Low interception radar signal detection-enhancement method based on LPI-DE-GAN

Abstract

The invention discloses a low-interception-probability radar signal detection and enhancement method based on LPI-DE-GAN, and the method comprises the following steps: S1, obtaining a low-interception-probability radar signal sequence and a 0 sequence, and forming a clean signal sequence; obtaining a noise and marine navigation signal sequence, and superposing the noise and marine navigation signal sequence with the clean signal sequence to obtain a mixed signal sequence; s2, constructing a label of the signal detection task; s3, constructing an LPI-DE-GAN model, wherein the LPI-DE-GAN model comprises a generator and a discriminator; s32, constructing a discriminator; and S4, constructing a loss function, and training an LPI-DE-GAN model. According to the method, the received low-intercept-probability radar signals of various modulation types mixed with large-amplitude fishing boat navigation interference signals can be accurately detected, after the low-intercept-probability radar signals are detected, end-to-end time domain signal enhancement is performed on the low-intercept-probability radar signals, fishing boat navigation signals and noise are suppressed, and original low-intercept-probability radar signals are reconstructed.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar, and particularly relates to a method for detecting and enhancing low-intercept radar signals based on LPI-DE-GAN. Background Art

[0002] As an important part of electronic countermeasures, radar countermeasures play a key role in modern warfare and are one of the important means to achieve the right to information dominance on the battlefield. Radar reconnaissance is the basis of radar countermeasures. By intercepting enemy radar signals through radar reconnaissance, parameters required for electronic countermeasures can be obtained, such as information about the type, purpose, and carrier of the enemy radar, evaluating the threat level of the enemy radar, mastering the enemy's combat intention, perceiving the real-time situation of the battlefield, and providing an important basis for formulating effective countermeasures.

[0003] Low Probability of Interception (LPI) radar signals have characteristics such as low peak power, large time-bandwidth product, frequency agility, and complex in-pulse modulation methods. Therefore, low-intercept probability radar systems are widely used in modern warfare. Due to the above characteristics of LPI radar signals, they are usually submerged in noise and large-amplitude marine navigation signal interference. Since the power of noise and interference is much greater than the power of LPI radar signals, forming ultra-weak signals, it is difficult for reconnaissance aircraft to detect LPI radar signals.

[0004] Currently, common radar signal detection methods include the matched filtering method, time-domain energy detection method, and frequency-domain energy detection method. The matched filtering method requires prior information of the target signal, and this method is not applicable to non-cooperative LPI radar signals; the time-domain energy detection method and the frequency-domain energy detection method need to observe noise and interference before the arrival of the target signal to determine the threshold, which requires rich engineering experience and prior information of noise and interference. In addition, the two energy-based detection methods have poor performance under low signal-to-noise ratios.

[0005] Under the background of large-amplitude marine navigation signal interference, non-cooperative LPI radar signals are weak, the signal-to-interference-plus-noise ratio is low, and they are submerged in noise and interference. At this time, the performance of common radar signal detection methods deteriorates, and the detection performance will be significantly worse. Therefore, there is an urgent need to study a method that can accurately detect and effectively enhance non-cooperative LPI radar signals with low signal-to-interference-plus-noise ratio under the interference of large-amplitude marine navigation signals, laying a foundation for subsequent signal processing processes such as modulation mode recognition and parameter estimation. Summary of the Invention

[0006] The present invention provides a low-intercept radar signal detection and enhancement method based on LPI-DE-GAN, which can accurately detect low-intercept probability radar signals received with various modulation types and mixed with large-amplitude fishing vessel navigation interference signals. After detecting the low-intercept probability radar signals, end-to-end time-domain signal enhancement is performed on them to suppress fishing vessel navigation signals and noise and reconstruct the original low-intercept probability radar signals.

[0007] The object of the present invention is achieved through the following technical solutions: A low-intercept radar signal detection and enhancement method based on LPI-DE-GAN includes the following steps:

[0008] S1. Obtain low-intercept radar probability signal sequences with various different modulations, and obtain a 0 sequence with the same total length as the low-intercept radar signal sequence. Then, jointly form a clean signal sequence with the low-intercept probability signal and the 0 sequence samples; further obtain noise and marine navigation signal sequences, and superimpose them on the clean signal sequence to obtain a mixed signal sequence;

[0009] S2. Slice the clean signal sequence and the mixed signal sequence, and construct labels for the signal detection task; jointly form a data set with the sliced mixed signal sequence, the clean signal sequence, and the labels of the signal detection task;

