A Method and System for Underwater Acoustic Target Recognition Based on Mel-Cepstral and Attention Residual Networks

Abstract

This invention provides a method and system for underwater acoustic target recognition based on Mel-Cepstral Interpretation (MCI) and Attention Residual Networks (ARDNs), comprising: Step 1: acquiring labeled underwater acoustic target samples, extracting MCI features, first-order difference features, and second-order difference features from the samples to construct a three-dimensional MCI feature vector; Step 2: constructing a target recognition network based on an attention residual module; Step 3: training the target recognition network using the MCI features to obtain an underwater acoustic target recognition model, and using this model to identify the target underwater acoustic object to determine its category. This invention not only enhances the temporal information of the original features but also effectively utilizes various excellent recognition networks and pre-trained models in the field of image recognition, exhibiting strong generalization, stability, and high recognition efficiency. Furthermore, it is applicable to underwater acoustic target recognition with different sampling rates and signal lengths.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic target recognition technology, and more specifically, to an underwater acoustic target recognition method and system based on Mel-Cepstral and Attention Residual Networks. Background Technology

[0002] Underwater target identification is a crucial task in underwater acoustics, providing essential support for underwater acoustic information acquisition and countermeasures, and also representing a significant challenge in the field of underwater acoustic signal processing. The increasing automation of underwater acoustic equipment places higher demands on underwater target identification technology. However, due to the complexity of the marine environment and the influence of noise from various underwater organisms and vehicles, current domestic sonar technology still lags considerably behind the requirements for intelligent underwater acoustic equipment in terms of automatic target identification.

[0003] Patent document CN113435276A (application number: CN202110665957.8) discloses a method for underwater acoustic target recognition based on an adversarial residual network, belonging to the field of passive underwater acoustic target reconnaissance technology. This method first preprocesses the target signal data collected by the hydrophone using sonar processing technology to obtain a time-frequency feature spectrum. Then, a novel adversarial residual network is used to construct an recognition architecture targeting the underwater acoustic target's time-frequency feature spectrum. The sea state label and target identity label used during sample collection are used as dual labels for the underwater acoustic target's time-frequency feature spectrum. The network architecture is trained using artificial intelligence methods to ultimately achieve underwater acoustic target recognition.

[0004] With the development of computer hardware technology, artificial intelligence technologies, represented by deep learning, have been widely applied in speech recognition, speech synthesis, and image classification. For underwater acoustic target recognition, research teams both domestically and internationally have applied deep neural networks such as CNNs and RNNs to acoustic target recognition, achieving preliminary results in the recognition of various targets. However, existing models are relatively simple and have not been studied in depth to address the specific characteristics of underwater acoustic targets. Utilizing deep learning methods based on the acoustic characteristics of targets to extract more discriminative and deeper feature vectors and constructing a high-precision, high-efficiency target recognition model provides a new technical approach for underwater acoustic target recognition. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for underwater acoustic target recognition based on Mel-Cepstral and Attention Residual Networks.

[0006] The underwater acoustic target recognition method based on Mel-frequency cepstral and attention residual networks provided by the present invention includes:

[0007] Step 1: Obtain labeled underwater acoustic target samples, extract Mel-Cepstral features, first-order difference features, and second-order difference features from the underwater acoustic target samples to form a three-dimensional Mel-Cepstral feature vector;

[0008] Step 2: Construct a target recognition network based on the attention residual module;

[0009] Step 3: Train the target recognition network using the Mel-Cepstral features of the underwater acoustic target to obtain the underwater acoustic target recognition model. Use the underwater acoustic target recognition model to identify the underwater acoustic target under test and obtain the underwater acoustic target category.

[0010] Preferably, step 1 includes:

[0011] Step 1.1: Using a sampling frequency f s The original underwater acoustic target sample is resampled, and the resampled signal x(n) is pre-emphasized to obtain the signal y(n), where y(n) = x(n) - 0.97x(n-1);

[0012] Step 1.2: With a frame length of 0.025f s Point, frame shift 0.001f s The signal y(n) is divided into frames, and a Hamming window is added to each frame of the signal before a Fast Fourier Transform is performed to obtain the signal energy spectrum E. (P,Q) Where P is the number of frames and Q is the number of Fourier transform points;

[0013] Step 1.3: Multiply the signal energy spectrum by M Mel filter banks, the result is represented as H, and perform discrete cosine transform and cepstral operation K(·) on H to obtain the original Mel cepstral feature representation as follows: Where p∈{1,2,...,P}, m∈{1,2,...,L}; p is an integer from 1 to P, where P is the maximum number of frames; m is an integer from 1 to L, where L is the filter order; i represents an integer from 1 to M;

[0014] Step 1.4: Perform first and second derivatives on the original Mel cepstral features, and combine the obtained first and second difference coefficients with the original Mel cepstral features to obtain the three-dimensional Mel cepstral feature matrix.

