Radar active interference suppression method, system, medium and equipment

Abstract

The invention discloses a radar active interference suppression method and system, a medium and equipment. The method comprises the following steps: constructing a training set and a test set; constructing a self-supervised training ViT network; training the self-supervised training ViT network by using the training set to obtain a trained self-supervised training ViT network; performing interference feature extraction on the time-frequency graph in the test set by using the trained self-supervised training ViT network to obtain a self-supervised interference feature graph; interference positioning processing is carried out on the self-supervised interference characteristic pattern to obtain a self-supervised interference characteristic pattern with an interference positioning frame; and performing interference suppression on the time-frequency graph according to the self-supervised interference characteristic graph with the interference positioning frame. According to the method, dependence on data manual marking during CNN training can be eliminated, the defect that the CNN only pays attention to local information is overcome, the generalization of the network and the accuracy of interference positioning are improved, and therefore the interference suppression effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal interference suppression, and in particular to a radar active interference suppression method, system, medium and equipment. Background Art

[0002] At present, there are many methods to counter radar intermittent sampling and forwarding interference, such as frequency agile waveform and mismatch filter design, interference reconstruction and elimination, bandpass filter design, etc. Most of these methods require parameter estimation of the interference signal, so the anti-interference performance also depends on the accuracy of the interference signal parameter estimation. In the case of composite interference, the parameter estimation accuracy of the interference signal is more difficult and less accurate than that of single interference, resulting in poor interference suppression effect.

[0003] In recent years, some CNN networks have been proposed for detecting and locating interference in radar signal time-frequency diagrams, and have achieved some impressive results. However, most CNN networks still have the following shortcomings:

[0004] First, a large amount of manually labeled radar data is needed for training, but radar data in reality is difficult to obtain, and labeling also requires a lot of manual work;

[0005] Second, the CNN network only considers local information, and has a certain bias when processing data, which has a certain impact on the final result;

[0006] Third, CNN networks are generally unable to process data types that have not been seen during training, have difficulty coping with scenarios where complex interference methods change rapidly, and have poor generalization. Summary of the invention

[0007] In order to solve the technical problems existing in the background technology, the present invention proposes a radar active interference suppression method, system, medium and device.

[0008] The present invention proposes a radar active interference suppression method based on self-supervised training ViT network, comprising:

[0009] Obtain radar active interference time-frequency atlas without interference type labels;

[0010] Extract radar active interference time-frequency graphs from the radar active interference time-frequency graph set without interference type labels, and combine them with the ImageNet dataset to form a training set;

[0011] The remaining radar active interference time-frequency graphs in the radar active interference time-frequency graph set without interference type labels are used to form a test set;

[0012] Construct a self-supervised training ViT network;

[0013] The self-supervised training ViT network is trained using the training set to obtain a trained self-supervised training ViT network;

[0014] The trained self-supervised ViT network is used to extract interference features from the time-frequency graph in the test set to obtain a self-supervised interference feature graph.

[0015] Perform interference location processing on the self-supervised interference feature map to obtain a self-supervised interference feature map with an interference location frame;

[0016] Interference suppression is performed on the time-frequency graph based on the self-supervised interference feature map with interference localization box.

[0017] Preferably, the self-supervised training ViT network includes a teacher network and a student network, and both the teacher network and the student network use the VIT-tiny network model as a basic model.

[0018] Preferably, the VIT-tiny network model comprises: an image segmentation module, a patch encoding module and a transformer encoder connected in sequence;

[0019] Among them, the transformer encoder includes 6 transformer layers connected in sequence; among them, each transformer layer includes a first normalization layer, a multi-head self-attention layer, a second normalization layer and a multi-layer perceptron, the input end of the first normalization layer is used as the input end of the transformer layer, the output end of the first normalization layer is connected to the input end of the multi-head self-attention layer, and the input end of the first normalization layer and the output end of the multi-head self-attention layer are added and connected to the input end of the second normalization layer, the output end of the second normalization layer is connected to the input end of the multi-layer perceptron, and the input end of the second normalization layer and the output end of the multi-layer perceptron are added and connected as the output end of the transformer layer.

