A sample expansion method for inhibiting data set imbalance in underwater acoustic target recognition

Abstract

The application provides a sample expansion method for inhibiting data set imbalance in underwater acoustic target recognition, fully utilizes learning and decision performance of a convolutional neural network, applies the convolutional neural network to a generator and a discriminator of a CGAN network, adopts a three-layer convolutional neural network in the generator, and generates a small number of class samples according to learned features; the discriminator adopts a one-layer convolutional neural network structure, judges the authenticity of time-frequency features generated by the generator and time-frequency features of real samples in each training, and feeds back the judgment result to the generator to make the generator generate time-frequency feature samples with higher quality; in the CGAN network, input is replaced by time-frequency features of underwater acoustic samples instead of original underwater acoustic signals, so that the generator can better learn data features. The network can autonomously define the class of generated data, supplement the data set in a directional manner, and solve the problem of low overall recognition accuracy caused by the imbalance problem that the proportion of a certain type of data in the data set is too small.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic signal processing technology, specifically relating to a sample augmentation method for suppressing imbalance in underwater acoustic target recognition datasets. Background Technology

[0002] Underwater acoustic target recognition is a comprehensive subject involving multiple disciplines such as signal processing, sonar technology, and pattern recognition technology. It has been widely used in many fields, including marine environmental sounding and modeling, seabed target location and recognition, and the detection and recognition of underwater targets such as mines and submarines.

[0003] With the further development of computer hardware technology and signal processing technology, artificial intelligence (AI) technologies, represented by machine learning (ML), deep learning (DL), and big data, have been widely used in the research of intelligent target recognition technology for underwater acoustic signals, which can effectively improve the generalization ability and environmental adaptability of autonomous recognition systems.

[0004] In recent years, deep neural networks have been widely applied in various fields and have shown good results. Many researchers have also applied them to underwater acoustic target recognition. Although the amount of underwater acoustic target noise data is limited, deep learning can still leverage its advantages in feature learning and noise resistance. In particular, the integrated network structure of deep neural networks provides conditions for achieving joint parameter optimization.

[0005] However, in practical applications, the number of underwater acoustic samples is often limited, and there is often a problem of imbalanced datasets, meaning that the proportion of each category in the dataset varies significantly. When this problem exists, deep neural networks tend to favor the category with the larger number of samples, which negatively impacts the recognition results and prevents deep learning methods from achieving optimal recognition performance in real-world applications.

[0006] Conditional Generative Adversarial Networks (CGANs) are an extension of generative adversarial networks that control the generation process by introducing conditional information into the generator and discriminator. This conditional information can be class labels, attribute vectors, or other relevant auxiliary information. Unlike existing methods for dealing with imbalanced datasets, such as sampling, weighting, and single-classification, CGAN does not mechanically copy existing samples nor discard sample information. It also avoids the phenomenon of the model favoring a large proportion of samples during recognition. Its generator part uses deep learning models such as convolutional neural networks to fully learn sample information, and the discriminator part is generally handled by a deep learning classifier with strong recognition capabilities, further prompting the generator to produce high-quality samples.

[0007] Currently, CGAN networks are only used in the image domain for sample supplementation, and in current applications, the supplementary samples generated by CGAN networks are directly generated, without the ability to generate time-frequency features.

[0008] Underwater acoustic signals are significantly affected by the environment during propagation, exhibiting characteristics such as difficulty in data acquisition, limited sample data, and strong noise interference. In practical applications, it is challenging to collect sufficient data for every underwater target, leading to an imbalanced underwater acoustic sample dataset. This imbalance often causes trained deep neural networks to fail to achieve their training performance during testing and application. The generative model CGAN network effectively avoids the overfitting or underfitting problems of subsequent recognition models and effectively focuses on information from the less represented categories in the data samples.

[0009] In conclusion, how to combine CGAN networks with neural networks to suppress dataset imbalance in underwater acoustic target recognition is a topic worthy of exploration by technical personnel. Summary of the Invention

[0010] The purpose of this invention is to address the problem of poor training performance of deep learning models due to insufficient data collection for certain categories of underwater acoustic samples in existing technologies. This invention proposes a sample augmentation method to suppress dataset imbalance in underwater acoustic target recognition. Convolutional neural networks (CNNs) can effectively learn features and classify characteristics, and a CGAN network based on CNNs can be constructed. When the number of underwater acoustic samples of a certain category is insufficient to support the training of the recognition network, this CGAN network can be used to selectively generate a smaller number of data samples for that category, thus supplementing the dataset and achieving balance.

