Radar active composite jamming signal recognition and parameter measurement method based on Yolov8 network with convolutional attention mechanism and rejection loss

Abstract

A radar active composite interference signal identification and parameter measurement method based on the Yolov8 network with a convolutional attention mechanism and repulsive loss. To address the problems of the single working mode and poor robustness of existing interference signal identification methods, the present invention converts the interference identification problem into a target detection problem, and applies the target detection framework to detect, identify, and measure the parameters of the interference signal. The CBAM attention mechanism is introduced to enhance the network's image feature extraction capabilities and improve the recognition rate. In addition, a repulsive loss function is introduced to maintain good regression prediction results even when the interference signal's time-frequency images overlap significantly. The network in this paper has good detection and recognition performance for most signals at low interference-to-noise ratios. The present invention belongs to the field of radar interference signal detection and recognition technology.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar jamming signal detection and recognition, and specifically relates to a radar active composite jamming signal recognition and parameter measurement method based on a Yolov8 network with a convolutional block attention module (CBAM) and repulsion loss (REP). Background Art

[0002] With the increasing number of diverse radiation sources on modern battlefields, the electromagnetic environment has become increasingly complex, characterized by high signal density and a high degree of pulse sequence interleaving, resulting in intense competition and confrontation across the spectrum. Radars can receive multiple interference signals, which alias in the time or frequency domain to form composite interference signals. Since the amount of interference in these mixed signals is unknown, this can significantly impact radar detection. To implement more effective and targeted anti-interference measures, accurate and rapid signal detection, type identification, and parameter measurement of potential interference signals in radar echoes are necessary to ensure radar performance.

[0003] By searching the current relevant literature on compound interference identification, it was found that Qu, Qizhe proposed an interference classification network JRNet in "JRNet: Jamming recognition networks for radar compound suppression jamming signals" published in "IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY" (2020, 69(12): 15035-15045). It uses power spectrum features as network input and can identify 4 single interference types and 6 compound interference types. Although this method has good recognition performance under low interference-to-noise ratio, it cannot perform multi-task processing of signal detection, type recognition, and parameter measurement on interference signals at the same time. JiaXiang Zhang published "Radar Compound Jamming Cognition Based on a Deep Object Detection Network" in "IEEE TRANSACTIONS ONAEROSPACE AND ELECTRONIC SYSTEMS" (2023, 59(3): 3251-3263). He applied a target detection network to identify the interference type based on the time-frequency image and used the prediction box to locate the position of the interference signal time-frequency map on the time-frequency image. He also gave the corresponding parameter estimation method. However, he did not conduct experiments for low interference-to-noise ratio (INR) below 0 dB. Summary of the Invention

[0004] In order to solve the problem of missed detection caused by overlapping time-frequency images with low drying ratio when targeting radar jamming signals, the present invention proposes a radar active composite jamming signal recognition and parameter measurement method based on the Yolov8 network with convolutional attention mechanism and rejection loss.

[0005] The technical solution adopted by the present invention to solve the above problems is:

[0006] The present invention comprises the steps of:

[0007] Step 1: Establish a mathematical model of radar composite interference signal;

[0008] The input signal model is:

[0009]

[0010] Where k represents the number of components that constitute a single interference signal, which is composed of at most q interference components, and J i (n) represents the received interference signal, A i is the amplitude coefficient of the modulation when propagating through the environment, awgn(n) represents Gaussian white noise, x t (n) represents the real target echo signal.

[0011] Step 2: pre-processing the composite interference signal;

[0012] First, perform time-frequency transformation on the composite interference signal to obtain its time-frequency image X(t,ω), which is expressed as follows:

[0013]

[0014] Where ω represents the frequency and h(τ-t) is the analysis window function.