[0020] Preferably, a focused linear attention module is used to replace the self-attention module in the softmax function of the VIT-tiny network.

[0021] Preferably, the interference feature extraction of the time-frequency graph in the test set is performed using the trained self-supervised training ViT network to obtain a self-supervised interference feature graph, specifically including:

[0022] The time-frequency graphs in the test set were subjected to two different random transformations;

[0023] The two different randomly transformed time-frequency graphs are input into the teacher network and the student network of the self-supervised training ViT network respectively to obtain two output features;

[0024] The two output features are normalized using the softmax function to obtain the self-supervised interference features.

[0025] Preferably, in the training process of the self-supervised training ViT network, the loss function is loss = -P2logP1, where P1 is the output of the teacher network, P2 is the output of the student network, and the stop gradient operator is applied to the teacher network, the gradient is only propagated through the student network, and the teacher parameters are updated using the exponential moving average of the student network parameters; the optimizer uses AdamW, and the learning rate uses cosine scheduling to decay the learning rate.

[0026] Preferably, interference suppression is performed on the time-frequency graph according to the self-supervised interference feature graph with the interference positioning frame, specifically including:

[0027] According to the self-supervised interference feature map with the interference positioning frame, the coordinates of the interference positioning frame are obtained;

[0028] According to the coordinates of the interference positioning frame, the self-supervised interference feature map with the interference positioning frame is segmented to obtain an interference segmentation result map;

[0029] According to the interference segmentation result map, an interference mask map corresponding to the interference in the time-frequency map is obtained;

[0030] According to the interference mask, the interference is filtered out from the time-frequency diagram.

[0031] The present invention also proposes a radar active interference suppression system based on self-supervised training ViT network, comprising:

[0032] An acquisition module, used to acquire a radar active interference time-frequency atlas without interference type labels;

[0033] The training set construction module is used to extract radar active interference time-frequency maps from the radar active interference time-frequency map set without interference type labels, and combine it with the ImageNet dataset to form a training set;

[0034] A test set building module, used for forming a test set from the remaining radar active interference time-frequency graphs in the radar active interference time-frequency graph set without interference type labels;

[0035] Model building module, used to build self-supervised training ViT network;

[0036] A training module is used to train the self-supervised training ViT network using the training set to obtain a trained self-supervised training ViT network;

[0037] The testing module is used to extract interference features from the time-frequency graph in the test set using the trained self-supervised training ViT network to obtain a self-supervised interference feature graph;

[0038] A positioning module is used to perform interference positioning processing on the self-supervised interference feature map to obtain a self-supervised interference feature map with an interference positioning frame;

[0039] The suppression module is used to suppress interference on the time-frequency graph according to the self-supervised interference feature map with the interference positioning box.

[0040] The present invention also proposes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the radar active interference suppression method based on the self-supervised training ViT network as described in any one of the above is implemented.

[0041] The present invention also proposes an electronic device, comprising: a processor and a memory, the memory being used to store one or more programs, and when the one or more programs are executed by the processor, a radar active interference suppression method based on a self-supervised training ViT network as described in any one of the above is implemented.

[0042] In the present invention, the proposed radar active interference suppression method, system, medium and device adopt a self-supervised ViT network to extract interference features, make full use of radar active interference time-frequency diagrams without interference type labels in a deep comparative learning manner, perform self-supervised comparative learning on a training set consisting of radar active interference time-frequency diagrams without interference type labels and an ImageNet data set, effectively extract self-supervised interference feature maps, eliminate the reliance on manual data labeling during CNN network training and avoid the drawback that CNN only focuses on local information, improve the generalization of the network and the accuracy of interference positioning, and thus improve the effect of interference suppression. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The figure is a flow chart of a radar active interference suppression method based on self-supervised training of a ViT network in one embodiment of the present invention.

[0044] Figure 2 It is a block diagram of a radar active interference suppression system based on self-supervised training ViT network in one embodiment of the present invention.

[0045] Figure 3 A schematic diagram of interference in an embodiment of the present invention.

[0046] Figure 4 This is a comparison chart of the results of feature extraction of radar active jammers using different networks.

