Radar signal intra-pulse modulation recognition method based on denoising guided decoupling network

Abstract

The present invention discloses a method for identifying pulse intra-modulation of radar signals based on a denoising guided decoupling network, and specifically proposes a denoising guided decoupling network DGDNet based on the Inception structure, which simultaneously completes the denoising and identification of radar signals in an end-to-end manner. The global features of the radar signal time-frequency diagram are obtained by a global feature extractor, and the feature decoupler separates the pure radar signal expression (PSR) from the noise signal expression (NSR) for learning a modulation pattern recognizer in a low signal-to-noise ratio environment, and finally the classification task is completed by the modulation pattern recognizer. In order to expand the distance between the pure radar signal expression (PSR) and the noise signal expression (NSR), the concept of signal-noise mutual information (SNMI) loss is proposed. Experimental results show that the recognition accuracy of this method is 98.75% under a signal-to-noise ratio of ‑8dB, and the recognition accuracy of radar signals of 12 modulation formats under a ‑10dB environment is 89.25%.

Description

Technical Field

[0001] The invention belongs to the field of pattern recognition and information technology and can be used in radar electronic warfare. More specifically, the invention relates to a radar signal pulse intra-modulation recognition method based on a denoising guided decoupling network. Background Art

[0002] Pulse modulation recognition of radar signals is a key technology in radar electronic warfare. It plays an important role in modern electronic support measurement systems, electronic intelligence systems and radar early warning receivers. Accurately identifying the pulse modulation mode of radar signals helps to estimate the performance of radar transmitters and improve the accuracy of radar signal parameter estimation. However, the pulse compression technology used to improve the range resolution of pulse radars greatly reduces the power spectral density of radar signals. Therefore, under normal radar working conditions, the signal-to-noise ratio of the received radar signal is always significantly reduced, which seriously affects the recognition accuracy of the radar signal modulation mode. The accurate recognition of a wide range of modulation types of radar signals in low signal-to-noise ratio environments remains to be solved.

[0003] Therefore, it is necessary to preprocess the radar signal to eliminate the interference of the noise signal on the original radar signal. In the process of radar signal denoising, most of the focus is on traditional denoising methods, and not enough work has been done on radar signal denoising and recognition using neural network methods. In addition, in some studies on radar signal denoising using neural network, the extracted time-frequency features still need to be optimized when the noise is large, and the denoising effect needs to be strengthened. Summary of the invention

[0004] The technical problem to be solved by the present invention is mainly to provide a radar signal modulation recognition method which can achieve obvious denoising effect and accurately identify the signal modulation format.

[0005] In order to achieve the above objectives, the present invention constructs a decoupling network DGDNet based on denoising guidance, which is based on the Inception structure and simultaneously completes the denoising and recognition of radar signals in an end-to-end manner. The present invention provides a radar signal pulse intra-modulation recognition method based on a denoising guided decoupling network, comprising the following steps:

[0006] The radar signal is converted into a time-frequency image through the time-frequency analysis method CTFD;

[0007] Using the time-frequency image as an input of a global feature extractor of a denoising guided decoupling network to obtain a global feature that incorporates radar signals and noise signals;

[0008] Using the global feature as an input of a feature decoupler of a denoising guided decoupling network to obtain a pure radar signal feature expression and a noise signal feature expression, wherein the noise signal feature expression is used to train and optimize the feature decoupler;

[0009] The pure radar signal feature expression is used as the input of the modulation mode recognizer of the denoising guided decoupling network, and the modulation mode recognizer is trained to perform the radar signal modulation format recognition and classification task to obtain the modulation format recognition result.

[0010] Preferably, the step of converting the radar signal into a time-frequency image by using a time-frequency analysis method CTFD comprises:

[0011] Based on the time-frequency analysis method CWD, the kernel function is improved to obtain the time-frequency analysis method CTFD. The expression of the improved kernel function is:

[0012]

[0013] Where τ is the horizontal axis delay, ν is the vertical axis frequency shift, γ is the shape coefficient, and ξ is the scale coefficient, which are used to adjust the shape and size of the kernel function respectively;

[0014] The time-frequency analysis method CTFD is used to perform time-frequency analysis on radar signals of various modulation formats to obtain a time-frequency image.

