Deep learning based sea clutter small target detection method

Abstract

The application provides a sea clutter small target detection method based on deep learning, in the training process, clutter simulation is used, so that the discriminator of the WaveGAN model learns the deep characteristics of the sea clutter; in the detection process: the completed discriminator performs binary classification discrimination on whether the echo signal is real sea clutter data; when the discriminator output is true, it means that the echo signal is sea clutter and no target is detected; when the discriminator output is false, it means that the echo signal is sea clutter plus target and a target is detected. The application provides an end-to-end method, which does not need to perform excessive preprocessing on the echo data; does not depend on the statistical distribution model of the sea clutter, and has good detection effect under any distribution model; and has good detection performance on low signal-to-clutter ratio targets and low-speed targets.

Description

Technical Field

[0001] This invention relates to target detection technology, and particularly to the detection technology of small targets in sea clutter. Background Technology

[0002] Radar is the primary means of target detection. Its working principle involves emitting electromagnetic waves into the detection area and then determining the presence of a target based on the reflected echo. Target detection at sea differs from land-based detection. On land, the background is relatively fixed, and the reflected echo of a target is significantly different from the ambient echo. However, at sea, waves are constantly changing due to factors such as wind speed, making target echoes difficult to distinguish within the changing environmental echoes. Furthermore, our demands for marine target detection are becoming increasingly stringent, requiring the detection of small targets such as periscopes and drones. Therefore, we need to further explore more efficient and stable methods for detecting marine targets.

[0003] Detection of small targets amidst sea clutter is a critical issue in maritime surveillance. Traditional sea surface target detection algorithms initially focus on suppressing sea clutter, primarily relying on prior assumptions about its distribution to improve the signal-to-clutter ratio (SNR). They then utilize the difference in Doppler frequencies between the clutter and the target to make a decision on the target. However, sea clutter exhibits complex time-varying and dynamic characteristics, with its statistical distribution displaying non-Gaussian properties, often making it difficult to match prior assumptions. Furthermore, in conditions of high wind speeds and large waves, the average echo power of sea clutter increases, leading to greater Doppler frequency broadening and a decrease in the SNR. Slow-moving targets, after Doppler processing, are easily aliased with sea clutter, making them difficult to detect. The complex marine environment causes sea clutter to exhibit non-Gaussian, non-stationary, and nonlinear characteristics, posing significant challenges to the detection of small targets on the sea surface. Therefore, new sea surface target detection methods based on these new characteristic differences are needed.

[0004] With the development of deep learning, its methods have also been applied to the simulation and generation of radar sea clutter. Deep neural networks have the characteristic of fitting arbitrary functions and can generate simulated clutter that is highly similar to the input clutter.

[0005] Traditional target detection methods under clutter conditions require continuous estimation of the noise variance in the sea clutter background and constant adjustment of the threshold value based on the noise power to ensure that the false alarm rate remains at a constant level. This method of continuously adjusting the threshold value to maintain a constant false alarm probability is called constant false alarm rate (CFAR) detection.

[0006] CFAR detection process: first analyze the clutter background power in the reference cell, set the threshold of the detected cell, use the idea of binary decision, if the power of the detected cell is greater than the threshold, it is considered to have a target. Reference: Siddiq K, Irshad M. Analysis of the cell averaging CFAR in Weibull background using a distribution approximation, "[C]. 2009 2nd International Conference on Computer Control and Communication, Karachi Pakistan, 2009: 1-5. The contrast method used in the reference is a CFAR detection method under Weibull distribution condition, the threshold factor is related to the shape parameter of Weibull distribution. When CFAR detection is applied to sea clutter target detection, it needs to rely on the statistical distribution model of sea clutter. When the statistical model is mismatched, the performance will be seriously degraded, and the performance of low signal-to-clutter ratio target detection and low-speed target or zero Doppler target detection will be poor.

