Modulation recognition method of non-cooperative underwater acoustic communication signals based on improved residual network

Abstract

The present invention discloses a non-cooperative underwater acoustic communication signal modulation recognition method based on an improved residual network, including: pre-processing the collected underwater acoustic communication signals to obtain complex baseband signals; constructing an underwater acoustic communication signal recognition neural network, extracting representation features of the complex baseband signals using feature extraction formulas and convolutional neural networks, extracting implicit features of the complex baseband signals using an improved residual module, fusion of the representation features and implicit features, and then inputting them into a fully connected layer and a classification layer; training the underwater acoustic communication signal recognition neural network; inputting the pre-processed underwater acoustic communication signal into the trained underwater acoustic communication signal recognition neural network, and finally outputting a classification result. The present invention adopts an improved residual module to improve the effect of feature extraction while ensuring system efficiency; splicing and fusion of the representation features and implicit features improves the comprehensiveness of the underwater acoustic communication signal features and the performance of the subsequent classification network.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic communication countermeasures, and in particular to a non-cooperative underwater acoustic communication signal modulation recognition method based on an improved residual network. Background Art

[0002] Hydroacoustic communication is capable of long-distance underwater communication tasks. With the continuous development of hydroacoustic communication in recent years, it has become the most important long-distance, high-speed underwater wireless communication means. Various military and commercial hydroacoustic communication equipment have been put into use one after another. From the perspective of marine information security and marine resource development, research on monitoring, identification, interference, and decryption of non-cooperative hydroacoustic communication has received attention.

[0003] Three types of methods have been developed for modulation recognition of underwater acoustic communication signals: those based on maximum likelihood function, those based on artificial feature extraction, and those based on deep learning. Among them, the method based on maximum likelihood function has high computational complexity and requires a lot of prior information to construct the likelihood function and select the appropriate decision threshold, which is difficult to implement in a non-cooperative environment. The classification system design of the method based on feature extraction is difficult and performs poorly under low signal-to-noise ratio. The method based on deep learning mainly sends two-dimensional images such as constellation diagrams, time-frequency diagrams, and cyclic spectrograms, or sends in-phase orthogonal symbols to the neural network for training. The method based on deep learning is currently a more cutting-edge recognition method, but constellation diagrams and in-phase orthogonal symbols are difficult to obtain under non-cooperative conditions, the complexity of generating time-frequency diagrams and cyclic spectrograms is high, and the resolution needs to be manually adjusted. At the same time, the relevant deep learning classification methods are mostly concentrated in the wireless field, and the targeted optimization of the serious attenuation, multipath interference, Doppler frequency shift and other problems in the field of underwater acoustic communication is insufficient. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes a non-cooperative underwater acoustic communication signal modulation recognition method based on an improved residual network, which extracts the characterization features of the signal, mines the implicit features of the signal, fuses the two types of features to generate a feature matrix, and uses the feature matrix as the classification input of the deep neural network, so as to comprehensively utilize the complete characteristics of the signal, reduce the impact of low signal-to-noise ratio and complex channels on the classification system, and improve the accuracy of modulation type recognition of underwater communication signals.

[0005] The specific technical solutions are as follows:

[0006] A non-cooperative underwater acoustic communication signal modulation recognition method based on an improved residual network is characterized by comprising the following steps:

[0007] S1: Collect underwater acoustic communication signal data and pre-process it as a training set: use the frequency domain estimation method and zero-crossing detection method to preliminarily estimate the carrier frequency of the underwater acoustic communication signal, use the frequency domain symmetry as the weight to assign the weight to the preliminarily estimated carrier frequency, and obtain the final result of the carrier frequency estimation; use the power spectrum density to estimate the bandwidth of the signal; down-mix the signal after the carrier frequency and bandwidth estimation is completed, convert the signal from the passband to the baseband, and then filter, standardize and normalize it to obtain the complex baseband signal;

[0008] S2: Construct and train the underwater acoustic communication signal recognition neural network with the training set, which is implemented through the following sub-steps:

[0009] (2.1) Characterization feature extraction: using a statistical feature extraction formula and a morphological feature extraction formula, artificial feature extraction is performed on the complex baseband signal and feature mining is performed using a convolutional neural network to obtain a characterization feature matrix;

[0010] At the same time, implicit feature extraction is performed: using an improved residual module to extract implicit features of the complex baseband signal to obtain an implicit feature matrix;

[0011] (2.2) Feature fusion and classification: The representation features and the latent features are fused to obtain a feature fusion matrix containing the representation features and the latent features, and the feature fusion matrix is ​​input into the fully connected layer and the classification layer;

[0012] S3: The underwater acoustic communication signal to be classified is preprocessed in the same manner as S1, and then input into the underwater acoustic communication signal recognition neural network trained in S2, and finally the recognition result is output.

