Distributed radar multi-interference identification method based on Transform network

Abstract

The invention provides a distributed radar multi-interference identification method based on a Transform network, and the method aims at a multi-interference coexistence scene, is based on a long baseline distributed radar configuration, and comprises the steps: firstly estimating an interference source angle by each radar through employing a multi-signal classification (MUSIC) algorithm, carrying out the self-adaptive beam forming, separating each interference echo from an airspace, and carrying out the self-adaptive beam forming; and on the basis, correlation coefficients between interference echoes are calculated to realize interference sample correlation. And then converting the interference echo of each radar node into a time-frequency graph by using short-time Fourier transform, inputting the time-frequency graph into a Transform network to detect the interference type and estimate the interference working parameter based on the relative position of the interference in the time-frequency graph, and finally fusing the multi-node sensing information to output the interference type and the fused interference parameter. According to the method, multi-interference-source identification and parameter estimation can be realized, interference identification is carried out after the interference samples are separated in the spatial domain for the multi-interference-source overlapping problem in the time-frequency domain, and then multi-interference-source identification and parameter estimation are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of interference source identification, and in particular relates to a distributed radar multi-interference identification method based on a Transformer network. Background Art

[0002] The electromagnetic environment of modern battlefields is becoming increasingly complex, and active interference with variable parameters and complex forms has seriously restricted the working efficiency of radars. Accurate interference identification can provide prior information for subsequent targeted interference suppression measures, which is the premise and basis for good anti-interference effects.

[0003] The main steps of the interference identification algorithm include feature extraction and interference classification. Among them, the feature extraction method extracts the characteristic parameters of the interference signal based on the transformation domains such as time domain, frequency domain, time-frequency domain, polarization domain, etc. Common features include time domain moment skewness, envelope fluctuation, information entropy, time-frequency diagram, etc. Interference classification uses the feature differences corresponding to different interferences to identify interference. Traditional classifiers distinguish interference by manually defining thresholds, but under different interference-to-noise ratios, the interference identification performance is greatly affected by the threshold. Existing classifiers implement interference classification based on artificial intelligence algorithms such as decision trees, support vector machines, and convolutional neural networks.

[0004] When facing a single interference, the traditional single radar can achieve good results based on the existing interference identification algorithm, but when there are multiple interferences overlapping in the transform domain, its identification performance drops significantly. Distributed radar can obtain multi-perspective observation information by deploying multiple unit radars to collaboratively detect targets, and has great potential in interference identification. Summary of the invention

[0005] In view of this, an object of the present invention is to provide a distributed radar multi-interference identification method based on Transformer network.

[0006] A distributed radar multi-interference identification method, comprising:

[0007] Step 1: Build a distributed system consisting of one main radar and multiple auxiliary radars. The distributed system adopts a one-transmit-multiple-receive mechanism, where the main radar transmits signals and all radars receive echoes.

[0008] Step 2: Estimate the number of interference sources using the minimum description length (MDL) criterion, and then use the MUSIC algorithm to measure the interference source angle. After separating the interference samples based on spatial adaptive beamforming, realize the interference correlation between multiple radar nodes through cross-correlation processing, that is, find the echoes belonging to the same interference source in the interference echoes of each radar and perform correlation matching;

[0009] Step 3: using the interference echoes after correlation matching, identify the interference type of each interference echo of each radar, and estimate the interference parameters;

[0010] Step 4: For each interference, based on the interference type identification results of each radar, select the output with the largest number of times to obtain the interference type after multi-radar fusion;

[0011] For each interference, based on the interference parameter estimation results of the radar nodes identified as the correct interference type, the fused parameters of the interference are obtained.