[0007] With the improvement of radar resolution, the characteristics of sea clutter become more complex and it is difficult to use statistical distribution model for fitting. In order to improve the radar detection performance and break through the detection signal-to-clutter ratio, it is necessary to have a deep understanding of the characteristics of sea clutter and fully utilize different characteristics. Therefore, many researchers begin to change their research ideas and propose feature-based detection methods, which no longer focus on amplitude distribution, but select other features in time domain or frequency domain for target detection. Because sea clutter and target have different characteristics, there will be great differences in some feature analysis. The received radar echo is preprocessed first, the features are extracted, and then the features of pure sea clutter are compared to determine whether there is a target. When a single feature cannot be stably detected, multiple features can be used for comprehensive analysis to realize target detection under sea clutter condition. According to the different features, it can be divided into time-frequency features, fractal features and chaotic features. How to extract more stable and more universal features and how to improve the detection ability of small targets in strong clutter environment are problems that need further study. And the feature detection method needs to extract features from the radar echo first, which will involve a lot of calculation and preprocessing, and the actual application process requires as low calculation complexity as possible. And whether the feature detection method can still be applied in dynamic environment needs to be studied. SUMMARY

[0008] The technical problem to be solved by the present application is to provide a small target detection method in sea clutter which can select features with universality and does not need preprocessing.

[0009] The technical scheme adopted by the present application to solve the above technical problems is a sea clutter small target detection method based on deep learning, which applies a WaveGAN model to radar target detection, the WaveGAN model including a generator and a discriminator, and the specific steps are as follows:

[0010] The WaveGAN model training process: on the one hand, the received original sea clutter data is first subjected to a far distance echo and target data elimination and then normalized, and the normalized sea clutter data is subjected to distance dimension splitting, and the distance dimension split sea clutter data is extracted as real sea clutter data and input into the discriminator; on the other hand, the generator of the WaveGAN model generates simulated clutter data using input Gaussian white noise and inputs the simulated clutter data into the discriminator of the WaveGAN model; the discriminator performs binary classification discrimination on the input data to determine whether it is real sea clutter data; and in the training process, the clutter simulation enables the discriminator to learn the deep characteristics of the sea clutter.

[0011] The WaveGAN model detection process: the input echo signal to be detected is subjected to distance dimension splitting and then input into the trained discriminator, and the discriminator performs binary classification discrimination to determine whether it is real sea clutter data; when the discriminator outputs true, it indicates that the echo signal is sea clutter and no target is detected; and when the discriminator outputs false, it indicates that the echo signal is sea clutter plus target and a target is detected.

[0012] The present application is different from the general target detection method in that it learns target signals in various scenes to complete detection, and the present application innovatively uses the discriminator in an end-to-end deep learning network model that has learned different distributed clutter backgrounds, the discriminator does not learn target signals and can only accurately and efficiently complete the classification of whether the current signal belongs to clutter, and the result of detecting the target signal is associated with the classification result of the discriminator output not belonging to clutter. Experiments show that the present application can improve the generality of small target detection models in various complex clutter backgrounds, reduce the constraint of mismatched clutter distribution assumptions, and improve the detection performance of slow small targets in low signal-to-clutter ratio.

[0013] The present application has the following beneficial effects:

[0014] 1. An end-to-end method is provided, which does not require excessive preprocessing of echo data;

[0015] 2. It does not depend on a statistical distribution model of sea clutter and can have good detection effect under any distribution model;

[0016] 3. It has good detection performance for low signal-to-clutter ratio targets and low speed targets. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1Raw sea clutter data; (a) 2D time domain; (b) 2D frequency domain;

[0018] Figure 2 Preprocessed sea clutter data; (a) time domain; (b) range Doppler;

[0019] Figure 3 Target simulation result; (a) echo of the cell where the target is located (b) range Doppler after adding the target;

[0020] Figure 4 WaveGAN training process schematic diagram;

[0021] Figure 5 Discrimination process schematic diagram of the embodiment;

