Intermittent sampling and forwarding interference suppression method based on Fast-SCNN

Abstract

The invention belongs to the technical field of radars, and particularly relates to an intermittent sampling and forwarding interference suppression method based on Fast-SCNN. According to the method, a time-frequency analysis and deep learning semantic segmentation method is adopted, interference and target superposed receiving signals are used as input data, and regional distribution of interference signals is used as label data. And after preprocessing, the features are converted into time-frequency domain two-dimensional features, so that the interference can be conveniently extracted and suppressed. And subsequently, utilizing a Fast-scnn semantic segmentation network to more accurately predict the interference region in the time-frequency two-dimensional matrix. And in combination with a signal reconstruction method, the target echo signal is further restored, so that the suppression effect on intermittent sampling forwarding interference is better realized.

Description

Technical Field

[0001] The invention belongs to the field of radar technology, and in particular relates to an intermittent sampling forwarding interference suppression method based on Fast-SCNN. Background Art

[0002] In recent years, the emergence of Digital Radio Frequency Memory (DRFM) has brought significant challenges to radar anti-interference technology. Based on the DRFM technology, the intermittent sampling and forwarding interference pattern adopts a fragment storage and multiple forwarding interference strategy. This method can accurately generate interference signals that are highly coherent with the linear frequency modulation radar transmission wave, and form single or clustered false targets after pulse compression processing, thereby misleading and suppressing the radar system. Therefore, the identification and suppression technology of the intermittent sampling and forwarding interference pattern is of great significance.

[0003] Since the signal parameters of the interferer are random and unknown in actual application scenarios, if the interference suppression technology can only process interference signals with fixed parameter styles, its anti-interference performance will not meet the requirements. This study proposes a radar intermittent sampling forwarding interference suppression method based on the Fast-scnn semantic segmentation network. In the preprocessing stage, the algorithm uses the Short-Time Fourier Transform (STFT) to separate the interference signal and the target echo in the time-frequency two-dimensional image, and then uses the Fast-scnn semantic segmentation network to generate the interference area mask, and finally achieves interference suppression, improving the anti-interference and generalization capabilities of the radar system. Summary of the invention

[0004] The purpose of the present invention is to propose a radar intermittent sampling and repetitive forwarding interference suppression method based on Fast-scnn semantic segmentation network based on STFT time-frequency analysis for intermittent sampling and repetitive forwarding interference mode based on radio frequency storage and forwarding.

[0005] The technical solution of the present invention is:

[0006] A method for suppressing intermittent sampling and forwarding interference based on Fast-SCNN comprises the following steps:

[0007] S1. Get sample data:

[0008] Obtain interference-free target echo signals and intermittent sampling forwarding interference signals, which are divided into a training set and a verification set. Assume that the training set includes N1 interference-free target echo signals and intermittent sampling forwarding interference signals, respectively, and superimpose the two to obtain the superimposed training set signal; the verification set includes N2 interference-free target echo signals and intermittent sampling forwarding interference signals, respectively, and superimpose the verification set signals to obtain the superimposed verification set signals;

[0009] S2. Preprocess the sample data:

[0010] At the same time, the interference signal and the superimposed signal are preprocessed. The obtained sample data is de-linearly modulated, and then the de-linearly modulated signal is converted into a two-dimensional time-frequency feature image S(t,f) using short-time Fourier transform, where the time-frequency image obtained by intermittent sampling and forwarding the interference signal is used as the label set, the interference area label category is 1, and the label category of the remaining areas is 0; the time-frequency image obtained by the superimposed target echo signal is used as the data set, and the data set and the label set constitute the training data set; similarly, the data set obtained by the verification set is used as the verification data set;

[0011] S3. Build a semantic segmentation neural network based on Fast-SCNN:

[0012] The semantic segmentation neural network includes a learning downsampling module, a global feature extractor, a feature fusion module and a standard classifier; the learning downsampling module downsamples the input data, including a standard convolution layer and two distribution shift convolution layers; the global feature extractor extracts features from the data, extracts features through a continuous eight-layer bottleneck residual block, and then aggregates the context information of different regions through a pyramid pooling module; the feature fusion module upsamples the extracted features, adjusts the number of channels of the features of the two branches through a convolution operation, and then adds the branches to achieve feature fusion; the standard classifier contains two distribution shift convolution layers and a convolution layer with a convolution kernel size of 1x1, and the softmax layer is used as the final output layer in the training stage. In the inference stage, the argmax operation is used instead of the softmax to obtain the pixel-level interference area mask T(t,f);

[0013] The loss function uses cross entropy loss:

[0014]

[0015] Among them, y i represents the label of each pixel, p i Indicates the prediction result of each pixel, the pixel belonging to the interference area is 1, and the rest is 0;

[0016] S4, using the training data set to train the constructed semantic segmentation neural network, and adjusting the parameters through the verification data set to obtain the trained semantic segmentation neural network;

[0017] S5. Use the trained semantic segmentation neural network to suppress the interference of the echo signal superimposed with interference and realize echo signal reconstruction. Specifically, use the interference signal coordinates marked in the interference area mask T(t,f) to set the interference area of ​​the time-frequency two-dimensional image S(t,f) to 0 to obtain the suppressed target echo time-frequency characteristics. After using the inverse short-time Fourier transform to restore the time-frequency domain feature map to the time domain waveform, the time domain signal is subjected to linear frequency modulation restoration processing to realize the reconstruction of the target echo.

[0018] Furthermore, the parameters of the intermittent sampling forwarding interference signal are set to different interference forwarding intervals and forwarding segment durations, and the styles of the interference signal include direct forwarding, repeated forwarding and cyclic forwarding.

[0019] The beneficial effects of the present invention are as follows: the present invention adopts the method of time-frequency analysis and deep learning semantic segmentation, takes the received signal of interference and target superposition as input data, and takes the regional distribution of interference signal as label data. After preprocessing, it is converted into two-dimensional features in the time-frequency domain, which is convenient for extracting and suppressing interference. Subsequently, the Fast-scnn semantic segmentation network is used to more accurately predict the interference area in the time-frequency two-dimensional matrix. Combined with the signal reconstruction method, the target echo signal is further restored, thereby better achieving the suppression effect of intermittent sampling forwarding interference. DETAILED DESCRIPTION

[0020] Figure 1 Flowchart of radar intermittent sampling and forwarding interference suppression method based on Fast-scnn semantic segmentation network.

[0021] Figure 2 A Fast-scnn based interference mask generation network.

[0022] Figure 3 Comparison of pulse compression effects before and after interference suppression.

[0024] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0025] like Figure 1 As shown, the present invention mainly comprises the following steps:

[0026] S1. Get sample data:

[0027] The radar transmission signal is simulated as a linear frequency modulation signal using MATLAB, and the intermittent sampling forwarding interference signal is simulated based on the transmission signal. Among them, the parameters of the interference signal are set to different interference forwarding intervals and forwarding segment durations. The styles of the interference signal include direct forwarding, repeated forwarding, and cyclic forwarding. Interference signals and target echoes at different positions are randomly generated, Gaussian white noise is added, and superimposed as received signal sample data. The training set contains N1 samples, where the target echo and interference signal are recorded as:

[0028] train_sig={train_sig i |i=1,2,...,N1}

[0029] train-jam={train-jam i |i=1,2,...,N1}

[0030] The received signal after superposition is recorded as:

[0031] train-r = {train-r i |i=1,2,...,N1}

[0032] The validation set contains N2 samples, where the target echo and interference signal are recorded as:

[0033] val_sig={val_sig i |i=1,2,...,N2}

[0034] val-jam={val-jam i |i=1,2,...,N2}

[0035] The received signal after superposition is recorded as:

[0036] val-r={val-r i |i=1,2,...,N2}

[0037] The set used to test the interference suppression effect of the algorithm contains N3 samples, recorded as:

[0038] test-r={test-r i |i=1,2,...,N3}

[0039] S2. Preprocess the sample data:

[0040] At the same time, the interference signal and the superimposed signal are preprocessed. For the sample data generated by simulation, the linear frequency modulation is firstly processed to eliminate the frequency ramp characteristics of the signal, and then the de-linear frequency modulation signal is converted into a time-frequency two-dimensional feature image S(t,f) using short-time Fourier transform.

