A method for signal enhancement in an underwater acoustic jamming environment

Abstract

This invention provides a signal enhancement method for underwater acoustic interference environments. Through the game-like process between the generator and discriminator in a WGAN (Generative Adversarial Network), the generator learns the characteristics of the artificial interference signal and the target signal in the time-frequency diagram of the mixed signal, and outputs an enhanced signal to achieve anti-interference. This invention combines the data representation characteristics of STFT (Short Time Fourier Transform) and applies Generative Adversarial Networks (GANs) to anti-interference in artificial interference scenarios. It fully utilizes the advantages of mutual learning between the generator and discriminator in GANs to achieve Nash equilibrium, the ability of the short-time Fourier transform to adapt well to the data structure of neural networks, and the attention gate's ability to focus on key data points. The trained generator can enhance the target signal in the mixed signal, achieving a good anti-interference effect.

Description

Technical Field

[0001] This invention relates to the fields of signal processing and deep learning, and in particular to a method for extracting target signals from a mixture of artificial interference and target signals based on generative adversarial networks. Background Technology

[0002] Anti-jamming is a hot research topic in the field of underwater acoustic countermeasures. Currently, underwater acoustic detection often faces both natural and man-made interference, severely impacting detection performance. Besides natural interference such as marine environmental noise, underwater vehicle and ship noise, wake scattering, and marine reverberation, man-made interference is unavoidable in military applications. With the continuous iteration and upgrading of underwater jammers, the types of man-made interference are becoming increasingly diverse, such as frequency sweeping jamming, broadband noise jamming, and agile noise jamming. Furthermore, underwater detection equipment exhibits different failure patterns for each type of interference due to different detection algorithms. This necessitates considering multiple scenarios when designing anti-jamming methods, significantly increasing the difficulty of anti-jamming. Therefore, there is an urgent need for an anti-jamming technology that can adapt to various interference types, and deep learning, with its powerful abstract feature learning and fitting capabilities, is gradually demonstrating its advantages in the field of anti-jamming.

[0003] Underwater acoustic detection commonly uses relatively simple signal types, such as single-frequency signals and various frequency-modulated signals. The working principle involves an underwater acoustic transducer emitting a detection signal. After reflection from the target, the signal undergoes Doppler effect and time delay based on the target's velocity, and is then received by a receiver. The receiver can analyze the frequency components and time delay of the echo signal to determine the distance and velocity between the target and the detector. However, in certain scenarios, underwater acoustic detection equipment can be affected by artificial interference signals, leading to significant errors in the estimation of target parameters or even misjudging the presence or absence of the target. This severely impacts the equipment's ability to identify the target, resulting in a degraded detector performance.

[0004] Generative Adversarial Networks (GANs), a hot research area in deep learning, have seen mature applications in various fields. In enhancement tasks, existing research includes SEGAN for speech enhancement and the pix2pix architecture for image enhancement. However, these studies use samples with good geometric similarity, allowing GANs to perform tasks relatively stably. For this task, the non-stationarity of the artificial interference signal itself causes the GAN to collapse. Therefore, a new approach is needed to address this technical deficiency. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a signal enhancement method for underwater acoustic interference environments. Addressing the problem that existing generative adversarial network (GAN) architectures cannot stably train under artificial interference samples, this invention proposes a generator incorporating an attention gate for GAN training. By setting up a mixed signal dataset and a target signal dataset under a specific interference scenario, a generator and discriminator are designed and trained based on the datasets. Ultimately, the generator and discriminator reach Nash equilibrium, and the generator combined with the attention gate exhibits strong signal enhancement capabilities.

[0006] The purpose of this invention is to provide a method for enhancing signals in the face of artificial interference based on Generative Adversarial Networks (WGANs). Through the game-like process between the generator and discriminator in a WGAN, the generator learns the characteristics of the artificial interference signal and the target signal in the time-frequency diagram of the mixed signal, and outputs an enhanced signal to achieve anti-interference. This method is applicable to communication or military confrontation scenarios and can accurately reconstruct the original target signal under strong interference environments.

[0007] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0008] Step 1: Determine the interference status.