[0022] Figure 6 Loss function change curve; (a) discriminator; (b) generator;

[0023] Figure 7 Training process generation effect diagram; (a) epoch = 1; (b) epoch = 2001; (c) epoch = 4001; (d) epoch = 6001;

[0024] Figure 8 Sea clutter data with added target;

[0025] Figure 9 Detection result schematic diagram;

[0026] Figure 10 WaveGAN detection result; (a) detection probability curve; (b) false alarm probability curve;

[0027] Figure 11 Range Doppler diagram with added target;

[0028] Figure 12 Detection result under different Doppler frequencies;

[0029] Figure 13 Change of detection probability under different Doppler frequencies;

[0030] Figure 14 Signal-to-clutter ratio detection probability curve under different simulated clutter;

[0031] Figure 15 CFAR detection under real sea clutter;

[0032] Figure 16 Detection probability comparison curve under real sea clutter. DETAILED DESCRIPTION

[0033] I. Embodiment

[0034] The data used in the experiment is sea clutter data measured at sea. The original data is in the form of a two-dimensional matrix, and the two dimensions of the matrix represent the range dimension and the pulse dimension of the radar echo.

[0035] Figure 1 The sea clutter data is a clutter overview graph, a time domain two-dimensional graph, and a frequency domain two-dimensional graph. It can be seen from the overview graph that the clutter amplitude in the first 300 range cells is very high. From the combination of the pulse dimension and the range dimension, due to the influence of different sea conditions and waves, some sea spikes and strong clutter regions appear. After that, the sea clutter intensity will rapidly decrease, and the sea clutter at a far distance will be submerged in the environmental noise, showing white noise characteristics.

[0036] The data preprocessing is mainly to eliminate the range influence and remove target data and far distance data. The specific method is: first, remove the far distance echo and target data in the original sea clutter data, then perform normalization processing, and then perform range dimension splitting on the normalized data.

[0037] Figure 2 The training data statistical graph after preprocessing is given, Figure 2 (a) gives the data graph after normalization. It can be seen from the graph that the clutter amplitude values of different range cells are not much different. Figure 2 (b) gives the distance Doppler graph after preprocessing. It can be seen from the graph that the target echo and the removed far distance white noise range cell are removed.

[0038] The obtained sea clutter data has a total of 26280 range cell data, and the pulse number of each range cell is 4096. During model training, 64 range cell data are randomly taken as a training batch of data. When detecting the target, 2200 range dimension data are taken as a data matrix, and each range cell is repeated 64 times as a batch of data.

[0039] The above process generates pure sea clutter data, which is used for training the WaveGAN model to make the discriminator learn the deep features of the pure sea clutter data. In the embodiment, the generator and trainer of the WaveGAN model use the training method in the applicant's public number CN113702923, a radar clutter simulation method and electronic equipment.

[0040] The test data of the detection model adds targets with different signal-to-clutter ratios in the range dimension on the basis of the pure sea clutter data.

[0041] The radar transmitting signal is S(t) = u(t) * exp(j2πf c t), where u(t) is the signal complex envelope, f c is the operating frequency, when the target distance is R, the light speed is c = 3 × 108 , the echo of the target is obtained after a delay of 2*R / c. When the target speed is v, the Doppler frequency of the target is The radar wavelength λ = c / f c The overall target echo expression is S r (t) = K*u(t-τ)exp(j2π(f c +f d )(t-τ)), where τ = 2*R / c is the two-way delay time of the target echo, and K is the echo amplitude.

