Cross-linear process radar anti-main lobe interference detection method

Abstract

The invention belongs to the technical field of radars, and particularly relates to a cross-linear process radar anti-main lobe interference detection method. The invention provides a U-shaped anti-interference detection deep network model MSU-net of a multi-scale attention mechanism, realizes a radar anti-interference detection mode of a cross-linear processing flow in a radar search mode based on a monopulse multi-target environment, and completes a local end-to-end process from an interference-containing input end signal to real target information. During model design, a multi-scale feature extraction network, a CBAM attention module and a target detection network are adopted, and feature synchronous extraction and feature selectivity enhancement are realized. According to the design scheme, the features of different scales in the radar echo signals can be enhanced, various kinds of feature information effectively act in the training process, the robustness of a network model can be improved, then the anti-interference detection task of a cross-linear processing flow is completed, and the problem of performance superposition loss caused by a traditional staged radar signal processing mode is solved.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a radar anti-mainlobe interference detection method across linear processes. Background Art

[0002] Radar target detection in the presence of interference generally adopts a linear staged signal processing process, which requires the execution of independent processing modules such as interference perception, interference suppression, and target detection in sequence. It has the inherent defects of multi-stage error accumulation and insufficient real-time performance. Summary of the Invention

[0003] To address the above problems, the present invention proposes a radar anti-mainlobe interference target detection scheme based on a cross-linear processing flow. By constructing an integrated deep model of interference suppression and target detection, the intermediate result output links are reduced, and a local end-to-end solution from interference-containing data input to target detection result output is realized, reducing the performance loss caused by staged processing, and ultimately achieving high-precision detection of multiple targets under complex interference conditions.

[0004] The technical solution of the present invention is to propose a U-shaped anti-interference detection deep network model MSU-net (Multi-scale Attention U-shaped net) based on a multi-scale attention mechanism, use multiple parallel convolution processing paths to extract the multi-scale features of the target, combine the CBAM attention module to select features from the channel and spatial dimensions respectively, and finally realize single-pulse multi-target anti-mainlobe interference detection through a target detection network with a symmetrical U-shaped structure and jump connections; in addition, a targeted hybrid loss function BICE (Binary Cross Entropy Union DICE) is proposed to take into account both pixel-level classification capabilities and spatial structure optimization. Specifically, the following steps are included:

[0005] A radar anti-mainlobe interference detection method across linear processes is characterized by comprising the following steps:

[0006] S1. Generate radar target echo data with interference (can be generated through simulation):

[0007] Set the radar transmitter parameters and randomly generate the parameters of the jammer, radar target, and noise environment. The interference types are noise amplitude modulation interference, noise frequency modulation interference, intermittent sampling direct forwarding interference, intermittent sampling repeated forwarding interference, and intermittent sampling cyclic forwarding interference. The obtained training data set is recorded as:

[0008]

[0009] Where J represents 5 different interferences, N J is the number of samples of the J-th type of interference, is the total number of samples, is the i-th sample of the J-th type of interference, M is the number of distance units of the sample, and The detection result label is recorded as Right now:

[0010]

[0011] Where y(k) = 1 means there is a target at the k-th distance unit, and y(k) = 0 means there is no target at the k-th distance unit;

[0012] S2. Preprocess the obtained sample data:

[0013] Take samples from X After pulse compression preprocessing, the feature S is obtained, which is the input sample for the subsequent model

[0014] S3. Construct an interference detection network (defined as MSU-net interference detection network in the present invention), including a multi-scale convolution module, an attention mechanism module CBAM and a U-net network, wherein:

[0015] The input data of the multi-scale convolution module is The multi-scale convolution module contains four parallel paths. The input data passes through a 1×1 average pooling layer and a 1×3 convolution layer in the first path, with the number of convolution kernel channels being 1 and 4 respectively; the input data passes through a 1×1 convolution layer in the second path, with the number of convolution kernel channels being 6; the input data passes through a 1×1 convolution layer and a 1×3 convolution layer in the third path, with the number of convolution kernel channels being 4 and 6 respectively; the input data passes through a 1×1 convolution layer and two 1×3 convolution layers in the fourth path, with the number of convolution kernel channels being 6, 8, and 8 respectively; finally, the outputs of the four paths are connected through Concat to obtain the extracted features.

