Convolutional neural network-based target component extraction method under interference background
By applying convolutional neural networks in ISRJ technology to analyze the characteristics of interference and targets, the problems of model mismatch and target information loss in ISRJ technology are solved, and efficient target component extraction and anti-interference effect are achieved.
Patent Information
- Application Number
- CN202411879728.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-19
- Publication Date
- 2025-03-21
AI Technical Summary
The existing anti-intermittent sampling direct forwarding interference (ISRJ) technology is prone to model mismatch and loss of target information, and it is difficult to effectively extract target components in the context of interference.
The target component extraction method based on convolutional neural network (CNN) is adopted, and by establishing an echo model, analyzing the characteristics of interference and targets, designing generators and discriminators, and using an adversarial learning process to complete the extraction of target components in mixed signals.
The target component extraction in the context of interference is achieved, the anti-interference ability is improved, the target information is lost, and the effectiveness of the method is verified.
Smart Images
Figure CN119669728A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target component extraction, and in particular to a target component extraction method based on a convolutional neural network under an interference background. Background Art
[0002] The intermittent sampling forwarding interference generated by the digital radio frequency memory is generated by intercepting, delaying and forwarding the transmitted signal. It is partially coherent with the transmitted signal and can obtain a specific gain after signal processing to deceive the radar[1][2][3]. Moreover, by utilizing the digital processing capability of DRFM, the interference parameters can be easily adjusted to form different interference strategies, posing a huge threat to the radar system[4].
[0003] There are many strategies for ISRJ, including intermittent sampling direct forwarding jamming (ISDRJ), intermittent sampling repeated forwarding jamming (ISPRJ), and intermittent sampling cyclic forwarding jamming (ISCRJ) [5][6]. The most typical jamming strategy is ISDRJ, which intercepts a segment from the transmitted signal, forwards it once, and then repeats the sampling and forwarding process until the pulse ends. After the jamming signal is processed by pulse compression (PC), a main false target and multiple secondary false targets will be formed in the fast time dimension. On this basis, [6] proposed the jamming strategies of ISPRJ and ISCRJ, in which the former forwards the intercepted segment multiple times to form multiple false target groups. Each false target group consists of a main false target and multiple secondary false targets. ISCRJ not only transmits the current intercepted segment forward, but also transmits all previously intercepted segments backward. The false targets after PC have a more comprehensive and irregular distribution range. Since then, more researchers have analyzed the performance of ISRJ and further proposed improved jamming strategies. Reference [7] proposed adding a modulated convolution kernel to the ISRJ segment. This kind of interference is not only highly deceptive, but also highly suppressive, greatly increasing the difficulty of radar confrontation.
[0004] Radar countermeasures against ISRJ technology mainly follow a two-step strategy: 1) interference parameter estimation; 2) interference suppression. For the first step, [8] and [9] proposed parameter estimation methods based on the time domain and time-frequency domain, respectively. Considering the variability of ISRJ strategies and parameters, the above interference parameter estimation methods are prone to model mismatch problems. For the second step,
[10] and
[11] proposed interference suppression methods based on time domain interference reconstruction and time-frequency domain filter design, respectively. Interference filtering is performed on the basis of the first step. Therefore, if the parameter estimation is inaccurate, interference filtering will be insufficient. In addition, when the interference overlaps with the target signal, interference suppression will also cause the loss of target information. In short, the existing anti-ISRJ technical methods are prone to model mismatch and target information loss. Therefore, a new anti-interference idea is urgently needed. Summary of the invention
[0005] In view of the fact that the above-mentioned existing anti-ISRJ technical methods are prone to model mismatch and target information loss, the present invention is proposed.
[0006] Therefore, the purpose of the present invention is to provide a method for extracting target components based on convolutional neural networks under interference background, which aims to: analyze the changing characteristics of interference and the stable characteristics of the target by establishing an echo model, and then model the anti-interference task by analogy with the speech signal separation task. Finally, the network structure and loss function are designed to complete the task of extracting the target component in the mixed signal.
[0007] To solve the above technical problems, the present invention provides the following technical solutions: A method for extracting target components based on convolutional neural networks under interference background, including S1, a generator, and a G network to perform target component extraction. The input of G is a mixed signal x and a potential representation z, which represents environmental noise, in order to introduce the distribution characteristics of noise n, so as to obtain x faster and more accurately. t =xx j -n; the output of G is a pure signal Here we expect to input the mixed signal x into the network and obtain the clean target signal component x t ;
[0008] S2, discriminator, D is a typical binary classifier, whose input is the real target echo x t Or the target component generated by G The output is 0 or 1, where 0 means the detector judges the input as a generated sample, and 1 means the detector judges the input as a real sample. D contains multiple "one-dimensional convolution + BN + LeakyReLU" structures to continuously extract features from the input. Finally, the output is constrained to 0 to 1 through the Sigmoid activation function;
[0009] S3, loss function, the purpose of D is to find the difference between real samples and generated samples. On the contrary, the purpose of G is to generate samples that D cannot distinguish as much as possible. This adversarial learning process is described as a minimax game between G and D, the least squares GAN (LSGAN).