[0013] Furthermore, in step S1, the method for estimating the carrier frequency and bandwidth of the underwater acoustic communication signal specifically includes the following sub-steps:

[0014] (1.1) Perform Fourier transform on the collected underwater acoustic communication signal s(i) of length N to obtain the signal spectrum S(k), and use a weighted method to estimate the center frequency of S(k) Take f0 as the initial value of the first carrier frequency estimate;

[0015] (1.2) Since the accuracy of the frequency domain estimation method is affected by the symmetry of the spectrum, the estimate of the symmetry of the spectrum is used as the weight of the initial value f0 of the first carrier frequency estimate W0 = 1-|P left -P right |, where The corresponding serial number of the carrier frequency in the frequency domain is f s is the sampling rate;

[0016] (1.3) Filter to obtain the signal strong interval centered on the estimated initial value f0 of the first carrier frequency, and use the zero-crossing detection method to obtain the second estimated value of the carrier frequency. Assign it an initial weight W1=1-W0; where M z is the number of zero-crossing points of signal s(i); y(i) represents the difference sequence of the number of zero-crossing points of the signal, y(i) = s(i+1)-s(i);

[0017] (1.4) The final carrier frequency estimate is f c =W0f0+W1f1;

[0018] (1.5) With carrier frequency f c The bandwidth is estimated by taking the distance between different frequency components of the underwater acoustic communication signal in the power spectrum and the center frequency as the center and performing weighted summation:

[0019]

[0020] In the formula, is the serial number corresponding to the carrier frequency in the frequency domain; F(k) is the power spectrum.

[0021] Furthermore, in the step (2.1), the characterization feature extraction is specifically implemented by the following sub-steps:

[0022] (2.1.1) Artificial feature extraction is performed on the complex baseband signal to obtain high-order moments (M 20 ,M 21 ,M 40 ,M 41 ,M 42 ,M 43 ,M 60 ,M 61 ,M 62 ,M 63 ), high-order cumulants (C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 61 ,C 62 ,C 63 ), 24 eigenvalues ​​including spectral correlation function, spectral symmetry estimation value near the carrier, maximum spectral power density, discrete Fourier maximum value of double amplitude, and standard deviation of instantaneous frequency absolute value, and the above eigenvalues ​​are combined into a 24×1 numerical matrix;

[0023] (2.1.2) The 24×1 numerical matrix obtained in step (2.1.1) is input into a convolutional neural network containing two convolutional layers and one dropout layer, and the matrix X composed of the characteristic values ​​representing the output signal is [x1, x2, ...x m ], where m is the number of convolution kernels in the last convolution layer.

[0024] Furthermore, in the step (2.1), the implicit feature extraction is specifically implemented through the following sub-steps:

[0025] (2.1.3) Divide the complex baseband signal obtained in step S1 into several segments 2 N ×2 data matrix, wherein the first row of the matrix is ​​the in-phase signal of the complex baseband signal, and the second row is the orthogonal signal of the complex baseband signal;

[0026] (2.1.4) The 2 generated in step (2.1.3) N The data matrix of size ×2×n is input into the (N-6) layer improved residual module, and then into the fully connected layer. The output matrix is ​​the matrix composed of hidden eigenvalues: Y = [y1, y2, ...y n ], where n is the number of dimensions of the fully connected layer;

[0027] Each improved residual module includes a convolution layer with a 1×1 convolution kernel added in front of the standard residual unit. The number of channels in the convolution layer is n. Through this convolution layer, 2 N The ×2 data matrix will be increased to 2 N ×2×n matrix; a 2×1 maximum pooling layer and a discarding layer are added after the standard residual unit; the convolution layer dimension-increasing module, the standard residual unit, the maximum pooling layer, and the discarding layer constitute an improved residual module.