[0012] Preferably, in step 1: assume that each radar is a planar array with M array elements, of which M is the number of azimuth array elements. x , the number of array elements in the elevation direction is M y , when there are K interference sources, the element-level interference echo of each radar is expressed as:

[0013]

[0014] Among them, s 0 (t) represents the interference echo reaching the main radar; η k,i represents the amplitude coefficient of the kth interference signal received by the i-th radar; j k (t) represents the kth interference source signal; τ k,i represents the time delay of the received kth interference signal reaching the ith radar; n i (t) is the zero-mean complex Gaussian white noise of the ith radar, is the array steering vector of the kth interference signal, expressed as:

[0015]

[0016]

[0017]

[0018] in, are the corresponding steering vectors in azimuth and elevation respectively, is the Kronecker product, λ is the signal wavelength, x m ,y m are the azimuth distance and elevation distance of the mth array element relative to the phase center of the array.

[0019] Preferably, in step 2: when estimating the number of interference sources, first calculate the echo covariance matrix R 0 =E{s 0 s 0 H} Perform eigendecomposition to obtain M eigenvalues ​​λ m ,m=0,1,...,M-1, the cost function based on the MDL criterion is expressed as:

[0020]

[0021]

[0022] Where N is the number of sources to be estimated, L is the number of snapshots; N' is the N that minimizes MDL(N), and the number of interference sources is N MDL =N'-1.

[0023] Preferably, in step 2: the association matching method includes:

[0024] From λ m Get N MDL The eigenvectors corresponding to the largest eigenvalues ​​constitute the interference signal subspace U j , the remaining eigenvectors form the noise subspace U n , then the covariance matrix is ​​expressed as:

[0025]

[0026] Among them, Λ j ,Λ n is the diagonal matrix of the eigenvalues ​​of the interference signal and noise; since R 0 is a Hermitian matrix, and the interference signal corresponding to the steering vector and the interference signal subspace satisfies a linear transformation, then the steering vector is orthogonal to the noise subspace, so the MUSIC space spectrum is used:

[0027]

[0028] Perform spectrum peak search to obtain the interference source azimuth With pitch angle k=1,2,...,N MDL ;

[0029] Then based on Adaptive beamforming is performed using MVDR; the beamformer minimizes the total output power while ensuring that the desired angle reception is not distorted, and the weight vector is:

[0030]

[0031] Beamforming is performed on each interference separately to obtain Each radar performs the above operations and obtains the interference echo after beamforming. i=0,1,…,P-1;

[0032] Finally, normalize each interference echo to get Calculate the cross-correlation function between each interference echo of each auxiliary radar and the kth interference echo of the main radar and take the maximum value, which is expressed as:

[0033]

[0034] Where l = 1, 2, ..., N MDL ; When l = k, The maximum value is obtained, which means that the interference echo of the auxiliary radar matches the kth interference echo of the main radar, that is, it comes from the same interference source as the kth interference echo; MDL The interference echoes of each auxiliary radar are used as references, and the maximum value of the cross-correlation function is calculated with each interference echo of each auxiliary radar, so that the interference echoes belonging to the same interference source are matched, that is, the interference echo matching between the auxiliary radar and the main radar is achieved.

[0035] Preferably, in step 3: the interference type identification method includes:

[0036] First, the interference echoes of each radar after matching are short-time Fourier transformed:

[0037]

[0038] Where w(t) represents the Hamming window, and τ represents the center position of the window function. Then, the time-frequency graphs are input into the interference identification network respectively. The interference time-frequency graph first passes through the Swin Transformer backbone feature extraction network to output a shared feature graph. Then, the proposal box generation network convolves the feature graph to preliminarily estimate the interference area proposal frame. Then, the coarsely estimated proposal frame and the shared feature map are input into the ROI pooling layer. The pooling layer intercepts the feature map based on the proposal frame and performs regional pooling to obtain a local feature map of the same size. On this basis, a fully connected layer is used to classify the interference type and predict the interference area of ​​the target in the time-frequency graph.

