Lightweight radar active composite jamming fast identification method

Abstract

The application discloses a kind of lightweight radar active composite jamming fast identification method.First, the time domain, frequency domain, time-frequency domain characteristics of single jamming and composite jamming are analyzed, and the interference time-frequency diagram dataset is constructed.Second, build lightweight identification network based on MobileNet v3: adopt parallel 1×5 and 5×1 asymmetric depth separable convolution to constitute multi-scale residual block, extract time-frequency graph features;Embedded coordinate attention module refines features;Classification and identification are completed through global average pooling, fully connected layer and Dropout layer.Finally, the model is trained and tested.The application realizes high accuracy and fast identification of radar composite jamming, the model parameter quantity and calculation amount decrease significantly, effectively solves the problem of multiple composite jamming types and difficult identification, and is suitable for resource-limited engineering applications.

Description

Technical Field

[0001] This invention relates to the fields of radar signal processing and electronic countermeasures, specifically to a method for rapid identification of active composite radar interference based on a lightweight deep learning network model. Background Technology

[0002] Radar technology plays a crucial role in target detection, target classification and identification, and radar imaging. By emitting electromagnetic waves and receiving target echo signals, radar detects and locates enemy targets, enabling real-time tracking and observation of enemy operations and distribution. However, with the increasing types of radar interference and the combined effects of various interference methods, radar applications are hampered. Traditional interference identification methods primarily rely on manual feature extraction or traditional network identification, resulting in low accuracy and excessively complex models, making them unsuitable for deployment in electronic warfare environments.

[0003] To overcome the aforementioned research bottlenecks, radar interference identification technology employs novel neural networks, such as CNNs, to identify and classify various types of interference. However, a balance between accuracy and model complexity is difficult to achieve. Traditional networks struggle to identify complex interference because it possesses characteristics of multiple types of interference, making it impossible to accurately capture and classify the time-frequency plot features. Therefore, there is an urgent need to develop a lightweight neural network that combines low model complexity with high accuracy, capable of precisely identifying both single and complex interference. Summary of the Invention

[0004] The technical problem to be solved by this invention is that existing algorithms can only identify single interferences and fail when faced with complex interferences. This invention provides a novel lightweight network identification method that not only has a small number of model parameters, but also achieves high-precision identification of complex interferences, thereby improving the model's adaptability and identification accuracy in complex electromagnetic environments.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solution, including the following steps:

[0006] A lightweight radar active composite interference rapid identification method includes the following steps:

[0007] S1. Perform mathematical modeling on various single active interferences and composite interferences composed of at least two single active interferences, analyze the characteristics of various interferences in the time domain, frequency domain, and time-frequency domain, generate time-frequency images of interferences through time-frequency analysis methods, and randomly generate interference samples within a preset signal-to-noise ratio range to construct an interference dataset based on time-frequency images.

[0008] S2. Construct a lightweight radar interference identification network based on asymmetric multi-scale features and coordinate attention, wherein: a multi-scale residual block is constructed, consisting of a first asymmetric depthwise separable convolutional branch and a second asymmetric depthwise separable convolutional branch in parallel, wherein the kernel sizes of the two branches are transposes of each other and are not equal, for extracting frequency domain spectral features from the time-frequency map; a coordinate attention mechanism module is embedded to refine the extracted features and achieve accurate localization of composite interference features; and an interference type discrimination is completed through a classification and identification sub-network consisting of a global average pooling layer, a first fully connected layer, a Dropout random deactivation layer, and a second fully connected layer.

[0009] S3. Train and test the lightweight radar jamming identification model to verify the effectiveness and speed of the model's feature extraction, and finally achieve rapid identification of active composite radar jamming for resource-constrained equipment.

[0010] Preferably, the single active interference in step S1 includes intermittent sampling forwarding interference, comb spectrum interference, and noise amplitude modulation interference; the composite interference includes intermittent sampling forwarding interference + comb spectrum interference, intermittent sampling forwarding interference + noise amplitude modulation interference, comb spectrum interference + noise amplitude modulation interference, and intermittent sampling forwarding interference + comb spectrum interference + noise amplitude modulation interference.

[0011] Preferably, the preset signal-to-noise ratio range in step S1 is -6dB to 10dB.