[0015] Step 3: Input the training set into the Yolov8 network with convolutional attention mechanism and rejection loss for training;

[0016] The Yolov8 network with the convolutional attention mechanism and exclusion loss includes a context-aware cross-level fusion network module, a pyramid pooling module, a convolutional attention module, a splicing module, and a detection head module, among which the classification branch uses a cross-entropy loss function, and the regression branch uses an exclusion loss function and an intersection-over-union loss function.

[0017] Step 4: Input the training set video images into the backbone part of the network for feature extraction, which includes:

[0018] Step 4.1: The input image is operated through the convolution module to extract features;

[0019] Step 4.2: Input the feature map obtained by the convolutional layer into the context-aware cross-level fusion network module, which includes a 1x1 convolutional layer and n BottleNeck layers.

[0020] Step 4.3: Input the features obtained above into the pyramid pooling module. Pyramid pooling is improved from SPP in Yolov5 and consists of three maximum pooling layers, two convolutional layers and a feature fusion layer. It performs pooling operations of different sizes on the feature map and integrates the results to obtain a fixed-size output.

[0021] Step 5: Input the feature map into the Neck part of the network, perform upsampling, splicing and other operations to perform feature fusion, and introduce the convolutional attention module to enhance the feature extraction capability and improve the recognition rate.

[0022] Step 5.1: Upsample the feature map so that the output size is twice the input size for easy subsequent processing.

[0023] Step 5.2: Fuse the upsampled feature map with the feature map obtained by the context-aware cross-level fusion network module, and concatenate the 80*80*256 and 80*80*512 feature maps into a 80*80*768 feature map. The same applies to the subsequent concatenation modules.

[0024] Step 5.3: After the splicing module completes the feature map, it passes through the context-aware cross-level fusion network module and is input into the convolutional attention module. The convolutional attention module consists of a channel attention module and a spatial attention module.

[0025] Step 6: The fused features from the Neck section are input into the detection head module. The features are further extracted through the convolution module and the two-dimensional convolution module, and their dimensions are changed to facilitate recognition and regression prediction operations. This section introduces a rejection loss function to maintain good regression prediction results when there is a lot of overlap in the time-frequency images of the interference signal. The cross-entropy loss function, rejection loss, and intersection-over-union loss function are used to continuously correct the prediction results for the category and prediction box. Parameters such as pulse width, bandwidth, and center frequency are calculated based on the coordinates of the prediction box.

[0026] Step 7: Input the test set into the trained Yolov8 network with convolutional attention mechanism and rejection loss to obtain the interference signal components and interference signal parameters contained in the input signal.

[0027] The beneficial effects of the present invention are:

[0028] 1. Compared with the existing technology, this invention addresses the problems of single working mode and poor robustness of existing methods and designs a more intelligent composite interference signal detection, identification and parameter measurement method. It transforms the interference identification problem into a target detection problem and applies the target detection framework to detect, identify and measure the parameters of the interference signal.

[0029] 2. This paper conducts combined modeling analysis on multiple interference signals and designs a Yolov8 network with a convolutional attention mechanism and a rejection loss from the perspective of signal time-frequency images. The convolutional attention mechanism is introduced to enhance the network's ability to extract image features and improve the recognition rate. In addition, a rejection loss function is introduced to maintain good regression prediction results even when the interference signal time-frequency images overlap a lot.

[0030] 3. This invention uses multiple modules such as context-aware cross-level fusion network, convolutional attention, and pyramid pooling to extract features from the input time-frequency graph, solving the problem of a single working mode in traditional methods.

[0031] 4. The present invention can simultaneously complete the tasks of detection and identification and parameter measurement, while maintaining good detection and identification performance under low interference-to-noise ratio. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is the overall flow chart of the present invention;

[0033] Figure 2 Schematic diagram of the structure of Yolov8 with convolutional attention mechanism and rejection loss of the present invention;

[0034] In the figure, the time-frequency graph generated by the time-frequency transformation of the signal is first input into the Backbone module of the network for feature extraction, then the features are input into the Neck module for feature fusion, and finally the processed features are input into the detection head module for classification and regression.