[0011] To achieve the above objectives, the technical solution provided by this invention is: a sample augmentation method for suppressing dataset imbalance in underwater acoustic target recognition, comprising the following steps:

[0012] Step 1: Classify the underwater acoustic signals with mismatched sample sizes according to their type information to form a dataset, and divide the dataset into training and testing sets;

[0013] Step 2: Frame and window the dataset obtained in Step 1, and perform Fast Fourier Transform within each frame to calculate the energy spectrum and obtain the time-frequency characteristics of the underwater acoustic signal;

[0014] Step 3: Construct the generator of the CGAN network, which consists of a three-layer convolutional neural network, with the LeakyReLU activation function between each layer. The formula for the LeakyReLU activation function is:

[0015]

[0016] Where 'a' represents the output of each layer of the neural network;

[0017] Step 4: Build the discriminator of the CGAN network. The discriminator consists of a single convolutional neural network layer with the sigmoid activation function, as shown in the formula:

[0018]

[0019] Where b is the output of each layer of the neural network, and the output of the Sigmoid function is between (0,1);

[0020] Step 5: Input the desired category labels along with normally distributed random noise into the generator;

[0021] Step 6: Input the spurious time-frequency features generated by the generator and the real time-frequency features obtained in Step 2 into the discriminator;

[0022] Step 7: Train the generator established in Step 3 and the discriminator established in Step 4;

[0023] Step 8: After training is complete, the generated fake samples of a specific type are merged into the dataset with a mismatch in the number of samples to balance the dataset.

[0024] Furthermore, in step 1, the underwater acoustic data is divided into N classes, and the N classes of underwater acoustic data together form a dataset; the dataset of the class with a sufficient number of samples is divided into training and test sets in a 7:3 ratio. From the class with an insufficient number of samples, a test set with a size equivalent to the test set of the class with a sufficient number of samples is created, and the remaining data is used as the training set.

[0025] Furthermore, step 2 includes the following sub-steps:

[0026] Step 2.1: Sample the underwater acoustic dataset at a sampling rate of 5kHz-10kHz and normalize it. Divide the data into short time windows with a window size of 30-50 milliseconds and an overlap rate of 50%.

[0027] Step 2.2: Apply a Hamming window to each window and use Python library functions to apply a short-time Fourier transform to each window, converting the time-domain signal into a time-frequency domain signal;

[0028] Step 2.3: Use library functions to perform a short-time Fourier transform and extract the amplitude spectrum from the transform result; perform a square operation on the amplitude spectrum to obtain the energy spectrum;

[0029] Step 2.4: Normalize the energy spectrum characteristics to obtain the time-frequency characteristics of the underwater acoustic signal.

[0030] Furthermore, in step 3, the three-layer convolutional neural network consists of three convolutional layers sandwiched between two batch normalization layers; the first convolutional layer is a one-dimensional convolutional layer with 1 input channel, 64 output channels, a kernel size of 4, and a stride of 4; followed by a one-dimensional batch normalization layer with 64 feature channels, then another one-dimensional convolutional layer with 64 input channels, 128 output channels, a kernel size of 4, and a stride of 4; then a one-dimensional batch normalization layer with 128 feature channels; the third convolutional layer is a one-dimensional convolutional layer with 128 input channels, 800 output channels, a kernel size of 4, and a stride of 4.

[0031] Furthermore, in step 4, a convolutional neural network discriminator consists of a dropout layer, a one-dimensional convolutional layer, a fully connected layer, and a batch normalization layer; the one-dimensional convolutional layer has 2 input channels, 16 output channels, a kernel size of 64, and a stride of 1; the fully connected layer has 16 input features and 4 output features; followed by a one-dimensional batch normalization layer with 16 feature channels.

