A method and related apparatus for identifying underwater acoustic signal modulation patterns

Abstract

This application discloses a method and related apparatus for identifying underwater acoustic signal modulation patterns, relating to the field of signal modulation pattern identification in non-cooperative underwater acoustic communication. The method first acquires the original passband signal and processes it to obtain two complex signals, I and Q. Using a residual convolutional network module, the spatial features of the I and Q complex signals are extracted to obtain a spatial feature sequence. Then, using a bidirectional gated recurrent unit module, information from the spatial feature sequence is learned from both forward and backward directions, and the learned information is concatenated to obtain a temporal feature vector. The temporal feature vector is then subjected to feature mapping and classification to obtain classification confidence scores for different modulation patterns. The modulation pattern corresponding to the highest classification confidence score is determined as the classification result. This application enables high-precision, automated identification of underwater acoustic signal modulation patterns.

Description

Technical Field

[0001] This application relates to the field of signal modulation pattern identification in non-cooperative underwater acoustic communication, and in particular to a method and related apparatus for underwater acoustic signal modulation pattern identification. Background Technology

[0002] Modulation pattern identification in underwater acoustic communication is a key technology in non-cooperative underwater acoustic communication systems, and it is of great significance for promoting the intelligent development of marine information acquisition and underwater communication systems. Accurate identification of the modulation pattern provides a foundation for subsequent derivation of information such as the symbol rate and frame structure of the communication signal, ultimately enabling the analysis and utilization of the information carried by the non-cooperative underwater acoustic communication signal. With traditional methods increasingly unable to cope with complex underwater acoustic channels and diverse modulation types, deep learning, with its powerful data-driven feature extraction and end-to-end classification capabilities, is gradually becoming the mainstream technology in the field of modulation pattern identification. Some researchers have combined convolutional neural networks (CNNs) with fractional Fourier transforms and noise reduction algorithms, achieving an accuracy rate of over 90% in linear frequency modulation (LFM) signal detection; other studies have fused compressed sensing with CNNs, effectively reducing computational complexity while improving detection performance, demonstrating the application potential of deep learning in this field.

[0003] Despite the progress made in deep learning methods, existing typical network structures still have significant limitations. For example, while Visual Geometry Group 19 (VGG19) possesses strong nonlinear expressive power and a wide receptive field coverage, its large parameter count makes it difficult to deploy in resource-constrained underwater equipment. Long Short-Term Memory (LSTM) networks can effectively capture temporal dependencies, but they have a heavy computational burden and poor real-time performance. DenseNet alleviates the vanishing gradient problem and improves feature reuse efficiency through dense connections, but its complex connection structure results in a huge number of parameters, high memory consumption, and slow training speed, limiting its application in practical underwater acoustic systems. Summary of the Invention

[0004] The purpose of this application is to provide a method and related apparatus for identifying underwater acoustic signal modulation patterns, which can achieve high-precision and robust automatic identification of signal modulation patterns in complex underwater acoustic environments.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] In a first aspect, this application provides a method for identifying underwater acoustic signal modulation patterns, including:

[0007] The original passband signal is acquired and processed to obtain two complex signals, I and Q.

[0008] Using a residual convolutional network module, the spatial features of the I and Q complex signals are extracted to obtain a spatial feature sequence;

[0009] Using a bidirectional gated recurrent unit module, information about spatial feature sequences is learned from both the forward and backward directions, and the learned information is concatenated to obtain a temporal feature vector;

[0010] The time-series feature vectors are subjected to feature mapping and classification to obtain classification confidence scores for different modulation schemes, and the modulation scheme corresponding to the highest classification confidence score is determined as the classification result.

[0011] The residual convolutional network module includes two connected improved joint attention mechanism residual units, each of which includes a convolutional layer and a joint attention mechanism layer introduced after each convolutional layer.

[0012] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the underwater acoustic signal modulation mode identification method described in any one of the above.

[0013] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the underwater acoustic signal modulation mode recognition method described above.