[0039] Preferably, in step 3: the interference parameter estimation method includes: let W represent the width of the spectrum graph during prediction, H represent the height of the spectrum graph during prediction, T total Indicates the total length of the time axis of the time-frequency diagram, F total Represents the total length of the frequency axis. The parameters of each active interference type are calculated as follows:

[0040] 1) Interference pulse width

[0041]

[0042] in, and They represent the leftmost and rightmost x-axis coordinates of the region where the first interference slice is located;

[0043] 2) Interference bandwidth

[0044]

[0045] in, and They respectively represent the leftmost and rightmost y-axis coordinates of the region where the first interference slice is located;

[0046] 3) Interference frequency

[0047]

[0048] Among them, f 0 is the radar carrier frequency, △F is the minimum interval of the frequency axis of the time-frequency diagram;

[0049] 4) Interference pulse repetition period

[0050]

[0051] in, is the leftmost x-axis coordinate of the area where the second interference slice is located.

[0052] Preferably, in step 4: based on the interference estimation parameters of the radar node identified as the correct interference type, outliers that deviate from the mean by twice the variance are eliminated for each interference parameter, and then the remaining parameters are averaged and fused to obtain the fused parameters of the interference.

[0053] The present invention has the following beneficial effects:

[0054] The present invention proposes a distributed radar multi-interference identification method based on a Transformer network. The method is aimed at scenarios where multiple interferences coexist. Based on the long baseline distributed radar configuration, each radar first estimates the interference source angle using the multiple signal classification (MUSIC) algorithm and performs adaptive beamforming, separates each interference echo from the spatial domain, and on this basis calculates the correlation coefficient between the interference echoes to achieve interference sample association. Then, the interference echo of each radar node is converted into a time-frequency diagram using a short-time Fourier transform, input into the Transformer network to detect the interference type, and estimate the interference working parameters based on the relative position of the interference in the time-frequency diagram. Finally, the multi-node perception information is fused to output the interference type and the fused interference parameters. The present invention can realize multi-interference source identification and parameter estimation. For the problem of multiple interference sources overlapping in the time-frequency domain, interference identification is performed after the interference samples are separated in the spatial domain, thereby realizing multi-interference source identification and parameter estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of the transformer-based distributed radar multi-interference identification method;

[0056] Figure 2 This is a schematic diagram of the distributed radar interference source recognition system;

[0057] Figure 3 Identify network structures for interference;

[0058] Figure 4 It is the Swin Transformer feature extraction network structure;

[0059] Figure 5 Mapping network output results to parameter values;

[0060] Figure 6 It is a flow chart of multi-node interference perception type and parameter fusion method;

[0061] Figure 7 shows the interference sample sorting results in the simulation example. Figure 7(a) is the time-frequency diagram of the main radar interference sample before sorting, and Figure 7(b)(c)(d)(e)(f) are the 5 interference samples of the main radar after sorting.

[0062] Figure 8(a) shows the changes in loss value and learning rate during training in a simulation example; Figure 8(b) shows the changes in mAP during training;

[0063] Figure 9(a), Figure 9(b), Figure 9(c), Figure 9(d), and Figure 9(e) are the interference identification results of the main radar under five interference sources in the simulation example;

[0064] Fig.10 It is the distributed radar configuration (northeast celestial coordinate system) in the embodiment of the present invention. DETAILED DESCRIPTION

[0065] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0066] A distributed radar multi-interference identification method based on Transformer network, the flow chart is as follows Figure 1 As shown, the specific steps include:

[0067] Step 1: Receive multiple interference echoes

[0068] A distributed system consisting of a main radar and P-1 auxiliary radars is constructed. The system adopts a one-transmit-multiple-receive mechanism, in which the main radar transmits signals and all radars receive echoes. Figure 2 shown.