[0015] Preferably, the global feature extractor comprises: three filter cascade layers and convolution kernels of different sizes, and two small branches before each cascade layer are used to extract features of different scales.

[0016] Preferably, the feature decoupler comprises: a pure radar feature extractor and a noise feature extractor.

[0017] Preferably, the pure radar feature extractor includes: 4 Inception_A modules, 1 Reduction_A module, 7 Inception_B modules and 1 deconvolution module.

[0018] Preferably, the noise feature extractor includes: 1 Inception_A module, 1 Reduction_A module, 2 Inception_B modules and 1 deconvolution module.

[0019] Preferably, the modulation pattern recognizer includes: 1 Reduction_B module, 3 Inception_C modules and 1 FC layer, and the FC layer is connected to the SoftMax layer for converting the pure radar signal feature expression into a probability value.

[0020] Preferably, the step of training the modulation mode identifier to perform the radar signal modulation format recognition and classification task further includes:

[0021] The prediction results are compared with the actual modulation pattern labels using cross entropy as the loss function;

[0022] The modulation pattern recognizer is trained and optimized by minimizing the cross entropy loss function.

[0023] Preferably, after the step of using the global feature as the input of the feature decoupler to obtain the pure radar signal feature expression and the noise signal feature expression, the method further comprises:

[0024] Constructing a signal-noise mutual information loss function according to the pure radar signal expression and the noise signal characteristic expression;

[0025] The feature decoupler is trained and optimized by minimizing the signal-noise mutual information loss function.

[0026] Preferably, the formula of the signal-noise mutual information loss function is as follows:

[0027]

[0028]

[0029] Where X PSR and X NSR Respectively represent the characteristic expression of pure radar signal and the characteristic expression of noise signal, P p_n represents the joint probability distribution between two feature expressions, P psr and P nsr They are the marginal probability distributions of the two feature expressions, and are minimized by SN To make the two feature expressions more independent.

[0030] The technical solution provided by the present invention has the following beneficial effects:

[0031] (1) This paper proposes a new denoising guided decoupling network DGDNet based on the Inception structure, which simultaneously completes the denoising and recognition of radar signals in an end-to-end manner.

[0032] (2) The feature map of the radar signal time-frequency image is obtained through the global feature extractor, and the pure radar signal expression (PSR) is separated from the noise signal expression (NSR) through the feature decoupler. This is used to train the radar signal modulation pattern recognizer in a low signal-to-noise ratio environment. Finally, the classification task is completed through the modulation pattern recognizer.

[0033] (3) In order to expand the distance between the pure radar signal representation (PSR) and the noise signal representation (NSR), the concept of signal-noise mutual information (SNMI) loss function is proposed. Experimental results show that the recognition accuracy of this method is 98.75% under -8dB SNR and 89.25% under -10dB environment for radar signals with 12 modulation formats. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0035] Figure 1 Flow chart of a radar signal modulation identification method according to an embodiment of the present invention;

[0036] Figure 2 It is an overall framework diagram of the system in the embodiment of the present invention;

[0037] Figure 3 is a contour map of the CTFD kernel function of the time-frequency analysis in the embodiment of the present invention, wherein Figure 3 (a) is the overall contour map, Figure 3 (b) Figure 3 (a) A partial enlarged view;

[0038] Figure 4 A network structure diagram of DGDNet in an embodiment of the present invention;

[0039] Figure 5 This is a structural diagram of a global feature extractor in an embodiment of the present invention;

[0040] Figure 6 is a structural diagram of a characteristic decoupler in an embodiment of the present invention, wherein Figure 6 (a) is Inception_A, Figure 6 (b) is Inception_B, Figure 6 (c) is Reduction_A, Figure 6 (d) is Deconv-module;

[0041] Figure 7 : is a structural diagram of a modulation mode identifier in an embodiment of the present invention, wherein Figure 7 (a) is Reduction_B, Figure 7 (b) is Inception_C;