[0012] Preferably, the mathematical model of the noise amplitude modulation interference is expressed as follows:

[0013]

[0014] in, It is an exponential function with base e. It is zero-mean Gaussian white noise. It is the carrier voltage. The modulation coefficient of the noise. For the carrier frequency of the interference, phase exist Evenly distributed on top and with They are independent of each other;

[0015] The mathematical model for the comb-like spectral interference is expressed as follows:

[0016]

[0017] Among them, the interval between adjacent peaks is ,like This creates a false target in the distance direction. It has a suppressive effect;

[0018] The mathematical model for the intermittent sampling forwarding interference is expressed as follows:

[0019]

[0020] in, For rectangle functions, The width of the intermittent sampling pulse. For radar signal pulse width, The sampling period is... This represents the intermittent sampling duty cycle.

[0021] Preferably, in step S2, the kernel size of the first asymmetric depth-separable convolution branch is 1×5, and the kernel size of the second asymmetric depth-separable convolution branch is 5×1.

[0022] Preferably, the lightweight radar interference identification network in step S2 is built on MobileNet v3.

[0023] Preferably, the multi-scale residual block composed of the parallel first asymmetric depthwise separable convolutional branch and the second asymmetric depthwise separable convolutional branch in step S2 is specifically processed as follows: the input feature map is split into two parallel asymmetric depthwise separable convolutional branches, wherein the horizontal branch uses 1×5 convolution to capture the narrow-band spectral features of comb-like spectral interference using the long receptive field on the frequency axis; the vertical branch uses 5×1 convolution to extract the pulse slices and duty cycle patterns of intermittent sampling forwarding interference along the time dimension; the output feature maps of the two branches are added and fused pixel by pixel by the Add operator, and then output by a 1×1 pointwise convolutional layer.

[0024] Preferably, the specific processing method of the coordinate attention mechanism module in step S2 is as follows: directional pooling is performed on the input feature map along the horizontal and vertical directions to encode the spatial information into two one-dimensional vectors. and It captures the feature distribution of rows and columns in the time-frequency map; it encodes the information in the two directions through joint convolution, then separates and generates horizontal and vertical attention weight maps; it multiplies these two weight maps and applies them to the original features to achieve precise weighting of specific spatial locations in the feature map.

[0025] Preferably, in step S3, the model is trained for 50 rounds using a loss function combining Label Smoothing and Focal Loss.

[0026] Preferably, in step S3, accuracy, number of parameters, computational load, and average processing time per frame are used as evaluation metrics to compare the network with the unmodified MobileNet v3 network, thereby verifying the network's lightweight and fast processing advantages.

[0027] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are as follows:

[0028] 1. This invention addresses the multi-scale and anisotropic characteristics of composite interference time-frequency maps by designing a multi-scale residual block composed of parallel 1×5 and 5×1 asymmetric depth-separable convolutional branches. The lateral branches utilize the long receptive field on the frequency axis to accurately capture the narrow-band spectral features of comb-like interference, while the vertical branches deeply aggregate along the time dimension to extract pulse slices and duty cycle patterns of intermittent sampling and forwarding interference. This effectively overcomes the deficiency of traditional symmetric convolutional kernels in simultaneously capturing two types of directional features, significantly improving the radar's accuracy in identifying single and composite interference.

[0029] 2. This invention employs a depthwise separable convolution and asymmetric decomposition strategy, splitting the residual block into two parallel asymmetric depthwise separable convolution branches, and introducing a Dropout random deactivation layer in the classification sub-network. Compared with existing lightweight networks, this invention further reduces the number of model parameters and computational load, significantly reduces storage space and computing resource requirements, and shortens the average processing time of a single frame time-frequency map, enabling deployment in resource-constrained electronic warfare equipment such as UAVs and embedded systems.

[0030] 3. This invention embeds a coordinate attention mechanism into a lightweight network. By performing directional pooling on the feature map along the horizontal and vertical directions, spatial information is encoded into two one-dimensional vectors, which are then jointly convolutionally encoded to generate horizontal and vertical attention weight maps. This mechanism can accurately weight specific time-frequency coordinate regions where composite interference signals exist, effectively suppressing background noise and irrelevant features, and improving the robustness of the model in complex electromagnetic environments.

[0031] 4. This invention employs a loss function combining Label Smoothing and Focal Loss to alleviate the class imbalance problem of difficult-to-classify samples. Experimental verification shows that under different low signal-to-noise ratio conditions, this invention maintains a high level of recognition accuracy for both multiple complex interferences and single interferences, demonstrating good noise resistance and stability.

[0032] 5. This invention balances lightweight design with high performance. While maintaining high recognition accuracy, it significantly reduces model complexity and processing latency, achieving a good balance between recognition speed and accuracy. It is suitable for rapid electronic warfare scenarios in practical engineering. Attached Figure Description

[0033] Figure 1 This is a flowchart of the overall identification method of the present invention.