[0047] Figure 5 This is a comparison chart of interference location using different networks to extract features from radar active interference.

[0048] Figure 6This is a comparison chart of the time-frequency results after using different networks to suppress radar active interference.

[0049] Figure 7 This is a graph showing the pulse compression results after using different networks to suppress radar active interference. DETAILED DESCRIPTION

[0050] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0051] Reference Figure 1 The present invention proposes a radar active interference suppression method based on self-supervised training ViT network, comprising:

[0052] Obtain radar active interference time-frequency atlas without interference type labels;

[0053] Extract radar active interference time-frequency graphs from the radar active interference time-frequency graph set without interference type labels, and combine them with the ImageNet dataset to form a training set;

[0054] The remaining radar active interference time-frequency graphs in the radar active interference time-frequency graph set without interference type labels are used to form a test set;

[0055] Construct a self-supervised training ViT network;

[0056] The self-supervised training ViT network is trained using the training set to obtain a trained self-supervised training ViT network;

[0057] The trained self-supervised ViT network is used to extract interference features from the time-frequency graph in the test set to obtain a self-supervised interference feature graph.

[0058] Perform interference location processing on the self-supervised interference feature map to obtain a self-supervised interference feature map with an interference location frame;

[0059] Interference suppression is performed on the time-frequency graph based on the self-supervised interference feature map with interference localization box.

[0060] Compared with the traditional CNN network, the present invention adopts a self-supervised ViT network to extract interference features, fully utilizes the radar active interference time-frequency diagram without interference type labels in a deep contrastive learning manner, performs self-supervised contrastive learning on the training set consisting of the radar active interference time-frequency diagram without interference type labels and the ImageNet data set, effectively extracts the self-supervised interference feature map, eliminates the dependence on manual data labeling during CNN network training and avoids the drawback that CNN only focuses on local information, improves the generalization of the network and the accuracy of interference positioning, thereby improving the effect of interference suppression.

[0061] In this embodiment, obtaining a radar active interference time-frequency atlas without an interference type label specifically includes:

[0062] Obtain radar active jammer dataset without jammer type labels;

[0063] The radar active interference data set without interference type labels is subjected to time-frequency processing to obtain a radar active interference time-frequency atlas without interference type labels.

[0064] Among them, 80% of the radar active jamming time-frequency atlas collection is mixed with the ImageNet dataset as a training set, and the remaining 20% ​​of the radar active jamming time-frequency atlas collection is used as a test set.

[0065] In this embodiment, obtaining a radar active interference data set without an interference type label specifically includes:

[0066] Use matlab to simulate the signal and get the radar echo signal;

[0067] Radar echo signals are organized into a radar active interference time-frequency atlas without interference type labels.

[0068] Among them, Matlab is used to perform signal simulation to obtain radar echo signals, including:

[0069] Set the radar signal waveform; the radar signal waveform is a linear frequency modulation (LFM) signal, and its time domain expression is as follows:

[0070]

[0071] Where A is the signal amplitude, f0 is the frequency, T and B are the time width and bandwidth respectively; k is the frequency modulation slope, k = T / B; is a rectangular function,

[0072] In one specific embodiment, f0 is set to 3×10 7 Hz, T = 20 × 10 6 s, B = 10 6 Hz. At the same time, considering the long range of radar, the target is set as a point target in the simulation.

[0073] In this embodiment, time-frequency processing is performed on the radar active interference time-frequency atlas without the interference type label to obtain the radar active interference time-frequency atlas without the interference type label, which specifically includes:

[0074] Short-time Fourier transform (STFT) is used to perform time-frequency processing on the radar active interference time-frequency atlas without interference type labels to obtain the radar active interference time-frequency atlas without interference type labels.

[0075] Among them, short-time Fourier transform (STFT) is a time-frequency analysis method, which uses a window function to segment the signal and performs Fourier transform to convert the one-dimensional time domain data into two-dimensional data in the time-frequency domain, and uses a drawing function to convert the two-dimensional data in the time-frequency domain into interference time-frequency diagram data.