[0016] The input data of the attention mechanism module CBAM is the extracted features The spatial mapping is performed by channel and spatial attention in turn; the channel attention feature selection process is: The first path passes through a global maximum pooling layer and an MLP module in sequence, and the second path passes through a global average pooling layer and an MLP module in sequence. The MLP module includes a 1×1 convolution layer, a ReLU set function and a 1×1 convolution layer. The number of convolution kernel channels is (24 / r) and 24 respectively, where r is the compression coefficient. The output results of the two paths are added together and then passed through a Sigmoid activation function and then added to the Element-by-element multiplication yields the channel attention feature selection result The spatial attention feature selection process is: The two paths are passed in parallel. The first path performs maximum pooling and the second path performs average pooling. The results of the two paths are connected through Concat. Then they pass through a 1×7 convolution layer with a convolution kernel channel of 1. After passing the Sigmoid activation function, they are finally combined with The result of the attention mechanism module CBAM is obtained by element-by-element multiplication and is defined as the intermediate feature The calculation formula is as follows:

[0017]

[0018] Among them, M c (*) represents channel attention selection, M s (*) represents spatial attention selection;

[0019] The input of the U-net network is the intermediate feature U-net network pair The processing process is as follows: First, the encoder module is used to perform multi-layer downsampling operations. The first layer includes two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 64, and the output is obtained. Downsampling is achieved through a maximum pooling layer to obtain the downsampling result The second layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 128, and the output is Downsampling is achieved through a maximum pooling layer to obtain the downsampling result The third layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, the number of convolution kernel channels is 256, and then a Dropout layer is used to obtain the output. Finally, downsampling is achieved through a maximum pooling layer to obtain the downsampling result. The fourth layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 512. The encoder module output is Then use the decoder module to implement multi-layer upsampling operations. The first layer is upsampled through an Upsample layer to obtain the upsampled result. Then First, pass through a 1×3 convolution layer and ReLU activation function, the number of convolution kernel channels is 256, and then Connect through Concat to get the output The second layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, the number of convolution kernel channels is 256, and upsampling is achieved through an Upsample layer to obtain the upsampling result. Then First, pass through a 1×3 convolution layer and ReLU activation function, the number of convolution kernel channels is 128, and then Connect through Concat to get the output The third layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, the number of convolution kernel channels is 128, and upsampling is achieved through an Upsample layer to obtain the upsampling result. Then First, pass through a 1×3 convolution layer and ReLU activation function, the number of convolution kernel channels is 64, and then Connect through Concat to get the output The fourth layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 64. The decoder module output is Finally, through a 1×3 convolution layer, a ReLU activation function, a 1×1 convolution layer and a linear layer, where the number of convolution kernel channels is 256 and 1 respectively, the final output of the interference detection network is obtained, which is defined as the distance unit map

[0020] S4. Train the constructed interference detection network and use the gradient descent method to update the network parameters. The loss function used is:

[0021]

[0022] Where λ1 and λ2 are hyperparameters that control the contribution of each part; The binary cross entropy loss is defined as:

[0023]

[0024] is the DICE loss, defined as:

[0025]

[0026] Where yi is whether there is a real target in the i-th distance unit, which is 1 if there is, and 0 if not, pi is the probability that the model predicts that there is a real target in the i-th distance unit, N is the total number of samples, and ∈ is the smoothing factor;

[0027] S5. Use the trained interference detection network to perform anti-interference target detection on the input samples.

[0028] The beneficial effects of the present invention are as follows: in response to the information superposition loss problem caused by the linear stage of traditional radar systems, the present invention proposes a U-shaped anti-interference detection deep network model MSU-net with a multi-scale attention mechanism based on a radar anti-mainlobe interference detection scheme of a cross-linear processing flow, which realizes a radar anti-interference detection method of a cross-linear processing flow in a radar search mode based on a single-pulse multi-target environment, and completes the local end-to-end process from the input signal containing interference to the real target information. The model design adopts a multi-scale feature extraction network, a CBAM attention module and a target detection network to achieve feature synchronization extraction and feature selectivity enhancement. This design scheme can enhance the features of different scales in the radar echo signal. The multiple feature information interacts effectively with each other during training, which can improve the robustness of the network model, and then complete the anti-interference detection task of the cross-linear processing flow, solving the performance superposition loss problem caused by the traditional staged radar signal processing method. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is the MSU-net interference detection network flow chart.

[0030] Figure 2 It is a schematic diagram of the multi-scale convolution module.

