CN116540189A - Radar main lobe forwarding interference resisting method based on reversible residual error network - Google Patents

Abstract

The invention belongs to the technical field of radar anti-interference, and particularly relates to an anti-radar main lobe forwarding interference method based on a reversible residual error network. The invention provides a set of radar main lobe forwarding interference suppression deep learning framework Inv-NMF net suitable for a multi-target multi-interference environment based on a non-Negative Matrix Factorization (NMF) algorithm and a reversible residual neural network. The method can decompose the interference signal by utilizing radar emission signal information, and characterize the target signal information and the interference signal information at different positions of a coefficient matrix. And then analyzing the coefficient matrix by using a reversible residual error network to generate a Mask matrix, and separating the interference signal component and the target signal component. The invention can ensure that the target signal strength is not influenced to the greatest extent because of the information transmission non-destructive property of the reversible network. The method is not only not affected by the number of radar targets, but also has better adaptability to different radar emission waveforms.

Description

A Reversible Residual Network Based Anti-jamming Method of Radar Main Lobe Relay

technical field

The invention belongs to the technical field of radar anti-interference, and in particular relates to an anti-radar main lobe forwarding interference method based on a reversible residual network.

Background technique

The radar system is one of the core detection methods in modern communication. By transmitting and receiving electromagnetic wave signals, it can monitor the target area in real time around the clock. At present, with the continuous development of digital radio frequency memory (Digital Radio Frequency Memory, DRFM) technology, the radar main lobe forwarding interference is highly correlated with the radar transmitted signal and enters the radar receiver from the antenna main lobe. The radar anti-jamming strategy fails, and the target detection performance of the radar drops sharply. Therefore, research on suppression methods for radar main lobe forwarding interference is an important research direction to ensure radar detection capability, and has important research significance. Most of the traditional signal processing methods are limited to the suppression of forwarding interference under a single target, and the few algorithms that can suppress the forwarding interference under multiple targets are difficult to have a good interference suppression effect. Therefore, so far there is no suitable anti-radar main lobe forwarding jamming algorithm that can well solve the problem of radar jamming under multiple targets.

Contents of the invention

In view of the above problems, the present invention starts from the time-domain signal itself, uses the radar transmission signal information to reduce the signal separation complexity, refers to the non-negative matrix decomposition method commonly used in the blind source separation theory, uses the radar transmission signal as a dictionary matrix, and obtains the radar return The eigenmatrix of the wave. The characteristic matrix is used to replace the time-frequency map matrix obtained by short-time Fourier transform for subsequent interference signal separation. And for the problem of missing information in the multi-target signal recovery, try to use the reversible residual network to extract the signal features to ensure the approximate lossless recovery of the signal.

Based on a non-negative matrix factorization (NMF) algorithm and a reversible residual neural network, the present invention proposes a deep learning framework Inv-NMF net suitable for radar main lobe forwarding interference suppression in a multi-target and multi-interference environment. It can decompose the interference signal by using the radar transmission signal information, and represent the target signal information and the interference signal information in different positions of the coefficient matrix. Then the reversible residual network is used to analyze the coefficient matrix to generate a Mask mask matrix to separate the interference signal component and the target signal component. Due to the lossless information transmission of the reversible network, the present invention can ensure that the target signal strength is not affected to the greatest extent. This method is not only not affected by the number of radar targets, but also has good adaptability to different radar transmission waveforms.