[0035] Figure 3 is the performance change curve of the network with the number of iterations in the present invention;

[0036] A composite interference signal dataset was constructed by combining six signal types: slice reconstruction interference, white noise convolution interference, white noise product interference, spectrum dispersion interference, intermittent sampling forwarding interference, and echo signals. Time-frequency images were generated for each combination at 11 interference-to-noise ratios (INRs) of -10dB, -8dB, -6dB, -4dB, -2dB, 0dB, 2dB, 4dB, 6dB, 8dB, and 10dB, totaling 10,400 time-frequency images. The training results show that the proposed method achieves high recognition performance, with an mAP50:95 ratio of 0.94.

[0037] Figure 4 Identify parameter measurement results for detection of NCN, CI, and SMSP composite interference;

[0038] The labels on the upper side of the detection box in the figure are the signal category, confidence, pulse width, bandwidth and center frequency respectively.

[0039] Figure 5 This is the curve of mAP50:95 changing with the interference-to-noise ratio (JNR).

[0040] Depend on Figure 5 It can be seen that the model maintains good performance under various interference-to-noise ratios, with mAP50:95 reaching 0.943 at 10dB and 0.85 at -10dB. DETAILED DESCRIPTION

[0041] like Figure 1 The figure shows the overall flow chart of the present invention. First, the composite interference signal is transformed into a time-frequency image and a data set is generated. Then, the data set is divided into a training set and a test set. The training set is used to train the network of this paper, and the test set is used to verify the performance of the network. Specifically, the following steps are included:

[0042] Step 1: Establish a mathematical model of radar composite interference signal.

[0043] The simulation generates radar composite jamming signals, which include radar echo signals, suppression jamming, smart noise jamming, deception jamming, etc. The input signal model is:

[0044]

[0045] Where k represents the number of components constituting a single interference signal, J i (n) represents the received interference signal, A i is the amplitude coefficient of the modulation when propagating through the environment, awgn(n) represents Gaussian white noise, x t (n) represents the real target echo signal.

[0046] Step 2: Preprocess the composite interference signal.

[0047] The composite interference signal is transformed into a time-frequency image to obtain the signal's time-frequency joint distribution image. The interference signal is transformed into a two-dimensional time-frequency image X(t,ω) by using short-time Fourier transform.

[0048]

[0049] Where ω represents the frequency and h(τ-t) is the analysis window function.

[0050] The time-frequency image is preprocessed and the image is scaled normalized to a 640*640 image.

[0051] Step 3: Input the training set into the Yolov8 network with convolutional attention mechanism and rejection loss for training.

[0052] Step 4: Input the image into the backbone part of the network for feature extraction.

[0053] Step 4.1: The input image is operated through the convolution module to extract features. The convolution module consists of a convolution layer, a normalization layer, and then activated with the SiLU function. Compared with the ReLU function, the SiLU function has a smoother curve when it is close to zero, and because it uses the sigmoid function, the output range of the network can be between 0 and 1. The expression of the SiLU function is:

[0054]

[0055] Step 4.2: Input the feature map generated by the convolutional layer into the context-aware cross-level fusion network module, which consists of a 1x1 convolutional layer and n BottleNeck layers. First, the input feature map is transformed by the first convolutional layer, and the output feature map is divided into two parts. These two parts are processed by different convolutional layers before being merged. This process helps the model capture more contextual information, leading to more accurate object recognition.

[0056] The context-aware cross-level fusion network module improves the model's performance and accuracy. By introducing this module, the model can better capture complex features in images, achieving better results in object detection tasks. Furthermore, this module offers excellent scalability, further enhancing the model's feature extraction capabilities without significantly increasing computational costs.

[0057] Step 4.3: Input the features obtained above into the pyramid pooling module. Pyramid pooling is an improvement on the SPP in Yolov5. It consists of three max pooling layers, two convolutional layers, and a feature fusion layer. It performs pooling operations of varying sizes on the feature map and combines the results to produce a fixed-size output. This technique can effectively handle objects of varying sizes, thereby improving the generalization and robustness of the neural network.