[0042] The radar target simulation parameter settings in the embodiment include a working frequency of 9.5 GHz, a pulse width of 40 ns, a PRF of 1697 Hz, and a sampling frequency of 60 MHz. The amplitude model of the target uses the Swerling 4 model, and the signal-to-clutter ratio is defined as the ratio of the power of the target in the distance unit to the power of the clutter. The clutter power is defined as the sum of the squares of the amplitudes of the data in a single distance unit. Figure 3 (a) gives the amplitude diagram of the target echo in the same distance dimension when the signal-to-clutter ratio is 8 dB, Figure 3 (b) gives the range-Doppler diagram after adding a target with a speed of 10 m / s at the distance dimension 1000. It can be seen from the diagram that the target has a Doppler frequency. Subsequent target addition uses this simulation echo.

[0043] Figure 4 A training process for training WaveGAN is as follows:

[0044] 1) First, input the random noise vector and pure clutter data into the discriminator;

[0045] 2) The discriminator outputs a discrimination result, the generator optimizes the parameters according to the discrimination result, and outputs the false clutter data. The false clutter data is returned to the discriminator to repeat step 2) until the discriminator reaches the maximum training round;

[0046] 3) After the discriminator reaches the maximum training round, the discriminator detects and scores each distance dimension of the false clutter data output by the generator and the real sea clutter data;

[0047] During the training process of WaveGAN, the generator continuously optimizes the parameters according to the scoring of the discriminator, so that the generation effect of the generator is continuously enhanced, the false clutter is consistent with the real distributed clutter, and the discriminator continuously enhances the recognition ability of the model for distinguishing real sea clutter and generated false clutter. After the model converges, the discriminator can distinguish whether the high-dimensional hidden features of the input data are the same as the sea clutter.

[0048] Therefore, after the WaveGAN training is completed, the echo data is input into the discriminator alone, and the discriminator scores, according to the output value of the discriminator, a certain threshold is set as the detection threshold of the discriminator, so as to judge whether the target exists. Figure 5 For the discrimination process, i.e., the target detection process, the input structure of the discriminator is [N, L, 2], and there are N samples of distance dimension data input in each batch. In order to detect the same distance dimension, the data of the discrimination process is repeated N times for each distance dimension, so that the data input into the discriminator each time is the same distance dimension data, and the output data dimension is [N, 1], representing the scores of the N sample data by the discriminator. In the discrimination process, the output of the discriminator is averaged to obtain a scalar output, representing the score of the distance dimension data.

[0049] The WaveGAN generator includes an input layer, an intermediate layer and an output layer. The input layer is composed of a fully connected layer Dense, a matrix reshaping function reshape, a batch normalization batchnom and a ReLu activation function. The intermediate layer includes 4 repeated structures, each layer being composed of a one-dimensional deconvolution layer Conv1d_transpose, a batch normalization batchnom and a ReLu activation function. The output layer is composed of a one-dimensional deconvolution layer Conv1d_transpose and a tanh activation function. The convolution kernel parameters of each module of the generator are as follows:

[0050]

[0051] The WaveGAN discriminator includes an input layer, an intermediate layer and an output layer. The input layer is composed of a one-dimensional convolution layer Conv1d and an LReLu activation function. The intermediate layer includes 4 repeated structures, each layer being composed of a one-dimensional convolution layer Conv1d and an LReLu activation function. The output layer is composed of a matrix reshaping function reshape and a fully connected layer Dense. The convolution kernel parameters of each module of the discriminator are as follows:

[0052]

[0053] The specific parameter settings in the WaveGAN training are as follows: the batch_size of the sample data in each round of training is set to 64; the length of the sample data in each batch, slice_len, is set to 4096; the radar echo data is complex data, including IQ double channels, and the real part and the imaginary part of the complex number are respectively input into the model as a data channel; the model learning rate is set to 5e-5; the model optimization algorithm adopts the Adam optimization algorithm, and the two parameters of the Adam algorithm are β1=0.9 and β2=0.99. The model aims to use the discrimination ability of the discriminator for detection, so the ratio of the discriminator to the generator training coefficient in the model is set to 5:1, that is, after 5 rounds of parameter updating of the discriminator, one round of parameter updating of the generator is performed.