[0031] Figure 3 This is a schematic diagram of the attention mechanism module CBAM.

[0032] Figure 4 This is a schematic diagram of the U-net network. DETAILED DESCRIPTION

[0033] The present invention will be described in detail below with reference to the accompanying drawings.

[0034] Generate radar echo data by simulating according to Table 1 below:

[0035] Table 1. Training sample generation parameters

[0036]

[0037] Specifically: set the radar operating parameters (the radar transmission signal is a linear frequency modulation signal, with a bandwidth of 10MHz and a time width of 50μs); the number of targets is randomly generated from 1 to 5, and the distance between each target and the radar is random within 15Km; the interference types include noise amplitude modulation interference, noise frequency modulation interference, intermittent sampling direct forwarding interference, intermittent sampling repeated forwarding interference, and intermittent sampling cyclic forwarding interference. The jammer operating parameters are shown in the table. The interference-to-signal ratio of each sample is randomly generated between 5 and 30dB. The sampling time of each pulse echo is 100μs, that is, the length of each simulated echo data is 10,000 sampling points. The training data set is recorded as:

[0038]

[0039] Where J represents 5 different interferences, N J is the number of samples of the J-th type of interference, is the total number of samples, is the i-th sample of the J-th type of interference, M is the number of distance units of the sample, and The detection result label is recorded as Right now:

[0040]

[0041] Where y(k)=1 represents that there is a target on the k-th distance unit, and y(k)=0 represents that there is no target on the k-th distance unit.

[0042] First, take the input signal from the sample set X obtained in step S1 After pulse compression preprocessing, the feature S is obtained, which is the input sample of the subsequent interference detection network.

[0043] like Figure 2 As shown, the characteristics The dimension is 1×1×10000. It first enters the multi-scale convolution module: it contains four parallel paths. The first path passes through a 1×1 average pooling layer and a 1×3 convolution layer, with the number of convolution kernel channels being 1 and 4 respectively, to obtain features with a dimension of 4×1×5000; the second path passes through a 1×1 convolution layer with a number of convolution kernel channels being 6, to obtain features with a dimension of 6×1×5000; the third path passes through a 1×1 convolution layer and a 1×3 convolution layer with a number of convolution kernel channels being 4 and 6 respectively, to obtain features with a dimension of 8×1×5000; the fourth path passes through a 1×1 convolution layer and two 1×3 convolution layers with a number of convolution kernel channels being 6, 8, and 8 respectively, to obtain features with a dimension of 8×1×5000; finally, the four paths are connected through Concat to obtain the extracted features with a dimension of 24×1×5000 The specific calculation formula is as follows:

[0044]

[0045] in It represents a convolution operation with a convolution kernel size of 1×3, 1 input feature channel, and 4 output feature channels.

[0046] The features after merging and extraction are of dimension 24×1×5000 Enter the attention mechanism module CBAM: spatial mapping is performed through channel and spatial attention in turn. The calculation formula is as follows:

[0047]

[0048] First, the channel attention feature selection: The first path passes through a global maximum pooling layer to make the feature dimension 24×1×1, and then passes through an MLP module. The second path passes through a global average pooling layer to make the feature dimension 24×1×1, and then passes through an MLP module, wherein the MLP module includes a 1×1 convolution layer, a ReLU set function and a 1×1 convolution layer, and the number of convolution kernel channels is (24 / r) and 24 respectively, where r is the compression coefficient, and r=24 is taken in the present invention; the results of the two paths are added and then passed through a Sigmoid activation function, and then Element-by-element multiplication yields the channel attention feature selection result Its dimensions are 24×1×5000, and the specific calculation formula is as follows:

[0049]

[0050] M c (*)=Sigmoid(MLP(AdaptiveMaxPool(*))+MLP(AdaptiveAvgPool(*)))

[0051]

[0052] Then comes the spatial attention feature selection: The first path is used for maximum pooling, and the second path is used for average pooling. Both paths make the feature dimension become 1×1×5000. The results of the two paths are connected along the channel dimension through Concat to make the dimension become 2×1×5000. Then, a 1×7 convolution layer is passed with the convolution kernel channel number of 1. After the Sigmoid activation function, the convolution kernel is activated and the convolution kernel is activated. The result of the attention mechanism module CBAM is obtained by element-by-element multiplication, which is the intermediate feature Its dimensions are 24×1×5000, and the specific calculation formula is as follows:

[0053]

[0054]

