respectively extract features from the communication signal, and connect the output features through a residual structure, and the remaining pairs of RAC and DC modules are used to extract features from the output features; the 1 RAC module is used to continue to extract features from the features output by the last pair of RAC and DC modules.

3. The method for identifying end-to-end modulation type of underwater acoustic communication signals according to claim 2, characterized in that: Each RAC specifically includes: the first one-dimensional convolution layer Conv1, the first ReLU6 activation layer, the second one-dimensional convolution layer Conv2, the one-dimensional convolution layer Attention with a channel attention mechanism, the second ReLU6 activation layer and the maximum pooling layer MaxPooling connected in sequence; The Conv1 is used to use y i =w i ·x+b i The formula extracts features from underwater acoustic communication signals, where x is the input data of the convolutional layer, and w i and b i The output features y are i The corresponding convolution kernel and bias, i is the number of convolution layers; The first ReLU6 activation layer is used to activate the extracted features to obtain activated features; The Conv2 is used to continue to extract features from the activated features using the formula; The Attention is used to reduce the dimension of the features extracted by Conv2 using global pooling, output a set of vectors with the same dimension as the number of channels, and then add two dense layers using RELU activation and Sigmoid activation respectively, using the formula To learn the weight of each channel, we get the attention weighted feature map, where p i is the feature output by Conv2, σ i is the channel weight, is the attention weighted feature map; The second ReLU6 activation layer is used to activate the attention weighted features to obtain activated features; The MaxPooling is used to downsample the output features to 1 / 1 of the original length using a pooling kernel of size l×1, thereby removing redundant information and reducing model parameters.

4. The method for identifying end-to-end modulation type of underwater acoustic communication signals according to claim 2, characterized in that: Each DC specifically includes: 4 ordinary one-dimensional convolutional layers and 1 ReLU6 activation layer connected in sequence; The four common one-dimensional convolutional layers connected in sequence are used to continue extracting features from the underwater acoustic communication signal; The ReLU6 activation layer is used to activate the input features using the ReLU6(x)=min(6,max(x,0)) formula to obtain the activated features, where x represents the activation function input value, min(·) represents the minimum value function, and max(·) represents the maximum value function.

5. The method for identifying end-to-end modulation type of underwater acoustic communication signals according to claim 1, characterized in that: The classification module includes: a global average pooling layer, a fully connected layer and a Softmax activation layer connected in sequence; The global average pooling layer is used to perform channel average downsampling on the extracted features; The fully connected layer is used to generate a modulation type label of the underwater acoustic communication signal based on the downsampled features; The Softmax activation layer is used to calculate the modulation type label of the underwater acoustic communication signal using an activation function to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type.

6. The method for identifying end-to-end modulation type of underwater acoustic communication signals according to claim 5, characterized in that: The activation function is: Among them, Softmax(z f ) is the probability that the underwater acoustic communication signal belongs to each modulation type, z f is the modulation type label of the output of the fth node of the fully connected layer, k is the modulation type index of the underwater acoustic communication signal, and K is the total number of modulation types of the underwater acoustic communication signal.

7. The method for identifying end-to-end modulation type of underwater acoustic communication signals according to claim 1, characterized in that: The end-to-end modulation type recognition model of underwater acoustic communication signals is trained in the following manner: Construct training dataset D train And the test dataset D test : Detect the communication signal in the acquired underwater acoustic signal to obtain N one-dimensional signals X1, X2, ..., X of the same length of K-type underwater acoustic communication signals N Samples and corresponding true label vectors Y1,Y2,...,Y N ; Randomly select N1 samples and corresponding label vectors to form the training data set D train , the remaining N-N1 samples and their corresponding label vectors constitute the test data set D test ; Initialize the modulation type recognition model: The parameters of each network layer are Gaussian distributed with a mean of 0 and a variance of 1 (μ,σ 2 ) Random initialization; Training the modulation type recognition model: Divide the training data set into batches according to the set batch size; Perform one round of training on all batches of data in the model to obtain the optimized model, and then use all batches of data to perform the next round of training on the optimized model until the number of training rounds reaches the set number of epochs; The first round of training for the model includes: selecting a batch from the batch data, inputting the initialized end-to-end modulation type recognition model of underwater acoustic communication signals, combining the true label vector, using the cross entropy loss function to calculate the function loss value Loss of the model, using the error back propagation BP algorithm to update the modulation type recognition model parameters, and using the Adam optimizer with an adaptive learning rate to optimize the model parameters, repeatedly selecting a batch of data from the training set, and inputting the optimized modulation type recognition model until all batches of data are selected; Other rounds of training for the model include: selecting a batch from the batch data, inputting the end-to-end modulation type recognition model of underwater acoustic communication signals obtained from the previous round of training, combining the true label vector, using the cross entropy loss function to calculate the function loss value Loss of the model, using the error back propagation BP algorithm to update the modulation type recognition model parameters, and using the Adam optimizer with an adaptive learning rate to optimize the model parameters, repeatedly selecting a batch of data from the training set, and inputting the optimized modulation type recognition model until all batches of data are selected.

8. The method for identifying end-to-end modulation type of underwater acoustic communication signals according to claim 7, characterized in that: The calculation formula of Loss is: Among them, N1 represents the number of training set samples, K is the total number of underwater acoustic communication signal modulation types, k is the modulation type index, j is the sample index, q(x j,k ) represents the sample x j,k The predicted modulation type, p(x j,k ) represents the sample x j,k is the probability of the current predicted modulation type.

9. A modulation type identification system based on the end-to-end modulation type identification method of underwater acoustic communication signals according to claim 1, characterized in that: include: An acquisition module, used for detecting the communication signal in the underwater acoustic signal to obtain the underwater acoustic communication signal; The recognition module uses a pre-established and trained end-to-end modulation type recognition model for underwater acoustic communication signals based on residual attention convolution to recognize the modulation type of the acquired underwater acoustic communication signal; wherein the trained end-to-end modulation type recognition model for underwater acoustic communication signals based on residual attention convolution specifically includes: a feature extraction module and a classification module; the feature extraction module is used to extract features of the underwater acoustic communication signal based on residual attention; the classification module is used to classify the extracted features to obtain the probability that the underwater acoustic communication signal belongs to a certain modulation type.

10. The modulation type recognition system according to claim 9, characterized in that include: The feature extraction module includes: M pairs of RAC and DC modules consisting of M+1 residual attention convolution blocks RAC and M depth convolution blocks DC and 1 RAC module, which are connected in sequence; wherein the first pair of RAC and DC modules respectively extract features from the communication signal, and connect the output features through a residual structure, and the remaining pairs of RAC and DC modules are used to extract features from the output features; the 1 RAC module is used to continue to extract features from the features output by the last pair of RAC and DC modules.

Images