[0014] According to the specific embodiments provided in this application, this application has the following technical effects:

[0015] This application provides a method and related apparatus for identifying the modulation scheme of underwater acoustic signals, effectively balancing high identification accuracy, low computational complexity, and strong generalization ability, significantly improving the modulation scheme identification performance of underwater acoustic signals in complex channel environments. By introducing a joint attention mechanism (CBAM), the network can autonomously focus on key feature regions, enhancing its ability to model channel and spatial information, thereby achieving more accurate identification under various channel conditions. By optimizing the network structure, reducing redundant 1×1 convolution operations and avoiding the stacking of a large number of residual units, the number of parameters and computational overhead are significantly reduced, improving operating efficiency. In addition, by using a bidirectional gated recurrent unit (Bi-GRU) structure, the model can simultaneously capture forward and backward temporal dependencies, significantly enhancing its ability to express the temporal characteristics of underwater acoustic signals, enabling it to exhibit superior robustness and generalization ability when dealing with complex underwater acoustic environments such as multipath propagation, time-varying characteristics, and strong noise interference. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is an application environment diagram of an underwater acoustic signal modulation mode identification method according to an embodiment of this application;

[0018] Figure 2 A flowchart illustrating a method for identifying underwater acoustic signal modulation patterns, provided in an embodiment of this application;

[0019] Figure 3 A simplified flowchart illustrating a method for identifying underwater acoustic signal modulation patterns according to an embodiment of this application;

[0020] Figure 4 A schematic diagram of the overall ResBiGRUNet network structure provided for another embodiment of this application;

[0021] Figure 5 A schematic diagram of the specific structure of an improved joint attention mechanism residual unit provided in another embodiment of this application;

[0022] Figure 6 A schematic diagram of the specific structure of the joint attention mechanism layer provided in another embodiment of this application;

[0023] Figure 7 This is a schematic diagram of the structure of a residual convolutional network module provided in another embodiment of this application;

[0024] Figure 8 This is a schematic diagram of the structure of a bidirectional gated loop unit module provided in another embodiment of this application;

[0025] Figure 9 This application provides another embodiment of the recognition accuracy curves for a single-path channel dataset, showing the recognition accuracy of seven different network models under different signal-to-noise ratios.

[0026] Figure 10 This application provides another embodiment of the recognition accuracy curves for a deep-sea channel dataset, showing the recognition accuracy of seven different network models under different signal-to-noise ratios.

[0027] Figure 11 This application provides another embodiment of the recognition accuracy curves for a shallow sea channel dataset, showing the recognition accuracy of seven different network models under different signal-to-noise ratios.

[0028] Figure 12This application provides another embodiment of the recognition accuracy curves for a time-varying channel dataset, showing the recognition accuracy of seven different network models under different signal-to-noise ratios.

[0029] Figure 13 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application.

[0030] Attached image labels: 102 Terminal, 104 Storage Server. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] The applicant's research found that existing network structures still suffer from insufficient recognition accuracy, low computational efficiency, and limited generalization ability when dealing with complex environments such as multipath propagation, time-varying characteristics, and strong noise in underwater acoustic channels. Therefore, there is an urgent need for a new modulation scheme recognition method that can balance high recognition accuracy, low computational complexity, strong generalization ability, and effectively cope with multipath propagation, time-varying characteristics, and noise interference in underwater acoustic channels.

[0034] To address this, this invention proposes a method for identifying underwater acoustic signal modulation patterns based on a residual bidirectional gated recurrent unit network (ResBiGRUNet). By introducing a convolutional block attention mechanism (CBAM) into the residual units, the method enhances the model's ability to perceive and extract key features. By simplifying the network structure, redundant 1×1 convolution operations and residual unit stacking are reduced, significantly decreasing the number of parameters and improving computational efficiency. Furthermore, the introduction of a bidirectional gated recurrent unit (Bi-GRU) module captures both forward and backward temporal information, making it more suitable for classification tasks with strong temporal correlation, such as underwater acoustic communication signals.

[0035] This application provides a method for identifying underwater acoustic signal modulation patterns, relating to the field of signal modulation pattern identification in non-cooperative underwater acoustic communication, and particularly to a method for identifying underwater acoustic signal modulation patterns based on a Resent Bidirectional Gated Recurrent Unit (ResBiGRUNet).

[0036] The underwater acoustic signal modulation mode identification method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the raw passband signal to server 104. After receiving the raw passband signal, server 104 processes it to obtain I and Q complex signals. Using a residual convolutional network module, it extracts the spatial features of the I and Q complex signals to obtain a spatial feature sequence. Using a bidirectional gated recurrent unit module, it learns the information of the spatial feature sequence from both forward and backward directions, and concatenates the learned information to obtain a temporal feature vector. The temporal feature vector is then used for feature mapping and classification to obtain the classification confidence scores for different modulation schemes, and the modulation scheme corresponding to the highest classification confidence score is determined as the classification result. Server 104 can feed back the obtained classification results to terminal 102. In addition, in some embodiments, the underwater acoustic signal modulation mode identification method can also be implemented by the server 104 or the terminal 102 separately. For example, the terminal 102 can directly process the original passband signal, or the server 104 can obtain the original passband signal from the data storage system and process it.

