A method for optimizing an underwater acoustic target recognition model based on a sample expansion network

Abstract

This invention proposes an optimization method for underwater acoustic target recognition models based on sample augmentation networks, belonging to the field of underwater acoustic target recognition model optimization. Addressing the problem of low classification accuracy due to insufficient target data, this invention utilizes a mask-based sample generation approach. While ensuring real-time model performance, it constructs two pairs of structurally symmetrical generators and discriminators to map source domain samples to the target domain. Experimental results show that this invention, by constructing a recurrent adversarial generative network using mask prompts, generates reliable real samples from the target domain while maintaining a clear model structure. These samples are then added to the training set to optimize the recognition model, thereby improving recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic target recognition model optimization, and specifically to an underwater acoustic target recognition model optimization method based on sample augmentation network. Background Technology

[0002] Sonar remains the most effective and popular method for long-range target detection because sound waves are signals with minimal propagation loss in water. Therefore, using sonar to collect underwater information for automatic underwater target identification remains the mainstream method. Passive sonar, as a form of underwater identification, only receives signals and does not transmit them, offering excellent stealth capabilities. With the increasing needs of national maritime strategies and defense, the intelligent processing and analysis of sonar signals remains a research hotspot in many countries, especially in unmanned environments such as the deep sea.

[0003] A fundamental requirement in current intelligent sonar signal processing methods is the use of massive amounts of data to train neural networks. However, the scarcity and difficulty in collecting underwater target data leads to insufficient sample data, severely impacting the recognition efficiency of neural networks and the performance of automated identification equipment. This issue has created a bottleneck in the integration of artificial intelligence into sonar signals. Therefore, given the limited sonar signal acquisition, how to incorporate prior knowledge to expand the sample size while maintaining high classification accuracy for identification equipment remains a highly challenging problem.

[0004] Currently, researchers worldwide are dedicating themselves to in-depth research in this area. While we started relatively late, our potential for development is enormous. Traditional underwater acoustic target recognition methods, such as line spectrum analysis and matched filtering, do not require large amounts of sonar signal data. Although these methods have achieved good recognition results, they have extremely high requirements for the sonar's signal-to-noise ratio, exhibit significant differences in their ability to distinguish different types of acoustic signals, and have slow detection speeds. Therefore, they are difficult to apply to real-time scenarios with complex noise levels. However, by utilizing sample augmentation networks and introducing prior knowledge to augment a small amount of underwater acoustic data, the performance of the underwater acoustic target recognition model can be improved through adversarial methods. Simultaneously, the augmented data can also be used to improve the realism of ship noise samples. The task of augmenting underwater acoustic target samples using recurrent generative adversarial networks (GANs) mainly involves four steps: First, MFCC spectral feature extraction is performed on the underwater acoustic samples. Second, two generators and four discriminators are trained using the sample features with known labels. Third, the model generated in the second step is used to generate target domain samples from given source domain samples, and audio files are obtained through the MFCC spectral feature decoder to test the model. Fourth, the model is used for target classification, i.e., the data generated in the third step is added to the sample set for downstream recognition model training. However, the limited number of real samples in passive underwater acoustic target recognition tasks makes the above method difficult to implement effectively. Summary of the Invention

[0005] To address the problem of insufficient real samples in passive underwater acoustic target recognition tasks, this invention provides an optimization method for underwater acoustic target recognition models based on sample augmentation networks.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An optimization method for an underwater acoustic target recognition model based on a sample augmentation network includes the following steps:

[0008] (1) Use the Mel filter bank to extract the Mel frequency of the target ship's acoustic signal and calculate the cepstrum;

[0009] (2) Construct two generators that transform between source domain samples and target domain samples and four corresponding discriminators, and train them using source domain samples and target domain samples with known labels; the two generators are XY generator and YX generator;

[0010] (3) After the model training is completed, the Mel frequency cepstral coefficient spectrum features of the existing target domain samples are used to generate the Mel frequency cepstral coefficient spectrum features of the new samples, and the Mel frequency cepstral coefficient spectrum feature decoder is used to decode the new time domain samples and the corresponding audio files.