[0034] Figure 2 This is a schematic diagram of the structure of the core module AMD-BLOCK of this invention.

[0035] Figure 3 The time-frequency diagrams are for seven types of interference.

[0036] Figure 4 This represents the overall recognition confusion matrix under full signal-to-noise ratio conditions. Detailed Implementation

[0037] To address the shortcomings of existing methods for identifying radar interference based on various neural networks, such as high model complexity and difficulty in accurately identifying multiple types of complex interference, this invention proposes a solution based on the existing lightweight network MobileNet v3. This solution introduces asymmetric multi-scale depthwise separable convolutional AMD-Block and a spatially perceptual attention mechanism CA, thereby reducing the number of parameters and meeting the requirement for accurate classification of various types of complex interference.

[0038] The present invention proposes a lightweight radar active composite interference rapid identification method, comprising the following steps:

[0039] S1. Mathematical modeling is performed on multiple single active interferences and composite interferences composed of at least two single active interferences. Short-time Fourier transform is used to generate a two-dimensional time-frequency image to simulate the echo signal received by the radar receiver. An interference dataset is constructed based on the time-frequency image for training and testing the neural network.

[0040] S2. Construct a lightweight radar interference identification network based on asymmetric multi-scale features and coordinate attention. After the feature map is input, it is dimensionality reduced by the Stem layer and then passed through the AMD-BLOCK module to identify the spectral features in the time-frequency map. The feature map flows through CA to perform position encoding on the horizontal and vertical sides, enhancing the specific time-frequency coordinate regions where the interference signal exists. At this point, the time-domain feature extraction is completed. It then undergoes two more downsampling steps with a step size of 2. It flows through a classification and identification subnetwork consisting of a global average pooling layer, a first fully connected layer, a Dropout random deactivation layer, and a second fully connected layer to finally determine the type of interference.

[0041] S3. Train and test the lightweight radar jamming identification model to verify the effectiveness and speed of the model's feature extraction, and finally achieve rapid identification of active composite radar jamming for resource-constrained equipment.

[0042] The technical solution of the present invention will be further described in detail below:

[0043] Three common types of simple interference in the current environment were selected: intermittent sampling forwarding interference, comb spectrum interference, and noise stripe interference. Four types of composite interference were synthesized by combining two or all three simultaneously: intermittent sampling forwarding interference + comb spectrum interference, intermittent sampling forwarding interference + noise amplitude modulation interference, comb spectrum interference + noise amplitude modulation interference, and intermittent sampling forwarding interference + comb spectrum interference + noise amplitude modulation interference, for a total of seven types of interference. The above single and composite interferences cover various single active interferences and their combinations.

[0044] Mathematical analysis of the three types of simple disturbances yields the following results:

[0045] The mathematical model for amplitude modulation noise interference is expressed as follows:

[0046]

[0047] From the functional expression, we can deduce that the disturbance is a generalized stationary random process, where exp is an exponential function with base e. It is zero-mean Gaussian white noise. It is the carrier voltage and the modulation coefficient of the noise. control power, The carrier phase of the interference It is uniformly distributed on [0,2] and and They are independent random variables.

[0048] The mathematical model for comb-like spectral interference is expressed as follows:

[0049]

[0050] Where the interval between adjacent peaks is t, if the range resolution of the LFM radar is represented by 1 / B, then if This will create false targets in the direction of distance; the more false targets generated, the better the deception effect. The receiver has difficulty distinguishing multiple spikes, and the resulting comb-like interference has a suppressing effect.

[0051] The mathematical model for intermittent sampling forwarding interference is expressed as follows:

[0052]

[0053] in, It is a rectangle function. It is the width of the intermittently sampled pulse, and T is the pulse width of the radar signal. It is the sampling period. This indicates the intermittent sampling duty cycle.

[0054] After obtaining the mathematical relationship, time-frequency analysis was performed using short-time Fourier transform to construct an interference time-frequency graph dataset. The signal-to-noise ratio was set to vary from -6dB to 10dB, and 900 samples of each type of interference were synthesized, for a total of 6300 samples.

[0055] Then, the initial processing stage of the feature map begins. The 128×128×3 dimensional time-frequency image is input into the initial layer of Stem, and a 3×3 convolution kernel with a stride of 2 is used to complete the rapid downsampling of the spatial dimension, which suppresses high-frequency background noise while expanding the feature channels to 16 dimensions.