[0041] A label set is created based on the time-frequency images of the interference dataset, where the label category of the interference area is 1 and the label category of the remaining areas is 0.

[0042] S3. Construct a Fast-scnn semantic segmentation neural network to generate a mask of the interference signal area. The specific structure is as follows Figure 2 shown.

[0043] The overall network architecture of Fast-SCNN consists of four parts: learning downsampling module, global feature extractor, feature fusion module and standard classifier. First, three layers of convolution are used in the learning downsampling module, of which the first layer is a standard convolution layer and the other two layers are distribution shift convolution layers; second, the global feature extractor module contains eight bottleneck residual blocks and combines the pyramid pooling module (PPM) to aggregate contextual information of different regions; then, the feature fusion module upsamples the extracted features, adjusts the number of channels of the features of the two branches through convolution operations, and then adds the branches to achieve feature fusion; finally, the classifier contains two distribution shift convolution layers and a convolution layer with a convolution kernel size of 1x1. The softmax layer is used as the final output layer in the training stage, and the argmax operation is used instead of softmax in the inference stage to obtain the pixel-level interference region mask T(t,f).

[0044] The loss function uses cross entropy loss as follows:

[0045]

[0046] where y i represents the label of each pixel, p i Represents the prediction result of each pixel, which is 1 if it belongs to the interference area and 0 otherwise.

[0047] S4. According to the validation set samples, the Adam method is used to complete parameter training and learning rate optimization of the above Fast-scnn network to obtain the final semantic segmentation model.

[0048] S5. Post-process the output of the neural network to achieve echo signal reconstruction:

[0049] The echo signal reconstruction includes the following steps: first, the interference area of ​​the time-frequency two-dimensional image S(t,f) is set to 0 using the interference signal coordinates marked in the interference area mask T(t,f) to obtain the suppressed target echo time-frequency characteristics; then, the time-frequency domain feature map is restored to the time domain waveform using the inverse short-time Fourier transform; finally, the time domain signal is subjected to linear frequency modulation restoration processing to achieve the reconstruction of the target echo.

[0050] The specific structure of the semantic segmentation neural network based on Fast-scnn is that it contains a total of 48 layers of network structure except the input layer, namely: a learning downsampling module composed of 3 layers of network, a global feature extractor composed of 37 layers of network, a feature fusion module composed of 4 layers of network, and the last 4 layers are output classification layers.

[0051] The structure of the learning downsampling module is:

[0052] A two-dimensional convolution layer (Layer1) with 32 channels, a convolution kernel of 3*3, and a stride of 2; the activation function is Relu.

[0053] Two distribution shift convolution layers (Layer2-3) with 128 and 64 channels respectively, 3*3 convolution kernel and 2 stride.

[0054] The structure of the global feature extractor is:

[0055] Eight bottleneck residual blocks (Layer4-35), each of which contains the following four layers: a convolution layer with a 1*1 kernel, 32 channels, and Relu activation function; a depthwise separable convolution layer with a 3*3 kernel, 32 channels; a convolution layer with a 1*1 kernel, 64 channels; and a batch normalization layer. The stride of the depthwise separable convolution layer in the first and fourth bottleneck residual blocks is 2, and the stride of all other layers is 1.

[0056] A pyramid pooling module (Layer36-40), which contains 4 pooling layers and 1 output convolution layer. The pooling sizes are 1*1, 2*2, 3*3 and 6*6 respectively; the convolution kernel of the output convolution layer is 1*1, the stride is 1, and the number of channels is 128.