[0011] (4) Add the new time-domain samples obtained in step (3) to the sample set of the target domain to optimize the training of the recognition model, thereby improving the accuracy and robustness of the recognition model.

[0012] Furthermore, the specific method of step (1) is as follows:

[0013] (101) Obtain the Mel frequency cepstral coefficient spectral characteristics of the audio signals of the source domain sample X and the target domain sample Y;

[0014] (102) Normalize and batch process the characteristics of the Mel frequency cepstral coefficient spectrum to obtain the cepstral spectrum.

[0015] Furthermore, the specific method of step (2) is as follows:

[0016] (201) Generate a mask X with the same dimension as the source domain sample X and a mask Y with the same dimension as the target domain sample Y. Generate a target domain sample Y1 from the source domain sample X and the mask X using the XY generator. Generate a loop sample X2 from Y1 using the YX generator. Generate a source domain sample X1 from the target domain sample Y and the mask Y using the YX generator. Generate a loop sample Y2 from X1 using the XY generator. Generate an identity sample X3 from the source domain sample X using the YX generator. Generate an identity sample Y3 from the target domain sample Y using the XY generator.

[0017] (202) Using four discriminators, X1, Y1, X2, and Y2 are discriminated respectively to obtain the corresponding loss functions; then the distances between X and X3, and between Y and Y3 are calculated respectively.

[0018] (203) Fix the parameters of the discriminator, and use the loss function and distance obtained in step (202) to update the parameters of the generator through the Adam optimizer;

[0019] (204) Regenerate X1, Y1, X2, Y2, and then use the corresponding discriminator to discriminate X, Y and the regenerated X1, Y1, X2, Y2 again to obtain the corresponding loss function;

[0020] (205) Fix the parameters of the generator and use the loss function obtained in step (204) to update the parameters of the discriminator through the Adam optimizer.

[0021] Furthermore, the specific method of step (3) is as follows:

[0022] (301) Input sample X into the trained generator XY to generate the corresponding fake Y;

[0023] (302) Obtain the audio file of fake Y by using the Mel frequency cepstral coefficient spectrum feature decoder.

[0024] Furthermore, the specific method of step (4) is as follows:

[0025] (1) Test the sample generation effect of the audio file obtained in step (302) using the pre-trained classifier;

[0026] (2) Add the audio file obtained in step (302) to the sample set of the target domain for training the classifier and improving the accuracy of the classifier.

[0027] The beneficial effects of this invention are as follows:

[0028] 1. This invention extracts MFCC spectral features from ship noise, optimizes model training, and uses a recurrent generative adversarial network to learn the features, thereby enabling the generator in the network to generate MFCC spectral features similar to the required categories.

[0029] 2. This invention can generate the required acoustic signal and produce real samples through inverse MFCC transform. Specifically, this invention adds the new samples obtained by inverse MFCC transform to the sample set of the target domain for training of downstream recognition models, thereby improving the recognition accuracy of deep neural networks. Attached Figure Description

[0030] Figure 1 This is a schematic diagram illustrating the principle of an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram illustrating the extraction of MFCC spectral features in an embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram illustrating the construction of the generator and discriminator in an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0034] An optimization method for an underwater acoustic target recognition model based on a sample augmentation network includes the following steps:

[0035] S1, Mel Frequency Cepstrum Coefficient (MFCC) Spectral Feature Extraction: After extracting the Mel frequency of the target ship's acoustic signal using the MFCC spectral feature filter bank, the cepstrum is calculated.

[0036] S2, Model Training Optimization: Build two generators (XY generator and YX generator) that transform between source domain samples and target domain samples, and four corresponding discriminators (i.e., 2*2), and train them using source domain samples and target domain samples with known labels.

[0037] S3, Generate audio file: After the model training is completed, the MFCC spectrogram features of the existing target domain samples are used to generate the MFCC spectrogram features of the new samples, and the new time domain samples and corresponding audio files are obtained by decoding them through the MFCC spectrogram feature decoder.

[0038] S4, Model Testing Application: The new time-domain samples obtained in S3 are added to the target domain sample set to optimize the training of the recognition model, thereby improving the accuracy and robustness of the recognition model.

