A multi-domain feature fusion underwater acoustic communication modulation mode autonomous discrimination method

Abstract

This application belongs to the field of underwater acoustic communication signal and digital signal processing technology. This application provides a method for autonomously identifying underwater acoustic communication modulation modes through multi-domain feature fusion. The embodiments of this disclosure preprocess the data, extracting the orthogonal and in-phase components of the communication signal, and inputting them into a time-domain network branch. The frequency-domain waveform obtained after Fourier transform of the signal is input into a frequency-domain network branch. The time-frequency map obtained by continuous wavelet transform of the signal is input into a time-frequency map network branch. Correlation is calculated between the time-domain network branch, the frequency-domain network branch, and the continuous wavelet transform time-frequency map network branch, and the features obtained from the network branches are weighted and fused together. After feature fusion, the modulation modes of the underwater acoustic communication signal are classified through a fully connected layer and a softmax layer. By using the feature information from the time domain, frequency domain, and wavelet packet time-frequency map as input to the neural network, the feature results are flexibly fused, improving the robustness of autonomous identification of underwater acoustic communication signal modulation modes.

Description

Technical Field

[0001] This disclosure relates to the fields of underwater acoustic communication signals and digital signal processing technology, and in particular to a method for autonomously identifying underwater acoustic communication modulation modes by fusing multi-domain features. Background Technology

[0002] With the rapid development of fields such as adaptive modulation and coding in the ocean, non-cooperative communication, communication countermeasures, and military intelligence gathering and reconnaissance, the importance of autonomous modulation pattern discrimination in underwater acoustic communication is becoming increasingly prominent. In the military field, autonomous modulation pattern discrimination in underwater acoustic communication can efficiently collect and analyze enemy communication signals in complex underwater environments, providing crucial intelligence support and thus offering important basis for military decision-making. Simultaneously, by quickly and accurately identifying the modulation mode of enemy signals, effective jamming and anti-jamming strategies can be formulated, improving the effectiveness of communication countermeasures. In the civilian field, adaptive modulation and coding technology dynamically adjusts parameters such as modulation order and modulation mode by sensing changes in communication conditions, thereby improving bandwidth utilization while ensuring the reliability of the communication system. At the receiving end, autonomous discrimination of underwater acoustic communication modulation patterns is required to improve the reliability and stability of communication nodes. Therefore, there is an urgent need for an autonomous modulation pattern discrimination method for underwater acoustic communication that can quickly and accurately estimate the modulation mode of received underwater acoustic signals.

[0003] Traditional methods for autonomously identifying modulation patterns in underwater acoustic communication include statistical classifiers, support vector machines, and decision trees. While these traditional algorithms have achieved some success in identifying modulation patterns, they are highly dependent on feature selection and extraction. For unknown or complex signal features, their identification performance is poor, limiting their application to restricted environments. When facing complex and ever-changing underwater acoustic environments and increasingly diverse modulation methods, their limitations in performance and robustness become increasingly apparent. Deep learning, with its powerful autonomous feature extraction capabilities, excellent generalization ability, and ability to handle high-dimensional features and nonlinear relationships, has already seen some applications in autonomously identifying modulation patterns in underwater acoustic communication. However, most existing deep learning-based identification methods only use single-domain information (such as time-domain, frequency-domain, or time-frequency-domain information) of the signal waveform for feature representation. When processing complex underwater acoustic communication signals, they ignore the complementarity of feature representations across multiple domains and fail to consider the specific expressions of different communication modes in specific domains. This results in an inability to fully capture all signal features, thus affecting identification performance and robustness.

[0004] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.

[0005] It should be noted that this section is intended to provide background or context for the technical solutions of this disclosure as set forth in the claims. The description herein does not constitute an admission that it is prior art simply because it is included in this section. Summary of the Invention

[0006] The purpose of this disclosure is to provide a method for autonomously identifying modulation modes in underwater acoustic communication by fusing multi-domain features, thereby overcoming, to at least some extent, one or more problems caused by the limitations and defects of related technologies.

[0007] According to an embodiment of this disclosure, a method for autonomously identifying modulation modes in underwater acoustic communication by multi-domain feature fusion is provided, the method comprising:

[0008] Construct a dataset of underwater acoustic communication modulation modes and preprocess the dataset.