[0037] The terminal 102 can be, but is not limited to, various desktop computers, laptops, and IoT devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.

[0038] In one exemplary embodiment, such as Figure 2 As shown, a method for identifying underwater acoustic signal modulation patterns is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 204. Wherein:

[0039] Step 201: Obtain the original passband signal and process the original passband signal to obtain two complex signals, I and Q.

[0040] Step 202: Using the residual convolutional network module, extract the spatial features of the I and Q complex signals to obtain a spatial feature sequence.

[0041] Step 203: Using the bidirectional gated recurrent unit module, information of spatial feature sequences is learned from the forward and backward directions respectively, and the learned information is concatenated to obtain the temporal feature vector.

[0042] Step 204: Perform feature mapping and classification on the time-series feature vector to obtain the classification confidence of different modulation schemes, and determine the modulation scheme corresponding to the maximum classification confidence as the classification result.

[0043] By implementing steps 201 to 204 above, the residual convolution module enhanced by the joint attention mechanism (CBAM) effectively focuses on key spatial features in the I and Q signals, significantly improving the robustness of feature extraction in complex underwater acoustic channels (such as multipath and time-varying) and low signal-to-noise ratio environments. Through the bidirectional gated recurrent unit (Bi-GRU) module, the forward and backward temporal dependencies of the feature sequence are captured simultaneously, overcoming the deficiency of unidirectional networks in modeling the long-term continuity of underwater acoustic signals, and greatly improving the expressive power of temporal features. Finally, through a fully connected layer and a Softmax classifier, the high-dimensional temporal features are mapped to the classification confidence of different modulation schemes, realizing end-to-end high-precision and automated modulation scheme identification.

[0044] In another exemplary embodiment of this application, step 201 described above is replaced by the following steps:

[0045] The original passband signal is sampled, down-converted, filtered to remove high-frequency components, and normalized to obtain two complex signals, I and Q.

[0046] In this embodiment, taking the MPSK signal under a single-path channel as an example, the received MPSK signal under a single-path channel is represented as follows: ;

[0047] in, It is the amplitude. It is the carrier frequency. It is the modulation phase. It is the initial phase of the signal. It is additive white Gaussian noise. t is A continuous independent variable representing time.

[0048] The received MPSK signal under the single-path channel is sampled at 512 points and down-converted to obtain I and Q signals. The down-conversion process involves multiplying the received MPSK signal under the single-path channel with the local carrier through a mixer and shifting it back to the baseband.

[0049] The expression for down-converting the received MPSK signal from a single-path channel is as follows:

[0050] ;

[0051] ;

[0052] in, and These are I-channel data and Q-channel data, It is the center frequency of the downconverter. It is the time-domain representation of in-phase noise. It is the time-domain representation of orthogonal noise.

[0053] Furthermore, after filtering out the high-frequency components, only the low-frequency complex baseband components remain. The resulting I and Q signals after filtering out the high-frequency components are expressed as follows:

[0054] ;

[0055] ;

[0056] The two signals mentioned above are combined into a matrix. This yields a 2×512 long IQ sequence matrix.

[0057] ;

[0058] In practical non-cooperative underwater acoustic communication signals, the amplitude of long IQ sequence matrices may vary significantly due to signal strength, noise, modulation methods, etc. To unify the input scale under different modulation methods and improve the model's classification performance and generalization ability, long IQ sequence matrices need to be normalized before being fed into the neural network for training and testing. This embodiment uses the maximum absolute value normalization method. If we consider the I-channel data as the x-axis and the Q-channel data as the y-axis, the normalization process can be understood as proportionally scaling the distribution of all IQ sequence points to a circle with a radius of 1 centered at the origin. The normalization process includes the following steps:

[0059] For unnormalized long IQ sequence matrices , and the I-channel data Treating it as the real part of a complex signal, the Q-channel data... If we consider it as the imaginary part of a complex signal, then the magnitude of the complex signal can be calculated as:

[0060] ;

[0061] Next, the signal magnitude at each time step is normalized to its maximum value using the following formula:

[0062] ;

[0063] in, This represents the normalized signal magnitude. This represents the maximum magnitude of the signal throughout the entire sequence. This operation scales the magnitude of the signal to the range of [-1, 1], ensuring that the amplitude of all data points is compressed to a uniform scale.