[0010] S3. Construct an LPI-DE-GAN model for the detection and enhancement of low-intercept radar signals; this model includes a generator G and a discriminator D; it includes the following processes:

[0011] S31. Construct the generator G: The generator G consists of an encoder, a decoder, and a signal detector;

[0012] The encoder is composed of 11 convolutional layers connected in series, and the PReLU activation function is used in each convolutional layer; the mixed signal s noisy is compressed into an encoded vector c by the encoder. Then, a random noise vector z with the same size as c and conforming to the Gaussian distribution is selected, and the encoded vector c is concatenated with the noise vector z as the input of the decoder;

[0013] The decoder is composed of 11 transposed convolutional layers connected in series; except for the last transposed convolutional layer, the PReLU activation function is used in the remaining transposed convolutional layers; the 1st to Nth convolutional layers of the encoder are respectively connected in a skip connection with the 1st to Nth transposed convolutional layers of the decoder, where N is the number of convolutional layers / transposed convolutional layers;

[0014] Based on the concatenated vector of the encoded vector c and the noise vector z, the decoder generates an enhanced signal G(z, s noisy )

[0015] The signal detector consists of a Flatten layer and four fully-connected layers connected in sequence. Except for the last fully-connected layer, the ReLU activation function is used in the other fully-connected layers, and the Dropout technique is adopted to prevent overfitting. The Sigmoid activation function is used in the last fully-connected layer. The signal detector weights and integrates different features of the low-intercept radar signal contained in the encoded vector c, uses the integrated features as the basis for signal detection to input into the Sigmoid activation function, and outputs the test statistic T Net ∈[0,1];

[0016] S32. Construct a discriminator D. The discriminator D is a binary classifier used to discriminate the input signals. The discriminator D discriminates the enhanced signals generated by the generator G as fake, discriminates the clean signals as real, and feeds back the discriminated results to the generator G;

[0017] The discriminator D uses 11 convolutional layers to perform convolutional operations on the data, and introduces a batch normalization layer and a LeakyReLU activation function after each convolutional layer. Then, the features extracted by the convolutional layers are determined through a fully-connected layer and a Softmax activation function;

[0018] S4. Construct the loss function of the LPI-DE-GAN model, set the hyperparameters of the network, initialize the LPI-DE-GAN parameters, and train the LPI-DE-GAN model.

[0019] The beneficial effects of the present invention are as follows: The present invention proposes a low-intercept probability radar signal detection and enhancement method based on LPI-DE-GAN. Aiming at the problem that the detection performance of traditional methods decreases due to the low signal-to-interference-plus-noise ratio of low-intercept probability radar signals under the background of strong marine navigation signal interference, an LPI-DE-GAN model is constructed. Through the mutual confrontation between the generator and the discriminator, the model can accurately detect and enhance low-intercept probability radar signals in a complex electromagnetic environment. When training the network model, the signal enhancement and detection tasks are jointly optimized, improving the working efficiency of the network model. When performing inference, the LPI-DE-GAN model directly detects the received mixed signals. If the test statistic is higher than the threshold, signal enhancement is performed; otherwise, the task terminates. Compared with the traditional method of first filtering the received mixed signals for noise and interference and then detecting, the end-to-end detection process proposed by the present invention saves resources. The present invention improves the detection and enhancement performance of low-intercept LPI radar signals under low signal-to-interference-plus-noise ratio, has strong generalization ability, and is suitable for engineering implementation. Description of the Drawings

[0020] Figure 1 It is the implementation flowchart of the present invention;

[0021] Figure 2For the inference process of LPI-DE-GAN;

[0022] Figure 3 For the generator structure of LPI-DE-GAN;

[0023] Figure 4 For the discriminator structure of LPI-DE-GAN

[0024] Figure 5 For the detection probability broken line graph of the present invention under different signal-to-interference-plus-noise ratios;

[0025] Figure 6 For the false alarm probability broken line graph of the present invention under different signal-to-interference-plus-noise ratios;

[0026] Figure 7 For the enhancement result graph of the low probability of intercept signal of the present invention when the signal-to-interference-plus-noise ratio is -20 dB. Specific implementation manners

[0027] In order to enable those of ordinary skill in the art to better understand the technical solution of the present invention, the technical solution of the present application will be further described below in conjunction with the accompanying drawings and specific embodiments. This example mainly uses the scientific computing software MATLAB R2023a and Pycharm 2023.2.5 to conduct simulation experiments to verify its signal enhancement and detection effects.