Technical scheme of the present invention is:

An anti-radar main lobe forwarding interference method based on reversible residual network, comprising the following steps:

S1. Using the linear frequency modulation signal as the radar transmission signal, the interference signal is obtained through the radar main lobe forwarding interference simulation, and then the echo signal Signal _data with interference added, the emission signal Signal, the interference signal label Jam _target and the non-interference signal label Signal are obtained. _target :

Signal _data ＝{x ⁰ _i |i＝1,2,…,L}∈C ^L

Signal＝{x ¹ _i |i＝1,2,…,L}∈C ^L

Jam _target ＝{x ² _i |i＝1,2,…,L}∈C ^L

Signal _target ＝{x ³ _i |i＝1,2,…,L}∈C ^L

Among them, L is the signal length;

S2. Perform linear normalization processing on Signal _data , Signal and Signal _target respectively, map the signal strength to (0, 1), and use the maximum and minimum values of Signal _data to normalize the Jam _target to ensure that both are normalized The ratio is consistent, and then use the sliding window method to segment each M-long time slice signal in Signal _data and Signal, and then splicing in the next dimension to form the corresponding M×N two-dimensional matrix X∈R ^M×N and S∈R ^M×N , L=M×N;

S3. Set S as dictionary matrix W, perform NMF decomposition on matrix X, and obtain coefficient matrix H corresponding to signal sample matrix X. The specific method is:

Randomly initialize the coefficient matrix H∈R ^N×N to ensure that H _ij ≥ 0|i=1,...,N; j=1,...,N;

Update the H matrix parameters and calculate the error e:

If the error e satisfies the set condition, then output the coefficient matrix H, otherwise continue to update iteratively until the maximum number of iterations is reached;

S4. Construct an interference suppression network based on a reversible residual network, including a reversible residual network layer with the same three-layer structure. The forward propagation input of the network is the normalized coefficient matrix H obtained by NMF decomposition, and the expected output is to separate The interference signal mask matrix Mask; each layer network is composed of a group of symmetrical Unet networks, specifically: define the inputs of both ends of each layer network as X ₁ and X ₂ , and use the coefficient matrix H as the X ₁ of the first layer network and X ₂ input, X ₁ extracts the interference signal information in the original signal through the first Unet network in the first layer network, the first Unet network includes five CNN layers, of which the first layer and the fifth layer, the second layer The size of the input matrix is the same as that of the fourth layer, and they use layer-skip connections to interact with each other. The output of the previous layer and the output of this layer are spliced in the channel dimension, and they are jointly input to the next layer, specifically the first layer. The input is X ₁ , the input of the second layer is the output of the first layer, the input of the third layer is the input of the second layer, the input of the fourth layer is the output of the second layer and the third layer, and the input of the fifth layer It is the input of the first layer and the fourth layer; the output of the fifth layer is combined with X ₂ to obtain the target echo information Y ₂ that initially filters out the interference, and Y ₂ is extracted by the second Unet network in the first layer network. Pure target echo information, the second Unet network includes five CNN layers, the same as the first Unet network, the output of the fifth layer of the second Unet network is combined with _X1 to obtain the output _Y1 ; the first layer network The outputs Y ₁ and Y ₂ are respectively used as the input X ₂ and X ₁ of the next layer network, and the output of the last layer network is passed through the Sigmoid function to obtain the mask matrix Mask of the interference signal;

Multiply the output interference matrix Mask and the coefficient matrix H to obtain the coefficient matrix of the interference signal, and then multiply the coefficient matrix of the interference signal with the dictionary matrix W to obtain the interference signal J;

Subtract the interference signal J from the Signal _data to initially restore the interference-free target signal, and then pass the initially restored interference-free target signal through the convolutional neural network to obtain the final interference-free target signal y, and the convolutional neural network is a group The U-net model structure is similar to the first Unet network structure in S4. Its downsampling times are increased to 5 times, and the construction of each downsampling layer consists of a residual convolution layer, an activation function layer, and a batch normalization layer. Downsampling is achieved by increasing the step size of the convolution kernel, the sampling factor is 2 times, and the number of channels in each layer increases layer by layer. The upsampling process is the same, but the number of channels is reduced layer by layer, and finally a pure and interference-free signal is restored.

S5. Training the constructed interference suppression network, the loss function adopted includes:

L1 error loss for interfering signals:

Among them, J represents the interference signal recovered by using the mask matrix, and N is the length of the interference signal;

L1 error loss for the target signal:

Among them, y represents the non-interference target signal finally output by the model, and N is the length of the target signal;

Hilbert transform loss:

Among them, Hilbert() is the Hilbert transform function;

When the network loss basically converges, the final deep network model is obtained;

S6. Using the obtained deep network model to process the radar received signal to realize anti-interference from radar main lobe forwarding.