[0064] Normalized I and Q signals and The calculation formula is:

[0065] ;

[0066] ;

[0067] After normalization, the normalized IQ sequence matrix is ​​finally obtained. It can be represented as:

[0068] ;

[0069] This serves as the input to the subsequent residual convolutional network module, consisting of two complex signals, I and Q, which include the normalized I and Q signals. This normalization step ensures that the amplitudes of different modulation signals are processed under the same standard, providing a uniform scale for subsequent modulation scheme recognition tasks.

[0070] In another exemplary embodiment of this application, the residual convolutional network module in step 202 above includes two connected improved joint attention mechanism residual units, namely, a first improved joint attention mechanism residual unit and a second improved joint attention mechanism residual unit. The first improved joint attention mechanism residual unit extracts local features using two 3×3 convolutional kernels and enhances channel attention and spatial attention mechanisms, expanding the original single channel to 32 channels. The second improved joint attention mechanism residual unit further expands the number of channels to 64 channels through the same convolutional operation. Then, a two-dimensional max pooling layer is used for 2×2 spatial dimensionality reduction, making the data size 1×256×64, i.e., the length and width are halved, but the number of channels remains unchanged. Finally, a reshape layer is used to convert the data into a sequence format adapted to the input of the bidirectional gated recurrent unit module, i.e., the final output is 256 time steps, each time step containing a 64-dimensional feature vector, which is a 256×64 spatial feature sequence. Here, 256 represents the time step size, and 64 represents the channel dimension.

[0071] The improved joint attention mechanism residual unit includes a convolutional layer and a joint attention mechanism (Convolutional Block Attention Module, CBAM) layer introduced after each convolutional layer. A detailed structural diagram of the improved joint attention mechanism residual unit can be found in [link to diagram]. Figure 5 .

[0072] For a detailed structural diagram of the residual convolutional network module, please refer to [link / reference]. Figure 7Specifically, the first improved joint attention mechanism residual unit includes: an input layer, a first main path, a first shortcut path, a first adder, and a third activation function layer.

[0073] The first main path includes a first convolutional layer, a first batch normalization layer, a first joint attention mechanism module, a first activation function layer, a second convolutional layer, a second batch normalization layer, a second joint attention mechanism layer, and a second activation function layer connected in sequence; the input end of the first convolutional layer is connected to the output end of the input layer; the input end of the input layer is used to input the I and Q complex signals.

[0074] The first shortcut path is set in parallel with the first main path. To ensure that the residual connection part can effectively capture key information, the present invention also adds a 1×1 convolution on the first shortcut path and applies a joint attention mechanism to further enhance the feature extraction capability. The first shortcut path includes a third convolutional layer, a third batch normalization layer, and a third joint attention mechanism module connected in sequence; the input end of the third convolutional layer is connected to the output end of the input layer.

[0075] The output of the second activation function layer of the first main path and the output of the third joint attention mechanism module of the first shortcut path are connected to the input of the first adder, and the output of the first adder is connected to the input of the third activation function layer; wherein, the output of the third activation function layer is the output of the first improved joint attention mechanism residual unit.

[0076] The second improved joint attention mechanism residual unit includes: a second main path, a second shortcut path, a second adder, a two-dimensional max pooling layer, a reshape layer, and an output layer.

[0077] The second main path includes a fourth convolutional layer, a fourth batch normalization layer, a fourth joint attention mechanism layer, a third activation function layer, a fifth convolutional layer, a fifth batch normalization layer, a fifth joint attention mechanism layer, and a fourth activation function layer connected in sequence; the input of the fourth convolutional layer is connected to the output of the first improved joint attention mechanism residual unit.

[0078] The second shortcut path is set in parallel with the second main path. To ensure that the residual connection part can effectively capture key information, the present invention also adds a 1×1 convolution to the second shortcut path and applies a joint attention mechanism to further enhance the feature extraction capability. The second shortcut path includes a sixth convolutional layer, a sixth batch normalization layer, and a sixth joint attention mechanism layer connected in sequence; the input of the sixth convolutional layer is connected to the output of the first improved joint attention mechanism residual unit.