[0032] Furthermore, in step 7, the training method includes the following sub-steps:

[0033] Step 7.1: z is random noise that follows a normal distribution. X is generated by the generator. fake The discriminator is responsible for determining whether the input data is the generated sample X. fake Or the real sample X real ;

[0034] Step 7.2: The objective function for optimization is as follows:

[0035]

[0036] Where V(D,G) represents the degree of difference between the real sample and the generated sample, and the metric is the cross-entropy loss, which is expressed by the formula:

[0037]

[0038] Where N represents the number of classes of samples to be classified, x i y represents the i-th element of the real label. i This represents the probability that the model predicts x belongs to the i-th class; max D V(D,G) is a fixed generator G, and the discriminator is repeatedly trained to maximize its ability to distinguish whether a sample comes from real data or generated data. Let L = max D V(D,G),min G L is the generator G obtained under the condition of a fixed discriminator D, and G is required to minimize the difference between the real sample and the generated sample.

[0039] Furthermore, the generator and discriminator are trained alternately, that is, the discriminator D is trained first, and then the generator G is trained.

[0040] Furthermore, the total number of training rounds is 2000-3000, with the learning rate set to 0.0015 for the first 1000-1500 rounds and 0.0012 for the last 1000 rounds.

[0041] Beneficial effects

[0042] The advantages of this invention are:

[0043] (1) This invention solves the problem of class imbalance in underwater acoustic target recognition by generative means. It is the first to apply CGAN network to the field of underwater acoustic target recognition. CGAN generates time-frequency features of underwater acoustic samples instead of directly generating underwater acoustic samples. Experiments show that the effect of generating time-frequency features is much better than directly generating samples.

[0044] (2) This invention fully utilizes the learning and decision-making performance of convolutional neural networks and applies convolutional neural networks to the generator and discriminator of the CGAN network. The generator uses a three-layer convolutional neural network, which can learn as many features as possible in a small number of samples. Based on the learned features, it generates a small number of class samples. The discriminator uses a one-layer convolutional neural network structure. In each training, it judges the authenticity of the time-frequency features generated by the generator and the time-frequency features of the real samples and feeds the judgment results back to the generator to encourage the generator to generate higher quality time-frequency feature samples. In this CGAN network, the input is replaced by the time-frequency features of the underwater sound samples, which enables the generator to learn data features better. In addition, the network can autonomously define the category of the generated data and supplement the dataset in a targeted manner to solve the problem of low overall recognition accuracy caused by the imbalance of the data in the dataset, where the proportion of a certain type of data is too small.

[0045] (3) According to the appendix Figure 2 It can be concluded that when using CNN as the recognition network alone, the recognition accuracy is 83%; after supplementing the samples with CGAN and then using CNN for recognition, the accuracy reaches 93%. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the use of a CGAN network to suppress the imbalance problem in underwater acoustic target recognition and the validation using a CNN in this embodiment.

[0047] Figure 2 To supplement the recognition accuracy of the dataset before and after the sample

[0048] Figure 3 CGAN network architecture diagram for the problem of imbalanced underwater acoustic datasets Detailed Implementation

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0050] See Figures 1-3 The technical solution of the present invention includes the following steps:

[0051] Step 1: Assign corresponding category labels to the underwater acoustic data to form a dataset, and then divide the dataset into training and testing sets.

[0052] Step 2: Framing and Windowing the Underwater Acoustic Signal. Since the frequency of the underwater acoustic signal changes in real time, it can only be processed by framing. Fourier transforms are performed within each frame, and the frequency profile of the signal is obtained by connecting adjacent frames. After framing the signal, a window function (Hamming window) is applied to each frame to counteract the assumption of infinite data in the FFT and reduce spectral leakage. Specifically, a Fast Fourier Transform (FFT) is performed on the sample data, and the energy spectrum is calculated to obtain the time-frequency characteristics. Since it is difficult to observe signal characteristics in the time domain, it is usually converted to an energy distribution in the frequency domain for observation. Different energy distributions represent different acoustic signal characteristics. Therefore, after windowing, each frame must undergo another FFT to obtain the energy distribution in the spectrum. FFTs are performed on each frame of the windowed signal to obtain the spectrum of each frame, and the energy spectrum of the signal is obtained by taking the square of the modulus of the acoustic signal's spectrum, thus obtaining the time-frequency characteristics of the acoustic signal.

[0053] Step 3: Build the CGAN network generator. The main structure of the generator consists of three convolutional neural network layers, using the LeakyReLU activation function between each layer. The formula for the LeakyReLU activation function is:

[0054]

[0055] Where 'a' represents the output of each layer of the neural network. This loss function speeds up the convergence of the network while solving the problem that the commonly used ReLU activation function, when the gradient is less than 0, has zero and no longer responds to any data, causing those parameters to never be updated.