[0069] Assume that each radar is a planar array with M array elements, of which M is the number of azimuth array elements. x , the number of array elements in the elevation direction is M y , when there are K interference sources, the element-level interference echo of each radar can be expressed as

[0070]

[0071] Among them, s0 (t) represents the interference echo reaching the main radar; η k,i represents the amplitude coefficient of the kth interference signal received by the i-th radar; j k (t) represents the kth interference source signal; τ k,i represents the time delay of the received kth interference signal reaching the ith radar; n i (t) is the zero-mean complex Gaussian white noise of the ith radar, is the array steering vector of the kth interference signal, expressed as:

[0072]

[0073]

[0074]

[0075] in, are the corresponding steering vectors in azimuth and elevation respectively, is the Kronecker product, λ is the signal wavelength, x m ,y m are the azimuth distance and elevation distance of the mth array element relative to the phase center of the array.

[0076] Step 2: Interference sample association

[0077] The number of interference sources is estimated by the minimum description length (MDL) criterion, and then the interference source angle is measured using the MUSIC algorithm. The interference samples are separated based on spatial adaptive beamforming, and the interference correlation between multiple radar nodes is achieved through cross-correlation processing.

[0078] First, estimate the number of interference sources. Taking the main radar as an example, first calculate the echo covariance matrix R 0 =E{s 0 s 0 H} Perform eigendecomposition to obtain M eigenvalues ​​λ m ,m=0,1,...,M-1, the cost function based on the MDL criterion can be expressed as:

[0079]

[0080]

[0081] Where N is the number of sources to be estimated, and L is the number of snapshots. Let N' be the N that minimizes MDL(N), then the number of interference sources is N MDL =N'-1.

[0082] From λ m Get N MDLThe eigenvectors corresponding to the largest eigenvalues ​​constitute the interference signal subspace U j , the remaining eigenvectors form the noise subspace U n , then the covariance matrix can be expressed as

[0083]

[0084] Among them, Λ j ,Λ n is the diagonal matrix of the eigenvalues ​​corresponding to the interference signal and noise. 0 is a Hermitian matrix, and the interference signal corresponding to the steering vector and the interference signal subspace satisfy a linear transformation, then the steering vector is orthogonal to the noise subspace. Therefore, using the MUSIC space spectrum:

[0085]

[0086] Perform spectrum peak search to obtain the interference source azimuth With pitch angle k=1,2,...,N MDL .

[0087] Then based on Adaptive beamforming is performed using MVDR. The beamformer minimizes the total output power while ensuring that the desired angle reception is not distorted. The weight vector is

[0088]

[0089] Beamforming is performed on each interference separately to obtain k=1,2,...,N MDL After all radars perform the above operations, the interference echo after beamforming is obtained. i=0,1,…,P-1.

[0090] Finally, normalize each interference echo to get Calculate the cross-correlation function between each interference echo of each auxiliary radar and the kth interference echo of the main radar and take the maximum value, which is expressed as

[0091]

[0092] Where l = 1, 2, ..., N MDL When l = k, The maximum value is obtained, which means that the interference echo of the auxiliary radar matches the kth interference echo of the main radar, that is, the interference echo comes from the same interference source. MDLThe interference echoes of each auxiliary radar are used as references, and the maximum value of the cross-correlation function is calculated with each interference echo of each auxiliary radar, so that the interference echoes belonging to the same interference source are matched, that is, the interference echo matching between the auxiliary radar and the main radar is achieved.

[0093] Step 3: Interference identification and parameter estimation

[0094] The interference echo is converted into a time-frequency diagram through short-time Fourier transform, and then the Transformer network is used for interference identification and parameter estimation.

[0095] First, the interference echoes of each radar after matching are short-time Fourier transformed:

[0096]

[0097] Where w(t) represents the Hamming window and τ represents the center position of the window function. Then the time-frequency graphs are input into the interference identification network respectively.