[0076] In this embodiment, the window function uses a Hamming window, the sampling point is 1024, and the expression of the short-time Fourier transform is:

[0077]

[0078] Where τ is φ * The center position of (τ-t), * indicates the conjugate operation.

[0079] The interference types include single interference signals of different categories and composite slice interference signals of different slice combinations.

[0080] like Figure 3 As shown, in one specific embodiment, the interference signal types include five single interferences and four composite interferences. Among them, the five single interference signals are Interrupted Sampling and Repeater Jamming With Overlay Forwarding (ISRJOF), Interrupted-sampling Frequency Shift Repeater jamming (ISFSRJ), Mono-Pulse Dense Forwarding Jamming (MDFJ), Non-uniform Interrupted Sampling Repeater Jamming (NUISRJ) and Modulated Interrupt Sampling Repeater Jamming (MISRJ).

[0081] It is important to know that ISRJOF Figure 3 As shown in (a), ISFSRJ Figure 3 As shown in (b), MDFJ Figure 3 As shown in (c), NUISRJ Figure 3 As shown in (d), MISRJ Figure 3(e) as shown.

[0082] Among them, the four composite interferences in this embodiment are ISFSRJ and NUISRJ composite interference, MDFJ and NUISRJ composite interference, ISFSRJ and MISRJ composite interference, and ISRJOF and NUISRJ composite interference, respectively. Figure 3 (f) Figure 3 (g) Figure 3 (h) and Figure 3 (i) as shown.

[0083] The working principle of ISRJOF is to repeat the forwarding multiple times after intercepting a radar signal. As the number of superposition forwarding increases, the next PRT may have an interference signal. The signal time domain model is expressed as follows:

[0084]

[0085] S * =S(t-(mτ+(m-1)T s )+(M+1)(n-1)T s -L Jam );

[0086] L Jam =N(M+1)T w ;

[0087] Where N is the number of slices, T s is the slice width, M is the number of forwarding times, τ r is the jammer sampling time, and S(t) is the radar transmitting signal.

[0088] ISFSRJ uses a segmented frequency shift method for sampling signals whose frequency components are located at the edge of the bandwidth within the sampling period. Its signal time domain model can be expressed as:

[0089]

[0090] In the formula, f j represents frequency shift, which will produce lead or lag interference. j >0, leading interference is formed. j When <0, delayed interference is formed.

[0091] MDFJ samples only a portion of the radar signal and forwards it multiple times to form a dense group of false targets. The signal time domain model is as follows:

[0092]

[0093] NUISRJ samples the received signal at non-uniform time intervals and then forwards these sampled values ​​to other locations. The signal time domain model is as follows:

[0094]

[0095] Where, T n =(M+1)T s represents the period of each sampling pulse, Indicates the time when the nth sampling pulse starts, τ n Indicates delay.

[0096] MISRJ first converts the receiving antenna into a received radar signal, and then obtains a digital signal through bandpass filtering, down-conversion and analog-to-digital conversion. Then, the sampled part of the transmitted signal is modulated by a pre-set filter to obtain a modulated interference signal. The signal time domain model is as follows:

[0097] S s =p(t)*S(t);

[0098]

[0099] S MISRJ (t) = IFT(H(j)*S s (f));

[0100] Where p(t) represents the sampling function, T c represents the interference patch width, τ c Indicates the delay, S s (f) and IFT(·) represent S s (t) Fourier transform and inverse Fourier transform operations.

[0101] In this embodiment, after forming the training set and the test set, data enhancement is performed on the time-frequency graphs in the training set and the test set, respectively, wherein the data enhancement includes color jittering, Gaussian blurring, and multiple cropping.

[0102] In this embodiment, the self-supervised training ViT network includes a teacher network and a student network, and both the teacher network and the student network use the VIT-tiny network model as a basic model.