[0056] The convolutional neural network structure generates time-frequency features of underwater acoustic samples based on the input category labels and random noise that conforms to a normal distribution. However, the generated time-frequency features are still quite different from the time-frequency features of real underwater acoustic data. Feedback from the discriminator and loss function is needed to continue iterative training to obtain time-frequency features similar to real underwater acoustic data.

[0057] Step 4: Build the discriminator of the CGAN network, which consists of a single convolutional neural network layer. The activation function is the sigmoid activation function, with the following formula:

[0058]

[0059] Where b is the output of each layer of the neural network. The output of the Sigmoid function is between (0,1), with a limited output range, stable optimization, and meets the output requirements of the discriminator.

[0060] Step 5: Input the desired category labels along with normally distributed random noise into the generator.

[0061] Step 6: Input the spurious time-frequency features generated by the generator and the real time-frequency features in the dataset into the discriminator.

[0062] Step 7: Train the generator and discriminator. z is random noise (Gaussian noise) that follows a normal distribution. X is generated by the generator. fake The discriminator is responsible for determining whether the input data is the generated sample X. fake Or the real sample X real The objective function for optimization is as follows:

[0063]

[0064] V(D,G) represents the degree of difference between the real sample and the generated sample, and the metric is the cross-entropy loss, which is expressed by the formula:

[0065]

[0066] Where C represents the number of sample categories, x i y represents the i-th element of the real label. i This represents the probability that the model predicts x belongs to the i-th class. (max) D V(D,G) fixes the generator G, aiming to maximize the discriminator's ability to distinguish between samples from real and generated data. Let L = max D V(D,G),min G L is used to obtain the generator G under the condition of a fixed discriminator D. This G is required to minimize the difference between real samples and generated samples. The generator and discriminator are trained alternately, that is, the discriminator D is trained first, and then the generator G is trained, and this process is repeated continuously.

[0067] Step 8: After the CGAN network is trained, the generated fake samples of specific types are merged into the dataset with a mismatch in the number of samples to balance the dataset.

[0068] The sample supplementation method of the present invention will now be described in detail with reference to examples and accompanying drawings. In this example, a real acoustic dataset of underwater acoustic targets is used to verify the performance of the sample supplementation method of the present invention. This dataset is experimental data from Danjiangkou Lake, and the data includes three types of vessels.

[0069] Step 1: Construct a dataset containing three classes of data. The ratio of training set to test set for classes A and B is 7:3. Each class's training set contains 280 10-second data points, and each test set contains 120 10-second data points. For class C, the training set contains 70 10-second data points, and the test set is the same as classes A and B, containing 120 10-second data points.

[0070] Step 2: Sample the underwater acoustic data at an 8kHz sampling rate, normalize it, and divide the data into short-time windows of 50 milliseconds each with a 50% overlap. Apply a Hamming window to each window and use Python library functions to perform a Short-Time Fourier Transform (STFT) to convert the time-domain signal to the time-frequency domain. Use library functions (Librosa, NumPy, etc.) to perform the STFT operation. Extract the amplitude spectrum (i.e., the magnitude of the STFT result) from the STFT result, square the amplitude spectrum to obtain the energy spectrum. Then normalize the energy spectrum features to obtain the time-frequency characteristics of the underwater acoustic data.

[0071] Step 3: Build the generator. The generator structure consists of three convolutional layers sandwiched between two batch normalization layers. The first convolutional layer is a one-dimensional convolutional layer with 1 input channel, 64 output channels, a kernel size of 4, and a stride of 4. Next is a one-dimensional batch normalization layer with 64 feature channels, followed by another one-dimensional convolutional layer with 64 input channels, 128 output channels, a kernel size of 4, and a stride of 4. Then, another one-dimensional batch normalization layer with 128 feature channels. The third convolutional layer is a one-dimensional convolutional layer with 128 input channels, 800 output channels, a kernel size of 4, and a stride of 4. The LeakyReLU activation function is used (with a slope of 0.2 on the negative part). The formula for the LeakyReLU activation function is:

[0072]

[0073] Where 'a' represents the output of each layer of the neural network. This loss function accelerates the convergence speed of the network while solving the problem that the commonly used ReLU activation function, when the gradient is less than 0, becomes 0 and no longer responds to any data, causing those parameters to never be updated. The convolutional neural network structure generates time-frequency features of underwater acoustic samples based on the input class labels and normally distributed random noise. However, the generated time-frequency features are still quite different from the time-frequency features of real underwater acoustic data, requiring feedback from the discriminator and the loss function to continue iterative training to obtain time-frequency features similar to real underwater acoustic data.