[0015] Preferably, step 2 includes:

[0016] Step 2.1: The three-dimensional Mel-Cepstral feature matrix is ​​fed into the convolutional module. The convolutional module consists of two 3×3 convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function. The feature matrix passing through the convolutional layers is reduced in dimensionality using an average pooling layer.

[0017] Step 2.2: The dimensionality-reduced features are fed into three cascaded attention residual modules to obtain classification features. Each attention residual module contains two convolution operations. After the first convolution operation, the dimensionality-reduced features are batch normalized and non-linearly activated using the ReLU activation function. The result is input into the second convolutional layer and the batch normalization layer. The attention module is used to automatically filter features, highlighting preset important features and suppressing the interference of preset unimportant information. Finally, a 1×1 convolution operation is used to change the dimension of the original features, and the result is integrated with the features filtered by the attention module to obtain the optimized features.

[0018] Step 2.3: Integrate the optimized features using a fully connected layer, enhance the nonlinearity of the model using an activation function, and alleviate the overfitting problem using a Dropout layer;

[0019] Step 2.4: Input the integrated features into the last fully connected layer to obtain the final classification features.

[0020] Preferably, step 3 includes:

[0021] The target recognition network is randomly initialized, and the three-dimensional feature matrix is ​​fed into the network to obtain the underwater acoustic target category prediction result. The difference between the network output and the real label is calculated using the cross-entropy loss function, and the Adam algorithm is used for optimization. The learning rate is set to 0.001, and the network parameters are updated through gradient backpropagation to optimize the network prediction result. After multiple iterations, the final recognition model is obtained.

[0022] The cross-entropy loss function is expressed as: Where N represents the number of samples and K represents the number of categories. Let represent the probability that sample i is of class c. This represents the probability that sample i is predicted to be of class c; i is an integer from 1 to N; and c is an integer from 1 to K.

[0023] Preferably, the dataset is acquired and divided into training and validation sets in an 8:2 ratio to complete training and validation. The recognition performance of the underwater acoustic target recognition model is evaluated using recognition accuracy, precision, recall, and the harmonic mean of precision and recall. Recognition accuracy is expressed as: Precision refers to the proportion of all correctly classified samples out of the total sample. Here, TP represents the number of positive samples correctly predicted by the model, FP represents the number of negative samples correctly predicted by the model, FN represents the number of positive samples correctly predicted by the model, and TN represents the number of negative samples correctly predicted by the model. Precision is expressed as... Recall rate is expressed as The harmonic mean F1 of precision and recall is expressed as: It is positively correlated with model performance.

[0024] The underwater acoustic target recognition system based on Mel-frequency cepstral and attention residual networks provided by the present invention includes:

[0025] Module M1: Acquires labeled underwater acoustic target samples, extracts Mel-Cepstral features, first-order difference features, and second-order difference features from the underwater acoustic target samples, and constructs a three-dimensional Mel-Cepstral feature vector;

[0026] Module M2: Constructs a target recognition network based on the attention residual module;

[0027] Module M3: Trains a target recognition network using the Mel-Cepstral features of underwater acoustic targets to obtain an underwater acoustic target recognition model. The underwater acoustic target to be tested is then identified using the underwater acoustic target recognition model to obtain the underwater acoustic target category.