[0042] Figure 8 It is the time-frequency image of 12 modulation modes used in the embodiment of the present invention;

[0043] Fig. 9 This is a comparison chart of the recognition accuracy of the method of the present invention and other three networks under different signal-to-noise ratios;

[0044] Fig.10 The figure is a comparison chart of the recognition accuracy of the method of the present invention, DGDNet (without using signal-noise mutual information loss SNMI loss), and Inceptionv4 under different signal-to-noise ratios. DETAILED DESCRIPTION

[0045] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0046] like Figure 1 , Figure 2 As shown, an embodiment of the present invention provides a radar signal pulse intra-modulation recognition method based on a denoising guided decoupling network, comprising the following steps:

[0047] S1: The radar signal is converted into a time-frequency image (TFI) through the time-frequency analysis method CTFD.

[0048] The time-frequency analysis method CTFD is a Cohen-type time-frequency analysis method. It is a new method that improves the kernel function based on the time-frequency analysis method CWD. The kernel function is distributed along the axis and is elliptical, which meets the characteristics of most frequency modulation functions such as sinusoidal frequency modulation signal SFM, quadratic frequency modulation signal EQFM, and phase modulation signal PSK fuzzy domain mutual terms distributed on the axis and around it, making the new kernel function time-frequency analysis method CTFD have better noise resistance;

[0049] The expression of the improved kernel function is:

[0050]

[0051] Where τ is the horizontal axis delay, ν is the vertical axis frequency shift, γ is the shape coefficient, and ξ is the scale coefficient, which is used to adjust the shape and size of the kernel function. The contour map of the kernel function of the time-frequency analysis method CTFD is shown in Figure 3 shown.

[0052] S2: A denoising guided decoupling network (DGDNet) is proposed. The network structure of the denoising guided decoupling network is as follows: Figure 4 As shown, it includes: a global feature extractor, a feature disentangler and a modulation mode recognizer.

[0053] S3: Using the time-frequency image as the input of the global feature extractor to obtain the global feature of the fused radar signal and the noise signal.

[0054] The global feature extractor is a stem module that aims to obtain deep features of the input time-frequency image. Figure 5As shown in the figure, it is roughly composed of nine layers, of which there are three filter cascade layers. Correspondingly, the two small branches before each cascade layer are used to extract features of different scales. The initial size of the time-frequency image (TFIs) is 1024×1024. After image resizing and normalization, it is converted into a 299×299 tensor. The output contains both useful information and noise, that is, the global features of the fused radar signal and the noise signal are obtained.

[0055] S4: Using the global feature as the input of the feature decoupler to obtain a pure radar signal feature expression and a noise signal feature expression.

[0056] The structure of the feature decoupler is as follows Figure 6 As shown, it includes: Inception_A, Inception_B, Reduction_A, and deconvolution module Deconv-module;

[0057] Specifically, Figure 4 As shown, the feature decoupler is divided into: Pure radar feature extractor and noise feature extractor;

[0058] The clean radar feature extractor consists of: 4 Inception_A modules, 1 Reduction_A module, 7 Inception_B modules, and 1 deconvolution module. The Inception modules (4 Inception_A modules and 7 Inception_B modules) are used to extract useful signal features hidden in the time-frequency images. A reduction layer (1 Reduction_A module) is applied to reduce the image size. The clean radar signal representation (PSR) can be used to reconstruct the denoised time-frequency images through the deconvolution module.

[0059] Noise feature extractor: includes: 1 Inception_A module, 1 Reduction_A module, 2 Inception_B modules and 1 deconvolution module. The output of the noise feature extractor is the noise signal representation (NSR), which can be used to reconstruct the noisy image through the deconvolution module.

[0060] In order to expand the distance between the pure radar signal representation (PSR) and the noise signal representation (NSR), the concept of signal-noise mutual information (SNMI) loss function is proposed, and its formula is as follows:

[0061]

[0062]

[0063] Where XPSR and X NSR Respectively represent the characteristic expression of pure radar signal and the characteristic expression of noise signal, P p_n represents the joint probability distribution between two feature expressions, P psr and P nsr They are the marginal probability distributions of the two feature expressions, and are minimized by SN To make the two feature expressions more independent.