[0039] Step S1 includes the following steps:

[0040] S11: Calculate the MFCC spectral features of the input audio signals from the source domain X and the target domain Y;

[0041] S12: Normalize and batch process the MFCC spectral features.

[0042] Step S2 includes the following steps:

[0043] S21: Generate Y1 (i.e., fake Y) from sample X and mask X using the XY generator, and then generate cyclic sample X2 from fake Y using the YX generator; Generate X1 (i.e., fake X) from sample Y and mask Y using the YX generator, and then generate cyclic sample Y2 from fake X using the XY generator; In addition, generate identity sample X3 from X using the YX generator, and generate identity sample Y3 from Y using the XY generator;

[0044] S22: Using four discriminators, X1, Y1, X2, and Y2 are judged respectively to obtain the corresponding loss functions; then the distance between X and X3, and between Y and Y3 is calculated.

[0045] S23: With the discriminator parameters fixed, use the loss function and distance obtained in S22 to update the generator parameters through the Adam optimizer;

[0046] S24: After regenerating X1, Y1, X2, and Y2, use the corresponding discriminators to distinguish X, Y, and the regenerated X1, Y1, X2, and Y2 respectively to obtain the corresponding loss functions;

[0047] S25: Fix the parameters of the generator, use the loss function obtained in S24, and update the parameters of the discriminator through the Adam optimizer.

[0048] Step S3 includes the following steps:

[0049] S31: Using the trained generator XY, generate the corresponding fake Y from sample X;

[0050] S32: Obtain the audio file of fake Y through the MFCC spectrogram decoder.

[0051] Step S4 includes the following steps:

[0052] S41: Test the sample generation effect of the audio file obtained in S32 using the pre-trained recognition model;

[0053] S42: Add the audio file obtained in S32 to the sample set of the target domain to train the recognition model and improve the accuracy of the recognition model.

[0054] Here is a more specific example:

[0055] like Figure 1 As shown, an optimization method for an underwater acoustic target recognition model based on a sample augmentation network includes the following steps:

[0056] Step S1, MFCC spectral feature extraction:

[0057] The MFCC spectral characteristics are Mel frequency cepstral coefficients. Mel frequencies are proposed based on the characteristics of human hearing and have a non-linear relationship with Hz frequencies. Mel frequency cepstral coefficients (MFCC) are the Hz spectral characteristics calculated using this relationship.

[0058] S11: Simultaneously with obtaining the audio energy spectrum, a Mel filter bank needs to be constructed, and its dot product with the energy spectrum is performed to obtain the Mel spectrogram. The purpose of the Mel filter bank is to convert the energy spectrum into Mel frequencies that are closer to human ear perception. In this method, the number of Mel filters is set to 128. The filtering process can be represented by the following formula:

[0059]

[0060] Where N represents the total number of signal points in each frame, f represents the frequency point, and H m (f) represents the coefficients of the Mel filter bank, p(f) is the amplitude of the signal, and E(m) is the Mel frequency;

[0061] S12: For very faint sounds, the human ear can perceive even a slight increase in loudness. However, once the loudness has reached a certain level, even a significant increase will not noticeably change the human ear's perception. Simulating the "logarithmic" characteristic of the human ear's response to sound loudness, taking the logarithm of the Mel-frequency spectrum yields:

[0062] E′(m)=lgE(m)

[0063] S13: For each frame of data, DCT is used to change the data distribution, concentrating most of the signal data in the low-frequency region. The first 64 low-frequency data points after transformation are taken as the MFCC spectral features of that frame. The 64 features of each frame in a single signal sample are concatenated over time to obtain a 64*64 MFCC spectral feature. To facilitate subsequent processing by the classification model, the last frame in the time dimension and the first frame in the feature dimension are copied to obtain the final 80*64 MFCC spectral feature, as shown below. Figure 2 As shown.

[0064] S14: After extracting the MFCC spectral features, calculate the batch mean and batch standard deviation for each class of samples, and batch normalize the MFCC spectral features:

[0065]

[0066] In the formula, x scale The first formula represents the normalized features, where x is the sample, μ is the sample mean, S is the sample variance, and N is the total number of samples. The second formula calculates the average of all samples, and the third formula calculates the average of each sample after subtracting the square of the sample mean.