[0103] In a further embodiment, the VIT-tiny network model includes: an image partitioning module, a patch encoding module, and a transformer encoder connected in sequence;

[0104] Among them, the transformer encoder includes 6 transformer layers connected in sequence; among them, each transformer layer includes a first normalization layer, a multi-head self-attention layer, a second normalization layer and a multi-layer perceptron, the input end of the first normalization layer is used as the input end of the transformer layer, the output end of the first normalization layer is connected to the input end of the multi-head self-attention layer, and the input end of the first normalization layer and the output end of the multi-head self-attention layer are added and connected to the input end of the second normalization layer, the output end of the second normalization layer is connected to the input end of the multi-layer perceptron, and the input end of the second normalization layer and the output end of the multi-layer perceptron are added and connected as the output end of the transformer layer.

[0105] In this embodiment, the self-supervised training ViT network with such a structure can improve the accuracy of the interference features in the extracted self-supervised interference feature map, which is beneficial to the subsequent detection, positioning and segmentation of interference, and thus beneficial to the subsequent interference suppression effect.

[0106] Specifically, the input of each transformer layer first passes through the first normalization layer, then passes through the multi-head self-attention layer to obtain the multi-head self-attention relative to itself, and uses the residual structure to fuse the input of the first normalization layer and the output of the multi-head self-attention layer and pass through the second normalization layer, and then the obtained data is sent to the multi-layer perceptron (MLP). Among them, the multi-layer perceptron contains two linear layers, a fully connected layer, and two random dropout layers to prevent overfitting in training.

[0107] The image segmentation module is used to segment the input image into image blocks of a patch size to better extract detail features. In this embodiment, the size of the input image is 240×240 (pixels), the patch size is a square with a side length of 6 pixels, and the number of patches is 1600.

[0108] Among them, the patch encoding module is used to encode the image block into a Patch Embedding, which is integrated with the learnable classification token and position encoding and then input into the transformer encoder; the transformer encoder is used to obtain a feature vector representing the entire image based on the integrated learnable classification token and position encoding, including the global meaning and information in the context of the entire image.

[0109] In a further embodiment, a focused linear attention module is used to replace the self-attention module in the softmax function of the VIT-tiny network to change the attention calculation method, improve the efficiency of the self-attention calculation, and thus optimize the efficiency of the VIT-tiny network model.

[0110] In a further embodiment, the interference feature extraction of the time-frequency graph in the test set is performed using the trained self-supervised training ViT network to obtain a self-supervised interference feature graph, specifically including:

[0111] The time-frequency graphs in the test set were subjected to two different random transformations;

[0112] The two different randomly transformed time-frequency graphs are input into the teacher network and the student network of the self-supervised training ViT network respectively to obtain two output features;

[0113] The two output features are normalized using the softmax function to obtain the self-supervised interference features.

[0114] In this embodiment, during the training process of the self-supervised training ViT network, the loss function is loss = -P2logP1, where P1 is the output of the teacher network, P2 is the output of the student network, and the stop gradient operator is applied to the teacher network, and the gradient is only propagated through the student network. The teacher parameters are updated using the exponential moving average (ema) of the student network parameters; the optimizer uses AdamW, and the learning rate uses cosine scheduling to decay the learning rate so that the loss value of the self-supervised training ViT network is minimized.

[0115] This embodiment fixes the teacher network parameters, and iterates the teacher network by the gradient update of the student network, which simplifies the gradient update and makes the network training easier; moreover, self-supervised training is adopted in the training process of the self-supervised training ViT network, and no labeled data is required, which saves a lot of manpower and time.

[0116] In order to detect and locate interference, in one embodiment, interference location processing is performed on the self-supervised interference feature map to obtain a self-supervised interference feature map with an interference location frame, specifically including:

[0117] According to the pixel value of each pixel in the self-supervised interference feature map and a preset pixel threshold, the pixel points above the pixel threshold are extracted as interference positioning frames, thereby obtaining a self-supervised interference feature map with an interference positioning frame.

[0118] In this embodiment, interference suppression is performed on the time-frequency graph according to the self-supervised interference feature graph with the interference positioning frame, specifically including:

[0119] According to the self-supervised interference feature map with the interference positioning frame, the coordinates of the interference positioning frame are obtained;

[0120] According to the coordinates of the interference positioning frame, the self-supervised interference feature map with the interference positioning frame is segmented to obtain an interference segmentation result map;

[0121] According to the interference segmentation result map, an interference mask map corresponding to the interference in the time-frequency map is obtained;

[0122] According to the interference mask, the interference is filtered out from the time-frequency diagram.