[0064] In this embodiment, α is used to normalize the signal-noise mutual information loss function to match other losses, where the value of α is set to 0.25.

[0065] S5: Using the pure radar signal feature expression as the input of the modulation mode identifier, training the modulation mode identifier to perform radar signal modulation format recognition and classification tasks, and obtaining a modulation format recognition result.

[0066] Use cross entropy as the loss function to compare the prediction results with the actual modulation mode labels;

[0067] The modulation pattern recognizer is trained and optimized by minimizing the cross entropy loss function.

[0068] The structure of the modulation pattern recognizer is as follows: Figure 7 As shown, it includes Reduction_B module and Inception_C module;

[0069] Specifically, Figure 4 As shown in the figure, the modulation pattern recognizer includes: 1 Reduction_B module, 3 Inception_C modules and 1 FC layer. The FC layer is followed by a SoftMax layer to convert the feature representation into a probability value. Based on the pure radar feature extractor, the modulation pattern recognizer is designed to extract distinguishable feature representations from PSR and realize the classification of at least 12 modulation formats.

[0070] In the step of training the modulation mode recognizer to perform the radar signal modulation format recognition and classification task through the pure radar signal feature expression, it also includes:

[0071] Use cross entropy as the loss function to compare the prediction results with the actual modulation mode labels;

[0072] The modulation pattern recognizer is trained and optimized by minimizing the cross entropy loss function.

[0073] In this embodiment, intra-pulse modulation format recognition is performed on radar signals of 12 modulation formats, and the simulation signal parameters are shown in Table 1:

[0074] Table 1 Simulation signal parameters

[0075]

[0076]

[0077] 12 time-frequency images are identified as follows Figure 8 As shown, it can be seen that the method of the present invention can accurately classify 12 modulation formats.

[0078] like Fig. 9 As shown, Fig. 9 This is a comparison chart of the recognition accuracy of the present invention and other three networks under different signal-to-noise ratios. Table 2 shows the specific indicators of the recognition rates of the four networks under different signal-to-noise ratios.

[0079] Table 2 Specific indicators of the recognition rate of the four networks under different signal-to-noise ratio conditions

[0080]

[0081] from Fig. 9 From Table 2, we can see that the recognition accuracy of the DGDNet proposed in the present invention is better than that of the other three networks in a low signal-to-noise ratio environment.

[0082] like Fig.10 As shown, Fig.10 This is a comparison chart of the recognition accuracy of the three networks of DGDNet (using signal-noise mutual information loss SNMI loss), DGDNet (not using signal-noise mutual information loss SNMI loss), and Inceptionv4 under different signal-to-noise ratios. Table 3 shows the specific indicators of the recognition accuracy of the three networks of DGDNet (using signal-noise mutual information loss SNMI loss), DGDNet (not using signal-noise mutual information loss SNMI loss), and Inceptionv4 under different signal-to-noise ratios:

[0083] Table 3 Specific indicators of recognition accuracy of three networks under different signal-to-noise ratios

[0084]

[0085]

[0086] from Fig.10 From Table 3, we can see that the recognition effect of the DGDNet (using signal-noise mutual information SNMI loss) network of the present invention is more prominent in a low signal-to-noise ratio environment.

[0087] The present invention discloses a radar signal pulse intra-modulation recognition method based on a denoising guided decoupling network, comprising: a time-frequency analysis method CTFD converts a radar signal into a time-frequency image; a global feature extractor extracts global features; a feature decoupler extracts a pure radar signal feature expression and a noise signal feature expression; and a modulation pattern recognizer performs a radar signal modulation format classification task. The present invention proposes a new denoising guided decoupling network DGDNet based on an Inception structure, which simultaneously completes the denoising and recognition of radar signals in an end-to-end manner. The feature map of the radar signal time-frequency map is obtained by a global feature extractor, and the feature decoupler separates the pure radar signal expression (PSR) from the noise signal expression (NSR) for learning a radar signal modulation pattern recognizer in a low signal-to-noise ratio environment. Finally, the modulation pattern recognizer completes the classification task. In order to expand the distance between the pure radar signal expression (PSR) and the noise signal expression (NSR), the concept of signal-noise mutual information (SNMI) loss is proposed. Experimental results show that the recognition accuracy of this method is 98.75% under -8dB signal-to-noise ratio, and the recognition accuracy of radar signals of 12 modulation formats under -10dB environment is 89.25%.