[0009] The preprocessed underwater acoustic communication modulation mode dataset is divided into a training set and a test set according to a preset ratio;

[0010] A multi-domain fusion communication pattern recognition network is constructed; wherein the multi-domain fusion communication pattern recognition network includes a time-domain branch network, a frequency-domain branch network, a time-frequency graph branch network, and a feature fusion network, wherein the time-domain branch network, the frequency-domain branch network, and the time-frequency graph branch network run in parallel and are all connected to the feature fusion network;

[0011] The multi-domain fusion communication pattern recognition network is trained using the training set and the test set respectively to obtain the trained multi-domain fusion communication pattern recognition network;

[0012] The data to be judged is input into the trained multi-domain fusion communication pattern recognition network to obtain the judgment result.

[0013] Furthermore, the step of constructing an underwater acoustic communication modulation mode dataset and preprocessing the underwater acoustic communication modulation mode dataset includes:

[0014] Based on the measured signals received from the underwater acoustic environment, the simulated signals generated by simulation, and the underwater acoustic dataset in the database, an underwater acoustic communication modulation mode dataset is constructed.

[0015] The measured signal and the simulated signal are demodulated to baseband, the quadrature components and in-phase components of the measured signal and the simulated signal are extracted, baseband signals of the same length are truncated, and the measured signal and the simulated signal are classified and labeled according to the signal-to-noise ratio and the type of communication signal.

[0016] All data samples in the underwater acoustic communication modulation mode dataset Normalization process is performed to obtain and each data sample Each element is moved to the range [0,1]; where the normalization expression is:

[0017]

[0018] In the formula, This is the normalized dataset of underwater acoustic communication modulation modes. The maximum amplitude in the original dataset. This represents the minimum amplitude value in the original dataset.

[0019] For each of the data samples Perform a 128-point Fourier transform to obtain a frequency domain signal of the same length, which is then used as the input to the frequency domain branch network; the expression for the Fourier transform is:

[0020]

[0021] In the formula, The nth element of a normalized single underwater acoustic communication data set. Let N be the k-th element of the frequency domain signal vector, N be the number of Fourier transform points, n be the coordinates of the time domain signal element, k be the coordinates of the frequency domain signal element, and j be an imaginary number.

[0022] For each of the data samples Perform continuous wavelet transform and plot the wavelet transform time-frequency graph, which serves as the input to the time-frequency graph branch network; wherein, the expression for the wavelet transform is:

[0023]

[0024] In the formula, The nth element of a normalized single underwater acoustic communication data set. Let be a wavelet function, 'a' be the 'a'-th row of the wavelet-transformed matrix, and 'b' be the 'b'-th column of the wavelet-transformed matrix. c o It is a constant. d o It is a constant.

[0025] Furthermore, the step of dividing the preprocessed underwater acoustic communication modulation mode dataset into a training set and a test set according to a preset ratio includes:

[0026] All data samples in the preprocessed underwater acoustic communication modulation mode dataset are randomly shuffled, and the training set and the test set are randomly selected according to a preset ratio.

[0027] Furthermore, the time-domain branching network includes:

[0028] The system consists of a temporal input layer, a temporal convolutional layer 1, a temporal pooling layer 1, a temporal convolutional layer 2, a temporal pooling layer 2, a temporal convolutional layer 3, a temporal pooling layer 3, a temporal convolutional layer 4, a temporal pooling layer 4, and a temporal output layer; among which,

[0029] The temporal input layer of the temporal branch network has a matrix size of 2×128 as input. The temporal convolutional layer 1 has a size of 2×128×256, the temporal pooling layer 1 has a size of 2×64×256, the temporal convolutional layer 2 has a size of 2×32×128, the temporal pooling layer 2 has a size of 2×32×128, the temporal convolutional layer 3 has a size of 2×32×64, the temporal pooling layer 3 has a size of 2×16×64, the temporal convolutional layer 4 has a size of 2×16×64, the temporal pooling layer 4 has a size of 2×8×32, and the temporal output layer outputs 128 temporal features.

[0030] Furthermore, the frequency domain branching network includes:

[0031] The system consists of a frequency domain input layer, a frequency domain convolutional layer 1, a frequency domain pooling layer 1, a frequency domain convolutional layer 1, a frequency domain pooling layer 2, a frequency domain convolutional layer 3, a frequency domain pooling layer 3, a frequency domain convolutional layer 4, a frequency domain pooling layer 4, and a frequency domain output layer; among which,

[0032] The frequency domain input layer of the frequency domain branch network has a matrix size of 1×128 as input. The frequency domain convolutional layer 1 has a size of 1×128×256, the frequency domain pooling layer 1 has a size of 1×64×256, the frequency domain convolutional layer 1 has a size of 1×32×128, the frequency domain pooling layer 2 has a size of 1×32×128, the frequency domain convolutional layer 3 has a size of 1×32×64, the frequency domain pooling layer 3 has a size of 1×16×64, the frequency domain convolutional layer 4 has a size of 1×16×64, the frequency domain pooling layer 4 has a size of 1×8×32, and the output layer outputs 64 frequency domain features.