[0123] This embodiment adopts the image segmentation algorithm to improve the efficiency of interference suppression.

[0124] Specifically, based on the self-supervised interference feature map with the interference positioning box, an image segmentation algorithm is applied, the image segmentation algorithm will extract the foreground target, add mask processing in the subsequent processing, add a mask to the generated image and use the mask to suppress interference.

[0125] It is important to know that in the image segmentation algorithm, a Gaussian mixture model (GMM) is used to model the background and foreground respectively, and undefined pixels are marked as possible foreground or possible background; each pixel in the image is connected to the surrounding pixels through virtual edges, and according to the similarity between the pixel and the connected pixels, the probability of all its virtual edges belonging to the foreground or background is assigned; by counting the probability of the virtual edges of the pixel points, it is determined whether the pixel points belong to the foreground, and each pixel will belong to the foreground or background. After all the pixels are connected (may be connected to the background or foreground), if there is a virtual edge belonging to two areas at the same time (that is, two pixels connected by a virtual edge, one of which belongs to the foreground and the other belongs to the background), the virtual edge between them will be cut off. After similar operations are performed on all pixels, the foreground and background of the image can be segmented.

[0126] The present invention also proposes a radar active interference suppression system based on self-supervised training ViT network, comprising:

[0127] An acquisition module, used to acquire a radar active interference time-frequency atlas without interference type labels;

[0128] The training set construction module is used to extract radar active interference time-frequency maps from the radar active interference time-frequency map set without interference type labels, and combine it with the ImageNet dataset to form a training set;

[0129] A test set building module, used for forming a test set from the remaining radar active interference time-frequency graphs in the radar active interference time-frequency graph set without interference type labels;

[0130] Model building module, used to build self-supervised training ViT network;

[0131] A training module is used to train the self-supervised training ViT network using the training set to obtain a trained self-supervised training ViT network;

[0132] The testing module is used to extract interference features from the time-frequency graph in the test set using the trained self-supervised training ViT network to obtain a self-supervised interference feature graph;

[0133] A positioning module is used to perform interference positioning processing on the self-supervised interference feature map to obtain a self-supervised interference feature map with an interference positioning frame;

[0134] The suppression module is used to suppress interference on the time-frequency graph according to the self-supervised interference feature map with the interference positioning box.

[0135] It should be noted that the function of each module in this embodiment is the same as the corresponding description in the above method embodiment, and will not be repeated here.

[0136] The present invention also proposes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the radar active interference suppression method based on the self-supervised training ViT network as described in any one of the above is implemented.

[0137] The present invention also proposes an electronic device, comprising: a processor and a memory, the memory being used to store one or more programs; when the one or more programs are executed by the processor, the radar active interference suppression method based on the self-supervised training ViT network as described in any one of the above is implemented.

[0138] The effects of the present invention can be further illustrated by the following examples and comparative examples.

[0139] Example 1

[0140] The interference types of the time-frequency diagram in the test set in this embodiment include five single interferences, namely ISRJOF, ISFSRJ, MDFJ, NUISRJ and MISRJ, and four compound interference types, namely ISFSRJ and NUISRJ, MDFJ and NUISRJ, ISFSRJ and MISRJ, and ISRJOF and NUISRJ.

[0141] This embodiment uses the radar active interference suppression method based on self-supervised training ViT network proposed in the present invention to extract features and locate the selected interference. The feature extraction result is as follows: Figure 4 As shown in (d), the positioning results are Figure 5 (d) as shown.

[0142] Comparative Example 1

[0143] The test set in this embodiment is the same as the test set in Embodiment 1.

[0144] This embodiment uses the LOST network method to extract and locate interference features. The feature extraction results are as follows: Figure 4 As shown in (a), the positioning results are Figure 5 (a) shown.

[0145] Comparative Example 2

[0146] The test set in this embodiment is the same as the test set in Embodiment 1.

[0147] This embodiment uses the DSS network method to extract and locate interference features. The feature extraction results are as follows: Figure 4 As shown in (b), the positioning results are Figure 5 (b) as shown.