[0054] Using the above training parameters, the WaveGAN model is continuously trained until the model converges. Figure 6 The change curves of the generator loss function and the discriminator loss function in the training process are given, wherein the abscissa is the training round, and the ordinate is the loss value. It can be seen that with the increase of the training round, the generator loss value stabilizes at about 0.3, and the discriminator loss value stabilizes at about -1.9, indicating that the generator and the discriminator have converged respectively.

[0055] In order to show the training effect in the training process, the embodiment makes statistical analysis on the generated samples of the generator and the real data samples in the training process. It can be seen that with the training of the model, the gap between the generated data and the real data. Figure 7 The power spectrum diagram, amplitude square root curve, amplitude real part distribution curve and amplitude imaginary part distribution curve are given. In the generator, 64 sample data are generated in each batch, and the average of the 64 generated data is taken in the analysis, so that the overall generation effect of the generator can be seen.

[0056] Figure 7 In the power spectrum diagram of the embodiment, the light curve is the power spectrum of the generated data, and the dark curve is the power spectrum of the real sea clutter. In the three amplitude analysis diagrams, the dark curve is the generated data, and the light curve is the real sea clutter data. It can be seen that when the epoch is 1, the power spectrum and the amplitude of the generated data are completely different from those of the real data. When the epoch is 2001, the statistical distribution of the generated data changes towards the real data distribution, but there is still a big difference. When the epoch is 4001, the amplitude of the generated data is lower than that of the real data, the middle part of the power spectrum is similar, and there is still a difference in the two sides of the data. When the epoch is 6001, the power spectrum and the amplitude of the generated data are well fitted with the real data, and only the details are different. It can be seen from the figure that the generation effect of the generator is very good, and then the discriminator also learns the distribution characteristics of the real sea clutter data.

[0057] II. Performance analysis of target detection results:

[0058] 1. Detection results under different signal-to-clutter ratios

[0059] First, a target with a signal-to-clutter ratio of 8 dB is added at the distance dimension 200, a target with a signal-to-clutter ratio of 6 dB is added at the distance dimension 800, and a target with a signal-to-clutter ratio of 10 dB is added at the distance dimension 1200 as shown in Figure 8 .

[0060] Figure 9 The detection results of the discriminator after the model converges are given. The discriminator is finally a fully connected layer, and the range of the output of the discriminator is not limited. After the model converges during the training process, the maximum value of the output of the discriminator after the pure sea clutter is input is selected as the classification threshold of the detection target. The curve is the output of the discriminator when the target echo is input, and the horizontal line is the decision threshold, and the threshold value is 2.32. The detection results of the targets added at 200, 800, and 1200 are 4.0257, 3.6569, and 4.4858, respectively. It can be seen that the targets are detected at the three distance dimensions, and no false alarm is generated under this condition.

[0061] Under the above experimental conditions, target data with different intensities of signal-to-clutter ratios of 0-20 dB are set for detection. The detection results are recorded by repeating the experiment 2000 times, and the detection probability is calculated as shown in Figure 10 .

[0062] As can be seen from the figure, as the signal-to-clutter ratio increases, the detection probability gradually increases. When the signal-to-clutter ratio is 6 dB, the target detection probability can reach more than 0.5, when the signal-to-clutter ratio is 10 dB, the detection probability can be close to 0.9, and when the signal-to-clutter ratio reaches more than 15 dB, the detection probability is close to 1. As can be seen from the figure, the method proposed in this paper has good detection ability for targets under sea clutter conditions.

[0063] The target detection model of deep learning classifies the input data according to the comparison of the score of the input data by the discriminator and the threshold value to detect the target. In the experiment, the false alarm rate fluctuates around 0.00454, and the maximum appears at 6 dB, and the minimum appears at 13 dB. Under this condition, the false alarm probability has no obvious relationship with the input signal-to-clutter ratio.