[0033] Furthermore, the time-frequency graph branch network includes:

[0034] The time-frequency graph consists of an input layer, a convolutional layer 1, a densely connected block 1, a pooling layer 1, a densely connected block 2, a convolutional layer 2, a pooling layer 2, a convolutional layer 3, a pooling layer 4, and an output layer; where,

[0035] The time-frequency graph input layer of the time-frequency graph branch network has a size of 128×128×3. The output of time-frequency graph convolutional layer 1 is 64×64×144. The output of dense connection block 1 is 64×64×144. The output of time-frequency graph pooling layer 1 is 32×32×72. The output of dense connection block 2 is 32×32×144. The output of time-frequency graph convolutional layer 2 is 16×16×72. The output of time-frequency graph pooling layer 2 is 16×16×36. The output of time-frequency graph convolutional layer 3 is 16×16×36. The output of time-frequency graph pooling layer 3 is 8×8×18. The time-frequency graph output layer outputs 64 time-frequency graph features.

[0036] Furthermore, the training steps of the multi-domain fusion communication pattern recognition network include:

[0037] The time-domain signal in the data sample is input into the time-domain branch network to obtain the time-domain features;

[0038] The frequency domain signal in the data sample is input into the frequency domain branch network to obtain the frequency domain features;

[0039] The time-domain graph in the data sample is input into the time-domain graph branch network to obtain the time-domain graph features;

[0040] Calculate the weight values ​​of the time-domain feature, the frequency-domain feature, and the time-domain graph feature respectively;

[0041] Based on the weight values ​​of each feature, the feature descriptors of the time-domain feature, the frequency-domain feature, and the time-frequency map feature are weighted and serially concatenated to obtain the fused feature;

[0042] The fused features are then used to classify all modulation modes through a fully connected layer and a softmax layer.

[0043] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects:

[0044] In the embodiments of this disclosure, the above-mentioned multi-domain feature fusion method for autonomous discrimination of underwater acoustic communication modulation modes involves, on the one hand, data preprocessing to extract the orthogonal and in-phase components of the communication signal, which are then input into the time-domain network branch. The frequency-domain waveform obtained after Fourier transform of the signal is input into the frequency-domain network branch. The time-frequency map obtained by continuous wavelet transform of the signal is input into the two-dimensional time-frequency map network branch. Correlation calculations are performed on the time-domain network branch, the frequency-domain network branch, and the continuous wavelet transform time-frequency map network branch, and the features obtained from the network branches are weighted and fused together. After feature fusion, the modulation modes of the underwater acoustic communication signal are classified through a fully connected layer and a softmax layer. On the other hand, by combining the feature information from the time domain, frequency domain, and wavelet packet time-frequency map, multiple deep learning networks are used to comprehensively extract features from the signal, ultimately fusing multi-domain data. This fully utilizes the complementarity between different domains and different transform features, filling the limitations of extracting high-dimensional features from single-domain data, and more comprehensively capturing the complex characteristics of the signal, thereby achieving higher-performance modulation recognition. Attached Figure Description

[0045] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0046] Figure 1 This diagram illustrates the steps of an autonomous discrimination method for underwater acoustic communication modulation modes based on multi-domain feature fusion in an exemplary embodiment of this disclosure.

[0047] Figure 2 This diagram illustrates a flowchart of the simulation signal generation process in an exemplary embodiment of this disclosure.

[0048] Figure 3 This diagram illustrates a specific flowchart of the autonomous discrimination method for underwater acoustic communication modulation modes using multi-domain feature fusion in an exemplary embodiment of this disclosure.

[0049] Figure 4 This diagram illustrates the structure of the convolutional and fully connected layers of a multi-domain fusion communication pattern recognition network in an exemplary embodiment of this disclosure.

[0050] Figure 5 This diagram illustrates the accuracy of multi-domain feature fusion communication modulation pattern recognition in an exemplary embodiment of this disclosure.

[0051] Figure 6 This diagram illustrates the average recognition accuracy of multi-domain feature fusion communication modulation modes in an exemplary embodiment of this disclosure.