[0079] The output of the fourth activation function layer and the output of the sixth joint attention mechanism layer are both connected to the input of the second adder. The output of the second adder is connected to the input of the two-dimensional max pooling layer. The output of the two-dimensional max pooling layer is connected to the input of the reshape layer. The output of the reshape layer is connected to the output layer. The output of the fourth activation function layer is the output of the second improved joint attention mechanism residual unit.

[0080] In another exemplary embodiment of this application, the specific structure of the joint attention mechanism layer is detailed in [reference needed]. Figure 6 It includes a connected Channel Attention Module (CAM) layer and a Spatial Attention Module (SAM) layer, which focus information in the channel and space respectively.

[0081] The calculation method for the output features of the joint attention mechanism layer is as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] in, For input features, This indicates a global average pooling operation. This represents the global max pooling operation. This represents the Sigmod activation function. A two-layer fully connected neural network These are the features after processing by the channel attention module. These are the features after processing by the spatial attention module. The output features are those obtained after the fusion of the channel attention module and the spatial attention module.

[0086] In another exemplary embodiment of this application, the calculation method of the output features of the improved joint attention mechanism residual unit is as follows:

[0087] ;

[0088] in, It is represented as a variation function of the main path of the residual unit in the improved joint attention mechanism; The features are input to the residual units of the improved joint attention mechanism. For all learnable convolutional kernel parameters in the residual units of the improved joint attention mechanism, The output characteristics of the residual unit in the improved joint attention mechanism.

[0089] In another exemplary embodiment of this application, the 256×64 spatial feature sequence obtained in the previous exemplary embodiment is used, and the obtained 256×64 spatial feature sequence is input into the bidirectional gated loop unit module. A detailed structural diagram of the bidirectional gated loop unit module can be found in [link to diagram]. Figure 8 Specifically, the bidirectional gated loop unit module includes a forward gated loop unit layer, a backward gated loop unit layer, and a splicing layer, with the forward and backward gated loop unit layers arranged in parallel. Compared to the traditional unidirectional gated loop unit, which can only propagate information from the past to the future, the bidirectional gated loop unit module can learn the information of the input sequence from both the forward and backward directions through two parallel gated loop unit layers.

[0090] Both the forward-gated recurrent unit layer and the backward-gated recurrent unit layer contain 128 hidden units. Using a bidirectional gated recurrent unit module, information from the spatial feature sequence is learned from both the forward and backward directions, and the learned information is concatenated to obtain a 128-dimensional temporal feature vector, specifically including:

[0091] Using the aforementioned forward-gated recurrent unit layer, the feature vectors are processed sequentially from the first time step to the last time step to obtain the forward hidden state at each time step.

[0092] Using the backward-gated recurrent unit layer, the feature vector is processed in chronological order from the first time step to the last time step to obtain the backward hidden state at each time step.

[0093] By using a concatenation layer, the forward hidden state and the backward hidden state corresponding to each time step are concatenated to obtain the temporal feature vector.

[0094] The calculation method for the bidirectional gated loop unit module is as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] in, and They represent The hidden states of the forward GRU and backward GRU at any given time. This represents the forward GRU computation function. This represents the backward GRU computation function. Then it means The hidden state that is ultimately output at any given moment. () indicates that the hidden states of the forward GRU and the backward GRU are concatenated at each time step.

[0099] In another exemplary embodiment of this application, feature mapping and classification are performed on the time-series feature vector to obtain classification confidence scores for different modulation schemes, specifically including:

[0100] A fully connected layer is used to perform feature mapping on the temporal feature vector to obtain a fully connected output vector; the dimension of the fully connected output vector is the same as the number of modulation schemes.

[0101] The fully connected output vector is sequentially input into the Dropout layer and the Softmax classification layer for processing to obtain the classification confidence scores for different modulation schemes. The overall ResBiGRUNet network structure, consisting of residual convolutional network modules and bidirectional gated recurrent unit modules, is shown in the figure below. Figure 4 As shown.

[0102] In this embodiment, the Softmax classification layer uses Softmax as the classifier for different categories of underwater acoustic signal modulation recognition classification tasks. The Softmax function can convert the network output into a probability distribution, defined as follows:

[0103] ;

[0104] in, This represents the last layer of the network. i One output, This represents the last layer of the network. j One output, The total number of categories, Indicates the first The probability distribution of the class.

[0105] The advantage of the Softmax classifier is that it can normalize the output of the neural network so that the sum of the probabilities of all classes is 1, that is:

[0106]

[0107] This ensures that the output of the neural network can be directly interpreted as classification confidence, making it easier for the model to predict the mode in underwater acoustic signal modulation mode recognition tasks.