[0028] As Figure 1 shown, a low probability of intercept radar signal detection-enhancement method based on LPI-DE (D stands for Detection, and E stands for Enhancement)-GAN (generative adversarial network) of the present invention includes the following steps:

[0029] S1. Obtain low probability of intercept radar probability signal sequences with various different modulations, and obtain 0 sequences with the same total length as the low probability of intercept radar signal sequences. Then, jointly form a clean signal sequence with the low probability of intercept signals and 0 sequence samples, and use the clean signal sequence as a signal enhancement label; then obtain noise (Gaussian white noise in the environment) and marine navigation signal sequences, and superimpose them with the clean signal sequence to obtain a mixed signal sequence; including the following processes:

[0030] S11. Set the parameter range of the low probability of intercept (LPI) radar signal pulse, and obtain LPI radar signal sequences with multiple different modulation methods. The parameter range used in this embodiment is shown in Table 1. In this embodiment, the sampling frequency is set to 50 MHz, and LPI radar signal sequences with four different modulation methods, namely LFM, Bark, Frank, and Costas, are obtained. The length of each LPI radar signal sequence is L; obtain a 0-sequence with the same total length as the multiple LPI radar signal sequences as the comparison sample signal sequence when there is no signal; jointly form a clean signal sequence with all the LPI signals and the 0-sequence samples where 8L is the length of the clean signal sequence, and 2 indicates that the signal contains two dimensions, namely the real part and the imaginary part.

[0031] Table 1

[0032]

[0033] S12. Set the signal-to-interference-plus-noise ratio (SINR) range as [SINR min , SINR max , where SINR min and SINR max are the upper and lower bounds of the SINR range respectively. In this embodiment, SINR min is set to -20 dB, and SINR max is set to 0 dB; the SINR change step size ΔSINR is set to 2 dB; within the SINR range, noise and marine navigation signals at this SINR are generated every other ΔSINR, and the noise and marine navigation signals are superimposed on the clean signal to obtain a mixed signal sequence

[0034] S2. Slice the clean signal sequence and the mixed signal sequence, and construct the labels for the signal detection task; jointly form a dataset with the sliced mixed signal sequence, the clean signal sequence, and the labels for the signal detection task; and divide the dataset into a training set, a validation set, and a test set; including the following processes:

[0035] S21. Set the window length w and the step size s. In this embodiment, the window length w is set to 32768, and the step size s is set to 16384; use the set window length and step size to slice S noisy and S clean to obtain n mixed signal sample blocks and n clean signal sample blocks respectively as the mixed signal sequence and the signal enhancement task label input to the LPI-DE-GAN;

[0036] S22. Obtain "0" or "1" as the label for the LPI radar signal detection task, and all labels y iConstruct the signal detection label sequence y;

[0037]

[0038] When one of the multiple low probability of intercept radar signals in step S1 exists in the mixed signal input to the network model, y i takes "1"; otherwise, y i takes "0";

[0039] S23. Divide the obtained sample blocks and labels into a training set, a validation set, and a test set according to a ratio of 8:1:1. The training set and the validation set are used for the subsequent training of the LPI-DE-GAN model, and the test set is used to verify the model effect.

[0040] S3. Construct an LPI-DE-GAN model for the detection and enhancement of low probability of intercept radar signals; this model includes a generator G and a discriminator D, and its principle is as Figure 2 shown; it includes the following processes:

[0041] S31. Construct the generator G: The generator G is composed of an encoder, a decoder, and a signal detector, and its structure is as Figure 3 shown;

[0042] The encoder is composed of 11 consecutive convolutional layers. The PReLU activation function is used in each convolutional layer, the convolutional kernel size is 32, and the stride is 2. The mixed signal s noisy is compressed into an encoded vector c with a size of 16×1024 by the encoder; then a random noise vector z with the same size as c and conforming to the Gaussian distribution is selected, and the encoded vector c and the noise vector z are concatenated as the input of the decoder;

[0043] The decoder is composed of 11 consecutive transposed convolutional layers; except for the last transposed convolutional layer, the PReLU activation function is used in the remaining transposed convolutional layers; the deconvolution kernel size of each transposed convolutional layer is 32, and the stride is 2; the first to Nth convolutional layers of the encoder are respectively connected in a skip connection with the first to Nth transposed convolutional layers of the decoder, where N is the number of convolutional layers / transposed convolutional layers; Figure 3 In it, the convolutional layers are sequentially denoted as conv1 to conv11 from bottom to top (from input to output), and the transposed convolutional layers are sequentially denoted as deconv11 to deconv1 from bottom to top (from input to output). conv1 is connected in a skip connection with deconv1, conv2 is connected in a skip connection with deconv2, and so on...