[0028] Furthermore, the step (2.2) is specifically implemented by the following sub-steps:

[0029] (2.2.1) concatenating the representation eigenvalue matrix and the implicit eigenvalue matrix to complete the feature fusion of the representation features and the implicit features;

[0030] (2.2.2) Inputting the feature fusion matrix obtained in step (2.2.1) into a fully connected layer with a dimension of l, the number of identifiable modulation types is l;

[0031] (2.2.3) Input the matrix output from step (2.2.2) into the classification layer to generate the output matrix Z = [z1, z2, ... z l ], obtain the maximum value Z in the output matrix p , the sequence number p indicates that the classification result is the pth modulation type.

[0032] Furthermore, the classification layer is a Softmax classification layer.

[0033] The beneficial effects of the present invention are:

[0034] (1) The present invention combines artificially extracted representation features with implicit features that are difficult to extract artificially, thereby improving the comprehensiveness of signal features and the performance of subsequent classification networks.

[0035] (2) The improved residual module used in the present invention improves the feature extraction effect, reduces the system complexity, avoids overfitting, and improves the feature extraction effect while ensuring the system efficiency.

[0036] (3) Based on the artificial feature extraction of underwater acoustic communication signals, the present invention uses convolutional neural networks to conduct a deeper mining of features, thereby improving the effectiveness of artificial features. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the method of the present invention.

[0038] Figure 2 It is a flow chart of the frequency carrier and bandwidth estimation of the present invention.

[0039] Figure 3 This is a schematic diagram of the standard residual unit structure.

[0040] Figure 4 Schematic diagram of the improved residual module structure of the present invention.

[0041] Figure 5 It is a confusion diagram of the classification results when the signal-to-noise ratio of the test set of the present invention is 2dB.

[0042] Figure 6 It is a confusion diagram of the classification results when the signal-to-noise ratio of the test set of the present invention is 20dB. DETAILED DESCRIPTION

[0043] The present invention will be described in detail below according to the accompanying drawings and preferred embodiments, and the purpose and effect of the present invention will become more clear. The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] like Figure 1 As shown, the non-cooperative underwater acoustic communication signal modulation recognition method based on the improved residual network includes the following steps:

[0045] S1: Collect underwater acoustic communication signal data, preprocess it, and randomly divide it into two groups: training set and test set.

[0046] The steps for preprocessing the collected underwater acoustic communication signals are as follows:

[0047] (1.1) Carrier frequency and bandwidth estimation: Figure 2 As shown, the frequency domain estimation method and the zero-crossing detection method are used to make an initial estimate of the carrier frequency of the underwater acoustic communication signal, and the frequency domain symmetry is used as the weight to allocate the weight of the initially estimated carrier frequency to obtain the final result of the carrier frequency estimation, and the power spectrum density is used to estimate the bandwidth of the signal. This is specifically achieved through the following sub-steps:

[0048] (1.1.1) Perform Fourier transform on the collected underwater acoustic communication signal s(i) of length N to obtain the signal spectrum S(k). Use a weighted method to estimate the center frequency f0 of S(k), and use f0 as the initial value of the first carrier frequency estimate. The expression of f0 is:

[0049]

[0050] (1.1.2) Since the accuracy of the frequency domain estimation method is affected by the symmetry of the spectrum, the estimation of the symmetry of the spectrum is used as the weight of the initial value f0 of the first carrier frequency estimate W0 = 1-|P left -P right |, where

[0051] The corresponding serial number of the carrier frequency in the frequency domain is where f s is the sampling rate.

[0052] (1.1.3) Filtering, obtain the signal strong interval centered on the estimated initial value f0 of the first carrier frequency, use the zero-crossing detection method to obtain the second estimated value f1 of the carrier frequency, assign it an initial weight W1=1-W0, and the expression of f1 is:

[0053]

[0054] Where M z is the number of zero-crossing points of signal s(i); y(i) represents the difference sequence of the number of zero-crossing points of the signal, y(i)=s(i+1)-s(i).

[0055] (1.1.4) The final carrier frequency estimate is f c =W0f0+W1f1.

[0056] (1.1.5) With carrier frequency f c The bandwidth is estimated by taking the distance between different frequency components of the underwater acoustic communication signal in the power spectrum and the center frequency as the center and performing weighted summation:

[0057]

[0058] In the formula, is the serial number corresponding to the carrier frequency in the frequency domain; F(k) is the power spectrum.

[0059] (1.2) The signal with completed carrier frequency and bandwidth estimation is down-mixed, the signal is converted from passband to baseband, and then filtered to eliminate noise outside the baseband. Finally, it is standardized and normalized to obtain a complex baseband signal.