Furthermore, the jamming signals obtained through the simulation of radar main lobe forwarding jamming in S1 include intermittent sampling forwarding jamming, smart noise jamming, spectrum dispersion jamming and comb spectrum jamming. The working method of intermittent sampling and forwarding jamming is to first sample a segment of the intercepted radar signal and forward it, and then sample and forward the next signal segment until the falling edge of the radar transmission signal is detected. According to different intermittent sampling methods, the intermittent sampling forwarding interference can be subdivided into three types: direct forwarding interference J _ISDJ , repeated forwarding interference J _ISRJ and cyclic forwarding interference J _ISIJ . The three interference signals are represented as:

In the formula, N is the number of slices, T _s is the slice width, M is the number of times each interference slice is forwarded, T _u =(M+1)T _s is the interception time interval between two adjacent interference slices. a _m T _s is the interception time of the mth slice, b _n T _s is the corresponding delay when the slice is forwarded for the nth time. Specifically:

Smart noise jamming works by storing intercepted radar transmissions in digital registers using digital radio frequency memory (DRFM). Then, according to the form of the radar signal, the noise unit is controlled to generate a noise signal with matching length and type. Finally, the noise signal and the radar signal are multiplied or convoluted on the signal synthesizer, and a smart noise jamming signal with better effect can be produced. According to different signal modulation methods, the interference can be divided into noise convolution interference J _SCN and noise product interference J _SPN , and their mathematical expressions are:

J _SCN (t)=s(tt)×n(t)

In the formula, the radar signal received by DRFM is s(t), the output after delay is s(t-t), and the narrow-band Gaussian noise is n(t).

The working method of spectrum dispersive jamming (SMSP) is: when the jammer intercepts the radar transmission signal, it first digitally processes the intercepted signal and stores it in the DRFM jammer. Then, the shift register is used to obtain N sub-signals whose modulation slope is N times that of the transmitted signal, and they are serially input into the digital-to-analog converter, and the SMSP interference signal can be obtained after mixing.

The first subsignal can be expressed as:

Wherein, k is the slope of the sub-signal, k _j is the FM slope of the sub-signal, k _j =Nk, after J _SMSP1 is copied N-1 times, they are combined to obtain SMSP interference:

Comb spectrum jamming works by multiplying the comb spectrum signal with the intercepted radar transmission and then modulating it to generate a jamming signal.

The expression of the comb spectrum signal is:

Among them, f _i corresponds to the frequency point where each comb tooth appears, and a _i is the amplitude at the i-th frequency point.

The mathematical model of radar comb interference is:

The beneficial effects of the present invention are: the present invention designs a multi-target interference suppression network model based on NMF and a reversible residual network for complex detection scenarios with multiple targets and multiple interference sources. First, the model uses the radar transmission signal as prior information to generate a dictionary matrix, calculates the coefficient matrix of the received signal through NMF decomposition, and maps the complex one-dimensional signal into a two-dimensional sparse matrix, which greatly reduces the complexity of network learning. Then the model separates the decomposed coefficient matrix through the reversible residual network, which can not only effectively suppress the interference signal, but also ensure the approximately lossless transmission of the target information. The present invention can achieve a good suppression effect under the condition of multiple targets and multiple interference sources, greatly improves the target detection capability of subsequent radars, and ensures accurate detection of multiple targets by the radar.

Description of drawings

Fig. 1 is a schematic diagram of the overall processing flow of the present invention.

Fig. 2 is a schematic diagram of the structure of the interference suppression network based on the reversible residual network in the present invention.

Fig. 3 is a schematic diagram of the structure of the reversible residual network in the present invention.

FIG. 4 is a schematic diagram of the suppression effect of the present invention on three kinds of ISRJ interference.