[0028] Preferably, the module M1 includes:

[0029] Module M1.1: at sampling frequency f s The original underwater acoustic target sample is resampled, and the resampled signal x(n) is pre-emphasized to obtain the signal y(n), where y(n) = x(n) - 0.97x(n-1);

[0030] Module M1.2: with a frame length of 0.025f s Point, frame shift 0.001f s The signal y(n) is divided into frames, and a Hamming window is added to each frame of the signal before a Fast Fourier Transform is performed to obtain the signal energy spectrum E. (P,Q) Where P is the number of frames and Q is the number of Fourier transform points;

[0031] Module M1.3: Multiplies the signal energy spectrum by M Mel filter banks, the result is represented as H, and performs a discrete cosine transform and cepstral operation K(·) on H to obtain the original Mel cepstral feature representation. Where p∈{1,2,...,P}, m∈{1,2,...,L}; p is an integer from 1 to P, where P is the maximum number of frames; m is an integer from 1 to L, where L is the filter order; i represents an integer from 1 to M;

[0032] Module M1.4: Perform first and second derivatives on the original Mel cepstral features, and combine the obtained first and second difference coefficients with the original Mel cepstral features to obtain a three-dimensional Mel cepstral feature matrix.

[0033] Preferably, the module M2 includes:

[0034] Module M2.1: The three-dimensional Mel-Cepstral feature matrix is ​​fed into the convolution module, which consists of two 3×3 convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function. The feature matrix passing through the convolutional layers is reduced in dimensionality using an average pooling layer.

[0035] Module M2.2: The dimensionality-reduced features are fed into three cascaded attention residual modules to obtain classification features. Each attention residual module contains two convolution operations. After the first convolution operation, the dimensionality-reduced features are batch normalized and non-linearly activated using the ReLU activation function. The result is input into the second convolutional layer and the batch normalization layer. The attention module is used to automatically filter features, highlighting preset important features and suppressing the interference of preset unimportant information. Finally, a 1×1 convolution operation is used to change the dimension of the original features, and the result is integrated with the features filtered by the attention module to obtain the optimized features.

[0036] Module M2.3: It integrates information from optimized features using fully connected layers, enhances the nonlinearity of the model using activation functions, and alleviates overfitting by using Dropout layers;

[0037] Module M2.4: Inputs the integrated features into the last fully connected layer to obtain the final classification features.

[0038] Preferably, the module M3 includes:

[0039] The target recognition network is randomly initialized, and the three-dimensional feature matrix is ​​fed into the network to obtain the underwater acoustic target category prediction result. The difference between the network output and the real label is calculated using the cross-entropy loss function, and the Adam algorithm is used for optimization. The learning rate is set to 0.001, and the network parameters are updated through gradient backpropagation to optimize the network prediction result. After multiple iterations, the final recognition model is obtained.

[0040] The cross-entropy loss function is expressed as: Where N represents the number of samples and K represents the number of categories. Let represent the probability that sample i is of class c. This represents the probability that sample i is predicted to be of class c; i is an integer from 1 to N; and c is an integer from 1 to K.

[0041] Preferably, the dataset is acquired and divided into training and validation sets in an 8:2 ratio to complete training and validation. The recognition performance of the underwater acoustic target recognition model is evaluated using recognition accuracy, precision, recall, and the harmonic mean of precision and recall. Recognition accuracy is expressed as: Precision refers to the proportion of all correctly classified samples out of the total sample. Here, TP represents the number of positive samples correctly predicted by the model, FP represents the number of negative samples correctly predicted by the model, FN represents the number of positive samples correctly predicted by the model, and TN represents the number of negative samples correctly predicted by the model. Precision is expressed as... Recall rate is expressed as The harmonic mean F1 of precision and recall is expressed as: It is positively correlated with model performance.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) This invention integrates Mel-Cepstral 3D feature vectors and deep convolutional neural networks, making full use of the compact information expression of Mel-Cepstral features and the excellent feature extraction and recognition capabilities of deep learning. By extracting the first-order and second-order difference information of Mel-Cepstral features, a three-dimensional matrix similar to image input is obtained. This not only enhances the temporal information of the original features, but also effectively utilizes a variety of excellent recognition networks and pre-trained models in the field of image recognition, providing more usable methods for subsequent network design.

[0044] (2) This invention improves the stability of the training process and accelerates model convergence by introducing an attention residual module, thereby effectively alleviating the overfitting problem and shortening the training time. The attention mechanism enables the automatic selection of features, effectively reducing the interference of redundant information and optimizing the extracted target features.

