A method for imitating underwater acoustic signals based on MelGAN

Abstract

The application provides a method for imitating underwater acoustic signals based on MelGAN. The method comprises the following steps: using MelGAN as an underwater acoustic signal imitation model and training; wherein the kernel size of the hole convolution in the generator is set to an integer multiple of the step size, and the frame shift of the sampling in the generator is set to 256; the training process specifically comprises the following steps: inputting a Mel spectrogram into the generator to generate an imitated underwater acoustic signal by learning the distribution of a real underwater acoustic signal; inputting the imitated underwater acoustic signal and the real underwater acoustic signal into a discriminator to make the MelGAN model converge by using a loss function, and the converged MelGAN model is the underwater acoustic signal imitation model; and inputting a test sample into the trained generator to obtain an imitated underwater acoustic signal. The imitation effect is evaluated by using two methods of listening identification and spectrogram identification, the rhythm of the imitated underwater acoustic signal is similar to that of the real signal, and the imitation effect is good.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of underwater acoustic communication, and in particular to an underwater acoustic signal impersonation method based on MelGAN. BACKGROUND

[0002] On the sea, ship detection is mainly carried out through underwater acoustic signals, and in order to enhance the concealment ability of ships, underwater acoustic signal impersonation technology needs to be used. With the rapid development of artificial intelligence technology, the application of underwater acoustic signal impersonation technology combined with artificial intelligence is required by the times.

[0003] However, the current traditional deep neural network is supervised learning, and a large amount of sample data with good signal quality is required for training the model, and it is very time-consuming and laborious, but the underwater acoustic channel is very unstable and is greatly affected by temperature, ocean current and ocean density, and has serious multipath fading, so the underwater acoustic signals measured in the actual environment generally have small signal-to-noise ratio and serious noise interference, and the signal quality is very poor, which is difficult to meet the training standard of the traditional deep neural network. Therefore, the effect of using traditional deep learning technology to impersonate underwater acoustic signals is not ideal. SUMMARY

[0004] In order to improve the impersonation effect of underwater acoustic signals, the present application provides an underwater acoustic signal impersonation method based on MelGAN.

[0005] The underwater acoustic signal impersonation method based on MelGAN provided by the present application comprises:

[0006] Step 1: using MelGAN as an underwater acoustic signal impersonation model and training; wherein the kernel size of the hole convolution in the generator is set to an integer multiple of its step size, and the frame shift of the sampling in the generator is set to 256; the training process specifically comprises: inputting the Mel spectrogram into the generator to generate impersonated underwater acoustic signals by learning the distribution of the real underwater acoustic signals; inputting the impersonated underwater acoustic signals and the real underwater acoustic signals into the discriminator, and using the loss function to make the MelGAN model converge, and the converged MelGAN model is the underwater acoustic signal impersonation model;

[0007] Step 2: inputting the test sample into the trained generator to obtain the impersonated underwater acoustic signal.

[0008] Further, the size of the overlapping area is controlled by increasing or decreasing the number of layers of the hole convolution in the generator.

[0009] Further, the Hinge loss function is used as the loss function.

[0010] Further, the real underwater acoustic signal is preprocessed before being inputted into the discriminator; the preprocessing comprises: frequency reduction and pre-emphasis.

[0011] The beneficial effects of the present application are as follows:

[0012] The present application provides a MelGAN-based underwater acoustic signal impersonation method, which first performs frequency reduction processing on real underwater acoustic signal samples, then performs pre-emphasis processing on the real signal samples to obtain underwater acoustic signals with obvious characteristics, and then converts them into Mel spectrogram format and inputs them into the MelGAN model for training to obtain an underwater acoustic signal impersonation model, and finally uses the trained model to impersonate underwater acoustic signals. The present application can produce fake impersonated underwater acoustic signals with good results. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A flowchart of a MelGAN-based underwater acoustic signal impersonation method provided by an embodiment of the present application is shown in the figure.

[0014] Figure 2 A comparison diagram of the spectrum of impersonated underwater acoustic signals and real underwater acoustic signals provided by an embodiment of the present application is shown in the figure.

[0015] Figure 3 A comparison diagram of the spectrogram of impersonated underwater acoustic signals and real underwater acoustic signals provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0016] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0017] As shown in the figure, the present application provides a MelGAN-based underwater acoustic signal impersonation method, which includes the following steps: Figure 1

[0018] S101: Use MelGAN as an underwater acoustic signal impersonation model and train it; wherein the kernel size of the hole convolution in the generator is set to an integer multiple of its step size, and the frame shift of the sampling in the generator is set to 256; the training process specifically includes: inputting the Mel spectrogram into the generator to generate impersonated underwater acoustic signals by learning the distribution of real underwater acoustic signals; inputting the impersonated underwater acoustic signals and the real underwater acoustic signals into the discriminator, and using the loss function to make the MelGAN model converge, and the converged MelGAN model is the underwater acoustic signal impersonation model;

[0019] ​Specifically, considering that the underwater acoustic signal generation is trained using the intermediate format Mel spectrogram, and the underwater acoustic signal is converted into the Mel spectrogram, which is a lossy compression process, therefore, there must be distortion in the process of inversely mapping back to the underwater acoustic signal from the Mel spectrogram; therefore, in order to reduce the format conversion between data and reduce unnecessary quality loss, in the embodiment of the present application, when training the underwater acoustic signal impersonation model, the input of the generator of MelGAN does not use random noise input conforming to the Gaussian distribution, but uses the Mel spectrogram; and in fact, the inventors have also found through experiments that if the input of the generator of MelGAN still inputs random noise conforming to the Gaussian distribution like the traditional GAN model, the generated underwater acoustic signal effect will be poor. In addition, in the embodiment, the frame shift of the sample is set to 256, so the time resolution of the Mel spectrogram is 1 / 256 of the underwater acoustic signal, so the entire generator needs to be up-sampled by 256x to make the generated underwater acoustic signal consistent with the time resolution of the real underwater acoustic signal.