Fig. 5 is a schematic diagram of the suppression effect of the present invention on SCN, SJN, and Comb interference.

Fig. 6 is a schematic diagram of the suppression effect of the model of the present invention on SMSP interference.

Detailed ways

Below in conjunction with accompanying drawing and embodiment the technical scheme of the present invention is described in further detail:

Firstly, the present invention simulates the radar jamming signal based on the working mechanism of the main lobe forwarding the jamming signal, and generates the jamming target signal, the jamming signal and the unjammed target signal as signal samples and label samples. Then, the signal is normalized and preprocessed, the signal is mapped to between 0 and 1, and then the two-dimensional matrix of the signal is obtained through sliding window processing. Secondly, the Inv-NMF net model is designed. Its internal structure is mainly composed of three parts, see Figure 1: The first part is NMF decomposition. Using the radar transmitted signal s(t)∈R ^1×L as the dictionary matrix W, we can calculate the coefficient matrix H of the radar received signal x(t)∈R ^1×L through the multiplication update rule. The second part is the forward propagation part of the reversible network. A reversible residual network is used to extract the signal features in the coefficient matrix H, and then a mask matrix Mask is generated. The coefficient matrix of the interference signal can be obtained by multiplying the Mask matrix with the coefficient matrix H, and then the interference signal y in the radar received signal can be generated by multiplying the coefficient matrix with the dictionary matrix. Finally, subtract the part of the interference signal from the received signal, and we can obtain the information of the target signal. The third part is the recovery part of the target signal. A simple CNN module is utilized to accurately recover the target signal. Finally, the built network model is trained. After the model converges, the model is saved, and the randomly generated radar main lobe is used to forward the interference signal for testing, and the interference suppression effect of the statistical model is calculated.

Example

The specific steps in this example are as follows:

S1: Establish an interference signal model: first, simulate the radar interference signal according to the parameters in Table 1

Table 1 Simulation parameters

Finally, the output of the simulated radar system is the echo signal Signal _data with added interference, the transmitted signal Signal, the interference signal Jam _target and the target signal Signal _target without interference.

Signal _data ＝{x ⁰ _i |i＝1,2,…,20000}∈C ²⁰⁰⁰⁰

Signal＝{x ¹ _i |i＝1,2,…,20000}∈C ²⁰⁰⁰⁰

Jam _target ={x ² _i |i=1,2,…,20000}∈C ²⁰⁰⁰⁰

Signal _target ={x ³ _i |i=1,2,…,20000}∈C ²⁰⁰⁰⁰

S2: Linearly normalize the signal sample Signal _data , the transmitted signal Signal and the sample label signal Signal _target respectively, map the signal strength to (0, 1), and use the maximum and minimum values of the Signal _data to normalize the Jam _target , to ensure that the normalization ratios of the two are consistent. Then use the sliding window method to segment each 200-time-slice signal in Signal _data and Signal, and then splicing in the next dimension to form a corresponding 200×100 two-dimensional matrix X∈R ^{200 ×100} and S∈ R ^200×100 .

S3: Set S as the dictionary matrix W, perform NMF decomposition on the matrix X, and obtain the coefficient matrix H corresponding to the signal sample matrix X. The specific process is as follows:

First, randomly initialize the coefficient matrix H∈R ^100×100 to ensure that H _ij ≥ 0|i=1,...,100; j=1,...,100

Then, update the H matrix parameters and calculate the error e:

Finally, if the error e<1e-5, then output the coefficient matrix H, otherwise continue to update iteratively until the maximum number of iterations is 500.

S4: Construct the Inv-U net interference suppression network, as shown in Figure 2. Its forward propagation input is the normalized coefficient matrix H obtained by NMF decomposition, and the expected output is the separated interference signal mask matrix Mask. First, I input the coefficient matrix H to the two ends X ₁ and X ₂ of Inv-U net respectively. The network consists of three layers, and the F and G functions of each layer are composed of a set of symmetrical Unet networks. Among them, the former Unet module F(X ₁ ) is used to extract the interference signal information in the original signal, and the target echo information Y ₂ which can preliminarily filter out the interference can be obtained by combining X ₂ with F(X ₁ ). The latter Unet module G(Y ₂ ) is used to extract more pure target echo information on the basis of Y ₂ , and combine it with X ₁ to obtain the output Y ₁ , and then pass the Sigmoid function to obtain the interference signal Mask matrix Mask.