[0052] Figure 7 This illustrates the overall confusion matrix of the convolutional network in an exemplary embodiment of this disclosure;

[0053] Figure 8 The confusion matrix at different signal-to-noise ratios is shown in the exemplary embodiments of this disclosure. Detailed Implementation

[0054] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0055] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0056] This example implementation provides a method for autonomously determining modulation patterns in underwater acoustic communication through multi-domain feature fusion. (Reference) Figure 1 As shown, the multi-domain feature fusion underwater acoustic communication modulation mode autonomous discrimination method may include steps S101 to S105.

[0057] Step S101: Construct an underwater acoustic communication modulation mode dataset and preprocess the underwater acoustic communication modulation mode dataset;

[0058] Step S102: Divide the preprocessed underwater acoustic communication modulation mode dataset into a training set and a test set according to a preset ratio;

[0059] Step S103: Construct a multi-domain fusion communication pattern recognition network; wherein the multi-domain fusion communication pattern recognition network includes a time-domain branch network, a frequency-domain branch network, a time-frequency graph branch network, and a feature fusion network, wherein the time-domain branch network, the frequency-domain branch network, and the time-frequency graph branch network run in parallel and are all connected to the feature fusion network;

[0060] Step S104: Train the multi-domain fusion communication pattern recognition network using the training set and the test set respectively to obtain the trained multi-domain fusion communication pattern recognition network;

[0061] Step S105: Input the data to be judged into the trained multi-domain fusion communication pattern recognition network to obtain the judgment result.

[0062] The above-described multi-domain feature fusion method for autonomous discrimination of underwater acoustic communication modulation modes involves two main steps. First, data preprocessing extracts the orthogonal and in-phase components of the communication signal, which are then input into the time-domain network branch. The frequency-domain waveform obtained after Fourier transform is input into the frequency-domain network branch. The time-frequency map obtained by continuous wavelet transform is input into the two-dimensional time-frequency map network branch. Correlation calculations are performed on the time-domain, frequency-domain, and continuous wavelet transform time-frequency map network branches, and the features obtained from the network branches are then weighted and fused together. After feature fusion, the modulation modes of the underwater acoustic communication signal are classified using fully connected layers and softmax layers. Second, by combining the feature information from the time domain, frequency domain, and wavelet packet time-frequency map, multiple deep learning networks are used to comprehensively extract features from the signal. Finally, multi-domain data fusion is performed, fully utilizing the complementarity between different domains and transform features. This overcomes the limitations of extracting high-dimensional features from single-domain data, more comprehensively capturing the complex characteristics of the signal, and thus achieving higher-performance modulation recognition.

[0063] Below, we will refer to Figures 1 to 8 The steps of the above-described multi-domain feature fusion underwater acoustic communication modulation mode autonomous discrimination method in this example embodiment will be described in more detail.

[0064] In step S101, an underwater acoustic communication modulation mode dataset is constructed. The measured signals obtained by sampling underwater communication signals, the simulation data obtained by simulating underwater acoustic channels, and existing publicly available communication signal datasets are collectively used as the dataset for the autonomous discrimination method of underwater acoustic communication modulation modes. The 11 modulation modes in the dataset constructed in this paper are 8PSK, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM, QPSK, and WBFM.

[0065] Necessary preprocessing is performed on the dataset to improve its reliability. The signal is demodulated to baseband, and the quadrature and in-phase components of the communication signal are extracted. Gaussian noise is added to the simulation data, and the dataset is classified according to the signal-to-noise ratio. For each data sample... Normalization process is performed to obtain Each element of each data sample is moved to the range [0,1] to ensure that each element has the same scale and to avoid the influence of different scales on the decision result. The normalization expression is shown below.

[0066]

[0067] In step S102, the dataset is randomly shuffled, and training and test sets are randomly selected. Each sample matrix has a size of (2, 128), representing both in-direction and orthogonal component data. These two data streams are then combined and subjected to a Fourier transform to obtain the frequency domain signal. The specific expression for the Fourier transform is as follows:

[0068]

[0069] The time-frequency diagram of the signal is obtained through continuous wavelet transform, and its expression is:

[0070]

[0071] In steps S103 to S105, a branch network model is constructed. The network in this application includes a time-domain branch, a frequency-domain branch, and a continuous wavelet transform time-frequency graph branch. The input to the time-domain branch is a matrix of size 2×128. Convolutional layer 1 is 2×128×256, pooling layer 1 is 2×64×256, convolutional layer 2 is 2×32×128, pooling layer 2 is 2×32×128, convolutional layer 3 is 2×32×64, pooling layer 3 is 2×16×64, convolutional layer 4 is 2×16×64, pooling layer 4 is 2×8×32, and finally, 128 features are output.