[0045] (3) The present invention can achieve a recognition accuracy of over 99%, has strong generalization, stability and high recognition efficiency, and can be applied to the recognition of underwater acoustic targets with different sampling rates and different signal lengths. Attached Figure Description

[0046] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0047] Figure 1 This is a flowchart of the present invention;

[0048] Figure 2 Here is a flowchart of the feature extraction process;

[0049] Figure 3 To identify the network model structure diagram;

[0050] Figure 4 This is a structural diagram of the attention module. Detailed Implementation

[0051] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0052] Example 1:

[0053] The technical problem this invention aims to solve is to overcome the shortcomings of existing technologies and provide an underwater acoustic target recognition method based on Mel-frequency cepstral features and attention residual networks. This method utilizes the first and second differences of Mel-frequency cepstral features to enhance the temporal information of target features; it employs residual convolutional neural networks to simplify the learning process, enhance gradient propagation, and reduce overfitting; and it uses an attention mechanism to adaptively focus on key features and reduce redundant information, thereby optimizing feature extraction. Experimental results show that this invention can effectively improve the accuracy of underwater acoustic target classification and recognition. To achieve the above objectives, the technical solution adopted by this invention includes the following steps:

[0054] Step 1: Obtain labeled underwater acoustic target samples, extract Mel-Cepstral features and their first-order and second-order difference features from the data samples to form a three-dimensional Mel-Cepstral feature vector;

[0055] Step 2: Design the attention module and residual module to build an underwater acoustic target recognition network;

[0056] Step 3: Train the attention residual network using the Mel-Cepstral features of the underwater acoustic target to obtain the underwater acoustic target recognition model.

[0057] Step 1 includes:

[0058] Step 1.1: Using a sampling frequency f s The original underwater acoustic target sample is resampled, and the resampled signal x(n) is pre-emphasized to obtain the signal y(n), where y(n) = x(n) - 0.97x(n-1);

[0059] Step 1.2: With a frame length of 0.025f s Point, frame shift 0.001f s The signal y(n) is divided into frames, and a Hamming window is added to each frame of the signal before a Fast Fourier Transform is performed to obtain the signal energy spectrum E. (P,Q) Where P is the number of frames and Q is the number of Fourier transform points;

[0060] Step 1.3: Multiply the signal energy spectrum obtained in Step 1.2 by M Mel filter banks, the result of which is denoted as H. Perform discrete cosine transform and cepstral operation K(·) on H to obtain the original Mel cepstral feature representation. Where p∈{1,2,...,P}, m∈{1,2,...,L};

[0061] p is an integer from 1 to P, where P is the maximum number of frames; m is an integer from 1 to L, where L is the filter order; i is an integer from 1 to M.

[0062] Step 1.4: Perform first and second derivatives on the original Mel cepstral features obtained in Step 1.3, and combine the obtained first and second difference coefficients with the original Mel cepstral features to obtain the three-dimensional Mel cepstral feature matrix.

[0063] Step 2 includes:

[0064] Step 2.1: The Mel-Cepstral 3D feature matrix obtained in Step 1 is fed into the convolutional module, which consists of two 3×3 convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function. The feature matrix passing through the convolutional layers is reduced in dimensionality using an average pooling layer.

[0065] Step 2.2: The dimensionality-reduced features are fed into three cascaded attention residual modules to obtain classification features. Each attention residual module contains two convolution operations. After the first convolution operation, the dimensionality-reduced features undergo batch normalization and ReLU activation for non-linear activation. The result is then input into the second convolutional layer and the batch normalization layer. The attention modules automatically filter features, highlighting important features and suppressing interference from unimportant information. Finally, a 1×1 convolution operation is used to change the dimensionality of the original features, and the result is integrated with the features filtered by the attention modules to obtain the optimized features.

[0066] Step 2.3: Integrate the information of the features obtained in Step 2.2 using a fully connected layer, enhance the nonlinearity of the model using an activation function, and effectively alleviate the overfitting problem using a Dropout layer;

[0067] Step 2.4: Input the features obtained in Step 2.3 into the last fully connected layer to obtain the final classification features.

[0068] Step 3 includes:

[0069] The recognition network designed in step 2 is randomly initialized, and the 3D feature matrix obtained in step 1 is fed into the network to obtain the underwater acoustic target category prediction result. The difference between the network output and the true label is calculated using the cross-entropy loss function, and optimized using the Adam algorithm with a learning rate set to 0.001. The network parameters are updated through gradient backpropagation to optimize the network prediction result. After 100 iterations, the final recognition model is obtained. The cross-entropy loss function is expressed as follows: Where N represents the number of samples and K represents the number of categories. Let represent the probability that sample i is of class c. This represents the probability that sample i is predicted to be of class c; i is an integer from 1 to N; and c is an integer from 1 to K.