S41: Each Unet module consists of five CNN layers, as shown in Figure 3, in which the input matrices of the first layer and the fifth layer, the second layer and the fourth layer have the same size, and they use skipping connections for interaction. The output of the previous layer and the output of this layer are spliced in the channel dimension and input to the next layer together. The specific network parameters are shown in Table 2:

Table 2Unet module parameters

Finally, the coefficient matrix of the interference signal can be obtained by dot-multiplying the output interference matrix Mask and the coefficient matrix H. Then multiply it with the dictionary matrix W to get the interference signal J.

S5: Subtracting the interference signal J from the signal samples can initially restore the target signal without interference. Then pass it through a set of CNN model structures to achieve finer restoration and obtain the final interference-free target signal y. Its model structure is similar to the F-function structure described in S4. The number of downsampling is 4 times, and the construction of each downsampling layer is composed of residual convolution layer, activation function layer and batch normalization layer, and the number of channels increases layer by layer. The upsampling process is the same, but the number of channels decreases layer by layer. The specific convolutional layer parameters are shown in Table 3:

Table 3 CNN module parameters

S6: Model output signal, calculate loss value with interference signal Jam _target and non-interference target signal Signal _target . The first loss function Loss ₁ is the L1 (average absolute error) error loss of the interference signal:

In the formula, J represents the jamming signal recovered by using the mask matrix, Jam _target represents the real jamming signal label, and N is the length of the jamming signal.

The second loss function Loss ₂ is the L1 error loss of the target signal:

In the formula, y represents the non-interference target signal finally output by the model, Signal _target represents the real target signal label, and N is the length of the target signal. The reason why L1 loss is used to measure the quality of signal recovery is that it has a stable gradient and will not cause the problem of gradient explosion. In this way, some outliers can be ignored well, and small differences can also be penalized greatly.

The third loss function Loss ₃ is the Hilbert transformation loss:

Where Hilbert() is the Hilbert transform function. In the training setting, Batch_size is selected as 16, the learning rate is adjusted within the range of 0.0005, and the optimizer is selected as Adam. Save the model after training is complete.

Finally, under various main-lobe forwarding interference modes, using randomly generated interference signals with different sampling duty ratios, the main-lobe interference suppression is performed through the two-stage deep network, and the target interference-to-signal ratio improvement before and after interference suppression is counted. As shown in Figure 4-6.

The interference signal-to-noise ratio improvement factor JSR-IF measures the interference suppression effect in a multi-target environment:

Where a _star is the minimum target amplitude among multiple targets after echo pulse compression; and a _jam is the maximum amplitude of jamming targets after pulse compression.

The interference-to-signal ratio improvement factor (JSR-IF) can be expressed as:

JSR-IF＝JSR _unfiltered -JSR _filtered

Among them, JSR _filtered is the JSR of the pulse pressure result after interference suppression; and JSR _unfiltered is the JSR of the pulse pressure result of the original signal without interference suppression processing.

Claims (2)