[0067] Step S2, Model Training Optimization:

[0068] The model of this method consists of two generators and four discriminators, with the following structure: Figure 3 As shown, this is quite complex and requires various constraints to train the network model parameters. Our method designs the following loss function:

[0069]

[0070] On the right side of the equation, the first two terms are two different adversarial losses, the third term is the identity mapping loss, and the fourth term is the cycle consistency loss. During training, these four types of losses are combined according to their weights as the final optimization function to ensure that the model learns in the direction of the target and gradually converges.

[0071] S21: For samples of class X, generate a mask X (80*64) of the corresponding dimension. One column of mask X is a vector of all 1s, and the remaining columns are vectors of all 0s. Combine the X samples and mask X with the XY generator to generate the corresponding Y1 (80*64). Then combine Y1 and the all-1 mask with the YX generator to generate X2. Finally, combine X with the YX generator to generate X3. Perform the same operation for samples of class Y to generate X1, Y2, and Y3 (all of size 84*64). The formula for combining the mask and the sample is as follows:

[0072]

[0073] In the formula, x represents the sample and m represents the mask. The combined sample is obtained by multiplying corresponding matrix elements.

[0074] S22: Using four discriminators, X1, Y1, X2, and Y2 are discriminated against to obtain adversarial loss and cycle-consistent loss. Then, the L1 distance between X and X3, and between Y and Y3 are calculated. m and m′ represent the random missing mask and the all-1 mask, respectively. x and y are real samples from the source domain and the target domain, respectively.

[0075]

[0076] Among them, D Y (y) represents the probability that the discriminator determines that the real sample in the target domain comes from the target domain; This is the probability that the discriminator will classify a generated sample as coming from the source domain; the sum of these probabilities is the probability obtained. The adversarial loss value of the discriminator. Consistent with its form.

[0077]

[0078] Among them, D′ X (x) represents the probability that the discriminator determines that the true sample in the source domain comes from the source domain, 1-D′ X (G Y→X (concat(y′,m′))) is the probability that the discriminator will classify the generated sample as coming from the target domain; the sum of these probabilities is the probability that the discriminator will classify the sample as coming from the target domain. This represents the cycle-consistent loss value of the model. Consistent with its form.

[0079]

[0080] Among them, G Y→X (G X→Y (concat(x, m′))) represents the sample obtained after passing through two generators from the source domain sample. The distance between the sample and the original sample is calculated using the 1-norm of the vector, which is the sum of the absolute values ​​of the differences between each element. The identity adversarial loss value for the model. Consistent with its form.

[0081]

[0082] in, This represents the sample obtained after the target domain sample has passed through the XY generator. The identity adversarial loss value for the model. Consistent with its form.

[0083] S23: After calculating the loss and distance in S22, fix the parameters of the discriminator and update the parameters of the generator through the Adam optimizer;

[0084] m t =μ*m t-1 +(1-μ)*g t

[0085]

[0086]

[0087]

[0088]

[0089] Among them, g t Let m be the gradient. t For the current g t The mean of each component, n t For the current g t The variance of each component, For the corrected m t , For the corrected n t μ t To The correction factor, ν t To The correction coefficients are used. The first two formulas are the first and second moment estimates of the gradient, respectively; the third and fourth formulas are corrections to the first and second moment estimates. The last formula is a dynamic constraint on the learning rate n, and it has a defined range.

[0090] S24: After updating the generator parameters, regenerate false A, false B, loop A, and loop B, and recalculate the loss of S22.

[0091]

[0092] S25: Fix the parameters of the generator and update the parameters of the discriminator through another Adam optimizer.

[0093] Step S3, generate audio file:

[0094] After the source domain feature map is transformed into the target domain feature map through the model, in order to verify the validity of the generated audio file, it is necessary to restore it to an audio file through the MFCC spectrogram feature decoder.

[0095] S31: Using the generator AB trained in S2, generate the required MFCC spectral features of sample A for the pseudo B.

[0096] S32: Obtain the audio file of fake B through the MFCC spectrogram feature decoder;

[0097] h (t) =f(h (t-1) y t-1 c)

[0098] P(y t |y t-1 y t-2 ,...,y1,c)=g(h (t) y t-1 c)

[0099] Where f is the activation function, g is the decoder function, P is the generated audio file, and h is the MFCC spectrogram.