[0074] Step 4: Build the discriminator. The discriminator consists of a dropout layer, a one-dimensional convolutional layer, a fully connected layer, and a batch normalization layer. The dropout layer is used to randomly deactivate the discriminator during training to prevent overfitting. The one-dimensional convolutional layer has 2 input channels, 16 output channels, a kernel size of 64, and a stride of 1. The fully connected layer has 16 input features and 4 output features. Next is a one-dimensional batch normalization layer with 16 feature channels. The activation function is the sigmoid activation function, with the formula:

[0075]

[0076] Here, b represents the output of each layer of the neural network. It is restricted to a value between 0 and 1. The processed result is returned as the output of the discriminator.

[0077] Step 5: Input the category label (C class) along with random noise that conforms to a normal distribution into the generator. In a specific embodiment, select the label of the category with missing numbers. For example, if category C is the least represented among categories A, B, and C, then select to generate category C to achieve the goal of balancing the number of categories A, B, and C.

[0078] Step 6: Input the spurious time-frequency features generated by the generator and the real time-frequency features in the dataset into the discriminator. It should be noted that spurious time-frequency features refer to generated, false time-frequency features of a specified category. "Spurious" is relative to "real." The underwater acoustic dataset used in this invention is real measured data; this one is not measured, hence the term "spurious." It's just a qualifier, not a proper noun.

[0079] Step 7: Train the generator and discriminator. z is random noise (Gaussian noise) that follows a normal distribution. X is generated by the generator. fake The discriminator is responsible for determining whether the input data is the generated sample X. fake Or the real sample X real The objective function for optimization is as follows:

[0080]

[0081] V(D,G) represents the degree of difference between the real sample and the generated sample, and the metric is the cross-entropy loss, which is expressed by the formula:

[0082]

[0083] Where 3 indicates that there are three classes of samples to be classified, x i y represents the i-th element of the real label. i This represents the probability that the model predicts x belongs to the i-th class. (max) D V(D,G) fixes the generator G, aiming to maximize the discriminator's ability to distinguish between samples from real and generated data. Let L = max D V(D,G),min G L is used to obtain the generator G under the condition of fixing the discriminator D. This G is required to minimize the difference between the real samples and the generated samples. The generator and discriminator are trained alternately, that is, the discriminator D is trained first, and then the generator G is trained. The total number of training rounds is 2000. The learning rate is set to 0.0015 for the first 1000 rounds and 0.0012 for the last 1000 rounds.

[0084] Step 8: After the CGAN network is trained, the generated C-class fake samples are merged into the dataset with a mismatch in the number of samples and then input into the CNN network for identification.

[0085] It should be noted that Convolutional Neural Networks (CNNs) are currently the most widely used deep learning algorithms. They can fully utilize the locality of data by combining three main characteristics: local receptive regions, shared weights, and spatial or temporal pooling downsampling. This allows for optimization of the network structure while maintaining a certain degree of shift invariance. This method selects a CNN as the classifier for underwater acoustic target recognition to verify the effectiveness of the CGAN network in solving the imbalanced dataset problem. However, verification methods are not limited to CNNs alone.

[0086] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A sample augmentation method for suppressing imbalanced datasets in underwater acoustic target recognition, characterized in that, Includes the following steps: Step 1: Classify the underwater acoustic signals with mismatched sample sizes according to their type information to form a dataset, and divide the dataset into training and testing sets; Step 2: Frame and window the dataset obtained in Step 1, and perform Fast Fourier Transform within each frame to calculate the energy spectrum and obtain the time-frequency characteristics of the underwater acoustic signal; Step 3: Construct the generator of the CGAN network, which consists of three layers of convolutional neural networks, and the LeakyReLU activation function between each layer is the LeakyReLU function; the formula for the LeakyReLU activation function is: Where 'a' represents the output of each layer of the neural network; Step 4: Build the discriminator of the CGAN network. The discriminator consists of a single convolutional neural network layer with the sigmoid activation function, as shown in the formula: Where b is the output of each layer of the neural network, and the output of the Sigmoid function is between (0,1); Step 5: Input the desired category labels along with normally distributed random noise into the generator; Step 6: Input the spurious time-frequency features generated by the generator and the real time-frequency features obtained in Step 2 into the discriminator; Step 7: Train the generator established in Step 3 and the discriminator established in Step 4; Step 8: After training is complete, the generated fake samples of a specific type are merged into the dataset with a mismatch in the number of samples to balance the dataset.