[0060] S2: Construct and train the underwater acoustic communication signal recognition neural network, send the training set to the underwater acoustic communication signal recognition neural network for training, and the test set is used to evaluate the performance of the underwater acoustic communication signal recognition neural network. This is achieved through the following sub-steps:

[0061] (2.1) The neural network for underwater acoustic communication signal recognition includes two parts: representation feature extraction and latent feature extraction.

[0062] Characterization feature extraction uses the statistical feature extraction formula and the morphological feature extraction formula to extract artificial features of the complex baseband signal, and obtains high-order moments (M 20 ,M 21 ,M 40 ,M 41 ,M 42 ,M 43 ,M 60 ,M 61 ,M 62 ,M 63 ), high-order cumulants (C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 61 ,C 62 ,C 63 ), spectral correlation function, spectral symmetry estimation value near the carrier, maximum spectral power density, maximum discrete Fourier transform of double amplitude, standard deviation of instantaneous frequency absolute value, and the above eigenvalues ​​are combined into a 24×1 numerical matrix. The numerical matrix is ​​input into a convolutional neural network with two convolutional layers and one discarding layer for feature mining, and the matrix composed of the characteristic values ​​of the output signal is X=[x1,x2,...x m ], where m is the number of convolution kernels in the last convolution layer.

[0063] Implicit feature extraction uses an improved residual module to extract implicit features from complex baseband signals and obtain an implicit feature matrix. The specific steps are:

[0064] (2.1.1) Divide the complex baseband signal obtained in step S1 into several segments 2N ×2 size data matrix, where the first row of the matrix is ​​the in-phase signal of the complex baseband signal, and the second row is the orthogonal signal of the complex baseband signal.

[0065] (2.1.2) The 2 generated in step (2.1.1) N The data matrix of size ×2×n is input into (N-6) improved residual modules, and then input into the fully connected layer. The output matrix is ​​the matrix composed of hidden eigenvalues: Y = [y1, y2, ...y n ], where n is the dimension of the fully connected layer.

[0066] like Figure 4 As shown in, each improved residual module includes a Figure 3 The standard residual unit shown in the figure adds a convolution layer with a convolution kernel of 1×1 in front of the standard residual unit. The number of channels in the convolution layer is n. Through this convolution layer, 2 N The ×2 data matrix will be increased to 2 N ×2×n matrix; adding a dimensionality-raising structure improves the feature extraction effect. Adding a 2×1 maximum pooling layer and a dropout layer after the standard residual unit reduces system complexity and avoids overfitting. The overall structure of the convolution layer dimensionality-raising structure, standard residual unit, maximum pooling layer, and dropout layer is an improved residual module. The improved residual module improves the feature extraction effect while ensuring system efficiency.

[0067] Figure 5 and Figure 6 The following figure shows the confusion map of the classification results using the test set. The ordinate of the figure is the actual modulation type of the signal, and the abscissa is the predicted modulation type of the improved residual network. There are 7 squares on the abscissa and 7 squares on the ordinate in the figure, with a total of 49 squares. The darker the color of the square, the greater the probability of classification into that situation. The classification is correct only when the abscissa and the ordinate are the same modulation mode. The darker the color of the square on the diagonal from the upper left to the lower right in the figure, the better the classification result.

[0068] (2.2) Feature fusion and classification: The representation features and the implicit features are fused, that is, the representation eigenvalue matrix and the implicit eigenvalue matrix are concatenated to obtain a feature fusion matrix containing representation and implicit information. The feature fusion matrix is ​​input into a fully connected layer with a dimension of l, and the number of modulation types that can be identified is l. The matrix output by the fully connected layer is input into the Softmax classification layer to generate the output matrix Z = [z1, z2, … z l ], get the maximum value Z in the output matrix p , the serial number p indicates that the classification result is the pth modulation type, realizing the recognition of underwater acoustic communication signals.

[0069] S3: The underwater acoustic communication signal to be classified is preprocessed in the same way as S1, and then input into the underwater acoustic communication signal recognition neural network trained in S2, and finally the recognition result is output.