[0108] In another exemplary embodiment of this application, gradient descent is used for model training. The core parameters are: 10 classes, 6000 training samples and 2000 test samples per class of modulation signal, 64 samples per batch, 100 training cycles in total, initial learning rate of 0.001, learning rate patience value of 3, learning rate descent parameter of 0.5, and early stopper patience value of 10.

[0109] The training network uses a Softmax classifier and an Adaptive Moment Estimation (Adam) optimizer, with the loss function set as the classification cross-entropy.

[0110] The Softmax classifier ensures that the output of the neural network can be directly interpreted as classification confidence, which facilitates the model in class prediction in underwater acoustic signal modulation identification tasks.

[0111] Adam exhibits fast convergence speed and good generalization ability across various neural network architectures. The Adam optimizer dynamically adjusts the learning rate through first-order and second-order moment estimations. Let the neural network's... The updated training gradient is Then the first moment of the gradient and second moment The calculation formula is as follows:

[0112] ;

[0113] ;

[0114] in, and Let these represent the first and second moments of the gradient at the previous moment, respectively. and It controls the decay rate of the first and second moments.

[0115] Next and Perform deviation correction:

[0116] ;

[0117] ;

[0118] in, and It is the decay rate of the first and second moments at the current moment. and These are the bias correction values ​​for the first and second moments, respectively. The final weights of the neural network at the current time step... The calculation formula is as follows:

[0119] ;

[0120] in, A very small constant added to ensure numerical stability. These are the weight parameters of the model from the previous time step. This is the learning rate.

[0121] Since this embodiment is a multi-class classification task, the loss function chosen is the class cross-entropy loss. The physical meaning of cross-entropy loss is to measure the difference between two probability distributions. This loss function is defined as follows:

[0122] ;

[0123] in, Indicates the loss value. Indicates the number of categories. Encoding for the true category, The predicted probability is calculated using Softmax.

[0124] When predicting probability Approximate True Category Coding At that time, the loss value Smaller values ​​indicate a larger value, and vice versa. Therefore, during training, the optimizer minimizes... To improve the classification accuracy of the model.

[0125] In another exemplary embodiment of this application, the recognition accuracy curves of Residual Bidirectional Gated Recurrent Unit Network (ResBiGRUNet) and other traditional networks for four datasets—single-path channel, deep-sea channel, shallow-sea channel, and time-varying channel—under different signal-to-noise ratios are shown below. Figure 9-12 As shown.

[0126] Because multipath effects have a relatively small impact on communication signals in deep-sea channels, the deep-sea channel dataset is very similar to the single-path channel dataset, and the performance of various neural network models on both datasets shows strong consistency. The ResBiGRUNet method in this invention achieves an average recognition accuracy of 95.30% on the single-path channel dataset and 95.11% on the deep-sea channel dataset, the highest among all models. Compared to DenseNet (94.49% and 93.06% respectively), ResBiGRUNet improves performance by 0.81 and 2.05 percentage points on the two datasets, demonstrating superior recognition capabilities.

[0127] from Figure 9 and Figure 10Further observation reveals that ResBiGRUNet in this invention maintains a consistently high recognition rate under varying signal-to-noise ratio (SNR) conditions, particularly excelling in the low SNR range (-20dB to 0dB). At SNR = -16dB, ResBiGRUNet's recognition rate exceeds 90%, while VGG19 and DenseNet remain around 85%. LSTM and CGDNN achieve recognition accuracies between 65% and 75%, while GRU and CLDNN perform even worse, with recognition accuracies failing to reach 60%. Furthermore, in the high SNR range (0dB and above), ResBiGRUNet rapidly approaches 100% recognition, significantly outperforming LSTM, GRU, and CLDNN, demonstrating its superior generalization ability.

[0128] The propagation characteristics of shallow sea environments are complex, and signals are significantly affected by multipath effects and non-stationary noise, making the identification task more challenging than on single-path and deep-sea channel datasets. On the shallow-sea channel dataset, different neural network models show significant performance differences. ResBiGRUNet maintains its leading position, achieving an average recognition accuracy of 94.82%, significantly outperforming DenseNet (84.07%) and CGDNN (88.94%). In particular, compared to CLDNN (52.96%) and GRU (60.74%), ResBiGRUNet's recognition accuracy is improved by 41.86 and 34.08 percentage points respectively, demonstrating a substantial advantage.