[0070] This specific implementation example was completed for five types of underwater acoustic target radiated noise datasets (A, B, C, D, E). Figure 1 The flowchart of this invention shows that the raw underwater acoustic data is first preprocessed to extract Mel-frequency cepstral features, and then the features are further optimized using the attention residual module and a network is designed for recognition.

[0071] according to Figure 2 The feature extraction process first uses a sampling frequency f s The radiated noise data of five types of underwater acoustic targets are resampled. The resampled signal x(n) is pre-emphasized to obtain the signal y(n). Then, the signal y(n) is divided into frames, and a Hamming window is added to each frame before a Fast Fourier Transform is performed to obtain the signal energy spectrum E. (P,Q) Then the signal energy spectrum E (P,Q) Multiplying with the M-order Mel filter bank yields H, and performing a discrete cosine transform and cepstral operation K(·) on H obtains the original Mel cepstral features. Finally, the first and second differences of the original Mel cepstral are solved, and the obtained first and second difference coefficients are combined with the original Mel cepstral features to obtain the three-dimensional feature matrix of the Mel cepstral.

[0072] Figure 3 A flowchart of the recognition network model design is provided. The entire recognition network consists of convolutional modules, pooling layers, attention residual modules, fully connected layers, dropout layers, and softmax layers. All convolutional operations in the designed network model are two-dimensional convolutions with kernel sizes of 3×3 and 1×1, and a stride of 1. Average pooling layers are used for feature dimensionality reduction, effectively reducing the number of parameters. Each attention residual module contains two convolutional operations. After the first convolutional operation, the features undergo batch normalization and non-linear activation. After processing by the second convolutional layer and batch normalization layer, the attention module is used to automatically filter the features. Finally, a 1×1 convolutional operation is used to change the dimensionality of the original features, and the result is integrated with the features filtered by the attention module to obtain the optimized features. A schematic diagram of the attention module structure is shown below. Figure 4 The fully connected layer integrates information from the optimized features and uses an activation function to enhance the model's non-linearity. The Dropout layer mitigates overfitting and improves the model's generalization performance. Finally, the fully connected layer yields the final classification features, and the Softmax layer outputs the probability value of each data point being identified as a different category, thus achieving the classification of underwater acoustic targets.

[0073] The dataset was divided into training and validation sets in an 8:2 ratio for training and validation. The model's recognition performance was evaluated using accuracy, precision, recall, and the harmonic mean of precision and recall. Accuracy was expressed as... Precision is the proportion of all correctly classified samples out of the total sample. Here, TP represents the number of positive samples correctly predicted by the model, FP represents the number of negative samples predicted as positive by the model, FN represents the number of positive samples predicted as negative by the model, and TN represents the number of negative samples correctly predicted by the model. Precision is expressed as... Recall rate is expressed as The harmonic mean F1 of precision and recall is expressed as: The results are positively correlated with model performance. The recognition results on the validation set are shown in Table 1. To further demonstrate the effectiveness of each innovation of this invention, the recognition results of different input features and different network models are compared, and the results are shown in Table 2.

[0074] Table 1. Identification results for various categories in the ShipsEar dataset.

[0075] category Accuracy Recall rate F1 score accuracy A 1.00 0.99 0.99 0.99 B 0.98 0.98 0.98 0.98 C 0.99 0.99 0.99 0.99 D 0.99 1.00 0.99 1.00 E 1.00 1.00 1.00 1.00

[0076] Table 2 Ablation Experiment

[0077] Model Accuracy Recall rate F1 score accuracy One-dimensional Mel-spectral features + baseline model 0.94890 0.94872 0.94857 0.94872 3D Mel spectrum features + baseline model 0.96948 0.96923 0.96918 0.96923 3D Mel-spectral features + residual module 0.98203 0.98205 0.98203 0.98205 3D Mel-spectral features + attention residual module 0.99234 0.99231 0.99231 0.99231

[0078] Example 2:

[0079] This invention also provides an underwater acoustic target recognition system based on Mel-Cepstral and Attention Residual Networks. The underwater acoustic target recognition system based on Mel-Cepstral and Attention Residual Networks can be implemented by executing the process steps of the underwater acoustic target recognition method based on Mel-Cepstral and Attention Residual Networks. That is, those skilled in the art can understand the underwater acoustic target recognition method based on Mel-Cepstral and Attention Residual Networks as a preferred embodiment of the underwater acoustic target recognition system based on Mel-Cepstral and Attention Residual Networks.