[0055] like Figure 4 As shown, the intermediate features Entering the U-net network, the encoder module first performs downsampling. The first layer includes two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 64. The output is Downsampling is achieved through a maximum pooling layer to obtain the downsampling result The feature dimension is reduced to 64×1×2500; the second layer includes two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 128, and the output is obtained. Downsampling is achieved through a maximum pooling layer to obtain the downsampling result The feature dimension is reduced to 128×1×1250; the third layer includes two 1×3 convolutional layers, each followed by a ReLU activation function, the number of convolution kernel channels is 256, and then a Dropout layer is used to obtain the output. Downsampling is achieved through a maximum pooling layer to obtain the downsampling result The feature dimension is reduced to 256×1×625; the fourth layer includes two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 512, and the encoder module output is obtained. Then the decoder module implements the upsampling operation. The first layer is upsampled through an Upsample layer to obtain the upsampled result. Restore the feature dimension to 512×1×1250, and then First, pass through a 1×3 convolution layer and ReLU activation function, the number of convolution kernel channels is 256, and then Connect through Concat to get the output The second layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, the number of convolution kernel channels is 256, and upsampling is achieved through an Upsample layer to obtain the upsampling result. Restore the feature dimension to 256×1×2500, and then First, pass through a 1×3 convolution layer and ReLU activation function, the number of convolution kernel channels is 128, and then Connect through Concat to get the output The third layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, the number of convolution kernel channels is 128, and upsampling is achieved through an Upsample layer to obtain the upsampling result. Restore the feature dimension to 128×1×5000, and then First, pass through a 1×3 convolution layer and ReLU activation function, the number of convolution kernel channels is 64, and then Connect through Concat to get the output The fourth layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 64. The decoder module output is Its feature dimension is 64×1×5000; finally, through a 1×3 convolution layer, a ReLU activation function, a 1×1 convolution layer and a linear layer, where the number of convolution kernel channels is 256 and 1 respectively, the final output of the MSU-net interference detection network is obtained, which is the distance unit map Its feature dimension is 1×1×1000, and the specific calculation formula is as follows:

[0056]

[0057]

[0058] The hybrid loss function BICE (Binary Cross Entropy Union DICE) is used for training. The specific calculation formula is as follows:

[0059]

[0060] Among them, λ1=0.2 and λ2=0.8 are taken as hyperparameters to control the contribution of each part, and training is unlimited until the model converges.

[0061] In different interference patterns, a total of 6 JSRs and 5 targets are set. Each target in a single sample has an interference and all interference types are consistent. Each sample generates 10,000 sets of echo signals as test samples. After preprocessing, the test samples are extracted through a multi-scale convolution module, and then spatially mapped through the attention mechanism module CBAM to obtain intermediate features. Finally, they enter the U-net network to obtain distance unit mapping, realize the detection of targets at different positions, return the probability of the existence of targets at different positions, and judge whether there is a target at that position by the threshold.

[0062] Finally, under the complex conditions of JSR = 0-30dB and number of targets = 1-5, the average detection rate of the model for intermittent sampling interference reached 93%, and it still maintained a detection rate of more than 80% in the 5-target scenario; the detection rate for suppression interference within 3 targets exceeded 82%, verifying the effectiveness and robustness of the present invention.

[0063] Tables 2 and 3 below are some of the test results of the present invention:

[0064] Table 2 Comparison of anti-interference detection results of intermittent sampling and forwarding interference types when the number of targets is 3 (%)

[0065]

[0066]

[0067] Table 3 Comparison of anti-interference detection results of noise interference types when the number of targets is 3 (%)

[0068]

[0069]