[0010] As a preferred solution of the target component extraction method based on convolutional neural network under interference background of the present invention, the ISDRJ can be expressed as:
[0011]
[0012] As a preferred solution of the target component extraction method based on convolutional neural network under interference background of the present invention, wherein: sis the sampling period of interference, τ is the pulse width of interference, is the repetition frequency of interference, s(t) is the transmitted signal;
[0013] The first component is the transmitted signal s(t) and the duty cycle The product of has the same form as the transmitted signal. The terms and represent the modulation of the initial transmitted signal by each harmonic of the sampling pulse.
[0014] As a preferred solution of the target component extraction method based on convolutional neural network under interference background described in the present invention, wherein: wherein 'Conv', 'Enc', and 'Dec' represent convolutional layer, encoder layer, and decoder layer respectively.
[0015] As a preferred solution of the target component extraction method based on convolutional neural network under interference background described in the present invention, wherein:
[0016]
[0017]
[0018] In order to more intuitively supervise the generator to generate more real samples, the L1 norm between the generated samples and the real samples is added to the loss function. The size of the L1 norm is controlled by a new hyperparameter λ. Therefore, we change the G loss of LSGAN (Eq.7) to the following form:
[0019]
[0020] The beneficial effects of the present invention are as follows: by establishing an echo model, the changing characteristics of the interference and the stable characteristics of the target are analyzed, and then the anti-interference task is modeled by analogy with the speech signal separation task. Finally, the network structure and loss function are designed to complete the task of extracting the target component in the mixed signal. At the same time, the experiments in the specific real-time mode of the present invention demonstrate the effectiveness of the proposed method. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:
[0022] Figure 1 This is a working principle diagram of a typical ISRJ of the target component extraction method based on a convolutional neural network under the interference background of the present invention.
[0023] Figure 2This is a diagram of the GAN training process of the target component extraction method based on a convolutional neural network under the interference background of the present invention.
[0024] Figure 3 Schematic diagram of the network structure of the target component extraction method based on convolutional neural network under the interference background of the present invention.
[0025] Figure 4 A diagram showing the signal processing flow of the target component extraction method based on a convolutional neural network under interference background of the present invention.
[0026] Figure 5 Figure 2 is a signal processing flow chart of the target component extraction method based on convolutional neural network under interference background of the present invention.
[0027] Figure 6 It is a detection probability map of the target component extraction method based on convolutional neural network under the interference background of the present invention. DETAILED DESCRIPTION
[0028] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.
[0029] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0030] Reference Figure 1-6 , is an embodiment of the present invention, and provides a method for extracting target components based on a convolutional neural network under interference background. The method for extracting target components based on a convolutional neural network under interference background includes:
[0031] The working principle of a typical ISRJ[9] (ISDRJ, ISPRJ and ISCRJ) is as follows: Figure 1 As shown in the figure, the variability of the ISRJ strategy is illustrated. Taking ISDRJ as an example, the diversity of jamming parameters and the deception and suppression characteristics of jamming are introduced [6]. ISDRJ can be expressed as:
[0032]
[0033] Where T s is the sampling period of interference, τ is the pulse width of interference, is the repetition frequency of interference, and s(t) is the transmitted signal.
[0034] The first component is the transmitted signal s(t) and the duty cycle The product of has the same form as the transmitted signal. The terms and represent the modulation of the initial transmitted signal by each harmonic of the sampling pulse.
[0035] Based on the digital processing capability of DRFM, the three classic jamming strategies can be switched arbitrarily, and the pulse width and sampling period can be changed. In particular, after adding random convolution kernel modulation, it brings a more significant threat to the radar. The jamming situation after adding convolution kernel modulation is given [such as Figure 1 as shown].
[0036]
[0037] in represents a noise convolution kernel with random amplitude and distribution, L is the convolution kernel length, A l is a random number sequence, δ(·) is a unit impulse function, t j is the convolution kernel j th Order moment.
[0038] After PC, we get formula (3):
[0039]
[0040] Among them, y(t) is the response of the target echo after matched filtering, and h(t) is the matched filter.
[0041] The first term in formula (3) is the main false target, which reflects the characteristics of deception interference. The second term in formula (3) is the false target group, where A l represents the amplitude randomness of the false target group, It indicates the expansion of the false target group in the time domain, reflecting the repressive characteristics.