[0056] The pre-processed feature map enters the AMD-BLOCK module, whose structure is as follows: Figure 2 As shown, this module consists of two parallel asymmetric depthwise separable convolutional branches, one first and one second. The kernel sizes of the two branches are transposes of each other and are not equal. In this embodiment, the feature map is first split into two parallel feature extraction paths. The horizontal path uses a 1×5 convolutional branch, which utilizes its continuous receptive field of 5 pixels on the frequency axis to extract spectral features from the time-frequency map and initially filter out randomly distributed impulse noise. At the same time, the vertical path uses a 5×1 branch, whose 5-pixel span on the time axis can perform deep aggregation of impulse slice patterns, thereby capturing the duty cycle of the interference signal in the time dimension. Subsequently, the anisotropic feature maps generated by these two paths are added and fused pixel by pixel at the Add operator, and then converged into a composite feature map containing multi-dimensional physical properties, which then flows through a 1×1 pointwise convolutional layer.

[0057] After initial feature extraction, the feature map enters the CA coordinate attention mechanism module. Specifically, the input feature map is directionally pooled along the horizontal and vertical directions, encoding spatial information into two one-dimensional vectors. and The algorithm captures the feature distribution of rows and columns in the time-frequency image; it then performs joint convolutional encoding on the information from both directions, and then separates and generates horizontal and vertical attention weight maps; these two weight maps are multiplied and applied to the original features to achieve precise weighting of specific spatial locations in the feature map. Through this mechanism, precise feature locations are reselected using channel weights, and the weighted output is a refined and accurate feature map.

[0058] The time-frequency map after feature extraction flows through three Mobile-Bottleneck structure stages. By alternately performing downsampling with a step size of 2 and channel doubling operations, the feature map size is gradually compressed to 16×16, while the channel depth is expanded to 112 dimensions. Finally, it flows through a global average pooling layer, compressing the feature information into a 112-dimensional one-dimensional feature vector. Finally, a classifier with a Dropout random deactivation mechanism, consisting of a global average pooling layer, a first fully connected layer, a Dropout random deactivation layer, and a second fully connected layer, outputs the probability distribution of 7 types of interference.

[0059] The effectiveness of the implementation method of the present invention is verified through simulation experiments.

[0060] Seven categories were set up, and 900 time-frequency plots with 9 different signal-to-noise ratios were analyzed under each category, totaling 6300 data points. After interference input, the network identified each interference and obtained the interference identification confusion matrix, as shown in the appendix. Figure 3 As shown, the recognition accuracy of each type of interference is high. Except for the noise amplitude modulation interference, which has a recognition accuracy of 94%, the accuracy of the other interferences is above 98%, and the overall recognition accuracy is as high as 98%, which significantly improves the accuracy of radar interference recognition.

[0061] During training, a loss function combining Label Smoothing and Focal Loss is used. Labels are first processed using Label Smoothing, and then the smoothed labels are substituted into Focal Loss for calculation. In this embodiment, the model is trained for 50 epochs to find the maximum accuracy, and the number of parameters and computational cost of this lightweight network are calculated.

[0062] The parameter count comparison between the lightweight network based on depthwise separable convolution and the MobileNet v3 network is shown in Table 1. This experiment uses accuracy, parameter count, computational cost, and average processing time per frame as evaluation metrics.

[0063] Table 1 Reference Table

[0064]

[0065] As shown in Table 1, the lightweight network proposed in this invention has significant advantages. Compared with the unmodified MobileNetv3, the total computation is only 28% of that of the unmodified network, which significantly reduces the model complexity. The total number of parameters is also much lower than that of the unmodified network, effectively solving the problem of high resource consumption.

[0066] Meanwhile, in terms of computational efficiency, the interference identification algorithm based on this lightweight network showed in the test under the NVIDIA GeForce RTX 4060 graphics processor that the average identification time of a single frame time-frequency map was only 2.11ms, which is short and only 68% of the benchmark network. The identification time has been greatly improved, and the rapid identification of complex interference has been achieved.

[0067] Compared with traditional methods, this invention has high recognition accuracy, low resource consumption, and short recognition time. It also has good adaptability in environments with complex interference, and has a wide range of application prospects.