[0064] 2. Detection results under different Doppler frequencies

[0065] Under real sea clutter conditions, the target may have a certain speed, and the FFT transform is performed on the echo data to obtain a range-Doppler two-dimensional image. The range dimension corresponding to the target contains a certain Doppler frequency. To study the target detection effect of WaveGAN detection model under different Doppler frequencies, different Doppler frequency targets are added to the original sea clutter data at three different distance dimensions, and the target amplitudes are consistent. The range-Doppler two-dimensional image is shown in Figure 11 .

[0066] The results of the input discriminator after adding the Doppler frequency are shown in FIG. 6. The detection results of the 260-dimensional, 500-dimensional and 1000-dimensional targets are 3.6866, 3.6538 and 3.7546 respectively, all of which exceed the threshold value 2.32 represented by the horizontal line, and the targets can be detected. Figure 12

[0067] In order to study the detection effect of WaveGAN on low-speed targets, the signal-to-clutter ratio is fixed at 10 dB under the above experimental conditions, the target speed range [-10, 10] is changed, and 2000 repeated experiments are conducted under each experimental condition to record the detection results. Figure 13 The change of the detection probability with the change of the Doppler frequency is given. It can be seen from the figure that the detection probability fluctuates around 0.9, and the difference is not large, which shows that the WaveGAN model also has good detection effect on low-speed targets.

[0068] 3. Detection results of different distribution clutter

[0069] The target is added to the simulation clutter with 0-20dB, and the detection results are recorded by repeating 2000 experiments. Figure 14 The target detection results under the simulation clutter condition are given. It can be seen from the figure that the detection probability can reach more than 0.5 when the signal-to-clutter ratio is 6dB, and the detection probability can reach more than 0.8 when the signal-to-clutter ratio is 9dB. From the detection results, it can be seen that the WaveGAN model has good detection effect on different distribution statistical distribution clutter model. The traditional CFAR detector usually has strong dependence on the sea clutter model, and the performance will decrease greatly when the statistical model is mismatched. The WaveGAN model has strong adaptability to different distribution sea clutter data, and does not depend on a certain distribution.

[0070] 4. Comparison results of WaveGAN and CFAR

[0071] The reference unit of the CFAR detection is set to 16, the protection unit is set to 2, the false alarm rate is set to 0.001, the SCR is set to 10dB, and the target is added at the distance dimension 100. Figure 15 The detection results under the real sea clutter are given. The lower curve in the figure is the one-dimensional sea clutter data, and the upper curve is the CFAR detection threshold.

[0072] Under the same experimental conditions, the signal-to-clutter ratio is set to 0-20dB, and 2000 repeated experiments are conducted to record the detection probability. Figure 16 ​The comparison curves of CFAR detection and WaveGAN detection model under the condition of measured sea clutter are given. It can be seen from the figure that the CFAR detection has poor detection ability under the condition of low signal-to-clutter ratio in the measured sea clutter. When the signal-to-clutter ratio is 0-5dB, the detection probability is less than 0.1. With the increase of signal-to-clutter ratio, the detection probability gradually increases. When the signal-to-clutter ratio reaches 10dB, the detection probability is above 0.6, which is a relatively low detection probability. The detection probability of WaveGAN model is higher. When the signal-to-clutter ratio is less than 4dB, there is still a probability to detect the target. When the signal-to-clutter ratio is 6dB and 10dB, the WaveGAN detection probability is 0.4 and 0.3 higher than the CFAR detection probability respectively.

[0073] The CFAR detection has poor detection ability under the condition of low signal-to-clutter ratio in the measured sea clutter. When the signal-to-clutter ratio is 0-5dB, the detection probability is less than 0.1. With the increase of signal-to-clutter ratio, the detection probability gradually increases. When the signal-to-clutter ratio reaches 10dB, the detection probability is above 0.6, which is a relatively low detection probability. The detection probability of WaveGAN model is higher. When the signal-to-clutter ratio is less than 4dB, there is still a probability to detect the target. When the signal-to-clutter ratio is 6dB and 10dB, the WaveGAN detection probability is 0.4 and 0.3 higher than the CFAR detection probability respectively.