[0148] Comparative Example 3

[0149] The test set in this embodiment is the same as the test set in Embodiment 1.

[0150] This embodiment uses the TokenCut network method to extract and locate interference features. The feature extraction results are as follows: Figure 4 As shown in (c), the positioning results are Figure 5 (c) as shown.

[0151] Example 2

[0152] The time-frequency diagram with interference selected in this embodiment is as follows: Figure 6 As shown in (a), the pulse compression result of the time-frequency diagram with interference is as follows Figure 7 (a) shown.

[0153] This embodiment uses the radar active interference suppression method based on self-supervised training ViT network proposed by the present invention to filter out the interference of the selected time-frequency graph with interference. The time-frequency graph after interference filtering is as follows: Figure 6 As shown in (f), the pulse compression result after interference filtering is as follows Figure 7 (f) as shown.

[0154] Comparative Example 4

[0155] The time-frequency diagram with interference in this embodiment is the same as the time-frequency diagram with interference in Embodiment 2.

[0156] This embodiment uses a subspace projection (SSL) method to reconstruct and cancel the time-frequency diagram with interference to filter out the interference. The time-frequency diagram after the interference is filtered out is as follows: Figure 6 (b) shows that the pulse compression result after interference filtering is as follows Figure 7 (b) as shown.

[0157] Comparative Example 5

[0158] The time-frequency diagram with interference in this embodiment is the same as the time-frequency diagram with interference in Embodiment 2.

[0159] This embodiment uses a Bayesian compressed sensing (BCS) method to filter out the time-frequency graph with interference. The time-frequency graph after the interference is filtered out is as follows: Figure 6 (c) shows that the pulse compression result after interference filtering is as follows Figure 7 (c) as shown.

[0160] Comparative Example 6

[0161] The time-frequency diagram with interference in this embodiment is the same as the time-frequency diagram with interference in Embodiment 2.

[0162] The adaptive filter method of this embodiment filters the time-frequency graph with interference to suppress the interference. The time-frequency graph after the interference is filtered out is as follows: Figure 6 As shown in (d), the pulse compression result after interference filtering is as follows Figure 7 (d) as shown.

[0163] Comparative Example 7

[0164] The time-frequency diagram with interference in this embodiment is the same as the time-frequency diagram with interference in Embodiment 2.

[0165] This embodiment uses the TokenCut network method to extract and locate the time-frequency graph with interference, and uses the extracted interference features to form a mask to suppress interference. The filtering effect is as follows: Figure 6 As shown in (e), the pulse compression result after interference filtering is Figure 7 (e) as shown.

[0166] from Figure 4 and Figure 5 It can be seen that the radar active interference suppression method based on self-supervised training ViT network proposed in the present invention shows good results in both interference feature extraction and interference positioning.

[0167] Depend on Figure 6 and Figure 7 It can be seen that the radar active interference suppression method based on the self-supervised training ViT network proposed in the present invention not only suppresses interference well, but also has the best protection for the target signal.

[0168] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A radar active interference suppression method based on self-supervised training ViT network, characterized in that: include: Obtain radar active interference time-frequency atlas without interference type labels; Extract radar active interference time-frequency graphs from the radar active interference time-frequency graph collection without interference type labels, and combine them with the natural dataset ImageNet to form a training set; The remaining radar active interference time-frequency graphs in the radar active interference time-frequency graph set without interference type labels are used to form a test set; Construct a self-supervised training ViT network; The self-supervised training ViT network is trained using the training set to obtain a trained self-supervised training ViT network; The trained self-supervised ViT network is used to extract interference features from the time-frequency graph in the test set to obtain a self-supervised interference feature graph. Perform interference location processing on the self-supervised interference feature map to obtain a self-supervised interference feature map with an interference location frame; Interference suppression is performed on the time-frequency graph based on the self-supervised interference feature map with interference localization box.

2. The radar active interference suppression method based on self-supervised training ViT network according to claim 1 is characterized in that: The self-supervised training ViT network includes a teacher network and a student network. Both the teacher network and the student network use the VIT-tiny network model as the basic model.