Claims

1. A radar anti-mainlobe interference detection method across linear processes, characterized in that: The following steps are involved: S1. Generate radar target echo data with interference: Set the radar transmitter parameters and randomly generate the parameters of the jammer, radar target, and noise environment. The interference types are noise amplitude modulation interference, noise frequency modulation interference, intermittent sampling direct forwarding interference, intermittent sampling repeated forwarding interference, and intermittent sampling cyclic forwarding interference. The obtained training data set is recorded as: Where J represents 5 different interferences, N J is the number of samples of the J-th type of interference, is the total number of samples, is the i-th sample of the J-th type of interference, M is the number of distance units of the sample, and The detection result label is recorded as Right now: Where y(k) = 1 means there is a target at the k-th distance unit, and y(k) = 0 means there is no target at the k-th distance unit; S2. Preprocess the obtained sample data: Take samples from X After pulse compression preprocessing, the feature S is obtained, which is the input sample for the subsequent model S3. Construct an interference detection network, including a multi-scale convolution module, an attention mechanism module (CBAM), and a U-net network. The input data of the multi-scale convolution module is The multi-scale convolution module contains four parallel paths. The input data passes through a 1×1 average pooling layer and a 1×3 convolution layer in the first path, with the number of convolution kernel channels being 1 and 4 respectively; the input data passes through a 1×1 convolution layer in the second path, with the number of convolution kernel channels being 6; the input data passes through a 1×1 convolution layer and a 1×3 convolution layer in the third path, with the number of convolution kernel channels being 4 and 6 respectively; the input data passes through a 1×1 convolution layer and two 1×3 convolution layers in the fourth path, with the number of convolution kernel channels being 6, 8, and 8 respectively; finally, the outputs of the four paths are connected through Concat to obtain the extracted features. The input data of the attention mechanism module CBAM is the extracted features The spatial mapping is performed by channel and spatial attention in turn; the channel attention feature selection process is: The first path passes through a global maximum pooling layer and an MLP module in sequence, and the second path passes through a global average pooling layer and an MLP module in sequence. The MLP module includes a 1×1 convolution layer, a ReLU set function and a 1×1 convolution layer. The number of convolution kernel channels is (24 / r) and 24 respectively, where r is the compression coefficient. The output results of the two paths are added together and then passed through a Sigmoid activation function and then added to the Element-by-element multiplication yields the channel attention feature selection result The spatial attention feature selection process is: The two paths are passed in parallel. The first path performs maximum pooling and the second path performs average pooling. The results of the two paths are connected through Concat. Then they pass through a 1×7 convolution layer with a convolution kernel channel of 1. After passing the Sigmoid activation function, they are finally combined with The result of the attention mechanism module CBAM is obtained by element-by-element multiplication and is defined as the intermediate feature The calculation formula is as follows: Among them, M c (*) represents channel attention selection, M s (*) represents spatial attention selection; The input of the U-net network is the intermediate feature U-net network pair The processing process is as follows: First, the encoder module is used to perform multi-layer downsampling operations. The first layer includes two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 64, and the output is obtained. Downsampling is achieved through a maximum pooling layer to obtain the downsampling result The second layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 128, and the output is Downsampling is achieved through a maximum pooling layer to obtain the downsampling result The third layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, the number of convolution kernel channels is 256, and then a Dropout layer is used to obtain the output. Finally, downsampling is achieved through a maximum pooling layer to obtain the downsampling result. The fourth layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 512. The encoder module output is Then use the decoder module to implement multi-layer upsampling operations. The first layer is upsampled through an Upsample layer to obtain the upsampled result. Then First, pass through a 1×3 convolution layer and ReLU activation function, the number of convolution kernel channels is 256, and then Connect through Concat to get the output The second layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, the number of convolution kernel channels is 256, and upsampling is achieved through an Upsample layer to obtain the upsampling result. Then First, pass through a 1×3 convolution layer and ReLU activation function, the number of convolution kernel channels is 128, and then Connect through Concat to get the output The third layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, the number of convolution kernel channels is 128, and upsampling is achieved through an Upsample layer to obtain the upsampling result. Then First, pass through a 1×3 convolution layer and ReLU activation function, the number of convolution kernel channels is 64, and then Connect through Concat to get the output The fourth layer consists of two 1×3 convolutional layers, each followed by a ReLU activation function, and the number of convolution kernel channels is 64. The decoder module output is Finally, through a 1×3 convolution layer, a ReLU activation function, a 1×1 convolution layer and a linear layer, where the number of convolution kernel channels is 256 and 1 respectively, the final output of the interference detection network is obtained, which is defined as the distance unit map S4. Train the constructed interference detection network and use the gradient descent method to update the network parameters. The loss function used is: Where λ1 and λ2 are hyperparameters that control the contribution of each part; The binary cross entropy loss is defined as: is the DICE loss, defined as: Where yi is whether there is a real target in the i-th distance unit, which is 1 if there is, and 0 if not, pi is the probability that the model predicts that there is a real target in the i-th distance unit, N is the total number of samples, and ∈ is the smoothing factor; S5. Use the trained interference detection network to perform anti-interference target detection on the input samples.

Images