[0072] The frequency domain network branch input is a matrix of size 1×128. Convolutional layer 1 is 1×128×256, pooling layer 1 is 1×64×256, convolutional layer 1 is 1×32×128, pooling layer 2 is 1×32×128, convolutional layer 3 is 1×32×64, pooling layer 3 is 1×16×64, convolutional layer 4 is 1×16×64, pooling layer 4 is 1×8×32, and finally outputs 64 features.

[0073] The time-frequency graph is first compressed to reduce network complexity and improve running speed. The input layer size is 128×128×3. Convolutional layer 1 outputs 64×64×144, dense connection block 1 outputs 64×64×144, pooling layer 1 outputs 32×32×72, dense connection block 2 outputs 32×32×144, convolutional layer 2 outputs 16×16×72, pooling layer 2 outputs 16×16×36, convolutional layer 3 outputs 16×16×36, pooling layer 3 outputs 8×8×18, and finally outputs 64 features.

[0074] A feature fusion network is constructed, with three branches outputting a total of 256 features. To achieve better results by cooperating with the three types of feature descriptors, feature fusion is required. Since the features output by different branches have different relevance to the final communication judgment, a weighted fusion of features from different networks is adopted to reflect the representational ability of each feature to the communication modulation mode. For the calculation of weight values, the Pearson correlation coefficient method is used, and the expression is as follows:

[0075]

[0076]

[0077]

[0078]

[0079] in, is the Pearson correlation coefficient; cov(X,Y) is the covariance of two variables X and Y, used to reflect the relationship between variables X and Y; Let X be the standard deviation of the variable X. Let Y be the standard deviation of the variable Y.

[0080] After calculating the weights, the feature descriptors of the time-domain features, frequency-domain features, and continuous wavelet transform time-frequency plot features are weighted and serially concatenated. The fused feature descriptor is represented as follows:

[0081]

[0082] Based on the analysis of correlation coefficients in the experiment, the weight of the branch feature of the time domain network is 0.35, the weight of the branch feature of the frequency domain network is 0.25, the weight of the continuous wavelet transform time-frequency graph feature is 0.4, and finally, through a fully connected layer and a softmax classification layer, the output size is 11 modulation mode classifications.

[0083] In one embodiment, the underwater acoustic communication modulation mode discrimination method of this application integrates time-domain signals, frequency-domain signals, and continuous wavelet transform time-frequency diagrams. High-dimensional features from different domains are effectively extracted, avoiding the limitations of using single-dimensional data for feature extraction.

[0084] In one embodiment, the correlation of the proposed time-domain network branch, frequency-domain network branch, and continuous wavelet transform time-frequency graph network branch is calculated using the Pearson correlation method. The calculated weights are as follows: the feature weight of the time-domain network branch is 0.35, the feature weight of the frequency-domain network branch is 0.25, and the feature weight of the continuous wavelet transform time-frequency graph is 0.4.

[0085] In one embodiment, the feature descriptors of time-domain features, frequency-domain features, and continuous wavelet transform time-frequency graph features are weighted and serially concatenated. After feature fusion, the 11 types of modulation modes are classified through a fully connected layer and a softmax layer.

[0086] In one embodiment, since the method utilizes the quadrature and in-phase components of the communication signal for processing and discrimination, it is applicable not only to the autonomous discrimination of underwater acoustic communication modulation modes, but also to other communication scenarios such as electromagnetic communication and optical communication.

[0087] In one specific embodiment, the present application will be further described below with reference to the accompanying drawings and embodiments. The present application includes, but is not limited to, the following embodiments.

[0088] 1. Construct the dataset:

[0089] To train a multi-domain fusion underwater acoustic communication modulation mode autonomous discrimination network, a rich dataset was constructed using various data acquisition methods. The dataset includes underwater acoustic communication signals received in real-world underwater environments, simulated signals, and existing publicly available datasets. The simulated transmitted signals, after passing through a simulated multipath channel and adding Gaussian white noise, yielded received signals with signal-to-noise ratios ranging from -20dB to 18dB. The specific simulation signal generation process is as follows: Figure 2 As shown. Existing public datasets include the RML2016.10a dataset and the RML2016.10b dataset. A total of 50,000 samples have been constructed.

[0090] 2. Data Preprocessing

[0091] For both measured and simulated signals, the signals are demodulated to baseband, and the quadrature and in-phase components of the communication signal are extracted. Baseband signals of equal length are truncated, and the data are classified and labeled according to the signal-to-noise ratio and the type of communication signal. Each data sample is normalized by shifting each element of each sample to the range [0,1] to ensure that all elements have the same scale and avoid the influence of different scales on the decision results. The normalized data is used as the input to the time-domain branch network.