[0009] Analyze the interference scenario, establish a channel model by combining environmental characteristics and target information, and determine the range of state parameters R of the interference scenario, the dynamic transfer function h = (t, r ∈ R) of the channel, and the background noise n. a ;

[0010] Step 2: Generate the training set;

[0011] Using the interference scenario state parameter range R from step 1 and the channel's dynamic transfer function h = (t, r ∈ R), generate k pairs of time-domain waveforms s for the mixed signals. m The time-domain waveform s of the target signal o Each mixed signal s m With target signal s o For the same interference scenario, the mixed signal contains background noise n a Interference signal s j and target signal s o ;

[0012] Step 3: Preprocess the training set data;

[0013] The STFT is used to convert the time-domain waveforms of each pair of mixed signals and the time-domain waveforms of the target signal into 1×M×N time-frequency matrices, respectively, to obtain the time-frequency matrix of the mixed signal and the time-frequency matrix of the target signal. The real and imaginary parts of each element in the time-frequency matrix of the mixed signal and the time-frequency matrix of the target signal are respectively used as two channels to transform the 1×M×N complex time-frequency matrix into a 2×M×N real matrix.

[0014] Step 4: Design the structure and training strategy of WGAN;

[0015] Due to the special characteristics of the training samples, an optimized WGAN training strategy is adopted.

[0016] Step 5: Train WGAN.

[0017] The k pairs of mixed signal time-frequency matrices S m Input generator G, and use the training strategy in step 4 to train generator G and discriminator D simultaneously to obtain trained generator G'.

[0018] Step 6: Perform signal enhancement;

[0019] Mixed signal S m The input generator G yields the time-frequency matrix S of the enhanced signal. g =G′(S m Then, perform an inverse transform on the time-frequency matrix to obtain the time-domain waveform s of the enhanced signal. g .

[0020] In step 3, the sample undergoes an SFTF transformation. The expression for the STFT is:

[0021]

[0022] in, For Fourier transform, ω is the angular frequency, w(t) is the window function, and s(t) is the sample; s(t) is the discrete signal. It is a two-dimensional matrix, where each column represents the result sequence after windowing the signal at different positions and performing Fourier transform on the resulting segments.

[0023] The w(t) uses a Kaiser window, and its expression is:

[0024]

[0025] Where I0(·) is the first kind of Bessel function, N is the total length of the window function, the time-frequency matrix generated by STFT is filled with zeros, and the real part and imaginary part of the matrix are treated as separate channels.

[0026] In step 4, the optimization steps for WGAN are as follows:

[0027] The generator G in WGAN is UNET. For training stability, UNET does not contain fully connected layers or dropout layers, and an attention gate AG is added during the upsampling process of generator G. The input of generator G is the time-frequency matrix S of the mixed signal. m The loss function for generator G is L2 loss.

[0028] The discriminator D in WGAN is an FCN, and the input of discriminator D is the time-frequency matrix S of the target signal. o Or the generator's output matrix S g The generator model obtained after training has the effect of anti-interference and signal enhancement, and can enhance the target signal when the input signal has background noise and artificial interference.

[0029] In step 4, after optimization, the discriminator loss function in WGAN is:

[0030]

[0031] in For the expectation, x and x c Let f represent the target signal and the mixed signal, respectively. w (.) represents the complex function obtained from training the discriminator network; the discriminator loss function of WGAN needs to be combined with the L2 loss, and its expression is:

[0032]

[0033] Where λ is the weighting factor of L2 loss, ||G(x) c )-x||2 is the L2 loss.

[0034] In step 4, the expression for the attention gate is as follows:

[0035]

[0036]

[0037] Where σ1 and σ2 are activation functions, W x W g Both ψ and b are convolution operations. g b ψ , where x and g are the outputs of the downsampling layer and the upsampling layer, respectively.

[0038] The beneficial effects of this invention lie in the model enhancement method for artificial interference based on generative adversarial networks proposed in this invention. Combining the characteristics of STFT in data representation, the generative adversarial network is applied to anti-interference in artificial interference scenarios. It makes full use of the advantages of mutual learning between generators and discriminators in generative adversarial networks to achieve Nash equilibrium, the ability of the results of short-time Fourier transform to adapt well to the data structure of neural networks, and the attention gate's ability to pay attention to key data in the data. The trained generator can enhance the target signal in mixed signals, achieving a good anti-interference effect. Attached Figure Description

[0039] Figure 1 This is the overall flowchart for training WGAN.

[0040] Figure 2 shows the anti-interference effect of the generator when the signal-to-interference ratio is -9.8dB in the verification set. Figure 2(a) is the target signal time-frequency diagram, Figure 2(b) is the enhanced signal time-frequency diagram, and Figure 2(c) is the received signal time-frequency diagram.