[0088] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0089] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.

[0090] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A radar signal pulse intra-modulation recognition method based on a denoising guided decoupling network, characterized in that: include: The radar signal is converted into a time-frequency image through the time-frequency analysis method CTFD; Using the time-frequency image as an input of a global feature extractor of a denoising guided decoupling network to obtain a global feature that incorporates radar signals and noise signals; Using the global feature as an input of a feature decoupler of a denoising guided decoupling network to obtain a pure radar signal feature expression and a noise signal feature expression, wherein the noise signal feature expression is used to train and optimize the feature decoupler; Using the pure radar signal feature expression as the input of the modulation mode recognizer of the denoising guided decoupling network, training the modulation mode recognizer to perform the radar signal modulation format recognition and classification task, and obtaining the modulation format recognition result; The step of converting the radar signal into a time-frequency image by using the time-frequency analysis method CTFD comprises: Based on the time-frequency analysis method CWD, the kernel function is improved to obtain the time-frequency analysis method CTFD. The expression of the improved kernel function is: Where τ is the horizontal axis delay, ν is the vertical axis frequency shift, γ is the shape coefficient, and ξ is the scale coefficient, which are used to adjust the shape and size of the kernel function respectively; Performing time-frequency analysis on radar signals of various modulation formats by using the time-frequency analysis method CTFD to obtain a time-frequency image; After the step of using the global feature as the input of the feature decoupler to obtain the pure radar signal feature expression and the noise signal feature expression, the method further includes: Constructing a signal-noise mutual information loss function according to the pure radar signal expression and the noise signal characteristic expression; Training and optimizing the feature decoupler by minimizing the signal-noise mutual information loss function; The formula of the signal-noise mutual information loss function is as follows: in and Respectively represent the characteristic expression of pure radar signal and the characteristic expression of noise signal, represents the joint probability distribution between two feature expressions, and They are the marginal probability distributions expressed by the two features, and are minimized To make the two feature expressions more independent.

2. The radar signal intra-pulse modulation recognition method based on denoising guided decoupling network as claimed in claim 1 is characterized in that: The global feature extractor includes: three filter cascade layers and convolution kernels of different sizes, and two small branches before each cascade layer are used to extract features of different scales.

3. The radar signal intra-pulse modulation recognition method based on denoising guided decoupling network as claimed in claim 1 is characterized in that: The feature decoupler includes: a pure radar feature extractor and a noise feature extractor.

4. The radar signal intra-pulse modulation recognition method based on denoising guided decoupling network as claimed in claim 3 is characterized in that: The pure radar feature extractor includes: 4 Inception_A modules, 1 Reduction_A module, 7 Inception_B modules and 1 deconvolution module.

5. The radar signal intra-pulse modulation recognition method based on denoising guided decoupling network as claimed in claim 3 is characterized in that: The noise feature extractor includes: 1 Inception_A module, 1 Reduction_A module, 2 Inception_B modules and 1 deconvolution module.

6. The radar signal intra-pulse modulation identification method based on denoising guided decoupling network as claimed in claim 1 is characterized in that: The modulation pattern recognizer includes: 1 Reduction_B module, 3 Inception_C modules and 1 FC layer, and the FC layer is connected to the SoftMax layer to convert the pure radar signal feature expression into a probability value.

7. The radar signal intra-pulse modulation recognition method based on denoising guided decoupling network as claimed in claim 1 is characterized in that: The step of training the modulation mode identifier to perform the radar signal modulation format recognition and classification task also includes: The prediction results are compared with the actual modulation pattern labels using cross entropy as the loss function; The modulation pattern recognizer is trained and optimized by minimizing the cross entropy loss function.

Images