[0092] Perform a 128-point Fourier transform on each data sample to obtain a frequency domain signal of the same length, which serves as the input to the frequency domain branch network. Perform a continuous wavelet transform on each data sample and plot the wavelet transform time-frequency graph, which serves as the input to the time-frequency graph branch network.

[0093] 3. Flowchart of the multi-domain feature fusion method for autonomous discrimination of underwater acoustic communication modulation modes is shown below. Figure 3 As shown. The input to the time-domain branch network is a 2×128 time-domain data signal. The input to the frequency-domain branch network is a 128-point frequency-domain data signal. The input to the time-frequency plot branch network is a 128×128×3 time-frequency image.

[0094] Constructing a branch network model:

[0095] Multi-domain feature fusion network convolutional layer and fully connected layer design, such as Figure 4 As shown in the diagram, the network comprises a time-domain branch, a frequency-domain branch, and a continuous wavelet transform time-frequency plot branch. The time-domain branch takes a 2×128 matrix as input, uses a 2×128×256 convolutional layer 1, a 2×64×256 pooling layer 1, a 2×32×128 convolutional layer 2, a 2×32×128 pooling layer 2, a 2×32×64 convolutional layer 3, a 2×16×64 pooling layer 3, a 2×16×64 convolutional layer 4, and a 2×8×32 pooling layer 4, and finally outputs 128 features. The frequency domain network branch input is a matrix of size 1×128. Convolutional layer 1 is 1×128×256, pooling layer 1 is 1×64×256, convolutional layer 1 is 1×32×128, pooling layer 2 is 1×32×128, convolutional layer 3 is 1×32×64, pooling layer 3 is 1×16×64, convolutional layer 4 is 1×16×64, pooling layer 4 is 1×8×32, and finally outputs 64 features. The time-frequency graph is first subjected to dimensionality reduction and compression to reduce network complexity and improve running speed. The input layer size is 128×128×3. Convolutional layer 1 outputs 64×64×144. After passing through dense connection block 1, the output size is 64×64×144. Pooling layer 1 outputs 32×32×72. After passing through dense connection block 2, the output size is 32×32×144. Convolutional layer 2 outputs 16×16×72. Pooling layer 2 outputs 16×16×36. Convolutional layer 3 outputs 16×16×36. Pooling layer 3 outputs 8×8×18. Finally, 64 features are output.

[0096] Constructing a feature fusion network:

[0097] The three branch networks output a total of 256 features. To achieve better results by combining the three types of feature descriptors, feature fusion is required. Since the features output by different branch networks have different relevance to the final communication judgment, a weighted fusion of features from different networks is adopted to reflect the representational ability of each feature to the communication modulation mode. For the calculation of weight values, the Pearson correlation coefficient method is used, and the expression is as follows:

[0098]

[0099]

[0100]

[0101]

[0102] in, is the Pearson correlation coefficient; cov(X,Y) is the covariance of two variables X and Y, used to reflect the relationship between variables X and Y; Let X be the standard deviation of the variable X. Let Y be the standard deviation of the variable Y.

[0103] After calculating the weights, the feature descriptors of the time-domain features, frequency-domain features, and continuous wavelet transform time-frequency plot features are weighted and serially concatenated. The fused feature descriptor is represented as follows:

[0104]

[0105] Based on the analysis of correlation coefficients in the experiment, the weight of the branch feature of the time domain network is 0.35, the weight of the branch feature of the frequency domain network is 0.25, the weight of the continuous wavelet transform time-frequency graph feature is 0.4, and finally, through a fully connected layer and a softmax classification layer, the output size is 11 modulation mode classifications.

[0106] Model training:

[0107] During each model training iteration, the dataset is randomly rearranged. 490,000 data points are extracted as the training set, and 10,000 data points are used as the validation set. The multi-domain feature fusion underwater acoustic communication modulation mode network model is fully trained using the training signal dataset until the network model's detection error loss value drops below a preset threshold. Training is then terminated, and the network model is saved.

[0108] Underwater acoustic pattern recognition:

[0109] The trained network model is validated on a validation dataset. The accuracy of multi-domain feature fusion communication modulation pattern recognition is as follows: Figure 5 As shown. The average recognition accuracy of multi-domain feature fusion communication modulation mode is as follows: Figure 6 As shown in the figure, the recognition accuracy gradually increases with the increase of the signal-to-noise ratio (SNR), and the average recognition rate is 84% ​​when the SNR is greater than 15 dB, which is better than traditional methods. This application effectively realizes the recognition of the dataset.