[0058] Step 5: Input the feature map into the Neck part of the network for feature fusion.

[0059] Step 5.1: Upsample the feature map so that the output size is twice the input size for easy subsequent processing.

[0060] Step 5.2: Fuse the upsampled feature map with the feature map obtained by the context-aware cross-level fusion network module, and concatenate the feature maps of 80*80*256 and 80*80*512 into a feature map of 80*80*768. The same applies to the subsequent concatenation modules.

[0061] Step 5.3: The feature map after splicing by the splicing module passes through the context-aware cross-level fusion network module and is input into the convolutional attention module. The convolutional attention consists of the channel attention module and the spatial attention module.

[0062] In the channel attention module, two different pooling operations are performed on the input features of dimension H×W×C to obtain two channel information, which are then input into the multi-layer perceptron for dimensionality reduction to obtain two feature maps of dimension 1×1×C. After element-by-element summation and Sigmoid activation function processing, the channel attention features are obtained.

[0063] The feature map obtained by element-wise multiplication of the channel attention feature and the input feature map is input into the spatial attention module. The two pooling operations mentioned above are also performed to obtain two feature maps with dimensions of H×W×1. The two feature maps are connected and a 7×7 convolution operation is performed. The spatial attention feature is then obtained by processing them with the Sigmoid activation function.

[0064] Finally, the spatial attention features are multiplied by the input features of the spatial attention module to obtain the final generated features. The introduction of convolutional attention strengthens the network's feature extraction capabilities, allowing the network to focus more on key information, ultimately improving signal recognition rates.

[0065] Step 6: The results of the last three context-aware cross-level fusion network modules are input into the detection head module for classification and regression.

[0066] The detection head module consists of two layers: one for classification and the other for regression prediction. Each layer consists of two convolutional modules and one two-dimensional convolutional module. The cross-entropy function is used as the loss function for the classifier, while the exclusion loss function and the intersection-over-union loss function are used for the regression part.

[0067] The expression of the rejection loss function is:

[0068] L=L Attr +α*L RepGT +β*L RepBox

[0069] Where, LAttr It is the attraction between the predicted box and its true target box, expressed as follows

[0070]

[0071] in,

[0072]

[0073] Where, P + is the set of detection boxes P divided according to the IoU threshold, Smooth L1 is the SmoothL1 distance. B P Represents the given P∈P + The predicted box of regression, represents the ground-truth bounding box with the maximum intersection-over-union (IoU) with P.

[0074]

[0075] Where, L RepGT It's B P and The loss of the overlapping part between the prediction box P and the surrounding target box G is made as far away as possible. RepBox It is the loss function for the overlapping parts of the boxes of different targets i, j, so that different prediction boxes P i and P j Stay as far away as possible.

[0076] The IoU metric is a common metric in target detection. The intersection-over-union loss function is defined as the negative logarithm of the ratio of the intersection and union between the predicted box and the true box, which can measure the difference between the predicted box and the true box. Assume that the coordinates of the true box are The coordinates of the prediction box are x=(x t ,x b ,x l ,x r ).

[0077]

[0078] Step 7: Obtain the categories of the input composite interference signal and its position on the time-frequency image, and calculate the signal parameters based on the prediction box.

[0079] The mAP50:95 metric is used to evaluate the performance of the model, where mAP (Mean Average Precision) is the mean average precision. Based on the precision-recall curve, the average precision value corresponding to each recall value is calculated to obtain the average precision (AP). The expression of AP is as follows:

[0080]

[0081] where r i is the recall value corresponding to the i-th precision interpolation segment arranged in ascending order.

[0082] mAP calculates the AP of all categories and then takes the average value, which can measure the detection ability of the model in all categories.

[0083]

[0084] Where K is the number of types.