[0080] The underwater acoustic target recognition system based on Mel-Cepstral and Attention Residual Networks provided by the present invention includes: Module M1: acquiring labeled underwater acoustic target samples, extracting Mel-Cepstral features and their first-order and second-order difference features from the underwater acoustic target samples to form a three-dimensional Mel-Cepstral feature vector; Module M2: constructing a target recognition network based on the attention residual module; Module M3: training the target recognition network using the underwater acoustic target Mel-Cepstral features to obtain an underwater acoustic target recognition model, and recognizing the underwater acoustic target under test through the underwater acoustic target recognition model to obtain the underwater acoustic target category.

[0081] The module M1 includes: Module M1.1: with a sampling frequency f sThe original underwater acoustic target sample is resampled, and the resampled signal x(n) is pre-emphasized to obtain signal y(n), where y(n) = x(n) - 0.97x(n-1); Module M1.2: with a frame length of 0.025f s Point, frame shift 0.001f s The signal y(n) is divided into frames, and a Hamming window is added to each frame of the signal before a Fast Fourier Transform is performed to obtain the signal energy spectrum E. (P,Q) Where P is the number of frames and Q is the number of Fourier transform points; Module M1.3: Multiplies the signal energy spectrum by M Mel filter banks, the result is represented as H, and performs discrete cosine transform and cepstral operation K(·) on H to obtain the original Mel cepstral feature representation as Where p∈{1,2,...,P}, m∈{1,2,...,L}; p is an integer from 1 to P, where P is the maximum number of frames; m is an integer from 1 to L, where L is the filter order; i represents an integer from 1 to M; Module M1.4: Perform first-order and second-order derivatives on the original Mel cepstral features, and combine the obtained first-order and second-order difference coefficients with the original Mel cepstral features to obtain the three-dimensional Mel cepstral feature matrix.

[0082] Module M2 includes: Module M2.1: The three-dimensional Mel-frequency cepstral feature matrix is ​​fed into a convolutional module, which consists of two 3×3 convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function. The feature matrix passing through the convolutional layers is reduced in dimensionality using an average pooling layer. Module M2.2: The dimensionality-reduced features are fed into three cascaded attention residual modules to obtain classification features. Each attention residual module contains two convolutional operations. The dimensionality-reduced features are batch normalized after the first convolutional operation and non-linearly activated using the ReLU activation function. The result is input into the second... The module consists of a convolutional layer and a batch normalization layer. An attention module is used to automatically filter features, highlighting important preset features and suppressing interference from unimportant preset information. Finally, a 1×1 convolution operation is used to change the dimension of the original features, and the result is integrated with the features filtered by the attention module to obtain optimized features. Module M2.3: The fully connected layer is used to integrate the optimized features, and an activation function is used to enhance the nonlinearity of the model. A dropout layer is used to alleviate the overfitting problem. Module M2.4: The integrated features are input into the last fully connected layer to obtain the final classification features.

[0083] The module M3 includes: randomly initializing the target recognition network, feeding the three-dimensional feature matrix into the network to obtain the underwater acoustic target category prediction result, calculating the difference between the network output and the true label using the cross-entropy loss function, optimizing it using the Adam algorithm with a learning rate set to 0.001, updating the network parameters through gradient backpropagation to optimize the network prediction result, and obtaining the final recognition model after multiple iterations; the cross-entropy loss function is expressed as: Where N represents the number of samples and K represents the number of categories. Let represent the probability that sample i is of class c. This represents the probability that sample i is predicted to be of class c; i is an integer from 1 to N; and c is an integer from 1 to K.

[0084] The dataset was acquired and divided into training and validation sets in an 8:2 ratio to complete training and validation. The recognition performance of the underwater acoustic target recognition model was evaluated using recognition accuracy, precision, recall, and the harmonic mean of precision and recall. Recognition accuracy was expressed as: Precision refers to the proportion of all correctly classified samples out of the total sample. Here, TP represents the number of positive samples correctly predicted by the model, FP represents the number of negative samples correctly predicted by the model, FN represents the number of positive samples correctly predicted by the model, and TN represents the number of negative samples correctly predicted by the model. Precision is expressed as... Recall rate is expressed as The harmonic mean F1 of precision and recall is expressed as: It is positively correlated with model performance.