[0098] Interference identification network structure Figure 3 As shown in the figure. The interference time-frequency map first passes through the Swin Transformer backbone feature extraction network to output a shared feature map, and then the Region Proposal Network (RPN) convolves the feature map to preliminarily estimate the interference region proposal box, and then the roughly estimated proposal box and the shared feature map are input into the ROI (Region of Interest) pooling layer. The pooling layer intercepts the feature map based on the proposal box and performs regional pooling to obtain a local feature map of the same size. On this basis, the fully connected layer is used to classify the interference type and predict the interference region of the target in the time-frequency map.

[0099] Swin Transformer feature extraction network model Figure 4 As shown in the figure, the time-frequency graph is first split into multiple non-overlapping blocks (or sub-graphs) of the same size and projected to an arbitrary dimension C using a linear embedding layer (Linear Embedding) and then input into the Swin Transformer Block module. This module contains a window multi-head self-attention layer (Window Multi-head Self-Attention, W-MSA) and a shift window multi-head self-attention layer (Shift-Window Multi-head Self-Attention, SW-MSA), which appear alternately to achieve information transfer within and between windows. W-MSA and SW-MSA divide the samples into windows and calculate their corresponding attention weights through the multi-head attention mechanism:

[0100]

[0101]

[0102] Among them, Q, K, V are obtained by taking the inner product of the input and the network feature extraction matrix, and B represents the location information of each block. M represents the image size, d k Representation vector QK T The calculation process discussed above occurs in each sub-graph. In order to better extract global features, the Swin Transformer model introduces a sliding window (shiftwindow). In this way, the relative position of the divided windows is different each time the calculation is performed. Each sub-graph will participate in the calculation process of the attention mechanism, and the network will not only "care" about local features. By shifting the feature map and setting the corresponding mask for Attention, the result of calculating the equivalent window movement while maintaining a fixed number of windows is achieved. Finally, the sub-graphs are weighted and fused based on the attention weights to obtain the shared feature map after feature extraction.

[0103] The interference parameter estimation module is mainly based on the time-frequency graph features and the neural network model prediction information, and estimates the interference parameters by geometric conversion. Taking dense false target interference as an example to analyze the network model parameter estimation results, such as Figure 5 As shown in the figure. In the figure, W represents the width of the spectrum graph during prediction, H represents the height of the spectrum graph during prediction, and T total Indicates the total length of the time axis of the time-frequency diagram, which is 4ms in the figure. total Indicates the total length of the frequency axis, which is 5MHz in the figure above (negative frequency is because the signal has changed to baseband). The parameters of each active interference type are calculated as follows:

[0104] 1) Interference pulse width

[0105]

[0106] in, and They respectively represent the leftmost and rightmost x-axis coordinates of the area where the first interference slice is located.

[0107] 2) Interference bandwidth

[0108]

[0109] in, and Represents the leftmost and rightmost y-axis coordinates of the area where the first interference slice is located.

[0110] 3) Interference frequency

[0111]

[0112] Among them, f 0 is the radar carrier frequency, △F is the minimum interval of the frequency axis of the time-frequency diagram.

[0113] 4) Interference pulse repetition period

[0114]

[0115] in, is the leftmost x-axis coordinate of the area where the second interference slice is located.

[0116] Step 4: Multi-radar node fusion

[0117] Multi-radar node interference perception type and parameter fusion method Figure 6 As shown. For the kth interference, the output with the most times is selected based on the interference identification results of each radar node to obtain the interference type after fusion of multiple radar nodes. Then, based on the interference estimation parameters of the radar nodes identified as the correct interference type, outliers with deviations of twice the variance from the mean are removed for parameters such as slice width, frequency, and bandwidth, and then the remaining parameters are averaged and fused to obtain the fused parameters of the interference.

[0118] Example:

[0119] The simulation is implemented using MATLAB R2021b software. The distributed radar parameters and waveform parameters are set as follows:

[0120] Distributed radar parameters and waveform parameters

[0121]

[0122] Consider a simulation scenario: five jammers are located in each radar beam, and the scenario is characterized by the presence of five angularly resolvable jammers.