[0070] The present invention is specifically described below by way of examples:

[0071] S1: Collect underwater acoustic communication signal data, add Gaussian white noise to obtain a data set with a signal-to-noise ratio SNR = -10dB ~ 20dB, pre-process it, and randomly divide it into two groups: training set and test set: the data set contains BPSK, QPSK, 8PSK, 16QAM, OFDM, 2FSK, 4FSK, a total of 7 modulation modes, and the modulation mode categories are assigned serial numbers 1 to 7. . After carrier frequency estimation and bandwidth estimation, the underwater acoustic communication signal is down-mixed, filtered and normalized to obtain a complex baseband signal, which is divided into several data matrices of size 2048×2.

[0072] S2: Build and train a neural network for underwater acoustic communication signal recognition:

[0073] (2.1) The training set is sent to the underwater acoustic communication signal recognition neural network for training, and the complex baseband signal of size 2048×2 is artificially extracted to obtain the high-order moments (M 20 ,M 21 ,M 40 ,M 41 ,M 42 ,M 43 ,M 60 ,M 61 ,M 62 ,M 63 ), high-order cumulants (C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 61 ,C 62 ,C 63 ), 24 eigenvalues ​​including spectral correlation function, spectral symmetry estimation value near the carrier, maximum spectral power density, discrete Fourier maximum of double amplitude, and standard deviation of instantaneous frequency absolute value are combined into a 24×1 numerical matrix, which is normalized and sent to a convolutional neural network with 64 passbands and a convolution kernel size of 3×1, and outputs a 64×1 numerical matrix.

[0074] At the same time, the complex baseband signal of size 2048×2 is input into the improved residual module. The improved residual module includes: a dimensionality-enhancing module with a convolution kernel of 1×1 and 8 channels, a standard residual unit with a convolution kernel size of 3×2, a maximum pooling layer with a core size of 2×1, and a discarding layer. After the first layer of the improved residual module, the data matrix size is 1024×2×8, the second layer output size is 512×2×8, and so on. The output size of the fifth layer is 64×2×8, and the output of the fully connected layer is a 128×1 numerical matrix.

[0075] (2.2) Concatenate the two numerical matrices generated in step (2.1) to obtain a 192×1 numerical matrix, and then pass it through the fully connected layer to obtain a 7×1 feature combination matrix. Send the feature combination matrix to the Softmax classification layer to generate the output matrix Z = [z1, z2, ... z7], and find the maximum value Z in the output matrix p , the sequence number p indicates that the classification result is the pth modulation type.

[0076] S3: Input the test set into the underwater acoustic communication signal recognition neural network trained in S2, and finally output the recognition result. The detailed classification result is output as a confusion map, such as Figure 5 and Figure 6 shown.

[0077] As can be seen from the figure, in this example, when the signal-to-noise ratio of the test set is low (2dB), the overall classification accuracy is 87.12%; when the signal-to-noise ratio of the test set is high (20dB), the overall classification accuracy is 91.47%, which completes the classification task well.

[0078] The present invention performs targeted preprocessing on underwater acoustic communication signals based on common modulation types and underwater acoustic features, and uses a combination of signal feature extraction and neural networks to complete the recognition of underwater acoustic communication signal modulation types.

[0079] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the invention.