[0044] The decoder generates an enhanced signal G(z, s noisy ) based on the concatenated vector of the encoded vector c and the noise vector z;

[0045] The signal detector consists of a Flatten layer and four fully connected layers connected in sequence. Except for the last fully connected layer, the ReLU activation function is used in the other fully connected layers, and the Dropout technique is adopted to prevent overfitting; the Sigmoid activation function is used in the last fully connected layer. The signal detector weights and integrates different features of the low-intercept radar signal contained in the encoded vector c, and uses the integrated features as the basis for signal detection to input into the Sigmoid activation function, and outputs the test statistic T Net ∈[0,1]. When the test statistic is higher than the preset threshold τ Net , it is determined that a low-intercept probability radar signal is detected, and an enhanced signal is generated through the model constructed by the present invention to realize the reconstruction of the low-intercept probability radar signal, achieving the effect of suppressing noise and interference of marine navigation signals; when the test statistic is lower than the threshold τ Net , it is considered that there is no low-intercept probability radar signal in the input mixed signal sequence, and the mixed signal is discarded to process the next group of input signals.

[0046] S32. Construct a discriminator D. The discriminator D is a binary classifier, and its structure is as Figure 4 shown; the discriminator is used to discriminate the input signals. The input signals include enhanced signals and clean signals; the former is the enhanced signal generated by the generator, including the enhanced low-intercept probability radar signal and the enhanced 0 sequence (still approaching 0); the latter clean signal includes the low-intercept probability radar signal and the 0 sequence; at the same time, the clean signal is used as a label.

[0047] The discriminator D discriminates the enhanced signal generated by the generator G as false and the clean signal as true, and feeds back the discriminated result to the generator G, so that the generator G can adjust its own weights to improve the signal enhancement ability.

[0048] When an enhanced signal is input (at the beginning of training), the discriminator determines it as false, and when a label signal is input, the discriminator discriminates it as true. Generally speaking, the discriminator will force the generator to generate a higher-quality low-intercept probability radar signal at this time, that is, the generated enhanced image is getting closer and closer to the low-intercept probability radar signal.

[0049] The discriminator D uses 11 convolutional layers to perform convolutional operations on the data, and introduces a batch normalization layer and a LeakyReLU activation function after each convolutional layer; then the features extracted by the convolutional layers are determined through a fully connected layer and a Softmax activation function.

[0050] The role of the discriminator D is to feedback the quality of the enhanced signal during the training process, so that the generator G can generate a better enhanced signal and improve the signal enhancement ability of the generator. After the training is over, the generator G is used to perform signal enhancement and signal detection on the signal to be recognized, and the discriminator D can be deleted.

[0051] S4. Construct the loss function of the LPI-DE-GAN model, set the hyperparameters of the network, initialize the LPI-DE-GAN parameters, and train the LPI-DE-GAN model using the training set and the validation set; the process includes the following:

[0052] S41. Construct the loss functions of the LPI-DE-GAN generator and discriminator; the generator of the LPI-DE-GAN needs to detect and enhance the LPI radar signal submerged in noise and marine navigation signals. Therefore, the loss function of the generator should include the loss functions of two tasks: signal detection and enhancement. The target loss function of the generator is:

[0053]

[0054] In the formula, V Detect (G) and V Enhance (G) represent the loss functions of the signal detection task and the signal enhancement task respectively, and α and β represent the weights in the signal detection task and the signal enhancement task respectively; denotes taking the average of the input, that is, the expectation; z represents a sample sampled from random noise, and p z (z) refers to the probability distribution of the input noise z; s noisy represents the mixed signal, and p data (s noisy ) is the probability distribution of the mixed signal s noisy ; s clean represents the clean signal, ‖‖‖‖ 2 represents the L 2 loss, and λ is the hyperparameter that controls the L 2 loss; denotes the value of the test statistic T Net output by the detector, y i is the detection label, taking values of 0 or 1, M represents the number of samples, and μ is the hyperparameter that controls the loss of the detection task; D(G(z, s noisy ), s noisy ) represents the output of the discriminator D. The input of the discriminator is the enhanced signal G(z, s noisy ) and the noise signal s noisy , and the output is 1 or 0. If G(z, s noisy ) is closer to the true signal, then D(G(z, s noisy ), s noisy ) approaches 1, then the first term approaches 0, and the loss will be smaller.