3. The radar active interference suppression method based on self-supervised training ViT network according to claim 2 is characterized in that: The VIT-tiny network model includes: an image segmentation module, a patch encoding module, and a transformer encoder connected in sequence; Among them, the transformer encoder includes 6 transformer layers connected in sequence; among them, each transformer layer includes a first normalization layer, a multi-head self-attention layer, a second normalization layer and a multi-layer perceptron, the input end of the first normalization layer is used as the input end of the transformer layer, the output end of the first normalization layer is connected to the input end of the multi-head self-attention layer, and the input end of the first normalization layer and the output end of the multi-head self-attention layer are added and connected to the input end of the second normalization layer, the output end of the second normalization layer is connected to the input end of the multi-layer perceptron, and the input end of the second normalization layer and the output end of the multi-layer perceptron are added and connected as the output end of the transformer layer.

4. The radar active interference suppression method based on self-supervised training ViT network according to claim 2 is characterized in that: The self-attention module in the softmax function of the VIT-tiny network is replaced by a focused linear attention module.

5. The radar active interference suppression method based on self-supervised training ViT network according to claim 1 is characterized in that: The trained self-supervised training ViT network is used to extract interference features from the time-frequency graph in the test set to obtain a self-supervised interference feature graph, which specifically includes: The time-frequency graphs in the test set were subjected to two different random transformations; The two different randomly transformed time-frequency graphs are input into the teacher network and the student network of the self-supervised training ViT network respectively to obtain two output features; The two output features are normalized using the softmax function to obtain the self-supervised interference features.

6. The radar active interference suppression method based on self-supervised training ViT network according to claim 1 is characterized in that: In the training process of the self-supervised training ViT network, the loss function is loss = -P2logP1, where P1 is the output of the teacher network, P2 is the output of the student network, and the stop gradient operator is applied to the teacher network, and the gradient is only propagated through the student network. The teacher parameters are updated using the exponential moving average of the student network parameters; the optimizer uses AdamW, and the learning rate uses cosine scheduling to decay the learning rate.

7. The radar active interference suppression method based on self-supervised training ViT network according to claim 1 is characterized in that: According to the self-supervised interference feature map with the interference positioning frame, interference suppression is performed on the time-frequency map, specifically including: According to the self-supervised interference feature map with the interference positioning frame, the coordinates of the interference positioning frame are obtained; According to the coordinates of the interference positioning frame, the self-supervised interference feature map with the interference positioning frame is segmented to obtain an interference segmentation result map; According to the interference segmentation result map, an interference mask map corresponding to the interference in the time-frequency map is obtained; According to the interference mask, the interference is filtered out from the time-frequency diagram.

8. A radar active interference suppression system based on self-supervised training ViT network, characterized in that: include: An acquisition module, used to acquire radar active interference time-frequency atlas without interference type labels and supervised time-frequency atlas with interference type labels; The training set construction module is used to extract radar active interference time-frequency maps from the radar active interference time-frequency map set without interference type labels, and combine it with the ImageNet dataset to form a training set; A test set building module, used for forming a test set from the remaining radar active interference time-frequency graphs in the radar active interference time-frequency graph set without interference type labels; Model building module, used to build self-supervised training ViT network; A training module is used to train the self-supervised training ViT network using the training set to obtain a trained self-supervised training ViT network; The testing module is used to extract interference features from the time-frequency graph in the test set using the trained self-supervised training ViT network to obtain a self-supervised interference feature graph; A positioning module is used to perform interference positioning processing on the self-supervised interference feature map to obtain a self-supervised interference feature map with an interference positioning frame; The suppression module is used to suppress interference on the time-frequency graph according to the self-supervised interference feature map with the interference positioning box.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the radar active interference suppression method based on the self-supervised training ViT network as described in any one of claims 1 to 7 is implemented.

10. An electronic device comprising: A processor and a memory, the memory being used to store one or more programs, characterized in that when the one or more programs are executed by the processor, the radar active interference suppression method based on the self-supervised training ViT network as described in any one of claims 1 to 7 is implemented.

Images