[0129] from Figure 11 As seen in the curves, ResBiGRUNet maintains strong noise resistance in the low SNR (-18dB to -10dB) range of the shallow sea channel dataset, with its recognition rate consistently above 70%, while other models (such as GRU and CLDNN) only achieve around 50% recognition in this range. Furthermore, ResBiGRUNet achieves a 95% recognition rate at SNR = -14dB, significantly higher than LSTM, CGDNN, DenseNet, and VGG19 (approximately 55%, 65%, 80%, and 85%, respectively). At higher SNR levels (-10dB and above), ResBiGRUNet's recognition rate rapidly approaches 100%, while CLDNN, GRU, DenseNet, and LSTM, even at SNRs above 0dB, still cannot achieve 100% accuracy, exhibiting certain recognition bottlenecks. This suggests that the complexity of the shallow sea environment significantly impacts networks with weaker feature extraction capabilities.

[0130] On time-varying channel datasets, the overall performance of each model declined compared to other datasets, mainly due to the non-stationary nature of time-varying channels across different time periods. The ResBiGRUNet proposed in this study maintains its leading position with an average recognition accuracy of 92.27% on this dataset. Compared to the second-best performing VGG19 (89.31%), it improved by 2.96 percentage points. Compared to DenseNet (83.40%) and CGDNN (85.79%), it improved by 8.87 and 6.48 percentage points respectively. Compared to LSTM (80.48%) and GRU (55.58%), ResBiGRUNet improved by 11.79% and 36.69% respectively, fully demonstrating its adaptability to time-varying environments.

[0131] Depend on Figure 12 As can be seen, within the SNR range of -18dB to -10dB, ResBiGRUNet's recognition rate is above 65%, while CLDNN and GRU are both below 50%. At SNR = -12dB, ResBiGRUNet's recognition rate exceeds 90%, while DenseNet, CGDNN, and VGG19's recognition rates remain below 85%, indicating that ResBiGRUNet can reach a stable recognition state earlier in complex time-varying environments. Even with an SNR above -4dB, ResBiGRUNet still demonstrates superior recognition capabilities, with a recognition rate approaching 100%, significantly higher than CLDNN, GRU, DenseNet, and LSTM.

[0132] In summary, ResBiGRUNet outperforms other traditional neural network models on all datasets, demonstrating excellent noise resistance and generalization ability.

[0133] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 13As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores and processes data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for identifying underwater acoustic signal modulation patterns.

[0134] Those skilled in the art will understand that Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0135] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0136] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0137] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0138] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0139] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0140] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0141] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0142] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying a water acoustic signal modulation mode, characterized in that, The underwater acoustic signal modulation mode recognition method comprises: An original passband signal is acquired, and the original passband signal is processed to obtain I and Q two-channel complex signals; A residual convolution network module is used to extract spatial features of the I and Q two-channel complex signals to obtain a spatial feature sequence; A bidirectional gated recurrent unit module is used to learn information of the spatial feature sequence from a forward direction and a backward direction respectively, and the learned information is spliced to obtain a time sequence feature vector; The time sequence feature vector is subjected to feature mapping and classification to obtain classification confidence of different modulation modes, and a modulation mode corresponding to a maximum classification confidence is determined as a classification result; The residual convolution network module comprises two connected improved joint attention mechanism residual units, and the improved joint attention mechanism residual unit comprises a convolution layer and a joint attention mechanism layer introduced after each convolution layer. The residual convolution network module comprises a first improved joint attention mechanism residual unit and a second improved joint attention mechanism residual unit; the first improved joint attention mechanism residual unit comprises an input layer, a first main path, a first shortcut path, a first adder and a third activation function layer. The first main path comprises a first convolution layer, a first batch normalization layer, a first joint attention mechanism module, a first activation function layer, a second convolution layer, a second batch normalization layer, a second joint attention mechanism layer and a second activation function layer which are connected in sequence; an input end of the first convolution layer is connected with an output end of the input layer; and an input end of the input layer is configured to input the I and Q two-channel complex signals. The first shortcut path is arranged in parallel with the first main path, and the first shortcut path comprises a third convolution layer, a third batch normalization layer and a third joint attention mechanism module which are connected in sequence; an input end of the third convolution layer is connected with the output end of the input layer. An output end of the second activation function layer of the first main path and an output end of the third joint attention mechanism module of the first shortcut path are connected with an input end of the first adder in common; an output end of the first adder is connected with an input end of the third activation function layer; and an output end of the third activation function layer is an output end of the first improved joint attention mechanism residual unit.