[0020] Further, in order to reduce the amount of data and facilitate network modeling, and to remove the environmental or other marine biological noise in the ship-radiated noise signal, in the embodiment of the present application, the real underwater acoustic signal is preprocessed before being input to the discriminator; the preprocessing includes: frequency reduction and pre-emphasis. Through the preprocessing operation, the extracted underwater acoustic signal features can be more obvious and effective.

[0021] Further, one of the difficulties in the generation process of underwater acoustic signals is to make the underwater acoustic signals only have correlation in the time scale, in order to solve this difficulty, the original structure design of the MelGAN network is to make the Mel spectrogram first pass through the up-sampling and then pass through the residual block, this structure design actively introduces the inductive bias, so that there is an overlapping area between the signal frames, thereby producing correlation. However, the size of the overlapping area is closely related to the quality of the generated underwater acoustic signal, in order to be able to improve the quality of the generated underwater acoustic signal, the inventors found that the size of the overlapping area can be controlled by increasing or decreasing the number of layers of the hollow convolution to affect the signal quality, but it is not that the more the number of layers is, the better, experiments have proved that when the number of layers is too much, the signal quality is actually worse, and the training time is very long, so the selection of the number of layers should be determined according to the actual situation.

[0022] In addition, in order to solve the problem that harsh squeaking sound often appears in the signal when starting to generate underwater acoustic signals, the kernel size and step length of the hollow convolution are adjusted in the embodiment, specifically: the kernel size is set to an integer multiple of the step length.

[0023] As shown in Figure 2 The discriminator in MelGAN adopts a multi-scale architecture with three discrimination modules, and the three discrimination modules jointly discriminate the signal to improve the discrimination ability. The network structures of the three discrimination models are exactly the same, and each is composed of three convolution layers and four 4x down-sampling layers.

[0024] In the present application, each discriminant module is used to discriminate underwater acoustic signals of different resolutions. The first discriminant module is used to discriminate the original underwater acoustic signal; the second and third discriminant modules are respectively used to discriminate the signals after 2-fold and 4-fold pooling sampling of the original underwater acoustic signal, and the kernel of the pooling layer is 4.

[0025] When the discriminator works, each discriminant module learns the data distribution of the small signal block for classification, so that the discriminator can learn the data distribution of the entire underwater acoustic signal for classification. The discriminator uses the method of grouped convolution to well learn the high-frequency data in the underwater acoustic signal, and can also learn underwater acoustic signal sequences of different lengths. The network layer of the discriminator is similar to that of the generator. All network layers in the discriminator also use weight normalization, and the activation function is LeakyReLU, and the negative interval gradient is set to 0.2. The optimizer is Adam, and the learning rate is 0.0001.

[0026] As an implementable manner, the present embodiment uses Hinge loss as the loss function, as shown in formula (1):

[0027]

[0028] Wherein, x represents the real underwater acoustic signal, z represents the input of the generator, and i represents the discriminant module. A feature matching loss is also added, which minimizes the L1 distance between the underwater acoustic signal sample and the generated underwater acoustic signal in the discriminator.

[0029] S102: input the test sample into the trained generator to obtain the counterfeit underwater acoustic signal.

[0030] The underwater acoustic signal counterfeiting method based on MelGAN provided by the present application innovatively applies the MelGAN generative adversarial network to the field of counterfeit generation of underwater acoustic signals, and evaluates the counterfeiting effect through two methods of listening identification and spectrogram identification. The rhythm of the counterfeit underwater acoustic signal is similar to that of the real signal. By comparing the spectrograms (such as Figure 2 and Figure 3 ) of the two, it can be seen that the signal energy of both is mainly between 2500Hz-8000Hz, and the signal energy of other frequency bands is very low, indicating that the counterfeit underwater acoustic signal and the real underwater acoustic signal are in almost the same frequency band, and the counterfeiting effect is good.

[0031] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for underwater acoustic signal impersonation based on MelGAN, characterized in that, The application comprises the following steps: Step 1: MelGAN is used as a water sound signal imitation model and is trained; wherein the kernel size of the hole convolution in the generator is set to an integer multiple of the step size, and the frame shift of the sampling in the generator is set to 256; the training process specifically comprises: inputting a Mel spectrogram into the generator to generate an imitation water sound signal by learning the distribution of a real water sound signal; inputting the imitation water sound signal and the real water sound signal into a discriminator to make the MelGAN model converge by using a loss function, and the converged MelGAN model is the water sound signal imitation model; before inputting the real water sound signal into the discriminator, the real water sound signal is preprocessed; the preprocessing comprises: frequency reduction and pre-emphasis; Step 2: a test sample is inputted into the trained generator to obtain an imitation water sound signal.

2. The method of claim 1, wherein the MelGAN-based underwater acoustic signal imitation method is characterized by, The size of the overlapping area is controlled by increasing or decreasing the number of layers of the hole convolution in the generator.

3. The method of claim 1, wherein the MelGAN-based underwater acoustic signal imitation method is characterized by, The Hinge loss function is used as the loss function.

Images