[0041] Figure 3(a) shows the frequency error of the enhanced signal and the original target signal as a function of the signal-to-interference ratio in the validation set. Figure 3(b) shows the frequency error of the enhanced signal and the original target signal as a function of the signal-to-interference ratio in the validation set.

[0042] Figure 4 This is a schematic diagram showing the relationship between the cross-correlation coefficients of the enhanced signal and the local oscillator signal of each training strategy and the signal-to-interference ratio. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] Consider an underwater target detection mission under near-field conditions, where an active sonar emits a detection signal and uses the echo signal to sense the target's range and velocity information. Assume the sonar detection signal is a single-frequency pulse and is subject to interference from noise amplitude-modulated signals.

[0045] Step 1: Assume the detection scenario is near-field detection, the detector working distance is within 20m, multipath, interface reverberation and interface reflection are not considered, and the background noise is Gaussian white noise.

[0046] Step 2: Assume the single-frequency probe signal has a frequency of 12kHz, a duration of 0.04s, and an amplitude of 1. The carrier frequency of the noise amplitude-modulated signal is 15kHz, the baseband noise bandwidth is 13kHz–17kHz, the duration is the same as the probe signal, and the amplitude ratio of the noise noise signal to the probe signal is in the range of -20dB to 10dB. The filter generating the baseband noise is a Butterworth filter with a maximum passband ripple of 1dB, a minimum stopband attenuation of 40dB, and a baseband noise power of 1. The sampling frequency is 10 times the probe signal frequency. The expressions for the single-frequency signal and the noise amplitude-modulated signal are as follows:

[0047] s o (t)=A o cos(2πf o t)

[0048] s j (t)=A j (u1+u b (t))cos(2πf j t)

[0049] A o A represents the amplitude of the target echo received by the detector. j The amplitude of the interference signal received by the detector is u1, and the amplitude of the carrier signal is u. b For baseband noise, f o and f j These are the target signal frequency and the center frequency of the interference signal, respectively; the expression for the mixed signal is:

[0050] s n (t)=s o (t)+s j (t)+n a (t)

[0051] In this scenario, the interference scenario state refers to the difference between the interference signal amplitude and the echo delay and the interference delay, with the delay being randomly generated using a mean of 0.04. 2000 training samples and 50 test samples are generated using the above parameters.

[0052] Step 3: Perform SFTF transformation on the sample. The expression for STFT is:

[0053]

[0054] in, For Fourier transform, ω is the angular frequency, w(t) is the window function, and s(t) is the sample; s(t) is the discrete signal. This is a two-dimensional matrix, where each column represents the result sequence after Fourier transform of the obtained segments after windowing the signal at different locations; w(t) uses a Kaiser window, and its expression is:

[0055]

[0056] Where I0(·) is the Bessel function of the first kind, N is the total length of the window function, β = 5, N = 256, and the number of overlap points between each window is 240; the resulting sample is a matrix of size 1×256×M. Since the window function is fixed, the number of frequency axis points of STFT is unique, but since the echo signal delay is different from the interference signal delay in each interference scenario, the number of sample points in the time domain varies. In order to make the STFT result more suitable for the input of WGAN, the time-frequency matrix generated by STFT is padded with zeros to a size of 1×256×512, and the real and imaginary parts of the matrix are treated as separate channels, so the final sample size is 2×256×512.

[0057] Step 4: Select the generator and discriminator from WGAN. WGAN introduces Wasserstein distance into GAN and adds an attention gate during the generator's upsampling phase. This solves the gradient vanishing and mode collapse problems encountered in traditional GANs during training. After optimization, the discriminator loss function in WGAN is:

[0058]

[0059] in For the expectation, x and x c Let f represent the target signal and the mixed signal, respectively. w (.) represents the complex function obtained from training the discriminator network; the discriminator loss function of WGAN needs to be combined with the L2 loss, and its expression is:

[0060]

[0061] Where λ is the weighting factor of L2 loss, ||G(x) c The L2 loss is λ = 7. In this example, we take λ = 7.

[0062] The expression for the attention gate is as follows:

[0063]

[0064]

[0065] Where σ1 and σ2 are activation functions, W x W g Both ψ and b are convolution operations. gb ψ , where x and g are the outputs of the downsampling layer and the upsampling layer, respectively.