[0042] The echo signal with interference can be expressed as:
[0043] x=x t +x jc +n (4)
[0044] Where: x t Indicates the target echo signal:
[0045] x t =As(t-t0) (5)
[0046] in Target echo signal x t is highly similar to the transmitted signal s. The difference is that x t Added a range modulation.
[0047] Compared with the interference signal with changing strategies and parameters, the characteristics of the target signal are relatively stable and have a higher correlation with the matched filter. Therefore, the target detection task under interference background can be transformed from the existing two-stage idea of "interference parameter estimation + interference filtering" to extracting the target component, which can be expressed as x t +x jc +n->x t . Similar to this task, speech signal separation also extracts the signal of interest from the mixed signal, that is, s1+s2+n->s1, where s1 and s2 represent speaker 1 and speaker 2 respectively, and n represents noise. In recent years, deep learning has surpassed traditional signal processing algorithms in speech separation, which also makes it possible to apply deep learning to the task of extracting target signals in interference backgrounds.
[0048] Target component extraction method based on CNN.
[0049] The network structure draws on a classic method in speech signal separation, namely the Speech Enhancement Generative Adversarial Network (SEGAN), which includes a generator and a discriminator\cite{12}. The training process and network structure are as follows: Figure 2 and Figure 3 shown.
[0050] 1) Generator
[0051] In our case, the G network performs target component extraction. The input of G is the mixed signal x and the latent representation z, which represents the environmental noise, in order to introduce the distribution characteristics of the noise n in order to obtain x faster and more accurately. t =xx j -n. The output of G is a pure signal Here we expect to input the mixed signal x into the network and obtain the clean target signal component x t .
[0052] The G network structure follows the encoder-decoder principle. The included convolutional layers give it a strong ability to capture time series features and a large receptive field. The network structure is as follows Figure 3 In the encoding stage, the input signal passes through several dilated convolutional layers and parameter rectified linear units (PReLUs)
[14] The projection is compressed into a new space. The output of the encoding phase is represented by c, which is a compressed vector. The dilated convolution is used here, and studies have shown that this structure increases the stability of GAN training.
[15] . Subsequently, a latent vector z with the same length as the pulse repetition time (PRT) is introduced and concatenated with the vector c as the input of the decoding stage. In the decoding stage, multiple "stepped deconvolution + PReLUs" structures are used to decode the compressed vector c. During encoding, if all information flows are forcibly compressed to the over-compression bottleneck, a lot of low-level detail information may be lost, thereby incorrectly reconstructing the target information. Therefore, a skip connection is added before and after the compressed vector to pass the underlying fine-grained information to the decoding stage. In addition, this structure provides better training behavior because the gradient can be transferred throughout the structure.
[16] Interactive flow.
[0053] 2) Discriminator
[0054] D is a typical binary classifier, whose input is the true target echo x t Or the target component generated by G The output is 0 or 1, where 0 means the detector judges the input as a generated sample, and 1 means the detector judges the input as a real sample. D contains multiple "one-dimensional convolution + BN + LeakyReLU" structures to continuously extract features from the input. Finally, the output is constrained to 0 to 1 through the Sigmoid activation function.
[0055] Figure 3 Here, 'Conv', 'Enc', and 'Dec' represent the convolutional layer, encoder layer, and decoder layer, respectively.
[0056] 3) Loss Function
[0057] The goal of D is to find the difference between real samples and generated samples. In contrast, the goal of G is to generate samples that D cannot distinguish as much as possible. This adversarial learning process is described as a minimax game between G and D, the least squares GAN (LSGAN).
[0058]
[0059] In order to more intuitively supervise the generator to generate more real samples, the L1 norm between the generated samples and the real samples is added to the loss function. The size of the L1 norm is controlled by a new hyperparameter λ. Therefore, we change the G loss of LSGAN (Eq.7) to the following form:
[0060]
[0061] simulation
[0062] This section simulates and verifies the target component extraction method under the ISRJ background through the parameters shown in Table 1.
[0063] Table 1 Simulation parameters
[0064]
[0065] like Figure 4 As shown, where JSR=9dB, SNR=18dB, (a) interference echo signal, (b) PC result of echo signal, (c) target component, (d) PC result of target component.