[0085] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0086] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for underwater acoustic target recognition based on Mel-frequency cepstrum and attention residual network, characterized in that, include: Step 1: Obtain labeled underwater acoustic target samples, extract Mel-Cepstral features, first-order difference features, and second-order difference features from the underwater acoustic target samples to form a three-dimensional Mel-Cepstral feature vector; Step 2: Construct a target recognition network based on the attention residual module; Step 3: Train the target recognition network using the Mel-Cepstral features of the underwater acoustic target to obtain the underwater acoustic target recognition model. Use the underwater acoustic target recognition model to identify the underwater acoustic target under test and obtain the underwater acoustic target category. Step 1 includes: Step 1.1: Resample the original underwater acoustic target sample at a sampling frequency Resample the original underwater acoustic target sample at a sampling frequency After pre-emphasis, the signal is obtained wherein ; Step 1.2: Frame length Point, frame shift Point to signal Frame, and after Hamming window to each frame signal, fast Fourier transform, signal energy spectrum is , wherein Frame number, Fourier transform point number; Step 1.3: multiply the signal energy spectrum with a mel filter bank, the result is denoted as , and do discrete cosine transform and cepstrum operation on , get the original mel-cepstral feature denoted as , where , , ; p is an integer from 1 to P, P is the maximum frame number; m is an integer from 1 to L, L is the filter order; i represents an integer from 1 to M; Step 1.4: Perform first and second derivatives on the original Mel cepstral features, and combine the obtained first and second difference coefficients with the original Mel cepstral features to obtain the three-dimensional Mel cepstral feature matrix; Step 2 includes: Step 2.1: The three-dimensional Mel-Cepstral feature matrix is ​​fed into the convolutional module. The convolutional module consists of two 3×3 convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function. The feature matrix passing through the convolutional layers is reduced in dimensionality using an average pooling layer. Step 2.2: The dimensionality-reduced features are fed into three cascaded attention residual modules to obtain classification features. Each attention residual module contains two convolution operations. After the first convolution operation, the dimensionality-reduced features are batch normalized and non-linearly activated using the ReLU activation function. The result is input into the second convolutional layer and the batch normalization layer. The attention module is used to automatically filter features, highlighting preset important features and suppressing the interference of preset unimportant information. Finally, a 1×1 convolution operation is used to change the dimension of the original features, and the result is integrated with the features filtered by the attention module to obtain the optimized features. Step 2.3: Integrate the optimized features using a fully connected layer, enhance the nonlinearity of the model using an activation function, and alleviate the overfitting problem using a Dropout layer; Step 2.4: Input the integrated features into the last fully connected layer to obtain the final classification features; Step 3 includes: The target recognition network is randomly initialized, and the three-dimensional feature matrix is ​​fed into the network to obtain the underwater acoustic target category prediction result. The difference between the network output and the real label is calculated using the cross-entropy loss function, and the Adam algorithm is used for optimization. The learning rate is set to 0.001, and the network parameters are updated through gradient backpropagation to optimize the network prediction result. After multiple iterations, the final recognition model is obtained. The cross-entropy loss function is expressed as: ,in, Indicates the number of samples. Indicates the number of categories. Indicates sample The true category is the first The probability of a class Indicates sample The predicted number is the The probability of the class; i is an integer from 1 to N; c is an integer from 1 to K.

2. The underwater acoustic target recognition method based on Mel-frequency cepstral and attention residual networks according to claim 1, characterized in that, The dataset was acquired and divided into training and validation sets in an 8:2 ratio to complete training and validation. The recognition performance of the underwater acoustic target recognition model was evaluated using recognition accuracy, precision, recall, and the harmonic mean of precision and recall. Recognition accuracy was expressed as: This refers to the proportion of all correctly classified samples out of the total sample, where... This represents the number of positive samples that the model correctly predicted. This represents the number of negative samples that the model predicts to be positive. This represents the number of positive samples that the model predicts to be negative. The precision rate is expressed as: (The number of negative samples correctly predicted by the model is not specified in the original text.) Recall rate is expressed as The harmonic mean of precision and recall Represented as It is positively correlated with model performance.