[0055] The target loss function of the discriminator is:

[0056]

[0057] S42. Set the hyperparameters of LPI-DE-GAN and initialize the model weights; Set the RMSprop optimizer as the optimizer for the generator and discriminator, with an initial learning rate of 0.001, a batch size of 16, a maximum number of iterations of 250, α and β set to 1, λ set to 10, μ set to 100, and the dropout rate of Dropout set to 0.5; Initialize the LPI-DE-GAN weights using Xavier initialization;

[0058] S43. Use the training set and validation set data to train the LPI-DE-GAN model, perform forward propagation to calculate the loss function values of the LPI-DE-GAN generator and discriminator, and then update the parameters of the network model through the backpropagation algorithm. Repeat this process until the predetermined maximum number of iterations is reached. At the same time, use the RMSprop optimizer to reduce the loss function value during the training process, so that the distribution of the generated enhanced signal gradually approaches the distribution of the real LPI radar signal, realizing the detection and enhancement of low probability of intercept radar signals in the background of a large-amplitude marine navigation signal.

[0059] S5. Use the trained LPI-DE-GAN model for the detection and enhancement of mixed signals; The detector in the generator outputs the test statistic T Net , and set the detection threshold τ Net of the generator to 0.5. When the test statistic T Net is higher than the threshold τ Net , a low probability of intercept radar signal is detected, and the low probability of intercept radar signal is reconstructed to suppress noise and marine interference signals; When the test statistic is lower than the threshold τ Net , it is considered that there is no low probability of intercept radar signal in the input mixed signal sequence.

[0060] In this embodiment, in addition to the test set data, low signal-to-noise ratio and low probability of intercept radar signals that did not participate in the training are also set at the same time. The signal-to-interference-plus-noise ratio range of the signals participating in the test is -40 to 0 dB. When τ Net is 0.5, the changes in the detection probability and false alarm probability are respectively as shown in Figure 5 and Figure 6 . After completing the detection of the low probability of intercept radar signal, LPI-DE-GAN suppresses noise and marine interference signals, enhances the original low probability of intercept radar signal, and reconstructs its time-domain waveform. Figure 7 is the mixed signal, the clean low probability of intercept radar signal, and the enhanced signal when the signal-to-interference-plus-noise ratio is -20 dB.

[0061] Those of ordinary skill in the art will realize that the embodiments described herein are for the purpose of assisting the reader in understanding the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention based on these technical revelations disclosed in the present invention, and these deformations and combinations are still within the scope of protection of the present invention.

Claims

1. A low acquisition radar signal detection and enhancement method based on LPI-DE-GAN, characterized in that: The following steps are involved: S1. Obtain a plurality of differently modulated low intercept probability radar signal sequences, and obtain a 0 sequence with the same total length as the low intercept probability radar signal sequence, and then combine the low intercept probability signal and the 0 sequence samples to form a clean signal sequence; then obtain a noise and a ship navigation signal sequence, and superimpose them with the clean signal sequence to obtain a mixed signal sequence; S2, slice the clean signal sequence and the mixed signal sequence, and construct the label of the signal detection task; the sliced ​​mixed signal sequence and the clean signal sequence, as well as the label of the signal detection task, together constitute a data set; S3. Construct an LPI-DE-GAN model for detecting and enhancing low acquisition radar signals. The model includes a generator G and a discriminator D. The following processes are included: S31, constructing a generator G: the generator G consists of an encoder, a decoder and a signal detector; The encoder consists of 11 convolutional layers in sequence, and the PReLU activation function is used in each convolutional layer; the mixed signal s noisy The encoder compresses it into a coding vector c, then selects a random noise vector z of the same size as c and conforms to the Gaussian distribution, concatenates the coding vector c and the noise vector z as the input of the decoder; The decoder consists of 11 transposed convolutional layers connected in sequence. Except for the last transposed convolutional layer, the remaining transposed convolutional layers use the PReLU activation function. The 1st to Nth convolutional layers of the encoder are skip-connected with the 1st to Nth transposed convolutional layers of the decoder, respectively, where N is the number of convolutional layers / transposed convolutional layers. The decoder generates the enhanced signal G(z,s) based on the concatenation vector of the encoding vector c and the noise vector z. noisy ); The signal detector consists of a Flatten layer and four fully connected layers in sequence. Except for the last fully connected layer, the remaining fully connected layers use the ReLU activation function and the Dropout technology is used to prevent overfitting; the last fully connected layer uses the Sigmoid activation function; The signal detector weights and integrates the different features of the low intercept radar signal contained in the coding vector c, uses the integrated features as the basis for signal detection, and inputs the Sigmoid activation function to output the test statistic T Net ∈[0,1]; When the test statistic is higher than the preset threshold τ Net When τ is detected, it is judged that a low intercept probability radar signal is detected, and an enhanced signal is generated through the model to achieve the reconstruction of the low intercept probability radar signal; when the test statistic is lower than the threshold τ Net When , it is considered that there is no low intercept probability radar signal in the input mixed signal sequence, and the signal is discarded; S32. Construct a discriminator D. The discriminator D is a binary classifier used to identify the input signal. The discriminator D identifies the enhanced signal generated by the generator G as false and the clean signal as true, and feeds the identification result back to the generator G. The discriminator D uses 11 convolutional layers to perform convolution operations on the data, and introduces a batch normalization layer and a LeakyReLU activation function after each convolutional layer; then the features extracted by the convolutional layer are judged by a fully connected layer and a Softmax activation function; S4. Construct the loss function of the LPI-DE-GAN model, set the network hyperparameters, initialize the LPI-DE-GAN parameters, and train the LPI-DE-GAN model.