2. The sample augmentation method for suppressing dataset imbalance in underwater acoustic target recognition as described in claim 1, characterized in that, In step 1, the underwater acoustic data is divided into N categories, and the N categories of underwater acoustic data together form a dataset; whereby... For datasets with a sufficient number of samples in each category, the training and test sets are divided in a 7:3 ratio. In a class with insufficient sample size, a test set with a size equivalent to that in a class with sufficient sample size is created, and the remaining data is used as the training set.

3. The sample augmentation method for suppressing dataset imbalance in underwater acoustic target recognition as described in claim 1, characterized in that, Step 2 includes the following sub-steps: Step 2.1: Sample the underwater acoustic dataset at a sampling rate of 5kHz-10kHz and normalize it. Divide the data into short time windows with a window size of 30-50 milliseconds and an overlap rate of 50%. Step 2.2: Apply a Hamming window to each window and use Python library functions to apply a short-time Fourier transform to each window, converting the time-domain signal into a time-frequency domain signal; Step 2.3: Use library functions to perform a short-time Fourier transform and extract the amplitude spectrum from the transform result; perform a square operation on the amplitude spectrum to obtain the energy spectrum; Step 2.4: Normalize the energy spectrum characteristics to obtain the time-frequency characteristics of the underwater acoustic signal.

4. The sample augmentation method for suppressing dataset imbalance in underwater acoustic target recognition as described in claim 1, characterized in that, In step 3, the three-layer convolutional neural network consists of three convolutional layers sandwiched between two batch normalization layers. The first convolutional layer is a one-dimensional convolutional layer with 1 input channel, 64 output channels, a kernel size of 4, and a stride of 4. Next is a one-dimensional batch normalization layer with 64 feature channels, followed by a one-dimensional convolutional layer with 64 input channels, 128 output channels, a kernel size of 4, and a stride of 4. Then there is a one-dimensional batch normalization layer with 128 feature channels. The third convolutional layer is a one-dimensional convolutional layer with 128 input channels, 800 output channels, a kernel size of 4, and a stride of 4.

5. The sample augmentation method for suppressing dataset imbalance in underwater acoustic target recognition as described in claim 1, characterized in that, In step 4, the convolutional neural network discriminator consists of a dropout layer, a one-dimensional convolutional layer, a fully connected layer, and a batch normalization layer; the one-dimensional convolutional layer has 2 input channels, 16 output channels, a kernel size of 64, and a stride of 1; the fully connected layer has 16 input features and 4 output features. This is followed by a one-dimensional batch normalization layer with 16 feature channels.

6. The sample augmentation method for suppressing dataset imbalance in underwater acoustic target recognition as described in claim 1, characterized in that, In step 7, the training method includes the following sub-steps: Step 7.1: z is random noise that follows a normal distribution. X is generated by the generator. fake The discriminator is responsible for determining whether the input data is the generated sample X. fake Or the real sample X real ; Step 7.2: The objective function for optimization is as follows: Where V(D,G) represents the degree of difference between the real sample and the generated sample, and the metric is the cross-entropy loss, which is expressed by the formula: Where N represents the number of classes of samples to be classified, x i y represents the i-th element of the real label. i This represents the probability that the model predicts x belongs to the i-th class; max D V(D,G) is a fixed generator G, and the discriminator is repeatedly trained to maximize its ability to distinguish whether a sample comes from real data or generated data; let L = max D V(D,G),min G L is the generator G obtained under the condition of a fixed discriminator D, and G is required to minimize the difference between the real sample and the generated sample.

7. The sample augmentation method for suppressing dataset imbalance in underwater acoustic target recognition as described in claim 6, characterized in that, The generator and discriminator are trained alternately, that is, the discriminator D is trained first, and then the generator G is trained.

8. The sample augmentation method for suppressing dataset imbalance in underwater acoustic target recognition as described in claim 7, characterized in that, The total number of training rounds is 2000-3000. The learning rate is set to 0.0015 for the first 1000-1500 rounds and 0.0012 for the last 1000 rounds.

Images