[0110] The overall confusion matrix of a multi-domain feature fusion communication modulation pattern recognition network is as follows: Figure 7 As shown. The confusion matrix of the multi-domain feature fusion communication modulation pattern recognition network under different signal-to-noise ratios is as follows. Figure 8 As shown in the figure. The results show that, at a signal-to-noise ratio of 18dB, this application achieves over 92% accuracy for modulation modes such as 8PSK, AM-DSB, AM-SSB, and BPSK. It is more prone to confusion with WBFM and 16-QAM modulation modes; therefore, this application is more suitable for identifying modulation modes such as 8PSK and AM-DSB.

[0111] The above-described multi-domain feature fusion method for autonomous discrimination of underwater acoustic communication modulation modes first preprocesses the data to extract the orthogonal and in-phase components of the communication signal, which are then input into the time-domain network branch. The frequency-domain waveform obtained by Fourier transform of the signal is then input into the frequency-domain network branch. The time-frequency graph obtained by continuous wavelet transform of the signal is input into the two-dimensional time-frequency graph network branch. The Pearson correlation method is used to calculate the correlation between the proposed time-domain network branch, frequency-domain network branch, and continuous wavelet transform time-frequency graph network branch. The features obtained from the network branches are then weighted and fused together. After feature fusion, the modulation modes of the underwater acoustic communication signal are classified using a fully connected layer and a softmax layer.

[0112] The basic principles and implementation schemes of the method proposed in this application have been verified by computer numerical simulation. The results show that, in the dataset constructed in this application and under typical underwater acoustic channel environments, the method proposed in this application can effectively classify 11 types of underwater acoustic communication signal modulation modes. In the given typical embodiments, the method proposed in this application (compared to traditional methods) can reduce the modulation mode classification accuracy from 70% to 98%.

[0113] It should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", etc., in the above description indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of this disclosure and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this disclosure.

[0114] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, "a plurality of" means two or more, unless otherwise explicitly specified.

[0115] In the embodiments of this disclosure, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this disclosure according to the specific circumstances.