[0123] Jammer location (polar coordinates in the northeast celestial coordinate system)

[0124] Jammer number Distance / km Azimuth / ° Pitch angle / ° 1 120 -33.0 4.0 2 300 -15.0 2.5 3 120 -7.0 6.0 4 250 7.0 3.0 5 380 31.0 1.5

[0125] Interference parameters

[0126]

[0127]

[0128] Step 1: Receive multiple interference echoes

[0129] Assume that the normalized radar baseband transmission signal is

[0130]

[0131] in is the rectangular gate function; T p is the pulse width; is the phase modulation function.

[0132] The simulation example considers three types of narrow pulse interference, dense false target interference and smart noise interference. Narrow pulse interference is also called pulse train interference. The jammer transmits a series of high-frequency, high-energy pulse trains to the radar to achieve the purpose of suppression. This type of single interference is reflected as a very narrow pulse that lasts in time. Therefore, the instantaneous power requirement of the jammer is not high, and a good coverage and suppression effect can be achieved. The mathematical expression of its signal is:

[0133]

[0134] Where T np is the pulse width of the narrow pulse, τ np is the narrow pulse repetition period, which is the time interval from a narrow pulse to the next sub-pulse, N is the number of narrow pulses, Θ τ Indicates the time when the first narrow pulse appears in a sample.

[0135] Dense false target interference is a type of interference in which a digital storage transponder intercepts and analyzes the signal in a radar PRT, and then repeats it in large quantities, which can obtain a large degree of matching gain. After the radar is processed by matched filtering, a large number of equally spaced peaks will be obtained, which can have both deception effect and coverage suppression effect. Its expression is:

[0136]

[0137] Where N is the number of false targets, △τ is the interference forwarding interval, that is, the time interval between two intensive false target interferences. Θ τ Indicates the time when the first dense false target interference appears in a certain sample.

[0138] The signal model of smart noise interference can be expressed as

[0139]

[0140] Among them, T s is the slice pulse width, N is the number of slices, and N(t) is the pulse width T s noise, s(t) represents the intercepted radar signal, * represents the convolution process, n(t) is Gaussian white noise, Θ τ It indicates the time when the first dense false target interference appears in a certain sample. The length of this time usually varies in different samples, and the jammer directly determines this variable.

[0141] Step 2: Interference sample association

[0142] First, the number of interference sources is measured to be 5 based on the MDL algorithm. Then, each radar measures the interference angle based on the MUSIC method and performs adaptive beamforming based on the angle to sort interference samples. Taking the main radar as an example, the time-frequency diagram before and after interference sample sorting is shown in Figure 7.

[0143] The interference samples of each radar are matched using cross-correlation processing to obtain the interference angle under the perspective of each radar.

[0144] Interference angle estimation results

[0145]

[0146]

[0147] Step 3: Interference identification and parameter estimation

[0148] The data set used in the present invention contains single interference and compound interference samples of the above three types of interference. The compound interference is formed by combining different single interference types in pairs. Each data sample is a time-frequency diagram obtained by short-time Fourier transform, and its size is 201*400. The total number of samples is 3000, which is annotated according to the interference position. The annotation format refers to the standard xml file format of the VOC data set.

[0149] The present invention uses mAP (mean Average Precision) as the evaluation index for model training, which is obtained by weighted average of the average precision (AP) of all categories. AP is calculated by accuracy and recall, which represents the percentage of correctly identified objects to the total number of identified objects. The higher the AP, the better the target detection effect. The definition of precision and recall is derived from the intersection over union (IOU). IOU is calculated by dividing the intersection of the true frame and the predicted frame by the union. The larger the IOU, the closer the predicted frame is to the true frame. Generally, when IOU is greater than 0.5, the system determines it as a correct detection frame, and when it is less than 0.5, it is an incorrect detection frame, and then the accuracy can be calculated.