[0074] It can be verified by the test that the detection method based on WaveGAN model has better detection performance for low signal-to-clutter ratio and low Doppler frequency target; and is not dependent on a certain distributed clutter, and has stronger applicability.

Claims

1. A method for detecting small targets in sea clutter based on deep learning, characterized in that, The specific steps are: WaveGAN model training process: on the one hand, the received original sea clutter data is first normalized after removing long-distance echoes and target data, and then the normalized sea clutter data is split in the distance dimension, and the split sea clutter data is extracted as sample data of real sea clutter and input into the discriminator; on the other hand, the generator of the WaveGAN model generates simulated clutter data using input Gaussian white noise and inputs the simulated clutter data into the discriminator of the WaveGAN model; the discriminator performs binary classification discrimination on the input data to determine whether it is real sea clutter data; During the training process, the clutter simulation enables the discriminator to learn the deep characteristics of the sea clutter; WaveGAN model detection process: the input echo signal to be detected is split in the distance dimension and input into the trained discriminator, which performs binary classification discrimination to determine whether it is real sea clutter data; when the discriminator outputs true, it means that the echo signal is sea clutter and no target is detected; when the discriminator outputs false, it means that the echo signal is sea clutter plus target and a target is detected.

2. The method of claim 1, wherein, The generator includes an input layer, an intermediate layer and an output layer, the input layer is cascaded by a Dense full connection layer, a matrix reshaping function reshape, a batch normalization and a ReLu activation function; the intermediate layer includes 4 layers of repeated structures, each layer is cascaded by a one-dimensional deconvolution layer, a batch normalization batchnom and a ReLu activation function; the output layer is cascaded by a one-dimensional deconvolution layer and a tanh activation function; The WaveGAN discriminator includes an input layer, an intermediate layer and an output layer, the input layer is cascaded by a one-dimensional convolution layer and an LReLu activation function; the intermediate layer includes 4 layers of repeated structures, each layer is cascaded by a one-dimensional convolution layer and an LReLu activation function; the output layer is cascaded by a matrix reshaping function reshape and a Dense full connection layer.

3. The method of claim 1, wherein, The convolution kernel of the Dense full connection layer of the input layer of the generator is 1, 100, 4096, and the bias is 256; the one-dimensional deconvolution layer of the 4 layers of repeated structures in the intermediate layer has convolution kernels of 25, 1024, 512, 25, 512, 256, 25, 256, 128 and 25, 128, 64 from low to high, and biases of 512, 256, 128 and 64 from low to high; the convolution kernel of the one-dimensional deconvolution layer of the output layer is 25, 64, 2, and the bias is 2; The convolution kernel of the one-dimensional convolution layer of the input layer of the discriminator is 25, 2, 64, and the bias is 64; the one-dimensional convolution layer of the 4 layers of repeated structures in the intermediate layer has convolution kernels of 25, 64, 5128, 25, 128, 256, 25, 256, 512 and 25, 512, 1024 from low to high, and biases of 128, 256, 512 and 1024 from low to high; the convolution kernel of the Dense full connection layer of the output layer is 1, 4096, 1, and the bias is 1.

4. The method of claim 1, wherein, The ratio of the training coefficients of the discriminator to the generator in the WaveGAN model training process is 5:

1.

5. The method of claim 1, wherein, The sample data batch size in each round of training in the WaveGAN model training process is set to 64; the length of the sample data in each batch (slice_len) is set to 4096; the echo signal is complex data, including IQ double channels, and the real part and the imaginary part of the complex number are respectively input into the WaveGAN model as a data channel; the learning rate of the WaveGAN model is set to 5e-5; the model optimization algorithm adopts the Adam optimization algorithm, and the two parameters of the Adam algorithm are β1=0.9 and β2=0.99.

Images