Claims

1. A non-cooperative underwater acoustic communication signal modulation recognition method based on an improved residual network, characterized in that: The following steps are involved: S1: Collect and pre-process underwater acoustic communication signal data as a training set: use frequency domain estimation and zero-crossing detection methods to preliminarily estimate the carrier frequency of the underwater acoustic communication signal, use frequency domain symmetry as a weight to assign weights to the preliminarily estimated carrier frequency, and obtain the final result of carrier frequency estimation; Use power spectral density to estimate the bandwidth of the signal; Down-mixing the signal after carrier frequency and bandwidth estimation, converting the signal from passband to baseband, and then filtering, standardizing and normalizing it to obtain a complex baseband signal; S2: Construct and train the underwater acoustic communication signal recognition neural network with the training set, which is implemented through the following sub-steps: (2.1) Characterization feature extraction: using a statistical feature extraction formula and a morphological feature extraction formula, artificial feature extraction is performed on the complex baseband signal and feature mining is performed using a convolutional neural network to obtain a characterization feature matrix; At the same time, implicit feature extraction is performed: using an improved residual module to extract implicit features of the complex baseband signal to obtain an implicit feature matrix; (2.2) Feature fusion and classification: The representation features and the latent features are fused to obtain a feature fusion matrix containing the representation features and the latent features, and the feature fusion matrix is ​​input into the fully connected layer and the classification layer; S3: pre-processing the underwater acoustic communication signal to be classified in the same way as S1, and then inputting it into the underwater acoustic communication signal recognition neural network trained in S2, and finally outputting the recognition result; In step S1, the method for estimating the carrier frequency and bandwidth of the underwater acoustic communication signal specifically includes the following sub-steps: (1.1) Perform Fourier transform on the collected underwater acoustic communication signal s(i) of length N to obtain the signal spectrum S(k), and use a weighted method to estimate the center frequency of S(k) Take f0 as the initial value of the first carrier frequency estimate; (1.2) Since the accuracy of the frequency domain estimation method is affected by the symmetry of the spectrum, the estimate of the symmetry of the spectrum is used as the weight of the initial value f0 of the first carrier frequency estimate W0 = 1-|P left -P right |, where The corresponding serial number of the carrier frequency in the frequency domain is f s is the sampling rate; (1.3) Filter to obtain the signal strong interval centered on the estimated initial value f0 of the first carrier frequency, and use the zero-crossing detection method to obtain the second estimated value of the carrier frequency. Assign it an initial weight W1=1-W0; where M z is the number of zero-crossing points of signal s(i); y(i) represents the difference sequence of the number of zero-crossing points of the signal, y(i) = s(i+1)-s(i); (1.4) The final carrier frequency estimate is f c =W0f0+W1f1; (1.5) With carrier frequency f c The bandwidth is estimated by taking the distance between different frequency components of the underwater acoustic communication signal in the power spectrum and the center frequency as the center and performing weighted summation: In the formula, is the serial number corresponding to the carrier frequency in the frequency domain; F(k) is the power spectrum; In the step (2.1), the characterization feature extraction is specifically implemented through the following sub-steps: (2.1.1) Artificial feature extraction is performed on the complex baseband signal to obtain high-order moments (M 20 ,M 21 ,M 40 ,M 41 ,M 42 ,M 43 ,M 60 ,M 61 ,M 62 ,M 63 ), high-order cumulants (C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 61 ,C 62 ,C 63 ), 24 eigenvalues ​​including spectral correlation function, spectral symmetry estimation value near the carrier, maximum spectral power density, discrete Fourier maximum value of double amplitude, and standard deviation of instantaneous frequency absolute value, and the above eigenvalues ​​are combined into a 24×1 numerical matrix; (2.1.2) The 24×1 numerical matrix obtained in step (2.1.1) is input into a convolutional neural network containing two convolutional layers and one dropout layer, and the matrix X composed of the characteristic values ​​representing the output signal is [x1, x2, ...x m ], where m is the number of convolution kernels in the last convolution layer; In the step (2.1), the implicit feature extraction is specifically implemented through the following sub-steps: (2.1.3) Divide the complex baseband signal obtained in step S1 into several segments 2 N ×2 data matrix, wherein the first row of the matrix is ​​the in-phase signal of the complex baseband signal, and the second row is the orthogonal signal of the complex baseband signal; (2.1.4) The 2 generated in step (2.1.3) N The data matrix of size ×2×n is input into the (N-6) layer improved residual module, and then into the fully connected layer. The output matrix is ​​the matrix composed of hidden eigenvalues: Y = [y1, y2, ...y n ], where n is the number of dimensions of the fully connected layer; Each improved residual module includes a convolution layer with a 1×1 convolution kernel added in front of the standard residual unit. The number of channels in the convolution layer is n. Through this convolution layer, 2 N The ×2 data matrix will be increased to 2 N ×2×n matrix; adding a 2×1 maximum pooling layer and a discarding layer after the standard residual unit; the convolutional layer dimension-increasing module, the standard residual unit, the maximum pooling layer, and the discarding layer constitute an improved residual module; The step (2.2) is specifically implemented by the following sub-steps: (2.2.1) concatenating the representation eigenvalue matrix and the implicit eigenvalue matrix to complete the feature fusion of the representation features and the implicit features; (2.2.2) Inputting the feature fusion matrix obtained in step (2.2.1) into a fully connected layer with a dimension of l, the number of identifiable modulation types is l; (2.2.3) Input the matrix output from step (2.2.2) into the classification layer to generate an output matrix Z = [z1, z2, ... z l ], obtain the maximum value Z in the output matrix p , the sequence number p indicates that the classification result is the pth modulation type.

2. The non-cooperative underwater acoustic communication signal modulation recognition method based on improved residual network according to claim 1 is characterized in that: The classification layer is a Softmax classification layer.

Images