Claims

1. A lightweight radar active composite interference rapid identification method, characterized in that, Includes the following steps: S1. Perform mathematical modeling on various single active interferences and composite interferences composed of at least two single active interferences, analyze the characteristics of various interferences in the time domain, frequency domain, and time-frequency domain, generate time-frequency images of interferences through time-frequency analysis methods, and randomly generate interference samples within a preset signal-to-noise ratio range to construct an interference dataset based on time-frequency images. S2. Construct a lightweight radar interference identification network based on asymmetric multi-scale features and coordinate attention mechanism, wherein: construct a multi-scale residual block composed of a first asymmetric depth separable convolution branch and a second asymmetric depth separable convolution branch in parallel, wherein the convolution kernel sizes of the two branches are transposes of each other and are not equal, and are used to extract frequency domain spectral features in the time-frequency plot. An embedded coordinate attention mechanism module is used to refine the extracted features and achieve accurate localization of composite interference features; the interference type is determined by a classification and recognition subnetwork consisting of a global average pooling layer, a first fully connected layer, a Dropout random deactivation layer, and a second fully connected layer. S3. Train and test the lightweight radar jamming identification model to verify the effectiveness and speed of the model's feature extraction, and finally achieve rapid identification of active composite radar jamming for resource-constrained equipment.

2. The lightweight radar active composite interference rapid identification method according to claim 1, characterized in that, The single active interference in step S1 includes intermittent sampling forwarding interference, comb spectrum interference, and noise amplitude modulation interference; the composite interference includes intermittent sampling forwarding interference + comb spectrum interference, intermittent sampling forwarding interference + noise amplitude modulation interference, comb spectrum interference + noise amplitude modulation interference, and intermittent sampling forwarding interference + comb spectrum interference + noise amplitude modulation interference.

3. The lightweight radar active composite interference rapid identification method according to claim 1, characterized in that, In step S1, the preset signal-to-noise ratio range is -6dB to 10dB.

4. The lightweight radar active composite interference rapid identification method according to claim 2, characterized in that: The mathematical model for the amplitude-modulated noise interference is expressed as follows: ； in, It is an exponential function with base e. It is zero-mean Gaussian white noise. It is the carrier voltage. The modulation coefficient of the noise. For the carrier frequency of the interference, phase exist Evenly distributed on top and with They are independent of each other; The mathematical model for the comb-like spectral interference is expressed as follows: ； Among them, the interval between adjacent peaks is ,like This creates a false target in the distance direction. It has a suppressive effect; The mathematical model for the intermittent sampling forwarding interference is expressed as follows: ； in, For rectangle functions, The width of the intermittent sampling pulse. For radar signal pulse width, The sampling period is This represents the intermittent sampling duty cycle.

5. The lightweight radar active composite interference rapid identification method according to claim 1, characterized in that, In step S2, the kernel size of the first asymmetric horizontal depth separable convolution branch is 1×5, and the kernel size of the second asymmetric vertical depth separable convolution branch is 5×1.

6. The lightweight radar active composite interference rapid identification method according to claim 1, characterized in that, The lightweight radar jamming identification network in step S2 is built on MobileNet v3.

7. The lightweight radar active composite interference rapid identification method according to claim 1, characterized in that, In step S2, the multi-scale residual block composed of the first and second parallel asymmetric depth separable convolution branches is specifically processed as follows: the input feature map is split into two parallel asymmetric depth separable convolution branches, wherein the horizontal branch adopts 1×5 convolution, and the long receptive field on the frequency axis is used to capture the narrow band spectral features of comb-like spectral interference. The vertical branch uses a 5×1 convolution to deeply aggregate and extract the pulse slices and duty cycle patterns of intermittent sampling and forwarding interference along the time dimension; the output feature maps of the two branches are added and fused pixel by pixel by the Add operator, and then output through a 1×1 pointwise convolutional layer.

8. The lightweight radar active composite interference rapid identification method according to claim 1, characterized in that, The specific processing method of the coordinate attention mechanism module in step S2 is as follows: directional pooling is performed on the input feature map along the horizontal and vertical directions to encode the spatial information into two one-dimensional vectors. and It captures the feature distribution of rows and columns in the time-frequency map; it encodes the information in the two directions through joint convolution, then separates and generates horizontal and vertical attention weight maps; it multiplies these two weight maps and applies them to the original features to achieve precise weighting of specific spatial locations in the feature map.

9. A lightweight radar active composite interference rapid identification method according to claim 1, characterized in that, In step S3, the model is trained for 50 rounds using a loss function combining Label Smoothing and Focal Loss.

10. A lightweight radar active composite interference rapid identification method according to claim 1, characterized in that, In step S3, accuracy, number of parameters, computational load, and average processing time per frame are used as evaluation metrics to compare the network with the unmodified MobileNet v3 network, verifying the network's lightweight and fast processing advantages.

Images