[0066] The 2000 samples generated in step 2 are input into WGAN for training. After training, the 50 test samples generated in step 2 are used to test the generator's performance. The enhanced time-frequency matrix output by the generator is transformed to the time domain through inverse transformation to detect its anti-interference effect in the scene.

[0067] The signal enhancement method proposed in this invention was simulated, and the flowchart is as follows. Figure 1 As shown in Figures 2 and 3, the interference scenario was established and an underwater acoustic detection task was performed under near-field conditions. The final anti-interference effect is shown in Figures 2 and 3. Figure 4 As shown. From Figure 4 It can be seen that among various enhancement strategies, the method proposed in this invention has the best effect.

Claims

1. A method for signal enhancement in an underwater acoustic interference environment, characterized in that... Includes the following steps: Step 1: Determine the interference status; Analyze the interference scenario, establish a channel model by combining environmental characteristics and target information, and determine the range of state parameters for the interference scenario. Dynamic transfer function of the channel and background noise ; Step 2: Generate the training set; The range of state parameters of the interference scenario in step 1 With the dynamic transfer function of the channel ,generate Time-domain waveform of mixed signals Time-domain waveform of the target signal Each mixed signal With target signal The interference scenarios are the same, and the mixed signal includes background noise. Interference signals and target signal ; Step 3: Preprocess the training set data; The STFT is used to convert the time-domain waveforms of each pair of mixed signals and the time-domain waveforms of the target signal into... The time-frequency matrix is ​​obtained by dividing the time-frequency matrix of the mixed signal and the time-frequency matrix of the target signal. The real and imaginary parts of each element in the time-frequency matrix of the mixed signal and the time-frequency matrix of the target signal are then used as two channels. The complex time-frequency matrix transformation is 0 real matrix; Step 4: Design the structure and training strategy of WGAN; Due to the special characteristics of the training samples, an optimized WGAN training strategy is adopted. In step 4, the optimization steps for WGAN are as follows: The generator G in WGAN is a UNET, which does not contain fully connected layers or dropout layers. An attention gate AG is added during the upsampling process of generator G. The input to generator G is the time-frequency matrix of the mixed signal. The loss function for generator G is L2 loss. The discriminator D in WGAN is an FCN, and the input of discriminator D is the time-frequency matrix of the target signal. or the output matrix of the generator The final trained generator model has the effects of anti-interference and signal enhancement, and can enhance the target signal even when the input signal has background noise and artificial interference. Step 5: Train WGAN; Will For the time-frequency matrix of the mixed signal Input generator G, and use the training strategy in step 4 to train generator G and discriminator D simultaneously to obtain trained generator G'. Step 6: Perform signal enhancement; Mixed signals Input the generator G to obtain the time-frequency matrix of the enhanced signal. Then, an inverse transform is performed on the time-frequency matrix to obtain the time-domain waveform of the enhanced signal. .

2. The signal enhancement method in an underwater acoustic interference environment according to claim 1, characterized in that: In step 3, the sample undergoes an SFTF transformation. The expression for the STFT is: ； in, For Fourier transform, Angular frequency, For window functions, For the sample; For discrete signals, It is a two-dimensional matrix, where each column represents the result sequence after windowing the signal at different positions and performing Fourier transform on the resulting segments.

3. The signal enhancement method in an underwater acoustic interference environment according to claim 2, characterized in that: The Using the Kaiser window, the expression is: ； in Let N be the Bessel function of the first kind, and N be the total length of the window function. The time-frequency matrix generated by STFT is filled with zeros, and the real and imaginary parts of the matrix are treated as separate channels.

4. The signal enhancement method in an underwater acoustic interference environment according to claim 1, characterized in that: In step 4, after optimization, the discriminator loss function in WGAN is: ； in For the expectation, and Representing the target signal and the mixed signal respectively. (.) denotes the complex function obtained from training the discriminator network; the discriminator loss function of WGAN needs to be combined with the L2 loss, and its expression is: ； in The weighting factor for L2 loss. This is known as L2 loss.

5. The signal enhancement method in an underwater acoustic interference environment according to claim 1, characterized in that: In step 4, the expression for the attention gate is as follows: ； in and For activation function, , , All are convolution operations. , For bias terms, and These are the outputs of the downsampling layer and the upsampling layer, respectively.

Images