[0150]

[0151] Among them, TP represents the number of correctly detected boxes, and FP represents the number of incorrectly detected boxes. Recall refers to the number of correctly predicted positive samples in the prediction data set divided by the number of actual positive samples. Its calculation process is expressed as

[0152]

[0153] Among them, FN represents the number of missed inspection boxes. After obtaining the precision and recall rate, we can draw the PR curve to obtain AP

[0154]

[0155] Then the labeled samples are sent to the neural network model for training. The training hardware equipment is a computer equipped with Windows system, the processor is Intel(R) Core(TM) i7-10875H CPU@2.30GHz, equipped with NVIDIA GeForce RTX 2060 graphics card with 16GB RAM and 6GB video memory, and the training network software configuration environment is Pytorch1.7.1, CUDA 11.0. The fine-tuning batch size (batch_size) is set to 4, the number of iterations is 50, and it takes about 7 hours. The changes in the loss value (loss), learning rate (learning rate), and validation set mAP (IOU=0.5) indicators during training are shown in Figure 8.

[0156] Finally, the trained neural network is used to identify the interference samples of each radar and estimate the parameters. The recognition result of the main radar is shown in Figure 9.

[0157] Step 4: Multi-radar node fusion

[0158] For each interference node, the output with the most number of interference identification results is selected to obtain the interference type after multi-node fusion. Then, the interference estimation parameters of the radar nodes identified as the correct interference type are averaged after removing outliers. The fused interference identification results and parameter estimation are shown in the following table. Among them, the average estimation errors of pulse width, frequency, bandwidth and forwarding period are 6.42%, 0.04%, 6.20% and 2.81% respectively.

[0159] Distributed radar interference identification and parameter estimation results (interference type, pulse width, frequency)

[0160]

[0161] Distributed radar interference identification and parameter estimation results (bandwidth, forwarding period)

[0162]

[0163] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A distributed radar multi-interference identification method, characterized in that: include: Step 1: Build a distributed system consisting of one main radar and multiple auxiliary radars. The distributed system adopts a one-transmit-multiple-receive mechanism, where the main radar transmits signals and all radars receive echoes. Step 2: Estimate the number of interference sources using the minimum description length (MDL) criterion, and then use the MUSIC algorithm to measure the interference source angle. After separating the interference samples based on spatial adaptive beamforming, realize the interference correlation between multiple radar nodes through cross-correlation processing, that is, find the echoes belonging to the same interference source in the interference echoes of each radar and perform correlation matching; Step 3: using the interference echoes after correlation matching, identify the interference type of each interference echo of each radar, and estimate the interference parameters; Step 4: For each interference, based on the interference type identification results of each radar, select the output with the largest number of times to obtain the interference type after multi-radar fusion; For each interference, based on the interference parameter estimation results of the radar nodes identified as the correct interference type, the fused parameters of the interference are obtained.

2. A distributed radar multi-interference identification method as claimed in claim 1, characterized in that: In the step 1, it is assumed that each radar is a planar array with M array elements, of which M is the number of azimuth array elements. x , the number of array elements in the elevation direction is M y , when there are K interference sources, the element-level interference echo of each radar is expressed as: Where s0(t) represents the interference echo reaching the main radar; ηk,i represents the amplitude coefficient of the kth interference signal received by the i-th radar; j k (t) represents the kth interference source signal; τ k,i represents the time delay of the received kth interference signal reaching the ith radar; n i (t) is the zero-mean complex Gaussian white noise of the ith radar, is the array steering vector of the kth interference signal, expressed as: in, are the corresponding steering vectors in azimuth and elevation respectively, is the Kronecker product, λ is the signal wavelength, x m ,y m are the azimuth distance and elevation distance of the mth array element relative to the phase center of the array.

3. A distributed radar multi-interference identification method as claimed in claim 2, characterized in that: In the step 2, when estimating the number of interference sources, firstly calculate the echo covariance matrix R0=E{s0s0 H } Perform eigendecomposition to obtain M eigenvalues ​​λ m ,m=0,1,...,M-1, the cost function based on the MDL criterion is expressed as: Where N is the number of sources to be estimated, L is the number of snapshots; N' is the N that minimizes MDL(N), and the number of interference sources is N MDL =N'-1.