[0100] Step S4, Model Testing and Application:

[0101] The authenticity of the audio files generated by S3 was tested using an unoptimized recognition model; the generated audio files were then added to the training set to verify the effectiveness of the sample augmentation model. The specific method is as follows:

[0102] S41: Test the sample generation effect of the audio file obtained in S32 using the recognition model. The recognition model classifies it into the corresponding class with a probability of 100%.

[0103]

[0104] In the formula, exp(f y ) is the output value of the current node. It is the sum of the outputs of all nodes.

[0105] S42: Add the audio file obtained in S32 to class B, retrain the recognition model, and improve the accuracy of the recognition model from 93% to 97%.

[0106] In summary, this invention addresses the problem of low classification accuracy caused by insufficient target data. Utilizing a mask-based sample generation approach, it constructs two pairs of structurally symmetrical generators and discriminators to map source domain samples to the target domain, while maintaining real-time model performance. Experimental results demonstrate that this invention, through the mask-based approach to construct a recurrent adversarial generative network, generates reliable, realistic samples from the target domain while ensuring a clear model structure. These samples are then added to the training set to optimize the recognition model, thereby improving recognition accuracy.

[0107] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A method for optimizing an underwater acoustic target recognition model based on a sample augmentation network, characterized in that, Includes the following steps: (1) The Mel frequency of the target ship's acoustic signal is extracted using the Mel filter bank, and the cepstrum is calculated; the specific method is as follows: (101) Obtain the characteristics of the Mel frequency cepstral coefficients of the audio signals of the source domain sample X and the target domain sample Y; (102) Normalize and batch process the characteristics of the Mel frequency cepstral coefficients to obtain the cepstral spectrum; (2) Construct two generators that transform between source domain samples and target domain samples, and four corresponding discriminators, trained using source domain samples and target domain samples with known labels; the two generators are an XY generator and a YX generator; the specific method is as follows: (201) Generate a mask X with the same dimension as the source domain sample X and a mask Y with the same dimension as the target domain sample Y. Generate a target domain sample Y1 from the source domain sample X and the mask X using the XY generator. Generate a cyclic sample X2 from Y1 using the YX generator. Generate a source domain sample X1 from the target domain sample Y and the mask Y using the YX generator. Generate a cyclic sample Y2 from X1 using the XY generator. Generate an identity sample X3 from the source domain sample X using the YX generator. Generate an identity sample Y3 from the target domain sample Y using the XY generator. (202) Using four discriminators, X1, Y1, X2, and Y2 are discriminated respectively to obtain the corresponding loss functions; then the distances between X and X3, and between Y and Y3 are calculated respectively. (203) Fix the parameters of the discriminator, and use the loss function and distance obtained in step (202) to update the parameters of the generator through the Adam optimizer; (204) Regenerate X1, Y1, X2, Y2, and then use the corresponding discriminator to discriminate X, Y and the regenerated X1, Y1, X2, Y2 again to obtain the corresponding loss function; (205) Fix the parameters of the generator and use the loss function obtained in step (204) to update the parameters of the discriminator through the Adam optimizer; (3) After the model training is completed, the Mel frequency cepstral coefficient spectrum features of the existing target domain samples are used to generate the Mel frequency cepstral coefficient spectrum features of the new samples, and then decoded by the Mel frequency cepstral coefficient spectrum feature decoder to obtain the new time-domain samples and the corresponding audio files; the specific method is as follows: (301) Input sample X into the trained generator XY to generate the corresponding fake Y; (302) Obtain the audio file of fake Y by using the Mel frequency cepstral coefficient spectrum feature decoder; (4) Add the new time-domain samples obtained in step (3) to the sample set of the target domain to optimize the training of the recognition model, thereby improving the accuracy and robustness of the recognition model; the specific method is as follows: (1) Test the sample generation effect of the audio file obtained in step (302) using a pre-trained classifier; (2) Add the audio file obtained in step (302) to the sample set of the target domain for training the classifier and improving the accuracy of the classifier.

Images