1. an anti-radar mainlobe forwarding interference method based on reversible residual network, is characterized in that, comprises the following steps: S1. Using the linear frequency modulation signal as the radar transmission signal, the interference signal is obtained through the radar main lobe forwarding interference simulation, and then the echo signal Signal _data with interference added, the emission signal Signal, the interference signal label Jam _target and the non-interference signal label Signal are obtained. _target : Signal _data ＝{x ⁰ _i |i＝1,2,…,L}∈C ^L Signal＝{x ¹ _i |i＝1,2,…,L}∈C ^L Jam _target ＝{x ² _i |i＝1,2,…,L}∈C ^L Signal _target ＝{x ³ _i |i＝1,2,…,L}∈C ^L Among them, L is the signal length; S2. Perform linear normalization processing on Signal _data , Signal and Signal _target respectively, map the signal strength to (0, 1), and use the maximum and minimum values of Signal _data to normalize the Jam _target to ensure that both are normalized The ratio is consistent, and then use the sliding window method to segment each M-long time slice signal in Signal _data and Signal, and then splicing in the next dimension to form the corresponding M×N two-dimensional matrix X∈R ^M×N and S∈R ^M×N , L=M×N; S3. Set S as dictionary matrix W, perform NMF decomposition on matrix X, and obtain coefficient matrix H corresponding to signal sample matrix X. The specific method is: Randomly initialize the coefficient matrix H∈R ^N×N to ensure that H _ij ≥ 0|i=1,...,N; j=1,...,N; Update the H matrix parameters and calculate the error e: If the error e satisfies the set condition, then output the coefficient matrix H, otherwise continue to update iteratively until the maximum number of iterations is reached; S4. Construct an interference suppression network based on a reversible residual network. The interference suppression network includes a reversible residual network layer with the same three-layer structure. The forward propagation input of the network is the normalized coefficient matrix H obtained by NMF decomposition, and the expected output is is the separated interference signal mask matrix Mask; each layer network is composed of a group of symmetrical Unet networks, specifically: define the inputs of both ends of each layer network as X ₁ and X ₂ , and use the coefficient matrix H as the first layer network X ₁ and X ₂ are input, and X ₁ extracts the interference signal information in the original signal through the first Unet network in the first layer network. The first Unet network includes five CNN layers, of which the first layer and the fifth layer, The size of the input matrix of the second layer and the fourth layer is the same, and they use a layer-skip connection to interact with each other. The output of the previous layer and the output of this layer are spliced in the channel dimension, and they are jointly input to the next layer. Specifically, the first layer The input of one layer is X ₁ , the input of the second layer is the output of the first layer, the input of the third layer is the input of the second layer, the input of the fourth layer is the output of the second layer and the third layer, and the fifth layer The input of the layer is the input of the first layer and the fourth layer; the output of the fifth layer is combined with X ₂ to obtain the target echo information Y ₂ that initially filters out the interference, and Y ₂ passes through the second Unet network in the first layer network Extract more pure target echo information, the second Unet network includes five CNN layers, same as the first Unet network, the output of the fifth layer of the second Unet network is combined with _X1 to obtain output _Y1 ; The output Y ₁ and Y ₂ of one layer of network are respectively used as the input X ₂ and X ₁ of the next layer of network, and the output of the last layer of network is passed through the Sigmoid function to obtain the mask matrix Mask of the interference signal; Multiply the output interference matrix Mask and the coefficient matrix H to obtain the coefficient matrix of the interference signal, and then multiply the coefficient matrix of the interference signal with the dictionary matrix W to obtain the interference signal J; Subtract the interference signal J from the Signal _data Signal _data to initially restore the interference-free target signal, and then pass the initially restored interference-free target signal through the convolutional neural network to obtain the final interference-free target signal y; S5. Training the constructed interference suppression network, the loss function adopted includes: L1 error loss for interfering signals: Among them, J represents the interference signal recovered by using the mask matrix, and N is the length of the interference signal; L1 error loss for the target signal: Among them, y represents the non-interference target signal finally output by the model, and N is the length of the target signal; Hilbert transform loss: Among them, Hilbert() is the Hilbert transform function; When the network loss basically converges, the final deep network model is obtained; S6. Using the obtained deep network model to process the radar received signal to realize anti-interference from radar main lobe forwarding. 2. A kind of anti-radar main lobe forwarding interference method based on reversible residual network according to claim 1, it is characterized in that, the interference signal obtained by radar main lobe forwarding interference simulation in S1 comprises intermittent sampling forwarding interference, smart Noise interference, spectral dispersion interference and comb spectrum interference.