[0057] The structure of the feature fusion module is:

[0058] The pyramid pooling module is connected to a 4*4 upsampling layer (Layer41); a depthwise separable convolution layer (Layer42) with 128 channels and a 3*3 convolution kernel; and a two-dimensional convolution layer (Layer43) with 128 channels and a 3*3 convolution kernel.

[0059] The learning downsampling module is connected to a two-dimensional convolutional layer (Layer44) with 128 channels and a convolution kernel of 3*3.

[0060] The outputs of the two modules are superimposed as fusion features.

[0061] The structure of the classifier module is:

[0062] Two distribution shift convolution layers (Layer45-46), with 128 channels, 3*3 convolution kernel, and stride 1.

[0063] A two-dimensional convolution layer (Layer47) with 1 channel, 3*3 convolution kernel and 1 stride.

[0064] An 8*8 upsampling layer (Layer48) connected to the argmax layer as the final output layer.

[0065] Example

[0066] Step 1: Use MATLAB to simulate sample data. The radar transmission wave selects a linear frequency modulation signal with a bandwidth of 40*10 6 Hz, pulse width is 20*10 -6 s, the modulation slope K is 2*10 12 The parameters of intermittent sampling forwarding interference are set to three groups: the first group sampling duration is 2*10 -6 s, the forwarding interval is 8*10 -6 s; the second set of sampling duration is 1.5*10 -6 s, the forwarding interval is 6*10 -6 s; the third group of sampling duration is 1*10 -6 s, the forwarding interval is 5*10 -6 s. Each set of interference parameters randomly generates 200 samples, randomly generates target echoes and interference signals at different positions, and sets the interference-signal ratio to 5-20dB, with a step distribution of 5dB intervals, and the signal-to-noise ratio of additive Gaussian white noise to 10dB. In the simulation signal, the signal with superimposed interference is used to generate the data set of the neural network, and the interference signal is used to generate the corresponding label set.

[0067] Step 2: De-linear frequency modulation is performed on the simulated signal, that is, the simulated signal r(t) is mixed with the local transmission wave reference signal to eliminate the linear frequency modulation component. The local reference signal can be expressed as:

[0068]

[0069] The signal after de-linear frequency modulation is:

[0070] s de-chirp (t) = r(t)*s LO (t)

[0071] After the processing is completed, time-frequency analysis is performed. A 400-point Hamming window is used to slide point by point, and the signal is short-time Fourier transformed to obtain the time-frequency diagram S(t,f).

[0072] S(t,f)=STFT[s de-chirp (t)]

[0073] Step 4: Convert the time-frequency two-dimensional feature matrix obtained in the previous step into dB format according to the amplitude, and normalize the value to 255. The time-frequency map of the target echo with interference is used as the input data of the neural network, and the time-frequency map containing only the interference area is used as the label data. Construct a 48-layer Fast-scnn semantic segmentation network, in which the last layer is a softmax classifier. Train the model according to the training samples, and obtain the final deep interference suppression network model through iteration.

[0074] Step 5: Reconstruct the target echo signal based on the interference area distribution matrix output by the above network. First, set the pixels in the interference area in S(t,f) generated by the superimposed signal to 0 to obtain the time-frequency feature map S after interference suppression. tar (t,f):

[0075] S tar (t,f)=S(t,f)*[1-T(t,f)]

[0076] Then, the inverse short-time Fourier transform (ISTFT) is used to restore the time-frequency feature map to a time domain waveform:

[0077] s tar (t) = ISTFT[S tar (t,f)]

[0078] Among them, the parameters of the inverse transformation need to be consistent with the forward transformation, that is, the window function is selected as a Hamming window with a length of 200 for point-by-point sliding, and the OLA method is used for signal recovery. Finally, the recovered signal is directly mixed with the linear frequency modulation transmission wave reference to restore the linear frequency modulation of the signal and achieve the target reconstruction. The reference signal can be expressed as:

[0079]

[0080] The target echo reconstruction signal after restoration is:

[0081] s rec (t) = s tar (t)*s LFM (t)

[0082] The above method is used to verify the effect of the radar intermittent sampling forwarding interference suppression method based on the Fast-scnn semantic segmentation network of the present invention. The target echo signal and intermittent sampling forwarding interference are simulated, and the sample data are randomly allocated in a ratio of 5:2:3 to obtain the training set, verification set and test set. By verifying the samples in the test set, Figure 3 It can be seen that before interference suppression, the target echo is suppressed by the intermittent sampling forwarding interference. After model processing, the target echo is completely reconstructed, and the interference-to-signal ratio is improved by about 25dB.

Claims

1. A method for suppressing intermittent sampling and forwarding interference based on Fast-SCNN, characterized in that: The following steps are involved: S1. Get sample data: Obtain interference-free target echo signals and intermittent sampling forwarding interference signals, which are divided into a training set and a verification set. The training set includes N1 interference-free target echo signals and intermittent sampling forwarding interference signals, respectively. The training set signals are superimposed to obtain a superimposed training set signal. The verification set includes N2 interference-free target echo signals and intermittent sampling forwarding interference signals, respectively. The verification set signals are superimposed to obtain a superimposed verification set signal. S2. Preprocess the sample data: At the same time, the interference signal and the superimposed signal are preprocessed, specifically, the obtained sample data is de-linearly modulated, and then the de-linearly modulated signal is converted into a time-frequency two-dimensional feature image S(t,f) using short-time Fourier transform, wherein the time-frequency two-dimensional feature image obtained through the training set is divided into a data set and a label set, wherein the time-frequency image obtained by intermittent sampling and forwarding the interference signal is used as the label set, the interference area label category is 1, and the label category of the remaining areas is 0; the time-frequency image obtained by the superimposed target echo signal is used as the data set, and the data set and the label set constitute the training data set; similarly, the data set obtained through the verification set is used as the verification data set; S3. Build a semantic segmentation neural network based on Fast-SCNN: The semantic segmentation neural network includes a learning downsampling module, a global feature extractor, a feature fusion module and a standard classifier; the learning downsampling module downsamples the input data, including a standard convolution layer and two distribution shift convolution layers; the global feature extractor extracts features from the data, extracts features through a continuous eight-layer bottleneck residual block, and then aggregates the context information of different regions through a pyramid pooling module; the feature fusion module upsamples the extracted features, adjusts the number of channels of the features of the two branches through a convolution operation, and then adds the branches to achieve feature fusion; the standard classifier contains two distribution shift convolution layers and a convolution layer with a convolution kernel size of 1x1, and the softmax layer is used as the final output layer in the training stage. In the inference stage, the argmax operation is used instead of the softmax to obtain the pixel-level interference area mask T(t,f); The loss function uses cross entropy loss: Among them, y i represents the label of each pixel, p i Indicates the prediction result of each pixel, the pixel belonging to the interference area is 1, and the rest is 0; S4, using the training data set to train the constructed semantic segmentation neural network, and adjusting the parameters through the verification data set to obtain the trained semantic segmentation neural network; S5. Use the trained semantic segmentation neural network to suppress the interference of the echo signal superimposed with interference and realize echo signal reconstruction. Specifically, use the interference signal coordinates marked in the interference area mask T(t,f) to set the interference area of ​​the time-frequency two-dimensional image S(t,f) to 0 to obtain the suppressed target echo time-frequency characteristics. After using the inverse short-time Fourier transform to restore the time-frequency domain feature map to the time domain waveform, the time domain signal is subjected to linear frequency modulation restoration processing to realize the reconstruction of the target echo.

2. The method for suppressing interference of intermittent sampling and forwarding based on Fast-SCNN according to claim 1, characterized in that: The parameters of the intermittent sampling forwarding interference signal are set to different interference forwarding intervals and forwarding segment durations. The styles of the interference signal include direct forwarding, repeated forwarding and cyclic forwarding.

Images