[0085] mAP50:95 is the mean of mAP for all confidence levels between 50 and 95 with a step size of 5.

[0086] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement of the above embodiments made according to the technical essence of the present invention, within the spirit and principles of the present invention, without departing from the content of the technical solution of the present invention, shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. Radar active composite jamming signal recognition and parameter measurement method based on Yolov8 network with convolutional attention mechanism and rejection loss, characterized by: The steps include: Step 1: Establish a mathematical model of radar composite interference signal; Step 2: pre-processing the composite interference signal; Step 3: Input the training set into the Yolov8 network with convolutional attention mechanism and rejection loss for training; Step 4: Input the training set video images into the backbone part of the network for feature extraction; Step 5: Input the feature map into the Neck part of the network for feature fusion; Step 5.1: Upsample the feature map so that the output size is twice the input size for easy subsequent processing. Step 5.2: Fuse the upsampled feature map with the feature map obtained by the context-aware cross-level fusion network module, and concatenate the 80*80*256 and 80*80*512 feature maps into a 80*80*768 feature map. The same applies to the subsequent concatenation modules. Step 5.3: After the spliced ​​feature map passes through the context-aware cross-level fusion network module, it is input into the convolutional attention module, which consists of a channel attention module and a spatial attention module. In the channel attention module, two different pooling operations are performed on the input features of dimension H×W×C to obtain two channel information, which are then input into the multi-layer perceptron for dimensionality reduction to obtain two feature maps of dimension 1×1×C. After element-by-element summation and Sigmoid activation function processing, the channel attention features are obtained. The feature map obtained by element-wise multiplication of the channel attention feature and the input feature map is input into the spatial attention module. The two pooling operations mentioned above are also performed to obtain two feature maps with dimensions of H×W×1. The two feature maps are concatenated and then subjected to a 7×7 convolution operation. The spatial attention feature is then obtained by processing with a Sigmoid activation function. Finally, the spatial attention feature is multiplied by the input feature of the spatial attention module to obtain the final generated feature; Step 6: Input the fused features of the Neck part into the detection head module for classification and regression; The detection head module consists of two layers, one for classification and the other for regression prediction. Each layer consists of two convolution modules and one two-dimensional convolution module. The cross entropy function is used as the loss function for the classifier; the exclusion loss function and the intersection-over-union loss function are used as the loss function for the regression part. Step 7: Input the test set into the trained Yolov8 network with convolutional attention mechanism and rejection loss to obtain the interference signal components and interference signal parameters contained in the input signal.

2. The radar active composite interference signal identification and parameter measurement method based on the Yolov8 network with convolutional attention mechanism and rejection loss according to claim 1 is characterized in that, In step 1, the input signal model is: Where k represents the number of components constituting a single interference signal, J i (n) represents the received interference signal, A i is the amplitude coefficient of the modulation when propagating through the environment, awgn(n) represents Gaussian white noise, x t (n) represents the real target echo signal.

3. The radar active composite interference signal identification and parameter measurement method based on the Yolov8 network with convolutional attention mechanism and rejection loss according to claim 1 is characterized in that, In step 2, the composite interference signal is transformed into a time-frequency image X(t,ω) by the following expression: Where ω represents the frequency and h(τ-t) is the analysis window function.

4. The radar active composite jamming signal identification and parameter measurement method based on the Yolov8 network with convolutional attention mechanism and rejection loss according to claim 1 is characterized in that, Step 4 includes: Step 4.1: The input image is operated through the convolution module to extract features; Step 4.2: Input the feature map obtained by the convolution module layer into the context-aware cross-level fusion network module, which consists of a 1x1 convolution module layer and n BottleNeck layers. Step 4.3: Input the features obtained above into the pyramid pooling module. Pyramid pooling is improved from SPP in Yolov5 and consists of three maximum pooling layers, two convolutional layers and a feature fusion layer. It performs pooling operations of different sizes on the feature map and integrates the results to obtain a fixed-size output.

Images