2. The low acquisition radar signal detection-enhancement method based on LPI-DE-GAN according to claim 1, characterized in that: The step S1 includes the following process: S11, setting the parameter range of low probability of intercept radar signal pulses, obtaining low probability of intercept radar signal sequences of multiple different modulation modes; obtaining a 0 sequence with the same total length as the multiple low probability of intercept radar signal sequences as a comparison sample signal sequence when there is no signal; and combining all low probability of intercept signals and 0 sequence samples into a clean signal sequence; S12, set the SINR range to [SINR min ,SINR max ], SINR min and SINR max are the upper and lower bounds of the signal to interference and noise ratio range respectively; the step size of the signal to interference and noise ratio change is ΔSINR; within the SINR range, noise and ship navigation signals under the signal to interference and noise ratio are generated at each interval ΔSINR, and the noise and ship navigation signals are superimposed on the clean signal to obtain a mixed signal sequence.

3. The low acquisition radar signal detection-enhancement method based on LPI-DE-GAN according to claim 1, characterized in that: The step S2 includes the following process: S21, set the window length w and step length s, and use the set window length and step length to calculate S noisy and S clean Slice to obtain n mixed signal sample blocks s noisy and n clean signal sample blocks s clean , respectively as the mixed signal sequence and signal enhancement task label of LPI-DE-GAN input; S22, obtain "0" or "1" as the label of the low probability of intercept radar signal detection task, all labels y i Constructing a signal detection tag sequence y; When the mixed signal input to the network model contains one of the multiple low probability of intercept radar signals in step S1, y i Take "1"; otherwise, y i Take "0".

4. The low acquisition radar signal detection-enhancement method based on LPI-DE-GAN according to claim 1, characterized in that: The step S4 includes the following process: S41. Construct the loss function of LPI-DE-GAN generator and discriminator. The loss function of the generator is: Where V Detect (G) and V Enhance (G) represents the loss functions of the signal detection task and the signal enhancement task, respectively. α and β represent the weights in the signal detection task and the signal enhancement task, respectively. represents the average value of the input; z represents the sample obtained from random noise, p z (z) refers to the probability distribution of the input noise z; s noisy represents a mixed signal, p data (s noisy ) is a mixed signal s noisy The probability distribution of clean represents the clean signal, ‖‖‖‖2 represents the L2 loss, and λ is the hyperparameter controlling the L2 loss. The test statistic T representing the detector output Net The value of y i is the detection label, which takes a value of 0 or 1, M represents the number of samples, and μ is a hyperparameter that controls the loss of the detection task; D(G(z,s noisy ),s noisy ) represents the output of the discriminator D; The objective loss function of the discriminator is: S42. Set the hyperparameters of LPI-DE-GAN and initialize the model weights; S43, train the LPI-DE-GAN model, perform forward propagation to calculate the loss function values ​​of the LPI-DE-GAN generator and discriminator, and then update the parameters of the network model through the back propagation algorithm, and repeat this process until the predetermined maximum number of iterations is reached.

Images