3. An underwater acoustic target recognition system based on Mel-frequency cepstral and attention residual networks, characterized in that, include: Module M1: Acquires labeled underwater acoustic target samples, extracts Mel-Cepstral features, first-order difference features, and second-order difference features from the underwater acoustic target samples, and constructs a three-dimensional Mel-Cepstral feature vector; Module M2: Constructs a target recognition network based on the attention residual module; Module M3: Trains a target recognition network using the Mel-Cepstral features of underwater acoustic targets to obtain an underwater acoustic target recognition model. The underwater acoustic target recognition model is then used to identify the underwater acoustic target under test and obtain the underwater acoustic target category. The module M1 includes: Module M1.1: Based on sampling frequency Resample the original underwater acoustic target sample and use the resampled signal. The signal is obtained after pre-emphasis. ,in ; Module M1.2: by frame length Point, frame shift Point-to-point signal The signal is divided into frames, and a Hamming window is added to each frame of the signal before a Fast Fourier Transform is performed to obtain the signal energy spectrum. ,in For frame number, The number of points in the Fourier transform; Module M1.3: Combines the signal energy spectrum with... Multiplying a group of Mel filter banks, the result is expressed as: and to Perform Discrete Cosine Transform and Raised-Cepstral Operations The original Mel-Cepstral features are represented as follows: ,in , p is an integer from 1 to P, where P is the maximum number of frames; m is an integer from 1 to L, where L is the filter order; i is an integer from 1 to M. Module M1.4: Perform first and second derivatives on the original Mel cepstral features, and combine the obtained first and second difference coefficients with the original Mel cepstral features to obtain a three-dimensional Mel cepstral feature matrix; The module M2 includes: Module M2.1: The three-dimensional Mel-Cepstral feature matrix is ​​fed into the convolution module, which consists of two 3×3 convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function. The feature matrix passing through the convolutional layers is reduced in dimensionality using an average pooling layer. Module M2.2: The dimensionality-reduced features are fed into three cascaded attention residual modules to obtain classification features. Each attention residual module contains two convolution operations. After the first convolution operation, the dimensionality-reduced features are batch normalized and non-linearly activated using the ReLU activation function. The result is input into the second convolutional layer and the batch normalization layer. The attention module is used to automatically filter features, highlighting preset important features and suppressing the interference of preset unimportant information. Finally, a 1×1 convolution operation is used to change the dimension of the original features, and the result is integrated with the features filtered by the attention module to obtain the optimized features. Module M2.3: It integrates information from optimized features using fully connected layers, enhances the nonlinearity of the model using activation functions, and alleviates overfitting by using Dropout layers; Module M2.4: Inputs the integrated features into the last fully connected layer to obtain the final classification features; The module M3 includes: The target recognition network is randomly initialized, and the three-dimensional feature matrix is ​​fed into the network to obtain the underwater acoustic target category prediction result. The difference between the network output and the real label is calculated using the cross-entropy loss function, and the Adam algorithm is used for optimization. The learning rate is set to 0.001, and the network parameters are updated through gradient backpropagation to optimize the network prediction result. After multiple iterations, the final recognition model is obtained. The cross-entropy loss function is expressed as: ,in, Indicates the number of samples. Indicates the number of categories. Indicates sample The true category is the first The probability of a class Indicates sample The predicted number is the The probability of the class; i is an integer from 1 to N; c is an integer from 1 to K.

4. The underwater acoustic target recognition system based on Mel-frequency cepstral and attention residual networks according to claim 3, characterized in that, The dataset was acquired and divided into training and validation sets in an 8:2 ratio to complete training and validation. The recognition performance of the underwater acoustic target recognition model was evaluated using recognition accuracy, precision, recall, and the harmonic mean of precision and recall. Recognition accuracy was expressed as: This refers to the proportion of all correctly classified samples out of the total sample, where... This represents the number of positive samples that the model correctly predicted. This represents the number of negative samples that the model predicts to be positive. This represents the number of positive samples that the model predicts to be negative. The precision rate is expressed as: (The number of negative samples correctly predicted by the model is not specified in the original text.) Recall rate is expressed as The harmonic mean of precision and recall Represented as It is positively correlated with model performance.

Images