4. A distributed radar multi-interference identification method as claimed in claim 3, characterized in that: In the step 2: The association matching method includes: From λ m Get N MDL The eigenvectors corresponding to the largest eigenvalues ​​constitute the interference signal subspace U j , the remaining eigenvectors form the noise subspace U n , then the covariance matrix is ​​expressed as: Among them, Λ j ,Λ n is the diagonal matrix of the eigenvalues ​​corresponding to the interference signal and noise; since R0 is a Hermitian matrix, and the steering vector corresponding to the interference signal satisfies a linear transformation with the interference signal subspace, the steering vector is orthogonal to the noise subspace, so the MUSIC space spectrum is used: Perform spectrum peak search to obtain the interference source azimuth With pitch angle Then based on Adaptive beamforming is performed using MVDR; the beamformer minimizes the total output power while ensuring that the desired angle reception is not distorted, and the weight vector is: Beamforming is performed on each interference separately to obtain Each radar performs the above operations and obtains the interference echo after beamforming. Finally, normalize each interference echo to get Calculate the cross-correlation function between each interference echo of each auxiliary radar and the kth interference echo of the main radar and take the maximum value, which is expressed as: Where l = 1, 2, ..., N MDL ; When l = k, r i k,l The maximum value is obtained, which means that the interference echo of the auxiliary radar matches the kth interference echo of the main radar, that is, it comes from the same interference source as the kth interference echo; MDL The interference echoes of each auxiliary radar are used as references, and the maximum value of the cross-correlation function is calculated with each interference echo of each auxiliary radar, and the interference echoes belonging to the same interference source are matched, that is, the interference echo matching between the auxiliary radar and the main radar is achieved.

5. A distributed radar multi-interference identification method as claimed in claim 4, characterized in that: In step 3: the interference type identification method includes: First, the interference echoes of each radar after matching are short-time Fourier transformed: Where w(t) represents the Hamming window, and τ represents the center position of the window function. Then, the time-frequency graphs are input into the interference identification network respectively. The interference time-frequency graph first passes through the Swin Transformer backbone feature extraction network to output a shared feature graph. Then, the proposal box generation network convolves the feature graph to preliminarily estimate the interference area proposal frame. Then, the coarsely estimated proposal frame and the shared feature map are input into the ROI pooling layer. The pooling layer intercepts the feature map based on the proposal frame and performs regional pooling to obtain a local feature map of the same size. On this basis, a fully connected layer is used to classify the interference type and predict the interference area of ​​the target in the time-frequency graph.

6. A distributed radar multi-interference identification method as claimed in claim 5, characterized in that: In step 3: the interference parameter estimation method includes: let W represent the width of the spectrum graph during prediction, H represent the height of the spectrum graph during prediction, T total Indicates the total length of the time axis of the time-frequency diagram, F total Represents the total length of the frequency axis. The parameters of each active interference type are calculated as follows: 1) Interference pulse width in, and They respectively represent the leftmost and rightmost x-axis coordinates of the region where the first interference slice is located; 2) Interference bandwidth in, and They respectively represent the leftmost and rightmost y-axis coordinates of the region where the first interference slice is located; 3) Interference frequency Among them, f0 is the radar carrier frequency, △F is the minimum interval of the frequency axis of the time-frequency diagram; 4) Interference pulse repetition period in, is the leftmost x-axis coordinate of the area where the second interference slice is located.

7. A distributed radar multi-interference identification method as claimed in claim 6, characterized in that: In the step 4: based on the interference estimation parameters of the radar node identified as the correct interference type, outliers that deviate from the mean by twice the variance are removed for each interference parameter, and then the remaining parameters are averaged and fused to obtain the fused parameters of the interference.

Images