2. The method of claim 1, wherein, The original passband signal is processed to obtain the I and Q two-channel complex signals, specifically comprising: The original passband signal is sampled, subjected to down-conversion processing, filtered to remove high-frequency components and normalized to obtain the I and Q two-channel complex signals.

3. The method of claim 1, wherein the method further comprises: The residual convolution network module comprises a first improved joint attention mechanism residual unit and a second improved joint attention mechanism residual unit; the second improved joint attention mechanism residual unit comprises a second main path, a second shortcut path, a second adder, a two-dimensional maximum pooling layer, a reshape layer and an output layer. The second main path comprises a fourth convolutional layer, a fourth batch normalization layer, a fourth joint attention mechanism layer, a third activation function layer, a fifth convolutional layer, a fifth batch normalization layer, a fifth joint attention mechanism layer and a fourth activation function layer connected in sequence; an input end of the fourth convolutional layer is connected with an output end of the first improved joint attention mechanism residual unit; The second shortcut path is arranged in parallel with the second main path, and the second shortcut path comprises a sixth convolutional layer, a sixth batch normalization layer and a sixth joint attention mechanism layer connected in sequence; an input end of the sixth convolutional layer is connected with an output end of the first improved joint attention mechanism residual unit; An output end of the fourth activation function layer and an output end of the sixth joint attention mechanism layer are jointly connected to an input end of the second adder, an output end of the second adder is connected to an input end of the two-dimensional maximum pooling layer, an output end of the two-dimensional maximum pooling layer is connected to an input end of the Reshape layer, and an output end of the Reshape layer is connected to an output layer. An output end of the fourth activation function layer is an output end of the second improved joint attention mechanism residual unit.

4. The method of claim 1 or 3, wherein The joint attention mechanism layer comprises a connected channel attention layer and a spatial attention layer; The calculation manner of the output feature of the joint attention mechanism layer is specifically as follows: ； ； ； wherein, is an input feature, denotes a global average pooling operation, denotes a global max pooling operation, denotes a Sigmoid activation function, is a two-layer fully connected neural network, is a feature processed by a channel attention module, is a feature processed by a spatial attention module, is an output feature fused by the channel attention module and the spatial attention module.

5. The method of claim 4, wherein the water acoustic signal modulation mode is identified by: The calculation manner of the output feature of the improved joint attention mechanism residual unit is specifically as follows: ； wherein, is a function of the main path of the improved joint attention mechanism residual unit; is a feature input to the improved joint attention mechanism residual unit, is all the learnable convolution kernel parameters in the improved joint attention mechanism residual unit, is the output feature of the improved joint attention mechanism residual unit.

6. The method of claim 1, wherein, The spatial feature sequence comprises a plurality of time steps, and each time step comprises a multi-dimensional feature vector; The bidirectional gated recurrent unit module comprises a forward gated recurrent unit layer, a backward gated recurrent unit layer and a splicing layer, and the forward gated recurrent unit layer and the backward gated recurrent unit layer are arranged in parallel; the forward gated recurrent unit layer and the backward gated recurrent unit layer each comprise a plurality of hidden units; the bidirectional gated recurrent unit module is used to learn information of the spatial feature sequence from the front and back respectively, and the learned information is spliced to obtain a time sequence feature vector, and the method specifically comprises: The forward gated recurrent unit layer is used to process the feature vector in time sequence from the first time step to the last time step to obtain a forward hidden state of each time step; The backward gated recurrent unit layer is used to process the feature vector in time sequence from the first time step to the last time step to obtain a backward hidden state of each time step; The splicing layer is used to splice the forward hidden state and the backward hidden state corresponding to each time step to obtain a time sequence feature vector.

7. The method of claim 1, wherein the method further comprises: The time sequence feature vector is subjected to feature mapping and classification to obtain classification confidence of different modulation modes, and the method specifically comprises: A fully connected layer is used to perform feature mapping on the time sequence feature vector to obtain a fully connected output vector; the dimension of the fully connected output vector is the same as the number of modulation modes; The fully connected output vector is sequentially input into a Dropout layer and a Softmax classification layer for processing to obtain classification confidence of different modulation modes.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the method for identifying the modulation mode of underwater acoustic signals according to any one of claims 1-7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method for identifying the modulation mode of underwater acoustic signals according to any one of claims 1-7.

Images