[0116] In embodiments of this disclosure, unless otherwise expressly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature being directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature being directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0117] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0118] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A method for autonomously determining modulation modes in underwater acoustic communication by fusing multi-domain features, characterized in that, The method includes: Construct a dataset of underwater acoustic communication modulation modes and preprocess the dataset. The preprocessed underwater acoustic communication modulation mode dataset is divided into a training set and a test set according to a preset ratio; A multi-domain fusion communication pattern recognition network is constructed; wherein the multi-domain fusion communication pattern recognition network includes a time-domain branch network, a frequency-domain branch network, a time-frequency graph branch network, and a feature fusion network, wherein the time-domain branch network, the frequency-domain branch network, and the time-frequency graph branch network run in parallel and are all connected to the feature fusion network; The multi-domain fusion communication pattern recognition network is trained using the training set and the test set respectively to obtain the trained multi-domain fusion communication pattern recognition network; The data to be judged is input into the trained multi-domain fusion communication pattern recognition network to obtain the judgment result; wherein, The step of constructing an underwater acoustic communication modulation mode dataset and preprocessing the dataset includes: Based on the measured signals received from the underwater acoustic environment, the simulated signals generated by simulation, and the underwater acoustic dataset in the database, an underwater acoustic communication modulation mode dataset is constructed. The measured signal and the simulated signal are demodulated to baseband, the quadrature components and in-phase components of the measured signal and the simulated signal are extracted, baseband signals of the same length are truncated, and the measured signal and the simulated signal are classified and labeled according to the signal-to-noise ratio and the type of communication signal. All data samples in the underwater acoustic communication modulation mode dataset Normalization process is performed to obtain and each data sample Each element is moved to the range [0,1]; where the normalization expression is: In the formula, This is the normalized dataset of underwater acoustic communication modulation modes. The maximum amplitude in the original dataset. This represents the minimum amplitude value in the original dataset. For each of the data samples Perform a 128-point Fourier transform to obtain a frequency domain signal of the same length, which is then used as the input to the frequency domain branch network; the expression for the Fourier transform is: In the formula, The nth element of a normalized single underwater acoustic communication data set. Let N be the k-th element of the frequency domain signal vector, N be the number of Fourier transform points, n be the coordinates of the time domain signal element, k be the coordinates of the frequency domain signal element, and j be an imaginary number. For each of the data samples Perform continuous wavelet transform and plot the wavelet transform time-frequency graph, which serves as the input to the time-frequency graph branch network; wherein, the expression for the wavelet transform is: In the formula, The nth element of a normalized single underwater acoustic communication data set. Let be a wavelet function, 'a' be the 'a'-th row of the wavelet-transformed matrix, and 'b' be the 'b'-th column of the wavelet-transformed matrix. c o It is a constant. d o It is a constant; The time-domain branching network includes: The system consists of a temporal input layer, a temporal convolutional layer 1, a temporal pooling layer 1, a temporal convolutional layer 2, a temporal pooling layer 2, a temporal convolutional layer 3, a temporal pooling layer 3, a temporal convolutional layer 4, a temporal pooling layer 4, and a temporal output layer; among which, The temporal input layer of the temporal branch network has a matrix size of 2×128 as input, the temporal convolutional layer 1 has a size of 2×128×256, the temporal pooling layer 1 has a size of 2×64×256, the temporal convolutional layer 2 has a size of 2×32×128, the temporal pooling layer 2 has a size of 2×32×128, the temporal convolutional layer 3 has a size of 2×32×64, the temporal pooling layer 3 has a size of 2×16×64, the temporal convolutional layer 4 has a size of 2×16×64, the temporal pooling layer 4 has a size of 2×8×32, and the temporal output layer outputs 128 temporal features. The frequency domain branching network includes: The system consists of a frequency domain input layer, a frequency domain convolutional layer 1, a frequency domain pooling layer 1, a frequency domain convolutional layer 1, a frequency domain pooling layer 2, a frequency domain convolutional layer 3, a frequency domain pooling layer 3, a frequency domain convolutional layer 4, a frequency domain pooling layer 4, and a frequency domain output layer; among which, The frequency domain input layer of the frequency domain branch network has a matrix size of 1×128 as input, the frequency domain convolutional layer 1 has a size of 1×128×256, the frequency domain pooling layer 1 has a size of 1×64×256, the frequency domain convolutional layer 1 has a size of 1×32×128, the frequency domain pooling layer 2 has a size of 1×32×128, the frequency domain convolutional layer 3 has a size of 1×32×64, the frequency domain pooling layer 3 has a size of 1×16×64, the frequency domain convolutional layer 4 has a size of 1×16×64, the frequency domain pooling layer 4 has a size of 1×8×32, and the output layer outputs 64 frequency domain features. The time-frequency graph branch network includes: The time-frequency graph consists of an input layer, a convolutional layer 1, a densely connected block 1, a pooling layer 1, a densely connected block 2, a convolutional layer 2, a pooling layer 2, a convolutional layer 3, a pooling layer 4, and an output layer; where, The time-frequency graph input layer of the time-frequency graph branch network has a size of 128×128×3, the output of time-frequency graph convolutional layer 1 is 64×64×144, the output of dense connection block 1 is 64×64×144, the output of time-frequency graph pooling layer 1 is 32×32×72, the output of dense connection block 2 is 32×32×144, the output of time-frequency graph convolutional layer 2 is 16×16×72, the output of time-frequency graph pooling layer 2 is 16×16×36, the output of time-frequency graph convolutional layer 3 is 16×16×36, the output of time-frequency graph pooling layer 3 is 8×8×18, and the time-frequency graph output layer outputs 64 time-frequency graph features. The training steps of the multi-domain fusion communication pattern recognition network include: The time-domain signal in the data sample is input into the time-domain branch network to obtain the time-domain features; The frequency domain signal in the data sample is input into the frequency domain branch network to obtain the frequency domain features; The time-domain graph in the data sample is input into the time-domain graph branch network to obtain the time-domain graph features; Calculate the weight values ​​of the time-domain feature, the frequency-domain feature, and the time-domain graph feature respectively; Based on the weight values ​​of each feature, the feature descriptors of the time-domain feature, the frequency-domain feature, and the time-frequency map feature are weighted and serially concatenated to obtain the fused feature; The fused features are then used to classify all modulation modes through a fully connected layer and a softmax layer.

2. The multi-domain feature fusion underwater acoustic communication modulation mode autonomous discrimination method according to claim 1, characterized in that, The step of dividing the preprocessed underwater acoustic communication modulation mode dataset into a training set and a test set according to a preset ratio includes: All data samples in the preprocessed underwater acoustic communication modulation mode dataset are randomly shuffled, and the training set and the test set are randomly selected according to a preset ratio.

Images