[0066] When the interference signal ratio (JSR) is 9dB, the signal-to-noise ratio (SNR) is 18dB, the target distance unit is 201, the interference strategy is ISDRJ, and the interference distance unit is 388, the echo signal and its pulse compression results are as follows: Figure 4 (a) and Figure 4 (b) The output signal of the network and the pulse compression result of the signal are shown as follows: Figure 4 (c) and Figure 4 (d) as shown. Figure 5 and Figure 4 The difference is that the SNR is set to -9dB, and the target and interference distance units are set to 465 and 632 respectively. The reason for choosing JSR as -9dB here is to meet the actual scenario of deception interference. Figure 4 and Figure 5 The relationship between the interference and the target can be clearly seen in the figure; the purpose of selecting SNRs of 18dB and -9dB is to illustrate that the proposed method can extract the target component under both low SNR and high SNR, but the SNR cannot be too low, otherwise the positional relationship between the target and the interference signal after PC cannot be intuitively seen from the figure.
[0067] Figure 4 and Figure 5 It shows that under different signal-to-noise ratios, the proposed method can accurately extract the target signal from the mixed signal
[0068] Figure 5 JSR=9dB, SNR=-9dB, (a) interference echo signal, (b) PC result of echo signal, (c) target component, (d) PC result of target component.
[0069] At the same time, the detection probability under different SNR and JSR is given. It is worth noting that the interference strategies in this part of the experiment include ISDRJ, ISCRJ and ISPRJ. Specifically, the components extracted after PC and the matching coefficient results are sent to the Constant False Alarm Rate (CFAR) detector to count the target detection probability and interference detection probability under different false alarm probabilities. When the false alarm probability is 10 -4 When , the target detection probability and interference detection probability under different JSR and SNR are as follows: Figure 6 (a) and Figure 6 As shown in (b), the number of Monte Carlo experiments is 10,000, and the unprocessed echo signal is directly sent to the CFAR detector.
[0070] from Figure 6 (a) It can be seen that the detection probability of this method under different SNR and JSR is higher than that of direct detection without processing. In addition, Figure 6 (b) shows that, without processing, the probability of interference detection is very high. However, when the anti-interference method is used, the probability of interference detection is almost 0, which fully demonstrates the effectiveness of the proposed anti-interference method;
[0071] Figure 6 In the detection probability table, (a) target detection probability, (b) interference detection probability.
[0072] This paper proposes a target component extraction method based on CNN under interference background. The idea of target component extraction is used to realize anti-interference target detection, in which the CNN framework is introduced to reconstruct the target component and suppress the interference component. The experiment proves the effectiveness of the proposed method. CNN is used for the first time to extract target echo from mixed echo signals, providing a new anti-interference idea.
[0073] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.
Claims
1. A method for extracting target components based on convolutional neural network under interference background, characterized in that: include: S1, generator, G network performs target component extraction. The input of G is the mixed signal x and the potential representation z, which represents the environmental noise, in order to introduce the distribution characteristics of the noise n, so as to obtain x faster and more accurately. t =xx j -n; the output of G is a pure signal Here we expect to input the mixed signal x into the network and obtain the clean target signal component x t ; S2, discriminator, D is a typical binary classifier, whose input is the real target echo x t Or the target component generated by G The output is 0 or 1, where 0 means the detector judges the input as a generated sample, and 1 means the detector judges the input as a real sample. D contains multiple "one-dimensional convolution + BN + LeakyReLU" structures to continuously extract features from the input. Finally, the output is constrained to 0 to 1 through the Sigmoid activation function; S3, loss function, the purpose of D is to find the difference between real samples and generated samples. On the contrary, the purpose of G is to generate samples that D cannot distinguish as much as possible. This adversarial learning process is described as a minimax game between G and D, the least squares GAN (LSGAN).
2. The method for extracting target components based on a convolutional neural network under interference background according to claim 1 is characterized in that: The ISDRJ can be expressed as:
3. The method for extracting target components based on convolutional neural network under interference background according to claim 2 is characterized in that: Where T s is the sampling period of interference, τ is the pulse width of interference, is the repetition frequency of interference, s(t) is the transmitted signal; The first component is the transmitted signal s(t) and the duty cycle The product of has the same form as the transmitted signal. The terms and represent the modulation of the initial transmitted signal by each harmonic of the sampling pulse.
4. The method for extracting target components based on a convolutional neural network under interference background according to claim 3 is characterized in that: Wherein 'Conv', 'Enc', and 'Dec' represent convolutional layer, encoder layer, and decoder layer, respectively.
5. The method for extracting target components based on a convolutional neural network under interference background according to claim 4 is characterized in that: Said The goal of D is to find the difference between real samples and generated samples. In contrast, the goal of G is to generate samples that D cannot distinguish as much as possible. This adversarial learning process is described as a minimax game between G and D, the least squares GAN (LSGAN). In order to more intuitively supervise the generator to generate more real samples, the L1 norm between the generated samples and the real samples is added to the loss function. The size of the L1 norm is controlled by a new hyperparameter λ. Therefore, we change the G loss of LSGAN (Eq.7) to the following form: